DETAILED ACTION
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
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and the fee set forth in 37 CFR 1.17(e) has been timely paid, the finality of the previous Office action has been withdrawn pursuant to 37 CFR 1.114.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 08/05/2026 has been entered.
The status of the claims is as follows.
Claims 1-5, 18 are amended. Claims 1-20 are currently pending.
Claim Rejections - 35 USC § 103
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.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
Claims 1-8, 18-20 are rejected under 35 U.S.C. 103 as being unpatentable over Feng et al. (“ARIEL: Adversarial Graph Contrastive Learning” [2022], hereinafter “Feng”) in view of She et al. (US 20230088676 A1, hereinafter “She”) in view of Xie et al. (“A Hierarchical Generative Embedding Model for Influence Maximization in Attributed Social Networks” [2022], hereinafter “Xie”).
Regarding Claim 1,
Feng discloses A method of machine learning, comprising: receiving a graph including a node; generating a node embedding for the node based on the graph using a graph neural network (GNN); (Feng [Section 2.1];
PNG
media_image1.png
333
351
media_image1.png
Greyscale
Wherein a method for contrastive graphing is accomplished for learning a passed encoder neural network
Feng [Section 2.3.3]; “ In principle, our framework could be applied on any graph neural network (GNN) architecture for encoder as long as it could be attacked. For simplicity, we employ a two-layer Graph Convolutional Network (GCN) [29] for node-level contrastive learning and a three-layer Graph Isomorphism Network (GIN) [30] for graph-level contrastive learning in this work.” wherein contrastive graphing comprises, in part, receiving a graph G including nodes V; wherein the encoder is a Graph Convolution Network, thus graph representation learning comprising node generation dependent on a passed encoder that is being learned reads on generating a node embedding for the node based on the graph using a graph neural network (GNN))
computing a contrastive learning loss based on the node embedding, a positive sample, and a negative sample, wherein the positive sample and the negative sample each comprise node embeddings …; and updating parameters of the GNN based on the contrastive learning loss; (Feng [Section 2.3.1];
PNG
media_image2.png
51
535
media_image2.png
Greyscale
PNG
media_image3.png
100
540
media_image3.png
Greyscale
PNG
media_image4.png
187
343
media_image4.png
Greyscale
Wherein contrastive learning comprises computation of a contrastive learning loss based on a node embedding matrix
Feng [Algorithm 1];
PNG
media_image5.png
264
358
media_image5.png
Greyscale
Wherein the ARIEL contrastive learning model updates to minimize the contrastive learning loss reads on updating parameters of the GNN
Feng [Section 2.2]; “2.2 InfoNCE Loss InfoNCE loss [28] is the predominant work-horse of the contrastive learning loss, which maximizes the lower bound
PNG
media_image6.png
207
344
media_image6.png
Greyscale
”
Wherein the InfoNCE loss component of the contrastive learning loss reads on computing a node feature loss based on the positive and negative samples)
wherein the node feature loss is based on node features of the node and node features of a different node of the graph, (Feng [Section 2.2]; “2.2 InfoNCE Loss InfoNCE loss [28] is the predominant work-horse of the contrastive learning loss, which maximizes the lower bound
PNG
media_image6.png
207
344
media_image6.png
Greyscale
”
Wherein the InfoNCE loss component of the contrastive learning loss reads on computing a node feature loss based on the positive and negative samples)
and wherein the hierarchical community loss is based on a first node cluster associated with the node and a second node cluster (Feng [Section 2.3.1]; “2.3.1 Node-level Contrastive Learning Given a graph G, two views of the graph G1 = {A1, X1} and G2 = {A2, X2} are first generated. This step can be treated as the data augmentation on the original graph, and various augmentation methods can be used herein. We use random edge dropping and feature masking as GCA does. The node embedding matrix for each graph can be computed as H1 = f(A1, X1) and H2 = f(A2, X2). The corresponding node pairs in two graph views are the positive pairs and all other node pairs are negative. Define θ(u, v) to be the similarity function between vectors u and v, in practice, it is usually chosen as the cosine similarity on the projected embedding of each vector, using a two-layer neural network as the projection head. Denote ui = H1[i, :] and vi = H2[i, :], the contrastive loss is defined as
PNG
media_image7.png
125
343
media_image7.png
Greyscale
wherein this contrastive loss component defined between pairs of node clusters within the graph is interpreted as the hierarchical community loss)
Feng fails to explicitly disclose but She discloses wherein the positive sample and the negative sample each comprise node embeddings from the same graph (She [0093]; “The samples that are close to one another are referred to herein as positive samples, and the samples that are far apart are referred to herein as negative samples. According to this disclosure, and with respect to a heterogeneous graph (or subgraph), a set of positive samples and a set of negative samples are generated for each node using a particular type of sampling”
She [0095]; “To generate positive samples, the graph is traversed (e.g., using a breadth-first-search (BFS) algorithm) from a specified node until reaching another node with the same node type. The BFS searching ensures minimum distance between positive samples and the specified nodes because nodes with closer distance have more similar embeddings. To generate negative samples, a node that has a different node type to the specified node is randomly chosen. Across different epochs, the negative samples can be different nodes, which avoids overfitting to a single negative sample”)
It would have been obvious to modify Feng’s method of obtaining positive and negative samples from a plurality of graphs to instead obtain its positive and negative samples from the same singular graph in a similar fashion to She’s method. One would have been motivated to do so because “the local structure of a node in the graph is fully explored” (She [0096]).
Feng/She fails to explicitly disclose but Xie discloses wherein the contrastive learning loss comprises at least three terms including a first term representing a node feature loss, a second term representing a network homophily loss, and a third term representing a hierarchical community loss … wherein the network homophily loss is based on a neighboring node and a non- neighboring node (Xie [Section 4.2]; “In this subsection, we will introduce the learning procedure of the HGE model. Like most unsupervised embedding models, we formulate network structure preservation as an optimization problem. By optimizing the lower bound of the corresponding loss function, the HGE model converges. Our loss function can be divided into three independent parts, corresponding to the capture of the structure of the hierarchical network, general network and node attributes.”
Xie [Section 4.2.3];
PNG
media_image8.png
529
839
media_image8.png
Greyscale
It would have been obvious to modify Feng/She’s method of obtaining positive and negative samples from the same singular graph to modify its contrastive loss with Xie’s contrastive loss comprising all three of a network homophily, hierarchical loss, and node feature loss. One would have been motivated to do so because the “first one is how to preserve the hierarchical network structure, the second one is how to capture general network structure properties, and the third one is how to acquire node attribute features” (Xie [Section 4.1]) thus allowing the model to cohesively preserve the total plurality of relationships.
Regarding Claim 2,
Feng/She/Xie teaches the method of Claim 1 (and thus the rejection of Claim 1 is incorporated). Feng/She/Xie further discloses identifying node features of the node for the positive sample; identifying node features of a different node of the graph for the negative sample; computing the node feature loss based on the positive sample and the negative sample; (Feng [Section 2.2]; “2.2 InfoNCE Loss InfoNCE loss [28] is the predominant work-horse of the contrastive learning loss, which maximizes the lower bound
PNG
media_image6.png
207
344
media_image6.png
Greyscale
”
Wherein the InfoNCE loss component of the contrastive learning loss reads on computing a node feature loss based on the positive and negative samples)
Regarding Claim 3,
Feng/She/Xie teaches the method of Claim 1 (and thus the rejection of Claim 1 is incorporated). Feng/She/Xie further discloses identifying an edge of the graph; identifying a neighboring node for the positive sample based on the edge; identifying a non-neighboring node for the negative sample; (Feng [Section 2.3.1]; “Given a graph G, two views of the graph G1 = {A1, X1} and G2 = {A2, X2} are first generated. This step can be treated as the data augmentation on the original graph, and various augmentation methods can be used herein. We use random edge dropping and feature masking as GCA does. The node embedding matrix for each graph can be computed as H1 = f(A1, X1) and H2 = f(A2, X2). The corresponding node pairs in two graph views are the positive pairs and all other node pairs are negative.”)
and computing the network homophily loss based on the positive sample and the negative sample;
Xie [Section 4.2.3];
PNG
media_image8.png
529
839
media_image8.png
Greyscale
Wherein the attribute loss associated between neighboring nodes thus reads on a network homophily loss; wherein Feng/She discloses such nodes being of positive and negative samples))
Regarding Claim 4,
Feng/She/Xie teaches the method of Claim 1 (and thus the rejection of Claim 3 is incorporated). Feng/She/Xie further discloses identifying a node triangle based on the edge and the node, wherein the neighboring node is identified based on the node triangle (Feng [Section 3.2 Column 2 Paragraph 2]; “To finally obtain LA from L˜A, each element is independently sampled from a Bernoulli distribution as LA[i, j] ∼ Bernoulli(L˜A[i, j]). While to obtain a symmetric matrix, we only sample the upper triangular part (the elements on the diagonal are known to be 0 in our formulation) and obtain the lower triangular part through transposition.” wherein the adjacency matrix representative of nodes and edge relationships in the graph comprise node triangles)
Regarding Claim 5,
Feng/She/Xie teaches the method of Claim 1 (and thus the rejection of Claim 1 is incorporated). Feng/She/Xie further discloses identifying a first node cluster and a second node cluster, wherein the first node cluster is associated with the node; (Feng [Section 3.4];
PNG
media_image9.png
227
354
media_image9.png
Greyscale
Wherein the two views are read as identified node clusters associated with nodes)
identifying the positive sample based on the first node cluster; identifying the negative sample based on the second node cluster; and computing the hierarchical community loss based on the positive sample and the negative sample (Feng [Section 2.3.1]; “2.3.1 Node-level Contrastive Learning Given a graph G, two views of the graph G1 = {A1, X1} and G2 = {A2, X2} are first generated. This step can be treated as the data augmentation on the original graph, and various augmentation methods can be used herein. We use random edge dropping and feature masking as GCA does. The node embedding matrix for each graph can be computed as H1 = f(A1, X1) and H2 = f(A2, X2). The corresponding node pairs in two graph views are the positive pairs and all other node pairs are negative. Define θ(u, v) to be the similarity function between vectors u and v, in practice, it is usually chosen as the cosine similarity on the projected embedding of each vector, using a two-layer neural network as the projection head. Denote ui = H1[i, :] and vi = H2[i, :], the contrastive loss is defined as
PNG
media_image7.png
125
343
media_image7.png
Greyscale
wherein this contrastive loss component is interpreted as the hierarchical community loss
Feng [Equation 31];
PNG
media_image10.png
74
344
media_image10.png
Greyscale
wherein the contrastive learning loss includes the hierarchical community loss)
Regarding Claim 6,
Feng/She/Xie teaches the method of Claim 1 (and thus the rejection of Claim 1 is incorporated). Feng/She/Xie further discloses computing an updated node embedding for the node based on the updated parameters of the GNN; and computing a node cluster based on the updated node embedding (Feng [Algorithm 1];
PNG
media_image11.png
271
358
media_image11.png
Greyscale
Wherein the node embedding matrix updated for an updated model and its updated parameters is performed
Feng [Section 3.5]; “For a batch of graphs B and the batch of their augmentation views B +, we aim to generate a batch of adversarial views, which we denote as Badv. Denote the combined graph of each batch as G∗ , G+∗ and G∗ adv. The objective of the
PNG
media_image12.png
483
344
media_image12.png
Greyscale
”
wherein the generation of batches of adversarial views comprising computing perturbations of each graph dependent on the current iteration’s Node Embedding matrix reads on computing a node cluster based on updated node embeddings)
Regarding Claim 7,
Feng/She/Xie teaches the method of Claim 6 (and thus the rejection of Claim 6 is incorporated). Feng/She/Xie further discloses computing a cluster centroid based on the node cluster and the updated node embedding (Feng [Section 3.2 Paragraph 3];
PNG
media_image13.png
294
346
media_image13.png
Greyscale
PNG
media_image14.png
495
341
media_image14.png
Greyscale
Feng [Section 3.5]; “For a batch of graphs B and the batch of their augmentation views B +, we aim to generate a batch of adversarial views, which we denote as Badv. Denote the combined graph of each batch as G∗ , G+∗ and G∗ adv. The objective of the
PNG
media_image12.png
483
344
media_image12.png
Greyscale
”
wherein the generation of batches of adversarial views comprising computing perturbations of each graph dependent on the current iteration’s Node Embedding matrix reads on computing a node cluster based on updated node embeddings; wherein the calculation of a feature matrix representative of perturbations to the graph structure to compute an updated cluster thus inherently reads on calculating the cluster’s centroid through the feature matrix)
Regarding Claim 8,
Feng/She/Xie teaches the method of Claim 1 (and thus the rejection of Claim 1 is incorporated). Feng/She/Xie further discloses providing customized content to a user based on the updated parameters of the GNN (Feng [Section 4];
PNG
media_image15.png
317
356
media_image15.png
Greyscale
Feng [Section 4.1.1]; “For the node-level contrastive learning, we use eight datasets for the evaluation, including Cora, CiteSeer, AmazonComputers, Amazon-Photo, Coauthor-CS, Coauthor-Physics, Facebook and LastFM Asia. Cora and CiteSeer [39] are citation networks, where nodes represent documents and edges correspond to citations. Amazon-Computers and AmazonPhoto [35] are extracted from the Amazon co-purchase graph. In these graphs, nodes are the goods and they are connected by an edge if they are frequently bought together. Coauthor-CS and Coauthor-Physics [35] are the co-authorship graphs, where each node is an author and the edge indicates the co-authorship on a paper. Facebook [40] is a page-page graph of verified Facebook pages where edges corresponds to the likes of each other. LastFM Asia [41] is a social network of Asian users, each node represents a user and they are connected via friendship. For the graph-level contrastive learning, we evaluate ARIEL on four datasets from the benchmark TUDataset [42], including the biochemical molecules graphs NCI1, PROTEINS, DD and MUTAG.” wherein the updated GNN learned through ARIEL to provide node classification specific to each dataset (classification of Asian user nodes in the LastFM Asia graph) reads on providing customized content to a user based on the updated parameter of the GNN)
Regarding Claim 18,
Feng discloses a processor; a memory storing instructions executable by the processor; a graph neural network (GNN) configured to generate a node embedding for a node based on a graph; and a training component configured to compute a contrastive learning loss based on the node embedding, wherein the positive sample and the negative sample each comprise node embeddings … and update parameters of the GNN based on the contrastive learning loss (Feng [Section 4];
PNG
media_image16.png
95
352
media_image16.png
Greyscale
Feng [Section 4.6];
PNG
media_image17.png
50
349
media_image17.png
Greyscale
Wherein the NVIDIA Tesla V100S GPU with 32G memory configured to both generate a node embedding for a node based on a graph through ARIEL as well as train ARIEL to compute contrastive learning losses and parameter updates throughout training reads on a processor-based training component and memory
Feng [Section 2.1];
PNG
media_image1.png
333
351
media_image1.png
Greyscale
Wherein a method for contrastive graphing is accomplished for learning a passed encoder neural network
Feng [Section 2.3.3]; “ In principle, our framework could be applied on any graph neural network (GNN) architecture for encoder as long as it could be attacked. For simplicity, we employ a two-layer Graph Convolutional Network (GCN) [29] for node-level contrastive learning and a three-layer Graph Isomorphism Network (GIN) [30] for graph-level contrastive learning in this work.” wherein contrastive graphing comprises, in part, receiving a graph G including nodes V; wherein the encoder is a Graph Convolution Network, thus graph representation learning comprising node generation dependent on a passed encoder that is being learned reads on generating a node embedding for the node based on the graph using a graph neural network
Feng [Section 2.3.1];
PNG
media_image2.png
51
535
media_image2.png
Greyscale
PNG
media_image3.png
100
540
media_image3.png
Greyscale
PNG
media_image4.png
187
343
media_image4.png
Greyscale
Wherein contrastive learning comprises computation of a contrastive learning loss based on a node embedding matrix
Feng [Algorithm 1];
PNG
media_image5.png
264
358
media_image5.png
Greyscale
Wherein the ARIEL contrastive learning model updates to minimize the contrastive learning loss reads on updating parameters of the GNN
Feng [Section 2.2]; “2.2 InfoNCE Loss InfoNCE loss [28] is the predominant work-horse of the contrastive learning loss, which maximizes the lower bound
PNG
media_image6.png
207
344
media_image6.png
Greyscale
”
Wherein the InfoNCE loss component of the contrastive learning loss reads on computing a node feature loss based on the positive and negative samples; wherein the contrastive learning loss includes the node feature loss)
wherein the node feature loss is based on node features of the node and node features of a different node of the graph, (Feng [Section 2.2]; “2.2 InfoNCE Loss InfoNCE loss [28] is the predominant work-horse of the contrastive learning loss, which maximizes the lower bound
PNG
media_image6.png
207
344
media_image6.png
Greyscale
”
Wherein the InfoNCE loss component of the contrastive learning loss reads on computing a node feature loss based on the positive and negative samples)
and wherein the hierarchical community loss is based on a first node cluster associated with the node and a second node cluster (Feng [Section 2.3.1]; “2.3.1 Node-level Contrastive Learning Given a graph G, two views of the graph G1 = {A1, X1} and G2 = {A2, X2} are first generated. This step can be treated as the data augmentation on the original graph, and various augmentation methods can be used herein. We use random edge dropping and feature masking as GCA does. The node embedding matrix for each graph can be computed as H1 = f(A1, X1) and H2 = f(A2, X2). The corresponding node pairs in two graph views are the positive pairs and all other node pairs are negative. Define θ(u, v) to be the similarity function between vectors u and v, in practice, it is usually chosen as the cosine similarity on the projected embedding of each vector, using a two-layer neural network as the projection head. Denote ui = H1[i, :] and vi = H2[i, :], the contrastive loss is defined as
PNG
media_image7.png
125
343
media_image7.png
Greyscale
wherein this contrastive loss component defined between pairs of node clusters within the graph is interpreted as the hierarchical community loss)
Feng fails to explicitly disclose but She discloses wherein the positive sample and the negative sample each comprise node embeddings from the same graph (She [0093]; “The samples that are close to one another are referred to herein as positive samples, and the samples that are far apart are referred to herein as negative samples. According to this disclosure, and with respect to a heterogeneous graph (or subgraph), a set of positive samples and a set of negative samples are generated for each node using a particular type of sampling”
She [0095]; “To generate positive samples, the graph is traversed (e.g., using a breadth-first-search (BFS) algorithm) from a specified node until reaching another node with the same node type. The BFS searching ensures minimum distance between positive samples and the specified nodes because nodes with closer distance have more similar embeddings. To generate negative samples, a node that has a different node type to the specified node is randomly chosen. Across different epochs, the negative samples can be different nodes, which avoids overfitting to a single negative sample”)
It would have been obvious to modify Feng’s method of obtaining positive and negative samples from a plurality of graphs to instead obtain its positive and negative samples from the same singular graph in a similar fashion to She’s method. One would have been motivated to do so because “the local structure of a node in the graph is fully explored because a node can have multiple adjacent nodes, and a meta-path neighbor may be found through each adjacent node” (She [0096]).
Feng/She fails to explicitly disclose but Xie discloses wherein the contrastive learning loss comprises at least three terms including a first term representing a node feature loss, a second term representing a network homophily loss, and a third term representing a hierarchical community loss … wherein the network homophily loss is based on a neighboring node and a non- neighboring node (Xie [Section 4.2]; “In this subsection, we will introduce the learning procedure of the HGE model. Like most unsupervised embedding models, we formulate network structure preservation as an optimization problem. By optimizing the lower bound of the corresponding loss function, the HGE model converges. Our loss function can be divided into three independent parts, corresponding to the capture of the structure of the hierarchical network, general network and node attributes.”
Xie [Section 4.2.3];
PNG
media_image8.png
529
839
media_image8.png
Greyscale
It would have been obvious to modify Feng/She’s method of obtaining positive and negative samples from the same singular graph to modify its contrastive loss with Xie’s contrastive loss comprising a network homophily, hierarchical loss, and node feature loss. One would have been motivated to do so because the “first one is how to preserve the hierarchical network structure, the second one is how to capture general network structure properties, and the third one is how to acquire node attribute features” (Xie [Section 4.1]) thus allowing the model to preserve the plurality of relationships.
Regarding Claim 19,
Feng/She/Xie teaches the method of Claim 18 (and thus the rejection of Claim 18 is incorporated). Feng/She/Xie further discloses a clustering component configured to cluster nodes of the graph based on the node embedding (Feng [Algorithm 1];
PNG
media_image11.png
271
358
media_image11.png
Greyscale
Wherein the node embedding matrix updated for an updated model and its updated parameters is performed
Feng [Section 3.5]; “For a batch of graphs B and the batch of their augmentation views B +, we aim to generate a batch of adversarial views, which we denote as Badv. Denote the combined graph of each batch as G∗ , G+∗ and G∗ adv. The objective of the
PNG
media_image12.png
483
344
media_image12.png
Greyscale
”
wherein the generation of batches of adversarial views comprising computing perturbations of each graph dependent on the current iteration’s Node Embedding matrix reads on computing a node cluster based on updated node embeddings
Feng [Table 3];
PNG
media_image18.png
239
728
media_image18.png
Greyscale
Wherein the node classification component conducted on 32G GPUs reads on a clustering component configured to cluster nodes in classification classes)
Regarding Claim 20,
Feng/She/Xie teaches the method of Claim 18 (and thus the rejection of Claim 18 is incorporated). Feng/She/Xie further discloses a segmentation component configured to segment a plurality of graph snapshots based on an output of the GNN (Feng [Section 3.2];
PNG
media_image19.png
585
351
media_image19.png
Greyscale
Wherein the PGD attack for data augmentation of the original graph through generated additional segments to augment the original graph with reads on a segmentation component based on the GNN generated graph output)
Claims 9-13, 16-17 are rejected under 35 U.S.C. 103 as being unpatentable over Feng et al. (“ARIEL: Adversarial Graph Contrastive Learning” [2022], hereinafter “Feng”) in view of Cirstea et al. (“Graph Attention Recurrent Neural Networks for Correlated Time Series Forecasting” [2021], hereinafter “Cirstea”).
Regarding Claim 9,
Feng discloses A method for contrastive graphing, comprising: receiving a plurality of graph snapshots …; generating, using a graph neural network (GNN), a node embedding representing the at least one node for a graph snapshot of the plurality of graph snapshots (Feng [Section 2.1];
PNG
media_image1.png
333
351
media_image1.png
Greyscale
Wherein a method for contrastive graphing is accomplished for learning a passed encoder neural network
Feng [Section 2.3.2]; “2.3.2 Graph-level Contrastive Learning The graph-level contrastive learning is more close to the contrastive learning in the visual domain. For a batch of graphs B = {G1, · · · , Gb}, we obtain the augmentation of each graph as B + = {G + 1 , · · · , G+ b } through node dropping, subgraph sampling, edge perturbation and feature masking as in GraphCL [21]” wherein the batch of received graphs reads on a plurality of graph snapshots
Feng [Section 2.3.3]; “In principle, our framework could be applied on any graph neural network (GNN) architecture for encoder as long as it could be attacked. For simplicity, we employ a two-layer Graph Convolutional Network (GCN) [29] for node-level contrastive learning and a three-layer Graph Isomorphism Network (GIN) [30] for graph-level contrastive learning in this work.” wherein contrastive graphing comprises, in part, receiving a graph G including nodes V; wherein the encoder is a Graph Convolution Network, thus graph representation learning comprising node generation dependent on a passed encoder that is being learned reads on generating a node embedding for the node based on the graph using a graph neural network (GNN)); wherein such node embeddings being associated with nodes of a graph of the plurality of graph snapshots thus reads on a node embedding representing the at least one node for a graph snapshot of the plurality of graph snapshots
identifying a snapshot segment including a subset of the plurality of graph snapshots …; and generating a merged graph based on the subset of the plurality of graph snapshots in the snapshot segment (Feng [Section 2.3.2 Paragraph 3]; “ Specifically, we notice that a set of graphs with Gi = {Ai , Xi} can be combined into one graph as G∗ = {block diag(A1, · · · , Ab), Concat(X1, · · · , Xb)}. Under this transformation, graph embedding of Gi can be treated as the embedding of a supernode in G∗ . This observation helps us bridge the gap between the node-level contrastive learning and graph-level contrastive learning, where the only difference between them is the granularity of the instance in the contrastive learning loss. Therefore, we can build a universal framework for the graph contrastive learning which can be used for both node-level and graph-level downstream tasks” wherein the snapshot segment comprising the plurality of graphs with Gi = {Ai, Xi} reads on an identified snapshot segment including a subset of the plurality of graph snapshots; wherein the set of graphs combined into one graph reads on generating a merged graph based on the subset of the plurality of graph snapshots)
Feng fails to explicitly disclose but Cirstea discloses wherein each of the plurality of graph snapshots comprises a graph including at least one node and corresponds to a different time within a time span (Cirstea [Section II Subsection B];
PNG
media_image20.png
121
343
media_image20.png
Greyscale
PNG
media_image21.png
322
345
media_image21.png
Greyscale
Wherein each of the graph signals at different timestamps thus reads on the plurality of graph snapshots comprising a graph that includes at least one node
Cirstea [Figure 2];
PNG
media_image22.png
190
348
media_image22.png
Greyscale
wherein the plurality of graph signal snapshots associated with different timestamps of t…t+P time span reads on each of the plurality of graph snapshots corresponding to a different time within the time span)
identifying a snapshot segment including a subset of the plurality of graph snapshots by generating a segment embedding for the snapshot segment and comparing the segment embedding to the node embedding (Cirstea [Section II Subsection B];
PNG
media_image20.png
121
343
media_image20.png
Greyscale
PNG
media_image21.png
322
345
media_image21.png
Greyscale
Cirstea [Section III Subsection B];
PNG
media_image23.png
218
351
media_image23.png
Greyscale
PNG
media_image24.png
578
351
media_image24.png
Greyscale
Wherein the determination of attention scores derived through embeddings associated with select time-series segments thus is understood as the determination of which neighboring vertex’s time-series segments are more relevant (as evidenced by the calculated attention scores to be assigned to graph temporal snapshot segments) through comparison of their embeddings; wherein comparison of such segment embeddings against one another also reads on comparison of their associated node embeddings, and thus a comparison between time-series segment embeddings against other time-series segment embeddings and their implicit node embeddings is achieved)
It would have been obvious to modify the plurality of graph snapshots determined by Feng’s method to be correspondent to different times in a similar fashion to Cirstea’s method. One would have been motivated to do so because “Given historical statuses of all entities, we aim at predicting the future statuses of all entities” (Cirstea [Section II Subsection A]) thus allowing for Feng’s snapshots to comprise historical information useful for prediction of future snapshots.
Regarding Claim 10,
Feng/Cirstea teaches the method of Claim 9 (and thus the rejection of Claim 9 is incorporated). Feng/Cirstea further discloses computing a contrastive learning loss based on the node embedding; and updating parameters of the GNN based on the contrastive learning loss (Feng [Section 2.3.1];
PNG
media_image2.png
51
535
media_image2.png
Greyscale
PNG
media_image3.png
100
540
media_image3.png
Greyscale
PNG
media_image4.png
187
343
media_image4.png
Greyscale
Wherein contrastive learning comprises computation of a contrastive learning loss based on a node embedding matrix
Feng [Algorithm 1];
PNG
media_image5.png
264
358
media_image5.png
Greyscale
Wherein the ARIEL contrastive learning model updates to minimize the contrastive learning loss reads on updating parameters of the GNN)
Regarding Claim 11,
Feng/Cirstea teaches the method of Claim 10 (and thus the rejection of Claim 10 is incorporated). Feng/Cirstea further discloses identifying a node of a first graph snapshot of the snapshot segment; identifying a corresponding node of a second graph snapshot of the snapshot segment for a positive sample; identifying a non-corresponding node of the second graph snapshot for a negative sample (Feng [Section 2.2]; “2.2 InfoNCE Loss InfoNCE loss [28] is the predominant work-horse of the contrastive learning loss, which maximizes the lower bound
PNG
media_image6.png
207
344
media_image6.png
Greyscale
”
Wherein the InfoNCE loss component of the contrastive learning loss reads on computing a temporal consistency loss based on the positive and negative samples; wherein the contrastive learning loss includes the temporal consistency loss)
Regarding Claim 12,
Feng/Cirstea teaches the method of Claim 9 (and thus the rejection of Claim 9 is incorporated). Feng/Cirstea further discloses identifying a snapshot of the plurality of graph snapshots; generating a segment embedding for the snapshot segment; and computing a distance between the snapshot and the snapshot segment based on the node embedding for the snapshot and the segment embedding for the snapshot segment (Feng [Section 3.2];
PNG
media_image25.png
369
345
media_image25.png
Greyscale
Wherein the difference between the adjacency matrices for A and A’ and their feature matrices X and X’ (wherein H= f(A, X), thus reading on the computed distance being inherently based on their node embeddings) reads on computing a distance between the snapshot and the snapshot segment)
Regarding Claim 13,
Feng/Cirstea teaches the method of Claim 12 (and thus the rejection of Claim 12 is incorporated). Feng/Cirstea further discloses determining that the distance is less than a threshold distance; and adding the snapshot to the snapshot segment based on the determination (Feng [Section 3.2];
PNG
media_image25.png
369
345
media_image25.png
Greyscale
Wherein the difference between the adjacency matrices for A and A’ and their feature matrices X and X’ (wherein H= f(A, X), thus reading on the computed distance being inherently based on their node embeddings) reads on computing a distance between the snapshot and the snapshot segment; wherein the distance is thresholded by budget ∆A and ∆X; wherein the adversarial attack data augmentation based on its change thresholded by ∆A and ∆X reads on adding the snapshot to the snapshot segment based on the determination)
Regarding Claim 16,
Feng/Cirstea teaches the method of Claim 9 (and thus the rejection of Claim 9 is incorporated). Feng/Cirstea further discloses generating a plurality of merged graphs corresponding to the plurality of snapshot segments, respectively (Feng [Section 3.5]; “For a batch of graphs B and the batch of their augmentation views B +, we aim to generate a batch of adversarial views, which we denote as Badv. Denote the combined graph of each batch as G∗ , G+∗ and G∗ adv. The objective of the
PNG
media_image12.png
483
344
media_image12.png
Greyscale
”
wherein the generation of batches of adversarial views B* corresponding to snapshot segments B and augmentation snapshot segments B+ reads on generating a plurality of merged graphs corresponding to the plurality of snapshot segments)
Regarding Claim 17,
Feng/Cirstea teaches the method of Claim 9 (and thus the rejection of Claim 9 is incorporated). Feng/Cirstea further discloses clustering nodes of the merged graph to obtain a plurality of node clusters (Feng [Algorithm 1];
PNG
media_image11.png
271
358
media_image11.png
Greyscale
Wherein the node embedding matrix updated for an updated model and its updated parameters is performed
Feng [Section 3.5]; “For a batch of graphs B and the batch of their augmentation views B +, we aim to generate a batch of adversarial views, which we denote as Badv. Denote the combined graph of each batch as G∗ , G+∗ and G∗ adv. The objective of the
PNG
media_image12.png
483
344
media_image12.png
Greyscale
”
wherein the generation of batches of adversarial views comprising computing perturbations of each graph dependent on the current iteration’s Node Embedding matrix reads on computing node clusters)
Claims 14-15 are rejected under 35 U.S.C. 103 as being unpatentable over Feng et al. (“ARIEL: Adversarial Graph Contrastive Learning” [2022], hereinafter “Feng”) in view of Cirstea et al. (“Graph Attention Recurrent Neural Networks for Correlated Time Series Forecasting” [2021], hereinafter “Cirstea”) in view of Wu et al. (US20230334332A1, hereinafter “Wu”).
Regarding Claim 14,
Feng/Cirstea teaches the method of Claim 12 (and thus the rejection of Claim 12 is incorporated). Feng/Cirstea fails to explicitly disclose but Wu discloses determining that the distance is greater than a threshold distance; and adding the snapshot to a subsequent snapshot segment based on the determination (Wu [0090]; “In some embodiments, the construction of V (e.g., corresponding to the set of nodes of the graph) starts with the generating a k-nearest-neighbor graph (k-NNG) of the input z and the nodes in Zref each point in Zref∪{z} is a node in the graph, and an edge from node i to node j exists iff j is among i's top-k nearest neighbors in distance (e.g., Euclidean distance) over the embedding space. In some embodiments, the system then keeps the nodes whose graph distance from z in the kNNG is within a threshold l. For example, if l=1, then the system may only keep the immediate top-k nearest neighbors of z (one-hop neighbors); if l=2, then the system may also keep the k nearest neighbors for each z's one-hop neighbors. As depicted in graph iteration 908 of FIG. 9 (e.g., a first iteration, in which l=1 and k=4), four nodes are selected as nearest neighbors of z (e.g., the center node). In this case, three nodes are benign (e.g., white nodes), and one node is adversarial (e.g., a red node). In a second graph iteration 910, in which l=2, the k nearest neighbors for each of z's one-hop neighbors are determined.” wherein after determining whether graph distance from l is greater than a certain threshold (i.e. if l>1), the system adds the subsequent snapshot segments (neighbors associated with future neighboring nodes) to the current graph snapshot in construction)
It would have been obvious to incorporate Wu’s method of adding a graph snapshot to a subsequent snapshot segment when the calculated distance between the snapshot and the snapshot segment exceeds a certain threshold in Feng/Cirstea’s method of contrastive graphing comprising generating merged graphs through determined node embeddings and calculated distances between graph snapshots. One would have been motivated to do so because “the resulting adversarially-augmented reference set (e.g., including both the clean reference set 902 and the adversarial examples 904) will have twice as many points as the clean reference set … these adversarial samples are able to encode information regarding the layout of adversarial examples to benign examples in the local manifold” (Wu [0090]).
Regarding Claim 15,
Feng/Cirstea teaches the method of Claim 9 (and thus the rejection of Claim 9 is incorporated). Feng/Cirstea fails to explicitly disclose but Wu discloses generating a plurality of snapshot segments by iterating through the plurality of graph snapshots and adding a current snapshot either to a current snapshot segment or a next snapshot segment (Wu [0090]; “ For example, the system may select an attack algorithm, create adversarial examples 904 for all inputs in Zref against the given model, and add the adversarial examples to Zref. In this example, the resulting adversarially-augmented reference set (e.g., including both the clean reference set 902 and the adversarial examples 904) will have twice as many points as the clean reference set. In some embodiments, these adversarial samples are able to encode information regarding the layout of adversarial examples to benign examples in the local manifold. In FIG. 9 , the embedding 906 for the query image (z) corresponds to the object that is to be classified based on generating the LNG, which further corresponds to the center node of the LNG. In some embodiments, the construction of V (e.g., corresponding to the set of nodes of the graph) starts with the generating a k-nearest-neighbor graph (k-NNG) of the input z and the nodes in Zref each point in Zref∪{z} is a node in the graph, and an edge from node i to node j exists iff j is among i's top-k nearest neighbors in distance (e.g., Euclidean distance) over the embedding space. In some embodiments, the system then keeps the nodes whose graph distance from z in the kNNG is within a threshold l. For example, if l=1, then the system may only keep the immediate top-k nearest neighbors of z (one-hop neighbors); if l=2, then the system may also keep the k nearest neighbors for each z's one-hop neighbors. As depicted in graph iteration 908 of FIG. 9 (e.g., a first iteration, in which l=1 and k=4), four nodes are selected as nearest neighbors of z (e.g., the center node). In this case, three nodes are benign (e.g., white nodes), and one node is adversarial (e.g., a red node). In a second graph iteration 910, in which l=2, the k nearest neighbors for each of z's one-hop neighbors are determined. Similar to the first iteration, a combination of benign or adversarial objects may be selected in the second iteration. It should be understood that the system may utilize any suitable values for l and k (e.g., I=1, 2, 4, etc.; k=40, 200, etc.), which may correspond to parameters of a distance metric associated with determining nearest neighbors” wherein the generation of the adversarially-augmented reference set through the augmentation of graph snapshots with its neighbors (adjacent snapshots) reads on generating a plurality of snapshot segments by iterating through graph snapshots and adding a current snapshot to a next snapshot segment (keeping k nearest neighbors for z’s one hop neighbors in the generated graph snapshot segment))
It would have been obvious to incorporate Wu’s method of generating a plurality of snapshot segments by iteratively adding snapshot segments to current snapshots in Feng/Cirstea’s method of contrastive graphing comprising generating merged graphs through determined node embeddings and graph snapshots. One would have been motivated to do so because “the resulting adversarially-augmented reference set (e.g., including both the clean reference set 902 and the adversarial examples 904) will have twice as many points as the clean reference set … these adversarial samples are able to encode information regarding the layout of adversarial examples to benign examples in the local manifold” (Wu [0090]).
Response to Arguments
The Examiner acknowledges the Applicant’s amendments to Claims 1-5, 9, 18.
Applicant’s arguments filed January 8th, 2026, traversing the rejection of claims 1-20 under 35 U.S.C. § 103 have been fully considered, but are not fully persuasive.
Regarding Argument I recited on pages 9-11 of remarks, arguments are moot because the new ground of rejection does not rely on any reference applied in the prior rejection of record for any teaching or matter specifically challenged in this argument.
Regarding Argument II recited on pages 11-14 of remarks, applicant argues Feng and Cirstea do not teach “identifying a snapshot segment including a subset of the plurality of graph snapshots by generating a segment embedding for the snapshot segment and comparing the segment embedding to the node embedding.”
Examiner respectfully disagrees. Examiner interprets the “graph snapshot segment” as simply the series of Cirstea’s graph signal Xt across a particular time period. Each of the feature vectors xi(t) are representative of generated segment embeddings for t time, and the comparison of select vertex’s time-series to determine an attention score thus reads on selection of a select snapshot segment (identification of which time-series was most “relevant”) by generating a segment embedding (feature-vectors associated with time t within the time-series) and comparing the segment embedding to the node embedding (comparison of different time-series’ feature-vectors against one another thus reads on comparing of some segment embeddings to other segment embeddings comprising node embeddings). Thus, Feng/Cirstea is demonstrated to disclose under broadest reasonable interpretation “identifying a snapshot segment including a subset of the plurality of graph snapshots by generating a segment embedding for the snapshot segment and comparing the segment embedding to the node embedding.”.
Regarding Argument III recited on pages 14-16 of remarks, applicant argues there is no prima facie case that claims 1, 9 and 18 are obvious over the cited references due to an improper motivation to combine Feng and She.
Examiner respectfully disagrees. Although She involves meta-path neighbor sampling, She is only meant to demonstrate the advantages of generating positive and negative samples from a singular graph instead of a plurality of graphs. The motivation to combine lies in how She’s exploration is constrained to complete exploration of its local nodal structure and generation of samples in a consequently more local context, thus preserving local structural relationships stronger in comparison to Feng’s methodology which would determine such positive and negative samples in a broader multi-graph context.
The rejection of Claims 1, 9 and 18 under U.S.C. § 103 has been maintained.
The rejection of Claims 2-8, 10-17 and 19-20 under U.S.C. § 103, which depend directly or indirectly from Claims 1, 9 and 18 respectively, have been maintained.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant’s disclosure:
“Classifying Out-of-distribution Data using a Contrastive Loss” (US 12288393 B2) which discloses a contrastive loss comprising in-part a homophily loss as well as a node feature loss
“Systems and Methods for Noise-Robust Contrastive Learning” (US 20210374553 A1) which discloses contrastive loss and associated determined dimensional feature vector embeddings
Any inquiry concerning this communication or earlier communications from the examiner should be directed to JONATHAN J KIM whose telephone number is (571) 272-0523. The examiner can normally be reached 10-7.
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 on (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
/JONATHAN J KIM/Examiner, Art Unit 2141
/MATTHEW ELL/Supervisory Patent Examiner, Art Unit 2141
Prosecution Insights