DETAILED ACTION
This action is responsive to the Claims filed on 07/17/2025. Claims 1-20 are pending in the case. Claims 1, 8 and 14 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 Rejections - 35 U.S.C. § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. §§ 102 and 103 (or as subject to pre-AIA 35 U.S.C. §§ 102 and 103) is incorrect, any correction of the statutory basis for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of 35 U.S.C. § 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102 of this title, 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.
Claim(s) 1-5, 8-12, 14-18 and 20 are rejected under 35 U.S.C. § 103 as being unpatentable over of Xiang et al “Semi-supervised Credit Card Fraud Detection via Attribute-Driven Graph Representation” further in view of Tarvainen et al. “Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results”
Claim 1/8/14
Xiang teaches, A computer-implemented method for training an anomaly detection model, comprising, for each training iteration in a series of training iterations: [claim 8] A system comprising: at least one memory storing instructions; and at least one hardware processor interoperably coupled with the at least one memory, wherein execution of the instructions by the at least one hardware processor causes performance of operations comprising [claim 14] A non-transitory, computer-readable medium storing computer-readable instructions, that upon execution by at least one hardware processor, cause performance of operations, (pg 5 Experiments “The proposed gated temporal attention network model.1 We also evaluate two variants of our model, GTAN-A and GTAN-R… train the model with the Adam optimizer for 100 epochs with early stopping.” Pg 7 “Semi-supervised experiments demonstrate the excellent fraud detection performance of our model” as noted in the instant specification, fraud detection is an example of anomaly detection. The Adam optimizer in this context is computer implemented method for training a model. A computer necessarily having memory for storing instructions for performing the operations via a hardware processor) receiving unlabeled metrics data relating to a set of accounts; receiving, for a first set of unlabeled metrics data, a set of respective labels, wherein each label is associated with an account in the set of accounts; (pg 5 “To the best of our knowledge, we did not find any public semi-supervised credit card fraud detection dataset. Therefore, we collect the partially labeled records from our collaborated partners, namely Financial Fraud Semi supervised Dataset… The ground truth labels are obtained on cases reported by consumers and confirmed by financial domain experts. If a transaction is reported by a cardholder or identified by financial experts as fraudulent, we label it as 1; otherwise, it is labeled as 0” the records are partially labeled meaning that some metrics are unlabeled while others in the set are labeled, each label is associated with a cardholder, i.e an associated with an account) processing (1) the first set of unlabeled metrics data that is associated with the received set of labels and (pg 4 “we propose to take the manually annotated label as one of the categorical attributes of the transaction, and get the embedding of this categorical at tribute, which we call risk embedding. Specifically, we take the manually annotated label as the risk feature of each transaction, where the category of unlabeled data is ‘unlabeled’, and the category of the rest of the data is ‘fraud’ or ‘legitimate’. Then, we add this feature to the transaction data as one of our input categorical attributes” feature augmentation of the unlabeled metrics amounts to processing the first set of unlabeled metrics associated with labels as claimed.) processing…(2) a second set of unlabeled metrics data, to generate a set of network inputs, wherein each network input represents a respective account in the set of accounts; (pg 4 “Then, we propose to embed the partially observed risk attributes (i.e., labels) into the same space as the other node features… Then, we add the node features and risk embeddings together as input node features” the various features are added together to generate a set of network inputs, the respective inputs represent a respective account as they are representative of consumer account data.)
Xiang does not explicitly teach, processing a subset of network inputs that correspond to the second set of unlabeled metrics data using a trained teacher neural network that generates a respective initial anomaly prediction output for each network input; training a student neural network to optimize a loss function, wherein the student neural network processes the set of network inputs and generates a respective proposed anomaly prediction output for each network input, and wherein optimizing the loss function comprises, for each account in the set of accounts… minimizing a loss term that measures a difference between the initial anomaly prediction output of the teacher neural network and the proposed anomaly prediction output of the student neural network when the set of respective labels does not include a label associated with the account;
Tarvainen when combined with Xiang, addresses the use of student teacher network in semi-supervised learning teaches, processing a subset of network inputs that correspond to the second set of unlabeled metrics data using a trained teacher neural network that generates a respective initial anomaly prediction output for each network input (pg 5 “We used the rest of the primary training set together with the extra training set as unlabeled examples. We ran experiments with Mean Teacher and our baseline model, and used either 0, 100000 or 500000 extra examples. Table 3 shows the results” pg 6 “Mean Teacher uses unlabeled training data more efficiently than the Π model” while the network processes both the labeled data and unlabeled training data to generate predictive labels, as understood by examiner, when combined with Xiang such predictions are of fraud labels.) training a student neural network to optimize a loss function (figure 2 caption “The figure depicts a training batch with a single labeled example… Both the student and the teacher model evaluate the input applying noise (n, n’) within their computation. The softmax output of the student model is compared with the one-hot label using classification cost and with the teacher output using consistency cost… After the weights of the student model have been updated with gradient descent… A training step with an unlabeled example would be similar, except no classification cost would be applied.” The training of the student model is described according to two loss functions, PHOSTIA would understand that gradient descent is the process by which the loss function is optimized via parameter updates to the model, in this case the student model.) wherein the student neural network processes the set of network inputs and generates a respective proposed anomaly prediction output for each network input, (figure 2
PNG
media_image1.png
375
877
media_image1.png
Greyscale
as shown in the figure the student model processes input data to generate and output prediction, the input is processed through the depicted input layer (i.e network inputs). As stated when combined with Xiang such data is a label of fraud activity, which is an anomaly prediction) and wherein optimizing the loss function comprises, for each account in the set of accounts: (pg 4 “At each training step, all the EMA predictions of the examples in that minibatch are updated based on the new predictions” here each training step refers to the optimization of the loss function for each training example, which when applied to account information corresponds to each account in the set of accounts as claimed) minimizing a loss term that measures a difference between the initial anomaly prediction output of the teacher neural network and the proposed anomaly prediction output of the student neural network when the set of respective labels does not include a label associated with the account; (pg 3 Figure 2 caption “The figure depicts a training batch with a single labeled example…. The softmax output of the student model is compared with the one-hot label using classification cost and with the teacher output using consistency cost… After the weights of the student model have been updated with gradient descent,… A training step with an unlabeled example would be similar, except no classification cost would be applied.” Pg 3 “we define the consistency cost J as the expected distance between the prediction of the student model…and the prediction of the teacher model “ A typical labeled training step uses both a consistency and classification loss/cost, when the data does not include a label only the consistency loss is used. ) and minimizing a loss term that measures a difference between the proposed anomaly prediction output of the student neural network and the label associated with the account when the set of respective labels includes a label associated with the account. (pg 3 Figure 2 caption “The figure depicts a training batch with a single labeled example…. The softmax output of the student model is compared with the one-hot label using classification cost and with the teacher output using consistency cost… After the weights of the student model have been updated with gradient descent,…” classification cost is understood to be a description of the distance between a predicted label and the ground truth label, as known by one of ordinary skill in the art. This is also apparent by inspection of Figure 2 which shows the classification cost between the label and the student prediction output.)
Accordingly, it would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to modify the semi supervised fraud detecting system described by Xiang to comprise a student teacher network for classification as described by Tarvainen. One would have been motivated to make such a combination because both references describe systems for classification with semi supervised training data. Further as noted by Tarvainen “adding high-quality labels to training data manually is often expensive. Therefore, it is desirable to use regularization methods that exploit unlabeled data effectively to reduce over-fitting in semi-supervised learning” and “the more accurate target labels lead to a faster feedback loop between the student and the teacher models, resulting in better test accuracy. Second, the approach scales to large datasets and on-line learning.” (Tarvainen introduction pg 1-3)
Claim 2/9/15
Xiang/Tarvainen teaches claim 1/8/14
Further Tarvainen teaches, wherein the student neural network and the teacher neural network have a same architecture. (pg 1 introduction “In this case, the model assumes a dual role as a teacher and a student” pg 3 Figure 2
PNG
media_image2.png
320
878
media_image2.png
Greyscale
the art depicts the models as having the same architecture, noting all methods use the same architecture. pg 4 Table 1 caption “All the methods use a similar 13-layerConvNetarchitecture”)
Claim 3/10/16
Xiang/Tarvainen teaches claim 2/9/15
Further Tarvainen teaches, wherein the trained teacher neural network in a particular training iteration is the student neural network from a training iteration prior to the particular training iteration. (pg 1 introduction “In this case, the model assumes a dual role as a teacher and a student. As a student, it learns as before; as a teacher, it generates targets, which are then used by itself as a student for learning” pg 2 Section2 “To overcome the limitations of Temporal Ensembling, we propose averaging model weights instead of predictions. Since the teacher model is an average of consecutive student models, we call this the Mean Teacher method” the teacher is simply the average representations of prior student models.)
Claim 4/11/17
Xiang/Tarvainen teaches claim 2/9/15
Further Tarvainen teaches, evaluating the accuracy of proposed anomaly detection outputs from the student neural network from a previous training iteration to generate a top subset of proposed anomaly detection outputs (pg 6 Figure 4 “Figure 4: Validation error on 250-label SVHN over four runs per hyperparameter setting and their means.” The percent validation error on 250 labels is a representation of the top subset of predictions as training progresses which includes prior iterations. For example, a 15% validation error is an indication of the proportion of correct or top classification from the validation set.)
Xiang teaches, and actively querying human experts to label the top subset of proposed anomaly detection outputs. (pg 4 “we propose to take the manually annotated label as one of the categorical attributes… Specifically, we take the manually annotated label…” pg 7 conclusion “Semi-supervised experiments demonstrate the excellent fraud detection performance of our model with only a tiny proportion of manually annotated data” a portion of the label data is labeled manually, which in the context of partially labeled training data refers to human expert labels, experts with respect to the labeling accuracy. Labeling a portion of data amounts to labeling the top subset as claimed.)
Claim 5/12/18
Xiang/Tarvainen teaches claim 1/8/14
Tarvainen teaches, wherein the loss term that measures a difference between the initial anomaly prediction output of the teacher neural network and the proposed anomaly prediction output of the student neural network is a mean squared error. (pg 3 “More formally, we define the consistency cost J as the expected distance between the prediction of the student model and the prediction of the teacher model” and “pg 7 “we use mean squared error (MSE) as our consistency cost function”)
Claim 20
Xiang/Tarvainen teaches claim 1/8/14
Further Xiang teaches, wherein the anomaly detection model uses machine learning techniques (abstract pg 1 “we pass messages among the nodes through a Gated Temporal Attention Network (GTAN) to learn the transaction representation. We further model the fraud patterns through risk propagation among transactions…. The result shows that our proposed method, namely GTAN, outperforms other state-of-the-art baselines on three fraud detection datasets.” )
Claim(s) 6, 13, 19 are rejected under 35 U.S.C. § 103 as being unpatentable over of Xiang/Tarvainen, further in view of Laine et al. “temporal ensembling for semi-supervised learning”
Claim 6/13/19
Xiang/Tarvainen teaches claim 1/8/14
Tarvainen teaches, wherein the loss term that measures a difference between the proposed anomaly prediction output of the student neural network and the label associated with the account (as previously noted in the rejection of claim 1)
Xiang/Tarvainen does not explicitly teach, is a cross-entropy loss.
Laine, however, when addressing the use of classification loss for labeled data in a semi-supervised setting teaches, is a cross-entropy loss. (pg 2 figure 1 caption “Structure of the training pass in our methods… Labels yi are available only for the labeled inputs, and the associated cross-entropy loss component is evaluated only for those” pg 2 Section 2.1 “The first component is the standard cross entropy loss, evaluated for labeled inputs only” here the label loss is calculated using cross entropy, which is a standard loss function in machine learning)
Accordingly, it would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to modify the semi supervised fraud detecting system described by Xiang/Tarvainen to comprise a cross entropy classification loss as described by Laine. One would have been motivated to make such a combination because both references describe systems for classification with semi supervised training data. Further, Tarvainen describes it’s system as being the improved version of the system described by Laine (i.e temporal ensembling) noting “Without changing the network architecture, Mean Teacher achieves an error rate of 4.35% on SVHN with 250 labels, outperforming Temporal Ensembling trained with 1000 labels” (Tarvainen abstract). Both Tarvainen and Laine describe a classification cost, only Laine explicitly notes the loss is a cross entropy loss.
Claim(s) 7 is rejected under 35 U.S.C. § 103 as being unpatentable over of Xiang/Tarvainen, further in view of Shekhar “Incorporating Privileged Information to Unsupervised Anomaly Detection”
Claim 7
Xiang/Tarvainen teaches claim 1/8/14
Xiang/Tarvainen does not explicitly teach, wherein the teacher neural network implements an Isolation Forest anomaly detection algorithm.
Shekhar when addressing a teacher network for unlabeled example knowledge distillation teaches, wherein the teacher neural network implements an Isolation Forest anomaly detection algorithm. (pg 1 “Simply put, privileged information (PI) is additional data/knowledge/information that is available only at the learning/model building phase for (subset of) training examples, which however is unavailable for (future) test examples.” Pg 2 “The setup involves an Intelligent (or non-trivial) Teacher at learning phase, who provides the Student with privileged information… The only (perhaps straightforward) extension of LUPI to unsupervised anomaly detection… We note that the top performer in is the Isolation Forest (iForest) algorithm” pg 8 “Next, we show how to generalize these ideas to anomaly detection with no label supervision. Figure1 shows an overview. To this end, we utilize a state-of-the-art ensemble technique for anomaly detection, called Isolation Forest”)
Accordingly, it would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to modify the semi supervised fraud detecting system described by Xiang/Tarvainen to comprise a teacher which employs isolation forest method as described by Shekhar. One would have been motivated to make such a combination because as noted by Shekhar, Isolation forest is a top performer for anomaly detection when no labels are provided for supervision, which is the case in the learning phase described by Tarvainen of the student network. Further, as noted by Shekhar “Our SPI outperforms competition in detection performance in most of the datasets” (pg 13 Shekhar)
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to JOHNATHAN R GERMICK whose telephone number is (571)272-8363. The examiner can normally be reached M-F 9:30-4:30.
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, Kakali Chaki can be reached on 571-272-3719. 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.
/J.R.G./
Examiner, Art Unit 2122
/KAKALI CHAKI/ Supervisory Patent Examiner, Art Unit 2122
Prosecution Insights