Metapath-based heterogeneous network relationship supplementing method among member enterprises of enterprise group

Abstract

The invention provides a meta-path-based heterogeneous network relationship supplementing method between enterprise group member enterprises. The method comprises the following steps: constructing an enterprise group heterogeneous graph; the enterprise group heterogeneous graph is used for representing a heterogeneous network relationship and a node initial representation vector between enterprise group member enterprises; performing node neighborhood sampling based on a meta-path set by the enterprise group heterogeneous graph to obtain a neighborhood node set; on the basis of the neighborhood node set, a heterogeneous graph representation learning method is adopted to update the initial representation vector of the node, and a node embedding representation vector in the heterogeneous graph of the enterprise group is obtained; inputting the node pair to be subjected to link prediction to a preset link predictor to obtain a link prediction result output by the link predictor; and supplementing the heterogeneous network relationship according to a link prediction result. According to the method, the topological structure of the heterogeneous network and the attribute information of the nodes are coupled through the enterprise group heterogeneous graph, and link prediction and relation supplementation of the heterogeneous network relation between enterprises can be carried out by using more comprehensive structure information.

Description

Technical Field

[0001] The present invention relates to the technical field of heterogeneous networks, and in particular to a method for supplementing heterogeneous network relationships among member enterprises of an enterprise group based on a meta-path. Background Art

[0002] In the real world, various types of interactions between enterprises are one of the important references for corporate credit analysis, and the enterprise-related data representing these interactions are often scattered across multiple platforms and channels, making it difficult for a single channel to fully record and store them. Due to the barriers between channels and the cost of data acquisition, the data that can be obtained is quite limited. This part of data accounts for a very small proportion of all interactive data in the real world, so the coverage of the actual situation of the inter-enterprise related data that can be obtained is relatively small. In the scenario of using enterprise interaction data for data analysis or modeling, the lack of data will cause the analysis results to deviate greatly from the actual situation.

[0003] The link prediction task has a long history. Its goal is to predict the missing or generated links between pairs of nodes in the network. It has been widely used in social networks, biological networks, citation networks, recommendation systems and other fields. Existing link prediction methods include methods that rely on heuristic node similarity scores and node embedding methods that rely on deep walking algorithms. However, existing methods can only use the topological structure information in the prediction network, and cannot use the attribute information of the nodes in the prediction network.

[0004] Therefore, in order to achieve link prediction between member companies of a corporate group, it is urgent to design a method that can use more comprehensive structural information for link prediction. Summary of the invention

[0005] The present invention provides a method and system for supplementing heterogeneous network relationships between member enterprises of an enterprise group based on meta-paths, which is used to solve the defect that only the topological structure information in the prediction network can be used, but the attribute information of the nodes in the prediction network cannot be used. The present invention couples the topological structure of the heterogeneous network with the attribute information of the nodes through the node embedding representation vector in the heterogeneous graph of the enterprise group, and can use more comprehensive structural information to perform link prediction and relationship supplement of heterogeneous network relationships between enterprises.

[0006] The present invention provides a method for supplementing heterogeneous network relationships between member enterprises of an enterprise group based on a meta-path, comprising: constructing a heterogeneous graph of an enterprise group; the heterogeneous graph of the enterprise group is used to characterize the heterogeneous network relationships and initial node representation vectors between member enterprises of the enterprise group; performing node neighborhood sampling based on the meta-path set by the heterogeneous graph of the enterprise group to obtain a neighborhood node set; based on the neighborhood node set, using a heterogeneous graph representation learning method to update the initial node representation vector to obtain a node embedding representation vector; the node embedding representation vector is used to determine whether a node pair to be predicted for linking exists; inputting the node pair to be predicted for linking into a preset link predictor to obtain a link prediction result output by the link predictor; and supplementing the heterogeneous network relationship according to the link prediction result.

[0007] According to a meta-path-based method for supplementing heterogeneous network relationships between member enterprises of an enterprise group provided by the present invention, the construction of a heterogeneous graph of the enterprise group includes: determining existing enterprises in a selected enterprise group and the relationship types between enterprises; determining initial node representation vectors corresponding to the existing enterprises through feature engineering; generating bidirectional semantic edges between enterprises according to the relationship types between the enterprises; the bidirectional semantic edges are used to characterize the heterogeneous network relationships between enterprises; and constructing the heterogeneous graph of the enterprise group based on the initial node representation vectors and the bidirectional semantic edges.

[0008] According to a meta-path-based heterogeneous network relationship supplement method between member enterprises of an enterprise group provided by the present invention, node neighborhood sampling is performed based on the meta-path set up in the heterogeneous graph of the enterprise group to obtain a neighborhood node set, including: obtaining a candidate neighborhood node sequence based on a meta-path-guided random walk sampling strategy set up in the heterogeneous graph of the enterprise group; and selecting a preset number of neighborhood nodes from the candidate neighborhood node sequence to form the neighborhood node set.

[0009] According to a meta-path-based heterogeneous network relationship supplement method between member enterprises of an enterprise group provided by the present invention, the node embedding representation vector in the heterogeneous graph of the enterprise group is determined based on the neighborhood node set using a heterogeneous graph representation learning method, including: based on the neighborhood node set, determining the attention weights of the nodes in the meta-path through a node-level attention mechanism; aggregating the node information in the meta-path based on the attention weights of the nodes in the meta-path to obtain a node-level embedding representation vector; based on the node-level embedding representation vector, determining the semantic attention weights between meta-paths through a semantic-level attention mechanism; aggregating the semantic information between meta-paths based on the semantic attention weights between meta-paths to obtain a node embedding representation vector in the heterogeneous graph of the enterprise group.

[0010] According to a meta-path-based heterogeneous network relationship supplement method between member enterprises of an enterprise group provided by the present invention, the link predictor includes a node pair determination module, a similarity determination module and a result generation module; the node pair determination module is used based on the node pairs to be linked; the similarity determination module is used to calculate the similarity score of the node pairs to be linked by using a dot product method; the result generation module is used to generate the link prediction result based on the similarity score.

[0011] According to a meta-path-based method for supplementing heterogeneous network relationships between member enterprises of an enterprise group provided by the present invention, the heterogeneous network relationships are supplemented according to the link prediction results, including: corresponding bidirectional link prediction results to node pairs with bidirectional semantics of the same type of relationship; when the bidirectional link prediction results both predict the existence of a relationship, the heterogeneous network relationships represented by the node pairs are supplemented to the heterogeneous graph of the enterprise group.

[0012] The present invention also provides a system for supplementing heterogeneous network relationships between member enterprises of an enterprise group based on a meta-path, comprising: a construction module, used to construct a heterogeneous graph of an enterprise group; the heterogeneous graph of an enterprise group is used to characterize the heterogeneous network relationships between member enterprises of the enterprise group and the initial representation vector of nodes; a sampling module, used to perform node neighborhood sampling based on the meta-path set in the heterogeneous graph of the enterprise group to obtain a neighborhood node set; a representation learning module, used to update the initial node representation vector based on the neighborhood node set using a heterogeneous graph representation learning method to obtain a node embedding representation vector; the node embedding representation vector is used to determine whether a node pair to be predicted for linking exists; a prediction module, used to input the node pair to be predicted for linking into a preset link predictor to obtain a link prediction result output by the link predictor; and a supplement module, used to supplement the heterogeneous network relationship according to the link prediction result.

[0013] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method for supplementing heterogeneous network relationships between member enterprises of an enterprise group based on meta-paths as described in any one of the above-mentioned methods.

[0014] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the above-described methods for supplementing heterogeneous network relationships between member enterprises of an enterprise group based on meta-paths.

[0015] The present invention also provides a computer program product, including a computer program, which, when executed by a processor, implements any of the above-mentioned methods for supplementing heterogeneous network relationships between member enterprises of an enterprise group based on meta-paths.

[0016] The present invention provides a method for supplementing heterogeneous network relationships between member enterprises of an enterprise group based on a meta-path, the method comprising: constructing a heterogeneous graph of an enterprise group; the heterogeneous graph of an enterprise group is used to characterize the heterogeneous network relationships and initial node representation vectors between member enterprises of the enterprise group; performing node neighborhood sampling based on the meta-path set in the heterogeneous graph of the enterprise group to obtain a neighborhood node set; based on the neighborhood node set, using a heterogeneous graph representation learning method to update the initial node representation vector to obtain a node embedding representation vector in the heterogeneous graph of the enterprise group; inputting a node pair to be link predicted into a preset link predictor to obtain a link prediction result output by the link predictor; and supplementing the heterogeneous network relationship according to the link prediction result. The present invention couples the topological structure of the heterogeneous network with the attribute information of the node through the heterogeneous graph of the enterprise group, and can use more comprehensive structural information to perform link prediction and relationship supplement of heterogeneous network relationships between enterprises. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0018] Figure 1 It is a flow chart of a method for supplementing heterogeneous network relationships among member enterprises of an enterprise group based on meta-paths provided by the present invention.

[0019] Figure 2 It is a schematic diagram of the principle of a method for supplementing heterogeneous network relationships among member enterprises of an enterprise group based on a meta-path provided by the present invention.

[0020] Figure 3 It is a principle schematic diagram of the enterprise group heterogeneous graph provided by the present invention.

[0021] Figure 4 It is a schematic diagram of the principle of the meta-path provided by the present invention.

[0022] Figure 5 It is a schematic diagram of the principle of the node-level and semantic-level dual attention message aggregation mechanism provided by the present invention.

[0023] Figure 6 This is a logical diagram of supplementary determination of inter-enterprise relations provided by the present invention.

[0024] Figure 7 It is a structural schematic diagram of a heterogeneous network relationship supplement system among member enterprises of an enterprise group based on a meta-path provided by the present invention.

[0025] Figure 8It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0026] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0027] In the scenario where there is a lack of inter-enterprise association data, existing link prediction methods include Common Neighbors (CN), Adamic-Adar (AA), SimRank and other methods that rely on heuristic node similarity scores, as well as matrix factorization (MF) or node2vec and other node embedding methods that rely on the DeepWalk algorithm. However, the above methods can only use the topological structure information in the prediction network, and cannot use the attribute information of the nodes in the prediction network, making it difficult to achieve accurate prediction of inter-enterprise relationships.

[0028] Please refer to Figure 1 , Figure 1 A schematic flow chart of a method for supplementing heterogeneous network relationships among member enterprises of an enterprise group based on meta-paths provided by the present invention.

[0029] Please refer to Figure 2 , Figure 2 A schematic diagram of the principle of a method for supplementing heterogeneous network relationships among member enterprises of an enterprise group based on meta-paths provided by the present invention.

[0030] In order to solve the technical problems existing in the prior art, the present invention provides a method for supplementing heterogeneous network relationships among member enterprises of an enterprise group based on a meta-path, comprising: 101: Constructing a heterogeneous graph of enterprise groups; the heterogeneous graph of enterprise groups is used to represent the heterogeneous network relationships and node initial representation vectors among the member enterprises of the enterprise group; 102: Based on the meta-path set in the heterogeneous graph of the enterprise group, node neighborhood sampling is performed to obtain a neighborhood node set; 103: Based on the neighborhood node set, a heterogeneous graph representation learning method is used to update the initial node representation vector to obtain a node embedding representation vector; the node embedding representation vector is used to determine whether there is a link between the node pairs to be predicted; 104: Input the node pair to be link predicted to a preset link predictor, and obtain the link prediction result output by the link predictor; 105: Supplement heterogeneous network relationships based on link prediction results.

[0031] As a link prediction method based on network structure, graph neural network can use more comprehensive structural information to model link prediction tasks compared to traditional link prediction methods that rely on heuristic node similarity scores or node embeddings. Among them, the node-based method learns node embeddings containing graph structures through local neighborhoods, and then aggregates the learned node pair embeddings to construct link representations between nodes. Graph neural networks allow enterprise nodes to exchange information through their unique information propagation mechanism. In the graph neural network model, the update status of a node depends not only on its own attributes, but also on the status of its neighboring nodes. The present invention is proposed for scenarios where there is a lack of inter-enterprise correlation data, and aims to model the correlation patterns between enterprises through graph neural network technology, and to infer potential unknown correlations through known correlation data, so as to supplement the actual acquired data, so that the amount of data available for scenarios based on inter-enterprise correlation data analysis or modeling is more sufficient.

[0032] Specifically, the present invention first constructs a heterogeneous graph of an enterprise group (a heterogeneous graph neural network model), incorporates the heterogeneous network relationship (topological structure) between the member enterprises of the enterprise group and the initial node representation vector (node ​​attribute information) of the enterprise, so as to more accurately predict whether the enterprise has a credit risk at the moment and whether it is worth being absorbed as a potential high-quality customer through these relationships. In the context of the heterogeneous graph of the enterprise group, it is necessary to consider the heterogeneity brought by the multi-type edges in the graph and the large-scale characteristics brought by hundreds of thousands of nodes to build a link prediction graph neural network of multi-type relationships between the member enterprises of the enterprise group. For the heterogeneity of the enterprise group graph, since the distribution of the influence of the neighborhoods reached by different types of edges on the central node is not necessarily the same, it is necessary to use a suitable method to comprehensively incorporate the neighborhood information composed of different nodes reached by multiple types of edges; for the characteristics of a large number of nodes and a large scale in the enterprise group (representation learning problem of large-scale graphs), it is necessary to perform node neighborhood sampling based on the meta-path set in the heterogeneous graph of the enterprise group during the learning process to reduce the scale of message aggregation used for representation learning, so as to extract the neighborhood subgraph (neighborhood node set) with representative neighborhood information of the central node. In addition, the representation learning problem of heterogeneous graphs (heterogeneous graph representation learning method) focuses on how to properly aggregate heterogeneous semantic messages provided by different types of edges in the enterprise group graph to update the node representation based on the neighborhood node set, so as to obtain the appropriate node embedding representation vector, and determine whether there is a link between the node pairs to be predicted based on the node embedding representation vector. Then the node pairs to be predicted are input into a preset link predictor, the existence of various types of edges between enterprise nodes is evaluated, and the link prediction results output by the link predictor are obtained, thereby improving the prediction accuracy. According to the link prediction results, it is judged whether a certain relationship exists between the node pairs, so as to supplement the heterogeneous network relationships. The present invention couples the topological structure of the heterogeneous network with the attribute information of the nodes through the heterogeneous graph of the enterprise group, and can use more comprehensive structural information to perform link prediction and relationship supplementation of heterogeneous network relationships between enterprises.

[0033] As a preferred embodiment, constructing a heterogeneous graph of an enterprise group includes: determining existing enterprises in a selected enterprise group and the types of relationships between enterprises; determining initial node representation vectors corresponding to existing enterprises through feature engineering; generating bidirectional semantic edges between enterprises according to the types of relationships between enterprises; the bidirectional semantic edges are used to characterize the heterogeneous network relationships between enterprises; and constructing a heterogeneous graph of the enterprise group based on the initial node representation vectors and the bidirectional semantic edges.

[0034] Please refer to Figure 3 , Figure 3 A schematic diagram of the principles of the heterogeneous graph of enterprise groups provided by the present invention.

[0035] Taking the new energy automobile enterprise group that can represent the intersection of new energy and smart cars as an example, in this embodiment, the collected industry code and business scope are used as screening conditions to determine whether the enterprise belongs to the new energy automobile enterprise group, and the enterprise customers belonging to the enterprise group are constructed as existing enterprises in the form of keyword fuzzy matching. Existing enterprises, as the node source of the enterprise group heterogeneous graph, can not only meet the characteristics of strong correlation of enterprise groups, but also have rich association data between nodes, so that the enterprise group graph can capture relatively more information transmission relationships between enterprises.

[0036] In order to fully reflect the credit and risk of the enterprise, several data tables containing basic information of the enterprise such as business operation, credit, equity, transaction, investment, guarantee, industry, region, scale, etc. can be selected to construct the characteristics of the enterprise node. Using the original data of existing enterprises, the enterprise node characteristics are constructed through feature engineering to obtain the initial node representation vector. Feature engineering includes missing value processing, concentration elimination, outlier capping, variable binning and WOE (Weight of Evidence) coding, IV (Information Value) value variable screening and Pearson correlation coefficient elimination.

[0037] The enterprise group graph is defined as a heterogeneous directed graph, which can describe the enterprises in the enterprise group and the various relationships between enterprises. Nodes are used to describe enterprises, and edges are used to describe the relationships between enterprises. The heterogeneity of the enterprise group graph is reflected in the multiple categories of edges, while the category of nodes is only one type of enterprise. , is a node set, is the edge set, is the node category, is the edge category. In order to model the transmission of credit and risk, this embodiment samples three types of inter-enterprise related information: transaction, investment and financing, and guarantee to construct relationship edges between enterprise nodes. These three types of relationships can reflect the credit and risk transmission rules between enterprises from different angles. For transaction relationships, the circulation of products and services in enterprise groups can reflect the strong dependence between enterprises. For example, problems with raw material suppliers will affect the production of downstream enterprises, and risks will also be passed from suppliers to downstream production enterprises; for investment and financing relationships, equity investment relationships will form a holding structure. Generally speaking, the financial status and credit problems of the parent company will affect its subsidiaries to a certain extent; for guarantee relationships, the enterprise as the guarantor can bear certain risks for the guaranteed enterprise through the guarantee relationship. Correspondingly, if there is a problem with the guarantor, its credit changes will also be passed to the guaranteed party at a certain ratio.

[0038] In order to provide more association semantics for a single node, this embodiment extracts 6 types of relationship edges from the three types of relationships: transaction, investment and financing, and guarantee (bidirectional semantic edges are used to characterize heterogeneous network relationships between enterprises). Among them, a transaction has a financial impact on both parties of the transaction. For the borrower, it means an inflow of funds, and for the lender, it means an outflow of funds. Therefore, the lending transaction edge (Transaction Edge, TX) and the borrowing transaction edge (ReverseTransaction Edge, TXR) are extracted from the transaction relationship to bidirectionally propagate the impact of the transaction. The same is true for the investment and guarantee relationship. An investment behavior means an outflow of funds for the investor and an inflow of funds for the financing party. A guarantee behavior means credit lending for the guarantor and credit growth for the guaranteed party. Therefore, the investment edge (InvestmentEdge, IV) and the financing edge (Reverse Investment Edge, IVR), the guarantee edge (Guarantee Edge, GR) and the guaranteed edge (Reverse Guarantee Edge, GRR) are extracted respectively. Therefore, the types of relationships in the heterogeneous graph of enterprise groups can be expressed as .

[0039] As a preferred embodiment, node neighborhood sampling is performed based on the meta-path set in the enterprise group heterogeneous graph to obtain a neighborhood node set, including: obtaining a candidate neighborhood node sequence based on a meta-path guided random walk sampling strategy set in the enterprise group heterogeneous graph; selecting a preset number of neighborhood nodes from the candidate neighborhood node sequence to form a neighborhood node set.

[0040] Please refer to Figure 4 , Figure 4 A schematic diagram of the principle of the meta-path provided by the present invention.

[0041] For the problem of node representation learning in large-scale graphs, in this embodiment, a meta-path is introduced to describe the relationship between nodes in heterogeneous graphs, and the neighborhood range of node representation learning is narrowed by sampling the neighborhood of nodes based on the meta-path. First, a series of meta-paths are customized according to the six different types of edges in the enterprise group graph, and a random walk strategy guided by the meta-path is used to obtain the neighborhood node set of the central node collected under the guidance of different meta-paths. Before the representation learning of the enterprise group heterogeneous graph, the enterprise group heterogeneous graph structure containing multiple types of edges is effectively sampled, which can reduce the computational complexity of message passing in the graph.

[0042] Metapath is a key concept used in heterogeneous information network (HIN) analysis. It is defined as a path that expresses complex semantic information between different types of entities (nodes) and relationships (edges) in a network. Metapath is mainly used to express complex relationships between multiple types of nodes and edges in a heterogeneous graph. A metapath in a heterogeneous graph of a corporate group is defined as a sequence of node and edge types: , where the node ,side , the meta-path sequence represents the sequence from Type of node starts by Type of edge to Nodes of the same type, and so on, eventually reaching Considering that there is only one type of node in the heterogeneous graph of enterprise groups, the unified form of the meta-path designed for the enterprise group graph can be simplified to .

[0043] The random walk sampling strategy starts from a node in a random walk process, then randomly selects a neighboring node to move to the next node, forms a path through continuous walking, and after merging and removing duplicate points along all paths, the resulting graph can be used as a representative subset of the original graph. Using meta-paths to guide random walks is a biased sampling strategy, which means that the node sequence generation mechanism during the walk process is not to randomly select neighboring nodes of the current node, but to confirm candidate neighboring nodes that meet the triple type of <current node type, edge type, next node type> through the setting of meta-paths and the current node type, and then randomly select the next node in the candidate neighboring node to be placed in the current sequence. In the candidate neighboring node sequence obtained by the random walk process guided by meta-paths, a preset number of neighboring nodes are selected to construct a neighboring node set as the neighborhood of the central node under the semantics of the current meta-path.

[0044] This embodiment uses 6 types of edges to set up a series of meta-paths that are meaningful in the propagation of financial risks based on real-world meanings. Different relationships have different degrees of influence on the assets of central node enterprises. Among them, the direct impact of capital inflows from purchasing services from enterprises is the greatest. The edge corresponding to the central node is TXR, followed by outward guarantee or investment relationships, corresponding to edges IV and GR. If problems arise in the guaranteed or invested enterprises, it will also directly affect the assets and credit of the central enterprise. The other three edges TX, IVR, and GRR have indirect impacts on the economy of the central enterprise. In order to limit the number of model parameters and control the risk of overfitting, this embodiment sets 12 meta-paths with the highest importance for neighborhood nodes within three degrees, which are first-degree meta-paths, , two-dimensional path , three-dimensional path Before learning the heterogeneous graph representation, a series of meta-paths with realistic financial risk transmission semantics designed in this embodiment are used to sample neighborhood nodes with different semantics around the central node, providing a set of neighborhood nodes with high importance to the central node for the next step of neighborhood message aggregation.

[0045] As a preferred embodiment, based on the neighborhood node set, a heterogeneous graph representation learning method is used to determine the node embedding representation vector in the heterogeneous graph of the enterprise group, including: based on the neighborhood node set, determining the attention weights of the nodes in the meta-path through the node-level attention mechanism; aggregating the node information in the meta-path based on the attention weights of the nodes in the meta-path to obtain the node-level embedding representation vector; based on the node-level embedding representation vector, determining the attention weights of the semantics between meta-paths through the semantic-level attention mechanism; aggregating the semantic information between meta-paths based on the attention weights of the semantics between meta-paths to obtain the node embedding representation vector in the heterogeneous graph of the enterprise group.

[0046] Please refer to Figure 5 , Figure 5 Schematic diagram of the principle of the node-level and semantic-level dual attention message aggregation mechanism provided by the present invention.

[0047] For the problem of learning heterogeneous graph representation, in this embodiment, a dual attention mechanism for heterogeneous graphs is introduced. The dual attention mechanism is divided into a node-level attention mechanism and a semantic-level attention mechanism, which are used to model the attention weights of the central node to different neighborhood nodes within and between meta-paths. The node-level attention mechanism is used to assign attention weights to a set of nodes obtained by a meta-path, while the semantic-level attention is used to determine the weight distribution between different semantics represented by node sets extracted from different meta-paths. The attention weights learned by these two attention mechanisms represent the information of different nodes in the meta-path and the contribution of information brought by different meta-paths to the node representation. The neighborhood node messages are aggregated according to the attention weights, and a new global node representation (node ​​embedding representation vector) is updated and generated.

[0048] When aggregating neighbor node messages obtained by each node through the meta-path, it is necessary to note that different neighbor nodes have different importance to the central node. This embodiment uses a multi-head attention mechanism (node-level attention mechanism) to evaluate the importance of the neighbor nodes obtained under the guidance of each meta-path to the central node as the attention weight of the node in the meta-path, and aggregates the neighbor node representation based on the attention weight of the node in the meta-path to update the embedding of the central node (node-level embedding representation vector).

[0049] By a certain meta path In the obtained neighborhood node set, the principle formula for the single-head attention layer to use the representation of each neighborhood node to update the central node representation is:

[0050] in, is the node updated by the single-head attention layer represents a vector, concat is a vector concatenation function, As a nonlinear activation function, the present invention uses the ReLU function, For Node The set of neighbor nodes of For Node The node representation vector before updating, The normalized node With Node Attention weights, parameter matrix is the parameter matrix shared between nodes, For Node With Node The original attention coefficient between them.

[0051] First, use a shared parameter matrix All nodes are linearly transformed, the node representation is mapped to a high-dimensional feature space to obtain sufficient expression ability, and then input into a single-layer feedforward neural network To calculate the attention coefficient between adjacent nodes, the coefficient represents the node i The feature pair node j The attention mechanism used here is masked graph attention, which introduces the structural information of the graph by restricting only the neighboring nodes to participate in the attention mechanism of the current node. An attention layer can only calculate the information between a node and its one-hop neighboring node. k Hop adjacent nodes, need to include k In order to make the sum of the attention coefficients of each node to all its neighboring nodes equal to 1, the Softmax function is used to normalize the attention coefficients of all neighboring nodes of the central node.

[0052] Since most graphs in the real world have the property of being scale-free, that is, the degree distribution of their nodes is uneven, the corporate group graph constructed based on the relationship between real-world corporate groups is no exception. A few nodes have a large number of edges, while most nodes have only a small number of edges, and the variance of the node degrees in the graph is large. Since single-head attention can only learn a specific relationship pattern between nodes, it may not be enough for scale-free graphs to fully capture the interaction pattern of nodes in the graph. This embodiment uses a multi-head attention mechanism to solve this problem. In the multi-head attention mechanism, each "head" can be regarded as an independent attention focus point, and multiple heads combined can capture different aspects of information from the data.

[0053] The multi-head attention mechanism can be realized by repeating the calculation of single-head attention multiple times. Aggregating the results of each head by splicing or averaging can obtain the node representation of the aggregation of neighborhood node information under the same meta-path semantics. The formula for updating the node representation using the multi-head attention mechanism is:

[0054] in, Nodes updated by the multi-head attention layer The representation vector of is the vector concatenation function, is the normalized node under the k-th attention With Node The attention weights between is the parameter matrix shared between nodes under the k-th attention head, is the number of heads of multi-head attention.

[0055] In the aggregation of semantic information between meta-paths, in order to learn the importance of different meta-paths for node representation and obtain a more comprehensive node embedding that integrates the semantics of each meta-path, a semantic-level attention mechanism is used. The semantic-level attention mechanism projects the embedding representation of each node under the meta-path into a new vector space through a nonlinear transformation layer, and then multiplies it by a learnable semantic attention vector. q , and the weighted sum of all nodes is divided by the number of nodes to average the attention weight, and each meta-path is obtained The attention weight , all meta-paths share these learnable parameters.

[0056]

[0057] in, For each meta-path The attention weight, is the set of all nodes in the graph, is the learnable semantic attention vector, represents the transpose operation of a vector, For the meta path The node after the multi-head attention layer update The representation vector of is the learnable parameter matrix, is the learnable bias parameter vector.

[0058] After obtaining the weight coefficient, the attention weights under different meta-paths are normalized by Softmax. The attention weight is expressed as (Attention weights for semantics between meta-paths). The higher the value, the more important the meta-path is.

[0059]

[0060] in, P is the total number of meta-paths in the heterogeneous graph of this scheme.

[0061] Finally, the attention coefficients of each meta-path are used Node-level embedding of the central node on each meta-path The final node embedding can be obtained by weighted summation (aggregating semantic information between meta-paths based on the attention weights of the semantics between meta-paths). (Node embedding representation vectors in a heterogeneous graph of business groups).

[0062]

[0063] As a preferred embodiment, the link predictor includes a node pair determination module, a similarity determination module and a result generation module; the node pair determination module is used based on the node pairs to be linked; the similarity determination module is used to calculate the similarity scores of the node pairs to be linked using the dot product method; the result generation module is used to generate link prediction results based on the similarity scores.

[0064] In the link prediction task in this embodiment, link prediction tasks are set for different types of edges, which are combined to form a multi-label link predictor. In terms of prediction method, for the node representation generated based on the aggregation of itself and heterogeneous neighborhood information in the heterogeneous graph of the enterprise group, the dot product method is used to calculate the similarity score between node pairs. The dot product is a similarity measurement method that can quantify the similarity between two node embeddings, thereby providing a probability estimate for the existence of potential links between them. The similarity score is used to evaluate whether a certain type of edge exists between node pairs. A high similarity score indicates that there is a greater probability that there is an edge between the two nodes, and a low similarity score indicates that there is a lower probability that there is an edge. Because different types of edges in heterogeneous graphs have different edge formation patterns, it is necessary to set independent link prediction tasks for each type of edge to learn separately. Calculation The type of existence between two nodes is The edge existence score calculation formula of the edge is:

[0065] in, For Node i The node embedding representation vector of For Node j The node embedding representation vector of .

[0066] As a preferred embodiment, the heterogeneous network relationships are supplemented according to the link prediction results, including: corresponding bidirectional link prediction results to node pairs with bidirectional semantics of the same type of relationship; when the bidirectional link prediction results are both predicted to exist, the heterogeneous network relationships represented by the node pairs are supplemented to the heterogeneous graph of the enterprise group.

[0067] Please refer to Figure 6 , Figure 6 This is a schematic diagram of the supplementary determination logic for inter-enterprise relationships provided by the present invention.

[0068] In order to make the relationship predicted as existing more accurate, the judgment logic of edge supplementation is set. For potential missing edges in the graph, the prediction results of the two-way links in the link prediction of two types of edges of the same relationship (corresponding to the node pairs of the same relationship with bidirectional semantics) are "logically ANDed". If both are predicted to exist, the relationship can be defined as existing. Finally, the relationship corresponding to the node pair is supplemented into the heterogeneous graph of the enterprise group.

[0069] The following is a description of the heterogeneous network relationship supplementation system between member enterprises of an enterprise group based on a meta-path provided by the present invention. The heterogeneous network relationship supplementation system between member enterprises of an enterprise group based on a meta-path described below and the heterogeneous network relationship supplementation method between member enterprises of an enterprise group based on a meta-path described above can be referenced to each other.

[0070] Please refer to Figure 7 , Figure 7 A schematic diagram of the structure of a meta-path-based heterogeneous network relationship supplement system among member enterprises of an enterprise group provided by the present invention.

[0071] The present invention also provides a system for supplementing heterogeneous network relationships between member enterprises of an enterprise group based on a meta-path, comprising: a construction module 1, used to construct a heterogeneous graph of an enterprise group; the heterogeneous graph of an enterprise group is used to characterize the heterogeneous network relationships between member enterprises of the enterprise group and the initial representation vectors of nodes; a sampling module 2, used to perform node neighborhood sampling based on the meta-path set in the heterogeneous graph of the enterprise group to obtain a neighborhood node set; a representation learning module 3, used to update the initial node representation vector based on the neighborhood node set using a heterogeneous graph representation learning method to obtain a node embedding representation vector; the node embedding representation vector is used to determine whether a node pair to be predicted for linking exists; a prediction module 4, used to input the node pair to be predicted for linking into a preset link predictor to obtain a link prediction result output by the link predictor; a supplementation module 5, used to supplement the heterogeneous network relationship according to the link prediction result.

[0072] Figure 8 An example of a physical structure diagram of an electronic device is shown in FIG. Figure 8 As shown, the electronic device may include: a processor 810, a communication interface 820, a memory 830 and a communication bus 840, wherein the processor 810, the communication interface 820 and the memory 830 communicate with each other through the communication bus 840. The processor 810 may call the logic instructions in the memory 830 to execute a method for supplementing heterogeneous network relationships between member enterprises of an enterprise group based on a meta-path, the method comprising: constructing a heterogeneous graph of an enterprise group; the heterogeneous graph of an enterprise group is used to characterize the heterogeneous network relationships and the initial node representation vectors between the member enterprises of the enterprise group; performing node neighborhood sampling based on the meta-path set of the heterogeneous graph of the enterprise group to obtain a neighborhood node set; based on the neighborhood node set, using a heterogeneous graph representation learning method to update the initial node representation vector to obtain a node embedding representation vector; the node embedding representation vector is used to determine whether a node pair to be predicted for linking exists; inputting the node pair to be predicted for linking into a preset link predictor to obtain a link prediction result output by the link predictor; and supplementing the heterogeneous network relationship according to the link prediction result.

[0073] In addition, the logic instructions in the above-mentioned memory 830 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0074] On the other hand, the present invention also provides a computer program product, which includes a computer program, which can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the meta-path-based heterogeneous network relationship supplement method between member enterprises of an enterprise group provided by the above-mentioned methods, and the method includes: constructing an enterprise group heterogeneous graph; the enterprise group heterogeneous graph is used to characterize the heterogeneous network relationship and the node initial representation vector between the member enterprises of the enterprise group; based on the meta-path set in the enterprise group heterogeneous graph, node neighborhood sampling is performed to obtain a neighborhood node set; based on the neighborhood node set, the node initial representation vector is updated using a heterogeneous graph representation learning method to obtain a node embedding representation vector; the node embedding representation vector is used to determine whether a node pair to be linked is linked; the node pair to be linked is input into a preset link predictor to obtain a link prediction result output by the link predictor; and the heterogeneous network relationship is supplemented according to the link prediction result.

[0075] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the meta-path-based heterogeneous network relationship supplementation method between member enterprises of an enterprise group provided by the above-mentioned methods, the method comprising: constructing an enterprise group heterogeneous graph; the enterprise group heterogeneous graph is used to characterize the heterogeneous network relationship and node initial representation vector between member enterprises of the enterprise group; performing node neighborhood sampling based on the meta-path set in the enterprise group heterogeneous graph to obtain a neighborhood node set; based on the neighborhood node set, using a heterogeneous graph representation learning method to update the node initial representation vector to obtain a node embedding representation vector; the node embedding representation vector is used to determine whether a node pair to be linked is linked; inputting the node pair to be linked to a preset link predictor to obtain a link prediction result output by the link predictor; and supplementing the heterogeneous network relationship according to the link prediction result.

[0076] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0077] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0078] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for supplementing heterogeneous network relationships among member enterprises of an enterprise group based on meta-path, characterized in that: include: Constructing a heterogeneous graph of an enterprise group; the heterogeneous graph of the enterprise group is used to characterize the heterogeneous network relationships and node initial representation vectors between member enterprises of the enterprise group; Perform node neighborhood sampling based on the meta-path set in the heterogeneous graph of the enterprise group to obtain a neighborhood node set; Based on the neighborhood node set, the node initial representation vector is updated by using a heterogeneous graph representation learning method to obtain a node embedding representation vector; the node embedding representation vector is used to determine whether a node pair to be linked is linked; Inputting the node pair to be link predicted into a preset link predictor, and obtaining the link prediction result output by the link predictor; The heterogeneous network relationship is supplemented according to the link prediction result.

2. The method for supplementing heterogeneous network relationships among member enterprises of an enterprise group based on meta-path according to claim 1 is characterized in that: The construction of the enterprise group heterogeneous graph includes: Identify the existing companies in the selected corporate group and the types of relationships between them; Determine an initial node representation vector corresponding to the existing enterprise through feature engineering; According to the relationship types between the enterprises, a bidirectional semantic edge is generated between the enterprises; the bidirectional semantic edge is used to represent the heterogeneous network relationship between the enterprises; The enterprise group heterogeneous graph is constructed based on the node initial representation vector and the bidirectional semantic edge.

3. The method for supplementing heterogeneous network relationships among member enterprises of an enterprise group based on meta-path according to claim 1, characterized in that: The node neighborhood sampling is performed based on the meta-path set in the enterprise group heterogeneous graph to obtain a neighborhood node set, including: Based on the meta-path guided random walk sampling strategy set in the heterogeneous graph of the enterprise group, a candidate neighborhood node sequence is obtained; A preset number of neighboring nodes are selected from the candidate neighboring node sequence to form the neighboring node set.

4. The method for supplementing heterogeneous network relationships among member enterprises of an enterprise group based on meta-path according to claim 1, characterized in that: The step of determining the node embedding representation vector in the heterogeneous graph of the enterprise group based on the neighborhood node set by using a heterogeneous graph representation learning method includes: Based on the neighborhood node set, determining the attention weights of the nodes in the meta-path through a node-level attention mechanism; Aggregating node information within the meta-path based on the attention weights of the nodes within the meta-path to obtain a node-level embedding representation vector; Based on the node-level embedding representation vector, determining the attention weight of the semantics between meta-paths through a semantic-level attention mechanism; Based on the attention weights of the semantics between the meta-paths, semantic information aggregation between the meta-paths is performed to obtain a node embedding representation vector in the heterogeneous graph of the enterprise group.

5. The method for supplementing heterogeneous network relationships among member enterprises of an enterprise group based on meta-path according to claim 1, characterized in that: The link predictor includes a node pair determination module, a similarity determination module and a result generation module; The node pair determination module is used to determine the node pairs based on the node pairs to be linked; The similarity determination module is used to calculate the similarity score of the node pair to be linked and predicted by using a dot product method; The result generation module is used to generate the link prediction result based on the similarity score.

6. The method for supplementing heterogeneous network relationships among member enterprises of an enterprise group based on meta-paths according to any one of claims 1 to 5, characterized in that: The supplementing of the heterogeneous network relationship according to the link prediction result includes: For node pairs with similar relationships and mutual bidirectional semantics, the corresponding bidirectional link prediction results are obtained; In the case where the two-way link prediction results are both predicted to exist relationships, the heterogeneous network relationships represented by the node pairs are added to the heterogeneous graph of the enterprise group.

7. A system for supplementing heterogeneous network relationships among member enterprises of an enterprise group based on meta-paths, characterized in that: include: A building module for constructing a heterogeneous graph of corporate groups; The enterprise group heterogeneous graph is used to characterize the heterogeneous network relationships and node initial representation vectors among the member enterprises of the enterprise group; A sampling module, used for sampling node neighborhoods based on the meta-paths set in the heterogeneous graph of the enterprise group to obtain a neighborhood node set; A representation learning module, used to update the initial node representation vector based on the neighborhood node set by using a heterogeneous graph representation learning method to obtain a node embedding representation vector; the node embedding representation vector is used to determine whether a node pair to be linked is linked; A prediction module, used for inputting the node pair to be link predicted into a preset link predictor, and obtaining the link prediction result output by the link predictor; A supplementing module is used to supplement the heterogeneous network relationship according to the link prediction result.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the method for supplementing heterogeneous network relationships among member enterprises of an enterprise group based on meta-paths as described in any one of claims 1 to 6 is implemented.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for supplementing heterogeneous network relationships among member enterprises of an enterprise group based on meta-paths as claimed in any one of claims 1 to 6 is implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the method for supplementing heterogeneous network relationships among member enterprises of an enterprise group based on meta-paths as claimed in any one of claims 1 to 6 is implemented.

Images