Heterogeneous network node representation method and system based on biased walk of side information entropy

Abstract

The application discloses a heterogeneous network node representation method and system based on biased walk of edge information entropy, and comprises the following steps: receiving an input heterogeneous network containing multiple types of nodes and edges, obtaining an adjacency matrix and a degree matrix, and calculating an edge information entropy matrix; for each target node, normalizing the edge information entropy between the target node and different types of neighbor nodes; determining multiple meta-paths according to the semantic information composition of the heterogeneous network nodes; for each target node, obtaining the next node type according to the node sequence set by the meta-path, and performing biased sampling according to the normalized edge information entropy; taking the next node as the target node after obtaining the next node, and repeating the above steps until the set path length and walk times are reached, so that a walk path set of the node is finally obtained; finally, inputting the obtained node walk path set as the context semantic association sequence of the target node into a classical Skip-Gram model for training, so that the vector representation of each target node is finally obtained.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing and network embedding, and particularly relates to a heterogeneous network node representation method and system based on biased walk with edge information entropy. BACKGROUND

[0002] A graph is a very common data structure, which abstracts entities in the objective world as nodes in the network, and the association relationship between entities as edges in the network. Therefore, compared with structured Euclidean data, unstructured graph data has no restrictions on the form of the pre-defined data model, and is more flexible in expressing objective things, and contains rich information. At present, graph data has been widely used in representing social networks, protein interaction networks, and knowledge graphs in information retrieval, etc.

[0003] A heterogeneous graph, also known as a heterogeneous network, is a network with multiple different node types and connection relationships between nodes. Compared with a homogeneous network containing only a single node and edge type, a heterogeneous network reflects different semantic relationships through the connection behavior between different node types, thereby having stronger expression ability. Therefore, using comprehensive structural information and rich semantic information in a heterogeneous network helps to mine the essence of things hidden under the surface. At the same time, due to the characteristics of high complexity, structural diversity and information heterogeneity, there are still great challenges in analyzing and applying a heterogeneous network.

[0004] Graph representation learning, i.e. network embedding, is to project nodes in a network into a low-dimensional continuous space while preserving the network structure and inherent properties of nodes, in order to serve downstream machine learning tasks such as node classification [1], link prediction [2], community detection [3], etc. In recent years, due to the development of deep learning technology and the improvement of computing power, researchers have combined the random walk with restart [4] in network analysis with the classic Skip-Gram model in natural language processing to learn the embedding representation of network nodes, and representative methods such as DeepWalk [5], node2vec [6], etc. However, these methods can only handle homogeneous networks with single type of nodes and edges.

[0005] For the characteristics of heterogeneous network, researchers define meta-paths to explore the different semantics implied in these paths by limiting the connections of different types of paths. The meta-path-based heterogeneous network embedding method, such as metapath2vec[7], usually needs to define a set of meta-paths in advance, and then randomly sample the neighbor nodes of the current node under the constraint of the meta-path, so as to generate a series of node sequences with specific semantic information, that is, meta-path instances. Finally, the Skip-Gram model is used to train the embedding representation of the nodes. Although this method can effectively improve the semantic mining performance in heterogeneous networks, it uses a random sampling strategy when generating node sequences, only referring to the node types in the meta-path, and does not consider the edge attributes in the heterogeneous network, thus causing large semantic differences between meta-path instances due to edge attribute bias, which ultimately adversely affects the performance of node embedding representation. Therefore, a heterogeneous network node representation method based on biased walk of edge information entropy is proposed. The edge information entropy[8] can capture more critical information in the heterogeneous network, and it is used as the edge attribute value for meta-path-based node biased sampling, so that the generated node sequence is biased towards the neighbor nodes with richer semantics in the heterogeneous network, thereby improving the expression ability of node embedding representation for key information in the heterogeneous network.

[0006] 1. Kipf, T. N., & Welling, M. Semi-supervised classification with graph convolutional networks. International Conference on Learning Representations 2017.

[0007] 2. Wang, D., Cui, P., Zhu, W., & Yang, S. Structural Deep Network Embedding. International Conference on Knowledge Discovery and Data Mining 2016, 1225-1234.

[0008] 3. Fortunato, S. Community detection in graphs. Physics reports 2010, 486, (3-5), 75-174.

[0009] 4. Tong, H., Faloutsos, C., & Pan, J. Y. Random walk with restart: Fast solutions and applications. International Conference on Data Mining 2006, 613-622.

[0010] 5. Perozzi, B., Al-Rfou, R., & Skiena, S. DeepWalk: Online learning of social representations. International conference on Knowledge discovery and data mining 2014, 701-710.

[0011] 6. Grover, A., & Leskovec, J. Node2vec: Scalable feature learning for networks. International Conference on Knowledge Discovery and Data Mining 2016, 855-864.

[0012] 7. Dong, Y., Chawla, N. V., & Swami, A. Metapath2vec: Scalable representation learning for heterogeneous networks International Conference on Knowledge Discovery and Data Mining 2017, 135-144.

[0013] 8. Xu, Z., Pu, C., & Yang, J. Link prediction based on path entropy Physica A: Statistical Mechanics and its Applications 2016, 456, 294-301. SUMMARY

[0014] The application aims to overcome the problem that the traditional random sampling method in the prior art heterogeneous network representation learning method can ignore the edge attribute difference in the heterogeneous network, and proposes a heterogeneous network node representation method and system based on biased walk of edge information entropy, fully considers the influence of the edge attribute difference on semantic information mining, makes the walk path of the node biased to the neighbor node with richer semantic information in the heterogeneous network, thereby enhances the modeling ability of the relationship between different types of nodes in the heterogeneous network, captures the key information in the network, and improves the expression ability of the node embedding representation to the key information of the heterogeneous network.

[0015] The application adopts the following technical solutions:

[0016] In one aspect, a heterogeneous network node representation method based on biased walk of edge information entropy comprises:

[0017] An adjacency matrix and a degree matrix obtaining step S101 receives an input heterogeneous network G containing multiple types of nodes and edges,

[0018] obtains an adjacency matrix A and a degree matrix D of the heterogeneous network;

[0019] An edge information entropy matrix obtaining step S102 calculates an edge information entropy matrix EI of the heterogeneous network by using the adjacency matrix A and the degree matrix D.

[0020]

[0021] A normalization processing step S103 performs normalization on the edge information entropy between each target node v and neighbor nodes of different types according to the edge information entropy matrix EI, and stores the normalized edge information entropy in a dictionary form, taking a neighbor node set of a certain type of the target node as the key of the dictionary and the normalized edge information entropy as the value of the dictionary, to obtain a normalized edge information entropy dictionary .

[0022] A biased sampling step S104 determines a plurality of meta-paths according to the semantic information composition of the heterogeneous network nodes; for each target node, the normalized edge information entropy dictionary is queried according to the next node type given by the meta-path, and the Alias method is used to perform biased sampling on the edge list of the next node type to obtain the next node.

[0023] A walk path set obtaining step S105 takes the node obtained by the biased sampling as a target node, and repeats the biased sampling step until a set path length and a set walk number are reached, to finally obtain a walk path set of the target node.

[0024] A node vector representation step S106 inputs the walk path set of the target node as a context semantic association sequence of the target node into a classical Skip-Gram model for training, and finally obtains a vector representation of each target node.​

[0025] Preferably, the S101 specifically comprises:

[0026] Given a heterogeneous network , the number of nodes ; for each node , there is a mapping function and , where denotes the node type set of the heterogeneous network ; for each edge , there is a mapping function and , where denotes the edge type set of the heterogeneous network ; the adjacency matrix and the degree matrix of the heterogeneous network are defined as follows:

[0027] (1)

[0028] (2)

[0029] where, and denote nodes; denotes the node set; denotes the edge set; denotes the sum of each column of the matrix A.

[0030] Preferably, the S102 specifically comprises:

[0031] The probability that there is no edge between node and node is:

[0032] (3)

[0033] where, and are the degrees of node and node ; and is the total number of edges of the heterogeneous network ; and denotes the number of edge sets connected to node of the heterogeneous network ; and denotes the node connected to node in the heterogeneous network​ Number of disconnected edge sets The possibility of no connection between node and node is estimated

[0034] The existence of an edge between node and node is calculated by formula (3), i.e. :

[0035] (4)

[0036] From the information entropy theory, given an event and its occurrence probability , its information entropy is defined as:

[0037] (5)

[0038] Therefore, the edge information entropy matrix of the heterogeneous network is calculated by formula (3)-(5) as follows:

[0039] (6).

[0040] Preferably, the S103 specifically comprises:

[0041] For a target node in the graph, its set of the first type of neighbor nodes is denoted as , ; for each type of neighbor node, the edge information entropy is normalized respectively to obtain the normalized edge information entropy between the node and the first type of neighbor node , and the calculation formula is:

[0042] (7)

[0043] The value also represents the transition probability of the target node and the first type of neighbor node , i.e. ; in order to facilitate the biased walk in the next step, a dictionary is used to save the matrix, and a set of neighbor nodes of a certain type of the target node is taken as the key of the dictionary and the normalized edge information entropy is taken as the value of the dictionary represents the information entropy between node and node ​Represents a node Its first Neighbor-like nodes The information entropy between them.

[0044] Preferably, S104 specifically includes:

[0045] Based on practical application needs, multiple meta-paths to the target node are determined; a meta-path refers to a path in a graph composed of multiple node types and edge types; meta-path Defined as a sequence of alternating entity types and relations, as follows:

[0046]

[0047] in, Metapath The type of node in the middle, Indicates the type of edge in the metapath; a path that passes through The path of a node is an instance of a metapath;

[0048] Based on the node type defined in the metapath, the normalized edge information entropy calculated in S103 is used as the transition probability of the node. The Alias ​​method is then used to generate the node sequence sequentially, where the... The transition probabilities of each step are as follows:

[0049]

[0050] in, This indicates that the node type in the metapath is The node, Represents a node Jump to node type Neighbors The probability, Represents a node The node type.

[0051] Preferably, step S105 specifically includes:

[0052] Different types of meta-paths represent different topologies and semantic information in heterogeneous networks; multiple meta-paths are defined based on prior knowledge for heterogeneous networks. For each target node, repeat step S104 until the set path length and number of traversals are reached, and finally obtain the set of traversal paths for the target node.

[0053] Preferably, S106 specifically includes:

[0054] The set of walk paths in S105 is used as input to the Skip-Gram model. The weight matrix is ​​updated through gradient descent and backpropagation algorithms to learn node embeddings. The basic idea of ​​the Skip-Gram model is that, given a fixed-length text sequence, a sliding window is used to obtain several words before and after the center word as the target. Then, for each context word, a training sample containing a single center word and a context word is constructed. The likelihood function is optimized by using gradient descent with negative sampling to learn the feature representation of the words.

[0055] The objective function of the Skip-Gram model uses the negative log-likelihood loss function, defined as:

[0056]

[0057] in, As a positive sample, For negative samples, These are model parameters. It is the Sigmoid function. It is a node Embedded vector, It is a node Embedded vector, It is a node The embedding vector.

[0058] On the other hand, a heterogeneous network node representation system based on biased walks with edge information entropy includes:

[0059] The adjacency matrix and degree matrix acquisition module is used to receive input heterogeneous network G containing various types of nodes and edges.

[0060] Obtain the adjacency matrix A and degree matrix D of the heterogeneous network;

[0061] The edge information entropy matrix acquisition module is used to calculate the edge information entropy matrix of heterogeneous networks using the adjacency matrix A and the degree matrix D.

[0062] EI array;

[0063] The normalization module, based on the edge information entropy matrix EI, normalizes the edge information entropy between each target node v and its neighbor nodes of different types, and stores it in dictionary form. The dictionary uses a set of neighbor nodes of a certain type as the key and the normalized edge information entropy as the value, resulting in a normalized edge information entropy dictionary. ;

[0064] The biased sampling module is used to determine multiple meta-paths based on the semantic information of heterogeneous network nodes. For each target node, based on the next node type given by the meta-path, the normalized edge information entropy dictionary is queried, and the Alias ​​method is used to perform biased sampling on the edge list of the next node type to obtain the next node.

[0065] The path set acquisition module is used to take the nodes obtained by biased sampling as target nodes, and repeatedly execute the biased sampling module until the set path length and number of traversals are reached, and finally obtain the path set of the target node.

[0066] The node vector representation module is used to take the set of traversal paths of the target node as the contextual semantic association sequence of the target node, input it into the classic Skip-Gram model for training, and finally obtain the vector representation of each target node.

[0067] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0068] (1) This invention uses edge information entropy as the edge weight of heterogeneous network, and assigns higher weight to the edges that carry more information in the heterogeneous network, thereby making it easier to mine key semantic features in the network.

[0069] (2) The present invention uses multiple meta-paths to traverse and obtains semantic information of different meta-paths, which increases the scalability of node information;

[0070] (3) Based on edge information entropy matrix sampling, the present invention makes the node's traversal biased towards neighboring nodes with richer semantics in the heterogeneous network, thereby improving the node embedding representation's ability to express key information of the heterogeneous network. Attached Figure Description

[0071] Figure 1 This is a flowchart of a heterogeneous network node representation method based on biased walk with edge information entropy, according to an embodiment of the present invention.

[0072] Figure 2 This is a diagram showing the relationship between different types of nodes in DBLP data according to an embodiment of the present invention.

[0073] Figure 3 This invention presents the results of node classification analysis of node embedding representations in DBLP data using Macro-F1 and accuracy scores; wherein, (a) represents the results of node classification analysis of node embedding representations using Macro-F1 scores; and (b) represents the results of node classification analysis of node embedding representations using accuracy scores.

[0074] Figure 4This is the result of a visualization clustering analysis of node embedding representations in DBLP data using the t-SNE algorithm, as described in an embodiment of the present invention.

[0075] Figure 5 This is a structural block diagram of a heterogeneous network node representation system based on biased walks with edge information entropy, according to an embodiment of the present invention. Detailed Implementation

[0076] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0077] See Figure 1 As shown in the figure, this embodiment presents a method for representing heterogeneous network nodes based on biased walks with edge information entropy, which includes the following steps.

[0078] Step S101, obtaining the adjacency matrix and degree matrix, involves receiving the input heterogeneous network G containing various types of nodes and edges.

[0079] Obtain the adjacency matrix A and degree matrix D of the heterogeneous network.

[0080] Specifically, given a heterogeneous network Number of nodes For each node There exists a mapping function and ,in Representing heterogeneous networks The set of node types; for each edge There exists a mapping function and ,in Representing heterogeneous networks A set of edge types; heterogeneous networks adjacency matrix Sum-degree matrix The definition is as follows:

[0081] (1)

[0082] (2)

[0083] in, and Represents a node; Represents a set of nodes; Represents the set of edges; This represents the summation over each column of matrix A.

[0084] In this embodiment, heterogeneous network Taking the existing DBLP (Database systems and Logic Programming) network as an example, this dataset has 15,649 nodes and 51,363 edges, including four types of nodes: 5,915 author nodes (Class A), 5,237 paper nodes (Class P), 4,479 topic nodes (Class T), and 18 site nodes (Class V). See also Figure 2 As shown, this includes not only isomorphic edges between papers, but also heteromorphic edges between authors and papers, papers and sites, and papers and topics.

[0085] Step S102, obtaining the edge information entropy matrix, involves calculating the edge information entropy matrix of the heterogeneous network using the adjacency matrix A and the degree matrix D.

[0086] EI array.

[0087] Specifically, computing nodes and nodes There is no edge between them, that is probability for:

[0088] (3)

[0089] in, and It is a node and nodes The degree; It is a heterogeneous network The total number of sides; Representing heterogeneous networks With nodes The number of connected edge sets; Representing heterogeneous networks In and nodes Connect, and simultaneously with nodes The number of sets of disconnected edges; Nodes were estimated and nodes There is no possibility of connection between them;

[0090] The nodes are calculated using formula (3). and nodes There are edges between them, that is :

[0091] (4)

[0092] According to the theory of information entropy, given an event... and its probability of occurrence Its information entropy is defined as:

[0093] (5)

[0094] Therefore, heterogeneous networks can be calculated using formulas (3)-(5). The edge information entropy matrix is ​​as follows:

[0095] (6).

[0096] In the normalization process S103, based on the edge information entropy matrix EI, for each target node v, the edge information entropy between it and its neighbor nodes of different types is normalized and stored in dictionary form. The set of neighbor nodes of a certain type for the target node is used as the key of the dictionary, and the normalized edge information entropy is used as the value, resulting in the normalized edge information entropy dictionary. .

[0097] Specifically, for a target node in the graph Its first The set of neighbor nodes is represented as , For each type of neighbor node, the edge information entropy is normalized to obtain the node. With the Neighbor-like nodes Normalized edge information entropy between The calculation formula is:

[0098] (7)

[0099] This value also represents the target node. With the Neighbor-like nodes The transition probability, i.e. To facilitate the next biased walk, data is stored in dictionary format. Matrix, target node The set of neighbor nodes of a certain type is used as the key of the dictionary, and the normalized edge information entropy is used as the value of the dictionary; Represents a node and nodes Information entropy between them; Represents a node and nodes The information entropy between them.

[0100] In the biased sampling step S104, multiple meta-paths are determined according to the semantic information composition of heterogeneous network nodes. For each target node, the normalized edge information entropy dictionary is queried according to the next node type given by the meta-path, and the Alias ​​method is used to perform biased sampling on the edge list of the next node type to obtain the next node.

[0101] Specifically, based on the actual application needs, multiple meta-paths to the target node are determined; a meta-path refers to a path in a graph composed of multiple node types and edge types; meta-path Defined as a sequence of alternating entity types and relations, as follows:

[0102]

[0103] in, Metapath The type of node in the middle, Indicates the type of edge in the metapath; a path that passes through The path of a node is an instance of a metapath;

[0104] Based on the node type defined in the metapath, the normalized edge information entropy calculated in S103 is used as the transition probability of the node. The Alias ​​method is then used to generate the node sequence sequentially, where the... The transition probabilities of each step are as follows:

[0105]

[0106] in, This indicates that the node type in the metapath is The node, Represents a node Jump to node type Neighbors The probability, Represents a node The node type.

[0107] In this embodiment, based on the DBLP dataset, two metapaths, "APA" and "APVPA," are set. For each target node, the normalized edge information entropy dictionary is queried according to the next node type given by the metapath. The Alias ​​method is used to perform biased sampling on the edge list of the next node type to obtain the next node.

[0108] In step S105, the nodes obtained by biased sampling are used as target nodes. The biased sampling steps are repeated until the set path length and number of traversals are reached, and finally the traversal path set of the target node is obtained.

[0109] Specifically, different types of meta-paths represent different topologies and semantic information in heterogeneous networks; multiple meta-paths are defined based on prior knowledge for heterogeneous networks. For each target node, repeat step S104 until the set path length and number of traversals are reached, and finally obtain the set of traversal paths for the target node.

[0110] In this embodiment, the node sampled in S104 is used as the target node, and S104 is repeated until the set path length is reached. and number of walks Finally, the set of traversal paths to the target node is obtained. .

[0111] In step S106, the set of traversal paths of the target node is used as the contextual semantic association sequence of the target node and input into the classic Skip-Gram model for training, so as to obtain the vector representation of each target node.

[0112] Specifically, the set of walk paths in S105 is used as input to the Skip-Gram model. The weight matrix is ​​updated through gradient descent and backpropagation algorithms to learn node embeddings. The basic idea of ​​the Skip-Gram model is that, given a fixed-length text sequence, a sliding window is used to obtain several words before and after the center word as the target. Then, for each context word, a training sample containing a single center word and a context word is constructed. The likelihood function is optimized by using gradient descent with negative sampling to learn the feature representation of the words.

[0113] The objective function of the Skip-Gram model uses the negative log-likelihood loss function, defined as:

[0114]

[0115] in, As a positive sample, For negative samples, These are model parameters. It is the Sigmoid function. It is a node Embedded vector, It is a node Embedded vector, It is a node The embedding vector.

[0116] Finally, the effectiveness of the node vector representation in this embodiment was verified through the following experiments.

[0117] In this embodiment, the DBLP network has 1,909 labeled author nodes (Class A nodes), which can be divided into four categories: database, data mining, machine learning, and information retrieval. Based on the obtained node vector representations, the labeled author nodes (Class A nodes) in the DBLP network are used to perform node classification and clustering analysis tasks.

[0118] For the node classification task, a set of nodes is randomly selected as labeled nodes for training, and the rest are used for testing. Based on the embedding representations obtained from the training nodes, a logistic regression classifier is trained to predict the most likely label of the test nodes, and the predicted labels are compared with their true labels to obtain Macro-F1 and accuracy scores. The average Macro-F1 and accuracy scores are calculated for 50 repeated trials with training set proportions of 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, and 90%.

[0119] For node clustering analysis tasks, the t-SNE algorithm is used to reduce the node embedding representation to two dimensions, and then the nodes are visualized after being labeled with different colors according to different categories.

[0120] See Figure 3 As shown, the results for the node classification task are presented, where (a) represents the results of node classification analysis using the Macro-F1 score on the node embedding representation; and (b) represents the results of node classification analysis using the accuracy score on the node embedding representation. It is clear from the classification results that more training data generally leads to higher performance, and the scores are consistently better. See also... Figure 4 The results shown are for the node clustering analysis task. It can be seen that the visualization of the obtained node embedding representations effectively displays the node categories.

[0121] As can be seen from the above verification, the heterogeneous network node representation method based on biased walk with edge information entropy described in this invention performs well in both node classification and clustering analysis tasks.

[0122] See Figure 5 As shown, this embodiment also discloses a heterogeneous network node representation system based on biased walks with edge information entropy, including:

[0123] The adjacency matrix and degree matrix acquisition module 501 is used to receive input heterogeneous network G containing various types of nodes and edges.

[0124] Obtain the adjacency matrix A and degree matrix D of the heterogeneous network;

[0125] The edge information entropy matrix acquisition module 502 is used to calculate the edge information of heterogeneous networks using the adjacency matrix A and the degree matrix D.

[0126] Entropy matrix EI;

[0127] The normalization processing module 503 is used to normalize the edge information entropy between each target node v and its neighbor nodes of different types based on the edge information entropy matrix EI, and store it in dictionary form. The set of neighbor nodes of a certain type of target node is used as the key of the dictionary, and the normalized edge information entropy is used as the value of the dictionary, thus obtaining the normalized edge information entropy dictionary. ;

[0128] The biased sampling module 504 is used to determine multiple meta-paths based on the semantic information of heterogeneous network nodes. For each target node, the normalized edge information entropy dictionary is queried according to the next node type given by the meta-path, and the Alias ​​method is used to perform biased sampling on the edge list of the next node type to obtain the next node.

[0129] The traversal path set acquisition module 505 is used to take the nodes obtained by biased sampling as target nodes, repeatedly execute the biased sampling module until the set path length and number of traversals are reached, and finally obtain the traversal path set of the target node.

[0130] The node vector representation module 506 is used to take the set of walking paths of the target node as the context semantic association sequence of the target node, input it into the classic Skip-Gram model for training, and finally obtain the vector representation of each target node.

[0131] For the specific implementation of each module in a heterogeneous network node representation system based on edge information entropy biased walk, please refer to a heterogeneous network node representation method based on edge information entropy biased walk; this embodiment will not repeat the description.

[0132] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for representing nodes in heterogeneous networks based on biased walks with edge information entropy, characterized in that, This technology is applied to data processing and network embedding to address the problem of insufficient semantic information mining in heterogeneous networks due to differences in edge attributes. Nodes include author nodes, paper nodes, topic nodes, and site nodes; edges include isomorphic edges between papers, as well as heterogeneous edges between authors and papers, papers and sites, and papers and topics. The methods include: In step S101, the adjacency matrix and degree matrix are obtained by receiving the input heterogeneous network G containing multiple types of nodes and edges, and obtaining the adjacency matrix A and degree matrix D of the heterogeneous network. Step S102 for obtaining the edge information entropy matrix: Using the adjacency matrix A and the degree matrix D, calculate the edge information entropy matrix EI of the heterogeneous network. In the normalization process S103, based on the edge information entropy matrix EI, for each target node v, the edge information entropy between it and its neighbor nodes of different types is normalized and stored in dictionary form. The set of neighbor nodes of a certain type for the target node is used as the key of the dictionary, and the normalized edge information entropy is used as the value, resulting in the normalized edge information entropy dictionary. ; In the biased sampling step S104, multiple meta-paths are determined according to the semantic information composition of heterogeneous network nodes. For each target node, the normalized edge information entropy dictionary is queried according to the next node type given by the meta-path, and the Alias ​​method is used to perform biased sampling on the edge list of the next node type to obtain the next node. Step S105, which obtains the set of walking paths, takes the node obtained by biased sampling as the target node and repeats the biased sampling step until the set path length and number of walks are reached, and finally obtains the set of walking paths of the target node. In step S106, the set of traversal paths of the target node is used as the contextual semantic association sequence of the target node and input into the classic Skip-Gram model for training, so as to obtain the vector representation of each target node.

2. The heterogeneous network node representation method based on biased walk with edge information entropy according to claim 1, characterized in that, S101 specifically includes: Given a heterogeneous network Number of nodes For each node There exists a mapping function and ,in Representing heterogeneous networks The set of node types; for each edge There exists a mapping function and ,in Representing heterogeneous networks A set of edge types; heterogeneous networks adjacency matrix Sum-degree matrix The definition is as follows: (1) (2) in, and Represents a node; Represents a set of nodes; Represents the set of edges; This represents the summation over each column of matrix A.

3. The heterogeneous network node representation method based on biased walk with edge information entropy according to claim 2, characterized in that, S102 specifically includes: compute nodes and nodes There is no edge between them, that is probability for: (3) in, and It is a node and nodes The degree; It is a heterogeneous network The total number of sides; Representing heterogeneous networks With nodes The number of connected edge sets; Representing heterogeneous networks In and nodes Connect, and simultaneously with nodes The number of sets of disconnected edges; Nodes were estimated and nodes There is no possibility of connection between them; The nodes are calculated using formula (3). and nodes There are edges between them, that is : (4) According to the theory of information entropy, given an event... and its probability of occurrence Its information entropy is defined as: (5) Therefore, heterogeneous networks can be calculated using formulas (3)-(5). The edge information entropy matrix is ​​as follows: (6)。 4. The heterogeneous network node representation method based on biased walk with edge information entropy according to claim 3, characterized in that, S103 specifically includes: For a target node in the graph Its first The set of neighbor nodes is represented as , For each type of neighbor node, the edge information entropy is normalized to obtain the node. With the Neighbor-like nodes Normalized edge information entropy between The calculation formula is: (7) This value also represents the target node. With the Neighbor-like nodes The transition probability, i.e. To facilitate the next biased walk, data is stored in dictionary format. Matrix, target node The set of neighbor nodes of a certain type is used as the key of the dictionary, and the normalized edge information entropy is used as the value of the dictionary; Represents a node With nodes Information entropy between them; Represents a node Its first Neighbor-like nodes Information entropy between them.

5. The heterogeneous network node representation method based on biased walk with edge information entropy according to claim 4, characterized in that, S104 specifically includes: Based on practical application needs, multiple meta-paths to the target node are determined; a meta-path refers to a path in a graph composed of multiple node types and edge types; meta-path Defined as a sequence of alternating entity types and relations, as follows: ； in, Metapath The type of node in the middle, Indicates the type of edge in the metapath; a path that passes through The path of a node is an instance of a metapath; Based on the node type defined in the metapath, the normalized edge information entropy calculated in S103 is used as the transition probability of the node. The Alias ​​method is then used to generate the node sequence sequentially, where the... The transition probabilities of each step are as follows: ； in, This indicates that the node type in the metapath is The node, Represents a node Jump to node type Neighbors The probability, Represents a node The node type.

6. The heterogeneous network node representation method based on biased walk with edge information entropy according to claim 1, characterized in that, S105 specifically includes: Different types of meta-paths represent different topologies and semantic information in heterogeneous networks; multiple meta-paths are defined based on prior knowledge for heterogeneous networks. For each target node, repeat step S104 until the set path length and number of traversals are reached, and finally obtain the set of traversal paths for the target node.

7. The heterogeneous network node representation method based on biased walk with edge information entropy according to claim 1, characterized in that, S106 specifically includes: The set of walk paths in S105 is used as input to the Skip-Gram model. The weight matrix is ​​updated through gradient descent and backpropagation algorithms to learn node embeddings. The basic idea of ​​the Skip-Gram model is that, given a fixed-length text sequence, a sliding window is used to obtain several words before and after the center word as the target. Then, for each context word, a training sample containing a single center word and a context word is constructed. The likelihood function is optimized by using gradient descent with negative sampling to learn the feature representation of the words. The objective function of the Skip-Gram model uses the negative log-likelihood loss function, defined as: ； in, As a positive sample, For negative samples, These are model parameters. It is the Sigmoid function. It is a node Embedded vector, It is a node Embedded vector, It is a node The embedding vector.

8. A heterogeneous network node representation system based on biased walks with edge information entropy, characterized in that, This system, applied in the fields of data processing and network embedding technology, addresses the problem of insufficient semantic information mining in heterogeneous networks due to differences in edge attributes. Nodes include author nodes, paper nodes, topic nodes, and site nodes. Edges include isomorphic edges between papers, as well as heterogeneous edges between authors and papers, papers and sites, and papers and topics. The system comprises: The adjacency matrix and degree matrix acquisition module is used to receive input heterogeneous network G containing various types of nodes and edges. Obtain the adjacency matrix A and degree matrix D of the heterogeneous network; The edge information entropy matrix acquisition module is used to calculate the edge information entropy matrix of heterogeneous networks using the adjacency matrix A and the degree matrix D. EI array; The normalization module, based on the edge information entropy matrix EI, normalizes the edge information entropy between each target node v and its neighbor nodes of different types, and stores it in dictionary form. The dictionary uses a set of neighbor nodes of a certain type as the key and the normalized edge information entropy as the value, resulting in a normalized edge information entropy dictionary. ; The biased sampling module is used to determine multiple meta-paths based on the semantic information of heterogeneous network nodes. For each target node, based on the next node type given by the meta-path, the normalized edge information entropy dictionary is queried, and the Alias ​​method is used to perform biased sampling on the edge list of the next node type to obtain the next node. The path set acquisition module is used to take the nodes obtained by biased sampling as target nodes, and repeatedly execute the biased sampling module until the set path length and number of traversals are reached, and finally obtain the path set of the target node. The node vector representation module is used to take the set of traversal paths of the target node as the contextual semantic association sequence of the target node, input it into the classic Skip-Gram model for training, and finally obtain the vector representation of each target node.

Images