DETAILED ACTION
Continued Examination Under 37 CFR 1.114
A request for continued examination under 37 CFR 1.114, including the fee set forth in 37 CFR
1.17(e), was filed in this application after final rejection. Since this application is eligible for continued
examination under 37 CFR 1.114, and the fee set forth in 37 CFR 1.17(e) has been timely paid, the
finality of the previous Office action has been withdrawn pursuant to 37 CFR 1.114. Applicant's
submission filed on 24 March 2026 has been entered.
Response to Amendment
The amendment filed on 13 February 2026 has been entered.
Claims 1-2, 4-12, 14-22 are pending.
Claims 1-2, 4, 11-12, 14 are amended.
Claims 21-22 are cancelled.
Claims 1-2, 4-12, 14-20 will be pending.
Response to Arguments
Applicant's arguments filed on 13 February 2026 have been fully considered, but they are not persuasive.
Applicant’s remarks, regarding the rejections of claims under 35 USC 103, have been fully considered.
Applicant notes Claim 1 has been amended to clarify "generating, for each node in the set of nodes, using a memory-augmented graph convolutional network (GCN)" and "the node attributes of the neighbour node and a defined set of relationship types stored in a relational memory, wherein the relational memory includes a latent relation matrix that includes a plurality of learned relationship types and a key matrix that includes a respective key value for each of the learned relationship types" (emphasis added). The Applicant submits that Rossi and/or Shang, whether taken alone or in combination, do not disclose any "relational memory network" including a "latent relation matrix" as recited in amended Claim 1.
Applicant notes Claim 1 has been amended to incorporate the subject matter of former Claim 21. Support for the amendment can be found at least in [0036] of the Application as originally filed. More specifically, Claim 1 now recites "wherein the relational memory is configured to restrict a learning space for learning each first edge attribute and each first node representation, based on the defined set of relationship types".
Applicant submits that the Rossi, Shang and/or Miller, whether taken alone or in combination, do not arrive at all the features of amended Claim 1.
Applicant’s arguments have been considered, but 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 the argument.
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Claim Rejections - 35 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.
This application currently names joint inventors. In considering patentability of the claims the examiner presumes that the subject matter of the various claims was commonly owned as of the effective filing date of the claimed invention(s) absent any evidence to the contrary. Applicant is advised of the obligation under 37 CFR 1.56 to point out the inventor and effective filing dates of each claim that was not commonly owned as of the effective filing date of the later invention in order for the examiner to consider the applicability of 35 U.S.C. 102(b)(2)(C) for any potential 35 U.S.C. 102(a)(2) prior art against the later invention.
Claims 1-2, 4, 6, 11-12, 14, 16 are rejected under 35 U.S.C. 103 as being unpatentable over Rossi et al. (U.S. Pre-Grant Publication No. 2020/0082265, hereinafter 'Rossi'), in view of Shang et al. (NPL: “GAMENet: Graph Augmented MEmory Networks for Recommending Medication Combination”, hereinafter ‘Shang’), and further in view of Ma et al. (NPL: “Memory Augmented Graph Neural Networks for Sequential Recommendation”, hereinafter ‘Ma’).
Regarding claim 1 and analogous claim 11, Rossi teaches A computer implemented method for processing a graph that includes a set of nodes and a set of edges, node in the set of nodes each having an associated set of node attributes, edge in the set of edges each representing a relationship that connects two respective nodes in the set of nodes, the method comprising ([0023] FIG. 3 is a graphical diagram showing example relational feature operators, in accordance with some embodiments. In one embodiment, processing logic derives a set of base graph features using the graph topology and attributes (if available). In one embodiment, “graph feature” refers to an edge or node feature and includes features derived by meshing the graph structure with attributes.; [0027] Conversely, learning an edge representation (DeepGL-edge) given G and an initial set of base node features (and optionally an initial set of attributes), processing logic may derive edge features by applying each relational operator Φk∈Φ to the nodes at either end of the edge. In another embodiment, each relational operator Φk∈Φ can be applied to the various combinations of in/out/total neighbors of each pair of nodes i and j that form an edge. When the input attributes match the type of graph element (node, edge) for which a feature representation is learned, then the attributes are simply appended to the feature matrix X.):
generating a first neighborhood vector representation that aggregates information from the generated first edge attributes and the node attributes of the plurality of neighbour nodes ([0044] Once the relational functions (definitions) are learned, they can be generating a first neighborhood vector representation extracted directly (no learning) to obtain the feature vectors for the graph elements (e.g., nodes/edges).; [0023] In one embodiment, “graph feature” refers to an edge or node feature and includes aggregates information from the generated first edge attributes and the node attributes of the plurality of neighbour nodes features derived by meshing the graph structure with attributes.); and
Rossi fails to teach generating, for each node in the set of nodes, using a memory-augmented graph convolutional network (GCN), a respective first node embedding by: generating, for the node and each of a plurality of neighbour nodes, a respective first edge attribute defining a respective relationship between the node and the neighbour node based on the node attributes of the node, the node attributes of the neighbour node and a defined set of relationship types stored in a relational memory, wherein the relational memory includes a latent relation matrix that includes a plurality of learned relationship types and a key matrix that includes a respective key value for each of the learned relationship types, wherein the relational memory is configured to restrict a learning space for learning each first edge attribute and each first node representation, based on the defined set of relationship types; generating, at a first layer of the memory-augmented GCN, a first node embedding of the node based on the node attributes of the node and the generated first neighborhood vector representation.
Shang teaches generating, for each node in the set of nodes, using a memory-augmented graph convolutional network (GCN), a respective first node embedding by ([Abstract, pg. 1126] To fill this gap, we propose the Graph Augmented Memory Net works (GAMENet), which integrates the drug-drug interac tions knowledge graph by a memory module implemented as a graph convolutional networks, and models longitudinal patient records as the query.; [Definition 2 (EHR&DDI Graph), pg. 1127] EHR graph and DDI graph can be denoted as Ge = {V,Ee} and Gd = {V,Ed} respectively, where node set V = Cm = {cm1 ,cm2 ,··· ,cmn } represents the set of medications, Ee is the edge set of known combination medication in EHR database and Ed is the edge set of known DDIs between a pair of drugs.; [Graph Augmented Memory Module, pg. 1128] Then, graph node embeddingsZ1,Z2∈R|Cm|×d are generated using GCN. Finally we combine different node embeddings as Memory Bank Mb∈R|Cm|×d where β is a weighting variable to fuse different knowledge graphs.):
generating, for the node and each of a plurality of neighbour nodes, a respective first edge attribute defining a respective relationship between the node and the neighbour node based on the node attributes of the node ([Graph Convolutional Networks (GCN), pg. 1127] emerged for inducing informative latent feature representations of nodes from arbitrary graphs (Kipf and Welling 2017; Defferrard, Bresson, and Vandergheynst 2016; Hamilton, Ying, and Leskovec 2017; Chen, Ma, and Xiao 2018). GCN models learn node embeddings in the following manner: Given each graph node initially attached with a a respective first edge attribute defining a respective relationship between the node and the neighbour node based on the node attributes of the node feature vector, the embedding vector of generating, for the node and each of a plurality of neighbour nodes each node are the transformed weighted sum of the feature vectors of its neighbors. All nodes are simultaneously updated to perform a layer of forward propagation. The deeper the network, the larger the local neighborhood. Thus global information is disseminated to each graph node for learning better node embeddings. GCNs haven been successfully used to model biomedical networks such as drug-drug interaction (DDI) graphs. For example, (Ma et al. 2018) models each drug as a node and DDIs as node labels in the drug association network and extended the GCN to embed multi-view drug features and edges. (Zitnik, Agrawal, and Leskovec 2018) used GCN to model the drug interaction problems by constructing a large two-layer multimodal drug interaction graphs. In this paper, we use GCN to model medication as nodes and DDIs as links.),
generating, at a first layer of the memory-augmented GCN, a first node embedding of the node based on the node attributes of the node and the generated first neighborhood vector representation ([Figure 1: The GAMENet:, pg. 1128] At current t th visit, the multi-hot input c t d , c t p are input into Embedding Networks to generate embedding e t d , e t p using Eq. 1. Then Dual-RNN generates current hidden states h t d , h t p by accepting both embeddings from Embeddings Network and longitudinal hidden state h t−1 of RNN denoted by return arrow described in Eq. 2. We use concatenated h t d , h t p as query q t (a.k.a. patient representation) in Eq. 3 to output o t b by reading from Memory Bank (MB) Mb in Eq. 7 generated from late-fusion based multiple knowledge graph in Eq.4, 5. Meantime, the the node attributes of the neighbour node and a defined set of relationship types stored in a memory network Dynamic Memory (DM) stores key-value form history information along time by Eq. 6 and can be used to generate o t d in Eq. 7. Finally, query and memory outputs are concatenated in Eq. 8 to make recommendation. In training phase, combined loss Eq. 10 is optimized to find optimal model parameters.; [Graph Augmented Memory Module, pg. 1129] Then we applied a two-layer GCN on each graph to learn improved embeddings on drug combination usage and DDIs respectively. The output Mb is generated as a weighted sum of the two graph embeddings. Z1 = A˜ etanh(A˜ eWe1)W1 Z2 = A˜ dtanh(A˜ dWe2)W2 Mb = Z1 + βZ2 (5) where We1, We2 ∈ R |Cm|×d are medication embeddings from EHR graph and DDI graph (each contains |Cm| number of d-dimensional vectors), W1, W2 ∈ R d×d are hidden weight parameter matrices. All W∗ are updated during training phase. Then, generating, at a first layer of the memory-augmented GCN, a first node embedding of the node based on the node attributes of the node and the generated first neighborhood vector representation graph node embeddings Z1, Z2 ∈ R |Cm|×d are generated using GCN. Finally we combine different node embeddings as Memory Bank Mb ∈ R |Cm|×d where β is a weighting variable to fuse different knowledge graphs. For Dynamic Memory (DM) Mt d , the combined patient {q t ′ }(t ′ < t) (the keys) associated with corresponding multi-hot medication vector {c t ′ m} (the values) are inserted into DM as key-value pairs. This kind of design provides a way to locate most similar patient representation over time and retrieve the proper weighted medications set. Specifically, we can incrementally insert key value pair after each visit step and treat Mt d as a vectorized indexable dictionary as follows: Mt d = {q t ′ : c t ′ m} t−1 1 (6) where Mt d is empty when t = 1. For clarity, we use Mt d,k = [q 1 ; q 2 ; · · · ; q t−1 ] ∈ R |t−1|×d to denote the key vectors and Mt d,v = [c 1 m; c 2 m; · · · ; c t−1 m ] ∈ R |t−1|×|Cm| to denote the value vectors at t th visit.).
Rossi and Shang are considered to be analogous to the claimed invention because they are in the same field of machine learning. In view of the teachings of Rossi, it would have been obvious for a person of ordinary skill in the art to apply the teachings of Shang to Rossi before the effective filing date of the claimed invention in order to integrate knowledge graph interactions by a memory module implemented as a graph convolutional networks, to provide safe and personalized recommendations (cf. Shang, [Abstract, pg. 1126] Recent progress in deep learning is revolutionizing the healthcare domain including providing solutions to medication recommendations, especially recommending medication combination for patients with complex health conditions. Existing approaches either do not customize based on patient health history, or ignore existing knowledge on drug-drug interactions (DDI) that might lead to adverse outcomes. To fill this gap, we propose the Graph Augmented Memory Networks (GAMENet), which integrates the drug-drug interactions knowledge graph by a memory module implemented as a graph convolutional networks, and models longitudinal patient records as the query. It is trained end-to-end to provide safe and personalized recommendation of medication combination. We demonstrate the effectiveness and safety of GAMENet by comparing with several state-of-the-art methods on real EHR data. GAMENet outperformed all baselines in all effectiveness measures, and also achieved 3.60% DDI rate reduction from existing EHR data.).
Ma teaches the node attributes of the neighbour node and a defined set of relationship types stored in a relational memory, wherein the relational memory includes a latent relation matrix that includes a plurality of learned relationship types and a key matrix that includes a respective key value for each of the learned relationship types, wherein the relational memory is configured to restrict a learning space for learning each first edge attribute and each first node representation, based on the defined set of relationship types ([Long-term Interest Modeling] To capture the long-term user interest, we can use the node attributes of the neighbour node and a defined set of relationship types stored in a relational memory external memory units (Sukhbaatar et al. 2015; Zhang et al. 2017) to store the time-evolving user interests given the user accessed items in Hu,l = (I1,I2,...,Il−1). However, maintaining the memory unit for each user has a huge memory overhead to store the parameters. Meanwhile, the memory unit may capture information that is very similar to that represented by the user embedding pu. Therefore, we propose to use a wherein the relational memory includes a latent relation matrix that includes a plurality of learned relationship types memory network to store the latent interest representation shared by all users, where each memory unit rep resents a certain type of latent user interest, such as the user interest regarding different categories of movies. Given the items accessed by a user in the past Hu,l, we can learn a combination of different types of interest to reflect the user long-term interest (or state) before Lu,l. Instead of performing a summing operation to generate the query as in the original memory network (Sukhbaatar et al. 2015), we wherein the relational memory is configured to restrict a learning space for learning each first edge attribute and each first node representation, based on the defined set of relationship types apply a multi-dimensional attention model to generate the query embedding. This allows discriminating informative items that can better reflect the user preference to have a greater influence on the positioning of the corresponding external memory units. Formally, we denote the item embeddings in Hu,l as Hu,l ∈ Rd×| Hu,l|. The multi dimensional attention to generate the query embedding zu,l is computed as: Hu,l := Hu,l +PE(Hu,l), Su,l = softmaxW(3) a tanh(W(1) a Hu,l +(W(2) Zu,l = tanh(Su,l · Hu,l), zu,l = avg(Zu,l), a pu)⊗1φ) , (3) where PE(·) is the sinusoidal positional encoding function that maps the item positions into position embeddings, which is the same as the one used in Transformer (Vaswani et al. 2017). φ equals to |Hu,l|, ⊗ denotes the outer product. W(1) a ,W(2) a ∈Rd×d and W(3) a ∈Rh×d are the learnable parameters in the attention model, and h is the hyper parameter to control the number of dimensions in the attention model. Su,l ∈ Rh×| Hu,l| is the attention score matrix. Zu,l ∈ Rh× d is the matrix representation of the query, and each of the h rows represents a different aspect of the query. Finally, zu,l ∈ Rd is the combined query embedding that averages the different aspects. Given the query embedding zu,l, we use this query to find the appropriate combination of the shared user latent interest in the memory network. Formally, the and a key matrix that includes a respective key value for each of the learned relationship types keys and values of the memory network (Sukhbaatar et al. 2015; Miller et al. 2016) are denoted as K ∈ Rd×m and V ∈ Rd×m, respectively, where m is the number of memory units in the memory network. Therefore, the user long-term interest embedding can be modeled as: si = softmax(zu,l · ki), ou,l = si vi , (4) i pH u,l = zu,l + ou,l , where ki,vi ∈ Rd are the i-th memory unit and the super script H denotes the representation is from the user long term interest.);
Rossi, Shang, and Ma are considered to be analogous to the claimed invention because they are in the same field of machine learning. In view of the teachings of Rossi and Shang, it would have been obvious for a person of ordinary skill in the art to apply the teachings of Ma to Rossi before the effective filing date of the claimed invention in order to capture both long- and short-term user interests, applying a graph neural network to model item contextual information within a short-term period, and utilizing a shared memory network to capture the long-range dependencies between items (cf. Ma, [Abstract] The chronological order of user-item interactions can reveal time-evolving and sequential user behaviors in many recommender systems. The items that users will interact with may depend on the items accessed in the past. However, the substantial increase of users and items makes sequential recommender systems still face non-trivial challenges: (1) the hardness of modeling the short-term user interests; (2) the difficulty of capturing the long-term user interests; (3) the effective modeling of item co-occurrence patterns. To tackle these challenges, we propose a memory augmented graph neural network (MA-GNN) to capture both the long- and short-term user interests. Specifically, we apply a graph neural network to model the item contextual information within a short-term period and utilize a shared memory network to capture the long-range dependencies between items. In addition to the modeling of user interests, we employ a bilinear function to capture the co-occurrence patterns of related items. We extensively evaluate our model on five real-world datasets, comparing with several state-of-the-art methods and using a variety of performance metrics. The experimental results demonstrate the effectiveness of our model for the task of Top-K sequential recommendation.).
Regarding claim 2 and analogous claim 12, Rossi, as modified by Shang and Ma, teaches The method of claim 1 and The processing device of claim 11, respectively.
Shang teaches generating, for each node in the set of nodes, using the memory-augmented GCN, a second node embedding by: generating, for the node and each of a plurality of neighbour nodes, a respective second edge attribute defining a respective relationship between the node and the neighbour node, based on the first node embedding of the node, a first node embedding of the neighbour node and a vector of weighted relationship types; generating a second neighborhood vector representation that aggregates information from the generated second edge attributes and the first node embeddings of the neighbour nodes; and generating, at a second layer of the memory-augmented GCN, the second node embedding based on the first node embedding of the node and the second generated neighborhood vector representation ([Graph Convolutional Networks (GCN), pg. 1127] emerged for inducing informative latent feature representations of nodes from arbitrary graphs (Kipf and Welling 2017; Defferrard, Bresson, and Vandergheynst 2016; Hamilton, Ying, and Leskovec 2017; Chen, Ma, and Xiao 2018). GCN models learn node embeddings in the following manner: Given generating, for the node and each of a plurality of neighbour nodes, a respective second edge attribute defining a respective relationship between the node and the neighbour node, based on the first node embedding of the node, a first node embedding of the neighbour node and a vector of weighted relationship types each graph node initially attached with a feature vector, the embedding vector of each node are the transformed weighted sum of the feature vectors of its neighbors. All nodes are simultaneously updated to perform a layer of forward propagation. The deeper the network, the larger the local neighborhood.; Figure 1: The GAMENet: At current t th visit, the generating a second neighborhood vector representation that aggregates information from the generated second edge attributes and the first node embeddings of the neighbour nodes multi-hot input c t d , c t p are input into Embedding Networks to generate embedding e t d , e t p using Eq. 1. Then Dual-RNN generates current hidden states h t d , h t p by accepting both embeddings from Embeddings Network and longitudinal hidden state h t−1 of RNN denoted by return arrow described in Eq. 2. We use concatenated h t d , h t p as query q t (a.k.a. patient representation) in Eq. 3 to output o t b by reading from Memory Bank (MB) Mb in Eq. 7 generated from late-fusion based multiple knowledge graph in Eq.4, 5. Meantime, the
generating, at a second layer of the memory-augmented GCN, the second node embedding based on the first node embedding of the node and the second generated neighborhood vector representation (Dynamic Memory (DM) stores key-value form history information along time by Eq. 6 and can be used to generate o t d in Eq. 7. Finally, query and memory outputs are concatenated in Eq. 8 to make recommendation. In training phase, combined loss Eq. 10 is optimized to find optimal model parameters.).
Rossi, Shang, and Ma are combinable for the same rationale as set forth above with respect to claim 1.
Regarding claim 4 and analogous claim 14, Rossi, as modified by Shang and Ma, teaches The method of claim 2 and The processing device of claim 12, respectively.
Ma teaches wherein generating each first node edge attribute and each second node edge attribute comprises: determining the vector of weighted relationship types from the defined set of relationship types stored in the relational memory network ([Long-term Interest Modeling] To capture the long-term user interest, we can use external memory units (Sukhbaatar et al. 2015; Zhang et al. 2017) to store the time-evolving user interests given the user accessed items in Hu,l = (I1,I2,...,Il−1). However, maintaining the memory unit for each user has a huge memory overhead to store the parameters. Meanwhile, the memory unit may capture information that is very similar to that represented by the user embedding pu. Therefore, we propose to use a memory network to store the latent interest representation shared by all users, where each memory unit rep resents a certain type of latent user interest, such as the user interest regarding different categories of movies. Given the items accessed by a user in the past Hu,l, we can learn a combination of different types of interest to reflect the user long-term interest (or state) before Lu,l. Instead of performing a summing operation to generate the query as in the original memory network (Sukhbaatar et al. 2015), we apply a multi-dimensional attention model to generate the query embedding. This allows discriminating informative items that can better reflect the user preference to have a greater influence on the positioning of the corresponding external memory units. Formally, we denote the item embeddings in Hu,l as Hu,l ∈ Rd×| Hu,l|. The multi dimensional attention to generate the query embedding zu,l is computed as: Hu,l := Hu,l +PE(Hu,l), Su,l = softmaxW(3) a tanh(W(1) a Hu,l +(W(2) Zu,l = tanh(Su,l · Hu,l), zu,l = avg(Zu,l), a pu)⊗1φ) , (3) where PE(·) is the sinusoidal positional encoding function that maps the item positions into position embeddings, which is the same as the one used in Transformer (Vaswani et al. 2017). φ equals to |Hu,l|, ⊗ denotes the outer product. W(1) a ,W(2) a ∈Rd×d and W(3) a ∈Rh×d are the learnable parameters in the attention model, and h is the hyper parameter to control the number of dimensions in the attention model. Su,l ∈ Rh×| Hu,l| is the attention score matrix. Zu,l ∈ Rh× d is the matrix representation of the query, and each of the h rows represents a different aspect of the query. Finally, zu,l ∈ Rd is the combined query embedding that averages the different aspects. Given the query embedding zu,l, we use this query to find the appropriate combination of the shared user latent interest in the memory network. Formally, the keys and values of the memory network (Sukhbaatar et al. 2015; Miller et al. 2016) are denoted as from the defined set of relationship types stored in the relational memory network K ∈ Rd×m and V ∈ Rd×m, respectively, where m is the number of memory units in the memory network. Therefore, the user long-term determining the vector of weighted relationship types interest embedding can be modeled as: si = softmax(zu,l · ki), ou,l = si vi , (4) i pH u,l = zu,l + ou,l , where ki,vi ∈ Rd are the i-th memory unit and the super script H denotes the representation is from the user long term interest.).
Rossi, Shang, and Ma are combinable for the same rationale as set forth above with respect to claim 1.
Regarding claim 6 and analogous claim 16, Rossi, as modified by Shang and Ma, teaches The method of claim 4 and The processing device of claim 14, respectively.
Shang teaches wherein, for each node and neighbour node: generating the respective first edge attribute for the node and the neighbour node comprises applying a first function to combine the node attributes of the node with the node attributes of the neighbour node based on learned parameters, the vector of weighted relationship types being determined based on the output of the first function; and generating the respective second edge attribute for the node and the neighbour node comprises applying the first function to combine the first node embedding of the node with the first node embedding of the neighbour node based on the learned parameters, the vector of weighted relationship types being determined based on the output of the first function ([Definition 2 (EHR&DDI Graph), pg. 1127] applying a first function to combine the node attributes of the node with the node attributes of the neighbour node based on learned parameters EHR graph and DDI graph can be denoted as Ge = {V, Ee} and Gd = {V, Ed} respectively, where node set V = Cm = {cm1 , cm2 , · · · , cmn } represents the set of medications, Ee is the edge set of known combination medication in EHR database and Ed is the edge set of known DDIs between a pair of drugs. Adjacency matrix Ae, Ad ∈ R |Cm|×|Cm| are defined to clarify the construction of edge Ee, Ed. For Ae, we firstly create a bipartite graph with drug on one side and drug combination on the other side. Then Ae = AbA | b where Ab ∈ R |Cm|×l is the adjacency matrix of the bipartite graph, Ab[i, j] = 1 when i th medication exists in j th medications combination and the number of unique medications combination denotes as l. For Ad, only pair-wise drug-drug interactions are considered, Ad[i, j] = 1 when the i th medication has interaction with the j th one.; [Problem 1 (Medication Combination Recommendation), pg. 1127] Given medical codes of the current visit at time t (excluding medication codes) c t d , c t p , patient history P = [x1, x2, · · · , xt−1] and EHR graph Ge, and DDI graph Gd, we want to recommend multiple medications by generating multi-label output yˆt ∈ {0, 1} |Cm|; [Graph Augmented Memory Module, pg. 1129] Then we applied a two-layer GCN on each graph to learn improved embeddings on drug combination usage and DDIs respectively. The generating the respective second edge attribute for the node and the neighbour node comprises applying the first function to combine the first node embedding of the node with the first node embedding of the neighbour node based on the learned parameters, the vector of weighted relationship types being determined based on the output of the first function output Mb is generated as a weighted sum of the two graph embeddings. Z1 = A˜ etanh(A˜ eWe1)W1 Z2 = A˜ dtanh(A˜ dWe2)W2 Mb = Z1 + βZ2 (5) where We1, We2 ∈ R |Cm|×d are medication embeddings from EHR graph and DDI graph (vector of weighted relationship types being determined based on the output of the first function each contains |Cm| number of d-dimensional vectors), W1, W2 ∈ R d×d are hidden weight parameter matrices. All W∗ are updated during training phase.).
Rossi, Shang, and Ma are combinable for the same rationale as set forth above with respect to claim 1.
Claims 5, 8-9, 15, 18-19 are rejected under 35 U.S.C. 103 as being unpatentable over Rossi, in view of Shang, Ma, and further in view of Miller et al. (NPL: “Key-Value Memory Networks for Directly Reading Documents”, hereinafter ‘Miller’).
Regarding claim 5 and analogous claim 15, Rossi, as modified by Shang and Ma, teaches The method of claim 4 and The processing device of claim 14, respectively.
Rossi, as modified by Shang and Ma, fails to teach wherein determining the vector of weighted relationship types comprises: determining a probability value for each of the respective key values and applying the determined probability values as weights to each of the learned relationship types.
Miller teaches wherein determining the vector of weighted relationship types comprises: determining a probability value for each of the respective key values and applying the determined probability values as weights to each of the learned relationship types ([Key Addressing:] during addressing, determining a probability value for each of the respective key values each candidate memory is assigned a relevance probability by comparing the question to each key: phi = Softmax(AΦX(x) · AΦK(khi )) where Φ· are feature maps of dimension D, A is a d× D matrix and Softmax(zi) = e zi/ P j e zj . We discuss choices of feature map in Sec. 3.2.; [Value Reading:] in the final reading step, the applying the determined probability values as weights to each of the learned relationship types values of the memories are read by taking their weighted sum using the addressing probabilities, and the vector o is returned: o = X i phiAΦV (vhi ).)
Rossi, Shang, Ma, and Miller are considered to be analogous to the claimed invention because they are in the same field of machine learning. In view of the teachings of Rossi, Shang, and Ma, it would have been obvious for a person of ordinary skill in the art to apply the teachings of Miller to Rossi before the effective filing date of the claimed invention in order to utilize different encodings in the addressing and output stages of the memory read operation, in comparing knowledge bases (cf. Miller, [Abstract] Directly reading documents and being able to answer questions from them is an unsolved challenge. To avoid its inherent difficulty, question answering (QA) has been directed towards using Knowledge Bases (KBs) instead, which has proven effective. Unfortunately KBs often suffer from being too restrictive, as the schema cannot support certain types of answers, and too sparse, e.g. Wikipedia contains much more information than Freebase. In this work we introduce a new method, Key-Value Memory Networks, that makes reading documents more viable by utilizing different encodings in the addressing and output stages of the memory read operation. To compare using KBs, information extraction or Wikipedia documents directly in a single framework we construct an analysis tool, WIKIMOVIES, a QA dataset that contains raw text alongside a preprocessed KB, in the domain of movies. Our method reduces the gap between all three settings. It also achieves state-of-the-art results on the existing WIKIQA benchmark.).
Regarding claim 8 and analogous claim 18, Rossi, as modified by Shang and Ma, teaches The method of claim 2 and The processing device of claim 12, respectively.
Rossi, as modified by Shang and Ma, fails to teach wherein, for each node in the set of nodes, generating the first neighborhood vector comprises: for each of the plurality of neighbour nodes, applying a second function to combine, based on a set of learned second function parameters, the attributes of the neighbour node with the respective first edge attribute for the node and the neighbour node, and aggregating the outputs of second function to generate the first neighborhood vector.
Miller teaches wherein, for each node in the set of nodes, generating the first neighborhood vector comprises: for each of the plurality of neighbour nodes, applying a second function to combine, based on a set of learned second function parameters, the attributes of the neighbour node with the respective first edge attribute for the node and the neighbour node, and aggregating the outputs of second function to generate the first neighborhood vector ([Key Hashing:] the question can be used to preselect a small subset of the possibly large array. This is done using an inverted index that finds a subset (kh1 , vh1 ), . . . ,(khN , vhN ) of memories of size N where the key shares at least one word with the question with frequency < F = 1000 (to ignore stop words), following Dodge et al. (2016). More sophisticated retrieval schemes could be used here, see e.g. Manning et al. (2008), [Key Addressing:] during addressing, each candidate memory is assigned a relevance probability by comparing the question to each key: phi = Softmax(AΦX(x) · AΦK(khi )) where Φ· are feature maps of dimension D, A is a d× D matrix and Softmax(zi) = e zi/ P j e zj . We discuss choices of feature map in Sec. 3.2.; [Value Reading:] in the final reading step, the values of the memories are read by taking their weighted sum using the addressing probabilities, and the vector o is returned: o = X i phiAΦV (vhi ). The memory access process is conducted by the “controller” neural network using q = AΦX(x) as the query. for each of the plurality of neighbour nodes, applying a second function to combine, based on a set of learned second function parameters, the attributes of the neighbour node with the respective first edge attribute for the node and the neighbour node After receiving the result o, the query is aggregating the outputs of second function to generate the first neighborhood vector updated with q2 = R1(q + o) where R is a d × d matrix. The memory access is then repeated (specifically, only the addressing and reading steps, but not the hashing), using a different matrix Rj on each hop, j. The key addressing equation is transformed accordingly to use the updated query: phi = Softmax(q > j+1AΦK(khi )) . The motivation for this is that new evidence can be combined into the query to focus on and retrieve more pertinent information in subsequent accesses. Finally, after a fixed number H hops, the resulting state of the controller is used to compute a final prediction over the possible outputs: aˆ = argmaxi=1,...,CSoftmax(q > H+1BΦY (yi)) where yi are the possible candidate outputs, e.g. all the entities in the KB, or all possible candidate answer sentences in the case of a dataset like WIKIQA (see Sec. 5.2). The d × D matrix B can also be constrained to be identical to A. The whole network is trained end-to-end, and the model learns to perform the iterative accesses to output the desired target a by minimizing a standard cross-entropy loss between aˆ and the correct answer a. Backpropagation and stochastic gradient descent are thus used to learn the matrices A, B and R1, . . . , RH.).
Rossi, Shang, Ma, and Miller are combinable for the same rationale as set forth above with respect to claim 5.
Regarding claim 9 and analogous claim 19, Rossi, as modified by Shang and Ma, teaches The method of claim 2 and The processing device of claim 12, respectively.
Rossi, as modified by Shang and Ma, fails to teach wherein, for each node in the set of nodes, generating the first node embedding comprises: applying a third function to combine, based on a set of learned third function parameters, the attributes of the node with the first neighborhood vector; and for each node, generating the second node embedding comprises: applying the third function to combine, based on a further set of learned third function parameters, the first embedding of the node with the second neighborhood vector.
Miller teaches wherein, for each node in the set of nodes, generating the first node embedding comprises: applying a third function to combine, based on a set of learned third function parameters, the attributes of the node with the first neighborhood vector; and for each node, generating the second node embedding comprises: applying the third function to combine, based on a further set of learned third function parameters, the first embedding of the node with the second neighborhood vector (([Key Hashing:] the question can be used to preselect a small subset of the possibly large array. This is done using an inverted index that finds a subset (kh1 , vh1 ), . . . ,(khN , vhN ) of memories of size N where the key shares at least one word with the question with frequency < F = 1000 (to ignore stop words), following Dodge et al. (2016). More sophisticated retrieval schemes could be used here, see e.g. Manning et al. (2008), [Key Addressing:] during addressing, each candidate memory is assigned a relevance probability by comparing the question to each key: phi = Softmax(AΦX(x) · AΦK(khi )) where Φ· are feature maps of dimension D, A is a d× D matrix and Softmax(zi) = e zi/ P j e zj . We discuss choices of feature map in Sec. 3.2.; [Value Reading:] in the final reading step, the values of the memories are read by taking their weighted sum using the addressing probabilities, and the vector o is returned: o = X i phiAΦV (vhi ). The memory access process is conducted by the “controller” neural network using q = AΦX(x) as the query. applying a third function to combine, based on a set of learned third function parameters, the attributes of the node with the first neighborhood vector After receiving the result o, the query is aggregating the outputs of second function to generate the first neighborhood vector updated with q2 = R1(q + o) where R is a d × d matrix. The memory access is then repeated (specifically, only the addressing and reading steps, but not the hashing), using a different matrix Rj on each hop, j. The key addressing equation is transformed accordingly to use the updated query: phi = Softmax(q > j+1AΦK(khi )) . The motivation for this is that new evidence can be combined into the query to focus on and retrieve more pertinent information in subsequent accesses. Finally, and for each node, generating the second node embedding comprises: applying the third function to combine, based on a further set of learned third function parameters, the first embedding of the node with the second neighborhood vector after a fixed number H hops, the resulting state of the controller is used to compute a final prediction over the possible outputs: aˆ = argmaxi=1,...,CSoftmax(q > H+1BΦY (yi)) where yi are the possible candidate outputs, e.g. all the entities in the KB, or all possible candidate answer sentences in the case of a dataset like WIKIQA (see Sec. 5.2). The d × D matrix B can also be constrained to be identical to A. The whole network is trained end-to-end, and the model learns to perform the iterative accesses to output the desired target a by minimizing a standard cross-entropy loss between aˆ and the correct answer a. Backpropagation and stochastic gradient descent are thus used to learn the matrices A, B and R1, . . . , RH.).
Rossi, Shang, Ma, and Miller are combinable for the same rationale as set forth above with respect to claim 5.
Claims 7, 17 are rejected under 35 U.S.C. 103 as being unpatentable over Rossi, in view of Shang, Ma, and further in view of Bhagat et al. (NPL: “NODE CLASSIFICATION IN SOCIAL NETWORKS”, hereinafter ‘Bhagat’).
Regarding claim 7 and analogous claim 17, Rossi, as modified by Shang and Ma, teaches The method of claim 2 and The processing device of claim 12, respectively.
Rossi, as modified by Shang and Ma, fails to teach wherein: generating the first node embedding for each node comprises determining the plurality of neighbour nodes for the node by sampling a fixed-size uniform draw of nodes that are within a predefined degree of relationship with the node based on the edges; and generating the second node embedding for each node comprises determining the plurality of neighbour nodes for the node by further sampling a fixed-size uniform draw of nodes that are within the predefined degree of relationship with the node based on the edges.
Bhagat teaches wherein: generating the first node embedding for each node comprises determining the plurality of neighbour nodes for the node by sampling a fixed-size uniform draw of nodes that are within a predefined degree of relationship with the node based on the edges ([2.1 Representing data as a graph, pg. 8] Given a node vi , let NN(vi) denote the set of k nearest neighbors, i.e. the k other nodes which are closest based on some measure of distance: this could be cosine or euclidean distance based on the features of the nodes. The k Nearest Neighbor (kNN) graph then has an edge between a pair of nodes vi, vj if vj ∈ NN(vi).); and
generating the second node embedding for each node comprises determining the plurality of neighbour nodes for the node by further sampling a fixed-size uniform draw of nodes that are within the predefined degree of relationship with the node based on the edges (As shown in the image below, Bhagat teaches generating the second node embedding for each node first embeds each vi into a p-dimensional initial node representation comprises determining the plurality of neighbour nodes for the node by further sampling a fixed-size uniform draw of nodes that are within the predefined degree of relationship with the node based on the edges kNNGNN is introduced, which implements kNN learning in an end-to-end manner.) .
PNG
media_image1.png
711
537
media_image1.png
Greyscale
Rossi, Shang, Ma, and Bhagat are considered to be analogous to the claimed invention because they are in the same field of machine learning. In view of the teachings of Rossi, Shang, and Ma, it would have been obvious for a person of ordinary skill in the art to apply the teachings of Bhagat to Rossi before the effective filing date of the claimed invention in order to induce a graph structure for objects based on homophily or co-citation (cf. Bhagat, [Inducing a graph., pg. 8] In some applications, the input may be a set of objects with no explicit link structure, for instance, a set of images from Flickr. We may choose to induce a graph structure for the objects, based on the principles of homophily or co-citation regularity: we should link entities which have similar characteristics (homophily) or which refer to the same objects (co-citation regularity).).
Claims 10, 20 are rejected under 35 U.S.C. 103 as being unpatentable over Rossi, in view of Shang, Ma, and further in view of Mermoud et al. (Pre-Grant Publication No. 2021/0303598, hereinafter ‘Mermoud’).
Regarding claim 10 and analogous claim 20, Rossi, as modified by Shang and Ma, teaches The method of claim 1 and The processing device of claim 11, respectively.
Rossi, as modified by Shang and Ma, fails to teach wherein the nodes represent transceiver devices in a wireless network and the node attributes include communication properties implemented at or measured at the respective transceiver devices, and the edges represent interactions through the wireless network between two transceiver devices.
Mermoud teaches wherein the nodes represent transceiver devices in a wireless network and the node attributes include communication properties implemented at or measured at the respective transceiver devices, and the edges represent interactions through the wireless network between two transceiver devices ([0037] Example machine learning techniques that device classification process 248 can employ may include, but are not limited to, nearest neighbor (NN) techniques (e.g., k-NN models, replicator NN models, etc.), statistical techniques (e.g., Bayesian networks, etc.), clustering techniques (e.g., k-means, mean-shift, etc.), neural networks (e.g., reservoir networks, artificial neural networks, etc.), support vector machines (SVMs), logistic or other regression, Markov models or chains, principal component analysis (PCA) (e.g., for linear models), multi-layer perceptron (MLP) artificial neural networks (ANNs) (e.g., for non-linear models), replicating reservoir networks (e.g., for non-linear models, typically for time series), random forest classification, or the like.; [0016] Sensor networks, a type of smart object network, are typically shared-media networks, such as wireless networks. That is, in addition to one or more sensors, each wherein the nodes represent transceiver devices in a wireless network sensor device (node) in a sensor network may generally be equipped with a radio transceiver or other communication port, a microcontroller, and an energy source, such as a battery.; [0077] FIG. 7A illustrates an example concept graph 700 that CLB 602 may form. As shown, concept graph 700 takes the form of a DAG comprising nodes 702 that node attributes include communication properties implemented at or measured at the respective transceiver devices, and the edges represent interactions through the wireless network between two transceiver devices are interconnected by directed edges 704. The direction of an edge 704 defines a relationship between nodes such that a given edge 704 is directed from a “parent” node 702 towards its “child” node 702.).
Rossi, Shang, Ma, and Mermoud are considered to be analogous to the claimed invention because they are in the same field of machine learning. In view of the teachings of Rossi, Shang, and Ma, it would have been obvious for a person of ordinary skill in the art to apply the teachings of Mermoud to Rossi before the effective filing date of the claimed invention in order to obtain data indicative of device attributes of a plurality of devices for device classification (cf. Mermoud, [0014] According to one or more embodiments of the disclosure, a device classification service obtains data indicative of device attributes of a plurality of devices. The device classification service forms, based on the obtained data indicative of the device attributes, a concept graph that comprises nodes that represent different sets of the device attributes. The device classification service determines, by analyzing the concept graph, a relevance score for each of the device attributes that quantifies how relevant that attribute is to classifying a device by its device type. The device classification service uses the relevance scores for the device attributes to cluster the plurality of devices into device type clusters by their device attributes.).
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to MAGGIE MAIDO whose telephone number is (703) 756-1953. The examiner can normally be reached M-Th: 6am - 4pm.
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 Huntley can be reached on (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.
/MM/Examiner, Art Unit 2129
/MICHAEL J HUNTLEY/Supervisory Patent Examiner, Art Unit 2129
Prosecution Insights