Prosecution Insights
Last updated: November 16, 2026
Application No. 19/153,950

A Multi-Profile Fraud Detection System For Detecting Suspicious Airline Ticket Sales Activities

Non-Final OA §101§103
Filed
Aug 05, 2025
Priority
Oct 09, 2023 — TÜ 2023/012715 +1 more
Examiner
HASBROUCK, MERRITT J
Art Unit
3695
Tech Center
3600 — Transportation & Electronic Commerce
Assignee
Yildiz Teknoloji Transfer Ofisi Anonim Sirketi
OA Round
1 (Non-Final)
10%
Grant Probability
At Risk
1-2
OA Rounds
18%
With Interview

Examiner Intelligence

Grants only 10% of cases
10%
Career Allowance Rate
16 granted / 156 resolved
-41.7% vs TC avg
Moderate +8% lift
Without
With
+7.8%
Interview Lift
resolved cases with interview
Typical timeline
Avg Prosecution
30 currently pending
Career history
203
Total Applications
across all art units

Office Action

§101 §103
DETAILED ACTION Notice of Pre-AIA or AIA Status The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA . Claims 1-20 have been examined in this application. This communication is the first action on the merits. Priority Application 19/153,950 filed on 08/05/2025 is a 371 of PCT/TR2024/051123 10/01/2024 and claims benefit of TÜRKIYE 2023/012715 10/09/2023. Examiner Request The Applicant is requested to indicate where in the specification there is support for amendments to claims should Applicant amend. The purpose of this is to reduce potential 35 U.S.C. § 112(a) or § 112 1st paragraph issues that can arise when claims are amended without support in the specification. The Examiner thanks the Applicant in advance. Claim Rejections - 35 USC § 101 35 U.S.C. § 101 reads as follows: Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title. Claims 1-20 are rejected under 35 U.S.C. § 101 because the claimed invention is directed to a judicial exception (i.e., a law of nature, a natural phenomenon, or an abstract idea) without significantly more. (MPEP 2106). The claims are directed to a method, system, and apparatus which is one of the statutory categories of invention (Step 1: YES). The recitation of the claimed invention is analyzed as follows, in which the abstract elements are boldfaced. Claim 1 recites the limitations of: A computer-implemented method for detecting fraud activities in airline ticket and service sales, comprising the steps of: a. receiving a dataset comprising a series of airline ticket and service sale requests and sale transactions, b. receiving profiles comprising past transactions in chronological order of all objects included in each transaction in the dataset, c. updating the received profiles by adding a record of the last transaction, d. generating a profile-based dataset, e. feeding the profile-based dataset into a balanced random undersampling algorithm in order to obtain an undersampled dataset in which legitimate and fraudulent transactions are distributed in a balanced manner, f. feeding the dataset into a deep neural network, g. generating a complexity matrix for the outputs of the deep neural network, h. generating performance metrics from the complexity matrix, i. training the deep neural network with feedback of the performance metrics. The claim as a whole recites a method that, under its broadest reasonable interpretation, covers collecting, analyzing, and transmitting data to facilitate determining fraudulent activity in a financial transaction. This is a fundamental economic practice of a financial transaction; a commercial interaction, such as for business relations; and managing personal behavior or relationships or interactions between people, which are certain methods of organizing human activity. Furthermore, the claims cover the use of a computer for collecting, analyzing, and transmitting data to facilitate determining fraudulent activity in a financial transaction. As the steps could be performed by a human without a computer, the claim limitations fall within the mental processes grouping, and the claim recites an abstract idea. The claims also recite the use of a balanced random undersampling algorithm to facilitate determining fraudulent activity in a financial transaction. This is a mathematical concept or calculation. In the alternative, the balanced random undersampling algorithm is considered a technology that is recited at a high level of generality and merely applied as a tool to implement the abstract idea. Thus, the claims recite an abstract idea. (Step 2A, prong 1: YES). Moreover, the judicial exception is not integrated into a practical application. Other than reciting a “computer” and “deep neural network”, to perform the steps of “updating” and “generating”, nothing in the claim elements preclude the steps from practically being a certain method of organizing human activity or mental process or mathematical concept. The claim as a whole does not integrate the judicial exception into a practical application. The claim merely describes how to generally “apply” the concept of collecting, analyzing, and transmitting data to facilitate determining fraudulent activity in a financial transaction in a computer environment. The additional computer elements recited in the claim limitations are recited at a high-level of generality such that it amounts to no more than mere instructions to apply the exception utilizing generic computer components. For example, the specification discloses “[Spec. at pp. 2-3] An object of the present invention is to provide a method for detecting fraudulent transactions in airline ticket and service sales. In particular, the object is to provide a computer-implemented fraud detection method that uses historical data relations, which allows to overcome the performance limitations caused by imbalanced transactional datasets.” “In the fraud detection method of the invention, machine learning and deep learning techniques are used in combination with a profile-based approach to overcome the constraints of transactional datasets.” Thus, the specification supports that general purpose computers or computer components are utilized to implement the steps of the abstract idea. Merely implementing the abstract idea on a generic computer is not a practical application of the abstract idea. The claim as a whole, in viewing the additional elements both individually and in combination, does not integrate the judicial exception into a practical application. Additionally, with respect to the balanced random undersampling algorithm, patents that do no more than claim the application of generic machine learning to new data environments, without disclosing improvements to the machine learning models to be applied, are patent ineligible under § 101. (See Recentive Analytics, Inc. v. Fox Corp., 134 F.4th 1205). Accordingly, these additional elements do not integrate the abstract idea into a practical application because it does not impose any meaningful limits on practicing the abstract idea. The claim is directed to an abstract idea. (Step 2A prong two: No) The claim does not include additional elements, when considered both individually and as an ordered combination, that are sufficient to amount to significantly more than the judicial exception. As discussed above with respect to integration of the abstract idea into a practical application, the additional elements of using “computer” and “deep neural network”, to perform the steps of “updating” and “generating”, amounts to no more than mere instructions to apply the exception using generic computer component. The claim merely describes how to generally “apply” the concept of collecting, analyzing, and transmitting data to facilitate determining fraudulent activity in a financial transaction in a computer environment. Thus, even when viewed as a whole, nothing in the claim adds significantly more (i.e. an inventive concept) to the abstract idea. Such additional elements are determined to not contain an inventive concept according to MPEP 2106.05(f). It should be noted that (1) the “recitation of claim limitations that attempt to cover any solution to an identified problem with no restriction on how the result is accomplished and no description of the mechanism for accomplishing the result, does not provide significantly more because this type of recitation is equivalent to the words “apply it”, and (2) “Use of a computer or other machinery in its ordinary capacity for economic or other tasks (e.g., to receive, store, or transmit data) or simply adding a general purpose computer or computer components after the fact to an abstract idea (e.g., a fundamental economic practice, commercial interaction, or managing personal behavior or relationships or interactions between people, mental process, or mathematical calculation) does not integrate a judicial exception into a practical application or provide significantly more”. See MPEP 2106.05(g). Dependent claims 2-20 merely limit the abstract idea and do not recite any further additional elements beyond the cited abstract idea and the elements addressed above, thus, they do not amount to significantly more. The dependent claims are abstract for the reasons presented above because there are no additional elements that integrate the abstract idea into a practical application or are sufficient to amount to significantly more than the judicial exception when considered both individually and as an ordered combination. Thus, the dependent claims are directed to an abstract idea. (Step 2B: No) Therefore, claims 1-20 are not patent-eligible. Claim Rejections - 35 USC § 103 In the event the determination of the status of the application as subject to AIA 35 U.S.C. §§ 102 and 103 (or as subject to pre-AIA 35 U.S.C. §§ 102 and 103) is incorrect, any correction of the statutory basis for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status. The following is a quotation of 35 U.S.C. § 103 which forms the basis for all obviousness rejections set forth in this Office action: A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made. Claims 1-5 and 8-18 are rejected under 35 U.S.C. 103 as being unpatentable over Florimond, U.S. Patent Application Publication Number 2017/0262852; in view of Reddy, U.S. Patent Application Publication Number 2023/0351324; in view of Rashid, “Developing an Artificial Intelligence based system to Detect Fraud and Corruption in Government”, NPL Publication 2022; in view of Seo, WIPO. Patent Application Publication Number 2018/097653; in view of Gao, U.S. Patent Application Publication Number 2023/0342993. As per claim 1, Florimond explicitly teaches: A computer-implemented method for detecting fraud activities in airline ticket and service sales, comprising the steps of: a. receiving a dataset comprising a series of airline ticket and service sale requests and sale transactions, (Florimond US20170262852 at paras. 5-7, 17-18, 41-51) ("[0005] In an embodiment of the invention, a system is provided that includes one or more processors and a memory coupled to the processor. The memory includes program code that, when executed by the one or more processors, causes the system to detect a change to a travel record after an itinerary defined in the travel record has been booked, and in response to detecting the change, determine a pattern of changes to the travel record. The program code further causes the system to determine if the pattern matches a potentially fraudulent pattern, and in response to the pattern matching the potentially fraudulent pattern, flag the travel record as potentially fraudulent." "[0018] Referring now to FIG. 1, an operating environment 10 in accordance with an embodiment of the invention may include a Global Distribution System (GDS) 12, a reservation system 14, a seller system 16, a payment system 18, a database monitoring system 20, a fraud screening system 22, and a travel record database 24. Each of the GDS 12, reservation system 14, seller system 16, payment system 18, database monitoring system 20, fraud screening system 22, and travel record database 24 may communicate through a network 26. The reservation system 14 may include a Computer Reservation System (CRS) that enables the GDS 12 or seller system 16 to reserve and pay for airline tickets. The reservation system 14 may also interact with other reservation systems (not shown), either directly or through the GDS 12, to enable a validating carrier to sell tickets for seats provided by an operating carrier. The operating carrier may then bill the validating carrier for the products provided. Billing between sellers and travel product providers may be provided or otherwise facilitated by the payment system 18. The network 26 may include one or more private or public networks (e.g., the Internet) that enable the exchange of data between systems.") b. receiving profiles comprising past transactions.[in chronological order] of all objects included in each transaction in the dataset, (Florimond US20170262852 at paras. 5-7, 17-18, 38, 41-51) ("[0017] Embodiments of the invention are directed to systems, methods, and computer program products that determine whether a previously completed transaction for a product is fraudulent. This determination may involve identifying a pattern of post-purchase changes to one or more database records that is indicative of fraud. Embodiments of the invention may be implemented by a database monitoring system comprising one or more networked computers or servers. The networked computers may include a Global Distribution System (GDS), and may provide processing and database functions for travel-related systems and modules that analyze database records to identify transactions that may be fraudulent. The database records may include, for example, travel records such as Passenger Name Records (PNRs), payment records, ticket documents, and/or customer profiles. If the database monitoring system detects suspicious post-booking reservation activity, the system may flag the transaction as potentially fraudulent, request additional fraud screening, and/or cancel the booking." "[0038] The machine learning engine 70 may be configured to operate in a learning mode and a detection mode. Fraud detection by the machine learning engine 70 in the detection mode may be performed by a common traversal framework. This framework may allow different customers to define different typologies of fraud for different impacted areas. The framework may also automatically identify travel records which appear to be fraudulent using a machine learning algorithm, e.g., a neural network training algorithm. In addition to travel records, the fraud analysis topologies may analyze other types of records, such as payment records and user profiles, to further refine detection of fraudulent patterns." "[0051] To configure the machine learning engine 70 to identify fraudulent patterns of booking, a custom neural network may be constructed using techniques such as pre-training by auto-encoding. This may be accomplished, for example, by analyzing fraud results for historical transactions that include transactions in which a chargeback was received (i.e., known fraudulent transactions), and transactions for which a chargeback was not received (i.e., known legitimate transactions). An exemplary historical set of samples may include a number of known fraudulent transactions (e.g., 4500) and a number of known legitimate transactions (e.g., 100,000). From this historical set of sample of transactions, a training set may be defined that includes a number of known fraudulent transactions (e.g., 4000), a number of known legitimate transactions (e.g., 8000), and a testing set including a number of transactions (e.g., 4000) randomly selected from the historical sample. The post analysis engine 68 may retrieve features from transactions to use as inputs to the neural network 80. These features may include general information such as the ip-based city or region, and may include a large number (e.g., 732) of binary variables.") legitimate and fraudulent transactions... (Florimond US20170262852 at paras. 5-7, 17-18, 41-51) ("[0051] To configure the machine learning engine 70 to identify fraudulent patterns of booking, a custom neural network may be constructed using techniques such as pre-training by auto-encoding. This may be accomplished, for example, by analyzing fraud results for historical transactions that include transactions in which a chargeback was received (i.e., known fraudulent transactions), and transactions for which a chargeback was not received (i.e., known legitimate transactions). An exemplary historical set of samples may include a number of known fraudulent transactions (e.g., 4500) and a number of known legitimate transactions (e.g., 100,000). From this historical set of sample of transactions, a training set may be defined that includes a number of known fraudulent transactions (e.g., 4000), a number of known legitimate transactions (e.g., 8000), and a testing set including a number of transactions (e.g., 4000) randomly selected from the historical sample. The post analysis engine 68 may retrieve features from transactions to use as inputs to the neural network 80. These features may include general information such as the ip-based city or region, and may include a large number (e.g., 732) of binary variables.") f. feeding the dataset into a deep neural network, (Florimond US20170262852 at paras. 5-7, 17-18, 41-51) ("[0049] The weights of the links 92, 94 connecting the nodes 88 of hidden layer 84 to the nodes 88 of input and output layers 82, 86 may be adjusted by training algorithms that optimize the output to provide known correct results (i.e., fraud=true/false) in response to the input parameters (e.g., name change event=true/false) that produced the known result. For example, the post analysis engine 68 may use a deep learning neural network algorithm to train the neural network 80. [0050] The database monitoring engine 60 may use the machine learning engine 70 to recognize patterns and evaluate a probability that the recognized pattern is indicative of fraud. Because the machine learning engine 70 may effectively learn complex patterns autonomously, once trained, the machine learning engine 70 may operate as a “black-box” that predicts fraud based on the booking record and/or payment record feeds. [0051] To configure the machine learning engine 70 to identify fraudulent patterns of booking, a custom neural network may be constructed using techniques such as pre-training by auto-encoding. This may be accomplished, for example, by analyzing fraud results for historical transactions that include transactions in which a chargeback was received (i.e., known fraudulent transactions), and transactions for which a chargeback was not received (i.e., known legitimate transactions). An exemplary historical set of samples may include a number of known fraudulent transactions (e.g., 4500) and a number of known legitimate transactions (e.g., 100,000). From this historical set of sample of transactions, a training set may be defined that includes a number of known fraudulent transactions (e.g., 4000), a number of known legitimate transactions (e.g., 8000), and a testing set including a number of transactions (e.g., 4000) randomly selected from the historical sample. The post analysis engine 68 may retrieve features from transactions to use as inputs to the neural network 80. These features may include general information such as the ip-based city or region, and may include a large number (e.g., 732) of binary variables.") Florimond does not explicitly teach, however, Reddy does teach: c. updating the received profiles by adding a record of the last transaction, (Reddy US20230351324 at paras. 35-39, 42-44, 60-62) ("[0036] For example, location holds, inventory adjustments, and/or store order label cancels can appear in the inventory audit trail for a particular warehouse and/or distribution center and may indicate a high probability that a corresponding bin is defective. In brief, a location hold is when a worker puts a particular bin (e.g., location) on hold if they suspect something is wrong with item(s) and/or quantity of item(s) in that particular bin. This bin may not be accessibly for other operations to be performed until the hold is released. Inventory adjustments can include recorded adjustments (e.g., increases or decreases) in items and/or units in a particular bin. Sometimes, inventory adjustments can include resolving location holds in which a worker can adjust the items and/or units in a bin that was put on hold and then release the hold. Inventory adjustments can include cycle counts, in which a worker adjusts the items and/or units in a bin as part of routine cycle counts and/or verification audits. Inventory adjustments can also include consolidation efforts, in which a worker adjusts the items and/or units in a bin as part of a consolidation task. Inventory adjustments may also include item correction and/or expiration efforts, in which a worker adjusts an incorrect Vendor Case Pack (VCP)/Store Ship Pack (SSP) and/or updated expired items in a particular bin. In brief, the VCP can indicate a minimum number of units shipped from a vendor to the warehouse and/or distribution center. The SSP indicates a minimum number of units shipped to a retail environment. As an illustrative example, when a vendor ships a pack of soft drinks with 40 units to a warehouse, the VCP is 40. When the warehouse ships a pack of the soft drinks with 8 units to a store, the SSP is 8. Thus, one VCP of the soft drinks contains 5 SSPs of the soft drinks." "[0042] The bin and worker data 114 can include a variety of information that may be recorded by workers as they complete their assigned tasks in the distribution center 110. For example, the data 114 can include, for each bin 106A-N in the distribution center 110, a quantity of workers who access the bin, worker history, and bin history. In some implementations, as shown in FIG. 1, the worker history can indicate a propensity for workers to misplace and/or miss-pick items from the bin. The bin history can also indicate how recent a worker was last at the bin, when information was last validated or verified for the bin, and indications of whether any content of the bin was identified as inaccurate (e.g., item type, item quantity, etc.)." "[0061] Moreover, the computer system 102 can iteratively improve the model over time as inventory audit trails are updated with new information for the particular distribution center. In some implementations, the computer system 102 can iterate training across different parameter settings each day (or at other predetermined time intervals) and select a best or preferred model to use during runtime for that day. Accordingly, a distribution center can execute a different model each day based on different parameters that are relevant to activities in the distribution center for that day.") d. generating a profile-based dataset, (Reddy US20230351324 at paras. 35-39, 42-44, 60-62) ("[0039] Referring to the figures, FIG. 1 is a conceptual diagram for predicting defective bins in a distribution center 110 (e.g., replenishment network, warehouse, fulfillment network, etc.). A computer system 102, worker devices 104A-N, and data store 105 can communicate (e.g., wired and/or wireless) via network(s) 108. In brief, the computer system 102 can be configured to predict defective bins in the distribution center 110. The computer system 102 can be remote from the distribution center 110. For example, the computer system 102 can be a cloud-based service. Moreover, the computer system 102 can be configured to predict defective bins in one or more different distribution centers associated with a particular business, organization, company, or other entity. The worker devices 104A-N can be mobile devices, smart phones, computers, laptops, tablets, and/or one or more other types of user devices. The worker devices 104A-N can be used by workers in the distribution center 110 to track their activities as they perform tasks in the distribution center 110. For example, a worker can be tasked with auditing items in one or more bins in the distribution center 110. Results or other data from the worker's audit can be inputted via the worker device 104A-N and uploaded to the data store 105 and/or the computer system 102. These results or other data can then be used by the computer system 102 to train machine learning models to predict defective bins and/or as input to the machine learning models during runtime. The data store 105 can be a database, cloud-based storage, or other storage device/system that can store data about the distribution center 110, such as inventory audit trails. Data that is logged/inputted by the workers at the worker devices 104A-N can be stored in the data store 105 for later retrieval and use by the computer system 102." "[0042] The bin and worker data 114 can include a variety of information that may be recorded by workers as they complete their assigned tasks in the distribution center 110. For example, the data 114 can include, for each bin 106A-N in the distribution center 110, a quantity of workers who access the bin, worker history, and bin history. In some implementations, as shown in FIG. 1, the worker history can indicate a propensity for workers to misplace and/or miss-pick items from the bin. The bin history can also indicate how recent a worker was last at the bin, when information was last validated or verified for the bin, and indications of whether any content of the bin was identified as inaccurate (e.g., item type, item quantity, etc.). [0043] Additionally or alternatively, other information can be received in block A, such as location holds, inventory adjustments, store order label cancels, other worker activity data, item activity data, and/or other bin activity data that are part of inventory audit trails for the distribution center 110. Information about a particular worker who replenishes or picks items at a bin can be an indicator of whether the bin is defective. As an illustrative example, if that worker puts laptops in bin A5 instead of bin A7 every three out of 10 tasks associated with bin A5, then bin A5 can have a 30% increased chance of being defective. Moreover, in some implementations, any bin that the particular worker interacts with can be assigned a 30% increased chance of being defective, simply because the particular worker has a propensity to misplace items in bins. [0044] In block B, the computer system 102 can apply one or more models to the data 114. In some implementations, the computer system 102 can determine which model(s) to apply based on the distribution center 110. As described further below (e.g., refer to FIGS. 2 and 4A-B), a defective bin prediction model can be generated for each distribution center based on factors/features that are unique/specific to that distribution center." "[0061] Moreover, the computer system 102 can iteratively improve the model over time as inventory audit trails are updated with new information for the particular distribution center. In some implementations, the computer system 102 can iterate training across different parameter settings each day (or at other predetermined time intervals) and select a best or preferred model to use during runtime for that day. Accordingly, a distribution center can execute a different model each day based on different parameters that are relevant to activities in the distribution center for that day.") e. feeding the profile-based dataset into a balanced random undersampling algorithm in order to obtain an undersampled dataset in which [legitimate and fraudulent] transactions are distributed in a balanced manner, (Reddy US20230351324 at paras. 15-16, 42-44, 78-80) ("[0042] The bin and worker data 114 can include a variety of information that may be recorded by workers as they complete their assigned tasks in the distribution center 110. For example, the data 114 can include, for each bin 106A-N in the distribution center 110, a quantity of workers who access the bin, worker history, and bin history. In some implementations, as shown in FIG. 1, the worker history can indicate a propensity for workers to misplace and/or miss-pick items from the bin. The bin history can also indicate how recent a worker was last at the bin, when information was last validated or verified for the bin, and indications of whether any content of the bin was identified as inaccurate (e.g., item type, item quantity, etc.)." "[0044] In block B, the computer system 102 can apply one or more models to the data 114. In some implementations, the computer system 102 can determine which model(s) to apply based on the distribution center 110. As described further below (e.g., refer to FIGS. 2 and 4A-B), a defective bin prediction model can be generated for each distribution center based on factors/features that are unique/specific to that distribution center." "[0078] In block 406, the computer system can identify data in the remaining retrieved data for bins flagged as not defective. Over a last 4 weeks, a distribution center can have a defect rate (e.g., percent of bins flagged as defective) between 0.5% and 2%. Since this means that approximately 98% of bins may be flagged as non-defective, the computer system can undersample the bins flagged as non-defective to generate a more balanced dataset for training purposes. [0079] In block 408, the computer system can identify data in the remaining retrieved data for N bins that are flagged as defective. The computer system can use stratification techniques and/or stratified random sampling (block 410). N can be a portion of the bins that is less than all of the bins flagged as defective. With stratified random sampling, a total population of the bins flagged as defective can be divided into smaller subgroups or strata. The strata can be formed based on shared attributes or characteristics of the bins. Stratified random sampling therefore allows for obtaining a sample population of the negative cases that best represents an entire population of the negative cases to generate a more balanced training dataset. [0080] In block 411, the computer system can combine the identified data into a positive defective bin dataset. This dataset can, for example, be a 50% positive defective bin dataset. By retrieving all cases of defective bins and a smattering of other cases of non-defective bins, as an example, the training dataset can have stronger signals of accuracy since the cases of defective bins may not be drowned out by an abundance of the other cases of non-defective bins.") Therefore, it would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Florimond and Reddy, because it allows for improved systems for a defective transaction prediction tool that may not require additional resources, such as workers, for random sampling and/or auditing of transactions in a warehouse and/or distribution center to then predict the defective transactions. Instead, the disclosed techniques can leverage existing inventory audit trails as sample data for training the machine learning models described herein. The inventory audit trails can include an abundance of pick data, pull data, and other types of data associated with the warehouses and/or distribution centers that can be used to accurate and proactively predict transaction defectiveness. (Reddy at Abstract and paras. 19-21). Florimond and Reddy do not explicitly teach, however, Rashid does teach: g. generating a complexity matrix for the outputs of the deep neural network, (Rashid NPL 2022 at tables 3-4, pp. 4-6) ("III. RESULTS In the research, we will present the experiments which show the quality of our detection System. We discuss the results of experiments with various auto-encoder hyper parameters and show the impact of pre-training and feature scaling on the detection performance. Because we have access to the labels of the dataset, we are also going to present several performance measures to discuss the efficiency of our detection system along with comparison to existing work. A. Number of layers Although the number layers generally depend on the complexity of function we want to model with the neural network, picking the right number is not trivial. To illustrate that, we show three figures with different layer configurations, in each the number of neurons between layers’ decreases (and increases) by 5. To generate these curves, we have used the detection threshold of 0.001727. The f1 score was calculated against the same dataset and same threshold by using the labels we have. Figure 4 (left) shows a training phase of auto encoder with four layers (thus the latent representation has 20 dimensions). Figures 4 show auto-encoders which consist of six and ten layers respectively (thus the latent representations have 15 dimensions and 5 dimensions). We would favor the configuration on left, based on the shape of the curve and based on the maximum value of the mean squared feature recall." "B. Classification performance In this section we discuss the performance of our unsupervised detection system by using the labels from the dataset. Although we have already showed some performance measure in the previous sections, all of them used the whole dataset both for training and classification. This is not ideal, because we might over-fit on the data (e.g. by choosing certain parameters) and our solution might not generalize well. We would also like to classify incoming transactions immediately (so called online classification), but retraining the whole auto-encoder every time would be too slow. Because of these three reasons we used cross validation to simulate the situation of receiving an unseen dataset. We use the stratified k-fold method where k = 5, which means we split the dataset randomly into a training one, containing 80% original transactions, and testing one, containing 20% of the original transactions. Because the split is stratified, the ratio of fraudulent transactions remains always the same. This is done 5 times to obtain 5 different training and testing sets, which are used as an input to the detection system (with the detection threshold of 0.1727). We average the results and obtain a single confusion matrix which is depicted in Figure 6.") h. generating performance metrics from the complexity matrix, (Rashid NPL 2022 at tables 3-4, pp. 4-6) ("B. Classification performance In this section we discuss the performance of our unsupervised detection system by using the labels from the dataset. Although we have already showed some performance measure in the previous sections, all of them used the whole dataset both for training and classification. This is not ideal, because we might over-fit on the data (e.g. by choosing certain parameters) and our solution might not generalize well. We would also like to classify incoming transactions immediately (so called online classification), but retraining the whole auto-encoder every time would be too slow. Because of these three reasons we used cross validation to simulate the situation of receiving an unseen dataset. We use the stratified k-fold method where k = 5, which means we split the dataset randomly into a training one, containing 80% original transactions, and testing one, containing 20% of the original transactions. Because the split is stratified, the ratio of fraudulent transactions remains always the same. This is done 5 times to obtain 5 different training and testing sets, which are used as an input to the detection system (with the detection threshold of 0.1727). We average the results and obtain a single confusion matrix which is depicted in Figure 6.") Therefore, it would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Florimond, Reddy, and Rashid, because it allows for an improved system to facilitate an algorithm that will improve current detection of fraud techniques and contribute to the reduction of fraud in healthcare. (Rashid at Abstract and pp. 1). Florimond, Reddy, and Rashid do not explicitly teach, however, Seo does teach: i. training the deep neural network with feedback of the performance metrics. (Seo WO2018097653A1 at pp. 3-4) ("In order to solve the above problems, a method of predicting a chargeback fraud user according to an embodiment of the present invention includes: (1) transaction history data of a conventional normal user and a chargeback fraud user for each user; A data processing step of processing the data based on one case, (2) A data classification step of dividing the processed transaction history data into training data and test data, and (3) A chargeback fraud user among the training data. A data adjustment step of adjusting the number of data by oversampling the data for (4) training using a specific machine learning method using the training data in which the number of data is adjusted, and testing using the learned machine learning model Predictive classification step for predicting whether a chargeback fraud user is against data, (5) performance for predictive classification (6) Predictive classification step and performance by oversampling or undersampling data for chargeback fraud users in the training data until the performance for the predictive classification reaches the target value. Iteratively performing a repeating step of measuring the measurement, (7) predicting the chargeback fraud for the transaction history data of the new user using the machine learning model that has reached the target prediction classification performance.") Therefore, it would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Florimond, Reddy, Rashid, and Seo, because it allows for an improved system providing a machine learning model that satisfies the target prediction classification performance by processing transaction history data of a conventional user, and implements the implemented machine learning model. And a method and program for predicting chargeback fraud for a new user's transaction history data. (Seo at Abstract and pp. 2-3). Florimond, Reddy, Rashid, and Seo do not explicitly teach, however, Gao does teach: in chronological order... (Gao US20230342993 at paras. 34-37) ("[0035] Exemplary embodiments use a golden-angle-ordered radial k-space undersampled acquisition (shown as an example in FIG. 3), although any suitable MRI k-space data acquisition strategy can be used. Starting from 0°, each radial k-space spoke in undersampled k-space 30 is rotated by a golden angle, which is approximately 111.25°. In this example, five spokes are illustrated (labeled in time order 1-5), but it should be appreciated that many more spokes will be captured in practice, more than 100 in most embodiments. In some embodiments, substantially around 100 spokes (95-110) are captured, such that around 400 total spokes are captured or generated, making the resulting data set equivalent in size to existing fully-sampled radial capture schemes. The golden-angle acquisition ensures uniform coverage of the k-space for any arbitrary number of consecutive spokes so any newly acquired radial spoke is interleaved with the acquired ones. The spokes that are physically close (in azimuth angle) are dependent on each other. The angular dependencies between each of the spokes need not be given to a transformer network and can be learned by the attention modules in a transformer network to predict unacquired spokes and identify those acquired spokes that are most relevant to each spoke being predicted. For this reason, acquired spokes can be input to the network in any order. Some embodiments, rearrange the radial spokes based on the chronological order of acquisition and encode the temporal spoke index in the network. Note that the same concept can be extended to other non-Cartesian k-space trajectories such as the spiral trajectory or uniform-angle ordering. [0036] Graphs 32 and 34 shows pre-processing steps of radial k-space. For each single coil 2D slice, the golden-angle-ordered radial spokes are rearranged based on the chronological order of acquisition. The spokes in undersampled k-space 30 include a real part 35 and an imaginary part 36. These spokes can be ordered in any reasonable manor along an x-axis, representing the inputs to the transformer network, due to the self-attention layers in the network. In some embodiments, these are input in a chronological order. In some embodiments, the spokes can be order for 0-360 degrees. Graph 32 shows the simple case where spokes are arranged in chronological order, and only the real part 35 is input into the network. However, the transformer network can also process and predict the imaginary part of the k-space. Thus, in some embodiments, the input to the network is shown in graph 34, where the real part 35 is concatenated with the imaginary part 36 of each spoke, representing each spoke as a longer vector. In some embodiments, each of the k-space spokes is inverse Fourier transformed to generate a projection vector at each angle. The projection signals span the whole image compared to the centrally condensed k-space signals, so the network can better learn the image details and the reconstructed image is less sensitive to network errors. In some embodiments, the real and imaginary parts of each complex-valued projection are concatenated to a single long vector prior to input to the network. Because the frequency encoding line (kx) is typically two-times oversampled, the complex-valued projection vector is of length 512 and the concatenated projection vector is of length 1024. The input and output are of size 1024×100 with each projection vector as a token in the network.") Therefore, it would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Florimond, Reddy, Rashid, Seo, and Gao, because it allows for an improved system for removing or reducing artifacts within undersampled data sets. (Gao at Abstract and paras. 2-8). As per claim 2, Florimond and Reddy do not explicitly teach, however, Rashid does teach: wherein the complexity matrix is weighted based on a cost of the transactions. (Rashid NPL 2022 at tables 3-4, pp. 6-7) ("There is also a very straightforward way to quantify the cost There is also a very straightforward way to quantify the cost of various detection thresh-olds. If we assume that undetected fraud costs the bank as much money as was stolen, we can calculate the net gain from using our detection system based on the incurred expenses to verify the transaction authenticity. We will use the same average cost of 1$ per verification. Note that we don't know the currency of transactions in our dataset, so we use multiple European currencies (because the dataset is from European bank). The net gain of our solution for various currencies and thresholds are shown in the Table 4.") Therefore, it would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Florimond, Reddy, Rashid, Seo, and Gao, because it allows for an improved system to facilitate an algorithm that will improve current detection of fraud techniques and contribute to the reduction of fraud in healthcare. (Rashid at Abstract and pp. 1). As per claim 3, Florimond and Reddy do not explicitly teach, however, Rashid does teach: wherein the complexity matrix is weighted based on a number of the transactions. (Rashid NPL 2022 at tables 3-4, pp. 5-6) ("B. Classification performance In this section we discuss the performance of our unsupervised detection system by using the labels from the dataset. Although we have already showed some performance measure in the previous sections, all of them used the whole dataset both for training and classification. This is not ideal, because we might over-fit on the data (e.g. by choosing certain parameters) and our solution might not generalize well. We would also like to classify incoming transactions immediately (so called online classification), but retraining the whole auto-encoder every time would be too slow. Because of these three reasons we used cross validation to simulate the situation of receiving an unseen dataset. We use the stratified k-fold method where k = 5, which means we split the dataset randomly into a training one, containing 80% original transactions, and testing one, containing 20% of the original transactions. Because the split is stratified, the ratio of fraudulent transactions remains always the same. This is done 5 times to obtain 5 different training and testing sets, which are used as an input to the detection system (with the detection threshold of 0.1727). We average the results and obtain a single confusion matrix which is depicted in Figure 6.") Therefore, it would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Florimond, Reddy, Rashid, Seo, and Gao, because it allows for an improved system to facilitate an algorithm that will improve current detection of fraud techniques and contribute to the reduction of fraud in healthcare. (Rashid at Abstract and pp. 1). As per claim 4, Florimond, Reddy, and Rashid do not explicitly teach, however, Seo does teach: further comprising oversampling the undersampled dataset. (Seo WO2018097653A1 at pp. 11-12) ("The iteration step S60 may oversample or undersample the data for the chargeback fraud user of the training data until the performance of the prediction classification reaches the target value, thereby predicting classification S40 and The step of repeating the performance measurement step (S50). At this time, if the oversampling ratio is too high and the training data amount is excessively increased, an overload may occur when learning in the predictive classification step S40, and thus, the repeating step S60 may perform undersampling in addition to oversampling of the training data. Can be done. In addition, in the repeating step (S60), undersampling may be performed to reach a more accurate target value. The ratio of oversampling or undersampling at the time of performing the repetition step S60 may be regular or arbitrary, and is not particularly limited. For example, the oversampling ratio at the nth iteration can be defined as A × B n (where A and B are natural numbers and n is an integer). After the nth iteration, the undersampling ratio at the mth iteration can be defined as A × B n − (C × m) (where A, B and C are natural numbers, n and m are integers, n <m) That is, when "Recall with 0.940 or more" is set as the performance target value for the support vector machine SVM, the prediction classification step S40 and the performance measurement step S50 when the oversampling is 100% are performed. As a result, since Recall has 0.820 that is less than or equal to the target value, the oversampling is increased to 200% in the first iteration, and the prediction classification step S40 and the performance measurement step S50 are performed again. As a result, since Recall has 0.887 that is less than or equal to the target value, the oversampling is raised to 300% in the second iteration, and the prediction classification step S40 and the performance measurement step S50 are performed again. As a result, Recall has 0.948, which is above the target value. Subsequently, although it may be immediately skipped to the prediction step S70, in the iteration step S60, undersampling is performed in the third iteration, that is, the oversampling is set to less than 300%, for example, 280% to predict the classification step ( S40) and the performance measurement step S50 may be performed again.") Therefore, it would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Florimond, Reddy, Rashid, Seo, and Gao, because it allows for an improved system providing a machine learning model that satisfies the target prediction classification performance by processing transaction history data of a conventional user, and implements the implemented machine learning model. And a method and program for predicting chargeback fraud for a new user's transaction history data. (Seo at Abstract and pp. 2-3). As per claim 5, Florimond does not explicitly teach, however, Reddy does teach: wherein the balanced random undersampling algorithm maintains (Reddy US20230351324 at paras. 15-16, 42-44, 78-80) ("[0042] The bin and worker data 114 can include a variety of information that may be recorded by workers as they complete their assigned tasks in the distribution center 110. For example, the data 114 can include, for each bin 106A-N in the distribution center 110, a quantity of workers who access the bin, worker history, and bin history. In some implementations, as shown in FIG. 1, the worker history can indicate a propensity for workers to misplace and/or miss-pick items from the bin. The bin history can also indicate how recent a worker was last at the bin, when information was last validated or verified for the bin, and indications of whether any content of the bin was identified as inaccurate (e.g., item type, item quantity, etc.)." "[0044] In block B, the computer system 102 can apply one or more models to the data 114. In some implementations, the computer system 102 can determine which model(s) to apply based on the distribution center 110. As described further below (e.g., refer to FIGS. 2 and 4A-B), a defective bin prediction model can be generated for each distribution center based on factors/features that are unique/specific to that distribution center." "[0078] In block 406, the computer system can identify data in the remaining retrieved data for bins flagged as not defective. Over a last 4 weeks, a distribution center can have a defect rate (e.g., percent of bins flagged as defective) between 0.5% and 2%. Since this means that approximately 98% of bins may be flagged as non-defective, the computer system can undersample the bins flagged as non-defective to generate a more balanced dataset for training purposes. [0079] In block 408, the computer system can identify data in the remaining retrieved data for N bins that are flagged as defective. The computer system can use stratification techniques and/or stratified random sampling (block 410). N can be a portion of the bins that is less than all of the bins flagged as defective. With stratified random sampling, a total population of the bins flagged as defective can be divided into smaller subgroups or strata. The strata can be formed based on shared attributes or characteristics of the bins. Stratified random sampling therefore allows for obtaining a sample population of the negative cases that best represents an entire population of the negative cases to generate a more balanced training dataset. [0080] In block 411, the computer system can combine the identified data into a positive defective bin dataset. This dataset can, for example, be a 50% positive defective bin dataset. By retrieving all cases of defective bins and a smattering of other cases of non-defective bins, as an example, the training dataset can have stronger signals of accuracy since the cases of defective bins may not be drowned out by an abundance of the other cases of non-defective bins.") Therefore, it would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Florimond and Reddy, because it allows for improved systems for a defective transaction prediction tool that may not require additional resources, such as workers, for random sampling and/or auditing of transactions in a warehouse and/or distribution center to then predict the defective transactions. Instead, the disclosed techniques can leverage existing inventory audit trails as sample data for training the machine learning models described herein. The inventory audit trails can include an abundance of pick data, pull data, and other types of data associated with the warehouses and/or distribution centers that can be used to accurate and proactively predict transaction defectiveness. (Reddy at Abstract and paras. 19-21). Florimond, Reddy, Rashid, and Seo do not explicitly teach, however, Gao does teach: a chronological order in the undersampled dataset. (Gao US20230342993 at paras. 34-37) ("[0035] Exemplary embodiments use a golden-angle-ordered radial k-space undersampled acquisition (shown as an example in FIG. 3), although any suitable MRI k-space data acquisition strategy can be used. Starting from 0°, each radial k-space spoke in undersampled k-space 30 is rotated by a golden angle, which is approximately 111.25°. In this example, five spokes are illustrated (labeled in time order 1-5), but it should be appreciated that many more spokes will be captured in practice, more than 100 in most embodiments. In some embodiments, substantially around 100 spokes (95-110) are captured, such that around 400 total spokes are captured or generated, making the resulting data set equivalent in size to existing fully-sampled radial capture schemes. The golden-angle acquisition ensures uniform coverage of the k-space for any arbitrary number of consecutive spokes so any newly acquired radial spoke is interleaved with the acquired ones. The spokes that are physically close (in azimuth angle) are dependent on each other. The angular dependencies between each of the spokes need not be given to a transformer network and can be learned by the attention modules in a transformer network to predict unacquired spokes and identify those acquired spokes that are most relevant to each spoke being predicted. For this reason, acquired spokes can be input to the network in any order. Some embodiments, rearrange the radial spokes based on the chronological order of acquisition and encode the temporal spoke index in the network. Note that the same concept can be extended to other non-Cartesian k-space trajectories such as the spiral trajectory or uniform-angle ordering. [0036] Graphs 32 and 34 shows pre-processing steps of radial k-space. For each single coil 2D slice, the golden-angle-ordered radial spokes are rearranged based on the chronological order of acquisition. The spokes in undersampled k-space 30 include a real part 35 and an imaginary part 36. These spokes can be ordered in any reasonable manor along an x-axis, representing the inputs to the transformer network, due to the self-attention layers in the network. In some embodiments, these are input in a chronological order. In some embodiments, the spokes can be order for 0-360 degrees. Graph 32 shows the simple case where spokes are arranged in chronological order, and only the real part 35 is input into the network. However, the transformer network can also process and predict the imaginary part of the k-space. Thus, in some embodiments, the input to the network is shown in graph 34, where the real part 35 is concatenated with the imaginary part 36 of each spoke, representing each spoke as a longer vector. In some embodiments, each of the k-space spokes is inverse Fourier transformed to generate a projection vector at each angle. The projection signals span the whole image compared to the centrally condensed k-space signals, so the network can better learn the image details and the reconstructed image is less sensitive to network errors. In some embodiments, the real and imaginary parts of each complex-valued projection are concatenated to a single long vector prior to input to the network. Because the frequency encoding line (kx) is typically two-times oversampled, the complex-valued projection vector is of length 512 and the concatenated projection vector is of length 1024. The input and output are of size 1024×100 with each projection vector as a token in the network.") Therefore, it would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Florimond, Reddy, Rashid, Seo, and Gao, because it allows for an improved system for removing or reducing artifacts within undersampled data sets. (Gao at Abstract and paras. 2-8). As per claim 8, Florimond and Reddy do not explicitly teach, however, Rashid does teach: weighted based on a number of the transactions, and wherein the performance metrics are further generated (Rashid NPL 2022 at tables 3-4, pp. 6-7) ("There is also a very straightforward way to quantify the cost There is also a very straightforward way to quantify the cost of various detection thresh-olds. If we assume that undetected fraud costs the bank as much money as was stolen, we can calculate the net gain from using our detection system based on the incurred expenses to verify the transaction authenticity. We will use the same average cost of 1$ per verification. Note that we don't know the currency of transactions in our dataset, so we use multiple European currencies (because the dataset is from European bank). The net gain of our solution for various currencies and thresholds are shown in the Table 4." "B. Classification performance In this section we discuss the performance of our unsupervised detection system by using the labels from the dataset. Although we have already showed some performance measure in the previous sections, all of them used the whole dataset both for training and classification. This is not ideal, because we might over-fit on the data (e.g. by choosing certain parameters) and our solution might not generalize well. We would also like to classify incoming transactions immediately (so called online classification), but retraining the whole auto-encoder every time would be too slow. Because of these three reasons we used cross validation to simulate the situation of receiving an unseen dataset. We use the stratified k-fold method where k = 5, which means we split the dataset randomly into a training one, containing 80% original transactions, and testing one, containing 20% of the original transactions. Because the split is stratified, the ratio of fraudulent transactions remains always the same. This is done 5 times to obtain 5 different training and testing sets, which are used as an input to the detection system (with the detection threshold of 0.1727). We average the results and obtain a single confusion matrix which is depicted in Figure 6.") Therefore, it would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Florimond, Reddy, and Rashid, because it allows for an improved system to facilitate an algorithm that will improve current detection of fraud techniques and contribute to the reduction of fraud in healthcare. (Rashid at Abstract and pp. 1). Florimond, Reddy, and Rashid do not explicitly teach, however, Seo does teach: further comprising generating a second complexity matrix for the outputs of the deep neural network, wherein the second complexity matrix is (Seo WO2018097653A1 at pp. 11-12) ("The iteration step S60 may oversample or undersample the data for the chargeback fraud user of the training data until the performance of the prediction classification reaches the target value, thereby predicting classification S40 and The step of repeating the performance measurement step (S50). At this time, if the oversampling ratio is too high and the training data amount is excessively increased, an overload may occur when learning in the predictive classification step S40, and thus, the repeating step S60 may perform undersampling in addition to oversampling of the training data. Can be done. In addition, in the repeating step (S60), undersampling may be performed to reach a more accurate target value. The ratio of oversampling or undersampling at the time of performing the repetition step S60 may be regular or arbitrary, and is not particularly limited. For example, the oversampling ratio at the nth iteration can be defined as A × B n (where A and B are natural numbers and n is an integer). After the nth iteration, the undersampling ratio at the mth iteration can be defined as A × B n − (C × m) (where A, B and C are natural numbers, n and m are integers, n <m) That is, when "Recall with 0.940 or more" is set as the performance target value for the support vector machine SVM, the prediction classification step S40 and the performance measurement step S50 when the oversampling is 100% are performed. As a result, since Recall has 0.820 that is less than or equal to the target value, the oversampling is increased to 200% in the first iteration, and the prediction classification step S40 and the performance measurement step S50 are performed again. As a result, since Recall has 0.887 that is less than or equal to the target value, the oversampling is raised to 300% in the second iteration, and the prediction classification step S40 and the performance measurement step S50 are performed again. As a result, Recall has 0.948, which is above the target value. Subsequently, although it may be immediately skipped to the prediction step S70, in the iteration step S60, undersampling is performed in the third iteration, that is, the oversampling is set to less than 300%, for example, 280% to predict the classification step ( S40) and the performance measurement step S50 may be performed again.") from the second complexity matrix. (Seo WO2018097653A1 at pp. 11-12) ("The iteration step S60 may oversample or undersample the data for the chargeback fraud user of the training data until the performance of the prediction classification reaches the target value, thereby predicting classification S40 and The step of repeating the performance measurement step (S50). At this time, if the oversampling ratio is too high and the training data amount is excessively increased, an overload may occur when learning in the predictive classification step S40, and thus, the repeating step S60 may perform undersampling in addition to oversampling of the training data. Can be done. In addition, in the repeating step (S60), undersampling may be performed to reach a more accurate target value. The ratio of oversampling or undersampling at the time of performing the repetition step S60 may be regular or arbitrary, and is not particularly limited. For example, the oversampling ratio at the nth iteration can be defined as A × B n (where A and B are natural numbers and n is an integer). After the nth iteration, the undersampling ratio at the mth iteration can be defined as A × B n − (C × m) (where A, B and C are natural numbers, n and m are integers, n <m) That is, when "Recall with 0.940 or more" is set as the performance target value for the support vector machine SVM, the prediction classification step S40 and the performance measurement step S50 when the oversampling is 100% are performed. As a result, since Recall has 0.820 that is less than or equal to the target value, the oversampling is increased to 200% in the first iteration, and the prediction classification step S40 and the performance measurement step S50 are performed again. As a result, since Recall has 0.887 that is less than or equal to the target value, the oversampling is raised to 300% in the second iteration, and the prediction classification step S40 and the performance measurement step S50 are performed again. As a result, Recall has 0.948, which is above the target value. Subsequently, although it may be immediately skipped to the prediction step S70, in the iteration step S60, undersampling is performed in the third iteration, that is, the oversampling is set to less than 300%, for example, 280% to predict the classification step ( S40) and the performance measurement step S50 may be performed again.") Therefore, it would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Florimond, Reddy, Rashid, Seo, and Gao, because it allows for an improved system providing a machine learning model that satisfies the target prediction classification performance by processing transaction history data of a conventional user, and implements the implemented machine learning model. And a method and program for predicting chargeback fraud for a new user's transaction history data. (Seo at Abstract and pp. 2-3). As per claim 9, Florimond, Reddy, and Rashid do not explicitly teach, however, Seo does teach: further comprising oversampling the undersampled dataset. (Seo WO2018097653A1 at pp. 11-12) ("The iteration step S60 may oversample or undersample the data for the chargeback fraud user of the training data until the performance of the prediction classification reaches the target value, thereby predicting classification S40 and The step of repeating the performance measurement step (S50). At this time, if the oversampling ratio is too high and the training data amount is excessively increased, an overload may occur when learning in the predictive classification step S40, and thus, the repeating step S60 may perform undersampling in addition to oversampling of the training data. Can be done. In addition, in the repeating step (S60), undersampling may be performed to reach a more accurate target value. The ratio of oversampling or undersampling at the time of performing the repetition step S60 may be regular or arbitrary, and is not particularly limited. For example, the oversampling ratio at the nth iteration can be defined as A × B n (where A and B are natural numbers and n is an integer). After the nth iteration, the undersampling ratio at the mth iteration can be defined as A × B n − (C × m) (where A, B and C are natural numbers, n and m are integers, n <m) That is, when "Recall with 0.940 or more" is set as the performance target value for the support vector machine SVM, the prediction classification step S40 and the performance measurement step S50 when the oversampling is 100% are performed. As a result, since Recall has 0.820 that is less than or equal to the target value, the oversampling is increased to 200% in the first iteration, and the prediction classification step S40 and the performance measurement step S50 are performed again. As a result, since Recall has 0.887 that is less than or equal to the target value, the oversampling is raised to 300% in the second iteration, and the prediction classification step S40 and the performance measurement step S50 are performed again. As a result, Recall has 0.948, which is above the target value. Subsequently, although it may be immediately skipped to the prediction step S70, in the iteration step S60, undersampling is performed in the third iteration, that is, the oversampling is set to less than 300%, for example, 280% to predict the classification step ( S40) and the performance measurement step S50 may be performed again.") Therefore, it would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Florimond, Reddy, Rashid, Seo, and Gao, because it allows for an improved system providing a machine learning model that satisfies the target prediction classification performance by processing transaction history data of a conventional user, and implements the implemented machine learning model. And a method and program for predicting chargeback fraud for a new user's transaction history data. (Seo at Abstract and pp. 2-3). As per claim 10, Florimond does not explicitly teach, however, Reddy does teach: wherein the balanced random undersampling algorithm maintains (Reddy US20230351324 at paras. 15-16, 42-44, 78-80) ("[0042] The bin and worker data 114 can include a variety of information that may be recorded by workers as they complete their assigned tasks in the distribution center 110. For example, the data 114 can include, for each bin 106A-N in the distribution center 110, a quantity of workers who access the bin, worker history, and bin history. In some implementations, as shown in FIG. 1, the worker history can indicate a propensity for workers to misplace and/or miss-pick items from the bin. The bin history can also indicate how recent a worker was last at the bin, when information was last validated or verified for the bin, and indications of whether any content of the bin was identified as inaccurate (e.g., item type, item quantity, etc.)." "[0044] In block B, the computer system 102 can apply one or more models to the data 114. In some implementations, the computer system 102 can determine which model(s) to apply based on the distribution center 110. As described further below (e.g., refer to FIGS. 2 and 4A-B), a defective bin prediction model can be generated for each distribution center based on factors/features that are unique/specific to that distribution center." "[0078] In block 406, the computer system can identify data in the remaining retrieved data for bins flagged as not defective. Over a last 4 weeks, a distribution center can have a defect rate (e.g., percent of bins flagged as defective) between 0.5% and 2%. Since this means that approximately 98% of bins may be flagged as non-defective, the computer system can undersample the bins flagged as non-defective to generate a more balanced dataset for training purposes. [0079] In block 408, the computer system can identify data in the remaining retrieved data for N bins that are flagged as defective. The computer system can use stratification techniques and/or stratified random sampling (block 410). N can be a portion of the bins that is less than all of the bins flagged as defective. With stratified random sampling, a total population of the bins flagged as defective can be divided into smaller subgroups or strata. The strata can be formed based on shared attributes or characteristics of the bins. Stratified random sampling therefore allows for obtaining a sample population of the negative cases that best represents an entire population of the negative cases to generate a more balanced training dataset. [0080] In block 411, the computer system can combine the identified data into a positive defective bin dataset. This dataset can, for example, be a 50% positive defective bin dataset. By retrieving all cases of defective bins and a smattering of other cases of non-defective bins, as an example, the training dataset can have stronger signals of accuracy since the cases of defective bins may not be drowned out by an abundance of the other cases of non-defective bins.") Therefore, it would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Florimond and Reddy, because it allows for improved systems for a defective transaction prediction tool that may not require additional resources, such as workers, for random sampling and/or auditing of transactions in a warehouse and/or distribution center to then predict the defective transactions. Instead, the disclosed techniques can leverage existing inventory audit trails as sample data for training the machine learning models described herein. The inventory audit trails can include an abundance of pick data, pull data, and other types of data associated with the warehouses and/or distribution centers that can be used to accurate and proactively predict transaction defectiveness. (Reddy at Abstract and paras. 19-21). Florimond, Reddy, Rashid, and Seo do not explicitly teach, however, Gao does teach: a chronological order in the undersampled dataset. (Gao US20230342993 at paras. 34-37) ("[0035] Exemplary embodiments use a golden-angle-ordered radial k-space undersampled acquisition (shown as an example in FIG. 3), although any suitable MRI k-space data acquisition strategy can be used. Starting from 0°, each radial k-space spoke in undersampled k-space 30 is rotated by a golden angle, which is approximately 111.25°. In this example, five spokes are illustrated (labeled in time order 1-5), but it should be appreciated that many more spokes will be captured in practice, more than 100 in most embodiments. In some embodiments, substantially around 100 spokes (95-110) are captured, such that around 400 total spokes are captured or generated, making the resulting data set equivalent in size to existing fully-sampled radial capture schemes. The golden-angle acquisition ensures uniform coverage of the k-space for any arbitrary number of consecutive spokes so any newly acquired radial spoke is interleaved with the acquired ones. The spokes that are physically close (in azimuth angle) are dependent on each other. The angular dependencies between each of the spokes need not be given to a transformer network and can be learned by the attention modules in a transformer network to predict unacquired spokes and identify those acquired spokes that are most relevant to each spoke being predicted. For this reason, acquired spokes can be input to the network in any order. Some embodiments, rearrange the radial spokes based on the chronological order of acquisition and encode the temporal spoke index in the network. Note that the same concept can be extended to other non-Cartesian k-space trajectories such as the spiral trajectory or uniform-angle ordering. [0036] Graphs 32 and 34 shows pre-processing steps of radial k-space. For each single coil 2D slice, the golden-angle-ordered radial spokes are rearranged based on the chronological order of acquisition. The spokes in undersampled k-space 30 include a real part 35 and an imaginary part 36. These spokes can be ordered in any reasonable manor along an x-axis, representing the inputs to the transformer network, due to the self-attention layers in the network. In some embodiments, these are input in a chronological order. In some embodiments, the spokes can be order for 0-360 degrees. Graph 32 shows the simple case where spokes are arranged in chronological order, and only the real part 35 is input into the network. However, the transformer network can also process and predict the imaginary part of the k-space. Thus, in some embodiments, the input to the network is shown in graph 34, where the real part 35 is concatenated with the imaginary part 36 of each spoke, representing each spoke as a longer vector. In some embodiments, each of the k-space spokes is inverse Fourier transformed to generate a projection vector at each angle. The projection signals span the whole image compared to the centrally condensed k-space signals, so the network can better learn the image details and the reconstructed image is less sensitive to network errors. In some embodiments, the real and imaginary parts of each complex-valued projection are concatenated to a single long vector prior to input to the network. Because the frequency encoding line (kx) is typically two-times oversampled, the complex-valued projection vector is of length 512 and the concatenated projection vector is of length 1024. The input and output are of size 1024×100 with each projection vector as a token in the network.") Therefore, it would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Florimond, Reddy, Rashid, Seo, and Gao, because it allows for an improved system for removing or reducing artifacts within undersampled data sets. (Gao at Abstract and paras. 2-8). As per claim 11, Florimond, Reddy, and Rashid do not explicitly teach, however, Seo does teach: further comprising oversampling the undersampled dataset. (Seo WO2018097653A1 at pp. 11-12) ("The iteration step S60 may oversample or undersample the data for the chargeback fraud user of the training data until the performance of the prediction classification reaches the target value, thereby predicting classification S40 and The step of repeating the performance measurement step (S50). At this time, if the oversampling ratio is too high and the training data amount is excessively increased, an overload may occur when learning in the predictive classification step S40, and thus, the repeating step S60 may perform undersampling in addition to oversampling of the training data. Can be done. In addition, in the repeating step (S60), undersampling may be performed to reach a more accurate target value. The ratio of oversampling or undersampling at the time of performing the repetition step S60 may be regular or arbitrary, and is not particularly limited. For example, the oversampling ratio at the nth iteration can be defined as A × B n (where A and B are natural numbers and n is an integer). After the nth iteration, the undersampling ratio at the mth iteration can be defined as A × B n − (C × m) (where A, B and C are natural numbers, n and m are integers, n <m) That is, when "Recall with 0.940 or more" is set as the performance target value for the support vector machine SVM, the prediction classification step S40 and the performance measurement step S50 when the oversampling is 100% are performed. As a result, since Recall has 0.820 that is less than or equal to the target value, the oversampling is increased to 200% in the first iteration, and the prediction classification step S40 and the performance measurement step S50 are performed again. As a result, since Recall has 0.887 that is less than or equal to the target value, the oversampling is raised to 300% in the second iteration, and the prediction classification step S40 and the performance measurement step S50 are performed again. As a result, Recall has 0.948, which is above the target value. Subsequently, although it may be immediately skipped to the prediction step S70, in the iteration step S60, undersampling is performed in the third iteration, that is, the oversampling is set to less than 300%, for example, 280% to predict the classification step ( S40) and the performance measurement step S50 may be performed again.") Therefore, it would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Florimond, Reddy, Rashid, Seo, and Gao, because it allows for an improved system providing a machine learning model that satisfies the target prediction classification performance by processing transaction history data of a conventional user, and implements the implemented machine learning model. And a method and program for predicting chargeback fraud for a new user's transaction history data. (Seo at Abstract and pp. 2-3). As per claim 12, Florimond does not explicitly teach, however, Reddy does teach: wherein the balanced random undersampling algorithm maintains (Reddy US20230351324 at paras. 15-16, 42-44, 78-80) ("[0042] The bin and worker data 114 can include a variety of information that may be recorded by workers as they complete their assigned tasks in the distribution center 110. For example, the data 114 can include, for each bin 106A-N in the distribution center 110, a quantity of workers who access the bin, worker history, and bin history. In some implementations, as shown in FIG. 1, the worker history can indicate a propensity for workers to misplace and/or miss-pick items from the bin. The bin history can also indicate how recent a worker was last at the bin, when information was last validated or verified for the bin, and indications of whether any content of the bin was identified as inaccurate (e.g., item type, item quantity, etc.)." "[0044] In block B, the computer system 102 can apply one or more models to the data 114. In some implementations, the computer system 102 can determine which model(s) to apply based on the distribution center 110. As described further below (e.g., refer to FIGS. 2 and 4A-B), a defective bin prediction model can be generated for each distribution center based on factors/features that are unique/specific to that distribution center." "[0078] In block 406, the computer system can identify data in the remaining retrieved data for bins flagged as not defective. Over a last 4 weeks, a distribution center can have a defect rate (e.g., percent of bins flagged as defective) between 0.5% and 2%. Since this means that approximately 98% of bins may be flagged as non-defective, the computer system can undersample the bins flagged as non-defective to generate a more balanced dataset for training purposes. [0079] In block 408, the computer system can identify data in the remaining retrieved data for N bins that are flagged as defective. The computer system can use stratification techniques and/or stratified random sampling (block 410). N can be a portion of the bins that is less than all of the bins flagged as defective. With stratified random sampling, a total population of the bins flagged as defective can be divided into smaller subgroups or strata. The strata can be formed based on shared attributes or characteristics of the bins. Stratified random sampling therefore allows for obtaining a sample population of the negative cases that best represents an entire population of the negative cases to generate a more balanced training dataset. [0080] In block 411, the computer system can combine the identified data into a positive defective bin dataset. This dataset can, for example, be a 50% positive defective bin dataset. By retrieving all cases of defective bins and a smattering of other cases of non-defective bins, as an example, the training dataset can have stronger signals of accuracy since the cases of defective bins may not be drowned out by an abundance of the other cases of non-defective bins.") Therefore, it would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Florimond and Reddy, because it allows for improved systems for a defective transaction prediction tool that may not require additional resources, such as workers, for random sampling and/or auditing of transactions in a warehouse and/or distribution center to then predict the defective transactions. Instead, the disclosed techniques can leverage existing inventory audit trails as sample data for training the machine learning models described herein. The inventory audit trails can include an abundance of pick data, pull data, and other types of data associated with the warehouses and/or distribution centers that can be used to accurate and proactively predict transaction defectiveness. (Reddy at Abstract and paras. 19-21). Florimond, Reddy, Rashid, and Seo do not explicitly teach, however, Gao does teach: a chronological order in the undersampled dataset. (Gao US20230342993 at paras. 34-37) ("[0035] Exemplary embodiments use a golden-angle-ordered radial k-space undersampled acquisition (shown as an example in FIG. 3), although any suitable MRI k-space data acquisition strategy can be used. Starting from 0°, each radial k-space spoke in undersampled k-space 30 is rotated by a golden angle, which is approximately 111.25°. In this example, five spokes are illustrated (labeled in time order 1-5), but it should be appreciated that many more spokes will be captured in practice, more than 100 in most embodiments. In some embodiments, substantially around 100 spokes (95-110) are captured, such that around 400 total spokes are captured or generated, making the resulting data set equivalent in size to existing fully-sampled radial capture schemes. The golden-angle acquisition ensures uniform coverage of the k-space for any arbitrary number of consecutive spokes so any newly acquired radial spoke is interleaved with the acquired ones. The spokes that are physically close (in azimuth angle) are dependent on each other. The angular dependencies between each of the spokes need not be given to a transformer network and can be learned by the attention modules in a transformer network to predict unacquired spokes and identify those acquired spokes that are most relevant to each spoke being predicted. For this reason, acquired spokes can be input to the network in any order. Some embodiments, rearrange the radial spokes based on the chronological order of acquisition and encode the temporal spoke index in the network. Note that the same concept can be extended to other non-Cartesian k-space trajectories such as the spiral trajectory or uniform-angle ordering. [0036] Graphs 32 and 34 shows pre-processing steps of radial k-space. For each single coil 2D slice, the golden-angle-ordered radial spokes are rearranged based on the chronological order of acquisition. The spokes in undersampled k-space 30 include a real part 35 and an imaginary part 36. These spokes can be ordered in any reasonable manor along an x-axis, representing the inputs to the transformer network, due to the self-attention layers in the network. In some embodiments, these are input in a chronological order. In some embodiments, the spokes can be order for 0-360 degrees. Graph 32 shows the simple case where spokes are arranged in chronological order, and only the real part 35 is input into the network. However, the transformer network can also process and predict the imaginary part of the k-space. Thus, in some embodiments, the input to the network is shown in graph 34, where the real part 35 is concatenated with the imaginary part 36 of each spoke, representing each spoke as a longer vector. In some embodiments, each of the k-space spokes is inverse Fourier transformed to generate a projection vector at each angle. The projection signals span the whole image compared to the centrally condensed k-space signals, so the network can better learn the image details and the reconstructed image is less sensitive to network errors. In some embodiments, the real and imaginary parts of each complex-valued projection are concatenated to a single long vector prior to input to the network. Because the frequency encoding line (kx) is typically two-times oversampled, the complex-valued projection vector is of length 512 and the concatenated projection vector is of length 1024. The input and output are of size 1024×100 with each projection vector as a token in the network.") Therefore, it would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Florimond, Reddy, Rashid, Seo, and Gao, because it allows for an improved system for removing or reducing artifacts within undersampled data sets. (Gao at Abstract and paras. 2-8). As per claim 13, Florimond does not explicitly teach, however, Reddy does teach: wherein the balanced random undersampling algorithm maintains (Reddy US20230351324 at paras. 15-16, 42-44, 78-80) ("[0042] The bin and worker data 114 can include a variety of information that may be recorded by workers as they complete their assigned tasks in the distribution center 110. For example, the data 114 can include, for each bin 106A-N in the distribution center 110, a quantity of workers who access the bin, worker history, and bin history. In some implementations, as shown in FIG. 1, the worker history can indicate a propensity for workers to misplace and/or miss-pick items from the bin. The bin history can also indicate how recent a worker was last at the bin, when information was last validated or verified for the bin, and indications of whether any content of the bin was identified as inaccurate (e.g., item type, item quantity, etc.)." "[0044] In block B, the computer system 102 can apply one or more models to the data 114. In some implementations, the computer system 102 can determine which model(s) to apply based on the distribution center 110. As described further below (e.g., refer to FIGS. 2 and 4A-B), a defective bin prediction model can be generated for each distribution center based on factors/features that are unique/specific to that distribution center." "[0078] In block 406, the computer system can identify data in the remaining retrieved data for bins flagged as not defective. Over a last 4 weeks, a distribution center can have a defect rate (e.g., percent of bins flagged as defective) between 0.5% and 2%. Since this means that approximately 98% of bins may be flagged as non-defective, the computer system can undersample the bins flagged as non-defective to generate a more balanced dataset for training purposes. [0079] In block 408, the computer system can identify data in the remaining retrieved data for N bins that are flagged as defective. The computer system can use stratification techniques and/or stratified random sampling (block 410). N can be a portion of the bins that is less than all of the bins flagged as defective. With stratified random sampling, a total population of the bins flagged as defective can be divided into smaller subgroups or strata. The strata can be formed based on shared attributes or characteristics of the bins. Stratified random sampling therefore allows for obtaining a sample population of the negative cases that best represents an entire population of the negative cases to generate a more balanced training dataset. [0080] In block 411, the computer system can combine the identified data into a positive defective bin dataset. This dataset can, for example, be a 50% positive defective bin dataset. By retrieving all cases of defective bins and a smattering of other cases of non-defective bins, as an example, the training dataset can have stronger signals of accuracy since the cases of defective bins may not be drowned out by an abundance of the other cases of non-defective bins.") Therefore, it would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Florimond and Reddy, because it allows for improved systems for a defective transaction prediction tool that may not require additional resources, such as workers, for random sampling and/or auditing of transactions in a warehouse and/or distribution center to then predict the defective transactions. Instead, the disclosed techniques can leverage existing inventory audit trails as sample data for training the machine learning models described herein. The inventory audit trails can include an abundance of pick data, pull data, and other types of data associated with the warehouses and/or distribution centers that can be used to accurate and proactively predict transaction defectiveness. (Reddy at Abstract and paras. 19-21). Florimond, Reddy, Rashid, and Seo do not explicitly teach, however, Gao does teach: a chronological order in the undersampled dataset. (Gao US20230342993 at paras. 34-37) ("[0035] Exemplary embodiments use a golden-angle-ordered radial k-space undersampled acquisition (shown as an example in FIG. 3), although any suitable MRI k-space data acquisition strategy can be used. Starting from 0°, each radial k-space spoke in undersampled k-space 30 is rotated by a golden angle, which is approximately 111.25°. In this example, five spokes are illustrated (labeled in time order 1-5), but it should be appreciated that many more spokes will be captured in practice, more than 100 in most embodiments. In some embodiments, substantially around 100 spokes (95-110) are captured, such that around 400 total spokes are captured or generated, making the resulting data set equivalent in size to existing fully-sampled radial capture schemes. The golden-angle acquisition ensures uniform coverage of the k-space for any arbitrary number of consecutive spokes so any newly acquired radial spoke is interleaved with the acquired ones. The spokes that are physically close (in azimuth angle) are dependent on each other. The angular dependencies between each of the spokes need not be given to a transformer network and can be learned by the attention modules in a transformer network to predict unacquired spokes and identify those acquired spokes that are most relevant to each spoke being predicted. For this reason, acquired spokes can be input to the network in any order. Some embodiments, rearrange the radial spokes based on the chronological order of acquisition and encode the temporal spoke index in the network. Note that the same concept can be extended to other non-Cartesian k-space trajectories such as the spiral trajectory or uniform-angle ordering. [0036] Graphs 32 and 34 shows pre-processing steps of radial k-space. For each single coil 2D slice, the golden-angle-ordered radial spokes are rearranged based on the chronological order of acquisition. The spokes in undersampled k-space 30 include a real part 35 and an imaginary part 36. These spokes can be ordered in any reasonable manor along an x-axis, representing the inputs to the transformer network, due to the self-attention layers in the network. In some embodiments, these are input in a chronological order. In some embodiments, the spokes can be order for 0-360 degrees. Graph 32 shows the simple case where spokes are arranged in chronological order, and only the real part 35 is input into the network. However, the transformer network can also process and predict the imaginary part of the k-space. Thus, in some embodiments, the input to the network is shown in graph 34, where the real part 35 is concatenated with the imaginary part 36 of each spoke, representing each spoke as a longer vector. In some embodiments, each of the k-space spokes is inverse Fourier transformed to generate a projection vector at each angle. The projection signals span the whole image compared to the centrally condensed k-space signals, so the network can better learn the image details and the reconstructed image is less sensitive to network errors. In some embodiments, the real and imaginary parts of each complex-valued projection are concatenated to a single long vector prior to input to the network. Because the frequency encoding line (kx) is typically two-times oversampled, the complex-valued projection vector is of length 512 and the concatenated projection vector is of length 1024. The input and output are of size 1024×100 with each projection vector as a token in the network.") Therefore, it would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Florimond, Reddy, Rashid, Seo, and Gao, because it allows for an improved system for removing or reducing artifacts within undersampled data sets. (Gao at Abstract and paras. 2-8). As per claim 14, Florimond, Reddy, and Rashid do not explicitly teach, however, Seo does teach: further comprising oversampling the undersampled dataset. (Seo WO2018097653A1 at pp. 11-12) ("The iteration step S60 may oversample or undersample the data for the chargeback fraud user of the training data until the performance of the prediction classification reaches the target value, thereby predicting classification S40 and The step of repeating the performance measurement step (S50). At this time, if the oversampling ratio is too high and the training data amount is excessively increased, an overload may occur when learning in the predictive classification step S40, and thus, the repeating step S60 may perform undersampling in addition to oversampling of the training data. Can be done. In addition, in the repeating step (S60), undersampling may be performed to reach a more accurate target value. The ratio of oversampling or undersampling at the time of performing the repetition step S60 may be regular or arbitrary, and is not particularly limited. For example, the oversampling ratio at the nth iteration can be defined as A × B n (where A and B are natural numbers and n is an integer). After the nth iteration, the undersampling ratio at the mth iteration can be defined as A × B n − (C × m) (where A, B and C are natural numbers, n and m are integers, n <m) That is, when "Recall with 0.940 or more" is set as the performance target value for the support vector machine SVM, the prediction classification step S40 and the performance measurement step S50 when the oversampling is 100% are performed. As a result, since Recall has 0.820 that is less than or equal to the target value, the oversampling is increased to 200% in the first iteration, and the prediction classification step S40 and the performance measurement step S50 are performed again. As a result, since Recall has 0.887 that is less than or equal to the target value, the oversampling is raised to 300% in the second iteration, and the prediction classification step S40 and the performance measurement step S50 are performed again. As a result, Recall has 0.948, which is above the target value. Subsequently, although it may be immediately skipped to the prediction step S70, in the iteration step S60, undersampling is performed in the third iteration, that is, the oversampling is set to less than 300%, for example, 280% to predict the classification step ( S40) and the performance measurement step S50 may be performed again.") Therefore, it would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Florimond, Reddy, Rashid, Seo, and Gao, because it allows for an improved system providing a machine learning model that satisfies the target prediction classification performance by processing transaction history data of a conventional user, and implements the implemented machine learning model. And a method and program for predicting chargeback fraud for a new user's transaction history data. (Seo at Abstract and pp. 2-3). As per claim 15, Florimond does not explicitly teach, however, Reddy does teach: wherein the balanced random undersampling algorithm maintains (Reddy US20230351324 at paras. 15-16, 42-44, 78-80) ("[0042] The bin and worker data 114 can include a variety of information that may be recorded by workers as they complete their assigned tasks in the distribution center 110. For example, the data 114 can include, for each bin 106A-N in the distribution center 110, a quantity of workers who access the bin, worker history, and bin history. In some implementations, as shown in FIG. 1, the worker history can indicate a propensity for workers to misplace and/or miss-pick items from the bin. The bin history can also indicate how recent a worker was last at the bin, when information was last validated or verified for the bin, and indications of whether any content of the bin was identified as inaccurate (e.g., item type, item quantity, etc.)." "[0044] In block B, the computer system 102 can apply one or more models to the data 114. In some implementations, the computer system 102 can determine which model(s) to apply based on the distribution center 110. As described further below (e.g., refer to FIGS. 2 and 4A-B), a defective bin prediction model can be generated for each distribution center based on factors/features that are unique/specific to that distribution center." "[0078] In block 406, the computer system can identify data in the remaining retrieved data for bins flagged as not defective. Over a last 4 weeks, a distribution center can have a defect rate (e.g., percent of bins flagged as defective) between 0.5% and 2%. Since this means that approximately 98% of bins may be flagged as non-defective, the computer system can undersample the bins flagged as non-defective to generate a more balanced dataset for training purposes. [0079] In block 408, the computer system can identify data in the remaining retrieved data for N bins that are flagged as defective. The computer system can use stratification techniques and/or stratified random sampling (block 410). N can be a portion of the bins that is less than all of the bins flagged as defective. With stratified random sampling, a total population of the bins flagged as defective can be divided into smaller subgroups or strata. The strata can be formed based on shared attributes or characteristics of the bins. Stratified random sampling therefore allows for obtaining a sample population of the negative cases that best represents an entire population of the negative cases to generate a more balanced training dataset. [0080] In block 411, the computer system can combine the identified data into a positive defective bin dataset. This dataset can, for example, be a 50% positive defective bin dataset. By retrieving all cases of defective bins and a smattering of other cases of non-defective bins, as an example, the training dataset can have stronger signals of accuracy since the cases of defective bins may not be drowned out by an abundance of the other cases of non-defective bins.") Therefore, it would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Florimond and Reddy, because it allows for improved systems for a defective transaction prediction tool that may not require additional resources, such as workers, for random sampling and/or auditing of transactions in a warehouse and/or distribution center to then predict the defective transactions. Instead, the disclosed techniques can leverage existing inventory audit trails as sample data for training the machine learning models described herein. The inventory audit trails can include an abundance of pick data, pull data, and other types of data associated with the warehouses and/or distribution centers that can be used to accurate and proactively predict transaction defectiveness. (Reddy at Abstract and paras. 19-21). Florimond, Reddy, Rashid, and Seo do not explicitly teach, however, Gao does teach: a chronological order in the undersampled dataset. (Gao US20230342993 at paras. 34-37) ("[0035] Exemplary embodiments use a golden-angle-ordered radial k-space undersampled acquisition (shown as an example in FIG. 3), although any suitable MRI k-space data acquisition strategy can be used. Starting from 0°, each radial k-space spoke in undersampled k-space 30 is rotated by a golden angle, which is approximately 111.25°. In this example, five spokes are illustrated (labeled in time order 1-5), but it should be appreciated that many more spokes will be captured in practice, more than 100 in most embodiments. In some embodiments, substantially around 100 spokes (95-110) are captured, such that around 400 total spokes are captured or generated, making the resulting data set equivalent in size to existing fully-sampled radial capture schemes. The golden-angle acquisition ensures uniform coverage of the k-space for any arbitrary number of consecutive spokes so any newly acquired radial spoke is interleaved with the acquired ones. The spokes that are physically close (in azimuth angle) are dependent on each other. The angular dependencies between each of the spokes need not be given to a transformer network and can be learned by the attention modules in a transformer network to predict unacquired spokes and identify those acquired spokes that are most relevant to each spoke being predicted. For this reason, acquired spokes can be input to the network in any order. Some embodiments, rearrange the radial spokes based on the chronological order of acquisition and encode the temporal spoke index in the network. Note that the same concept can be extended to other non-Cartesian k-space trajectories such as the spiral trajectory or uniform-angle ordering. [0036] Graphs 32 and 34 shows pre-processing steps of radial k-space. For each single coil 2D slice, the golden-angle-ordered radial spokes are rearranged based on the chronological order of acquisition. The spokes in undersampled k-space 30 include a real part 35 and an imaginary part 36. These spokes can be ordered in any reasonable manor along an x-axis, representing the inputs to the transformer network, due to the self-attention layers in the network. In some embodiments, these are input in a chronological order. In some embodiments, the spokes can be order for 0-360 degrees. Graph 32 shows the simple case where spokes are arranged in chronological order, and only the real part 35 is input into the network. However, the transformer network can also process and predict the imaginary part of the k-space. Thus, in some embodiments, the input to the network is shown in graph 34, where the real part 35 is concatenated with the imaginary part 36 of each spoke, representing each spoke as a longer vector. In some embodiments, each of the k-space spokes is inverse Fourier transformed to generate a projection vector at each angle. The projection signals span the whole image compared to the centrally condensed k-space signals, so the network can better learn the image details and the reconstructed image is less sensitive to network errors. In some embodiments, the real and imaginary parts of each complex-valued projection are concatenated to a single long vector prior to input to the network. Because the frequency encoding line (kx) is typically two-times oversampled, the complex-valued projection vector is of length 512 and the concatenated projection vector is of length 1024. The input and output are of size 1024×100 with each projection vector as a token in the network.") Therefore, it would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Florimond, Reddy, Rashid, Seo, and Gao, because it allows for an improved system for removing or reducing artifacts within undersampled data sets. (Gao at Abstract and paras. 2-8). As per claim 16, Florimond does not explicitly teach, however, Reddy does teach: wherein the balanced random undersampling algorithm maintains (Reddy US20230351324 at paras. 15-16, 42-44, 78-80) ("[0042] The bin and worker data 114 can include a variety of information that may be recorded by workers as they complete their assigned tasks in the distribution center 110. For example, the data 114 can include, for each bin 106A-N in the distribution center 110, a quantity of workers who access the bin, worker history, and bin history. In some implementations, as shown in FIG. 1, the worker history can indicate a propensity for workers to misplace and/or miss-pick items from the bin. The bin history can also indicate how recent a worker was last at the bin, when information was last validated or verified for the bin, and indications of whether any content of the bin was identified as inaccurate (e.g., item type, item quantity, etc.)." "[0044] In block B, the computer system 102 can apply one or more models to the data 114. In some implementations, the computer system 102 can determine which model(s) to apply based on the distribution center 110. As described further below (e.g., refer to FIGS. 2 and 4A-B), a defective bin prediction model can be generated for each distribution center based on factors/features that are unique/specific to that distribution center." "[0078] In block 406, the computer system can identify data in the remaining retrieved data for bins flagged as not defective. Over a last 4 weeks, a distribution center can have a defect rate (e.g., percent of bins flagged as defective) between 0.5% and 2%. Since this means that approximately 98% of bins may be flagged as non-defective, the computer system can undersample the bins flagged as non-defective to generate a more balanced dataset for training purposes. [0079] In block 408, the computer system can identify data in the remaining retrieved data for N bins that are flagged as defective. The computer system can use stratification techniques and/or stratified random sampling (block 410). N can be a portion of the bins that is less than all of the bins flagged as defective. With stratified random sampling, a total population of the bins flagged as defective can be divided into smaller subgroups or strata. The strata can be formed based on shared attributes or characteristics of the bins. Stratified random sampling therefore allows for obtaining a sample population of the negative cases that best represents an entire population of the negative cases to generate a more balanced training dataset. [0080] In block 411, the computer system can combine the identified data into a positive defective bin dataset. This dataset can, for example, be a 50% positive defective bin dataset. By retrieving all cases of defective bins and a smattering of other cases of non-defective bins, as an example, the training dataset can have stronger signals of accuracy since the cases of defective bins may not be drowned out by an abundance of the other cases of non-defective bins.") Therefore, it would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Florimond and Reddy, because it allows for improved systems for a defective transaction prediction tool that may not require additional resources, such as workers, for random sampling and/or auditing of transactions in a warehouse and/or distribution center to then predict the defective transactions. Instead, the disclosed techniques can leverage existing inventory audit trails as sample data for training the machine learning models described herein. The inventory audit trails can include an abundance of pick data, pull data, and other types of data associated with the warehouses and/or distribution centers that can be used to accurate and proactively predict transaction defectiveness. (Reddy at Abstract and paras. 19-21). Florimond, Reddy, Rashid, and Seo do not explicitly teach, however, Gao does teach: a chronological order in the undersampled dataset. (Gao US20230342993 at paras. 34-37) ("[0035] Exemplary embodiments use a golden-angle-ordered radial k-space undersampled acquisition (shown as an example in FIG. 3), although any suitable MRI k-space data acquisition strategy can be used. Starting from 0°, each radial k-space spoke in undersampled k-space 30 is rotated by a golden angle, which is approximately 111.25°. In this example, five spokes are illustrated (labeled in time order 1-5), but it should be appreciated that many more spokes will be captured in practice, more than 100 in most embodiments. In some embodiments, substantially around 100 spokes (95-110) are captured, such that around 400 total spokes are captured or generated, making the resulting data set equivalent in size to existing fully-sampled radial capture schemes. The golden-angle acquisition ensures uniform coverage of the k-space for any arbitrary number of consecutive spokes so any newly acquired radial spoke is interleaved with the acquired ones. The spokes that are physically close (in azimuth angle) are dependent on each other. The angular dependencies between each of the spokes need not be given to a transformer network and can be learned by the attention modules in a transformer network to predict unacquired spokes and identify those acquired spokes that are most relevant to each spoke being predicted. For this reason, acquired spokes can be input to the network in any order. Some embodiments, rearrange the radial spokes based on the chronological order of acquisition and encode the temporal spoke index in the network. Note that the same concept can be extended to other non-Cartesian k-space trajectories such as the spiral trajectory or uniform-angle ordering. [0036] Graphs 32 and 34 shows pre-processing steps of radial k-space. For each single coil 2D slice, the golden-angle-ordered radial spokes are rearranged based on the chronological order of acquisition. The spokes in undersampled k-space 30 include a real part 35 and an imaginary part 36. These spokes can be ordered in any reasonable manor along an x-axis, representing the inputs to the transformer network, due to the self-attention layers in the network. In some embodiments, these are input in a chronological order. In some embodiments, the spokes can be order for 0-360 degrees. Graph 32 shows the simple case where spokes are arranged in chronological order, and only the real part 35 is input into the network. However, the transformer network can also process and predict the imaginary part of the k-space. Thus, in some embodiments, the input to the network is shown in graph 34, where the real part 35 is concatenated with the imaginary part 36 of each spoke, representing each spoke as a longer vector. In some embodiments, each of the k-space spokes is inverse Fourier transformed to generate a projection vector at each angle. The projection signals span the whole image compared to the centrally condensed k-space signals, so the network can better learn the image details and the reconstructed image is less sensitive to network errors. In some embodiments, the real and imaginary parts of each complex-valued projection are concatenated to a single long vector prior to input to the network. Because the frequency encoding line (kx) is typically two-times oversampled, the complex-valued projection vector is of length 512 and the concatenated projection vector is of length 1024. The input and output are of size 1024×100 with each projection vector as a token in the network.") Therefore, it would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Florimond, Reddy, Rashid, Seo, and Gao, because it allows for an improved system for removing or reducing artifacts within undersampled data sets. (Gao at Abstract and paras. 2-8). As per claim 17, Florimond does not explicitly teach, however, Reddy does teach: wherein the balanced random undersampling algorithm maintains (Reddy US20230351324 at paras. 15-16, 42-44, 78-80) ("[0042] The bin and worker data 114 can include a variety of information that may be recorded by workers as they complete their assigned tasks in the distribution center 110. For example, the data 114 can include, for each bin 106A-N in the distribution center 110, a quantity of workers who access the bin, worker history, and bin history. In some implementations, as shown in FIG. 1, the worker history can indicate a propensity for workers to misplace and/or miss-pick items from the bin. The bin history can also indicate how recent a worker was last at the bin, when information was last validated or verified for the bin, and indications of whether any content of the bin was identified as inaccurate (e.g., item type, item quantity, etc.)." "[0044] In block B, the computer system 102 can apply one or more models to the data 114. In some implementations, the computer system 102 can determine which model(s) to apply based on the distribution center 110. As described further below (e.g., refer to FIGS. 2 and 4A-B), a defective bin prediction model can be generated for each distribution center based on factors/features that are unique/specific to that distribution center." "[0078] In block 406, the computer system can identify data in the remaining retrieved data for bins flagged as not defective. Over a last 4 weeks, a distribution center can have a defect rate (e.g., percent of bins flagged as defective) between 0.5% and 2%. Since this means that approximately 98% of bins may be flagged as non-defective, the computer system can undersample the bins flagged as non-defective to generate a more balanced dataset for training purposes. [0079] In block 408, the computer system can identify data in the remaining retrieved data for N bins that are flagged as defective. The computer system can use stratification techniques and/or stratified random sampling (block 410). N can be a portion of the bins that is less than all of the bins flagged as defective. With stratified random sampling, a total population of the bins flagged as defective can be divided into smaller subgroups or strata. The strata can be formed based on shared attributes or characteristics of the bins. Stratified random sampling therefore allows for obtaining a sample population of the negative cases that best represents an entire population of the negative cases to generate a more balanced training dataset. [0080] In block 411, the computer system can combine the identified data into a positive defective bin dataset. This dataset can, for example, be a 50% positive defective bin dataset. By retrieving all cases of defective bins and a smattering of other cases of non-defective bins, as an example, the training dataset can have stronger signals of accuracy since the cases of defective bins may not be drowned out by an abundance of the other cases of non-defective bins.") Therefore, it would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Florimond and Reddy, because it allows for improved systems for a defective transaction prediction tool that may not require additional resources, such as workers, for random sampling and/or auditing of transactions in a warehouse and/or distribution center to then predict the defective transactions. Instead, the disclosed techniques can leverage existing inventory audit trails as sample data for training the machine learning models described herein. The inventory audit trails can include an abundance of pick data, pull data, and other types of data associated with the warehouses and/or distribution centers that can be used to accurate and proactively predict transaction defectiveness. (Reddy at Abstract and paras. 19-21). Florimond, Reddy, Rashid, and Seo do not explicitly teach, however, Gao does teach: a chronological order in the undersampled dataset. (Gao US20230342993 at paras. 34-37) ("[0035] Exemplary embodiments use a golden-angle-ordered radial k-space undersampled acquisition (shown as an example in FIG. 3), although any suitable MRI k-space data acquisition strategy can be used. Starting from 0°, each radial k-space spoke in undersampled k-space 30 is rotated by a golden angle, which is approximately 111.25°. In this example, five spokes are illustrated (labeled in time order 1-5), but it should be appreciated that many more spokes will be captured in practice, more than 100 in most embodiments. In some embodiments, substantially around 100 spokes (95-110) are captured, such that around 400 total spokes are captured or generated, making the resulting data set equivalent in size to existing fully-sampled radial capture schemes. The golden-angle acquisition ensures uniform coverage of the k-space for any arbitrary number of consecutive spokes so any newly acquired radial spoke is interleaved with the acquired ones. The spokes that are physically close (in azimuth angle) are dependent on each other. The angular dependencies between each of the spokes need not be given to a transformer network and can be learned by the attention modules in a transformer network to predict unacquired spokes and identify those acquired spokes that are most relevant to each spoke being predicted. For this reason, acquired spokes can be input to the network in any order. Some embodiments, rearrange the radial spokes based on the chronological order of acquisition and encode the temporal spoke index in the network. Note that the same concept can be extended to other non-Cartesian k-space trajectories such as the spiral trajectory or uniform-angle ordering. [0036] Graphs 32 and 34 shows pre-processing steps of radial k-space. For each single coil 2D slice, the golden-angle-ordered radial spokes are rearranged based on the chronological order of acquisition. The spokes in undersampled k-space 30 include a real part 35 and an imaginary part 36. These spokes can be ordered in any reasonable manor along an x-axis, representing the inputs to the transformer network, due to the self-attention layers in the network. In some embodiments, these are input in a chronological order. In some embodiments, the spokes can be order for 0-360 degrees. Graph 32 shows the simple case where spokes are arranged in chronological order, and only the real part 35 is input into the network. However, the transformer network can also process and predict the imaginary part of the k-space. Thus, in some embodiments, the input to the network is shown in graph 34, where the real part 35 is concatenated with the imaginary part 36 of each spoke, representing each spoke as a longer vector. In some embodiments, each of the k-space spokes is inverse Fourier transformed to generate a projection vector at each angle. The projection signals span the whole image compared to the centrally condensed k-space signals, so the network can better learn the image details and the reconstructed image is less sensitive to network errors. In some embodiments, the real and imaginary parts of each complex-valued projection are concatenated to a single long vector prior to input to the network. Because the frequency encoding line (kx) is typically two-times oversampled, the complex-valued projection vector is of length 512 and the concatenated projection vector is of length 1024. The input and output are of size 1024×100 with each projection vector as a token in the network.") Therefore, it would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Florimond, Reddy, Rashid, Seo, and Gao, because it allows for an improved system for removing or reducing artifacts within undersampled data sets. (Gao at Abstract and paras. 2-8). As per claim 18, Florimond does not explicitly teach, however, Reddy does teach: wherein the balanced random undersampling algorithm maintains (Reddy US20230351324 at paras. 15-16, 42-44, 78-80) ("[0042] The bin and worker data 114 can include a variety of information that may be recorded by workers as they complete their assigned tasks in the distribution center 110. For example, the data 114 can include, for each bin 106A-N in the distribution center 110, a quantity of workers who access the bin, worker history, and bin history. In some implementations, as shown in FIG. 1, the worker history can indicate a propensity for workers to misplace and/or miss-pick items from the bin. The bin history can also indicate how recent a worker was last at the bin, when information was last validated or verified for the bin, and indications of whether any content of the bin was identified as inaccurate (e.g., item type, item quantity, etc.)." "[0044] In block B, the computer system 102 can apply one or more models to the data 114. In some implementations, the computer system 102 can determine which model(s) to apply based on the distribution center 110. As described further below (e.g., refer to FIGS. 2 and 4A-B), a defective bin prediction model can be generated for each distribution center based on factors/features that are unique/specific to that distribution center." "[0078] In block 406, the computer system can identify data in the remaining retrieved data for bins flagged as not defective. Over a last 4 weeks, a distribution center can have a defect rate (e.g., percent of bins flagged as defective) between 0.5% and 2%. Since this means that approximately 98% of bins may be flagged as non-defective, the computer system can undersample the bins flagged as non-defective to generate a more balanced dataset for training purposes. [0079] In block 408, the computer system can identify data in the remaining retrieved data for N bins that are flagged as defective. The computer system can use stratification techniques and/or stratified random sampling (block 410). N can be a portion of the bins that is less than all of the bins flagged as defective. With stratified random sampling, a total population of the bins flagged as defective can be divided into smaller subgroups or strata. The strata can be formed based on shared attributes or characteristics of the bins. Stratified random sampling therefore allows for obtaining a sample population of the negative cases that best represents an entire population of the negative cases to generate a more balanced training dataset. [0080] In block 411, the computer system can combine the identified data into a positive defective bin dataset. This dataset can, for example, be a 50% positive defective bin dataset. By retrieving all cases of defective bins and a smattering of other cases of non-defective bins, as an example, the training dataset can have stronger signals of accuracy since the cases of defective bins may not be drowned out by an abundance of the other cases of non-defective bins.") Therefore, it would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Florimond and Reddy, because it allows for improved systems for a defective transaction prediction tool that may not require additional resources, such as workers, for random sampling and/or auditing of transactions in a warehouse and/or distribution center to then predict the defective transactions. Instead, the disclosed techniques can leverage existing inventory audit trails as sample data for training the machine learning models described herein. The inventory audit trails can include an abundance of pick data, pull data, and other types of data associated with the warehouses and/or distribution centers that can be used to accurate and proactively predict transaction defectiveness. (Reddy at Abstract and paras. 19-21). Florimond, Reddy, Rashid, and Seo do not explicitly teach, however, Gao does teach: a chronological order in the undersampled dataset. (Gao US20230342993 at paras. 34-37) ("[0035] Exemplary embodiments use a golden-angle-ordered radial k-space undersampled acquisition (shown as an example in FIG. 3), although any suitable MRI k-space data acquisition strategy can be used. Starting from 0°, each radial k-space spoke in undersampled k-space 30 is rotated by a golden angle, which is approximately 111.25°. In this example, five spokes are illustrated (labeled in time order 1-5), but it should be appreciated that many more spokes will be captured in practice, more than 100 in most embodiments. In some embodiments, substantially around 100 spokes (95-110) are captured, such that around 400 total spokes are captured or generated, making the resulting data set equivalent in size to existing fully-sampled radial capture schemes. The golden-angle acquisition ensures uniform coverage of the k-space for any arbitrary number of consecutive spokes so any newly acquired radial spoke is interleaved with the acquired ones. The spokes that are physically close (in azimuth angle) are dependent on each other. The angular dependencies between each of the spokes need not be given to a transformer network and can be learned by the attention modules in a transformer network to predict unacquired spokes and identify those acquired spokes that are most relevant to each spoke being predicted. For this reason, acquired spokes can be input to the network in any order. Some embodiments, rearrange the radial spokes based on the chronological order of acquisition and encode the temporal spoke index in the network. Note that the same concept can be extended to other non-Cartesian k-space trajectories such as the spiral trajectory or uniform-angle ordering. [0036] Graphs 32 and 34 shows pre-processing steps of radial k-space. For each single coil 2D slice, the golden-angle-ordered radial spokes are rearranged based on the chronological order of acquisition. The spokes in undersampled k-space 30 include a real part 35 and an imaginary part 36. These spokes can be ordered in any reasonable manor along an x-axis, representing the inputs to the transformer network, due to the self-attention layers in the network. In some embodiments, these are input in a chronological order. In some embodiments, the spokes can be order for 0-360 degrees. Graph 32 shows the simple case where spokes are arranged in chronological order, and only the real part 35 is input into the network. However, the transformer network can also process and predict the imaginary part of the k-space. Thus, in some embodiments, the input to the network is shown in graph 34, where the real part 35 is concatenated with the imaginary part 36 of each spoke, representing each spoke as a longer vector. In some embodiments, each of the k-space spokes is inverse Fourier transformed to generate a projection vector at each angle. The projection signals span the whole image compared to the centrally condensed k-space signals, so the network can better learn the image details and the reconstructed image is less sensitive to network errors. In some embodiments, the real and imaginary parts of each complex-valued projection are concatenated to a single long vector prior to input to the network. Because the frequency encoding line (kx) is typically two-times oversampled, the complex-valued projection vector is of length 512 and the concatenated projection vector is of length 1024. The input and output are of size 1024×100 with each projection vector as a token in the network.") Therefore, it would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Florimond, Reddy, Rashid, Seo, and Gao, because it allows for an improved system for removing or reducing artifacts within undersampled data sets. (Gao at Abstract and paras. 2-8). Claim 6 is rejected under 35 U.S.C. 103 as being unpatentable over Florimond, U.S. Patent Application Publication Number 2017/0262852; in view of Reddy, U.S. Patent Application Publication Number US20230351324; in view of Rashid, NPL Publication 2002; in view of Seo, WIPO. Patent Application Publication Number 2018/097653; in view of Gao, U.S. Patent Application Publication Number 2023/0342993; in view of Baysal, U.S. Patent Application Publication Number 2017/0076375. As per claim 6, Florimond and Reddy do not explicitly teach, however, Rashid does teach: the complexity matrix is generated. (Rashid NPL 2022 at tables 3-4, pp. 4-6) ("III. RESULTS In the research, we will present the experiments which show the quality of our detection System. We discuss the results of experiments with various auto-encoder hyper parameters and show the impact of pre-training and feature scaling on the detection performance. Because we have access to the labels of the dataset, we are also going to present several performance measures to discuss the efficiency of our detection system along with comparison to existing work. A. Number of layers Although the number layers generally depend on the complexity of function we want to model with the neural network, picking the right number is not trivial. To illustrate that, we show three figures with different layer configurations, in each the number of neurons between layers’ decreases (and increases) by 5. To generate these curves, we have used the detection threshold of 0.001727. The f1 score was calculated against the same dataset and same threshold by using the labels we have. Figure 4 (left) shows a training phase of auto encoder with four layers (thus the latent representation has 20 dimensions). Figures 4 show auto-encoders which consist of six and ten layers respectively (thus the latent representations have 15 dimensions and 5 dimensions). We would favor the configuration on left, based on the shape of the curve and based on the maximum value of the mean squared feature recall." "B. Classification performance In this section we discuss the performance of our unsupervised detection system by using the labels from the dataset. Although we have already showed some performance measure in the previous sections, all of them used the whole dataset both for training and classification. This is not ideal, because we might over-fit on the data (e.g. by choosing certain parameters) and our solution might not generalize well. We would also like to classify incoming transactions immediately (so called online classification), but retraining the whole auto-encoder every time would be too slow. Because of these three reasons we used cross validation to simulate the situation of receiving an unseen dataset. We use the stratified k-fold method where k = 5, which means we split the dataset randomly into a training one, containing 80% original transactions, and testing one, containing 20% of the original transactions. Because the split is stratified, the ratio of fraudulent transactions remains always the same. This is done 5 times to obtain 5 different training and testing sets, which are used as an input to the detection system (with the detection threshold of 0.1727). We average the results and obtain a single confusion matrix which is depicted in Figure 6.") Therefore, it would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Florimond, Reddy, and Rashid, because it allows for an improved system to facilitate an algorithm that will improve current detection of fraud techniques and contribute to the reduction of fraud in healthcare. (Rashid at Abstract and pp. 1). Florimond, Reddy, Rashid, Seo, and Gao do not explicitly teach, however, Baysal does teach: further comprising performing an exchange rate standardization before (Baysal US20170076375 at paras. 36-41) ("[0036] In step 203, module 142 calculates FXEURUSD_UP and FXEURUSD_DN, where each of FXEURUSD_UP and FXEURUSD_DN is an exchange rate for converting EUR to USD, with FXEURUSD_UP being greater than FXEURUSD_DN. To calculate FXEURUSD_UP and FXEURUSD_DN, module 142 first calculates a value ΔFXEURUSD by taking the 99% quantile of 5-day log changes of the spot EUR-to-USD exchange rate standardized by exponentially weighted moving average volatility. Module 142 then calculates FXEURUSD_UP and FXEURUSD_DN by applying ΔFXEURUSD to the current EUR-to-USD spot rate (FXEURUSD_Spot)." "[0040] In some embodiments, JTDUP is calculated in USD and may estimate losses, apart from spread risk losses, that may result from a default of a single one of the entities represented in portfolio PF. In some such embodiments, and for each of j credit entities named in one or more of the PF positions, the algorithm of step 303 first calculates a value JTDUP(j). The value for JTDUP(j) represents an exposure of portfolio PF if there is a default of the jth credit entity and payouts are made based on that default. In some embodiments, JTDUP(j) value may equal the sum of (1) the mark-to-market (MTM) values, as of default, for all PF positions corresponding to a single name CDS naming the jth credit entity, and (2) the MTM values, as of default, for all PF positions corresponding to an index CDS position including the jth credit entity as a constituent, but with each of those index CDS positions assumed to have a notional equal to 1/R of the index CDS position actual notional (as of the time of JTDUP(j) calculation), with R being the number of names in the corresponding index CDS. If the jth entity is denominated in USD, then (1) and (2) are initially calculated in USD. If the jth entity is denominated in EUR, then (1) and (2) may be initially calculated in EUR and converted to USD using FXEURUSD_UP. Depending on the nature of each PF position naming the naming the jth credit entity, the value of JTDUPp(j) could be positive or negative.") Therefore, it would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Florimond, Reddy, Rashid, Seo, Gao, and Baysal, because it allows for an improved system to standardize monetary transaction values denominated in different currencies into a common currency, thereby facilitating consistent and meaningful evaluation of transaction costs when generating the cost-weighted complexity matrix or confusion matrix. (Baysal at Abstract and paras. 2-12). Claim 7 is rejected under 35 U.S.C. 103 as being unpatentable over Florimond, U.S. Patent Application Publication Number 2017/0262852; in view of Reddy, U.S. Patent Application Publication Number US20230351324; in view of Rashid, NPL Publication 2002; in view of Seo, WIPO. Patent Application Publication Number 2018/097653; in view of Gao, U.S. Patent Application Publication Number 2023/0342993; in view of Baysal, U.S. Patent Application Publication Number 2017/0076375; in view of Hansen, U.S. Patent Application Publication Number 2017/0169514. As per claim 7, Florimond and Reddy do not explicitly teach, however, Rashid does teach: the complexity matrix is generated. (Rashid NPL 2022 at tables 3-4, pp. 4-6) ("III. RESULTS In the research, we will present the experiments which show the quality of our detection System. We discuss the results of experiments with various auto-encoder hyper parameters and show the impact of pre-training and feature scaling on the detection performance. Because we have access to the labels of the dataset, we are also going to present several performance measures to discuss the efficiency of our detection system along with comparison to existing work. A. Number of layers Although the number layers generally depend on the complexity of function we want to model with the neural network, picking the right number is not trivial. To illustrate that, we show three figures with different layer configurations, in each the number of neurons between layers’ decreases (and increases) by 5. To generate these curves, we have used the detection threshold of 0.001727. The f1 score was calculated against the same dataset and same threshold by using the labels we have. Figure 4 (left) shows a training phase of auto encoder with four layers (thus the latent representation has 20 dimensions). Figures 4 show auto-encoders which consist of six and ten layers respectively (thus the latent representations have 15 dimensions and 5 dimensions). We would favor the configuration on left, based on the shape of the curve and based on the maximum value of the mean squared feature recall." "B. Classification performance In this section we discuss the performance of our unsupervised detection system by using the labels from the dataset. Although we have already showed some performance measure in the previous sections, all of them used the whole dataset both for training and classification. This is not ideal, because we might over-fit on the data (e.g. by choosing certain parameters) and our solution might not generalize well. We would also like to classify incoming transactions immediately (so called online classification), but retraining the whole auto-encoder every time would be too slow. Because of these three reasons we used cross validation to simulate the situation of receiving an unseen dataset. We use the stratified k-fold method where k = 5, which means we split the dataset randomly into a training one, containing 80% original transactions, and testing one, containing 20% of the original transactions. Because the split is stratified, the ratio of fraudulent transactions remains always the same. This is done 5 times to obtain 5 different training and testing sets, which are used as an input to the detection system (with the detection threshold of 0.1727). We average the results and obtain a single confusion matrix which is depicted in Figure 6.") Therefore, it would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Florimond, Reddy, and Rashid, because it allows for an improved system to facilitate an algorithm that will improve current detection of fraud techniques and contribute to the reduction of fraud in healthcare. (Rashid at Abstract and pp. 1). Florimond, Reddy, Rashid, Seo, Gao, and Baysal do not explicitly teach, however, Hansen does teach: further comprising performing a price normalization before (Hansen US20170169514 at paras. 51-59) ("[0055] Returning to FIG. 6A, the process continues to step 605, where the server 230 normalizes the price of each RFQ Line Item (tally) quote from each vendor. The normalization of the vendor's price is a computation that evaluates the vendor's price utilizing data from a metric. The normalization process is carried out because each vendor may respond to the Line Items of an RFQ by quoting products that are different from a buyer's RFQ and/or have a different tally configuration. The normalization of the pricing allows the buyers to objectively compare the relative value of the different products offered by the plurality of vendors. For example, one vendor may produce a quote for an RFQ of one unit of 2×4×10, two units of 2×4×12, and three units of 2×4×16. At the same time, another vendor may submit a quote for three units of 2×4×10, one unit of 2×4×12, and two units of 2×4×16. Even though there is some difference between these two offerings, the price normalization process provides a means for the buyer to effectively compare and evaluate the different quotes even though there are variations in the products. The price normalization process 900 is described in more detail below in conjunction with the flow diagram of FIG. 9.") Therefore, it would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Florimond, Reddy, Rashid, Seo, Gao, Baysal, and Hansen, because it allows for an improved system that allows buyers to evaluate the price of commodity offerings possessing varying shipping parameters. (Hansen at Abstract and paras. 2-12). Claims 19 and 20 are rejected under 35 U.S.C. 103 as being unpatentable over Florimond, U.S. Patent Application Publication Number 2017/0262852; in view of Reddy, U.S. Patent Application Publication Number US20230351324; in view of Rashid, NPL Publication 2002; in view of Seo, WIPO. Patent Application Publication Number 2018/097653; in view of Gao, U.S. Patent Application Publication Number 2023/0342993; in view of Hansen, U.S. Patent Application Publication Number 2017/0169514. As per claim 19, Florimond and Reddy do not explicitly teach, however, Rashid does teach: the complexity matrix is generated. (Rashid NPL 2022 at tables 3-4, pp. 4-6) ("III. RESULTS In the research, we will present the experiments which show the quality of our detection System. We discuss the results of experiments with various auto-encoder hyper parameters and show the impact of pre-training and feature scaling on the detection performance. Because we have access to the labels of the dataset, we are also going to present several performance measures to discuss the efficiency of our detection system along with comparison to existing work. A. Number of layers Although the number layers generally depend on the complexity of function we want to model with the neural network, picking the right number is not trivial. To illustrate that, we show three figures with different layer configurations, in each the number of neurons between layers’ decreases (and increases) by 5. To generate these curves, we have used the detection threshold of 0.001727. The f1 score was calculated against the same dataset and same threshold by using the labels we have. Figure 4 (left) shows a training phase of auto encoder with four layers (thus the latent representation has 20 dimensions). Figures 4 show auto-encoders which consist of six and ten layers respectively (thus the latent representations have 15 dimensions and 5 dimensions). We would favor the configuration on left, based on the shape of the curve and based on the maximum value of the mean squared feature recall." "B. Classification performance In this section we discuss the performance of our unsupervised detection system by using the labels from the dataset. Although we have already showed some performance measure in the previous sections, all of them used the whole dataset both for training and classification. This is not ideal, because we might over-fit on the data (e.g. by choosing certain parameters) and our solution might not generalize well. We would also like to classify incoming transactions immediately (so called online classification), but retraining the whole auto-encoder every time would be too slow. Because of these three reasons we used cross validation to simulate the situation of receiving an unseen dataset. We use the stratified k-fold method where k = 5, which means we split the dataset randomly into a training one, containing 80% original transactions, and testing one, containing 20% of the original transactions. Because the split is stratified, the ratio of fraudulent transactions remains always the same. This is done 5 times to obtain 5 different training and testing sets, which are used as an input to the detection system (with the detection threshold of 0.1727). We average the results and obtain a single confusion matrix which is depicted in Figure 6.") Therefore, it would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Florimond, Reddy, and Rashid, because it allows for an improved system to facilitate an algorithm that will improve current detection of fraud techniques and contribute to the reduction of fraud in healthcare. (Rashid at Abstract and pp. 1). Florimond, Reddy, Rashid, Seo, and Gao do not explicitly teach, however, Hansen does teach: further comprising performing a price normalization before (Hansen US20170169514 at paras. 51-59) ("[0055] Returning to FIG. 6A, the process continues to step 605, where the server 230 normalizes the price of each RFQ Line Item (tally) quote from each vendor. The normalization of the vendor's price is a computation that evaluates the vendor's price utilizing data from a metric. The normalization process is carried out because each vendor may respond to the Line Items of an RFQ by quoting products that are different from a buyer's RFQ and/or have a different tally configuration. The normalization of the pricing allows the buyers to objectively compare the relative value of the different products offered by the plurality of vendors. For example, one vendor may produce a quote for an RFQ of one unit of 2×4×10, two units of 2×4×12, and three units of 2×4×16. At the same time, another vendor may submit a quote for three units of 2×4×10, one unit of 2×4×12, and two units of 2×4×16. Even though there is some difference between these two offerings, the price normalization process provides a means for the buyer to effectively compare and evaluate the different quotes even though there are variations in the products. The price normalization process 900 is described in more detail below in conjunction with the flow diagram of FIG. 9.") Therefore, it would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Florimond, Reddy, Rashid, Seo, Gao, and Hansen, because it allows for an improved system that allows buyers to evaluate the price of commodity offerings possessing varying shipping parameters. (Hansen at Abstract and paras. 2-12). As per claim 20, Florimond and Reddy do not explicitly teach, however, Rashid does teach: the complexity matrix is generated. (Rashid NPL 2022 at tables 3-4, pp. 4-6) ("III. RESULTS In the research, we will present the experiments which show the quality of our detection System. We discuss the results of experiments with various auto-encoder hyper parameters and show the impact of pre-training and feature scaling on the detection performance. Because we have access to the labels of the dataset, we are also going to present several performance measures to discuss the efficiency of our detection system along with comparison to existing work. A. Number of layers Although the number layers generally depend on the complexity of function we want to model with the neural network, picking the right number is not trivial. To illustrate that, we show three figures with different layer configurations, in each the number of neurons between layers’ decreases (and increases) by 5. To generate these curves, we have used the detection threshold of 0.001727. The f1 score was calculated against the same dataset and same threshold by using the labels we have. Figure 4 (left) shows a training phase of auto encoder with four layers (thus the latent representation has 20 dimensions). Figures 4 show auto-encoders which consist of six and ten layers respectively (thus the latent representations have 15 dimensions and 5 dimensions). We would favor the configuration on left, based on the shape of the curve and based on the maximum value of the mean squared feature recall." "B. Classification performance In this section we discuss the performance of our unsupervised detection system by using the labels from the dataset. Although we have already showed some performance measure in the previous sections, all of them used the whole dataset both for training and classification. This is not ideal, because we might over-fit on the data (e.g. by choosing certain parameters) and our solution might not generalize well. We would also like to classify incoming transactions immediately (so called online classification), but retraining the whole auto-encoder every time would be too slow. Because of these three reasons we used cross validation to simulate the situation of receiving an unseen dataset. We use the stratified k-fold method where k = 5, which means we split the dataset randomly into a training one, containing 80% original transactions, and testing one, containing 20% of the original transactions. Because the split is stratified, the ratio of fraudulent transactions remains always the same. This is done 5 times to obtain 5 different training and testing sets, which are used as an input to the detection system (with the detection threshold of 0.1727). We average the results and obtain a single confusion matrix which is depicted in Figure 6.") Therefore, it would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Florimond, Reddy, and Rashid, because it allows for an improved system to facilitate an algorithm that will improve current detection of fraud techniques and contribute to the reduction of fraud in healthcare. (Rashid at Abstract and pp. 1). Florimond, Reddy, Rashid, Seo, and Gao do not explicitly teach, however, Hansen does teach: further comprising performing a price normalization before (Hansen US20170169514 at paras. 51-59) ("[0055] Returning to FIG. 6A, the process continues to step 605, where the server 230 normalizes the price of each RFQ Line Item (tally) quote from each vendor. The normalization of the vendor's price is a computation that evaluates the vendor's price utilizing data from a metric. The normalization process is carried out because each vendor may respond to the Line Items of an RFQ by quoting products that are different from a buyer's RFQ and/or have a different tally configuration. The normalization of the pricing allows the buyers to objectively compare the relative value of the different products offered by the plurality of vendors. For example, one vendor may produce a quote for an RFQ of one unit of 2×4×10, two units of 2×4×12, and three units of 2×4×16. At the same time, another vendor may submit a quote for three units of 2×4×10, one unit of 2×4×12, and two units of 2×4×16. Even though there is some difference between these two offerings, the price normalization process provides a means for the buyer to effectively compare and evaluate the different quotes even though there are variations in the products. The price normalization process 900 is described in more detail below in conjunction with the flow diagram of FIG. 9.") Therefore, it would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Florimond, Reddy, Rashid, Seo, Gao, and Hansen, because it allows for an improved system that allows buyers to evaluate the price of commodity offerings possessing varying shipping parameters. (Hansen at Abstract and paras. 2-12). Conclusion The prior art made of record and not relied upon is considered pertinent to applicant's disclosure and is available for review on Form PTO-892 Notice of References Cited. Any inquiry concerning this communication or earlier communications from the examiner should be directed to MERRITT J HASBROUCK whose telephone number is (571)272-3109. The examiner can normally be reached M-F 9:00-5:00. Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice. If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Christine Tran can be reached on 571-272-8103. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300. Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000. /MERRITT J HASBROUCK/Examiner, Art Unit 3695 /CHRISTINE M Tran/Supervisory Patent Examiner, Art Unit 3695
Read full office action

Prosecution Timeline

Aug 05, 2025
Application Filed
Oct 22, 2026
Non-Final Rejection mailed — §101, §103 (current)

Precedent Cases

Applications granted by this same examiner with similar technology

Patent 12639710
SYSTEMS AND TECHNIQUES TO UTILIZE AN ACTIVE LINK IN A UNIFORM RESOURCE LOCATOR TO PERFORM A MONEY EXCHANGE
5y 0m to grant Granted May 26, 2026
Patent 12299690
Systems and methods for tracking, predicting, and mitigating advanced persistent threats in networks
5y 8m to grant Granted May 13, 2025
Patent 12141784
SYSTEM FOR WHEELCHAIR-BASED NEAR FIELD COMMUNICATION (NFC) PAYMENT EXTENSION AND STANDARD
1y 4m to grant Granted Nov 12, 2024
Patent 12112369
TRANSMITTING PROACTIVE NOTIFICATIONS BASED ON MACHINE LEARNING MODEL PREDICTIONS
3y 4m to grant Granted Oct 08, 2024
Patent 11887102
TEMPORARY VIRTUAL PAYMENT CARD
4y 6m to grant Granted Jan 30, 2024
Study what changed to get past this examiner. Based on 5 most recent grants.

Strategy Recommendation AI-generated — please review before filing

Get a prosecution strategy drawn from examiner precedents, rejection analysis, and claim mapping.
Typically takes 5-10 seconds — AI-generated, attorney review required before filing

Prosecution Projections

1-2
Expected OA Rounds
10%
Grant Probability
18%
With Interview (+7.8%)
Low
PTA Risk
Based on 156 resolved cases by this examiner. Grant probability derived from career allowance rate.

Sign in with your work email

Enter your email to receive a magic link. No password needed.

Personal email addresses (Gmail, Yahoo, etc.) are not accepted.

Free tier: 3 strategy analyses per month