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 .
Continued Examination Under 37 CFR 1.114
A request for continued examination under 37 CFR 1.114, including the fee set forth in 37 CFR 1.17(e), was filed in this application after final rejection. Since this application is eligible for continued examination under 37 CFR 1.114, and the fee set forth in 37 CFR 1.17(e) has been timely paid, the finality of the previous Office action has been withdrawn pursuant to 37 CFR 1.114. Applicant's submission filed on 12 May 2026 has been entered.
Claim Rejections - 35 USC § 103
The text of those sections of Title 35, U.S. Code not included in this action can be found in a prior Office action.
Claim(s) 1-7, 13-14, and 16-20 is/are rejected under 35 U.S.C. 103 as being unpatentable over Shekhar et al. (US Pub. No. 2021/0233080, published July 2021, hereinafter “Shekhar”) in view of Sun et al. (US Pub. No. 2023/0351215, filed Sept. 2021, hereinafter “Sun”), and further in view of Wadhwa et al. (US Pub. No. 2022/0020026, published Jan. 2022, hereinafter “Wadhwa”).
Regarding claim 1, Shekhar teaches a method, comprising:
receiving entity data for a plurality of entities (Shekhar, [0030] – “the term “identity attribute” refers to data associated with a digital identity. In particular, an identity attribute can refer to a characteristic of a digital identity, an activity performed in association with the digital identity, and/or details related to the activity” – teaches receiving entity data (identity attributes) for a plurality of entities);
receiving transaction data for transactions between corresponding entities included in the plurality of entities (Shekhar, [0016] – “the fraudulent transaction detection system identifies a plurality of digital identities corresponding to a plurality of digital transactions.” – teaches receiving transaction data between corresponding entities included in the plurality of entities (plurality of digital identities corresponding to digital transactions));
performing a self-supervised training of a graph neural network including by sampling the heterogeneous graph representation for positive samples and negative samples to learn embedding representations for the nodes of the heterogeneous graph representation (Shekhar, [0016] – “ the time-dependent graph convolutional neural network can generate, based on the edge connections of the transaction graph, node embeddings corresponding to the plurality of nodes utilizing a plurality of temporal random walks” – teaches self-supervised training of a graph neural network by learning embedding representations for the nodes of the heterogenous graph representation, and in [0104] – “The researchers followed a similar strategy for training and testing each model. For example, the researchers utilized 75% of the edges from the datasets to learn the embeddings. Further, the researchers used 25% of the edges as positive link examples and sampled an equal number of negative links. For the edge representation, the researchers used the Hadamard distance between the connecting nodes.” – teaches using positive and negative samples of the embeddings to train the model); and
utilizing the learned embedding representations for the nodes of the heterogeneous graph representation for automatic transaction analysis (Shekhar, [0016] – “The fraudulent transaction detection system can determine that a digital identity from the pair of digital identities corresponds to a fraudulent transaction based on the similarity probability” – teaches utilizing the learned embedding representations for the nodes of the heterogenous graph representation for automatic transaction analysis (determines fraudulent transactions using a similarity probability based on the embeddings)).
Shekhar fails to explicitly teach generating a heterogeneous graph representation, wherein nodes of the heterogeneous graph representation include a first type of node representing an entity of the plurality of entities and a second type of node representing a discrete financial transfer event of the transactions, wherein directed edges connect nodes of the first type representing entities to nodes of the second type representing the discrete financial transfer events; and wherein learning the embedding representations comprises severing a selected directed edge corresponding to a positive sample prior to generating the embedding representations for a source node and a destination node associated with the selected directed edge.
However, analogous to the field of the claimed invention, Sun teaches:
generating a heterogeneous graph representation, wherein nodes of the heterogeneous graph representation include a first type of node representing an entity of the plurality of entities and a second type of node representing a discrete financial transfer event of the transactions, wherein directed edges connect nodes of the first type representing entities to nodes of the second type representing the discrete financial transfer events (Sun, [0025] – “A “topological graph” can include a representation of a graph in a plane of distinct vertices connected by edges. The distinct vertices in a topological graph may be referred to as “nodes.” Each node may represent specific information for an event or may represent specific information for a profile of an entity or object… For example, a topological graph may represent a transaction network in which a node representing a transaction may be connected by edges to one or more nodes that are related to the transaction, such as nodes representing information of a device, a user, a transaction type, etc. An edge may be associated with a numerical value, referred to as a “weight,” that may be assigned to the pairwise connection between the two nodes. The edge weight may be identified as a strength of connectivity between two nodes and/or may be related to a cost or distance, as it often represents a quantity that is required to move from one node to the next.” – teaches generating a heterogenous graph representation (topological graph), wherein nodes of the heterogenous graph representation include a first type of node representing an entity of the plurality of entities (nodes related to a transaction, such as nodes representing information of a device, a user, etc.) and a second type of node representing a discrete financial transfer event of the transactions (node representing a transaction), wherein directed edges connect nodes of the first type representing entities to nodes of the second type representing the discrete financial transfer events (edge weights represent quantity required to move from one node to the next, thus teaching directed edges connecting nodes of the first type to nodes of the second type));
wherein learning the [[an]] embedding representations comprises severing a selected directed edge corresponding to a positive sample prior to generating the embedding representations for a source node and a destination node associated with the selected directed edge (Sun, [0088] – “Unlike some previous works that assume dynamic graphs only grow over time, embodiments of the disclosure can allow for both addition and deletion of links (e.g., edges). Embodiments can allow an analysis computer to learn latent representations evt.elementOf.Rd for each node v e V at time-steps t=1,2,…,T, such that the representations evt both preserves the local structure around a node v and models the local structure evolution over time. The latent representations evt can be final node representations” – teaches wherein learning an embedding representation (latent representations, which are embeddings as stated in Sun at [0039]) comprises severing a selected directed edge corresponding to a positive sample (allows for deletion of links, thus teaching severing a selected directed edge corresponding to a positive sample; Sun further teaches positive and negative samples at [0187]) prior to generating the embedding representations for a source node and a destination node associated with the selected directed edge (allows deletion of links prior to generation of latent representations, such that representations preserve local structure around each node v in V and models structure evolution over time))
determining, for at least one embedding representation including a node of the first type and a node of the second type, a likelihood of an edge existing between the node of the first type and the node of the second type (Sun, [0038] – “Embodiments of the disclosure, in some cases also referred to as Dynamic Graph Light Convolution Network (DGLC), can operate on dynamic graphs and learn node representations that capture both structural features and temporal evolution patterns.” and in [0204] – “In this section, the effectiveness of DGLC is evaluated for link prediction task on six real-world datasets comparing with five state-of-the-art baselines.” – teaches determining, for at least one embedding representation (learns node latent representations, latent representations are embeddings as stated in Sun at [0039]) including a node of the first type and a node of the second type (topological graph including nodes of the first type and nodes of the second type, as stated in Sun [0025]), a likelihood of an edge existing between the node of the first type and the node of the second type (evaluates DGLC for link prediction tasks, thus teaching determining a likelihood of an edge existing between the node of the first type and the node of the second type));
Therefore, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to incorporate the different nodes and learning of embedding representations of Sun to the transaction data, graphs, self-supervised training, and utilization of learned embeddings for automatic transaction analysis of Shekhar in order to create a method for receiving transaction data, creating a heterogenous graph based on the transaction data, learning embeddings based on the nodes of the heterogenous graph, and utilizing the embeddings to perform automatic transaction analysis. Doing so would learn latent representations of nodes in graphs for applications such as link prediction (Sun, [0002]) and learning such representations can benefit a plethora of graph analytical tasks (Sun, [0039]).
The combination of Shekhar and Sun fails to explicitly teach determining an anomaly score for at least one transaction of the transactions based at least in part on the determined likelihood of the edge existing between the node of the first type and the node of the second type, and flagging the at least one transaction of the transactions in response to a determination that the determined anomaly score exceeds a threshold.
However, analogous to the field of the claimed invention, Wadhwa teaches:
determining, for at least one embedding representation including a node of the first type and a node of the second type, a likelihood of an edge existing between the node of the first type and the node of the second type (Wadhwa, [0030] – “In one embodiment, the server system is configured to apply machine learning algorithms over the graph embedding vector for training a data model to facilitate prediction of missing links in the temporal knowledge graph. The missing links may be related to money laundering financial transactions.” and in [0058] – “In one example scenario, a party ‘X’ transfers $1000 to a party ‘Y’ who is a nephew of the party ‘X’. In the above example scenario, the temporal knowledge graph has two nodes depicting the party ‘X’ (i.e., source node) and the party ‘Y’ (i.e., destination node)” – teaches determining, for at least one embedding representation including a node of the first type and a node of the second type (applies ML algorithms over graph embedding vector to facilitate prediction, or likelihood estimation, of missing links, or edges, between nodes. Teaches nodes of two types: source node and destination node));
determining an anomaly score for at least one transaction of the transactions based at least in part on the determined likelihood of the edge existing between the node of the first type and the node of the second type (Wadhwa, [0071] – “The prediction engine 220 is configured to determine time-based probabilities associated with the flagged cluster. The time-based probabilities may include, but not limited to, a time-based probability of next edge formation within the flagged cluster, a time-based probability of next edge formation outside the flagged cluster with a nearby cluster. In one embodiment, the time-based probability of the next edge formation within the flagged cluster is determined by constructing a Long Short Term Memory (LSTM) network for the flagged cluster using the trained data model. In one embodiment, the time-based probability of next edge formation outside the flagged cluster with the nearby cluster is determined by generating a convolution network. These time-based probabilities are used to detect nodes/groups/transactions that might lead to the money laundering financial transaction.” and in [0074] – “In one embodiment, the processor 202 is configured to update fraud score of the flagged cluster and the particular node based on the time-based probabilities.” – teaches determining an anomaly score (time-based probability, node fraud score, and cluster fraud score are all anomaly scores; wherein the fraud scores are based at least in part on the time-based probability score) for at least one transaction of the transactions based at least in part on the determined likelihood of the edge existing between the node of the first type and the node of the second type (fraud scores updated based at least in part on time-based probabilities which are the likelihood estimations of an edge formation between the node of the first type (source node) and the node of the second type (destination node))); and
flagging the at least one transaction of the transactions in response to a determination that the determined anomaly score exceeds a threshold (Wadhwa, [0072] – “In one embodiment, if the time-based probability of the next edge formation leading to a source node is greater than a predetermined threshold value, the prediction engine 220 identifies an issuer associated with a particular node (i.e., a trailing node) related to the next edge (i.e., link) which may be linked in future money-laundering activities. The source node refers to a node from where all the financial transactions were initiated previously.” and in [0075] – “Additionally, the processor 202 is configured to generate a suspicious activity report (SAR) file and alert the identified issuer 102 for preventing fraudulent financial transactions based on the SAR file. The SAR file may include, but not limited to, a cluster fraud score, a node fraud score, and a prediction probability associated with the next financial transaction being the money laundering financial transaction.” – teaches flagging the at least one transaction of the transactions in response to a determination that the determined anomaly score exceeds a threshold (when time-based probability exceeds threshold value fraud scores are updated and an alert file is generated to flag the transaction, cluster, and node)).
Therefore, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to incorporate the link prediction and anomaly scores of Wadhwa to the first and second type nodes and automatic transaction analysis of Shekhar and Sun in order to generate anomaly scores based on link predictions between the first and second type nodes. Doing so would provide an automated system for predicting next financial transactions of suspicious customers in near real-time which can be used to take pre-emptive action and help in enriching the SAR file for AML systems (Wadhwa, [0034]).
Claims 19 and 20 incorporate substantively all the limitations of claim 1 in a system and a non-transitory computer readable medium, and are rejected on similar grounds as above. Shekhar teaches the processors and memory of these claims at [0138] – “As shown in FIG. 10, the computing device 1000 can include one or more processor(s) 1002, memory 1004, a storage device 1006, input/output interfaces 1008 (or “I/O interfaces 1008”), and a communication interface 1010, which may be communicatively coupled by way of a communication infrastructure (e.g., bus 1012).”.
Regarding claim 2, the combination of Shekhar, Sun, and Wadhwa teach the method of claim 1, further comprising using at least a portion of the embedding representations to cluster at least a subset of the transactions into a plurality of different cluster groups (Shekhar, [0032] – “The server classifies the plurality of nodes in the graph into a set of clusters based on a degree of similarity among the plurality of nodes. In other words, those nodes that are similar to each other are classified in the same cluster” – teaches using at least a portion of the embedding representations to cluster at least a subset of the transactions into a plurality of different cluster groups).
Regarding claim 3, the combination of Shekhar, Sun, and Wadhwa teach the method of claim 1, wherein performing the self-supervised training of the graph neural network includes using at least one of: an edge prediction task, a transaction similarity task, or a subgraph similarity task (Shekhar, [0092] – “the fraudulent transaction detection system 106 trains a time-dependent graph convolutional neural network to generate similarity probabilities for pairs of digital identities” – teaches wherein performing the self-supervised training of the graph neural network includes using at least one of a transaction similarity or subgraph similarity task (trains a neural network to generate similarity probabilities for pairs of digital identities)).
Regarding claim 4, the combination of Shekhar, Sun, and Wadhwa teach the method of claim 1, further comprising: predicting an anomaly based at least on the embedding representations (Shekhar, [0047] – “The fraudulent transaction detection system 106 can further utilize the time-dependent graph convolutional neural network to generate a similarity probability (e.g., a similarity score) for a pair of digital identities based on the node embeddings. Via the server(s) 102, the fraudulent transaction detection system 106 can determine that a digital identity from the pair of digital identities corresponds to a fraudulent transaction based on the similarity probability” – teaches predicting an anomaly based at least on the embedding representations (teaches predicting an anomaly based on a similarity score based on the node embeddings)).
Regarding claim 5, the combination of Shekhar, Sun, and Wadhwa teach the method of claim 4, further comprising predicting the anomaly using a multilayer perceptron (MLP) (Shekhar, [0015] – “the graph convolutional neural network can predict whether two digital identities correspond to the same user and further determine whether one of the digital identities is associated with a fraudulent transaction” – teaches predicting the anomaly using a multilayer perceptron, as supported in [0095] – “Algorithm 2, presented below, provides a characterization of how the fraudulent transaction detection system 106 executes training of the time-dependent graph convolutional neural network in accordance with one or more embodiments” – teaches an algorithm for training the time-dependent graph convolutional neural network, where the algorithm includes feed-forward layers and backpropagation indicating the time-dependent graph convolutional neural network is a differentiable model predicting the anomaly).
Regarding claim 6, the combination of Shekhar, Sun, and Wadhwa teach the method of claim 4,
wherein predicting the anomaly based at least on the embedding representations is based at least on a sigmoid of a weighted Hadamard product of a first embedding representation and a second embedding representation (Sun, [0148] – “In the above equation, W.sub.a.sup.l,W.sub.b.sup.l∈R.sup.(D.sup.l-1 x D.sup.l-1), a,b∈R.sup.D.sup.l-1 are the learnable parameters, σ.sub.glu is the sigmoid function, and .Math. is the hadamard product” – teaches a sigmoid of a weighted Hadamard product of a first embedding representation and a second embedding representation (representations a and b), and in [0192] – “For example, the analysis computer can compute a Hadamard product using two vectors: the first final vector representation of the first node and a second final vector representation of the second node (e.g., vectors corresponding to the latest snapshot). The Hadamard product can be used as a vector representing a potential link between the two node” and [0193] – “the analysis computer can send an advisory notice that a transaction is likely to take place, or that a current transaction being attempted was not likely to take place and may therefore be fraudulent” – teaches predicting the anomaly based at least on the embedding representation based at least on a sigmoid of a weighted Hadamard product of the a first and second embedding representation).
Therefore, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to incorporate the predicting of the anomaly based at least in part on a sigmoid of a weighted Hadamard product of the first and second embeddings of Sun to further modify the method of Shekhar, Sun, and Wadhwa. Doing so would capture the temporal evolution of dynamic graphs and provide for a more efficient method to learn temporal features (Sun, [0113]).
Regarding claim 7, the combination of Shekhar, Sun, and Wadhwa teach the method of claim 1, wherein the graph neural network is applied in one or more discrete, fixed snapshots containing transactions in time intervals (Shekhar, [0015] – “One or more embodiments described herein include a fraudulent transaction detection system that accurately determines whether a digital identity is associated with a fraudulent transaction utilizing time-based node embeddings generated by a graph convolutional neural network.” – teaches the graph neural network applied in one or more discrete, fixed snapshots (time-based node embeddings generated by a graph convolutional neural network), and in [0018] – “the fraudulent transaction detection system 106 associates a transaction timestamp for a digital transaction with the corresponding edge connection” – teaches snapshots containing transactions in time intervals (timestamp for digital transaction)).
Regarding claim 13, the combination of Shekhar, Sun, and Wadhwa teach the method of claim 1,
wherein an attention coefficient between a first node and a second node defines a weight of a corresponding interaction between the first node and the second node (Sun, [0025] – “An edge may be associated with a numerical value, referred to as a “weight,” that may be assigned to the pairwise connection between the two nodes. The edge weight may be identified as a strength of connectivity between two nodes and/or may be related to a cost or distance, as it often represents a quantity that is required to move from one node to the next.” and in [0044] – “In contrast, embodiments of the disclosure provide a novel neural architecture to efficiently learn node representations on dynamic graphs. Specifically, embodiments can employ self-attention for structural neighborhoods and temporal dynamics. Embodiments can employ a graph structural learning layer to effectively capture structural neighborhood information, and then employ a temporal convolution layer to efficiently capture the temporal evolution of graph sequences. For example, embodiments can allow for an analysis computer to generate a node representation by considering the node's neighbors following a self-attentional strategy and then node's historical representations following a temporal convolution strategy. Unlike static graph embedding methods that focus entirely on preserving structural proximity, embodiments can learn dynamic node representations that reflect the temporal evolution of graph structure over the varying number of historical snapshots. Embodiments can be capable of accurately capturing both structural properties and temporal evolution patterns. In contrast to temporal smoothness-based methods, embodiments can learn attention weights that capture temporal dependencies at a fine-grained node-level granularity.” – teaches wherein an attention coefficient between a first node and a second node defines a weight of a corresponding interaction between the first node and the second node (edges between nodes are associated with a weight, employs self-attention to efficiently learn node representations, embodiments can learn attention weights that capture temporal dependencies and generate node representations by considering node’s neighbors following a self-attentional strategy, thus teaching an attention coefficient between a first node and a second node that defines a weight of a corresponding interaction between the first node and the second node)).
Therefore, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to incorporate the attention coefficient defining a weight of a corresponding interaction between the first and second nodes of Sun to further modify the method of Shekhar, Sun, and Wadhwa to define a weight of a corresponding interaction between the first and second nodes. Doing so would efficiently learn node representations on dynamic graphs using attention mechanisms that consider nodes and their neighbors (Sun, [0044]).
Regarding claim 14, the combination of Shekhar, Sun, and Wadhwa teach the method of claim 1, wherein the sampling of the heterogenous graph representation includes uniform negative sampling (Shekhar, [0093] – “Accordingly, in one or more embodiments, the fraudulent transaction detection system 106 defines the loss function for a pair of node representations (z.sub.u, z.sub.ν) as follows where P.sub.n(u) is a negative sampling distribution for the out of context nodes for the node U:” – teaches sampling of the heterogenous graph representation including negative sampling).
Regarding claim 16, the combination of Shekhar, Sun, and Wadhwa teach the method of claim 1, wherein the embedding representations are based at least on a first layer of the graph neural network such that behavior divergence measures reflect a source entity's transactions (Shekhar, [0070] – “the node embeddings that make up the input to the first neural network layer 416 include the identity attributes of the digital identities corresponding to the nodes 408a-408c, 408e-408f. Thus, the time-dependent graph convolutional neural network 404 can generate the node embedding 414a for the node 408a based on the identity attributes of the plurality of digital identities represented in the transaction graph 402." – teaches wherein the embedding representations are based at least on a first layer of the graph neural network (node embedding 414a generated for node 408a at the first layer) such that behavior divergence measures reflect a source entity’s transactions (based on identity attributes of the plurality of digital identities in transaction graph, thus reflecting source entity’s transactions)).
Regarding claim 17, the combination of Shekhar, Sun, and Wadhwa teach the method of claim 1, wherein the embedding representations are based at least on a second layer of the graph neural network such that behavior divergence measures reflect counterparts interacted with (Shekhar, [0020] – “the time-dependent graph convolutional neural network generates the prior neural network layer node embeddings based on the identity attributes of a set of nodes associated with the neighbor nodes” – teaches generating embedding representation based at least on a second layer (generates prior neural network layer node embeddings) of the graph neural network such that behavior divergence measures reflect counterparts interacted with (based on identity attributes of a set of nodes associated with neighbor nodes)).
Regarding claim 18, the combination of Shekhar, Sun, and Wadhwa teach the method of claim 1, wherein the embedding representations are based at least on a deepest layer of the graph neural network such that a counterpart's transactions affect a representation of a source entity (Shekhar, [0096] – “As shown in Algorithm 2, the fraudulent transaction detection system 106 computes the temporal neighborhood of each node. The fraudulent transaction detection system 106 applies the K-layered convolution to generate the representation of the nodes. The fraudulent transaction detection system 106 passes the output of the final convolution layer through a fully connected layer to generate the final node embeddings z.sub.u,∀u∈V. Accordingly, the fraudulent transaction detection system 106 learns a set of parameters for the time-dependent graph convolutional neural network (Q.sup.(k), b.sub.q.sup.(k), W.sup.(k), b.sub.w.sup.(k)) along with the weights of the final fully connected layer.” – teaches wherein the embedding representations are based at least on a deepest later of the graph neural network (passes output of final layer to generate final node embeddings) and in [0020] – “the time-dependent graph convolutional neural network generates the prior neural network layer node embeddings based on the identity attributes of a set of nodes associated with the neighbor nodes.” – teaches such that a counterpart’s (neighbor node) transactions affect a representation (layer node embedding based on attributes on neighboring nodes) of a source entity).
Claim(s) 8 and 12 is/are rejected under 35 U.S.C. 103 as being unpatentable over Shekhar, Sun, and Wadhwa as applied to claims 1, 19, and 20 above, and further in view of Verma et al. (US Pub. No. 2023/0289610, effective filing date of March 2022 (all citations supported by foreign priority application IN 202241012724), hereinafter “Verma”).
Regarding claim 8, the combination of Shekhar, Sun, and Wadhwa teach the method of claim 7, wherein
embedding representations produced by a graph neural network (GNN) on each snapshot are sent as input to a recurrent neural network (RNN) (Shekhar, [0097] – “Upon generating the node embeddings, the fraudulent transaction detection system 106 can determine whether a digital identity corresponds to a fraudulent entity using a machine learning technique (e.g., a machine learning model). In particular, the fraudulent transaction detection system 106 can provide the node embeddings generated by the time-dependent graph convolutional neural network 404 to a trained machine learning model.” – teaches sending embedding representations produced by graph neural network on each snapshot to a machine learning model )
The combination of Shekhar, Sun, and Wadhwa fails to explicitly teach wherein a sliding window is applied such that each snapshot is offset from a previous snapshot by a time interval of the sliding window and a recurrent neural network (RNN) that combines the embedding representations on each snapshot s with a per-customer hidden state, maintained across snapshots.
However, analogous to the field of the claimed invention, Verma teaches:
wherein a sliding window is applied such that each snapshot is offset from a previous snapshot by a time interval of the sliding window (Verma, [0128] – “the TPP model 504 may utilize the actual time (real-time) of the next event for predicting the corresponding marker. In one embodiment, the TPP model 504 demonstrates high applicability in scenarios where the marker (e.g., fraudulent, or non-fraudulent transaction) is computed or derived after some time as opposed to being simultaneously available with the event occurrence. As shown in FIG. 5A, there is a sequence of events denoted by their time of occurrence and corresponding markers. Mathematically, each sequence is represented by S={(t.sub.1, y.sub.1), (t.sub.2, y.sub.2), . . . , (t.sub.n, y.sub.n)}, where n refers to the total sequence length. Here, (t.sub.j, y.sub.j) refers to the j.sup.th event represented by the time of the event (t.sub.j) and the corresponding marker (y.sub.j). By default, the events are ordered in time, such that t.sub.j+1≥t.sub.j. Given the sequence of last n events, the task should predict the next event time t.sub.n+1 and the corresponding marker y.sub.n+1.” – teaches a sliding window applied such that each snapshot is offset from the previous snapshot by a time interval of the sliding window (payment transactions performed within a particular time interval))
a recurrent neural network (RNN) that combines the embedding representations on each snapshot s with a per-customer hidden state, maintained across snapshots (Verma, [0129] – “the TPP model 504 utilizes a recurrent neural network (RNN) as backbone architecture to learn the first embedding (i.e., the temporal embedding 506) based on analysis of past events. In general, RNN is a feed-forward neural network structure. In RNN, additional edges (also known as recurrent edges) are added such that the outputs from the hidden units at the current time step are fed into them again as future inputs at the next time step. In consequence, the same feed-forward neural network structure is replicated at each time step, and the recurrent edges connect the hidden units of the network replicated at adjacent time steps across time.” – teaches combining representations on each snapshot with a per-customer hidden state (recurrent edges connect hidden units replicated at adjacent time steps across time)).
Therefore, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to incorporate the sliding window and recurrent neural network of Verma to modify the method of Shekhar, Sun, and Wadhwa in order to provide embedding representations produced by a GNN on each snapshot to a RNN that combines representations on each snapshot with a per-customer hidden state, such that each snapshot is offset by a time interval. Doing so would capture the temporal embedding of the real-time sequence of the transaction and create an internal state of the network to memorize influence of past data samples (Verma, [0129]).
Regarding claim 12, the combination of Shekhar, Sun, and Wadhwa teach the method of claim 1.
The combination of Shekhar, Sun, and Wadhwa fails to explicitly teach wherein the first type of node has a respective set of learnable parameters and the second type of node has a respective set of learnable parameters different from the set of learnable parameters associated with the first type of node.
However, analogous to the field of the claimed invention, Verma teaches:
wherein the first type of node has a respective set of learnable parameters and the second type of node has a respective set of learnable parameters different from the set of learnable parameters associated with the first type of node (Verma, [0121] – “The processor 206 is configured to calculate the second loss value that enables learning representations for different node types (i.e., the first nodes and the second nodes) independently and simultaneously” – teaches learning representations for different node types independently and simultaneously, and in [0122] – “all the model parameters are learned via stochastic gradient descent with the Adam optimizer” – teaches wherein the different nodes have respective sets of learnable parameters (all parameters are learned via stochastic gradient descent, and the different node types are learned independently, thus their sets of learnable parameters are different from each other)).
Therefore, it would have been obvious to a person to incorporate the first and second types of nodes having different sets of learnable parameters of Verma to modify the first and second type nodes of Shekhar, Sun, and Wadhwa in order to have nodes with different sets of learnable parameters. Doing so would preserve the structure of the bipartite graph and provide high mutual-information with self-node features (Verma, [0122]).
Claim(s) 9-10 is/are rejected under 35 U.S.C. 103 as being unpatentable over Shekhar, Sun, and Wadhwa as applied to claims 1, 19, and 20 above, and further in view of Arora et al. (US Pub. No. 2022/0101327, published March 2022, hereinafter “Arora”).
Regarding claim 9, the combination of Shekhar, Sun, and Wadhwa teach the method of claim 1.
The combination of Shekhar, Sun, and Wadhwa fail to explicitly teach: computing representations by repeatedly sending messages along edges of a local neighborhood of a node; aggregating the messages; and combining the aggregated message with information associated with a source node.
However, analogous to the field of the claimed invention, Arora teaches: wherein performing the self-supervised training of the graph neural network includes performing message passing including by:
computing representations by repeatedly sending messages along edges of a local neighborhood of a node (Arora, [0147] – “The historical transaction data includes information (e.g., various transaction messages) pertaining to the historical transactions executed between the merchants M1-Mn and the consumers C1-Cn. At step 1004, the fraud detection server 110 generates the first consumer-merchant graph 200 based on the historical transaction data 818” – teaches computing representations (generated graph) by repeatedly sending message along edges of a local neighborhood of a node (historical data comprises messages executed, or sent, between nodes in a neighborhood));
aggregating the messages (Arora, [0147] – “The historical transaction data includes information (e.g., various transaction messages) pertaining to the historical transactions executed between the merchants M1-Mn and the consumers C1-Cn.” – teaches aggregating the messages (messages are stored as transaction data for use in creating embeddings));
combining the aggregated message with information associated with a source node (Arora, [0057] – “The fraud detection server 110 then utilizes the processed historical transaction data to generate the first consumer-merchant graph 200. The fraud detection server 110 may utilize one or more graph databases known in the art (e.g., Neo4j, Janus Graph, or the like) to generate the first consumer-merchant graph 200 from the processed historical transaction data. The first consumer-merchant graph 200 includes a plurality of nodes N.sub.M1-N.sub.M3 and N.sub.C1-N.sub.C4 that are connected by way of a plurality of weighted edges E.sub.1-E.sub.10. Each merchant and consumer referenced in the historical transaction data is assigned a unique node in the first consumer-merchant graph 200.” – teaches combining the aggregate message (stored historical transaction data) with information associated with a source node (merchant and consumer data of the historical transaction data, which is an aggregation of passed messages, is assigned, and thus combined with, a unique node in consumer-merchant graph, thus combining the aggregated message with information associated with a source node))
Therefore, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to incorporate the message passing and aggregation of Arora to the method of Shekhar, Sun, and Wadhwa in order to train the graph neural network by computing representations by repeatedly sending messages along edges of a local neighborhood of a node. Doing so would capture complex interdependencies in transaction data to provide for accurate detection of fraudulent transactions (Arora, [0006]).
Regarding claim 10, the combination of Shekhar, Sun, Wadhwa, and Arora teach the method of claim 9, wherein the computed representations are based at least on a context of a respective node (Shekhar, [0099] – “Indeed, by determining that a digital identity corresponds to a fraudulent entity, the fraudulent transaction detection system 106 can determine that digital transactions originating from the fraudulent entity are also fraudulent. Thus, the fraudulent transaction detection system 106 can utilize the node embeddings that correspond to digital identities to identify fraudulent transactions.” – teaches wherein computed representations are based on a context of a respective node (the respective node being the digital entity determined as the fraudulent entity, where computed representations, such as the digital transactions, are based on the context of the respective node)).
Claim(s) 11 and 15 is/are rejected under 35 U.S.C. 103 as being unpatentable over Shekhar, Sun, and Wadhwa as applied to claims 1, 19, and 20 above, and further in view of Liu et al. (NPL: Anomaly Detection in Dynamic Graphs via Transformer, published Nov. 2021, hereinafter “Liu”).
Regarding claim 11, the combination of Shekhar, Sun, and Wadhwa teach the method of claim 1.
The combination of Shekhar, Sun, and Wadhwa fails to explicitly teach wherein at least one node has an associated receptive field is defined by the number of layers of the graph neural network such that the number of layers controls a neighborhood considered for message passing.
However, analogous to the field of the claimed invention, Liu teaches:
wherein at least one node has an associated receptive field is defined by the number of layers of the graph neural network such that the number of layers controls a neighborhood considered for message passing (Liu, Section 4.1 Paragraph 5 – “we consider a sequence of graphs Gt τ = {Gt−τ+1 , · · · , G t} with length τ , where the time window size τ is a hyperparameter and determines the receipt fields on time axis.” – teaches wherein at least one node has an associated receptive field defined by the number of layers of the graph neural network such that the number of layers controls a neighborhood considered for message passing (receipt field defined by τ which is also the length of sequence of graphs)).
Therefore, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to incorporate the receptive fields defined by the number of layers of the graph neural network of Liu to the method of Shekhar, Sun, and Wadhwa in order to define the receipt fields of the nodes as the number of layers in the graph neural network. Doing so would capture the dynamic evolution of the graph and provide sufficient receipt fields for the learning model (Liu, Introduction).
Regarding claim 15, the combination of Shekhar, Sun, and Wadhwa teach the method of claim 1.
The combination of Shekhar, Sun, and Wadhwa fails to explicitly teach wherein performing the self-supervised training of the graph neural network includes jointly training an encoder and a decoder through binary cross-entropy.
However, analogous to the field of the claimed invention, Liu teaches:
wherein performing the self-supervised training of the graph neural network includes jointly training an encoder and a decoder through binary cross-entropy (Liu, Section 4.3 – “With the multiple timestamps of node encoding as input, the dynamic graph transformer can simultaneously capture both spatial and temporal features with a single encoder.” – teaches performing self-supervised training of the graph neural network including jointly training an encoder and decoder, and in Fig. 2 – “The whole framework is trained with a binary cross-entropy loss in an end-to-end manner.” – teaches training the encoder and decoder through binary-cross entropy).
Therefore, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to incorporate the self-supervised training of the graph neural network including jointly training an encoder and a decoder through binary cross-entropy of Liu to the method of Shekhar, Sun, and Wadhwa in order to have a graph neural network trained by jointly training an encoder and a decoder through cross-entropy loss. Doing so would provide a comprehensive node encoding composed of three functional terms to distill global spatial, local spatial, and temporal information (Liu, Introduction).
Response to Arguments
Applicant’s arguments, see pp. 1-3, filed 12 May 2026, with respect to the rejection(s) of claim(s) 1 and 19-20 under Shekhar, Verma, and Wadhwa have been fully considered and are persuasive. Therefore, the rejection has been withdrawn. However, upon further consideration, a new ground(s) of rejection is made over Shekhar, in view of Sun et al. (US Pub. No. 2023/0351215, filed Sept. 2021), and further in view of Wadhwa. Sun teaches the amended limitations of claim 1 regarding “generating a heterogenous graph representation… wherein directed edges connect nodes of the first type representing entities to nodes of the second type representing the discrete financial transfer events” and “wherein learning the embedding representation comprises severing a selected edge…”.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
Jourdan et al. (NPL: Characterizing Entities in the Bitcoin Blockchain, published Feb. 2019) teaches methods for generating and analyzing a Bitcoin transaction graph. The entity-transaction graph is a weighted directed bipartite graph with a first type node corresponding to entities of the graph and a second type node representing transactions of the graph.
Yu et al. (NPL: Neighborhood Extended Dynamic Graph Neural Network, published 2018) teaches methods for representation learning on dynamic graphs. Teaches a temporal attention propagation module for generating message for n-hop neighbors through a self-attention mechanism. Teaches propagation of messages through the nodes of the bipartite directed graph. Teaches wherein the message propagation module considers a receptive field. Teaches wherein the dynamic graphs include edge or node deletion within events or observations of the graph.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to LOUIS C NYE whose telephone number is 571-272-0636. The examiner can normally be reached Monday - Friday 9:00AM - 5:00PM.
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, MATT ELL can be reached at 571-270-3264. 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.
/LOUIS CHRISTOPHER NYE/Examiner, Art Unit 2141 /MATTHEW ELL/Supervisory Patent Examiner, Art Unit 2141
Prosecution Insights