DETAILED ACTION
This nonfinal action is in response to application 18/524,614 filed on 11/30/2023.
Claims 1-20 are pending in the application. Claims 1, 8, and 15 are independent claims.
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 .
Claim Interpretation
As recited in MPEP § 2111, during patent examination, “the pending claims must
be given their broadest reasonable interpretation consistent with the specification”.
Under a broadest reasonable interpretation (BRI), claim terms must be given their plain
and ordinary meaning (i.e., the meaning that the term would have to a person of
ordinary skill in the art), unless applicant sets forth a special definition of a claim term
within the specification. The plain and ordinary meaning of a term “may be evidenced by
a variety of sources, including the words of the claims themselves, the specification,
drawings, and prior art”.
The claims recite the term “fuzzy attributes”, which does not appear to be a term commonly used in the prior art or a term with a standard meaning. The term is generally described in the specification as encompassing attributes that “do not have a fixed value” or “encompass a range of values” [¶ 00014]. Under broadest reasonable interpretation, the term is thereby interpreted to encompass any abstraction/generalization of data attributes that is represented in a non-fixed or ranged form (e.g., outputs of common data pre-processing techniques such as feature discretization or binning).
Claim Rejections - 35 USC § 112
The following is a quotation of 35 U.S.C. 112(b):
(b) CONCLUSION.—The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the inventor or a joint inventor regards as the invention.
The following is a quotation of 35 U.S.C. 112 (pre-AIA ), second paragraph:
The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the applicant regards as his invention.
Claims 1-7 and 15-20 are rejected under 35 U.S.C. 112(b) or 35 U.S.C. 112 (pre-AIA ), second paragraph, as being indefinite for failing to particularly point out and distinctly claim the subject matter which the inventor or a joint inventor (or for applications subject to pre-AIA 35 U.S.C. 112, the applicant), regards as the invention.
Regarding claims 1 and 15, they recite the limitation “generating, using a graph neural network, embeddings associated with the particular account based on the graph”. There is insufficient antecedent basis for the term “the particular account” in the claims; the claims previously recite “a plurality of accounts, wherein each account in the plurality of accounts is associated with one or more fuzzy attributes”, but do not previously recite or specify any “particular” account. Consequently, the intended scope of the claim is indefinite.
For purposes of examination, the limitation is interpreted as “generat[ing], using a graph neural network, embeddings associated with a particular account based on the graph”.
Regarding claims 2-7, and 16-20, they inherit the deficiencies of their parent claims. Consequently, they are also rejected under 35 U.S.C. 112(b) as being indefinite for depending on an indefinite parent claim.
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 an abstract idea without significantly more. The analysis of the claims will follow the 2019 Revised Patent Subject Matter Eligibility Guidance, 84 Fed. Reg. 50 (“2019 PEG”).
Independent Claims (Claim 1, Claim 8, Claim 15):
Step 1: Claim 1 is drawn to a system/apparatus, claim 8 is drawn to a method, and claim 15 is drawn to a product. Therefore, each of these claims falls under one of the four categories of statutory subject matter (process/method, machine/apparatus, manufacture/product, or composition of matter).
Step 2A Prong 1: Claims 1, 8, and 15 each recite a judicially recognized exception of an abstract idea.
Claim 1 recites, inter alia:
generating a graph based on the account data, wherein the graph represents relationships among the plurality of accounts and the set of fuzzy attributes – This limitation amounts to observing a set of data, including abstracted/generalized (i.e., fuzzy) data, and then constructing a graph to visually depict observed relationships, and thereby recites a process of evaluation that a human could reasonably perform using pen and paper.
identifying one or more communities of fuzzy attributes from the set of fuzzy attributes based on the graph – This limitation amounts to further observing groupings of data from a constructed graph, and thereby recites a process of evaluation that a human could reasonably perform in the mind or using pen and paper.
generating embeddings associated with the particular account based on the graph – This limitation amounts to a generic translation of observed information into a numerical representation, and thereby amounts to a series of algorithmic steps that could be reasonably performed by a human using pen and paper, and/or recites a process of mathematical calculation.
determining a risk score for the particular account based on the embeddings and the one or more communities of fuzzy attributes – This limitation amounts to an observation of calculated numerical representations to further determine a value representing “risk”, and thereby recites a process of evaluation that a human could reasonably perform using pen and paper.
performing an action associated with the particular account based on the risk score – This limitation amounts to a step of merely executing a generic action in response to a determined value, and thereby recites a process of evaluation that a human could reasonably perform in the mind or using pen and paper.
Claim 8 recites, inter alia:
accessing a graph representing relationships among the plurality of accounts and a set of fuzzy attributes; – This limitation amounts to observing a set of data, including abstracted/generalized (i.e., fuzzy) data, and then observing a constructed graph that visually depicts observed relationships, and thereby recites a process of evaluation that a human could reasonably perform using pen and paper
identifying one or more communities of fuzzy attributes from the set of fuzzy attributes based on the graph – This limitation amounts to further observing groupings of data from a constructed graph, and thereby recites a process of evaluation that a human could reasonably perform in the mind or using pen and paper.
generating embeddings associated with the particular account based on the graph – This limitation amounts to a generic translation of observed information into a numerical representation, and thereby amounts to a series of algorithmic steps that could be reasonably performed by a human using pen and paper, and/or recites a process of mathematical calculation.
determining a risk score for the particular account based on the embeddings and the one or more communities of fuzzy attributes – This limitation amounts to an observation of calculated numerical representations to further determine a value representing “risk”, and thereby recites a process of evaluation that a human could reasonably perform using pen and paper.
processing the transaction request based on the risk score – This limitation amounts to a process of updating a record of transactions (e.g., logging a transaction as “accepted” or “rejected”) in response to a determined value, and thereby recites a process of evaluation that a human could reasonably perform using pen and paper.
Claim 15 recites, inter alia:
generating a set of fuzzy attributes based on one or more attributes associated with each account in the plurality of accounts; – This limitation amounts to observing a set of account data and then modifying data to be more generalized/abstract (i.e., fuzzy) (e.g., as described in the specification [¶ 00014], changing an IP address value “192.0.1.31” to “192.0.1.X where ‘X’ can be any number), and thereby recites a process of evaluation that a human could reasonably perform using pen and paper.
generating a graph based on the account data, wherein the graph represents relationships among the plurality of accounts and the set of fuzzy attributes – This limitation amounts to observing a set of data, including abstracted/generalized (i.e., fuzzy) data, and then constructing a graph to visually depict observed relationships, and thereby recites a process of evaluation that a human could reasonably perform using pen and paper.
identifying one or more communities of fuzzy attributes from the set of fuzzy attributes based on the graph – This limitation amounts to further observing groupings of data from a constructed graph, and thereby recites a process of evaluation that a human could reasonably perform in the mind or using pen and paper.
generating embeddings associated with the particular account based on the graph – This limitation amounts to a generic translation of observed information into a numerical representation, and thereby amounts to a series of algorithmic steps that could be reasonably performed by a human using pen and paper, and/or recites a process of mathematical calculation.
determining a risk score for the particular account based on the embeddings and the one or more communities of fuzzy attributes – This limitation amounts to an observation of calculated numerical representations to further determine a value representing “risk”, and thereby recites a process of evaluation that a human could reasonably perform using pen and paper.
performing an action associated with the particular account based on the risk score – This limitation amounts to a step of merely executing a generic action in response to a determined value, and thereby recites a process of evaluation that a human could reasonably perform in the mind or using pen and paper.
Step 2A Prong 2: The following additional elements recited in claims 1, 8, and 15 also do not integrate the recited judicial exceptions into a practical application.
Claim 1 additionally recites:
A system, comprising: a non-transitory memory; and one or more hardware processors coupled with the non-transitory memory and configured to read instructions from the non-transitory memory to cause the system to perform operations – This limitation amounts to mere instructions to implement or “apply” an abstract idea on a computer or computer components.
accessing account data associated with a plurality of accounts, wherein each account in the plurality of accounts is associated with one or more fuzzy attributes from a set of fuzzy attributes; – This limitation amounts to no more than a pre-solution data gathering step utilized for enabling further analysis, and thereby recites insignificant extra-solution activity.
[generating embeddings] using a graph neural network – This limitation does no more than invoke a generic graph neural network as a mere tool to perform an existing abstract procedure, and thereby amounts to mere instructions to “apply” an exception.
Claim 8 additionally recites:
receiving a transaction request associated with a particular account from a plurality of accounts; – This limitation amounts to no more than a pre-solution data gathering step utilized for enabling further analysis, and thereby recites insignificant extra-solution activity.
[receiving/accessing/identifying/determining/processing] by a computer system – This limitation amounts to mere instructions to implement or “apply” an abstract idea on a computer or computer components.
[generating embeddings] using a graph neural network – This limitation does no more than invoke a generic graph neural network as a mere tool to perform an existing abstract procedure, and thereby amounts to mere instructions to “apply” an exception.
Claim 15 additionally recites:
A non-transitory machine-readable medium having stored therein machine-readable instructions executable to cause a machine to perform operations – This limitation amounts to mere instructions to implement or “apply” an abstract idea on a computer or computer components.
accessing account data associated with a plurality of accounts, – This limitation amounts to no more than a pre-solution data gathering step utilized for enabling further analysis, and thereby recites insignificant extra-solution activity.
[generating embeddings] using a graph neural network – This limitation does no more than invoke a generic graph neural network as a mere tool to perform an existing abstract procedure, and thereby amounts to mere instructions to “apply” an exception.
[determining a risk score] using a machine learning model – This limitation does no more than invoke a generic machine learning model as a mere tool to perform an existing abstract procedure, and thereby amounts to mere instructions to “apply” an exception.
Step 2B: The additional elements recited in claims 1, 8, and 15, viewed individually or as an ordered combination, do not provide an inventive concept or otherwise amount to significantly more than the recited abstract ideas themselves.
Claim 1 additionally recites:
A system, comprising: a non-transitory memory; and one or more hardware processors coupled with the non-transitory memory and configured to read instructions from the non-transitory memory to cause the system to perform operations – Mere instructions to implement or “apply” an abstract idea on a computer or computer components do not provide an inventive concept or significantly more to the recited abstract idea.
accessing account data associated with a plurality of accounts, wherein each account in the plurality of accounts is associated with one or more fuzzy attributes from a set of fuzzy attributes; – Receiving data is well-understood, routine, and conventional activity (see MPEP § 2106.05(d); “Receiving or transmitting data over a network”) and thereby does not provide an inventive concept or significantly more to the recited abstract idea.
[generating embeddings] using a graph neural network – Mere instructions to “apply” an exception on a generic graph neural network do not provide an inventive concept or significantly more to the recited abstract idea.
Claim 8 additionally recites:
receiving a transaction request associated with a particular account from a plurality of accounts; – Receiving data is well-understood, routine, and conventional activity (see MPEP § 2106.05(d); “Receiving or transmitting data over a network”) and thereby does not provide an inventive concept or significantly more to the recited abstract idea.
[receiving/accessing/identifying/determining/processing] by a computer system – Mere instructions to implement or “apply” an abstract idea on a computer or computer components do not provide an inventive concept or significantly more to the recited abstract idea.
[generating embeddings] using a graph neural network – Mere instructions to “apply” an exception on a generic graph neural network do not provide an inventive concept or significantly more to the recited abstract idea.
Claim 15 additionally recites:
A non-transitory machine-readable medium having stored therein machine-readable instructions executable to cause a machine to perform operations – Mere instructions to implement or “apply” an abstract idea on a computer or computer components do not provide an inventive concept or significantly more to the recited abstract idea..
accessing account data associated with a plurality of accounts, – Receiving data is well-understood, routine, and conventional activity (see MPEP § 2106.05(d); “Receiving or transmitting data over a network”) and thereby does not provide an inventive concept or significantly more to the recited abstract idea.
[generating embeddings] using a graph neural network – Mere instructions to “apply” an exception on a generic graph neural network do not provide an inventive concept or significantly more to the recited abstract idea.
[determining a risk score] using a machine learning model – Mere instructions to “apply” an exception on a generic machine learning model do not provide an inventive concept or significantly more to the recited abstract idea.
Even when considered as an ordered combination, the recited additional elements ultimately do no more than generically place the claims in the context of analyzing account data and utilizing generic graph neural networks and/or machine learning models to perform abstract steps. As such, claims 1, 8, and 15 are not patent eligible.
Dependent Claims (Claims 2-7, Claims 9-14, Claims 16-20):
Dependent claims 2-7, 9-14, and 16-20 narrow the scope of independent claims 1, 8, and 15, and likewise narrow the recited judicial exceptions. They recite abstract idea limitations that are similar to those recited within the independent claims (i.e., mental processes and/or mathematical concepts), and thereby merely expand on the already recited exceptions. The dependent claims also do not recite any further additional elements that successfully integrate the recited judicial exceptions into a practical application or provide significantly more than the recited abstract ideas themselves. Consequently, claims 2-7, 9-14, and 16-20 are also rejected under 35 U.S.C. 101.
Step 1: Claims 2-7 are drawn to a system/apparatus, claims 9-14 are drawn to a method, and claims 16-20 are drawn to a product. Therefore, each of these claims falls under one of the four categories of statutory subject matter (process/method, machine/apparatus, manufacture/product, or composition of matter).
Step 2A Prong 1: Claims 2-7, 9-14, and 16-20 each recite a judicially recognized exception of an abstract idea.
Claim 2 recites, inter alia:
determining an initial risk score associated with the particular account based on the embeddings, wherein the risk score is determined further based on the initial risk score – This limitation amounts to an observation of calculated numerical representations to determine a value representing “risk”, wherein the value determination involves more than one step. It thereby recites a process of evaluation that a human could reasonably perform using pen and paper.
Claim 3 recites, inter alia:
wherein each account in the plurality of accounts is associated with a corresponding set of attributes, and wherein the determining the initial risk score associated with the particular account is further based on the corresponding set of attributes associated with the particular account – This limitation amounts to an observation of data to determine a value representing “risk”, and thereby recites a process of evaluation that a human could reasonably perform using pen and paper.
Claim 4 recites, inter alia:
modifying the initial risk score based on characteristics associated with at least one of the one or more communities – This limitation amounts to a step of value determination involving modifying a previously determined value based on additional observed factors, and thereby recites a process of evaluation that a human could reasonably using pen and paper.
Claim 5 recites, inter alia:
wherein the plurality of accounts is associated with a set of attributes and wherein the operations further comprise: deriving the set of fuzzy attributes from the set of attributes based on the account data, wherein each fuzzy attribute in the set of fuzzy attributes represents an abstraction of a corresponding attribute in the set of attributes – This limitation amounts to observing a set of account data and then modifying data to be more generalized/abstract (i.e., fuzzy) (e.g., as described in the specification [¶ 00014], changing an IP address value “192.0.1.31” to “192.0.1.X where ‘X’ can be any number), and thereby recites a process of evaluation that a human could reasonably perform using pen and paper.
Claim 6 recites, inter alia:
wherein the graph comprises a first set of vertices representing the plurality of accounts, a second set of vertices representing the set of fuzzy attributes, and a set of edges representing the relationships among the plurality of accounts and the set of fuzzy attributes, wherein each edge in the set of edges connects a first vertex in the first set of vertex to a second vertex in the second set of vertices based on an association between a first account represented by the first vertex and a first fuzzy attribute represented by the second vertex – This limitation amounts to further specifying format and structure of the graph constructed based on observation of data, and thereby recites a process of evaluation that a human could reasonably perform using pen and paper.
Claim 7 recites, inter alia:
receiving a transaction request associated with the particular account, wherein the action comprises authorizing, requesting additional data, or denying the transaction request based on the risk score – This limitation amounts to a process of updating a record of transactions (e.g., logging a transaction as “accepted” or “rejected”) in response to a determined value, and thereby recites a process of evaluation that a human could reasonably perform using pen and paper.
Claim 9 recites, inter alia:
determining that the particular account is associated with a particular community of fuzzy attributes from the one or more communities of fuzzy attributes, wherein the risk score is determined further based on characteristics associated with the particular community of fuzzy attributes – This limitation amounts to further observing data to identify relationships/groupings and determining a value of “risk” based on the identified relationships, and thereby recites a process of evaluation that a human could reasonably perform using pen and paper.
Claim 10 recites, inter alia:
wherein the characteristics associated with the particular community of fuzzy attributes represent a distribution of different types of accounts associated with the particular community of fuzzy attributes – This limitation amounts to identification of relationships across different accounts based on shared attributes, and thereby recites a process of evaluation that a human could reasonably perform using pen and paper.
Claim 11 recites, inter alia:
wherein the different types of accounts comprise at least one of a first account type associated with a bad index or a second account type associated with a good index – This limitation amounts to identification of account types based on attributes, and thereby recites a process of evaluation that a human could reasonably perform using pen and paper.
Claim 12 recites, inter alia:
wherein the characteristics associated with the particular community of fuzzy attributes represent a distribution of different types of transactions associated with the particular community of fuzzy attributes – This limitation amounts to identification of relationships across different recorded transactions based on shared attributes, and thereby recites a process of evaluation that a human could reasonably perform using pen and paper.
Claim 13 recites, inter alia:
generating a modified graph based on the graph, wherein the modified graph represents attribute relationships among the set of fuzzy attributes based on common associated accounts, wherein the identifying the one or more communities of fuzzy attributes is further based on the modified graph – This limitation amounts to updating the previously constructed graph to represent newly observed relationships, and thereby recites a process of evaluation that a human could reasonably perform using pen and paper.
Claim 14 recites, inter alia:
wherein the plurality of accounts is associated with a set of attributes, and wherein the operations further comprise: deriving the set of fuzzy attributes from the set of attributes based on the account data, wherein each fuzzy attribute in the set of fuzzy attributes represents an abstraction of a corresponding attribute in the set of attributes –This limitation amounts to observing a set of account data and then modifying data to be more generalized/abstract (i.e., fuzzy) (e.g., as described in the specification [¶ 00014], changing an IP address value “192.0.1.31” to “192.0.1.X where ‘X’ can be any number), and thereby recites a process of evaluation that a human could reasonably perform using pen and paper.
Claim 16 recites, inter alia:
determining an initial risk score associated with the particular account based on the embeddings, wherein the risk score is determined further based on the initial risk score – This limitation amounts to an observation of calculated numerical representations to determine a value representing “risk”, wherein the value determination involves more than one step. It thereby recites a process of evaluation that a human could reasonably perform using pen and paper.
Claim 17 recites, inter alia:
wherein the determining the initial risk score associated with the particular account is further based on the corresponding one or more attributes associated with the particular account – This limitation amounts to an observation of data to determine a value representing “risk”, and thereby recites a process of evaluation that a human could reasonably perform using pen and paper.
Claim 18 recites, inter alia:
wherein the determining the risk score comprises modifying the initial risk score based on characteristics associated with at least one of the one or more communities – This limitation amounts to a step of value determination involving modifying a previously determined value based on additional observed factors, and thereby recites a process of evaluation that a human could reasonably using pen and paper.
Claim 19 recites, inter alia:
wherein the characteristics associated with the at least one of the one or more communities of fuzzy attributes represent a distribution of different types of transactions associated with the at least one of the one or more communities of fuzzy attributes – This limitation amounts to identification of relationships across different recorded transactions based on shared attributes, and thereby recites a process of evaluation that a human could reasonably perform using pen and paper.
Claim 20 recites, inter alia:
receiving a transaction request associated with the particular account, wherein the action comprises authorizing, requesting additional data, or denying the transaction request based on the risk score – This limitation amounts to a process of updating a record of transactions (e.g., logging a transaction as “accepted” or “rejected”) in response to a determined value, and thereby recites a process of evaluation that a human could reasonably perform using pen and paper.
Step 2A Prong 2: Claims 3-7, 9-14, and 17-20 do not recite any further additional elements besides those already recited in the independent claims, and the following additional elements recited in claims 2 and 16 also do not integrate the recited judicial exceptions into a practical application.
Claim 2 additionally recites:
[determining an initial risk score] using a machine learning model – This limitation does no more than invoke a generic machine learning model as a mere tool to perform an existing abstract procedure, and thereby amounts to mere instructions to “apply” an exception.
Claim 16 additionally recites:
[determining an initial risk score] using the machine learning model – This limitation does no more than invoke a generic machine learning model as a mere tool to perform an existing abstract procedure, and thereby amounts to mere instructions to “apply” an exception.
Step 2B: The additional elements recited in claims 2 and 16, viewed individually or as an ordered combination, do not provide an inventive concept or otherwise amount to significantly more than the recited abstract ideas themselves.
Claim 2 additionally recites:
[determining an initial risk score] using a machine learning model – Mere instructions to “apply” an exception on a generic machine learning model do not provide an inventive concept or significantly more to the recited abstract idea.
Claim 16 additionally recites:
[determining an initial risk score] using the machine learning model – Mere instructions to “apply” an exception on a generic machine learning model do not provide an inventive concept or significantly more to the recited abstract idea.
Even when considered as an ordered combination, the recited additional elements ultimately do no more than generically place the claims in the context of utilizing generic graph neural networks and/or machine learning models to perform abstract steps. As such, claims 2-7, 9-14, and 16-20 also 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 (i.e., changing from AIA to pre-AIA ) 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 1-20 are rejected under 35 U.S.C. 103 as being unpatentable over Lu et al. (“BRIGHT - Graph Neural Networks in Real-Time Fraud Detection”, available arXiv 24 Aug 2022), hereinafter Lu, in view of Cao et al. (“TitAnt: Online Realtime Transaction Fraud Detection in Ant Financial”, available arXiv 18 Jun 2019), hereinafter Cao, and Le et al., (Pub. No. US 20230134689 A1, “Minimizing Risks Posed to Online Services”, published 04 May 2023), hereinafter Le.
Regarding claim 1, Lu teaches A system (“Detecting fraudulent transactions is an essential component to control risk in e-commerce marketplaces. Apart from rule-based and machine learning filters that are already deployed in production, we want to enable efficient real-time inference with graph neural networks (GNNs), which is useful to catch multihop risk propagation in a transaction graph. However, two challenges arise in the implementation of GNNs in production. First, future information in a dynamic graph should not be considered in message passing to predict the past. Second, the latency of graph query and GNN model inference is usually up to hundreds of milliseconds, which is costly for some critical online services. To tackle these challenges, we propose a Batch and Real-time Inception GrapH Topology (BRIGHT) framework to conduct an end-to-end GNN learning that allows efficient online real-time inference. BRIGHT framework consists of a graph transformation module (Two-Stage Directed Graph) and a corresponding GNN architecture (Lambda Neural Network)… The Lambda Neural Network decouples inference into two stages: batch inference of entity embeddings and real-time inference of transaction prediction.” [Lu Abstract]), comprising:
a non-transitory memory (“Optimizing inference efficiency is another goal in our BRIGHT framework, which aims to minimize neighbor queries during GNN inference. In this evaluation, we compare LGB (Bench), end-to-end LNN (GCN), and RT Net... We use one machine for the database server and inference client. Regarding hardware, the machine is equipped with 32 Intel Xeon Gold 6230 CPUs, 64 GB memory, and one Nvidia Tesla V100 GPU” [Lu page 8 Experiment Setup]); and
one or more hardware processors coupled with the non-transitory memory and configured to read instructions from the non-transitory memory to cause the system to perform operations ([Lu page 8 Experiment Setup] as detailed above) comprising:
accessing account data associated with a plurality of accounts, wherein each account in the plurality of accounts is associated with one or more attributes from a set of attributes (“We conduct experiments on real-world transaction records sampled from eBay marketplace to evaluate the proposed framework. Our dataset contains 7 million labeled transactions.” [Lu page 7 Dataset and Preprocessing]; “In our previous generation of detection engine, it is observed that the entities linked to transactions, such as shipping addresses and device machine IDs, are key clues in detecting frauds. Based on these entities, hundreds of patterns are summarized as features in machine learning models or as rules in decision engines” [Lu page 1 Introduction]; The dataset is composed of real-world transaction records (i.e., data wherein each transaction is implicitly associated with a user/account), wherein each transaction is linked to a set of entities (i.e., attributes));
generating a graph based on the account data, wherein the graph represents relationships among the plurality of accounts and the set of attributes; (“In this work, transaction fraud detection is treated as a binary classification problem in an inductive setting on a dynamic heterogeneous graph. In a static transaction graph G (Fig. 1 (a)), a vertex 𝑣 ∈ V has a type 𝜏(𝑣) ∈ A, where A := {𝑡𝑟𝑎𝑛𝑠𝑎𝑐𝑡𝑖𝑜𝑛, 𝑒𝑛𝑡𝑖𝑡𝑦}. An edge 𝑒 ∈ E links from a 𝑡𝑟𝑎𝑛𝑠𝑎𝑐𝑡𝑖𝑜𝑛 vertex (𝑡𝑥𝑛 for short hereafter) to an 𝑒𝑛𝑡𝑖𝑡𝑦 vertex…To construct a Two-Stage Directed Graph (TD Graph) to support Lambda Neural Network (LNN), our graph construction consists of the following steps, as illustrated in Fig. 1. (a) Static Graph. The static graph is constructed from months of transaction logs by transforming them into a bi-graph (transaction-entity graph). Note that the edges here are bidirectional” [Lu page 3 Time-sensitive Graph Construction])
identifying one or more communities of attributes from the set of attributes based on the graph; (“(c) Two-Stage Directed Graph. A TD Graph is stored in target transaction partitions. A partition is made up of several timestamps that are the temporal snapshots in which all related transactions and entities are. A TD graph is equivalent to not only a snapshot graph, but also a simplified topology view for target transactions in the partition. The TD Graph is used for LNN learning and inference.” [Lu page 3 Time-sensitive Graph Construction]; “The mini-batch strategy is used in GNN learning. In each mini-batch, there is only one target partition. The maximum number of historical time snapshots (T) in the target partition is set to 30. The size of the mini-batch depends on the merged community8 node size, which is close to 32k. The transformed TD graph is used as input to the LNN model” [Lu page 8 Experiment Setup]; “Similar to ClusterGCN [4], graph clustering structure is used in the BRIGHT training process. Historical neighboring reference transactions are 2 hops away from the target transactions. For the connected components, whose node sizes are larger than 32k, they are cut into smaller communities through Louvain [2]. Before training, small communities are merged in the same target partition. The node sizes of the merged communities are close to 32k” [Lu page 8 footnote 8]; From the TD graph, which is determined from the static graph, a target partition (comprising identified communities of transactions and entities (i.e., attributes) related to a target transaction) is further determined for learning and inference)
generating, using a graph neural network, embeddings associated with the particular account based on the graph (“Two-Stage Directed (TD) graph simplifies the topology view from the target transaction side, making it easy to split the graph into Real-Time Graph G𝑅𝑇 𝑇 and Batch Graph G𝐵𝑎𝑡𝑐ℎ 𝑇 . We show in Fig. 1 (c) these two parts of the graphs that are taken later by LNN as input for various parts of the network. G𝑅𝑇 𝑇 is the subgraph with thick directed edges (the dark red circle), while G𝐵𝑎𝑡𝑐ℎ 𝑇 is the subgraph with bidirectional edges (the light red circle)…(2) Link 𝑡𝑥𝑛𝑟𝑒𝑓 𝑡 and 𝑒𝑛𝑡𝑖𝑡𝑦𝑡𝑔𝑡 𝑡 with bi-directed edges, which forms the Batch Graph G𝐵𝑎𝑡𝑐ℎ 𝑡 . It is used for batch inference of entity representations” [Lu page 4 Two-Stage Directed Graph]; “As illustrated in Fig. 2, LNN consists of a Batch Net (in blue) and a Real-Time (RT) Net (in orange). Batch Net is a stack of GNN layers, similar to DeepGCNs [14]. RT Net consists of a graph convolutional layer and a decoder, which could simply be a fully connected linear layer…The first stage is Batch Net inference, after which we obtain the entity embeddings learned from historical transactions and flush them into a key-value store” [Lu page 5 Network Architecture]; see Batch Net (highlighted in light blue) in Figure 2 – G𝐵𝑎𝑡𝑐ℎ 𝑇 is obtained from TD graph and processed by Batch Net GNN layers to generate entity embeddings [Lu page 5])
determining a risk score for the particular account based on the embeddings and the one or more communities of attributes (“The first stage is Batch Net inference, after which we obtain the entity embeddings learned from historical transactions and flush them into a key-value store. In the second stage, the RT Net fetches the entity embeddings from the key-value store, also takes the raw features from the target transactions, and then computes the transaction risk scores.”[Lu page 5 Network Architecture]; see RT Net (highlighted in orange) in Figure 2 – GRT 𝑇, obtained from TD graph, and entity embeddings, obtained from Batch Net output, are utilized to compute risk score for the target (i.e., particular) transaction (and thereby associated account) [Lu page 5])
However, Lu does not expressly teach utilization of fuzzy attributes, over exact attributes, for performing inference.
In the same field of endeavor, Cao teaches a means of applying machine learning techniques to enable detection of fraudulent transactions (“To tackle this problem, we introduce the TitAnt, a transaction fraud detection system deployed in Ant Financial, one of the largest Fintech companies in the world. The system is able to predict online real-time transaction fraud in mere milliseconds. We present the problem definition, feature extraction, detection methods, implementation and deployment of the system, as well as empirical effectiveness” [Cao Abstract]) that utilizes fuzzy attributes for performing inference (“We set 100 trees for IF and raw basic features are fed as attributes. As rule-based ID3 and C5.0 cannot support continuous values well, we discretize the data into different bins [32]” [Cao page 7 Experimental Setup]).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have incorporated utilization of fuzzy attributes as taught by Cao into Lu because they are both directed towards applying machine learning techniques to enable detection of fraudulent transactions. Feature discretization/binning is already a well-known, commonly utilized data pre-processing technique in the art for purposes of speeding up convergence and minimizing the influence of outliers on training (see Galli, “Data discretization in machine learning” [pages 1-3]). As such, a person of ordinary skill in the art would recognize the value of incorporating this technique to obtain known benefits and achieve a boost in predictive performance, as demonstrated in Cao for comparable learning procedures (“C5.0 has better than ID3 by 6.9% on average, probably because it takes better data discretization and segmentation mechanisms such as Gain Ratio. LR is implemented with discretization pre- processing which tremendously improves performance” [Cao page 8 Empirical Results on Transaction Fraud Detection]).
However, the combination of Lu and Cao does not expressly teach performing an action associated with the particular account based on the risk score.
In the same field of endeavor, Le teaches a means of applying machine learning techniques to enable detection of fraudulent transactions (“One innovative aspect of the subject matter described in this disclosure can be implemented as a method for selectively advancing funds based on risks posed by transactions associated with an online service. The method can be performed by one or more processors of a system coupled to or associated with the online service, and can include receiving a payment request for a transaction between a vendor and a consumer, and sending a first request to a database associated with the online service for historical transactions and personal attributes of the vendor concurrently with sending a second request to a number of third-party services for credit information and personal attributes of the consumer. The method includes receiving information responsive to the first and second requests from the database and the third-party services, respectively. The method includes obtaining a risk score for the transaction based on an application of one or more risk assessment rules to the received information by a machine learning model trained with at least the historical transactions and the personal attributes of the vendor” [Le ¶ 0005]) that perform[s] an action associated with the particular account based on the risk score (“The method includes determining whether or not to advance funds to the vendor for the transaction, prior to sending a payment request to a financial service designated by the consumer, based at least in part on the risk score” [Le ¶ 0005]; “In some implementations, the method may also include declining the transaction without sending the payment request to the designated financial service when the risk score is greater than a first level, and informing the designated financial service of the declined transaction when the risk score is greater than the first level by more than an amount or percentage. The method may also include sending the payment request to the designated financial service when the risk score is less than a second level, where the first level is greater than the second level, and advancing funds for the transaction to the vendor prior to or concurrently with sending the payment request to the designated financial service when the risk score is less than the second level by more than an amount or percentage“ [Le ¶ 0006]).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have incorporated performing an action associated with the particular account based on the risk score as taught by Le into the combination of Lu and Cao because they are all directed towards applying machine learning techniques to enable detection of fraudulent transactions. Given that Lu already discusses applicability of the disclosed architecture for real-time inference of transaction risk (“Similar to the two-“tower" model architecture [5, 24, 25, 27–29] (see Sec. 6), we design LNN to support two stages of processing (batch and real-time inference)… (Real-Time Inference Stage.) For real-time risk assessment, the second stage of LNN (namely RT Net) calculates the score by Eq. 4.” [Lu pages 6-7 Network Architecture]), a person of ordinary skill in the art would recognize the value of leveraging the calculated risk score to further determine appropriate action in real time, thereby handling associated transaction services efficiently (“By advancing funds to vendors only for transactions that pose less than a certain risk to the online service, implementations of the subject matter disclosed herein can minimize the number and amount of fraudulent chargeback transactions associated with the online service” [Le ¶ 0034]).
Regarding claim 2, the combination of Lu, Cao, and Le teaches the limitations of parent claim 1, and Le further teaches determining, using a machine learning model, an initial risk score associated with the particular account based on the embeddings, wherein the risk score is determined further based on the initial risk score (“When the risk score is between the first and second levels, indicating that the corresponding transaction has not been classified as acceptable or unacceptable, the risk assessment engine 118 can perform additional processing to determine the appropriate action to take. In some instances, the risk assessment engine 118 may send the payment request to the designated financial service 150 (without advancing funds to the vendor), thereby allowing the designated financial service 150 to determine whether to process the transaction. In other instances, the risk assessment engine 118 may update the structured data set to include additional input data (e.g., credit information of the vendor, additional personal attributes of the vendor, additional personal attributes of the consumer, etc.), and then generate another risk score based on the updated structured data set” [Le ¶ 0083]).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have incorporated determining, using a machine learning model, an initial risk score associated with the particular account based on the embeddings, wherein the risk score is determined further based on the initial risk score as taught by Le into the combination because Lu, Cao, and Le are all directed towards applying machine learning techniques to enable detection of fraudulent transactions. Given that feature discretization/binning is known to risk information loss due to inherent loss of precision (see Galli, “Data discretization in machine learning” [pages 1-3]), a person of ordinary skill in the art would understand the value of leveraging more granular forms of input data to update calculated risk score in instances where the appropriate associated action is not yet conclusively determined ([Le ¶ 0083]).
Regarding claim 3, the combination of Lu, Cao, and Le teaches the limitations of parent claim 2, and Lu further teaches wherein each account in the plurality of accounts is associated with a corresponding set of attributes (“In our previous generation of detection engine, it is observed that the entities linked to transactions, such as shipping addresses and device machine IDs, are key clues in detecting frauds. Based on these entities, hundreds of patterns are summarized as features in machine learning models or as rules in decision engines” [Lu page 1 Introduction]; “To collect neighbor features to assess the risks of transaction fraud, multiple entities used in the checkout sessions are considered neighbors of 𝑡𝑥𝑛 nodes. These entities, including shipping address, E-mail, IP address, device ID, contact phone, payment token, and user account, are represented as 𝑒𝑛𝑡𝑖𝑡𝑦 nodes in G” [Lu page 3 Static Graph Construction]). The dataset is composed of real-world transaction records (i.e., data wherein each transaction is implicitly associated with a user/account), wherein each transaction is linked to a set of entities (i.e., attributes)), wherein the determining the initial risk score associated with the particular account is further based on the corresponding set of attributes associated with the particular account (“The first stage is Batch Net inference, after which we obtain the entity embeddings learned from historical transactions and flush them into a key-value store. In the second stage, the RT Net fetches the entity embeddings from the key-value store, also takes the raw features from the target transactions, and then computes the transaction risk scores.”[Lu page 5 Network Architecture]; see RT Net (highlighted in orange) in Figure 2 – GRT 𝑇, obtained from TD graph, and entity embeddings, obtained from Batch Net output, are utilized to compute risk score for the target transaction (and thereby associated account) [Lu page 5]).
Regarding claim 4, the combination of Lu, Cao, and Le teaches the limitations of parent claim 2, and Le further teaches wherein the determining the risk score comprises modifying the initial risk score based on characteristics, (“In other instances, the risk assessment engine 118 may update the structured data set to include additional input data (e.g., credit information of the vendor, additional personal attributes of the vendor, additional personal attributes of the consumer, etc.), and then generate another risk score based on the updated structured data set” [Le ¶ 0083]) wherein the characteristics are associated with at least one of the one or more communities taught in Lu (“To collect neighbor features to assess the risks of transaction fraud, multiple entities used in the checkout sessions are considered neighbors of 𝑡𝑥𝑛 nodes. These entities, including shipping address, E-mail, IP address, device ID, contact phone, payment token, and user account, are represented as 𝑒𝑛𝑡𝑖𝑡𝑦 nodes in G” [Lu page 3 Static Graph Construction]; ““(c) Two-Stage Directed Graph. A TD Graph is stored in target transaction partitions. A partition is made up of several timestamps that are the temporal snapshots in which all related transactions and entities are.” [Lu page 3 Time-sensitive Graph Construction]; The characteristics utilized in Le to modify initial risk score (e.g., additional personal attributes of the vendor, additional personal attributes of the consumer) correspond to value types of the entity nodes (e.g., shipping address, E-mail, IP address, etc.) that make up snapshots of partitions (i.e., communities) for each target transaction in Lu).
Regarding claim 5, the combination of Lu, Cao, and Le teaches the limitations of parent claim 1, and Lu further teaches wherein the plurality of accounts is associated with a set of attributes ([Lu page 1 Introduction] as detailed in claim 1 above; The dataset is composed of real-world transaction records (i.e., data wherein each transaction is implicitly associated with a user/account), wherein each transaction is linked to a set of entities (i.e., attributes)). Cao further teaches deriving the set of fuzzy attributes from the set of attributes based on the account data, wherein each fuzzy attribute in the set of fuzzy attributes represents an abstraction of a corresponding attribute in the set of attributes ([Cao page 7 Experimental Setup] as detailed in claim 1 above; Feature binning inherently results in an abstraction of original feature attributes).
Regarding claim 6, the combination of Lu, Cao, and Le teaches the limitations of parent claim 1, and Lu further teaches wherein the graph comprises a first set of vertices representing the plurality of accounts, a second set of vertices representing the set of attributes, and a set of edges representing the relationships among the plurality of accounts and the set of attributes, wherein each edge in the set of edges connects a first vertex in the first set of vertex to a second vertex in the second set of vertices based on an association between a first account represented by the first vertex and a first attribute represented by the second vertex (“Each 𝑡𝑥𝑛 vertex represents a checkout transaction along with a unique transaction ID, linked with multiple 𝑒𝑛𝑡𝑖𝑡𝑦 vertices such as shipping addresses, E-mails, contact phones that buyers need to confirm on checkout pages. Most of the 𝑒𝑛𝑡𝑖𝑡𝑦 vertices are also linked to multiple 𝑡𝑥𝑛 vertices. Given a set of target transactions, a static graph G can be constructed from their records. A transaction record could be broken down into a 𝑡𝑥𝑛 vertex and several 𝑒𝑛𝑡𝑖𝑡𝑦 vertices. Edges 𝑒 are placed between entities and transactions that use these entities” [Lu page 3 Static Graph Construction]; Edges representing relationships are placed between a set of txn vertices representing transactions (and their associated accounts therein) and a set of entity vertices). Cao further teaches utilization of fuzzy attributes over exact attributes ([Cao page 7 Experimental Setup]), as detailed in claim 1 above.
Regarding claim 7, the combination of Lu, Cao, and Le teaches the limitations of parent claim 1, and Le further teaches receiving a transaction request associated with the particular account, wherein the action comprises authorizing, requesting additional data, or denying the transaction request based on the risk score (“The method can be performed by one or more processors of a system coupled to or associated with the online service, and can include receiving a payment request for a transaction between a vendor and a consumer…In some implementations, the method may also include declining the transaction without sending the payment request to the designated financial service when the risk score is greater than a first level, and informing the designated financial service of the declined transaction when the risk score is greater than the first level by more than an amount or percentage. The method may also include sending the payment request to the designated financial service when the risk score is less than a second level, where the first level is greater than the second level, and advancing funds for the transaction to the vendor prior to or concurrently with sending the payment request to the designated financial service when the risk score is less than the second level by more than an amount or percentage.” [Le ¶ 0006]; “When the risk score is between the first and second levels, indicating that the corresponding transaction has not been classified as acceptable or unacceptable, the risk assessment engine 118 can perform additional processing to determine the appropriate action to take…the risk assessment engine 118 may update the structured data set to include additional input data” [Le ¶ 0083]).
Regarding claim 8, Lu teaches A method (“Detecting fraudulent transactions is an essential component to control risk in e-commerce marketplaces. Apart from rule-based and machine learning filters that are already deployed in production, we want to enable efficient real-time inference with graph neural networks (GNNs), which is useful to catch multihop risk propagation in a transaction graph. However, two challenges arise in the implementation of GNNs in production. First, future information in a dynamic graph should not be considered in message passing to predict the past. Second, the latency of graph query and GNN model inference is usually up to hundreds of milliseconds, which is costly for some critical online services. To tackle these challenges, we propose a Batch and Real-time Inception GrapH Topology (BRIGHT) framework to conduct an end-to-end GNN learning that allows efficient online real-time inference. BRIGHT framework consists of a graph transformation module (Two-Stage Directed Graph) and a corresponding GNN architecture (Lambda Neural Network)… The Lambda Neural Network decouples inference into two stages: batch inference of entity embeddings and real-time inference of transaction prediction.” [Lu Abstract]) comprising:
accessing, by a computer system, a graph representing relationships among a plurality of accounts and a set of attributes (“In this work, transaction fraud detection is treated as a binary classification problem in an inductive setting on a dynamic heterogeneous graph. In a static transaction graph G (Fig. 1 (a)), a vertex 𝑣 ∈ V has a type 𝜏(𝑣) ∈ A, where A := {𝑡𝑟𝑎𝑛𝑠𝑎𝑐𝑡𝑖𝑜𝑛, 𝑒𝑛𝑡𝑖𝑡𝑦}. An edge 𝑒 ∈ E links from a 𝑡𝑟𝑎𝑛𝑠𝑎𝑐𝑡𝑖𝑜𝑛 vertex (𝑡𝑥𝑛 for short hereafter) to an 𝑒𝑛𝑡𝑖𝑡𝑦 vertex…To construct a Two-Stage Directed Graph (TD Graph) to support Lambda Neural Network (LNN), our graph construction consists of the following steps, as illustrated in Fig. 1. (a) Static Graph. The static graph is constructed from months of transaction logs by transforming them into a bi-graph (transaction-entity graph). Note that the edges here are bidirectional…Each 𝑡𝑥𝑛 vertex represents a checkout transaction along with a unique transaction ID, linked with multiple 𝑒𝑛𝑡𝑖𝑡𝑦 vertices such as shipping addresses, E-mails, contact phones that buyers need to confirm on checkout pages.” [Lu page 3 Time-sensitive Graph Construction]);
identifying, by the computer system, one or more communities of attributes from the set of attributes based on the graph; (“(c) Two-Stage Directed Graph. A TD Graph is stored in target transaction partitions. A partition is made up of several timestamps that are the temporal snapshots in which all related transactions and entities are. A TD graph is equivalent to not only a snapshot graph, but also a simplified topology view for target transactions in the partition. The TD Graph is used for LNN learning and inference.” [Lu page 3 Time-sensitive Graph Construction]; “The mini-batch strategy is used in GNN learning. In each mini-batch, there is only one target partition. The maximum number of historical time snapshots (T) in the target partition is set to 30. The size of the mini-batch depends on the merged community8 node size, which is close to 32k. The transformed TD graph is used as input to the LNN model” [Lu page 8 Experiment Setup]; “Similar to ClusterGCN [4], graph clustering structure is used in the BRIGHT training process. Historical neighboring reference transactions are 2 hops away from the target transactions. For the connected components, whose node sizes are larger than 32k, they are cut into smaller communities through Louvain [2]. Before training, small communities are merged in the same target partition. The node sizes of the merged communities are close to 32k” [Lu page 8 footnote 8]; From the TD graph, which is determined from the static graph, a target partition (comprising identified communities of transactions and entities (i.e., attributes) related to a target transaction) is further determined for learning and inference)
generating, using a graph neural network, embeddings associated with a particular account based on the graph; (“Two-Stage Directed (TD) graph simplifies the topology view from the target transaction side, making it easy to split the graph into Real-Time Graph G𝑅𝑇 𝑇 and Batch Graph G𝐵𝑎𝑡𝑐ℎ 𝑇 . We show in Fig. 1 (c) these two parts of the graphs that are taken later by LNN as input for various parts of the network. G𝑅𝑇 𝑇 is the subgraph with thick directed edges (the dark red circle), while G𝐵𝑎𝑡𝑐ℎ 𝑇 is the subgraph with bidirectional edges (the light red circle)…(2) Link 𝑡𝑥𝑛𝑟𝑒𝑓 𝑡 and 𝑒𝑛𝑡𝑖𝑡𝑦𝑡𝑔𝑡 𝑡 with bi-directed edges, which forms the Batch Graph G𝐵𝑎𝑡𝑐ℎ 𝑡 . It is used for batch inference of entity representations” [Lu page 4 Two-Stage Directed Graph]; “As illustrated in Fig. 2, LNN consists of a Batch Net (in blue) and a Real-Time (RT) Net (in orange). Batch Net is a stack of GNN layers, similar to DeepGCNs [14]. RT Net consists of a graph convolutional layer and a decoder, which could simply be a fully connected linear layer…The first stage is Batch Net inference, after which we obtain the entity embeddings learned from historical transactions and flush them into a key-value store” [Lu page 5 Network Architecture]; see Batch Net (highlighted in light blue) in Figure 2 – G𝐵𝑎𝑡𝑐ℎ 𝑇 is obtained from TD graph and processed by Batch Net GNN layers to generate entity embeddings [Lu page 5])
determining, by the computer system, a risk score for the particular account based on the embeddings and the one or more communities of attributes. (“The first stage is Batch Net inference, after which we obtain the entity embeddings learned from historical transactions and flush them into a key-value store. In the second stage, the RT Net fetches the entity embeddings from the key-value store, also takes the raw features from the target transactions, and then computes the transaction risk scores.”[Lu page 5 Network Architecture]; see RT Net (highlighted in orange) in Figure 2 – GRT 𝑇, obtained from TD graph, and entity embeddings, obtained from Batch Net output, are utilized to compute risk score for the target (i.e., particular) transaction (and thereby associated account) [Lu page 5])
However, Lu does not expressly teach utilization of fuzzy attributes, over exact attributes, for performing inference.
In the same field of endeavor, Cao teaches a means of applying machine learning techniques to enable detection of fraudulent transactions (“To tackle this problem, we introduce the TitAnt, a transaction fraud detection system deployed in Ant Financial, one of the largest Fintech companies in the world. The system is able to predict online real-time transaction fraud in mere milliseconds. We present the problem definition, feature extraction, detection methods, implementation and deployment of the system, as well as empirical effectiveness” [Cao Abstract]) that utilizes fuzzy attributes for performing inference (“We set 100 trees for IF and raw basic features are fed as attributes. As rule-based ID3 and C5.0 cannot support continuous values well, we discretize the data into different bins [32]” [Cao page 7 Experimental Setup]).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have incorporated utilization of fuzzy attributes as taught by Cao into Lu because they are both directed towards applying machine learning techniques to enable detection of fraudulent transactions. Feature discretization/binning is already a well-known, commonly utilized data pre-processing technique in the art for purposes of speeding up convergence and minimizing the influence of outliers on training (see Galli, “Data discretization in machine learning” [pages 1-3]). As such, a person of ordinary skill in the art would recognize the value of incorporating this technique to obtain known benefits and achieve a boost in predictive performance, as demonstrated in Cao for comparable learning procedures (“C5.0 has better than ID3 by 6.9% on average, probably because it takes better data discretization and segmentation mechanisms such as Gain Ratio. LR is implemented with discretization pre- processing which tremendously improves performance” [Cao page 8 Empirical Results on Transaction Fraud Detection]).
However, the combination of Lu and Cao does not expressly teach receiving, by a computer system, a transaction request associated with a particular account from a plurality of accounts and processing, by the computer system, the transaction request based on the risk score.
In the same field of endeavor, Le teaches a means of applying machine learning techniques to enable detection of fraudulent transactions (“One innovative aspect of the subject matter described in this disclosure can be implemented as a method for selectively advancing funds based on risks posed by transactions associated with an online service. The method can be performed by one or more processors of a system coupled to or associated with the online service, and can include receiving a payment request for a transaction between a vendor and a consumer, and sending a first request to a database associated with the online service for historical transactions and personal attributes of the vendor concurrently with sending a second request to a number of third-party services for credit information and personal attributes of the consumer. The method includes receiving information responsive to the first and second requests from the database and the third-party services, respectively. The method includes obtaining a risk score for the transaction based on an application of one or more risk assessment rules to the received information by a machine learning model trained with at least the historical transactions and the personal attributes of the vendor” [Le ¶ 0005]) that receiv[es], by the computer system, a transaction request associated with a particular account from a plurality of accounts (“The method can be performed by one or more processors of a system coupled to or associated with the online service, and can include receiving a payment request for a transaction between a vendor and a consumer” [Le ¶ 0005]) and process[es], by the computer system, the transaction request based on the risk score (“The method includes determining whether or not to advance funds to the vendor for the transaction, prior to sending a payment request to a financial service designated by the consumer, based at least in part on the risk score” [Le ¶ 0005]; “In some implementations, the method may also include declining the transaction without sending the payment request to the designated financial service when the risk score is greater than a first level, and informing the designated financial service of the declined transaction when the risk score is greater than the first level by more than an amount or percentage. The method may also include sending the payment request to the designated financial service when the risk score is less than a second level, where the first level is greater than the second level, and advancing funds for the transaction to the vendor prior to or concurrently with sending the payment request to the designated financial service when the risk score is less than the second level by more than an amount or percentage“ [Le ¶ 0006]).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have incorporated receiving, by a computer system, a transaction request associated with a particular account from a plurality of accounts and processing, by the computer system, the transaction request based on the risk score as taught by Le into the combination of Lu and Cao because they are all directed towards applying machine learning techniques to enable detection of fraudulent transactions. Given that Lu already discusses applicability of the disclosed architecture for real-time inference of transaction risk (“Similar to the two-“tower" model architecture [5, 24, 25, 27–29] (see Sec. 6), we design LNN to support two stages of processing (batch and real-time inference)… (Real-Time Inference Stage.) For real-time risk assessment, the second stage of LNN (namely RT Net) calculates the score by Eq. 4.” [Lu pages 6-7 Network Architecture]), a person of ordinary skill in the art would recognize the value of handling real-time transaction requests and leveraging the calculated risk score to further determine appropriate action in real time, thereby handling associated transaction services efficiently (“By advancing funds to vendors only for transactions that pose less than a certain risk to the online service, implementations of the subject matter disclosed herein can minimize the number and amount of fraudulent chargeback transactions associated with the online service” [Le ¶ 0034]).
Regarding claim 9, the combination of Lu, Cao, and Le teaches the limitations of parent claim 8, and Lu further teaches determining that the particular account is associated with a particular community of attributes from the one or more communities of attributes, wherein the risk score is determined further based on characteristics associated with the particular community of attributes (In the training phase, both the Batch Graph and the RT Graph are used for end-to-end LNN training. In each mini-batch, we sample one partition for the time window T, where T := {0, 1, ..., 𝑡 }. The timestamp of each target transaction is within the partition time window. All timestamps for reference transactions are prior to those for target transactions… As input, RT Net takes features from target transactions and entity embeddings learned from Batch Net inference…Furthermore, RT Net isolates the aggregation of messages between different target transaction nodes 𝑡𝑥𝑛 𝑡𝑔𝑡 𝑡 within the same target snapshot partition… The risk score for target transaction 𝑡𝑥𝑛𝑡𝑔𝑡 𝑡 is decoded from the real-time transaction embedding ℎ𝑅𝑇 𝑡𝑥𝑛𝑡𝑔𝑡 𝑡 through a fully connected linear layer” [Lu pages 5-6 End-to-end Training]; Each target transaction (and associated account therein) has a risk score determined based on messages aggregated from its respective partition at time window t (i.e., particular community)). Cao further teaches utilization of fuzzy attributes over exact attributes ([Cao page 7 Experimental Setup]), as detailed in claim 8 above.
Regarding claim 10, the combination of Lu, Cao, and Le teaches the limitations of parent claim 9, and Lu further teaches wherein the characteristics associated with the particular community of attributes represent a distribution of different types of accounts associated with the particular community of attributes (“The historical transactions that share the same entities 𝑒𝑛𝑡𝑖𝑡𝑦 used in the target transaction, such as emails and shipping addresses, we call them reference transactions, which do not have any labels in our experiments” [Lu page 3 Time-sensitive Graph Construction]; All reference transaction nodes sharing common entities are stored with the target transactions in the same partition. On the TD graph, there is only one role for transactions, either a target transaction or a reference transaction. There are multiple snapshots 𝑠𝑛𝑎𝑝𝑠ℎ𝑜𝑡𝑖 in one target partition 𝑝𝑎𝑟𝑡𝑖𝑡𝑖𝑜𝑛𝑡 , 𝑖 ≤ 𝑡 , as illustrated in Fig. 2 (a)... In Fig. 2 (a), we show an illustrative example of the risk assessment of the transaction on 𝑡𝑥𝑛𝑡𝑔𝑡 𝑡 using TD Graph. The features of historical reference transactions 𝑡𝑥𝑛𝑟𝑒 𝑓 𝑡 are sent to 𝑒𝑛𝑡𝑖𝑡𝑦𝑡𝑔𝑡 𝑡 (phone and buyer in this example). They are the entity nodes 1-hop away from the target transaction. After bi-directed message-passing through the dotted edges, the embeddings (for phone and buyer in this example) of 1-hop neighbors are obtained. Then these embeddings are sent to the target transaction 𝑡𝑥𝑛𝑡𝑔𝑡 𝑡 through the directed edges (in red), which forms the graph used in real-time inference” [Lu pages 5-6 Two-Stage Directed Graph]; see snapshots t (highlighted in orange) and t-1, t-2, t-3 (highlighted in blue) in Figure 2(a) – messages for target transaction at snapshot t are aggregated from reference transactions at different snapshots t-1, t-2, t-3 (i.e., different types of transactions, and accounts therein) into embeddings of 1-hop entity neighbors (i.e., community of attributes) [Lu page 5]). Cao further teaches utilization of fuzzy attributes over exact attributes ([Cao page 7 Experimental Setup]), as detailed in claim 8 above.
Regarding claim 11, the combination of Lu, Cao, and Le teaches the limitations of parent claim 10, and Lu further teaches wherein the different types of accounts comprise at least one of a first account type associated with a bad index or a second account type associated with a good index (“Nodes 𝑡𝑥𝑛 with unauthenticated chargeback claims from the customer system are marked as 1, which are considered fraudulent transactions. The others are marked as 0, which represents legitimate transactions. These labels are used in our binary classification task” [Lu page 3 Time-sensitive Graph Constructions]; Transactions (and their accounts therein) are classified as either legitimate (i.e., good index) or fraudulent (i.e., bad index)).
Regarding claim 12, the combination of Lu, Cao, and Le teaches the limitations of parent claim 9, and Lu further teaches wherein the characteristics associated with the particular community of attributes represent a distribution of different types of transactions associated with the particular community of attributes (“The historical transactions that share the same entities 𝑒𝑛𝑡𝑖𝑡𝑦 used in the target transaction, such as emails and shipping addresses, we call them reference transactions, which do not have any labels in our experiments” [Lu page 3 Time-sensitive Graph Construction]; All reference transaction nodes sharing common entities are stored with the target transactions in the same partition. On the TD graph, there is only one role for transactions, either a target transaction or a reference transaction. There are multiple snapshots 𝑠𝑛𝑎𝑝𝑠ℎ𝑜𝑡𝑖 in one target partition 𝑝𝑎𝑟𝑡𝑖𝑡𝑖𝑜𝑛𝑡 , 𝑖 ≤ 𝑡 , as illustrated in Fig. 2 (a)... In Fig. 2 (a), we show an illustrative example of the risk assessment of the transaction on 𝑡𝑥𝑛𝑡𝑔𝑡 𝑡 using TD Graph. The features of historical reference transactions 𝑡𝑥𝑛𝑟𝑒 𝑓 𝑡 are sent to 𝑒𝑛𝑡𝑖𝑡𝑦𝑡𝑔𝑡 𝑡 (phone and buyer in this example). They are the entity nodes 1-hop away from the target transaction. After bi-directed message-passing through the dotted edges, the embeddings (for phone and buyer in this example) of 1-hop neighbors are obtained. Then these embeddings are sent to the target transaction 𝑡𝑥𝑛𝑡𝑔𝑡 𝑡 through the directed edges (in red), which forms the graph used in real-time inference” [Lu pages 5-6 Two-Stage Directed Graph]; see snapshots t (highlighted in orange) and t-1, t-2, t-3 (highlighted in blue) in Figure 2(a) – messages for target transaction at snapshot t are aggregated from reference transactions at different snapshots t-1, t-2, t-3 (i.e., different types of transactions) into embeddings of 1-hop entity neighbors (i.e., community of attributes) [Lu page 5]). Cao further teaches utilization of fuzzy attributes over exact attributes ([Cao page 7 Experimental Setup]), as detailed in claim 8 above.
Regarding claim 13, the combination of Lu, Cao, and Le teaches the limitations of parent claim 8, and Lu further teaches generating a modified graph based on the graph, wherein the modified graph represents attribute relationships among the set of attributes based on common associated accounts, wherein the identifying the one or more communities of attributes is further based on the modified graph (“To construct a Two-Stage Directed Graph (TD Graph) to support Lambda Neural Network (LNN), our graph construction consists of the following steps, as illustrated in Fig. 1. a) Static Graph. The static graph is constructed from months of transaction logs by transforming them into a bi-graph (transaction-entity graph). Note that the edges here are bidirectional. • (b) Snapshot Graph. Each node in the static graph is placed in its corresponding timestamp snapshot. Reference transactions are linked to entities that are on the same timestamp snapshot as the corresponding target transactions. • (c) Two-Stage Directed Graph. A TD Graph is stored in target transaction partitions. A partition is made up of several timestamps that are the temporal snapshots in which all related transactions and entities are. A TD graph is equivalent to not only a snapshot graph, but also a simplified topology view for target transactions in the partition. The TD Graph is used for LNN learning and inference” [Lu page 3 Time-sensitive Graph Construction]; From the static graph, a snapshot graph (i.e., modified graph) of related entities and transactions is generated based on an identified corresponding snapshot for each target transaction (i.e., community), wherein related reference transactions (and their associated accounts therein) are further linked based on shared entities (i.e, attributes)). Cao further teaches utilization of fuzzy attributes over exact attributes ([Cao page 7 Experimental Setup]), as detailed in claim 8 above.
Regarding claim 14, the combination of Lu, Cao, and Le teaches the limitations of parent claim 8, and Lu further teaches wherein the plurality of accounts is associated with a set of attributes ([Lu page 1 Introduction] as detailed in claim 8 above; The dataset is composed of real-world transaction records (i.e., data wherein each transaction is implicitly associated with a user/account), wherein each transaction is linked to a set of entities (i.e., attributes)). Cao further teaches deriving the set of fuzzy attributes from the set of attributes based on the account data, wherein each fuzzy attribute in the set of fuzzy attributes represents an abstraction of a corresponding attribute in the set of attributes ([Cao page 7 Experimental Setup] as detailed in claim 8 above; Feature binning inherently results in an abstraction of original feature attributes).
Regarding claim 15, Lu teaches A non-transitory machine-readable medium having stored therein machine-readable instructions (“Optimizing inference efficiency is another goal in our BRIGHT framework, which aims to minimize neighbor queries during GNN inference. In this evaluation, we compare LGB (Bench), end-to-end LNN (GCN), and RT Net... We use one machine for the database server and inference client. Regarding hardware, the machine is equipped with 32 Intel Xeon Gold 6230 CPUs, 64 GB memory, and one Nvidia Tesla V100 GPU” [Lu page 8 Experiment Setup])) executable to cause a machine to perform operations (“Detecting fraudulent transactions is an essential component to control risk in e-commerce marketplaces. Apart from rule-based and machine learning filters that are already deployed in production, we want to enable efficient real-time inference with graph neural networks (GNNs), which is useful to catch multihop risk propagation in a transaction graph. However, two challenges arise in the implementation of GNNs in production. First, future information in a dynamic graph should not be considered in message passing to predict the past. Second, the latency of graph query and GNN model inference is usually up to hundreds of milliseconds, which is costly for some critical online services. To tackle these challenges, we propose a Batch and Real-time Inception GrapH Topology (BRIGHT) framework to conduct an end-to-end GNN learning that allows efficient online real-time inference. BRIGHT framework consists of a graph transformation module (Two-Stage Directed Graph) and a corresponding GNN architecture (Lambda Neural Network)… The Lambda Neural Network decouples inference into two stages: batch inference of entity embeddings and real-time inference of transaction prediction.” [Lu Abstract]) comprising:
accessing account data associated with a plurality of accounts; (“We conduct experiments on real-world transaction records sampled from eBay marketplace to evaluate the proposed framework. Our dataset contains 7 million labeled transactions.” [Lu page 7 Dataset and Preprocessing]; “In our previous generation of detection engine, it is observed that the entities linked to transactions, such as shipping addresses and device machine IDs, are key clues in detecting frauds. Based on these entities, hundreds of patterns are summarized as features in machine learning models or as rules in decision engines” [Lu page 1 Introduction]; The dataset is composed of real-world transaction records (i.e., data wherein each transaction is implicitly associated with a user/account), wherein each transaction is linked to a set of entities (i.e., attributes));
generating a graph based on the account data, wherein the graph represents relationships among the plurality of accounts and a set of attributes [associated with each account in the plurality of accounts]; (“In this work, transaction fraud detection is treated as a binary classification problem in an inductive setting on a dynamic heterogeneous graph. In a static transaction graph G (Fig. 1 (a)), a vertex 𝑣 ∈ V has a type 𝜏(𝑣) ∈ A, where A := {𝑡𝑟𝑎𝑛𝑠𝑎𝑐𝑡𝑖𝑜𝑛, 𝑒𝑛𝑡𝑖𝑡𝑦}. An edge 𝑒 ∈ E links from a 𝑡𝑟𝑎𝑛𝑠𝑎𝑐𝑡𝑖𝑜𝑛 vertex (𝑡𝑥𝑛 for short hereafter) to an 𝑒𝑛𝑡𝑖𝑡𝑦 vertex…To construct a Two-Stage Directed Graph (TD Graph) to support Lambda Neural Network (LNN), our graph construction consists of the following steps, as illustrated in Fig. 1. (a) Static Graph. The static graph is constructed from months of transaction logs by transforming them into a bi-graph (transaction-entity graph). Note that the edges here are bidirectional” [Lu page 3 Time-sensitive Graph Construction])
identifying one or more communities of attributes from the set of attributes based on the graph; (“(c) Two-Stage Directed Graph. A TD Graph is stored in target transaction partitions. A partition is made up of several timestamps that are the temporal snapshots in which all related transactions and entities are. A TD graph is equivalent to not only a snapshot graph, but also a simplified topology view for target transactions in the partition. The TD Graph is used for LNN learning and inference.” [Lu page 3 Time-sensitive Graph Construction]; “The mini-batch strategy is used in GNN learning. In each mini-batch, there is only one target partition. The maximum number of historical time snapshots (T) in the target partition is set to 30. The size of the mini-batch depends on the merged community8 node size, which is close to 32k. The transformed TD graph is used as input to the LNN model” [Lu page 8 Experiment Setup]; “Similar to ClusterGCN [4], graph clustering structure is used in the BRIGHT training process. Historical neighboring reference transactions are 2 hops away from the target transactions. For the connected components, whose node sizes are larger than 32k, they are cut into smaller communities through Louvain [2]. Before training, small communities are merged in the same target partition. The node sizes of the merged communities are close to 32k” [Lu page 8 footnote 8]; From the TD graph, which is determined from the static graph, a target partition (comprising identified communities of transactions and entities (i.e., attributes) related to a target transaction) is further determined for learning and inference)
generating, using a graph neural network, embeddings associated with the particular account based on the graph; (“Two-Stage Directed (TD) graph simplifies the topology view from the target transaction side, making it easy to split the graph into Real-Time Graph G𝑅𝑇 𝑇 and Batch Graph G𝐵𝑎𝑡𝑐ℎ 𝑇 . We show in Fig. 1 (c) these two parts of the graphs that are taken later by LNN as input for various parts of the network. G𝑅𝑇 𝑇 is the subgraph with thick directed edges (the dark red circle), while G𝐵𝑎𝑡𝑐ℎ 𝑇 is the subgraph with bidirectional edges (the light red circle)…(2) Link 𝑡𝑥𝑛𝑟𝑒𝑓 𝑡 and 𝑒𝑛𝑡𝑖𝑡𝑦𝑡𝑔𝑡 𝑡 with bi-directed edges, which forms the Batch Graph G𝐵𝑎𝑡𝑐ℎ 𝑡 . It is used for batch inference of entity representations” [Lu page 4 Two-Stage Directed Graph]; “As illustrated in Fig. 2, LNN consists of a Batch Net (in blue) and a Real-Time (RT) Net (in orange). Batch Net is a stack of GNN layers, similar to DeepGCNs [14]. RT Net consists of a graph convolutional layer and a decoder, which could simply be a fully connected linear layer…The first stage is Batch Net inference, after which we obtain the entity embeddings learned from historical transactions and flush them into a key-value store” [Lu page 5 Network Architecture]; see Batch Net (highlighted in light blue) in Figure 2 – G𝐵𝑎𝑡𝑐ℎ 𝑇 is obtained from TD graph and processed by Batch Net GNN layers to generate entity embeddings [Lu page 5])
determining, using a machine learning model, a risk score for the particular account based on the embeddings and the one or more communities of attributes; (“The first stage is Batch Net inference, after which we obtain the entity embeddings learned from historical transactions and flush them into a key-value store. In the second stage, the RT Net fetches the entity embeddings from the key-value store, also takes the raw features from the target transactions, and then computes the transaction risk scores.”[Lu page 5 Network Architecture]; see RT Net (highlighted in orange) in Figure 2 – GRT 𝑇, obtained from TD graph, and entity embeddings, obtained from Batch Net output, are utilized to compute risk score for the target (i.e., particular) transaction (and thereby associated account) [Lu page 5])
However, Lu does not expressly teach generating a set of fuzzy attributes based on one or more attributes associated with each account in the plurality of accounts.
In the same field of endeavor, Cao teaches a means of applying machine learning techniques to enable detection of fraudulent transactions (“To tackle this problem, we introduce the TitAnt, a transaction fraud detection system deployed in Ant Financial, one of the largest Fintech companies in the world. The system is able to predict online real-time transaction fraud in mere milliseconds. We present the problem definition, feature extraction, detection methods, implementation and deployment of the system, as well as empirical effectiveness” [Cao Abstract]) that generat[es] a set of fuzzy attributes based on one or more attributes associated with each account in the plurality of accounts (“We set 100 trees for IF and raw basic features are fed as attributes. As rule-based ID3 and C5.0 cannot support continuous values well, we discretize the data into different bins [32]” [Cao page 7 Experimental Setup]).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have incorporated generating a set of fuzzy attributes based on one or more attributes associated with each account in the plurality of accounts as taught by Cao into Lu because they are both directed towards applying machine learning techniques to enable detection of fraudulent transactions. Feature discretization/binning is already a well-known, commonly utilized data pre-processing technique in the art for purposes of speeding up convergence and minimizing the influence of outliers on training (see Galli, “Data discretization in machine learning” [pages 1-3]). As such, a person of ordinary skill in the art would recognize the value of incorporating this technique to obtain known benefits and achieve a boost in predictive performance, as demonstrated in Cao for comparable learning procedures (“C5.0 has better than ID3 by 6.9% on average, probably because it takes better data discretization and segmentation mechanisms such as Gain Ratio. LR is implemented with discretization pre- processing which tremendously improves performance” [Cao page 8 Empirical Results on Transaction Fraud Detection]).
However, the combination of Lu and Cao does not expressly teach performing an action associated with the particular account based on the risk score.
In the same field of endeavor, Le teaches a means of applying machine learning techniques to enable detection of fraudulent transactions (“One innovative aspect of the subject matter described in this disclosure can be implemented as a method for selectively advancing funds based on risks posed by transactions associated with an online service. The method can be performed by one or more processors of a system coupled to or associated with the online service, and can include receiving a payment request for a transaction between a vendor and a consumer, and sending a first request to a database associated with the online service for historical transactions and personal attributes of the vendor concurrently with sending a second request to a number of third-party services for credit information and personal attributes of the consumer. The method includes receiving information responsive to the first and second requests from the database and the third-party services, respectively. The method includes obtaining a risk score for the transaction based on an application of one or more risk assessment rules to the received information by a machine learning model trained with at least the historical transactions and the personal attributes of the vendor” [Le ¶ 0005]) that perform[s] an action associated with the particular account based on the risk score (“The method includes determining whether or not to advance funds to the vendor for the transaction, prior to sending a payment request to a financial service designated by the consumer, based at least in part on the risk score” [Le ¶ 0005]; “In some implementations, the method may also include declining the transaction without sending the payment request to the designated financial service when the risk score is greater than a first level, and informing the designated financial service of the declined transaction when the risk score is greater than the first level by more than an amount or percentage. The method may also include sending the payment request to the designated financial service when the risk score is less than a second level, where the first level is greater than the second level, and advancing funds for the transaction to the vendor prior to or concurrently with sending the payment request to the designated financial service when the risk score is less than the second level by more than an amount or percentage“ [Le ¶ 0006]).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have incorporated performing an action associated with the particular account based on the risk score as taught by Le into the combination of Lu and Cao because they are all directed towards applying machine learning techniques to enable detection of fraudulent transactions. Given that Lu already discusses applicability of the disclosed architecture for real-time inference of transaction risk (“Similar to the two-“tower" model architecture [5, 24, 25, 27–29] (see Sec. 6), we design LNN to support two stages of processing (batch and real-time inference)… (Real-Time Inference Stage.) For real-time risk assessment, the second stage of LNN (namely RT Net) calculates the score by Eq. 4.” [Lu pages 6-7 Network Architecture]), a person of ordinary skill in the art would recognize the value of leveraging the calculated risk score to further determine appropriate action in real time, thereby handling associated transaction services efficiently (“By advancing funds to vendors only for transactions that pose less than a certain risk to the online service, implementations of the subject matter disclosed herein can minimize the number and amount of fraudulent chargeback transactions associated with the online service” [Le ¶ 0034]).
Regarding claim 16, the combination of Lu, Cao, and Le teaches the limitations of parent claim 15, and Le further teaches determining, using the machine learning model, an initial risk score associated with the particular account based on the embeddings, wherein the risk score is determined further based on the initial risk score (“When the risk score is between the first and second levels, indicating that the corresponding transaction has not been classified as acceptable or unacceptable, the risk assessment engine 118 can perform additional processing to determine the appropriate action to take. In some instances, the risk assessment engine 118 may send the payment request to the designated financial service 150 (without advancing funds to the vendor), thereby allowing the designated financial service 150 to determine whether to process the transaction. In other instances, the risk assessment engine 118 may update the structured data set to include additional input data (e.g., credit information of the vendor, additional personal attributes of the vendor, additional personal attributes of the consumer, etc.), and then generate another risk score based on the updated structured data set” [Le ¶ 0083]).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have incorporated determining, using a machine learning model, an initial risk score associated with the particular account based on the embeddings, wherein the risk score is determined further based on the initial risk score as taught by Le into the combination because Lu, Cao, and Le are all directed towards applying machine learning techniques to enable detection of fraudulent transactions. Given that feature discretization/binning is known to risk information loss due to inherent loss of precision (see Galli, “Data discretization in machine learning” [pages 1-3]), a person of ordinary skill in the art would understand the value of leveraging more granular forms of input data to update calculated risk score in instances where the appropriate associated action is not yet conclusively determined ([Le ¶ 0083]).
Regarding claim 17, the combination of Lu, Cao, and Le teaches the limitations of parent claim 16, and Lu further teaches wherein the determining the initial risk score associated with the particular account is further based on the corresponding one or more attributes associated with the particular account (“The first stage is Batch Net inference, after which we obtain the entity embeddings learned from historical transactions and flush them into a key-value store. In the second stage, the RT Net fetches the entity embeddings from the key-value store, also takes the raw features from the target transactions, and then computes the transaction risk scores.”[Lu page 5 Network Architecture]; see RT Net (highlighted in orange) in Figure 2 – GRT 𝑇, obtained from TD graph, and entity embeddings, obtained from Batch Net output, are utilized to compute risk score for the target transaction (and thereby associated account) [Lu page 5]).
Regarding claim 18, the combination of Lu, Cao, and Le teaches the limitations of parent claim 16, and Le further teaches wherein the determining the risk score comprises modifying the initial risk score based on characteristics, (“In other instances, the risk assessment engine 118 may update the structured data set to include additional input data (e.g., credit information of the vendor, additional personal attributes of the vendor, additional personal attributes of the consumer, etc.), and then generate another risk score based on the updated structured data set” [Le ¶ 0083]) wherein the characteristics are associated with at least one of the one or more communities taught in Lu (“To collect neighbor features to assess the risks of transaction fraud, multiple entities used in the checkout sessions are considered neighbors of 𝑡𝑥𝑛 nodes. These entities, including shipping address, E-mail, IP address, device ID, contact phone, payment token, and user account, are represented as 𝑒𝑛𝑡𝑖𝑡𝑦 nodes in G” [Lu page 3 Static Graph Construction]; ““(c) Two-Stage Directed Graph. A TD Graph is stored in target transaction partitions. A partition is made up of several timestamps that are the temporal snapshots in which all related transactions and entities are.” [Lu page 3 Time-sensitive Graph Construction]; The characteristics utilized in Le to modify initial risk score (e.g., additional personal attributes of the vendor, additional personal attributes of the consumer) correspond to value types of the entity nodes (e.g., shipping address, E-mail, IP address, etc.) that make up snapshots of partitions (i.e., communities) for each target transaction in Lu).
Regarding claim 19, the combination of Lu, Cao, and Le teaches the limitations of parent claim 18, and Lu further teaches wherein the characteristics associated with the at least one of the one or more communities of attributes represent a distribution of different types of transactions associated with the at least one of the one or more communities of attributes (“The historical transactions that share the same entities 𝑒𝑛𝑡𝑖𝑡𝑦 used in the target transaction, such as emails and shipping addresses, we call them reference transactions, which do not have any labels in our experiments” [Lu page 3 Time-sensitive Graph Construction]; All reference transaction nodes sharing common entities are stored with the target transactions in the same partition. On the TD graph, there is only one role for transactions, either a target transaction or a reference transaction. There are multiple snapshots 𝑠𝑛𝑎𝑝𝑠ℎ𝑜𝑡𝑖 in one target partition 𝑝𝑎𝑟𝑡𝑖𝑡𝑖𝑜𝑛𝑡 , 𝑖 ≤ 𝑡 , as illustrated in Fig. 2 (a)... In Fig. 2 (a), we show an illustrative example of the risk assessment of the transaction on 𝑡𝑥𝑛𝑡𝑔𝑡 𝑡 using TD Graph. The features of historical reference transactions 𝑡𝑥𝑛𝑟𝑒 𝑓 𝑡 are sent to 𝑒𝑛𝑡𝑖𝑡𝑦𝑡𝑔𝑡 𝑡 (phone and buyer in this example). They are the entity nodes 1-hop away from the target transaction. After bi-directed message-passing through the dotted edges, the embeddings (for phone and buyer in this example) of 1-hop neighbors are obtained. Then these embeddings are sent to the target transaction 𝑡𝑥𝑛𝑡𝑔𝑡 𝑡 through the directed edges (in red), which forms the graph used in real-time inference” [Lu pages 5-6 Two-Stage Directed Graph]; see snapshots t (highlighted in orange) and t-1, t-2, t-3 (highlighted in blue) in Figure 2(a) – messages for target transaction at snapshot t are aggregated from reference transactions at different snapshots t-1, t-2, t-3 (i.e., different types of transactions) into embeddings of 1-hop entity neighbors (i.e., community of attributes) [Lu page 5]). Cao further teaches utilization of fuzzy attributes over exact attributes ([Cao page 7 Experimental Setup]), as detailed in claim 15 above.
Regarding claim 20, the combination of Lu, Cao, and Le teaches the limitations of parent claim 15, and Le further teaches receiving a transaction request associated with the particular account, wherein the action comprises authorizing, requesting additional data, or denying the transaction request based on the risk score (“The method can be performed by one or more processors of a system coupled to or associated with the online service, and can include receiving a payment request for a transaction between a vendor and a consumer…In some implementations, the method may also include declining the transaction without sending the payment request to the designated financial service when the risk score is greater than a first level, and informing the designated financial service of the declined transaction when the risk score is greater than the first level by more than an amount or percentage. The method may also include sending the payment request to the designated financial service when the risk score is less than a second level, where the first level is greater than the second level, and advancing funds for the transaction to the vendor prior to or concurrently with sending the payment request to the designated financial service when the risk score is less than the second level by more than an amount or percentage.” [Le ¶ 0006]; “When the risk score is between the first and second levels, indicating that the corresponding transaction has not been classified as acceptable or unacceptable, the risk assessment engine 118 can perform additional processing to determine the appropriate action to take…the risk assessment engine 118 may update the structured data set to include additional input data” [Le ¶ 0083]).
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
Galli (“Data discretization in machine learning”, available online 4 Jul 2022) discloses a review of data discretization/binning as a commonly utilized pre-processing technique in machine learning.
Behera et al. (“Credit Card Fraud Detection: A Hybrid Approach Using Fuzzy Clustering & Neural Network”, available conference 2015) discloses an approach towards credit card fraud detection which first applies a fuzzy c-means clustering algorithm to find out the normal usage patterns of credit card users based on their past activity, calculate a transaction suspicion score according to the extent of deviation from the normal patterns, and thereby classify the transaction as legitimate or suspicious or fraudulent. Once a transaction is found to be suspicious, neural network based learning mechanism is applied to determine whether it was actually a fraudulent activity or an occasional deviation by a genuine user.
Masihullah et al. (“Identifying Fraud Rings Using Domain Aware Weighted Community Detection”, available conference 2022) discloses a means of fraud ring detection that leverages community detection algorithms to determine community profiles, in addition to node features, as input to a downstream discriminator.
Wang et al. (“Modeling Heterogeneous Graph Network on Fraud Detection: A Community-based Framework with Attention Mechanism”, available conference Nov 2021) discloses a means of GNN-based fraud detection that determines node representations based on a concatenation of its own embedding and its neighborhood embedding.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to VIJAY M BALAKRISHNAN whose telephone number is (571) 272-0455. The examiner can normally be reached 10am-5pm EST Mon-Thurs.
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, JENNIFER WELCH can be reached on (571) 272-7212. 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.
/V.M.B./
Examiner, Art Unit 2143
/KC CHEN/Primary Patent Examiner, Art Unit 2143
Prosecution Insights