A multi-view-based graph neural network important node identification method

Abstract

The application discloses a kind of important node identification methods of multi-view based graph neural network, obtain user information from social platform;According to user information, all users are represented as nodes, the relationship between users is represented as edge, and the relevant attributes of users are represented as features, to establish network graph structure;Extension is multi-view, data is described from different angles, and the embedding of nodes is obtained using the method of graph neural network on different views;The relationship between nodes is represented using high-order topological structure matrix, and the final representation of nodes is obtained by using attention method to fuse multi-view information, to predict the importance of nodes, to obtain end-to-end training model, to identify important nodes;The application not only can improve the expression ability of model, but also can be generalized to data without obvious graph structure, extract new relationship, establish graph structure, use graph neural network model, and improve the performance of important node identification in network.

Description

TECHNICAL FIELD

[0001] The present application relates to the identification of important nodes in a network, mainly using the method of graph neural network, learning the representation of nodes from the perspective of multi-view, and then using multi-layer perceptron to obtain the importance score of the node. BACKGROUND

[0002] Network is a data form that describes the relationship between objects in the real world, such as neural system network, biological network, traffic network, social network, financial transaction network, etc. A large number of physicists have begun to study more complex networks in the real world, and have developed the research direction of complex networks. Researchers mainly study complex network models such as small-world network and scale-free network and their complex structures (such as community, hierarchy, node classification, etc.), dynamics on networks (such as information propagation, network evolution, etc.) and so on. In the past decade, thanks to the rapid development of machine learning, especially deep learning, researchers have begun to study networks from the perspective of machine learning.

[0003] At present, deep learning has become one of the most popular technologies in the field of artificial intelligence, and has shown the best performance in image, speech and natural language processing. Based on the universal approximation theory, the deep learning architecture can fit very complex functions in the real world, so the expression ability of deep learning model in extracting the underlying complex patterns of data is widely recognized. However, as a general form of representing relationship data, it is a difficulty for deep learning to process network data. How to use deep learning technology to analyze complex network data has attracted a large number of researchers.

[0004] The main challenges of extending traditional deep learning technology to network field are as follows. First, network is an irregular field. Unlike images and texts, which have clear grid structure, network data is in non-Euclidean space, and traditional convolution, pooling and other operations cannot be directly applied to network data. Second, scalability. In the era of big data, actual network data often has millions or even billions of scales, such as social networks and e-commerce transaction networks. Scalability is a key factor for the algorithm to be applied. Third, the variability of network data. Network can be undirected graph and directed graph, static and dynamic, signed and unsigned, homogeneous and heterogeneous, etc. Different data types require different model architectures. Finally, interdisciplinary. Network is often closely related to other disciplines such as chemistry and biology. It is more difficult to design a model combined with domain knowledge. For example, molecular graph generation, the objective function and chemical constraints are usually not differentiable, and cannot be trained by gradient descent.

[0005] Identifying important nodes is an important problem in network analysis. Traditionally, heuristic algorithms are used to identify important nodes based on the topology of the network, such as node degree, k-shell, closeness centrality, etc. Using graph neural network methods, the topology structure and node rich attribute information can be utilized simultaneously, and the powerful fitting ability of neural networks can be used to better identify important nodes. SUMMARY

[0006] The present application provides a method for identifying important nodes based on multi-view graph neural networks. The method uses graph neural networks to identify important nodes in a network. Potential application scenarios include crowdsourcing, where a company or organization outsources tasks that were previously performed by employees to a large number of volunteers. The method can be used to identify important volunteers in a social network and better complete tasks.

[0007] TECHNICAL SOLUTION

[0008] A method for identifying important nodes based on multi-view graph neural networks includes the following steps:

[0009] Step 1: Data collection phase, obtain user information from social platforms.

[0010] Step 2: Data preprocessing phase: preprocess user information, then represent all users as nodes, relationships between users as edges, and user attributes as features based on the preprocessed user information, and establish a network graph structure.

[0011] Step 3: Network embedding phase: based on the network graph structure obtained in step 2, extend to multi-view, describe data from different angles, and use graph neural network methods to obtain node embedding on different views.

[0012] Step 4: Calculate multi-view: use high-order topology structure matrix to represent the relationship between nodes.

[0013] Step 5: Important node identification phase: based on the high-order topology structure matrix representing the relationship between nodes obtained in step 4, use attention method to fuse multi-view information, obtain the final representation of nodes, predict the importance of nodes, obtain an end-to-end training model, and identify important nodes.

[0014] Preferably, in step 2, the network graph structure is established as follows:

[0015] Step 2a): fill in the missing attributes in the user information with 0.

[0016] Step 2b), all users are represented as nodes, and the relationship between users is represented as an edge to establish a graph structure.

[0017] Step 2c), the relevant attributes of the user are processed as the characteristics of the nodes in the network, and the network graph structure is obtained.

[0018] Preferably, the method for representing the relationship between nodes in step 4 using high-order topological structure matrix is as follows:

[0019] Step 4a), information propagation on the graph is a method to obtain the similarity between nodes, and the propagation can obtain the high-order topological structure similarity between node pairs. First, a personalized PageRank method is used to obtain a propagation matrix.

[0020] Step 4b), based on the feature vector matrix of the node, another view-feature similarity matrix of the graph data is calculated.

[0021] Step 4c), based on the network graph structure obtained in step 2, the propagation matrix obtained in step 4a), and the feature similarity matrix obtained in step 4b), three views are obtained, which respectively describe the relationship between nodes in the network from the aspects of local proximity, global proximity, and feature similarity.

[0022] Step 4d), the representations of the nodes obtained in step 4c) are obtained using a graph convolutional neural network.

[0023] Preferably, the method for identifying important nodes in step 5 is as follows:

[0024] Step 5a), a channel attention mechanism is used to fuse the representations of the nodes in the three views.

[0025] Step 5b), based on the representation of the node, a multilayer perceptron is used to calculate the importance score of the node.

[0026] Step 5c), based on the node importance score learned by the neural network, a ranking loss is calculated to obtain an end-to-end training model.

[0027] Step 5d), according to steps 5a)-5c), an end-to-end multi-view based graph neural network is constructed.

[0028] Preferably, the user information in step 1 includes age, gender, occupation, education background, fans, and attention.

[0029] Compared with the prior art, the present application has the following beneficial effects:

[0030] 1, compared with the traditional node recognition based on network topological information, the method of the present application can simultaneously use topological information and attribute information of the node.

[0031] 2, the multi-view based method fuses the multi-view information of network data, can obtain relatively complete data information, and describes the real data distribution. Meanwhile, the method based on channel attention can adaptively fuse the information of different views and learn more optimal node representation.

[0032] 3, the model is end-to-end trained and data-driven, and can more accurately identify important nodes in the network. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 The method of the present application is shown in the overall flowchart. DETAILED DESCRIPTION

[0034] The present application will be further illustrated below in combination with the drawings and specific embodiments, and it should be understood that these examples are only used to illustrate the present application and not to limit the scope of the present application. After reading the present application, those skilled in the art can make various equivalent modifications of the present application, which all fall within the scope defined by the appended claims.

[0035] An important node identification method based on multi-view graph neural network, as shown in Figure 1 , comprising the following steps:

[0036] Step 1, data acquisition stage, obtaining enough users from a social platform, extracting data of attributes such as age, gender, occupation, education background, fans, and attention.

[0037] Step 2, data preprocessing stage: preprocessing user information, then representing all users as nodes according to the preprocessed user information, representing the relationship between users as edges, representing the related attributes of users as features, and establishing a network graph structure.

[0038] Step 2a), 0 is added to the missing attributes in the user information.

[0039] Step 2b), all users are represented as nodes, and the relationship between users is represented as edges to establish a graph structure.

[0040] Step 2c), the related attributes of users are processed as features of nodes in the network, and the network graph structure is obtained.

[0041] Step 3, network embedding stage: based on the network graph structure obtained in step 2, it is expanded to multiple views to describe data from different angles, and the method of graph neural network is used on different views to obtain node embedding.

[0042] In view of the problem that the graph convolutional neural network as a shallow model can only learn local structure, a strategy of fusing multi-view information is proposed to improve the expression ability of the model.

[0043] Step 4, calculate multi-view: the adjacency matrix A is a representation of the relationship between nodes, the closer the distance between nodes, the closer the relationship between nodes, and the traditional GCN generally stacks two layers, which can only obtain the information of the second-order neighbor. On the graph, it is assumed that there are many paths between two nodes u and v, but the shortest path between the two nodes is greater than 2, although the relationship between the two nodes is close, but GCN cannot use the information of node v when obtaining the hidden representation of u. To this end, this paper proposes to use a high-order topological structure matrix to represent the relationship between nodes.

[0044] Preferably: the method for representing the relationship between nodes by using a high-order topological structure matrix in step 4 is as follows:

[0045] Step 4a), information propagation on the graph is a method to obtain the similarity between nodes, and propagation can obtain the high-order topological structure similarity between node pairs. First, use the personalized PageRank method to obtain the propagation matrix:

[0046] S PPR = α(I n -(1-α)D 1 / 2 AD 1 / 2 ) -1

[0047] Where S PPR represents the propagation matrix, α represents the coefficient, which is set to 0.15, I n represents the unit matrix, D represents the matrix whose diagonal elements are node degrees, and A represents the adjacency matrix.

[0048] Step 4b), based on the feature vector matrix of the node, another view-feature similarity matrix of the graph data is calculated.

[0049]

[0050] Where, represents the feature similarity between node i and node j, x i represents the feature of node i, x j represents the feature of node j, and n represents the number of nodes.

[0051] Step 4c), based on the network graph structure obtained in step 2, the propagation matrix obtained in step 4a), and the feature similarity matrix obtained in step 4b), three views are obtained, which respectively describe the relationship between nodes in the network from the aspects of local proximity, global proximity and feature similarity.

[0052] Step 4d), the three views obtained in step 4c) are used to obtain the representation of the node by using the graph convolutional neural network.

[0053] Step 5, important node identification stage: based on the high-order topological structure matrix obtained in step 4, the relationship between nodes is represented, the node embedding in three views is fused using a channel attention mechanism, the final representation of the node is obtained, the importance of the node is predicted, an end-to-end training model is obtained, and important nodes are identified.

[0054] Step 5a), the representations of nodes in three views are fused using a channel attention mechanism.

[0055] Step 5b), based on the representation of the node, the importance score of the node is calculated using a multilayer perceptron.

[0056] Step 5c), based on the node importance score learned by the neural network, a ranking loss is calculated to obtain an end-to-end training model.

[0057] Step 5d), according to steps 5a)-5c), an end-to-end multi-view based graph neural network is constructed.

[0058] In summary, the present application proposes a multi-view based graph neural network important node identification method. Compared with the traditional network important node identification method, the method based on the graph neural network simultaneously considers the network topological structure information and the node attribute information, and simultaneously aims at the graph convolutional neural network as a shallow model, can only learn to local structure problem, proposes a multi-view information fusion strategy, and improves the expression ability of the model.

[0059] The method based on the graph neural network of the present application identifies important nodes in the network, simultaneously considers the network topological structure information and the node attribute information, and simultaneously aims at the graph convolutional neural network as a shallow model, can only learn to local structure problem, proposes a multi-view information fusion strategy, and improves the expression ability of the model. Meanwhile, it can be generalized to data without obvious graph structure, extract new relationships, establish graph structure, use the graph neural network model, and improve the performance of important node identification in the network.

[0060] The above only describes the preferred embodiments of the present application, and it should be noted that: for ordinary skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, and these improvements and refinements should be considered as the protection scope of the present application.

Claims

1. A method for identifying important nodes based on a multi-view graph neural network, characterized in that, Includes the following steps: Step 1, Data Collection Phase: Obtain user information from social media platforms; Step 2, Data Preprocessing Stage: Preprocess the user information, and then represent all users as nodes, the relationships between users as edges, and the relevant attributes of users as features, and establish a network graph structure. Step 3, Network Embedding Stage: Based on the network graph structure obtained in Step 2, it is expanded into multiple views to describe the data from different perspectives. The embedding of nodes is obtained by using graph neural networks on different views. Step 4, Calculate multiple views: Use a high-order topology matrix to represent the relationships between nodes; Step 5, Important Node Identification Stage: Based on the high-order topology matrix obtained in Step 4, which represents the relationships between nodes, attention methods are used to fuse multi-view information to obtain the final node representation, predict the importance of nodes, obtain an end-to-end training model, and identify important nodes. In step 2: the method for establishing the network graph structure is as follows, Step 2a), pad missing attributes in the user information with 0; Step 2b) Represent all users as nodes and the relationships between users as edges to build a graph structure; Step 2c) The user's relevant attributes are processed into the features of nodes in the network to obtain the network graph structure; The method for representing the relationships between nodes using a higher-order topological matrix in step 4 is as follows: Step 4a) Propagate information on the graph to obtain the similarity between nodes. Propagation can obtain the higher-order topological similarity between node pairs. First, the propagation matrix is ​​obtained using the PageRank method. Step 4b) Based on the feature vector matrix of the nodes, calculate another view of the graph data - the feature similarity matrix; Step 4c) Based on the network graph structure obtained in Step 2, the propagation matrix obtained in Step 4a), and the feature similarity matrix obtained in Step 4b), three views are obtained, which describe the relationship between nodes in the network from the three aspects of local proximity, global proximity, and feature similarity, respectively. The three views obtained in steps 4d) and 4c) are used to obtain the node representations using graph convolutional neural networks.

2. The method for identifying important nodes based on a multi-view graph neural network according to claim 1, characterized in that: The method for identifying important nodes in step 5 is as follows: Step 5a) Use the channel attention mechanism to fuse the node representations from the three views; Step 5b): The importance score of a node is calculated using a multilayer perceptron based on the node representation. Step 5c): Based on the node importance scores learned by the neural network, calculate the ranking loss to obtain the end-to-end training model; Step 5d): Based on steps 5a)-5c), construct an end-to-end multi-view based graph neural network.

3. The method for identifying important nodes in a multi-view graph neural network according to claim 1, characterized in that: Step 1: User information includes age, gender, occupation, and educational background.

Images