DETAILED ACTION Notice of Pre-AIA or AIA Status The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA. Priority Acknowledgment is made of A pplicant’s claim for foreign priority under 35 U.S.C. 119 (a)-(d). Receipt is acknowledged of certified copies of papers required by 37 CFR 1.55. Information Disclosure Statement The information disclosure statements submitted on July 21, 2023 and January 7, 2026 have been considered by the Examiner. Claim Rejections - 35 USC § 112 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 appl icant regards as his invention. Claims 4-8 , 12-16 and 20 are 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. In particular, claim 4 recites “wherein j 2 represents the one or more respective first nodes.” However, claim 4 already previously recites that “ j 1 represents the one or more respective first nodes,” and that “the one or more respective second nodes correspond to j 2 ∈ N - (i) .” It is therefore unclear as to whether the claimed first nodes are represented by j 1 or j 2. In claim 5, there is no antecedent basis for “the scaled transformed dual vector representations that are received by the node .” Claim 5 depends from claim 1, which recites “perform[ ing ] message passing using the scaled transformed dual vector representations.” However, there is no prior recitation of any node particularly receiving scaled transformed dual vector representations. Claims 6-8 depend from claim 5 and thereby include all of the limitations of claim 5. Accordingly, claims 6-8 are considered indefinite for the same reasons as claim 5 noted above. Like claim 4 noted above, claim 12 recites “wherein j 2 represents the one or more respective first nodes.” However, claim 12 already previously recites that “ j 1 represents the one or more respective first nodes,” and that “the one or more respective second nodes correspond to j 2 ∈ N - (i) .” It is therefore unclear as to whether the claimed first nodes are represented by j 1 or j 2. Like with claim 5 noted above, in claim 13 there is no antecedent basis for “the scaled transformed dual vector representations that are received by the node .” Claim 1 3 depends from claim 9 , which recites “perform ing, by the graph convolutional layer, message passing using the scaled transformed dual vector representations.” However, there is no prior recitation of any node particularly receiving scaled transformed dual vector representations. Claims 14-16 depend from claim 13, and thereby include all of the limitations of claim 13. Accordingly, claims 14-16 are considered indefinite for the same reasons as claim 13 noted above. Like claims 4 and 12 noted above, claim 20 recites “wherein j 2 represents the one or more respective first nodes.” However, claim 20 already previously recites that “ j 1 represents the one or more respective first nodes,” and that “the one or more respective second nodes correspond to j 2 ∈ N - (i) .” It is therefore unclear as to whether the claimed first nodes are represented by j 1 or j 2. 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. 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 , 5- 9 , 17 and 19 are rejected under 35 U.S.C. 103 as being unpatentable over the article entitled, “ Learning asymmetric embedding for attributed networks via convolutional neural network ” by Radmanesh et al. (“ Radmanesh ”), and also over the article entitled, “ Modeling Relational Data with Graph Convolutional Networks ” by Schlichtkrull et al. (“ Schlichtkrull ”). Regarding claim s 1 , 9 and 17 , Radmanesh describes a “ novel deep asymmetric attributed network embedding model based on the convolutional graph neural network ” (Abstract). Radmanesh teaches that this network embedding model utiliz es a graph convolutional layer that comprises a plurality of nodes and which, like claimed, is configured to: g enerate transformed dual vector representations by applying a source weight matrix and a target weight matrix to input dual vector representations of the plurality of nodes, wherein the input dual vector representations comprise, for each node of the plurality of nodes, a source vector representation that corresponds to the node in its role as a source and a target vector representation that corresponds to the node in its role as a target ( Radmanesh describes a directed graph that comprises a plurality of nodes and edges, wherein each edge represents a directed connection from a source node to a target node: Let’s define a directed attributed graph as G V,E,A,X,C where V= v 1 ,v 2 ,… v n and E= e i,j i,j=1 n specify the set of nodes with the size of n a nd the set of edges respectively. e i,j = v i , v j represents a directed connection from node v i to node v j . A∈ R n×n denotes the adjacency matrix where a i = a i,1 , a i,2 ,… a i,n represents the i -th row of A that includes non-negative weights of edges; a i,j =1 if there is a directed connection from v i to v j , and a i,j =0 otherwise. X∈ R n×F defines the attribute matrix where F is the number of attributes associated to each node . C specify the node label vector. (Section 2. “Problem Definition . ” Emphasis added. ). Radmanesh discloses that the model is applied to such a graph to produce an embedding vector for each node therein , wherein the embedding vector comprises concatenated source and target vectors : In this work we propose a novel deep asymmetric attributed network embedding model , Asymmetric Attributed Graph Convolutional Network (AAGCN), that preserves the asymmetric proximity and asymmetric similarity of directed attributed networks during the embedding process. AAGCN, as the name implies, is based on CGNNs that means it possesses all properties of neural network-based embedding models in addressing the non-linear structure of a network and the model scalability. Fig. 2 illustrates the framework of AAGCN. To preserve the asymmetric characteristics of directed attributed networks, first, the proposed method separately aggregates the features of a node with the features of its in- and out- neighbo u rs . The neighbourhood features aggregation schemes address the challenge of node distribution (inaccessibility between pairs of nodes) by enriching the node representations. Then, it learns two embedding vectors for each node, one source embedding vector and one target embedding vector. At the final step, the model provides an asymmetric representation for each node by concatenating the corresponding learned source and target embeddings . The source code is publicly available for future use. (Section 1 “Introduction . ” Emphasis added. Footnotes omitted. ). Definition 4: (Asymmetric network embedding): aims at representing the original high-dimensional network in a low-dimensional vector space so that asymmetric characteristics (asymmetric proximity and similarity) are preserved. Mathematically, given a directed attributed network G V,E,A,X,C , asymmetric network embedding learns the embedding vector space by a mapping function F :A∈ R n×n × X∈ R n×F →Z∈ R n×2 d , where d≪n. The embeddings are represented as Z= z 1 S . z 1 T , z 2 S . z 2 T ,…, z n S . z n T , where z i S ∈ R d and z i T ∈ R d define source embedding vector and target embedding vector of v i respectively. The embedding of v i , Z i = z i S . z i T ∈ R 2 d , is obtained by concatenating z i S and z i T . (Section 2. “Problem Definition . ” Emphasis added. ). To produce the embedding vector for each node in the directed graph , Radmanesh particularly discloses that the model can comprise a number of graph convolutional layers , wherein for each node in the graph, a first graph convolutional layer ( i ) generates a source representation by, in part, applying a source weight matrix to the features of the node and its out-neighbors, and (ii) generates a target representation by applying a target weight matrix to the features of the node and its in-neighbors: In a directed attributed network, we expect network nodes to be mapped close to each other based on not only their structural similarities but also their attribute similarities. Involving the node attributes during the embedding process, adds more informative information to the node representations that prevents the similar nodes with in- or out- degree of zero to be mapped faraway in the embedding space. Spatial CGNNs define a neighbourhood feature aggregation scheme for each node based on the pairwise local spatial relationships. An obvious solution for feature aggregation is to take the average value of the features of node v i along with its neighbours . However, in a directed graph, this approach disregards the direction of edges. In a directed graph, in- and out- neighbours of node v i are defined as follow: Γ in = j|(j,i)∈E 1 Γ out = j|(i,j)∈E 2 where Γ in and Γ out are in-degree and out-degree of node v i respectively. Here, to preserve the asymmetric nature of directed networks (i.e., asymmetric proximity and asymmetric similarity between nodes), we propose two distinct feature aggregation schemes used in source and target embedding processes. To this end, we simultaneously train two graph convolutions by passing them the adjacency matrix A and the attribute matrix X of the network. Considering the graph convolutions with only one hidden layer, in- and out- neighbour (s) feature aggregations of node v i are defined as follows: h i S = σ j∈ Γ out i ∪i X j W S 0 3 h i T = σ j∈ Γ in i ∪i X j W T 0 4 where W S 0 and W T 0 are the learnable weight matrix for the single hidden layer of source and target graph convolutions, respectively , and σ . is a non-linear activation function. h i S ∈ R d represents the output of source graph convolution that aggregates the features of node v i along with its out- neighbour (s). Similarly, h i T ∈ R d represents the output of target graph convolution that aggregates the features of node v i along with its in- neighbour (s) . Involving the features of a node during the feature aggregation process, addresses the challenge of distribution of nodes on directed networks. In detail, when the number of in- or out- neighbour (s) of node v i are low or even zero, considering the features of the node v i itself during learning representations helps the pair of nodes including node v i with similar features to be mapped close to each other in the embedding space, even when there is no explicit edge between them in the input space. The i -th row of A includes the out- neighbour (s) of node v i and its i -th column includes the in- neighbour (s) of node v i . Defining A =A+ I n as adjacency matrix of network with added self-connections ( I n is the identity matrix), Eqs . (3) and (4) can be factored as follows: H S = σ A X W S 0 5 H T = σ A T X W T 0 6 where H S ∈ R n×d and H T ∈ R n×d are source and target representations, respectively, and A T is the transpose of A . The final asymmetric representations are obtained by concatenating H S and H T : Z=concat H S , H T 7 We consider ReLU . = max .,0 as activation function of the hidden layer. (Section 4.1 “Asymmetric spatial graph convolution method . ” Emphasis added ). Accordingly, Radmanesh teaches implementing a graph convolutional layer that comprises a plurality of nodes, i.e. a node for each node of the directed graph, and that is configured to generate transformed dual vector representations, i.e. a source representation and a target representation , for each node by applying a source weight matrix and a target weight matrix to input dual vector representations of the node, e.g. to features of out- and in- neighbors of the node, wherein the input dual vector representations comprise, for the node, a source vector representation that corresponds to the node in its role as a source and a target vector representation that corresponds to the node in its role as a target.) ; and perform message passing using the transformed dual vector representations ( Radmanesh teaches that the model can comprise additional convolutional layers, wherein each subsequent layer ( i ) generates a source representation for each node using , in part, the representations generated in the previous layer for its out-neighbors, and (ii) generates a target representation for each node using, in part, the representations generated in the previous layer for its in-neighbors: Increasing the number of hidden layers (i.e., convolutional layers) to l , extracts l -th order similarity among nodes. This is because of weight sharing over the whole network. The general forms of feature aggregation schemes with l layer are : H S l = σ A H S l-1 W S l-1 10 H T l = σ A T H T l-1 W T l-1 11 where H S 0 = H T 0 =X . If we represent the output of last hidden layer of source and target graph convolutions as Z S and Z T , the overall asymmetric representations are represented as Z=concat( Z S , Z T ) . (Section 4.1 “Asymmetric spatial graph convolution method.”). Each node thus generates a source and target representation, which are used by neighbor nodes in the subsequent layer to generate further source and target representations. The provision of the representations, i.e. the scaled transformed dual vector representations, generated by each node in order to generate the source and target representations for its neighbor nodes in the subsequent convolutional layer is considered a form of message passing.). Radmanesh teaches that the above-described model (including the convolutional layers) can be implemented by a computing machine comprising one or more memories and a processor coupled to the one or more memories , and whereby the processor executes instructions stored in the one or more memories to carry out the above-noted tasks (see e.g. section 5.2 “Experimental setup”). Radmanesh thus teaches a computer-implemented autoencoding method similar to that of claim 9, which is implemented by a n autoencoder device (i.e. the computing machine) . The device implementing the above-described teachings of Radmanesh is considered a directed graph autoencoder device similar to that of claim 1 , and the memory necessary for storing the program instructions to implement the teachings is considered a non-transitory computer-readable medium similar to that of claim 17 . However, Radmanesh does not explicitly disclose that the transformed dual vector representations are scaled to generate scaled transformed dual vector representations, as is required by claim s 1 , 9 and 17 . Similar to Radmanesh , Schlichtkrull describes a graph convolutional layer that comprises a plurality of nodes and that is configured to apply weight matrices (e.g. W r (l) in equation (2)) to vector representations (e.g. h j (l) in equation (2)) to generate transformed vector representations (see e.g. section 2.1. “Relational graph convolutional networks”). Schlichtkrull further teaches scaling the transformed vector representations (e.g. using a “normalization constant,” c i,r in equation (2)) , and performing message passing using the scaled transformed vector representations (see e.g. section 2.1. “Relational graph convolutional networks”). It would have been obvious to one of ordinary skill in the art, having the teachings of Radmanesh and Schlichtkrull before the effective filing date of the claimed invention, to modify the graph convolutional layer taught by Radmanesh so as to further scale the transformed vector representations (i.e. scale the transformed dual vector representations) to generate scaled transformed (dual) vector representations, which are used in the subsequent message passing, like taught by Schlichtkrull . It would have been advantageous to one of ordinary skill to utilize such a combination because it would normalize the resulting transformed vector representations like taught by Schlichtkrull ( see e.g. section 2.1. “Relational graph convolutional networks”) and thus benefit training. Accordingly, Radmanesh and Schlichtkrull are considered to teach, to one of ordinary skill in the art, a directed graph autoencoder device like that of claim 1 , a computer-implemented autoencoding method like that of claim 9, and non-transitory computer-readable medium like that of claim 17 . As per claim 5, Radmanesh further teaches that the graph convolutional layer is further configured to generate aggregated dual vector representations by aggregating, for each node of the plurality of nodes, corresponding transformed vector representations of transformed dual vector representations that are received by the node, wherein the aggregated dual vector representations comprise a respective aggregated source vector representation for each node of the plurality of nodes and a respective aggregated target vector representation for each node of the plurality of nodes ( see e.g. Section 4.1 “Asymmetric spatial graph convolution method”: like noted above Radmanesh teaches that the model can comprise a plurality of convolutional layers, wherein each layer ( i ) generates a source representation for each node using, in part, the representations generated in the previous layer for its out-neighbors, and (ii) generates a target representation for each node using, in part, the representations generated in the previous layer for its in-neighbors. Each node in particular aggregates the source representations (as modified by a weight matrix) that are generated in the previous layer by the node’s out-neighbors to generate the source representation – see e.g. Section 4.1 “Asymmetric spatial graph convolution method.” Similarly, each node aggregates the target representations generated in the previous layer by the node’s in-neighbors to generate the target representation – see e.g. Section 4.1 “Asymmetric spatial graph convolution method.” Accordingly, the graph convolutional layer is considered to generate aggregated dual vector representations by aggregating, for each node of the plurality of nodes, corresponding transformed vector representations of the transformed dual vector representations that are received by the node, wherein the aggregated dual vector representations comprise a respective aggregated source vector representation for each node of the plurality of nodes and a respective aggregated target vector representation for each node of the plurality of nodes.). As described above, it would have been obvious to modify the graph convolutional layer taught by Radmanesh so as to further scale the transformed vector representations like taught by Schlichtkrull . Accordingly, the above-described combination of Radmanesh and Schlichtkrull is further considered to teach a directed graph autoencoder like that of claim 5. As per claim 6, it would have been obvious, as is described above, to modify the graph convolutional layer taught by Radmanesh so as to further scale the transformed vector representations (i.e. scale the transformed dual vector representations) to generate scaled transformed vector representations like taught by Schlichtkrull . Schlichtkrull particularly teaches that the graph convolutional layer is configured to generate layer output vector representations by, in part, scaling (e.g. using a “normalization constant,” c i,r in equation (2)) aggregated vector representations ( see e.g. section 2.1 “Relational graph convolutional networks”). Accordingly, the above-described combination of Radmanesh and Schlichtkrull is further considered to teach a directed graph autoencoder like that of claim 6. As per claim 7, Radmanesh further teaches applying an activation function (e.g. ReLU ) to layer output vector representations to generate encoder output vector representations (see e.g. section 4.1 “Asymmetric spatial graph convolution method”). Schlichtkrull provides a similar teaching (see e.g. section 2.1 “Relational graph convolutional networks”). Accordingly, the above-described combination of Radmanesh and Schlichtkrull is further considered to teach a directed graph autoencoder like that of claim 7. As per claim 8, Radmanesh does not explicitly teach implementing a decoder configured to determine updated weight matrices by decoding the encoder output vector representations, as is claimed. Schlichtkrull nevertheless provides such a teaching. Schlichtkrull describes an autoencoder that comprises an encoder comprising a graph convolutional layer that is configured to generate vector representations corresponding to a graph, and a decoder configured to reconstruct graph edges based on the vector representations (see e.g. section 4 “Link prediction”). Schlichtkrull suggests that the decoder is used to determine updated weight matrices (i.e. using a loss function) by decoding the encoder ’s output vector representations (see e.g. section 4 “Link prediction”). It would have been obvious to one of ordinary skill in the art, having the teachings of Radmanesh and Schlichtkrull before the effective filing date of the claimed invention, to further modify the device taught by Radmanesh and Schlichtkrull so as to utilize the graph convolutional layer in an autoencoder like taught by Schlichtkrull , which comprises a decoder configured to determine updated weight matrices by decoding encoder output vector representations. It would have been advantageous to one of ordinary skill to utilize such a combination because it can be effective in particular applications, e.g. link prediction, as is taught by Schlichtkrull (see e.g. section 4 “Link prediction”). Accordingly, Radmanesh and Schlichtkrull are further considered to teach, to one of ordinary skill in the art, a directed graph autoencoder device like that of claim 8. As per claim 19, Radmanesh teaches that the graph convolutional layer is configured to perform message passing at least in part by sending, by each node of the plurality of nodes, a respective pair of the transformed dual vector representations that correspond to the node (like noted above, Radmanesh teaches that the model can comprise multiple convolutional layers, wherein each subsequent convolutional layer ( i ) generates a source representation for each node using, in part, the source representations generated in the previous layer for its out-neighbors, and (ii) generates a target representation for each node using, in part, the target representations generated in the previous layer for its in-neighbors : Increasing the number of hidden layers (i.e., convolutional layers) to l , extracts l -th order similarity among nodes. This is because of weight sharing over the whole network. The general forms of feature aggregation schemes with l layer are : H S l = σ A H S l-1 W S l-1 10 H T l = σ A T H T l-1 W T l-1 11 where H S 0 = H T 0 =X . If we represent the output of last hidden layer of source and target graph convolutions as Z S and Z T , the overall asymmetric representations are represented as Z=concat( Z S , Z T ) . (Section 4.1 “Asymmetric spatial graph convolution method.”). Each node thus generates a source and target representation, which are used by neighbor nodes in a subsequent layer to generate further source and target representations. The provision of the source and target representations generated by each node in order to generate the source and target representations for its neighbor nodes in the subsequent convolutional layer is considered a form of message passing in which each node sends a respective pair of transformed representations that corresponds to the node.). As described above, it would have been obvious to modify the graph convolutional layer taught by Radmanesh so as to further scale the transformed vector representations like taught by Schlichtkrull . Schlichtkrull also teaches performing message passing by sending, by each node to neighboring nodes, the transformed vector representations that correspond to the node (see e.g. section 2.1 “Relational graph convolutional networks”). Accordingly, the above-described combination of Radmanesh , and Schlichtkrull is further considered to teach a non-transitory computer-readable medium like that of claim 19. Claims 2-4 , 10-1 6 , 18 and 20 are rejected under 35 U.S.C. 103 as being unpatentable over the combined teachings of Radmanesh and Schlichtkrull , as are described above, and also over U.S. Patent Application Publication No. 2020/0285944 to Lee et al. (“Lee”) . Regarding claim s 2 , 10 and 18 , Radmanesh and Schlichtkrull teach a directed graph autoencoder device like that of claim 1 , a computer-implemented autoencoding method like that of claim 9, and a non-transitory computer-readable medium like that of claim 17 , as is described above, and which implement a graph convolutional layer that comprises a plurality of nodes. As particularly note d above, it would have been obvious to modify the graph convolutional layer taught by Radmanesh so as to scale the transformed dual vector representations generated thereby, as is taught by Schlichtkrull . Schlichtkrull generally teaches scaling the vector representations by applying, for each node, a normalization constant ( e.g. c i,r in equation (2) ) that is based on edge degrees of the node , i.e. the number of neighbors of the node (see e.g. section 2.1. “Relational graph convolutional networks,” which recites “… where N i r denotes the set of neighbor indices of node i under relation r ∈ R . c i,r is a problem-specific normalization constant that can either be learned or chosen in advance (such as c i,r = N i r ).”). Because Radmanesh teaches that the source and target vector representation s of each node are respectively based on the out -neighbors and the in -neighbors of the node (see e.g. section 4.1 “ Asymmetric spatial graph convolution method”), it would have particularly been apparent to scale the source vector representation by a normalization constant that is based on the out degrees of the node and to scale the target vector representation by a normalization constant that is based on the in degrees of the node. Radmanesh and Schlichtkrull are thus further considered to teach configuring the graph convolution al layer to scale the transformed dual vector representations, by applying, for each node of the plurality of nodes, different parameters based on outdegrees and indegrees of the node. Radmanesh and Schlichtkrull , however, do not teach that the transformed dual vector representations are scaled by applying, for each node, different parameters to outdegrees and indegrees of the node, as is required by claim s 2 , 10 and 18 . Lee nevertheless describes a graph convolutional layer that comprises a plurality of nodes and that transform s vector representations by applying, for each of the plurality of nodes, different parameters (e.g. the degrees of neighbor nodes) to the degree of the node (see e.g. paragraphs 0066-0068). It would have been obvious to one of ordinary skill in the art, having the teachings of Radmanesh , Schlichtkrull and Lee before the effective filing date of the claimed invention, to modify the graph convolutional layer taught by Radmanesh and Schlichtkrull so as to scale the transformed vector representations by applying, for each of the plurality of nodes, different parameters to the degree s of the node , as is taught by Lee . Because, like noted above, Radmanesh teaches that the source vector representation is based on the out -neighbors of the node and the target vector representation is based on the in -neighbors of the node, it would have particularly been obvious to apply different parameters to the outdegrees and the indegrees of the node. It would have been advantageous to one of ordinary skill to utilize such a combination because it can help to prevent vanishing or exploding gradients during training, as is suggested by Lee (see e.g. paragraphs 0066-0068). Accordingly, Radmanesh , Schlichtkrull and Lee are considered to teach, to one of ordinary skill in the art, a directed graph autoencoder device like that of claim 2 , a computer-implemented autoencoding method like that of claim 10, and non-transitory computer-readable medium like that of claim 18 . As per claim s 3 and 11, Radmanesh teaches that the graph convolutional layer is configured to perform message passing at least in part by sending, by each node of the plurality of nodes, a respective pair of the transformed dual vector representations that correspond to the node ( like noted above, Radmanesh teaches that the model can comprise multiple convolutional layers, wherein each subsequent convolutional layer ( i ) generates a source representation for each node using, in part, the source representations generated in the previous layer for its out-neighbors, and (ii) generates a target representation for each node using, in part, the target representations generated in the previous layer for its in-neighbors : Increasing the number of hidden layers (i.e., convolutional layers) to l , extracts l -th order similarity among nodes. This is because of weight sharing over the whole network. The general forms of feature aggregation schemes with l layer are : H S l = σ A H S l-1 W S l-1 10 H T l = σ A T H T l-1 W T l-1 11 where H S 0 = H T 0 =X . If we represent the output of last hidden layer of source and target graph convolutions as Z S and Z T , the overall asymmetric representations are represented as Z=concat( Z S , Z T ) . (Section 4.1 “Asymmetric spatial graph convolution method.”). Each node thus generates a source and target representation, which are used by neighbor nodes in a subsequent layer to generate further source and target representations. The provision of the source and target representations generated by each node in order to generate the source and target representations for its neighbor nodes in the subsequent convolutional layer is considered a form of message passing in which each node sends a respective pair of transformed representations that corresponds to the node.). As described above, it would have been obvious to modify the graph convolutional layer taught by Radmanesh so as to further scale the transformed vector representations like taught by Schlichtkrull , and to particularly scale the transformed vector representations for each node by applying different parameters to the degrees of the node like taught by Lee. Schlichtkrull also teaches performing message passing by sending, by each node to neighboring nodes, the transformed vector representations that correspond to the node (see e.g. section 2.1 “Relational graph convolutional networks”), as does Lee (see e.g. paragraphs 0067-0068 and 0072). Accordingly, the above-described combination of Radmanesh , Schlichtkrull and Lee is further considered to teach a directed graph autoencoder device like that of claim 3 and a computer-implemented autoencoding method like that of claim 11 . As per claim s 4 , 12 and 20 , Radmanesh further teaches that the graph convolutional layer is configured to perform the message passing by sending, by each node of the plurality of nodes: ( 1 ) a transformed source vector representation of the respective pair of the transformed dual vector representations that corresponds to the node to one or more respective first nodes, wherein the one or more respective first nodes corresponds to j1∈ N + (i) , wherein j1 represents the one or more respective first nodes, N + (i) = N + (i) i , i represents the node, and N + (i) represents a first set of nodes of the plurality of nodes that the node points to; and (2) a transformed target vector representation of the respective pair of the transformed dual vector representations that corresponds to the node to one or more respective second nodes, wherein the one or more respective second nodes corresponds to j2∈ N - (i) , wherein j2 represents the one or more respective second nodes, N - (i) = N - (i) i , i represents the node, and N - (i) represents a second set of nodes of the plurality of nodes that point to the node ( see e.g. Section 4.1 “Asymmetric spatial graph convolution method”: like noted above, Radmanesh teaches that the model can comprise multiple convolutional layers, wherein each subsequent convolutional layer ( i ) generates a source representation for each node using, in part, the source representations generated in the previous layer for its out-neighbors, and (ii) generates a target representation for each node using, in part, the target representations generated in the previous layer for its in-neighbors . Each node thus generates a source and target representation, wherein the source representation is used by each of the in-neighbors of the node in the subsequent layer to generate a source representation for the neighbor node , and wherein the target representation is used by each of the out-neighbors of the node to generate a target representation for the neighbor node . T he node thus sends a first transformed representation, i.e. a source vector representation like claimed, to one or more respective first nodes, i.e. the out-neighbors of the node, which would correspond to j1∈ N + (i) , wherein j1 represents the one or more respective first nodes, N + (i) = N + (i) i , i represents the node, and N + (i) represents a first set of nodes of the plurality of nodes that the node points to. The node also sends a second transformed representation, i.e. a target representation like claimed, to one or more respective second nodes, i.e. the in-neighbors of the node, which would correspond to j2∈ N - (i) , wherein j2 represents the one or more respective second nodes, N - (i) = N - (i) i , i represents the node, and N - (i) represents a second set of nodes of the plurality of nodes that point to the node.). As described above, it would have been obvious to modify the graph convolutional layer taught by Radmanesh so as to further scale the transformed vector representations like taught by Schlichtkrull , and to particularly scale the transformed vector representations for each node by applying different parameters to the degrees of the node like taught by Lee. Accordingly, the above-described combination of Radmanesh , Schlichtkrull and Lee is further considered to teach a directed graph autoencoder device like that of claim 4 , a computer-implemented autoencoding method like that of claim 12, and a non-transitory computer-readable medium like that of claim 20 . As per claim 13, Radmanesh further teaches that the graph convolutional layer is further configured to generate aggregated dual vector representations by aggregating, for each node of the plurality of nodes, corresponding transformed vector representations of transformed dual vector representations that are received by the node, wherein the aggregated dual vector representations comprise a respective aggregated source vector representation for each node of the plurality of nodes and a respective aggregated target vector representation for each node of the plurality of nodes ( see e.g. Section 4.1 “Asymmetric spatial graph convolution method”: like noted above Radmanesh teaches that the model can comprise a plurality of convolutional layers, wherein each layer ( i ) generates a source representation for each node using, in part, the representations generated in the previous layer for its out-neighbors, and (ii) generates a target representation for each node using, in part, the representations generated in the previous layer for its in-neighbors. Each node in particular aggregates the source representations (as modified by a weight matrix) that are generated in the previous layer by the node’s out-neighbors to generate the source representation – see e.g. Section 4.1 “Asymmetric spatial graph convolution method.” Similarly, each node aggregates the target representations generated in the previous layer by the node’s in-neighbors to generate the target representation – see e.g. Section 4.1 “Asymmetric spatial graph convolution method.” Accordingly, the graph convolutional layer is considered to generate aggregated dual vector representations by aggregating, for each node of the plurality of nodes, corresponding transformed vector representations of the transformed dual vector representations that are received by the node, wherein the aggregated dual vector representations comprise a respective aggregated source vector representation for each node of the plurality of nodes and a respective aggregated target vector representation for each node of the plurality of nodes.). As described above, it would have been obvious to modify the graph convolutional layer taught by Radmanesh so as to further scale the transformed vector representations like taught by Schlichtkrull , and to particularly scale the transformed vector representations for each node by applying different parameters to the degrees of the node like taught by Lee. Accordingly, the above-described combination of Radmanesh , Schlichtkrull and Lee is further considered to teach a computer-implemented method like that of claim 13. As per claim 14, it would have been obvious, as is described above, to modify the graph convolutional layer taught by Radmanesh so as to further scale the transformed vector representations (i.e. scale the transformed dual vector representations) to generate scaled transformed vector representations like taught by Schlichtkrull . Schlichtkrull particularly teaches that the graph convolutional layer is configured to generate layer output vector representations by, in part, scaling (e.g. using a “normalization constant,” c i,r in equation (2)) aggregated vector representations ( see e.g. section 2.1 “Relational graph convolutional networks”). Accordingly, the above-described combination of Radmanesh , Schlichtkrull and Lee is further considered to teach a computer-implemented autoencoding method like that of claim 14. As per claim 15 , Radmanesh further teaches applying an activation function (e.g. ReLU ) to layer output vector representations to generate encoder output vector representations (see e.g. section 4.1 “Asymmetric spatial graph convolution method”). Schlichtkrull provides a similar teaching (see e.g. section 2.1 “Relational graph convolutional networks”). Accordingly, the above-described combination of Radmanesh , Schlichtkrull and Lee is further considered to teach a computer-implemented autoencoding method like that of claim 15 . As per claim 16 , Radmanesh does not explicitly teach decoding the encoder output vector representations to determine updated weight matrices, as is claimed. Schlichtkrull nevertheless provides such a teaching. Schlichtkrull describes an autoencoder that comprises an encoder comprising a graph convolutional layer that is configured to generate vector representations corresponding to a graph, and a decoder configured to reconstruct graph edges based on the vector representations (see e.g. section 4 “Link prediction”). Schlichtkrull suggests that the decoder is used to determine updated weight matrices (i.e. using a loss function) by decoding the encoder’s output vector representations (see e.g. section 4 “Link prediction”). It would have been obvious to one of ordinary skill in the art, having the teachings of Radmanesh , Schlichtkrull and Lee before the effective filing date of the claimed invention, to further modify method taught by Radmanesh , Schlichtkrull and Lee so as to utilize the graph convolutional layer in an autoencoder like taught by Schlichtkrull , which comprises a decoder configured to determine updated weight matrices by decoding encoder output vector representations. It would have been advantageous to one of ordinary skill to utilize such a combination because it can be effective in particular applications, e.g. link prediction, as is taught by Schlichtkrull (see e.g. section 4 “Link prediction”). Accordingly, Radmanesh , Schlichtkrull and Lee are further considered to teach, to one of ordinary skill in the art, a computer-implemented autoencoding method like that of claim 16 . Conclusion The prior art made of record on form PTO-892 and not relied upon is considered pertinent to applicant’s disclosure. The applicant is required under 37 C.F.R. §1.111(C) to consider these references fully when responding to this action. In particular, the Chinese Patent document to Zhang cited therein describes a directed graph node updating method that includes obtaining a source characteristic and target characteristic of each of a plurality of nodes, and updating the target characteristic of each node according to the source characteristic s of adjacent no d es, and updating the source characteristic of each node according to the target characteristics of adjacent nodes. The article by Wang et al. cited therein (“ TGAE: Temporal Graph Autoencoder for Travel Forecasting ”) describes “ Temporal Graph Autoencoder ,” which encodes the fundamentally asymmetric nature of a directed graph via directed neighborhood aggregation and learns a pair of vector representations for each node. The article by Ma et al. cited therein (“ Graph autoencoder for directed weighted network ”) describes a graph autoencoder-based network representation model for directed and weighted network embedding. The article by Virinchi et al. cited therein (“ Recommending Related Products Using Graph Neural Networks in Directed Graphs ”) describes a Graph Neural Network (GNN) based framework for related product recommendation that learn s dual embeddings for each of a plurality of product s by appropriately aggregating features from its neighborhood . Any inquiry concerning this communication or earlier communications from the examiner should be directed to FILLIN "Examiner name" \* MERGEFORMAT BLAINE T BASOM whose telephone number is FILLIN "Phone number" \* MERGEFORMAT (571)272-4044 . The examiner can normally be reached FILLIN "Work Schedule?" \* MERGEFORMAT Monday-Friday, 9:00 am - 5:30 pm, EST . Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice. If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Matt Ell can be reached at FILLIN "SPE Phone?" \* MERGEFORMAT (571)270- 3264 . The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300. Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000. /BTB/ 3/20/2026 /MATTHEW ELL/ Supervisory Patent Examiner, Art Unit 2141