CTFR 18/014,287 CTFR 96329 Notice of Pre-AIA or AIA Status 07-03-aia AIA 15-10-aia The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA. Response to Arguments 07-37 AIA Applicant's arguments filed 3/06/2026 have been fully considered but they are not persuasive. Regarding applicant arguments for 35 U.S.C 103 rejection, applicant argues in page 13-14 of remarks “ As amended, claim 1 recites, at least in part, "A method for classifying assets as classifying assets as being from a malicious content provider by features of individual content provider entities," "wherein each distant neighborhood corresponds with an aggregator node of the plurality of aggregator nodes, a neighborhood is a subset of nodes surrounding the aggregator node within … Applicant respectfully contends that, as amended, claim 1 is non-obvious in view of the cited references. ” Applicant argues amended limitations that have not been examined. Applicant further argues in page 14 “ In particular, Applicant respectfully contends that, as discussed during the Interview, amended claim 1 recites "a distant neighborhood is a neighborhood of nodes separated from a seed node of the plurality of nodes by at least two intermediate nodes ... " (emphasis added). The Office contends that Hamilton allegedly teaches the element of a distant neighborhood being separated from a seed node by at least two intermediate nodes … As such, the disclosure in Hamilton would not equate to at least two intermediate nodes. ” – The applicant further argues does equate to at least two intermediate nodes, however due to amendments to claim the scope of the claim has changed. Thus the argument is moot. The applicant further argues in page 15 “ Furthermore, Applicant respectfully contends that Hamilton does not teach, nor does the Office cite to Hamilton as teaching, "wherein the label of the seed node is indicative of whether the asset represented by the seed node is from a malicious content provider so as to prevent or restrict provisioning of said asset to a client device." As such, Applicant respectfully contends that Hamilton does not teach the elements of amended claim 1. Moreover, Applicant respectfully contends that none of Park, Klicpera, Jin, or Sun teach the missing element(s) of amended claim 1, …. Each of the dependent claims is patent-eligible at least for the same reasons as its respective base claim. As such, Applicant respectfully contends that each of claims 1-8, 11, 14, and 16-25 are allowable under 35 U.S.C. § 103. ” The applicant further argues how the prior art of record does not teach the amended limitations, however the amended limitations have not been examined, rendering the argument moot. The argument presented by the applicant is not persuasive . Claim Objections 07-29-01 AIA Claim 1 and 16 are objected to because of the following informalities: Claim 1 recites the limitation in line 6 “ as being from a malicious content provider ”. The claim previously recites a malicious content provider and the second recitation should be the malicious content provider. Also in claim 1 line 25 it recites again “ a malicious content provider ” and should be “the malicious content provider ” Claim 16 is analogous to claim 1 and has the same issue. ( Examiner Note: The claim could recite “A method for classifying assets as being from malicious content providers ” allowing for the later identification of a malicious content provider. However the claim would have to be amended to identify a particular malicious content provider given that the specification supports the change. ) Appropriate correction is required. Claim Rejections - 35 USC § 112 07-30-02 AIA The following is a quotation of 35 U.S.C. 112(b): (b) CONCLUSION.—The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the inventor or a joint inventor regards as the invention. The following is a quotation of 35 U.S.C. 112 (pre-AIA), second paragraph: The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the applicant regards as his invention. 07-34-01 Claim 1-8, 11, 14 and 16-25 rejected under 35 U.S.C. 112(b) or 35 U.S.C. 112 (pre-AIA), second paragraph, as being indefinite for failing to particularly point out and distinctly claim the subject matter which the inventor or a joint inventor (or for applications subject to pre-AIA 35 U.S.C. 112, the applicant), regards as the invention. Claim 1 recites “ malicious or deceptive ” and “ behavior and/or behavior otherwise non-conforming with one or more policies associated with said asset rendered the claim indefinite as the use of “ and/or ” along with the “ or ” previously stated make the scope of the claim unascertainable. Thus limitation “ wherein assets are classified as being from a malicious content provider if associated with malicious or deceptive behavior and/or behavior otherwise non-conforming with one or more policies associated with said asset, ” in indefinite as to when a content is classified as being from a malicious content provider. Further the claim is render indefinite as the claim has 3 recitations of “ a malicious content provider ” in line 1 and 6 and line 25. Claims 16 is analogous to claim 1 and has the same issue. All dependent claims inherit the issue. Allowable Subject Matter Claim 11 and analogous claim 24 would be allowable if rewritten to overcome the rejection(s) under 35 U.S.C. 112(b) or 35 U.S.C. 112 (pre-AIA), 2nd paragraph, set forth in this Office action and to include all of the limitations of the base claim and any intervening claims. Claim 11 and analogous claim 24 depend upon rejected claims 8 and analogous 23. Regarding claim 11 and analogous claim 24 the cited prior art used to reject the claims 8 and 23 fail to teach wherein a vector for an entity type is defined as ∑ i r i f T ( x i ) , wherein r i is a personalized page ranking with starting node i to a respective aggregator node corresponding to a distant neighborhood, each x i is a feature of the features of node i, and f T is a learnt neural network embedding specific to entity type T. The closest prior art Kim et al., "Tripartite Heterogeneous Graph Propagation for Large-scale Social Recommendation“, (2019) (“Kim”) teaches a Tripartite graph that uses personalized page ranking (Kim page 3 4.1 CTR Prediction para 2, We split the nodes by types: X1,X2, ..., X | O | , apply each predicting neural network: H i = f i ( X i ) and concatenate the results: H = [H1,H2, ..., X | O | ]. Starting with Z r ( o ) = H ∈ R | V | x m for each edge type r ∈ R, HGP uses similar scheme as Equation 1. The purpose of APPNP is not to learn deep node embedding, but to learn a transformation from attributes to class labels in the semi-supervised setting. HGP instead uses non-linear propagation with additional learnable weights to learn deep node representations: PNG media_image1.png 58 468 media_image1.png Greyscale ). However in combination with the prior art used to reject claim 8 and analogous 23 fail to teach or suggest wherein a vector for an entity type is defined as ∑ i r i f T ( x i ) , wherein r i is a personalized page ranking with starting node i to a respective aggregator node corresponding to a distant neighborhood, each x i is a feature of the features of node i, and f T is a learnt neural network embedding specific to entity type T. In particular the limitation of a vector as defined by ∑ i r i f T ( x i ) is not found in the prior art and/or would not be obvious to one of ordinary skill in the art absent Impermissible hindsight. Claim Rejections - 35 USC § 103 07-20-aia AIA 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. 07-21-aia AIA Claim( s) 1, 5, 14 16, 20, and 25 are r ejected under 35 U.S.C. 103 as being unpatentable over T orkamani et al. (US11743282B1) (“Torkamani”) in view of William L. Hamilton, Rex Ying, Jure Leskovec "Inductive Representation Learning on Large Graphs" (“Hamilton”) and Grover, Aditya, and Jure Leskovec. "node2vec: Scalable Feature Learning for Networks." arXiv preprint arXiv:1607.00653 (2016) (“Grover”), and further in view of Namyong Park, Andrey Kan, Xin Luna Dong, Tong Zhao, and Christos Faloutsos. 2019. Estimating Node Importance in Knowledge Graphs Using Graph Neural Networks. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (KDD '19). Association for Computing Machinery, New York, NY, USA, 596–606 (“Park”). R egarding claim 1 and analogous claim 16, as best understood based on the 112(b) identified above, Torkamani teaches a method for classifying assets as being from a malicious content provider by features of individual content provider entities and relations of the individual content provider entities to the assets using a neural network configured to maintain a network of nodes including a plurality of nodes and edges, each node of the plurality of nodes representing a respective asset of a plurality of assets corresponding to a plurality of content sources of the individual content provider entities, wherein assets are classified as being from a malicious content provider if associated with malicious or deceptive behavior and/or behavior otherwise non-conforming with one or more policies associated with said asset, the method comprising (Torkamani PNG media_image2.png 738 597 media_image2.png Greyscale Col 3 line 52-58, In one or more embodiments, the system may use one or more techniques to determine a probability of a cloud-based environment entity being malicious. For example, one technique may include a semi-supervised graph convolution method with minimal parameters and features tailored to minimize the number of false positives when identifying malicious entities (e.g., a trust-based technique)[ using a neural network configured to maintain a network of nodes including a plurality of nodes and edges, each node of the plurality of nodes representing a respective asset of a plurality of assets corresponding to a plurality of content sources of the individual content provider entities, ]. Col 8 line 24-40, In one or more embodiments, as shown in FIG. 2, the k-partite graph extracted at step 130 may include multiple types of edges, such as edges between domains and IP addresses to which the domains have resolved, edges between elastic computing instances and domains that the instances have queried, edges between FQDNs and their pairs of second-level domains and top-level domains, edges between elastic computing instances and cloud-based environment with which accounts the instances are associated, and the like. By extending to new data types other than DNS, the one or more reputation scoring services 126 may determine reputation scores for other entities in the cloud-based environment 110. For example, the one or more reputation scoring services 126 may be able to identify that all executable binaries from a specific domain should be considered malicious, or all domains with a certain security certificate are related [ A method for classifying assets as being from a malicious content provider by features of individual content provider entities and relations of the individual content provider entities to the assets ]. Col 18 line 34-52, FIG. 4A illustrates a flow diagram for a process 400 for cloud-based entity reputation scoring, in accordance with one or more example embodiments of the present disclosure. At block 402, a system (or device, e.g., the cloud-based environment 110 of FIG. 1 and FIG. 3) may determine data, such as DNS data (e.g., queries and responses), flow log data, security certificate authorities, pipeline data, and the like, representing interactions involving entities in and/or external to a cloud-based environment. For example, a cloud-based computing instance (e.g., a serverless computing instance, au elastic computing instance, etc.) may query or otherwise interact with domains (e.g., FQDNs), other entities in the cloud-based environment, and/or other entities external to the cloud-based environment. The data may indicate the instances, domains, and/or other entities used, some of which may have known scores/labels indicating whether the entities are malicious or not, and some entities not having known scores/labels indicating whether an entity is malicious. FIG. 4A, PNG media_image3.png 636 467 media_image3.png Greyscale [ wherein assets are classified as being from a malicious content provider if associated with malicious or deceptive behavior and/or behavior otherwise non-conforming with one or more policies associated with said asset, the method comprising ]) : [and determining, by the one or more processors, a label of the seed node based on the state of the seed node,] wherein the label of the seed node is indicative of whether the asset represented by the seed node is from a malicious content provider so as to prevent or restrict provisioning of said asset to a client device (Torkamani FIG. 1, PNG media_image4.png 747 611 media_image4.png Greyscale Col 20 13-18, At block 414, the system may determine reputation scores for the entities based on the feature vector. The system may determine whether a node should be labeled/scored as malicious or non-malicious (e.g., using a -1 or+ 1 value as described above) based on whether the node is connected to any known malicious nodes. Col 20 line 28-45, At block 416, the system optionally may allow/approve or block communications with/using entities based on the reputation scores. For example, when users and/or entities query, call, or otherwise communicate with other entities, the system may determine the reputation score of the entity, and allow the communication when the reputation score satisfies (e.g., exceeds or is below) a score threshold. When the reputation score does not satisfy a threshold, indicative of the entity being malicious, the system may block communications using a malicious entity [ wherein the label of the seed node is indicative of whether the asset represented by the seed node is from a malicious content provider ]. The system additionally or alternatively may alert/warn users of possibly malicious entities. The system, additionally or alternatively, may reduce false positive identification of malicious entities by filtering low-scoring entities (e.g., domains) from a list of malicious domains (e.g., third-party lists of malicious domains). The system may send the reputation scores to another service that may monitor communications based on the scores [ a malicious content provider so as to prevent or restrict provisioning of said asset to a client device. ].) . Torkamani does not explicitly teach aggregating, by one or more processors and at each of a plurality of aggregator nodes in the network of nodes, data regarding features of the asset of each node in each distant neighborhood of a multiplicity of distant neighborhoods, wherein each distant neighborhood corresponds with an aggregator node of the plurality of aggregator nodes, a neighborhood is a subset of nodes surrounding the aggregator node within a predefined distance, and a distant neighborhood is a neighborhood of nodes separated from a seed node of the plurality of nodes by at least two intermediate nodes; updating, by the one or more processors and at each of the plurality of aggregator nodes, a state of the aggregator node by assigning a weight to each of the features of the asset of each node of the corresponding distant neighborhood; updating, by the one or more processors and at each of the plurality of aggregator nodes, a state of the aggregator node by assigning a weight to each of the features of the asset of each node of the corresponding distant neighborhood; updating, by the one or more processors and at the seed node, a state of the seed node by performing convolutional analysis of each node in a local neighborhood surrounding the seed node, the local neighborhood including each of the plurality of aggregator nodes; However Hamilton teaches aggregating, by one or more processors and at each of a plurality of aggregator nodes in the network of nodes, data regarding features of the asset of each node in each distant neighborhood of a multiplicity of distant neighborhoods (Hamilton Page 4 Algorithm 1, PNG media_image5.png 363 754 media_image5.png Greyscale [ aggregating, by one or more processors and at each of a plurality of aggregator nodes in the network of nodes,] ( i.e. see Algorithm 1 line 4 ) 3.1 Embedding generation (i.e.., forward propagation) algorithm para 2 line 1-6, Algorithm 1 describes the embedding generation process in the case where the entire graph, G = (V; E), and features for all nodes x v , ∀ v ∈ V , are provided as input. We describe how to generalize this to the minibatch setting below. Each step in the outer loop of Algorithm 1 proceeds as follows, where k denotes the current step in the outer loop (or the depth of the search) and h k denotes a node’s representation at this step: First, each node v v ϵ V aggregates the representations of the nodes in its immediate neighborhood, { h u k - 1 , ∀ u ∈ N ( v ) } , into a single vector h N ( u ) k - 1 [ data regarding features of each node in each distant neighborhood of a multiplicity of distant neighborhoods ] ( see Algorithm 1 Input Page 14 Hardware Except for DeepWalk, we ran experiments single a machine with 4 NVIDIA Titan X Pascal GPUs (12Gb of RAM at 10Gbps speed), 16 Intel Xeon CPUs (E5-2623 v4 @ 2.60GHz), and 256Gb of RAM. DeepWalk was faster on a CPU intensive machine with 144 Intel Xeon CPUs (E7-8890 v3 @ 2.50GHz) and 2Tb of RAM. Overall, our experiments took about 3 days in a shared resource setting. We expect that a consumer-grade single-GPU machine (e.g., with a Titan X GPU) could complete our full set of experiments in 4-7 days, if its full resources were dedicated.)) , wherein each distant neighborhood corresponds with an aggregator node of the plurality of aggregator nodes (Hamilton Page 4, 3.1 Embedding generation (i.e.., forward propagation) algorithm para 2 line 1-8, Algorithm 1 describes the embedding generation process in the case where the entire graph, G = (V; E), and features for all nodes x v , ∀ v ∈ V , are provided as input. We describe how to generalize this to the minibatch setting below. Each step in the outer loop of Algorithm 1 proceeds as follows, where k denotes the current step in the outer loop (or the depth of the search) and h k denotes a node’s representation at this step: First, each node v v ϵ V aggregates [ with an aggregator node of the plurality of aggregator nodes ] the representations of the nodes in its immediate neighborhood, { h u k - 1 , ∀ u ∈ N ( v ) } , into a single vector h N ( u ) k - 1 ). Note that this aggregation step depends on the representations generated at the previous iteration of the outer loop (i.e., k - 1), and the k = 0 (“base case”) representations are defined as the input node features., and the k = 0 (“base case”) representations are defined as the input node features [ distant neighborhood corresponds ].) , a neighborhood is a subset of nodes surrounding the aggregator node within a predefined distance (Hamilton Page 4 3.1 Embedding generation (i.e.., forward propagation) algorithm para 2 line 12-15 and Para 3, For notational convenience, we denote the final representations output at depth K as K a s z v ≡ h v K , ∀ v ∈ V . The aggregation of the neighbor representations can be done by a variety of aggregator architectures (denoted by the AGGREGATE placeholder in Algorithm 1), and we discuss different architecture choices in Section 3.3 below. To extend Algorithm 1 to the minibatch setting, given a set of input nodes, we first forward sample the required neighborhood sets (up to depth K) and then we run the inner loop (line 3 in Algorithm 1), but instead of iterating over all nodes, we compute only the representations that are necessary to satisfy the recursion at each depth (Appendix A contains complete minibatch pseudocode)) , updating, by the one or more processors and at each of the plurality of aggregator nodes, a state of the aggregator node [by assigning a weight to each of the features of the asset of each node of the corresponding distant neighborhood] (Hamilton page 12 algorithm 2, PNG media_image6.png 572 751 media_image6.png Greyscale [ a state of the aggregator node ] Appendices A Minibatch pseudocode In order to use stochastic gradient descent, we adapt our algorithm to allow forward and backward propagation for minibatches of nodes and edges. Here we focus on the minibatch forward propagation algorithm, analogous to Algorithm 1. In the forward propagation of GraphSAGE the minibatch B contains nodes that we want to generate representations for. Algorithm 2 gives the pseudocode for the minibatch approach. The main idea is to sample all the nodes needed for the computation first. Lines 2-7 of Algorithm 2 correspond to the sampling stage. Each set Bk contains the nodes that are needed to compute the representations of nodes v 2 Bk+1, i.e., the nodes in the (k + 1)-st iteration, or “layer”, of Algorithm 1 [ updating, by the one or more processors and at each of the plurality of aggregator nodes, a state of the aggregator node by assigning ].) ; updating, by the one or more processors and at the seed node, a state of the seed node by performing convolutional analysis of each node in a local neighborhood surrounding the seed node, the local neighborhood including each of the plurality of aggregator nodes (Hamilton page 2 Figure 1, PNG media_image7.png 302 743 media_image7.png Greyscale [ updating, by the one or more processors and at the seed node, ] Page 5 3.3 Aggregator Architectures, Mean aggregator. Our first candidate aggregator function is the mean operator, where we simply take the elementwise mean of the vectors in { h u h - 1 , ∀ u ∈ N ( u ) } . The mean aggregator is nearly equivalent to the convolutional propagation rule used in the transductive GCN framework [17]. In particular, we can derive an inductive variant of the GCN approach by replacing lines 4 and 5 in Algorithm 1 with the following:4 PNG media_image8.png 30 577 media_image8.png Greyscale We call this modified mean-based aggregator convolutional since it is a rough, linear approximation of a localized spectral convolution [17]. An important distinction between this convolutional aggregator and our other proposed aggregators is that it does not perform the concatenation operation in line 5 of Algorithm 1—i.e., the convolutional aggregator does concatenate the node’s previous layer representation h v k - 1 with the aggregated neighborhood vector h N v k . This concatenation can be viewed as a simple form of a “skip connection” [13] between the different “search depths”, or “layers” of the GraphSAGE algorithm, and it leads to significant gains in performance (Section 4) [ a state of the seed node by performing convolutional analysis of each node in a local neighborhood surrounding the seed node, the local neighborhood including each of the plurality of aggregator nodes; ].) ; Torkamani and Hamilton are considered to be analogous to the claim invention because they are in the same field of distributed machine learning using graph neural networks. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filling date of the claimed invention to have modified Torkamani in view of Hamilton use aggregation nodes to classify the category of unseen nodes. Doing so to generate embeddings for unseen nodes and train without taks-specific supervision (Hamilton Page 2, 1 Introduction para 6 line 1-7, Instead of training a distinct embedding vector for each node, we train a set of aggregator functions that learn to aggregate feature information from a node’s local neighborhood (Figure 1). Each aggregator function aggregates information from a different number of hops, or search depth, away from a given node. At test, or inference time, we use our trained system to generate embeddings for entirely unseen nodes by applying the learned aggregation functions. Following previous work on generating node embeddings, we design an unsupervised loss function that allows GraphSAGE to be trained without task-specific supervision.). Park teaches [updating, by the one or more processors and at each of the plurality of aggregator nodes, a state of the aggregator node] by assigning a weight to each of the features of the asset of each node of the corresponding distant neighborhood; Torkamani and Park are considered to be analogous to the claim invention because they are in the same field of machine learning using graph neural networks. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filling date of the claimed invention to have modified Torkamani in view of Park to disclose assigning weights to each of the features. Doing so to better represent an entity by its neighbors and determine the importance of node features (Park page 598 3 Method, Awareness: GNN normally propagates information between neighbors through node embedding. This is to model the assumption that an entity and its neighbors affect each other, and thus the representation of an entity can be better represented in terms of the representation of its neighbors. In the context of node importance estimation, neighboring importance scores play a major role on the importance of a node, whereas other neighboring features may have little effect, if any. We thus directly aggregate importance scores from neighbors (Section 3.1), and show empirically that it outperforms embedding propagation (Section 4.4)). Grover teaches and a distant neighborhood is a neighborhood of nodes separated from a seed node of the plurality of nodes by at least two intermediate nodes (Grover Page 2, Figure 1, PNG media_image9.png 273 627 media_image9.png Greyscale Page 4-5, Benefits of random walks. There are several benefits of random walks over pure BFS/DFS approaches. Random walks are computationally efficient in terms of both space and time requirements. The space complexity to store the immediate neighbors of every node in the graph is O(|E|). For 2nd order random walks, it is helpful to store the interconnections between the neighbors of every node, which incurs a space complexity of O( O ( a 2 | V | ) ) where a is the average degree of the graph and is usually small for real world networks. The other key advantage of random walks over classic search-based sampling strategies is its time complexity. In particular, by imposing graph connectivity in the sample generation process, random walks provide a convenient mechanism to increase the effective sampling rate by reusing samples across different source nodes. By simulating a random walk of length l > k we can generate k samples for l - k nodes at once due to the Markovian nature of the random walk. Hence, our effective complexity is O ( l k ( l - k ) ) per sample. For example, in Figure 1 we sample a random walk { u , s 4 , s 5 , s 6 , s 8 , s 9 } of length l = 6, which results in N S u = s 4 , s 5 , s 6 , N S s 4 = s 5 , s 6 , s 8 a n d N S s 5 = { s 6 , s 8 , s 9 } . Note that sample reuse can introduce some bias in the overall procedure. However, we observe that it greatly improves the efficiency [ and a distant neighborhood is a neighborhood of nodes separated from a seed node of the plurality of nodes by at least two intermediate nodes; ]) ; Torkamani and Grover are considered to be analogous to the claim invention because they are in the same field of machine learning using graphs. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filling date of the claimed invention to have modified Torkamani in view of Grover to disclose using random walk. Doing so to increase the sampling rate by reusing samples across different source nodes and having a lower runtime (Page 4-5, Benefits of random walks. There are several benefits of random walks over pure BFS/DFS approaches. Random walks are computationally efficient in terms of both space and time requirements. The space complexity to store the immediate neighbors of every node in the graph is O(|E|). For 2nd order random walks, it is helpful to store the interconnections between the neighbors of every node, which incurs a space complexity of O( O ( a 2 | V | ) ) where a is the average degree of the graph and is usually small for real world networks. The other key advantage of random walks over classic search-based sampling strategies is its time complexity. In particular, by imposing graph connectivity in the sample generation process, random walks provide a convenient mechanism to increase the effective sampling rate by reusing samples across different source nodes. By simulating a random walk of length l > k we can generate k samples for l - k nodes at once due to the Markovian nature of the random walk. Hence, our effective complexity is O ( l k ( l - k ) ) per sample. For example, in Figure 1 we sample a random walk { u , s 4 , s 5 , s 6 , s 8 , s 9 } of length l = 6, which results in N S u = s 4 , s 5 , s 6 , N S s 4 = s 5 , s 6 , s 8 a n d N S s 5 = { s 6 , s 8 , s 9 } . Note that sample reuse can introduce some bias in the overall procedure. However, we observe that it greatly improves the efficiency). Regarding claim 5 and analogous claim 20, Torkamani in view of Hamilton, Grover, and Park teach the method of claim 1. Torkamani, Hamilton, Grover, and Park are combine in the same rational as set forth above with respect to claim 1 and analogous claim 16. Hamilton teaches wherein each node of each distant neighborhood is outside of a convolutional window around the seed node (Hamilton page 4 3.1 Embedding generation (i.e., forward propagation) algorithm para 2, Algorithm 1 describes the embedding generation process in the case where the entire graph, G = (V; E), and features for all nodes x v , ∀ v ∈ V , are provided as input. We describe how to generalize this to the minibatch setting below. Each step in the outer loop of Algorithm 1 proceeds as follows, where k denotes the current step in the outer loop (or the depth of the search) and h k denotes a node’s representation at this step: First, each node v ∈ V aggregates the representations of the nodes in its immediate neighborhood, { h u k - 1 , ∀ v ∈ N v } , into a single vector h N ( v ) k - 1 . Note that this aggregation step depends on the representations generated at the previous iteration of the outer loop (i.e., ), and the k = 0 (“base case”) representations are defined as the input node features. After aggregating the neighboring feature vectors, GraphSAGE then concatenates the node’s current representation, hk1 v , with the aggregated neighborhood vector, h N ( v ) k - 1 ., and this concatenated vector is fed through fully connected layer with nonlinear activation function _, which transforms the representations to be used at the next step of the algorithm (i.e., h u k , ∀ v ∈ N v ). ( i.e. at each step of the process includes more nodes not in the current convolutional window in a distance neighborhood )) . Regarding claim 14 and analogous claim 25, Torkamani in view of Hamilton, Grover, and Park teach the method of claim 1. Torkamani, Hamilton, Grover, and Park are combine in the same rational as set forth above with respect to claim 1 and analogous claim 16. Hamilton teaches further comprising passing at least one message from each of the aggregator nodes to the seed node before determining a label of the seed node (Hamilton page 2 PNG media_image10.png 381 947 media_image10.png Greyscale page 4 3.1 Embedding generation (i.e., forward propagation) algorithm para 2, Algorithm 1 describes the embedding generation process in the case where the entire graph, G = (V; E), and features for all nodes x v , ∀ v ∈ V , are provided as input. We describe how to generalize this to the minibatch setting below. Each step in the outer loop of Algorithm 1 proceeds as follows, where k denotes the current step in the outer loop (or the depth of the search) and h k denotes a node’s representation at this step: First, each node v ∈ V aggregates the representations of the nodes in its immediate neighborhood, { h u k - 1 , ∀ v ∈ N v } , into a single vector h N ( v ) k - 1 . Note that this aggregation step depends on the representations generated at the previous iteration of the outer loop (i.e., ), and the k = 0 (“base case”) representations are defined as the input node features [ passing at least one message from each of the aggregator nodes ]. After aggregating the neighboring feature vectors, GraphSAGE then concatenates the node’s current representation, hk1 v , with the aggregated neighborhood vector, h N ( v ) k - 1 ., and this concatenated vector is fed through fully connected layer with nonlinear activation function _, which transforms the representations to be used at the next step of the algorithm (i.e., h u k , ∀ v ∈ N v ). ( i.e. all labels are aggregated to determine a label for the seed node )) . 07-21-aia AIA Claim (s) 2, 3, 4, 17, 18, 19 are rejected under 35 U.S.C. 103 as being unpatentable over Torkamani in view of Hamilton, Grover, and Park and further in view of Klicpera et al.,"Predict then Propagate: Graph Neural Networks meet Personalized PageRank," Cornell University Library (2019) (“Klicpera”) . Regarding claim 2 and analogous claim 17, Torkamani in view of Hamilton, Grover, and Park teach the method of claim 1. Torkamani, Hamilton, Grover, and Park are combine in the same rational as set forth above with respect to claim 1 and analogous claim 16. Torkamani does not explicitly teach wherein assigning the weight is in accordance with an influence function proportional to a personalized page ranking from a first node to a second node. However Klicpera teaches wherein assigning the weight is in accordance with an influence function proportional to a personalized page ranking from a first node to a second node (Klicpera page 2-3 Graph Convolutional Networks and Their Limited Range, We first introduce our notation and explain the problem our model solves. Let G = (V,E) be a graph with nodes V and edges E. Let n denote the number of nodes and m the number of edges. The nodes are described by the feature matrix X ∈ R n x f , with the number of features f per node, and the class (or label) matrix Y ∈ R n x f , with the number of classes c. The graph G is described by the adjacency matrix A ∈ R n x n . A ~ = A + I n denotes the adjacency matrix with added self-loops. One simple and widely used message passing algorithm for semi-supervised classification is the Graph Convolutional Network (GCN). In the case of two message passing layers its equation is PNG media_image11.png 178 1012 media_image11.png Greyscale With two GCN-layers, only neighbors in the two-hop neighborhood are considered. There are essentially two reasons why a message passing algorithm like GCN cannot be trivially expanded to use a larger neighborhood. First, aggregation by averaging causes oversmoothing if too many layers are used. It, therefore, loses its focus on the local neighborhood (Li et al., 2018)… We will start by concentrating on the first issue. Xu et al. (2018) have shown that for a k-layer GCN the influence score of node x on y, PNG media_image12.png 42 244 media_image12.png Greyscale , is proportional in PNG media_image13.png 388 923 media_image13.png Greyscale [ wherein assigning the weight is in accordance with an influence function proportional a personalized page ranking ] expectation to a slightly modified k-step random walk distribution starting at the root node x, P r w ' ( x → y , k ) . Hence, the information of node x spreads to node y in a random walk-like manner [ proportional to a personalized page ranking from a first node to a second node ]. If we take the limit k ! 1 and the graph is irreducible and aperiodic, this random walk probability distribution P r w ' ( x → y , k ) converges to the limit (or stationary) distribution Plim(! y). This distribution can be obtained by solving the equation PNG media_image14.png 37 132 media_image14.png Greyscale ( i.e. the weights are affected by the influence scores )) . Torkamani and Klicpera are considered to be analogous to the claim invention because they are in the same field of machine learning using graph neural networks. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filling date of the claimed invention to have modified Torkamani in view of Klicpera to use an influence function. Doing so to allow GCN to use infinitely many neighborhood aggregation layers which was not possible in classical message passing framework (Klicpera 3 PERSONALIZED PROPAGATION OF NEURAL PREDICTIONS Page 4 para 3, As a consequence, PPNP separates the neural network used for generating predictions from the propagation scheme. This separation additionally solves the second issue mentioned above: the depth of the neural network is now fully independent of the propagation algorithm. As we saw when connecting GCN to PageRank, personalized PageRank can effectively use even infinitely many neighborhood aggregation layers, which is clearly not possible in the classical message passing framework. Furthermore, the separation gives us the flexibility to use any method for generating predictions, e.g. deep convolutional neural networks for graphs of images). Regarding claim 3 and analogous claim 18, Torkamani in view of Hamilton, Grover, and Park and Klicpera teach the method of claim 2. Torkamani, Hamilton, Grover, and Park are combine in the same rational as set forth above with respect to claim 1 and analogous claim 16. Torkamani and Klicpera are combine in the same rational as set forth above with respect to claim 2 and analogous claim 17. Klicpera further teaches wherein the influence function is I x , y = ∑ i ∑ j ∂ y j / ∂ x i , wherein x is a first node in the distant neighborhood, each x i is a feature of node x, y is a second node in the distant neighborhood, and each y j is a feature of node y (Klicpera page 3, 2 GRAPH CONVOLUTIONAL NETWORKS AND THEIR LIMITED RANGE Para 2 We will start by concentrating on the first issue. Xu et al. (2018) have shown that for a k-layer GCN the influence score of node x on y, PNG media_image12.png 42 244 media_image12.png Greyscale [ wherein the influence function is I x , y = ∑ i ∑ j ∂ y j / ∂ x i , ], is proportional in PNG media_image13.png 388 923 media_image13.png Greyscale expectation to a slightly modified k-step random walk distribution starting at the root node x, P r w ' ( x → y , k ) . Hence, the information of node x spreads to node y in a random walk-like manner [ wherein x is a first node in the distant neighborhood, each x i is a feature of node x, ] [ y is a second node in the distant neighborhood, and each y j is a feature of node y ]) . Regarding claim 4 and analogous claim 19, Torkamani in view of Hamilton, Grover, and Park and Klicpera teach the method of claim 3. Torkamani, Hamilton, Grover, and Park are combine in the same rational as set forth above with respect to claim 1 and analogous claim 16. Torkamani and Klicpera are combine in the same rational as set forth above with respect to claim 2 and analogous claim 17. Klicpera further teaches wherein y is an aggregator node of the plurality of aggregator nodes (Klicpera page 3, 2 GRAPH CONVOLUTIONAL NETWORKS AND THEIR LIMITED RANGE Para 2 We will start by concentrating on the first issue. Xu et al. (2018) have shown that for a k-layer GCN the influence score of node x on y, PNG media_image12.png 42 244 media_image12.png Greyscale [ wherein the influence function is I x , y = ∑ i ∑ j ∂ y j / ∂ x i , ], is proportional in PNG media_image13.png 388 923 media_image13.png Greyscale expectation to a slightly modified k-step random walk distribution starting at the root node x, P r w ' ( x → y , k ) . Hence, the information of node x spreads to node y in a random walk-like manner [ wherein y is an aggregator node of the aggregator nodes ]) . 07-21-aia AIA Claim (s) 6 and 21 are rejected under 35 U.S.C. 103 as being unpatentable over Torkamani in view of Hamilton, Grover, and Park and further in view of Jin, Di, et al. "Universal graph convolutional networks." Advances in neural information processing systems 34 (2021): 10654-10664 (“Jin”) . Regarding claim 6 and analogous claim 21, Torkamani in view of Hamilton, Grover, and Park teach the method of claim 1. Torkamani, Hamilton, Grover, and Park are combine in the same rational as set forth above with respect to claim 1 and analogous claim 16. Torkamani does not explicitly teach wherein each node of each distant neighborhood is no more than a predetermined number of hops away from the corresponding aggregator node. However Jin teaches wherein each node of each distant neighborhood is no more than a predetermined number of hops away from the corresponding aggregator node (Jin Page 3 2 Notations and Preliminaries Para 3 line 1-3, Graph Convolutional Network. Graph Convolutional Network (GCN) [14] is a variant of multilayer convolutional neural networks that operates directly on networks. It learns embedding of each node by iteratively aggregating the information from its neighbors ( i.e. each nodes is an aggregator node ). Page 3 3 Motivating Observations para 1 line 1-6, Here, we present a simple yet intuitive case study to illustrate and analyze the performance of GCN changes with different propagation mechanisms. The main idea is that we will apply GCN to networks with different structural properties utilizing three types of nodes: 1-hop, 2-hop and k-nearest neighbor (kNN) neighbors, which are often believed to be the effective neighborhoods for node classification in networks [28, 35], to realize the information propagation, respectively [ a predetermined number of hops away from the corresponding aggregator node ]) . Torkamani and Jin are considered to be analogous to the claim invention because they are in the same field of distributed machine learning using graph neural networks. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filling date of the claimed invention to have modified Hamilton in view of Jin to use nodes within a predetermined number of hops form an aggregator. Doing so to have a strong homophily and obtain high accuracy (Jin page 4 3 Motivating Observations para 4, As shown in Figure 1, for networks with strong homophily (e.g., ), it is easy to obtain pout = 0:075 high accuracy using 1-hop network neighbors. However, as the inter-class edges increase, the accuracy is rapidly reduced. This mainly due to the homophily assumption, preventing GCN from effectively fusing information. On the other hand, for networks with strong heterophily (e.g., ), it pout = 0:165 is surprising that, the accuracy of GCN of using 2-hop neighbors as neighborhoods (i.e., 83.15%) is much higher than that of using 1-hop network neighbors (i.e., 32.85%). Since the homophily ratio of 2-hop neighbors may rise with the increase of inter-class edges, GCN of using 2-hop neighbors is more effective to some extent. Interestingly, we can find that GCN of utilizing kNN is easy to get the staple accuracy, i.e., 71.46%. In particular, it is much higher than those of using 1-hop and 2-hop neighbors on complete random network (i.e., ). pout = 0:125 Summary. This case study shows that the current propagation mechanism of GCN is not universal for general network data, but we can find that there are rules in several special situations (i.e., complete random network). This motivates us that networks with different structural properties may need adopt different propagation mechanisms . 07-21-aia AIA Claim (s) 7, 8, 22 and 23 are rejected under 35 U.S.C. 103 as being unpatentable over Torkamani in view of Hamilton, Grover, and Park and further in view of Jianing Sun, Yingxue Zhang, Wei Guo, Huifeng Guo, Ruiming Tang, Xiuqiang He, Chen Ma, and Mark Coates. 2020. Neighbor Interaction Aware Graph Convolution Networks for Recommendation. In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR '20). Association for Computing Machinery, New York, NY, USA, 1289–1298 (“Sun”) . Regarding claim 7 and analogous claim 22, Torkamani in view of Hamilton, Grover, and Park teach the method of claim 1. Torkamani, Hamilton, Grover, and Park are combine in the same rational as set forth above with respect to claim 1 and analogous claim 16. Torkamani does not explicitly teach further comprising identifying a combination of two or more attributes of the seed node; and wherein the state of the seed node is based on the combination of two or more attributes of the seed node. However Sun teaches further comprising identifying a combination of two or more attributes of the seed node; and wherein the state of the seed node is based on the combination of two or more attributes of the seed node (Sun Page 1289, 1. Introduction para 4, Graph Convolutional (Neural) Networks (GCNs) [1, 12, 17, 25, 35, 38, 40, 41] have proven to be among the best performing architectures for a variety of graph learning tasks. GCNs strive to learn how to iteratively aggregate feature information from local graph neighborhoods using neural networks. The convolution operation in a GCN is a two-stage process consisting of a neighborhood aggregation stage and a center-neighbor combination stage [ further comprising identifying a combination of two or more attributes of the seed node ]. The neighborhood aggregation stage learns the representation of a neighborhood by transforming and aggregating the feature information from the neighborhood. The center-neighbor combination stage learns the representation of the central node by combining the representation of the neighborhood with the features of the central node [ and wherein the state of the seed node is based on the combination of two or more attributes of the seed node .]). Torkamani and Sun are considered to be analogous to the claim invention because they are in the same field of distributed machine learning using graph neural networks. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filling date of the claimed invention to have modified Torkamani in view of Sun to disclose learning features of the central node in combination with the representation of the neighborhood during a center-neighbor combination stage. Doing so to In order to better incorporate information from heterogeneous interaction types (Sun Page 1290 1 Introduction para 4, By building an explicit graph representation on the user-item interactions, the effectiveness of applying GCNs has been demonstrated in [3, 5, 21, 37, 40]. PinSAGE [40] constructs a bipartite graph on the pin-board structure in the Pinterest framework; Neural Graph Collaborative Filtering (NGCF) [37] builds a user-item bipartite graph structure and further emphasizes the interactions between the neighbors and the center node during the neighborhood aggregation process. [21] clusters the users into communities and make predictions in each community collaboratively. In order to better incorporate information from heterogeneous interaction types (search, guide, click, etc.) or interaction motives,) several graph-based solutions have been proposed in [3, 5, 24]. Regarding claim 8 and analogous claim 23, Torkamani in view of Hamilton, Grover, and Park teach the method of claim 1. Torkamani, Hamilton, Grover, and Park are combine in the same rational as set forth above with respect to claim 1 and analogous claim 16. Torkamani and Sun are combine in the same rational as set forth above with respect to claim 7 and analogous claim 22. Sun further teaches wherein each node of each of the distant neighborhoods neighborhood shares an entity type with the corresponding aggregator node, and wherein each aggregator node has a different entity type (Sun Page 1292-1293 4.2 Parallel-GCNs, To address the heterogeneity of the user-item interaction graph, we design a parallel graph convolution mechanism (Parallel-GCN). After forward sampling the required neighborhood sets up to depth 𝐾 (for computational efficiency, due to the very large sizes of the graphs, we do not process all nodes in the neighborhood), we learn central node embedding vectors by aggregating independently from each of the 𝐾 neighborhoods (one per depth). That is, instead of recursively updating the node embedding at layer-( 𝑘 −1) with the neighbors in layer- 𝑘 , we learn multiple central node embeddings from neighbors at depths 1, ..., 𝐾 directly, as shown in Figure 4. The output embeddings of Parallel-GCNs for users and items at) different layer are { h u 1 , … , h u k } and { h v 1 , … , h v k } for the users and items, respectively [ wherein each node of each of the distant neighborhoods shares an entity type with the corresponding aggregator node ]. Most existing GCN algorithms [12, 17, 40] employ the same aggregation and transformation function at each layer for every node. Instead of sharing aggregators for both user nodes and item nodes, we process them separately by learning two sets of aggregators for two different entities on the bipartite graph [ and wherein each aggregator node has a different entity type ]. PNG media_image15.png 339 645 media_image15.png Greyscale ) . Pertinent Prior Art 07-96 AIA The prior art made of record and not relied upon is considered pertinent to applicant's disclosure. Hamilton, William L., Rex Ying, and Jure Leskovec. "Representation learning on graphs: Methods and applications." arXiv preprint arXiv:1709.05584 (2017) – teaches a method of using deep walk to explore further way form the node to more effectively capture the community structures . Conclusion 07-40 AIA Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL . See MPEP § 706.07(a). Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a). A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action. Any inquiry concerning this communication or earlier communications from the examiner should be directed to ALFREDO CAMPOS whose telephone number is (571)272-4504. The examiner can normally be reached 7:00 - 4:00 pm M - F. 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, Michael J. Huntley can be reached at (303) 297-4307. 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. /ALFREDO CAMPOS/Examiner, Art Unit 2129 /MICHAEL J HUNTLEY/Supervisory Patent Examiner, Art Unit 2129 Application/Control Number: 18/014,287 Page 2 Art Unit: 2129 Application/Control Number: 18/014,287 Page 3 Art Unit: 2129 Application/Control Number: 18/014,287 Page 4 Art Unit: 2129 Application/Control Number: 18/014,287 Page 5 Art Unit: 2129 Application/Control Number: 18/014,287 Page 6 Art Unit: 2129 Application/Control Number: 18/014,287 Page 7 Art Unit: 2129 Application/Control Number: 18/014,287 Page 8 Art Unit: 2129 Application/Control Number: 18/014,287 Page 9 Art Unit: 2129 Application/Control Number: 18/014,287 Page 10 Art Unit: 2129 Application/Control Number: 18/014,287 Page 11 Art Unit: 2129 Application/Control Number: 18/014,287 Page 12 Art Unit: 2129 Application/Control Number: 18/014,287 Page 13 Art Unit: 2129 Application/Control Number: 18/014,287 Page 14 Art Unit: 2129 Application/Control Number: 18/014,287 Page 15 Art Unit: 2129 Application/Control Number: 18/014,287 Page 16 Art Unit: 2129 Application/Control Number: 18/014,287 Page 17 Art Unit: 2129 Application/Control Number: 18/014,287 Page 18 Art Unit: 2129 Application/Control Number: 18/014,287 Page 19 Art Unit: 2129 Application/Control Number: 18/014,287 Page 20 Art Unit: 2129 Application/Control Number: 18/014,287 Page 21 Art Unit: 2129 Application/Control Number: 18/014,287 Page 22 Art Unit: 2129 Application/Control Number: 18/014,287 Page 23 Art Unit: 2129 Application/Control Number: 18/014,287 Page 24 Art Unit: 2129 Application/Control Number: 18/014,287 Page 25 Art Unit: 2129 Application/Control Number: 18/014,287 Page 26 Art Unit: 2129 Application/Control Number: 18/014,287 Page 27 Art Unit: 2129 Application/Control Number: 18/014,287 Page 28 Art Unit: 2129 Application/Control Number: 18/014,287 Page 29 Art Unit: 2129 Application/Control Number: 18/014,287 Page 30 Art Unit: 2129 Application/Control Number: 18/014,287 Page 31 Art Unit: 2129 Application/Control Number: 18/014,287 Page 32 Art Unit: 2129
Prosecution Insights