Paper recommendation method and system based on hierarchical graph attribute network representation learning

Abstract

The application provides a paper recommendation method and system based on hierarchical graph attribute network representation learning, and the method comprises the following steps: constructing a paper network based on a paper library; constructing a hierarchical graph through edge merging, structural equivalence merging, community merging and attribute merging based on the paper network; performing intra-layer network representation learning and inter-layer network representation learning on the hierarchical graph to generate paper node embedding representation, wherein the inter-layer network representation learning is performed through forward initialization and then backward initialization; and generating a paper recommendation list by calculating the similarity of embedding representation between paper nodes based on the paper node embedding representation, and recommending the papers in the paper recommendation list to users. The paper recommendation method based on hierarchical graph attribute network representation learning is adopted, so that the problem of low paper recommendation accuracy of the existing network representation learning method is solved from the root.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of personalized recommendation, and particularly relates to a paper recommendation method and system based on hierarchical graph attribute network representation learning. BACKGROUND

[0002] With the rapid development of the Internet and the rapid update of knowledge, the number of scientific papers published on the network has increased explosively, and the overload of scientific research information has caused researchers to spend a lot of time and effort to find the paper information they need, which indirectly causes waste of scientific research. Accurate paper recommendation to researchers is becoming an indispensable part of scientific research. Most of the algorithms related to paper recommendation are based on network representation learning. The paper recommendation algorithm based on network representation learning usually abstracts the papers in the paper database as nodes, the citation relationship between the papers as edges, and the title, abstract, keywords, authors, institutions, publication time, etc. of the papers as attribute information of the nodes. The network representation learning method is used to represent the nodes in the network as low-dimensional embedding representation, which is used for paper recommendation.

[0003] However, most of the current research on network representation learning is based on network structure, such as using the walk sequence generated by the network nodes, the first-order proximity, the second-order proximity, and the high-order proximity of the network structure for network representation learning and has achieved certain success. However, this kind of method does not consider the attribute information of the paper nodes, and it is difficult to accurately recommend papers. At the same time, the paper database generates a large amount of structural data in the paper network, and the network also contains various types of heterogeneous attribute information. The data amount is large, and the network structure is different in density and adjacency information. It is very difficult to balance the relationship between the topology structure and the attribute information in the network representation, obtain the embedding representation of each paper node in the network, and recommend papers.

[0004] Therefore, how to improve the accuracy of paper recommendation and improve user satisfaction is a problem to be solved in the field. SUMMARY

[0005] The purpose of the present application is to provide a paper recommendation method and system based on hierarchical graph attribute network representation learning to solve the problems of low accuracy of paper recommendation and low user satisfaction of the existing paper recommendation method based on network representation learning. The paper recommendation method based on hierarchical graph attribute network representation learning constructs a paper network according to a paper library, constructs a hierarchical graph through structure merging and attribute merging based on the paper network, performs intra-layer network representation learning and inter-layer network representation learning on the hierarchical graph, generates node embedding representation, generates a recommendation list by calculating the similarity of the embedding representation between the paper nodes, and recommends the papers in the recommendation list to the user, thereby solving the problems of low accuracy of paper recommendation and low user satisfaction of the existing paper recommendation method based on network representation learning from the root.

[0006] To achieve the above object, the present application adopts the following technical solutions:

[0007] A paper recommendation method based on hierarchical graph attribute network representation learning, characterized by comprising the steps of:

[0008] S1, constructing a paper network based on a paper library, wherein the nodes in the paper network represent papers, the edges represent the citation relationship between papers, and the attribute information of the nodes is the information of the papers, including title, abstract, keywords, authors, institutions, and publication time;

[0009] S2, based on the paper network, constructing a hierarchical graph G 0 , G 1 ,..., G H by structure merging and attribute merging;

[0010] S3, performing network representation learning on the hierarchical graph to generate paper node embedding representation, comprising:

[0011] S31, performing intra-layer network representation learning on the hierarchical graph;

[0012] S32, performing inter-layer network representation learning on the hierarchical graph, specifically comprising:

[0013] 1) Perform forward initialization: from layer 1 to layer H, the result embedding representation of multiple paper nodes in G i-1 perform Max-Pooling to initialize the initial embedding representation of the super node u in G i

[0014]

[0015] wherein, represents the result embedding representation of the i-1 layer paper node v in the forward initialization process, represents the initial embedding representation of the i layer paper node u in the forward initialization process, and the initial node embedding representation of the intra-layer representation learning is randomly initialized in the original paper network graph G 0

[0016] 2) Perform backward initialization: from layer H-1 to layer 0, given the newly learned result embedding of the paper node u in G i+1 and the result embedding generated in the forward initialization of the original node v constituting the super node u in G i ​​​​​performing the Mean-Pooling to obtain a new initialized embedding of the paper node v

[0017]

[0018] wherein, represents the result embedding representation of the paper node u in the i+1 layer in the backward initialization process, represents the initial embedding representation of the paper node v in the i layer in the backward initialization process, and on the H layer, the initialized embedding representation of the paper node v in the backward initialization process is equal to the result embedding representation of the paper node v in the forward initialization process

[0019] S33, the backward initialization network representations in each layer of the paper node are spliced in a cascaded manner to generate a paper node embedding representation, specifically:

[0020]

[0021] wherein, represents the backward initialization network representation of the paper node v in the i layer network, x v represents the embedding representation of the paper node v;

[0022] S4, according to the paper node embedding representation, a paper recommendation list is generated by calculating the similarity of the embedding representations between the paper nodes, and the paper recommendation list is recommended to the user.

[0023] Further, in step S2, the construction process of the hierarchical graph is specifically: according to the structural characteristics of the paper network, similar nodes are merged to realize structural merging by edge merging, structural equivalence merging and community merging, and the attribute information of the merged super node is calculated to realize attribute merging, and the two merging results are combined. This process is iterated until the number of nodes of the hierarchical graph reaches a set threshold to obtain a small enough network, and the construction of the hierarchical graph is completed.

[0024] Further, the edge merging is to select pairs of adjacent vertices in the paper network, and each pair of vertices is merged into a new vertex; the structural equivalence merging is to merge nodes sharing common neighbors in the paper network into a new super node; and the community merging is to merge nodes and edges within a community into a new node.

[0025] Further, the community merging adopts a Louvain algorithm to divide the communities by maximizing modularity, each paper node in the paper network is initialized as a community, if two paper nodes are merged into a community, the modularity can be improved, then the merging is performed until the merging no longer improves the modularity; wherein, the modularity is used to measure the density of edges within the community and between the communities, and the calculation formula of the modularity is:

[0026]

[0027] wherein, Q represents the modularity of the paper network, m is the weight sum of all edges in the paper network, A ij represents the edge weight between the paper node v i and v j , k i , k j represents the degree of the paper node v i and v j , c i and c j are the communities where the paper node v i and v j are located; when the paper node v i and v j are in the same community, δ is 1.

[0028] Further, a binary mapping matrix is obtained by edge merging, structure equivalent merging and community merging to save the merging result, wherein, V i represents the paper node set of the i-th hierarchical graph G i , the matrix retains the paper node mapping relationship of the hierarchical graph G i to G i+1 ; the value of the matrix is 1 if the paper node r of G i is merged into the supernode j of G i+1 , otherwise 0.

[0029] Further, in the hierarchical graph construction process, the attribute information of the supernode generated by each structure merging is obtained from the previously merged paper nodes through a set operation at the same time of structure merging, and the attribute of the supernode is the set of the attributes of the original paper node set before merging.

[0030] Further, in step S31, the intra-layer network representation learning of the hierarchical graph includes: for the i-th hierarchical graph G i and the initialized network representation S i of the hierarchical graph, using a TADW or CANE attribute network representation learning method to learn the network representation of the layer network, wherein, 0≤i≤H.

[0031] Further, step S4 specifically comprises: for the papers v i , calculating the similarity with other paper nodes in the paper network by using the embedding representation of the paper node, and the calculation formula is as follows:

[0032]

[0033] wherein x i represents the embedding representation of the paper node v i , represents the dot product operation, |x i | represents the modulus of the embedding representation of the paper node v i ; Sim i，j represents the similarity of the paper node v i and the paper node v j ;

[0034] By comparing the method with each paper in the paper library, the similarity vector sIM i of the final paper node v i and other papers in the paper network is obtained.

[0035] Select the papers v1, v2, v3,..., v n read in the time interval t, respectively calculate the similarity vector SIM1, SIM2, SIM3,..., SIM n of the paper node and other papers in the paper library, and sum the similarity vectors SIM1, SIM2, SIM3,..., SIM n to obtain the similarity of the set of papers read by the current user in the time interval t, and finally arrange the similarity in descending order to select top-k papers to generate a paper recommendation list, and recommend the paper recommendation list to the user.

[0036] A paper recommendation system based on hierarchical graph attribute network representation learning, comprising: a paper network construction module, a hierarchical graph construction module, a hierarchical network representation learning module, and a paper recommendation module;

[0037] The paper network construction module constructs a paper network based on a paper library, wherein the nodes in the paper network represent papers, the edges represent the citation relationship between papers, and the attribute information of the nodes is the information of the papers, and the information of the papers includes: title, abstract, keywords, authors, institutions, and publication time.

[0038] The hierarchical graph construction module constructs hierarchical graphs G 0 , G 1 ,..., and G H based on the paper network through structure merging and attribute merging, wherein the structure merging includes edge merging, structure equivalence merging, and community merging.

[0039] a hierarchical network representation learning module for performing network representation learning on the hierarchical graph to generate paper node embedding representations, including:

[0040] an intra-layer network representation learning module for performing intra-layer network representation learning on the hierarchical graph;

[0041] an inter-layer network representation learning module for performing inter-layer network representation learning on the hierarchical graph, specifically including:

[0042] 1) performing forward initialization: from layer 1 to layer H, the result embedding representation of the multiple paper nodes in G i-1 is initialized as performing Max-Pooling to initialize the initial embedding representation of the hypernode u in G i

[0043]

[0044] wherein, represents the result embedding representation of the paper node v in the i-1 layer in the forward initialization process, represents the initial embedding representation of the paper node u in the i layer in the forward initialization process, and in the original paper network graph G 0 , the initial node embedding representation of the intra-layer representation learning is randomly initialized;

[0045] 2) performing backward initialization: from layer H-1 to layer 0, given the result embedding of the paper node u newly learned in G i+1 and the result embedding of the original node v constituting the hypernode u in G i generated in the forward initialization, performing Mean-Pooling to obtain the new initial embedding of the paper node v.

[0046]

[0047] wherein, represents the result embedding representation of the paper node u in the i+1 layer in the backward initialization process, represents the initial embedding representation of the paper node v in the i layer in the backward initialization process, and on the H layer, the initial embedding representation of the paper node v in the backward initialization process is set to be equal to the result embedding representation of the paper node v in the forward initialization process.

[0048] ​The paper node embedding representation generation module adopts a cascading manner to splice the reverse initialization network representation in each level of the paper node to generate a paper node embedding representation, specifically:

[0049]

[0050] wherein, represents the reverse initialization network representation of the paper node v in the i-th layer network, x v represents the embedding representation of the paper node v.

[0051] The paper recommendation module generates a paper recommendation list by calculating the similarity of the embedding representations between the paper nodes according to the paper node embedding representation, and recommends the paper recommendation list to the user. The present application also provides a computer device, which comprises a memory and a processor, the memory stores a computer program, and the processor executes the computer program to realize the above method.

[0052] The present application discloses a paper recommendation method and system based on hierarchical graph attribute network representation learning. The paper recommendation method based on hierarchical graph attribute network representation learning constructs a paper network according to a paper library, constructs a hierarchical graph through structure merging and attribute merging based on the paper network, performs intra-layer network representation learning and inter-layer network representation learning on the hierarchical graph, generates node embedding representation, generates a recommendation list by calculating the similarity of the embedding representations between the paper nodes, and recommends the papers in the recommendation list to the user, thereby fundamentally solving the problems of low accuracy and low user satisfaction of the existing paper recommendation method based on network representation learning.

[0053] Compared with the prior art, the present application has the following beneficial effects:

[0054] 1. The paper recommendation method based on hierarchical graph attribute network representation learning of the present application constructs a hierarchical graph on the paper network through structure merging and attribute merging, realizes hierarchical network representation learning on the hierarchical graph, and obtains the final paper node embedding representation containing the structural features and attribute information of the original network of multiple orders through the cascade of inter-layer network representation learning, thereby enhancing the existing paper network representation learning method, improving the accuracy of paper recommendation, and further improving the user experience.

[0055] 2.The paper recommendation method based on hierarchical graph attribute network representation learning provided by the present application, in the process of constructing the hierarchical graph, through edge merging, structure equivalent merging and community merging, the first-order, second-order proximity and community characteristics of the network are maintained respectively, thereby maintaining the high-order proximity of the network, and the attribute information of the nodes is synchronously merged, the topological structure and attribute information of the network are integrated, a hybrid hierarchical graph construction method is proposed to maintain the multi-order network attribute information, and the accuracy of the paper recommendation is improved, and the user experience is improved.

[0056] 3.The paper recommendation method based on hierarchical graph attribute network representation learning provided by the present application, in the interlayer network representation learning, the forward initialization is performed by using the Max-Pooling method, the feature information with the greatest influence is input to the next layer, and the propagation of useless information is reduced; the reverse initialization is performed by using the Mean-Pooling method to reduce the error, and the reverse initialization network representations in each layer of the nodes are spliced in a cascading manner, the paper network representation learning method is optimized, and the accuracy of the paper recommendation is improved, and the user experience is improved. BRIEF DESCRIPTION OF DRAWINGS

[0057] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0058] Figure 1 The paper recommendation method based on hierarchical graph attribute network representation learning provided by the present application is shown in the schematic diagram.

[0059] Figure 2 The paper recommendation method based on hierarchical graph attribute network representation learning provided by the present application is shown in the schematic diagram.

[0060] Figure 3 The edge merging and structure equivalent merging provided by the present application are shown in the schematic diagram.

[0061] Figure 4 The paper recommendation system based on hierarchical graph attribute network representation learning provided by the present application is shown in the schematic diagram. DETAILED DESCRIPTION

[0062] With reference to the drawings and specific embodiments described below, the present application will be further illustrated. However, the present application is not limited to the embodiments described below.

[0063] It should be understood that the terms "comprise" and "include" as used in the specification and the appended claims indicate the presence of the described features, integers, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0064] The present application will be further illustrated below in conjunction with the drawings and specific embodiments, but is not limited to the embodiments described below.

[0065] It should also be understood that the terms used in the present application specification are only for the purpose of describing specific embodiments and are not intended to limit the present application. As used in the present application specification and the appended claims, the singular forms "a", "an" and "the" are intended to include the plural forms unless the context clearly indicates otherwise.

[0066] It should be further understood that the term "and / or" used in the present application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes these combinations.

[0067] The present application is a paper recommendation method based on hierarchical graph attribute network representation learning. As shown in Figure 1 , the paper recommendation method based on hierarchical graph attribute network representation learning includes the following steps S1 to S4. At the same time, the framework diagram of the paper recommendation method based on hierarchical graph attribute network representation learning is as shown in Figure 2 .

[0068] S1, constructing a paper network based on a paper library, wherein the nodes in the paper network represent papers, the edges represent the citation relationship between papers, and the attribute information of the nodes is the information of the papers, and the information of the papers includes title, abstract, keywords, authors, institutions, and publication time.

[0069] S2, based on the paper network, constructing hierarchical graphs G 0 , G 1 ,..., G H by structure merging and attribute merging, wherein the structure merging includes edge merging, structure equivalence merging, and community merging.

[0070] Specifically, in step S2, the construction process of the hierarchical graph is specifically: according to the structural characteristics of the paper network, similar nodes are merged by edge merging, structural equivalence merging and community merging to realize structural merging, and the attribute information of the merged supernodes is calculated to realize attribute merging, and the two merging results are combined, and the process is iterated until the number of nodes of the hierarchical graph reaches a set threshold to obtain a small enough network, and the construction of the hierarchical graph is completed.

[0071] (1) Structural merging

[0072] Specifically, the structural merging of the network includes three key steps, namely edge merging, structural equivalence merging and community merging, which respectively maintain the first-order, second-order proximity and community characteristics of the network. By merging layer by layer, the high-order proximity of the network is maintained.

[0073] The edge merging is to select pairs of adjacent vertices in the paper network, and each pair of vertices is merged into a new vertex. The first-order proximity in the network is the local pairwise proximity between two vertices. Let (v i , v j ) represent each pair of nodes connected by an edge, and the weight of the edge represents the first-order proximity of the nodes v i and v j . In order to maintain the first-order proximity, edge merging is therefore required. As shown in Figure 3 processes (a) to (b), nodes v1 and v2 are merged into a single node v 12 , and nodes v3 and v4 are merged into a single node v 34 , maintaining the first-order proximity. This method uses maximum matching to select edges to be merged, selecting as many edges as possible and ensuring that no two edges can be merged into the same node at the same time.

[0074] The structural equivalence merging is to merge nodes in the paper network that share common neighbors into a new supernode. Considering that edge merging cannot handle the ring structure in the network, structural equivalence merging is required after edge merging. In the process (a) to (b), the ring structure in the network cannot be handled by edge merging, and nodes v 12 and v 34 share neighbor nodes v 57 and v 68 . By merging nodes v 12 and v 34 that share similar neighbors into supernode v 1234 in the attached Figure 2 process (b) to (c), the ring structure of the network can be handled while maintaining the second-order proximity of the network.

[0075] The community merging is to merge the nodes and edges in a community into a new node. It is known that, in addition to the first-order, second-order, high-order structure similarity of the network, the network also has the characteristics of community. The network is further divided into a community structure with hierarchy through a community discovery method.

[0076] In a specific embodiment, the community merging adopts the Louvain algorithm, divides the communities by maximizing the modularity, initializes each paper node in the paper network as a community, and if two paper nodes are merged into a community, the modularity is improved, then the merging is performed until the merging no longer improves the modularity; wherein the modularity is used to measure the density of edges within the community and edges between the communities, and the calculation formula of the modularity is:

[0077]

[0078] wherein Q represents the modularity of the paper network, m is the weight sum of all edges in the paper network, A ij represents the edge weight of the paper node v i and v j , k i , k j represents the degree of the paper node v i and v j , c i and c j is the community in which the paper node v i and v j is located; when the paper node v i and v j are in the same community, δ is 1.

[0079] In a specific embodiment, a binary mapping matrix is obtained through edge merging, structure equivalent merging and community merging to save the merging result, wherein V i represents the paper node set of the i-th hierarchical graph G i , and the matrix retains the paper node mapping relationship of the hierarchical graph G i to G i+1 ; the value of the matrix is 1 if the paper node r of G i is merged into the super node j of G i+1 , otherwise 0. The mapping matrix can also be used for the attribute merging.

[0080] (2) Attribute merging

[0081] Specifically, during the hierarchical graph construction process, while the structures are merged, the attribute information of the supernode generated by each structural merger is obtained from the previously merged paper nodes through a union operation. The attributes of the supernode are the union of the attributes of the original paper node set before the merger.

[0082] Since the attribute information is not involved in the structure merging, in order to fully retain its attribute information, during the hierarchical graph construction process, the attribute information of the nodes (super nodes) generated by each structure merging is obtained from the previously merged paper nodes through the union operation. Figure 2 As shown, when paper nodes v1, v2v3, and v4 are merged into supernode v 1234 , supernode v 1234 The attribute of G is the union of node sets v1, v2v3, and v4. Therefore, when G i Multiple paper nodes in Merged into G i+1 Supernodes in Merge into supernode Thesis node sequence The mapping matrix M i，i+1 To query. Among them, the super node Attributes It can be the union of the original paper node set:

[0083]

[0084] Taking full account of the structural characteristics of the original network, the network is iteratively constructed into a series of hierarchical graphs G from fine to coarse by merging similar nodes. 0 , G 1 ,...,G H Combining the structural merging and attribute merging methods proposed above, we construct a hierarchical graph by following the steps of edge merging, structural equivalence merging, community merging, and attribute merging. We iterate this process until the number of nodes in the hierarchical graph reaches a set threshold, resulting in a sufficiently small network.

[0085] S3. Perform network representation learning on the hierarchical graph to generate a paper node embedding representation, including steps S31-S33.

[0086] S31, performing intra-layer network representation learning on the hierarchical graph;

[0087] Specifically, the layer-by-layer network representation learning for the hierarchical graph includes: for the i-th layer hierarchical graph G i and its initialization network representation s i , use TADW or CANE attribute network representation learning method to learn the network representation of this layer network, where 0≤i≤H.

[0088] S32, performing inter-layer network representation learning on the hierarchical graph;

[0089] To interact the representation learning results of each layer network, inter-layer representation learning is proposed. Inter-layer representation learning is iteratively executed in a round-robin manner. Specifically, the initialization is forwardly executed from layer 1 to layer H, and then the initialization is reversely executed from layer H-1 to layer 0; the intra-layer network representation learning of the i-th (0≤i≤H-1) layer is executed twice, and the network of the H-th layer is executed only once. The forward execution introduces the interaction of the feature information between nodes of the finer layer to the intra-layer network representation learning of the coarser layer, and the message passing manner of the reverse execution enables the intra-layer representation learning of the finer layer to understand the feature information of the coarser layer.

[0090] The inter-layer network representation learning on the hierarchical graph specifically comprises:

[0091] (1) performing forward initialization: from layer 1 to layer H, the result embedding representation of the multiple paper nodes in G i-1 is initialized to the result embedding representation of the corresponding hypernode u in G i ; performing Max-Pooling to initialize the initial embedding representation of the hypernode u in G i ;

[0092]

[0093] wherein, represents the result embedding representation of the i-1-th layer paper node v in the forward initialization process, represents the initial embedding representation of the i-th layer paper node u in the forward initialization process, and in the original paper network graph G 0 , the initial node embedding representation of the intra-layer representation learning is randomly initialized;

[0094] For the hierarchical graph G i , in order to input the most influential feature information to the next layer and reduce the propagation of useless information, the method of Max-Pooling is adopted in the forward execution process.

[0095] (2) performing reverse initialization: from layer H-1 to layer 0, the result embedding representation newly learned by the paper node u in G i+1 and the result embedding representation generated by the original node v constituting the hypernode u in G i in the forward initialization are given to obtain the new initialization embedding of the paper node v by performing Mean-Pooling;

[0096]

[0097] wherein, represents the result embedding representation of the i+1-th layer paper node u in the backward initialization process, represents the initial embedding representation of the i-th layer paper node v in the backward initialization process, on the H-th layer, let the initial embedding representation of the paper node v in the backward initialization process be equal to the result embedding representation of the paper node v in the forward initialization process

[0098] Considering that in the backward propagation process, the feature information of the current layer and the next layer both have certain contribution, therefore, the method of Mean-Pooling is adopted to average to reduce the error.

[0099] Summarizing the above process, the inter-layer network representation learning first performs forward initialization, and then performs backward initialization. When performing forward initialization, G i After generating the node embedding representation in the intra-layer representation learning, it is initialized to the node embedding representation of G i+1 by Max-Pooling, until i=H. After that, backward initialization is performed, G i+1 After generating the node embedding representation in the intra-layer representation learning, it is initialized to the node embedding of G i by Mean-Pooling, until i=0.

[0100] S33, the backward initialization network representations in each level of the paper node are spliced in a cascaded manner to generate the paper node embedding representation, specifically:

[0101]

[0102] wherein, represents the backward initialization network representation of the i-th layer network of the paper node v, x v represents the embedding representation of the paper node v. Assuming that the dimension of the node representation of each level is equal and a, then the dimension of the final node representation is (H+1)d.

[0103] After learning the node embedding representation of each layer in the backward initialization process in step S33, in order to make the final node embedding representation contain the structure information and attribute information of the hierarchical network.

[0104] S4, according to the paper node embedding representation, the similarity between the embedding representations of the paper nodes is calculated to generate a paper recommendation list, and the paper recommendation list is recommended to the user.

[0105] After obtaining the embedded representation of the node through hierarchical graph attribute network representation learning, the paper recommendation list is generated by calculating the similarity of the embedded representation of the paper node.

[0106] In a specific embodiment, for the papers v that the user has read i , use the embedding representation of the paper node to calculate the similarity with other paper nodes in the paper network. The calculation formula is as follows:

[0107]

[0108] Among them, x i Represents the paper node v i The embedding representation of ⊙ represents the dot product operation, |x i | represents the paper node v i The module of the embedded representation; Sim i，j Represents the paper node v i and paper node v j similarity;

[0109] By comparing this method with each paper in the paper library, the final paper node v is obtained. i Similarity vector SIM with other papers in the paper network i ;

[0110] Select the set of papers v1, v2, v3, ..., v read within the time interval t n , respectively calculate the similarity vectors SIM1, SIM2, SIM3...SIM between the paper node and other papers in the paper library n . And the similarity vector SIM1, SIM2, SIM3......SIM n The similarity of the set of papers read by the current user within the time interval t is obtained by summing up, and finally the similarity is sorted in descending order to select the top-k papers to generate a paper recommendation list, and the paper recommendation list is recommended to the user.

[0111] Figure 4 This is a paper recommendation system based on hierarchical graph attribute network representation learning provided by an embodiment of the present invention. Figure 4 As shown, the paper recommendation system based on hierarchical graph attribute network representation learning includes a paper network construction module, a hierarchical graph construction module, a hierarchical network representation learning module, and a paper recommendation module.

[0112] The paper network construction module builds a paper network based on the paper library, where the nodes in the paper network represent papers, the edges represent the citation relationships between papers, and the attribute information of the nodes is the paper information, which includes: title, abstract, keywords, author, institution, and publication time;

[0113] The hierarchical graph construction module, based on the paper network, constructs a hierarchical graph G through structure merging and attribute merging 0 , G 1 ,...,G H , wherein the structural merging includes edge merging, structural equivalence merging, and community merging;

[0114] The hierarchical network representation learning module performs network representation learning on the hierarchical graph to generate the paper node embedding representation, including the intra-layer network representation learning module, the inter-layer network representation learning module and the paper node embedding representation generation module. Specifically:

[0115] An intra-layer network representation learning module, performing intra-layer network representation learning on the hierarchical graph;

[0116] The inter-layer network representation learning module performs inter-layer network representation learning on the hierarchical graph, specifically including:

[0117] 1) Perform forward initialization: Starting from layer 1 to layer H, G i-1 The result embedding representation of multiple paper nodes in Perform Max-Pooling to initialize the corresponding G i The initialization embedding representation of super node u in

[0118]

[0119] in, represents the result embedding representation of the paper node v in the i-1 layer during the forward initialization process, Indicates the initial embedding representation of the i-layer paper node u in the forward initialization process, in the original paper network graph G 0 In the random initialization layer, the initial node embedding representation for learning is initialized;

[0120] 2) Perform reverse initialization: Starting from layer H-1 to layer 0, given G i+1 The newly learned result embedding of node u in the paper and G i The result embedding generated in the forward initialization of the original node v that constitutes the super node u is Perform Mean-Pooling to get the new initial embedding of the paper node v

[0121]

[0122] in, represents the result embedding representation of the paper node u in the i+1 layer during the reverse initialization process, The initial embedding representation of the i-layer paper node v in the backward initialization process, on the H-th layer, let the initialization embedding representation of the paper node v in the backward initialization process equal to the result embedding representation of the paper node v in the forward initialization process

[0123] The paper node embedding representation generation module adopts a cascading manner to splice the backward initialization network representation of each layer of the paper node to generate the paper node embedding representation, specifically:

[0124]

[0125] Among them, The backward initialization network representation of the i-layer network paper node v, x v The embedding representation of the paper node v;

[0126] The paper recommendation module generates a paper recommendation list by calculating the similarity of the embedding representations between the paper nodes according to the paper node embedding representation, and recommends the paper recommendation list to the user.

[0127] The above-mentioned paper recommendation system based on hierarchical graph attribute network representation learning can be implemented in the form of a computer program, which can run on a computer device.

[0128] The computer device can be a server, wherein the server can be a standalone server or a server cluster composed of multiple servers.

[0129] The computer device includes a processor, a memory and a network interface connected by a system bus, wherein the memory can include a non-volatile storage medium and an internal memory.

[0130] The non-volatile storage medium can store an operating system and a computer program. The computer program includes program instructions, which when executed, can cause the processor to execute a paper recommendation method based on hierarchical graph attribute network representation learning.

[0131] The processor is used to provide computing and control capabilities to support the operation of the entire computer device.

[0132] The internal memory provides an environment for the running of the computer program in the non-volatile storage medium, which when executed by the processor, can cause the processor to execute a paper recommendation method based on hierarchical graph attribute network representation learning.

[0133] The network interface is used for network communication with other devices. Those skilled in the art can understand that the above computer device structure is only part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different component arrangement.

[0134] The processor is configured to run a computer program stored in the memory, and the computer program implements the paper recommendation method based on hierarchical graph attribute network representation learning in Embodiment 1.

[0135] It should be understood that, in the embodiments of the present application, the processor can be a central processing unit (CPU), and the processor can also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor.

[0136] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiments can be completed by a computer program instructing relevant hardware. The computer program includes program instructions, and the computer program can be stored in a storage medium, which is a computer-readable storage medium. The program instructions are executed by at least one processor in the computer system to implement the process steps of the above-mentioned embodiments.

[0137] The present application also provides a storage medium. The storage medium can be a computer-readable storage medium. The storage medium stores a computer program, wherein the computer program is executed by a processor to make the processor execute the paper recommendation method based on hierarchical graph attribute network representation learning in Embodiment 1.

[0138] The storage medium can be a U disk, a mobile hard disk, a read-only memory (ROM), a magnetic disk or an optical disk, and various computer-readable storage media that can store program codes.

[0139] Those skilled in the art can understand that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized in electronic hardware, computer software or a combination of both. In order to clearly illustrate the interchangeability of hardware and software, the components and steps of each example have been described in the above description in a general manner. Whether the functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0140] In several embodiments provided by the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of each unit is only a logical function division, and actual implementation can have another division manner. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed.

[0141] The steps in the method embodiments of the present application can be sequentially adjusted, combined and deleted according to actual needs. The units in the device embodiments of the present application can be combined, divided and deleted according to actual needs. In addition, each functional unit in each embodiment of the present application can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit.

[0142] The integrated unit, if realized in the form of a software functional unit and sold or used as an independent product, can be stored in a storage medium. Based on such understanding, the technical solutions of the present application essentially or say the part that contributes to the prior art, or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a terminal or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application.

[0143] Note that the above is only the preferred embodiment of the present application and the technical principle applied. Those skilled in the art will understand that the present application is not limited to the specific embodiments described herein, and those skilled in the art can make various obvious changes, readjustments and substitutions without departing from the scope of the present application. Therefore, although the present application has been described in more detail through the above embodiments, the present application is not limited to the above embodiments, and can include more other equivalent embodiments without departing from the concept of the present application, and the scope of the present application is determined by the scope of the appended claims.

Claims

1. A paper recommendation method based on hierarchical graph attribute network representation learning, characterized by: Including steps: S1. Construct a paper network based on the paper database, wherein the nodes in the paper network represent papers, the edges represent the citation relationships between papers, and the attribute information of the nodes is the paper information, which includes: title, abstract, keywords, author, institution, and publication time; S2. Based on the paper network, construct a hierarchical graph G through structure merging and attribute merging 0 , G 1 ,...,G H , wherein the structural merging includes edge merging, structural equivalence merging, and community merging; S3. Perform network representation learning on the hierarchical graph to generate embedded representations of paper nodes, including: S31, performing intra-layer network representation learning on the hierarchical graph; S32, performing inter-layer network representation learning on the hierarchical graph, specifically comprising: 1) Perform forward initialization: Starting from layer 1 to layer H, G i-1 The result embedding representation of multiple paper nodes in Perform Max-Pooling to initialize the corresponding G i The initialization embedding representation of super node u in in, represents the result embedding representation of the paper node v in the i-1 layer during the forward initialization process, Indicates the initial embedding representation of the i-layer paper node u in the forward initialization process, in the original paper network graph G 0 In the random initialization layer, the initial node embedding representation for learning is initialized; 2) Perform reverse initialization: Starting from layer H-1 to layer 0, given G i+1 The newly learned result embedding of the paper node u and G i The result embedding generated in the forward initialization of the original node v that constitutes the super node u is Perform Mean-Pooling to get the new initial embedding of the paper node v in, represents the result embedding representation of the paper node u in the i+1 layer during the reverse initialization process, Represents the initial embedding representation of the paper node v in the i-th layer during the reverse initialization process. On the H-th layer, let the initial embedding representation of the paper node v in the reverse initialization process be The resulting embedding representation of the paper node v is equal to the forward initialization process S33. Use a cascade method to splice the reverse initialized network representations of each layer of the paper node to generate the embedded representation of the paper node, specifically: in, represents the reverse initialization network representation of the paper node v in the i-th layer network, x v represents the embedding representation of the paper node v; S4. Generate a paper recommendation list based on the embedding representation of the paper nodes by calculating the similarity of the embedding representation between the paper nodes, and recommend the paper recommendation list to the user.

2. The method according to claim 1, characterized in that In step S2, the construction process of the hierarchical graph is specifically as follows: according to the structural characteristics of the paper network, similar nodes are merged through edge merging, structural equivalence merging and community merging to achieve structural merging, and the attribute information of the merged super nodes is calculated to achieve attribute merging. The two merging results are combined and this process is iterated until the number of nodes in the hierarchical graph reaches the set threshold to obtain a sufficiently small network, thereby completing the construction of the hierarchical graph.

3. The method according to claim 2, characterized in that The edge merging is to select pairs of adjacent vertices in the paper network and merge each pair of vertices into a new vertex; the structural equivalence merging is to merge the nodes that share common neighbors in the paper network into a new supernode; The community merging is to merge the nodes and edges in a community into a new node.

4. The method according to claim 3, characterized in that The community merging adopts the Louvain algorithm to divide the community by maximizing the modularity. Each paper node in the paper network is initialized as a community. If merging two paper nodes into one community can improve the modularity, the merger is performed until the modularity is no longer improved. The modularity is used to measure the density of edges within the community and the edges between communities. The calculation formula of modularity is: Among them, Q represents the modularity of the paper network, m is the weight sum of all edges in the paper network, and A ij Represents the paper node v i With v j The edge weight, k i ,k j Represents the paper node v i and v j degree, c i and c j Is the paper node v i and v j The community where the paper node v i and v j When in the same community, δ is 1.

5. The method according to claim 2, characterized in that A binary mapping matrix is ​​obtained by edge merging, structural equivalent merging and community merging. Used to save the merged results, where V i Represents the i-th level graph G i The paper node set, the matrix retains the hierarchical graph G i to G i+1 The paper node mapping relationship; If the value is 1, it means G i The paper node r will be merged into G i+1 Otherwise, it is 0.

6. The method according to claim 2, characterized in that During the hierarchical graph construction process, while the structures are merged, the attribute information of the supernode generated by each structural merger is obtained from the previously merged paper nodes through a union operation. The attributes of the supernode are the union of the attributes of the original paper node set before the merger.

7. The method according to claim 1, characterized in that In step S31, the intra-layer network representation learning of the hierarchical graph includes: for the hierarchical graph G of the i-th layer i and its initialization network representation S i , use TADW or CANE attribute network representation learning method to learn the network representation of this layer network, where 0≤i≤H.

8. The method according to claim 1, characterized in that Step S4 specifically includes: i , use the embedding representation of the paper node to calculate the similarity with other paper nodes in the paper network. The calculation formula is as follows: Among them, x i Represents the paper node v i The embedding representation of ⊙ represents the dot product operation, |x i | represents the paper node v i The module of the embedded representation; Sim i，j Represents the paper node v i and paper node v j similarity; By comparing this method with each paper in the paper library, the final paper node v is obtained. i Similarity vector SIM with other papers in the paper network i ; Select the set of papers v1, v2, v3, ..., v read within the time interval t n , respectively calculate the similarity vectors SIM1, SIM2, SIM3...SIM between the paper node and other papers in the paper library n . And the similarity vector SIM1, SIM2, SIM3......SIM n The similarity of the set of papers read by the current user within the time interval t is obtained by summing up, and finally the similarity is sorted in descending order to select the top-k papers to generate a paper recommendation list, and the paper recommendation list is recommended to the user.

9. A paper recommendation system based on hierarchical graph attribute network representation learning, comprising: Paper network construction module, hierarchical graph construction module, hierarchical network representation learning module, paper recommendation module; The paper network construction module builds a paper network based on the paper library, where the nodes in the paper network represent papers, the edges represent the citation relationships between papers, and the attribute information of the nodes is the paper information, which includes: title, abstract, keywords, author, institution, and publication time; The hierarchical graph construction module, based on the paper network, constructs a hierarchical graph G through structure merging and attribute merging 0 , G 1 ,...,G H , wherein the structural merging includes edge merging, structural equivalence merging, and community merging; the hierarchical network representation learning module performs network representation learning on the hierarchical graph to generate the paper node embedding representation, including the intra-layer network representation learning module, the inter-layer network representation learning module, and the paper node embedding representation generation module: An intra-layer network representation learning module performs intra-layer network representation learning on the hierarchical graph; The inter-layer network representation learning module performs inter-layer network representation learning on the hierarchical graph, specifically including: 1) Perform forward initialization: Starting from layer 1 to layer H, G i-1 The result embedding representation of multiple paper nodes in Perform Max-Pooling to initialize the corresponding G i The initialization embedding representation of super node u in in, represents the result embedding representation of the paper node v in the i-1 layer during the forward initialization process, Indicates the initial embedding representation of the i-layer paper node u in the forward initialization process, in the original paper network graph G 0 In the random initialization layer, the initial node embedding representation for learning is initialized; 2) Perform reverse initialization: Starting from layer H-1 to layer 0, given G i+1 The newly learned result embedding of the paper node u and G i The result embedding generated in the forward initialization of the original node v that constitutes the super node u is Perform Mean-Pooling to get the new initial embedding of the paper node v in, represents the result embedding representation of the paper node u in the i+1 layer during the reverse initialization process, Represents the initial embedding representation of the paper node v in the i-th layer during the reverse initialization process. On the H-th layer, let the initial embedding representation of the paper node v in the reverse initialization process be The resulting embedding representation of the paper node v is equal to the forward initialization process The paper node embedding representation generation module uses a cascade method to splice the reverse initialized network representations of each layer of the paper node to generate the paper node embedding representation, specifically: in, represents the reverse initialization network representation of the paper node v in the i-th layer network, x v represents the embedding representation of the paper node v; The paper recommendation module generates a paper recommendation list based on the embedding representation of the paper nodes by calculating the similarity of the embedding representation between the paper nodes, and recommends the paper recommendation list to the user.

10. A computer device, characterized in that: The device includes a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the method according to any one of claims 1 to 8 when executing the computer program.

Images