Dynamic social network influence propagation maximization method based on deep learning and structure evolution agency

Abstract

The invention belongs to the technical field of digital economy and cyberspace security governance, and particularly relates to a dynamic social network influence propagation maximization method based on deep learning and structure evolution agency, which comprises the following steps: collecting social network data, modeling into a dynamic social network through a user relationship, and according to time intervals of different requirements, establishing a dynamic social network model; dividing the network into a plurality of time network snapshots; designing a network structure evolution proxy method, and selecting part of nodes as structure evolution proxy nodes according to the change of network topology and a node selection strategy; extracting a local topological structure of an agent node by using random diffusion and a convolutional neural network, and forming node embedded vector representation; a historical node embedding vector is updated by utilizing information increment of an agent node through a graph convolutional neural network, a node influence propagation expectation value is predicted, and a group of nodes with the maximum influence are selected as a seed set of a current network snapshot according to an influence value predicted by a model; and repeating the steps, selecting a group of seed sets for the whole dynamic network, and providing a decision basis for social network influence propagation and control.

Description

Technical Field

[0001] The present invention belongs to the technical field of digital economy and cyberspace security governance, and specifically is a method for maximizing the spread of influence in dynamic social networks based on deep learning and structural evolution agents. Background Art

[0002] The problem of maximizing social network influence aims to select a group of key users so that information, products or ideas can be widely spread in social networks to form economies of scale. Traditional research is mostly based on a single static network model, assuming that the network structure remains unchanged throughout the entire communication process. However, in reality, social networks change dynamically with time factors, and user behavior and connection relationships evolve dynamically. Therefore, static models often cannot fully describe the actual network communication situation.

[0003] The analysis and optimization of the problem of maximizing the spread of influence in dynamic social networks is one of the difficulties in the current field of social network analysis. In dynamic networks, the network structure and the influence of users evolve with the time factor, and it is necessary to consider the impact of the time dimension on network evolution and control optimization. Solving the problem of maximizing the spread of influence in dynamic social networks requires not only selecting the best seed users at different time points, but also adapting to the continuous evolution of the network to ensure the breadth and depth of information dissemination. At present, the effectiveness of existing methods has certain limitations. For example, the efficiency of the algorithm for selecting seeds at different time points is still an important bottleneck problem. While ensuring the effectiveness of the seed set, how to improve the utilization rate of network historical information and reduce repeated calculations is an urgent problem to be solved.

[0004] Deep learning is based on artificial neural networks and provides innovative solutions to practical problems by automatically learning features and patterns from large amounts of data. Using deep learning to solve the problem of maximizing influence propagation in social networks has significant advantages, mainly reflected in its powerful feature extraction capabilities, ability to handle complex data structures, and efficient prediction and decision-making capabilities. In particular, convolutional neural networks and graph convolutional neural networks can automatically learn useful features and patterns from large-scale data, which makes them uniquely advantageous in solving complex network problems. Traditional methods often rely on manually designed features, which is not only time-consuming and labor-intensive, but may also miss key features. Deep learning algorithms can process complex graph structure data and capture the complex relationships between nodes and network structure characteristics. Convolutional neural networks can automatically extract multi-level features from raw data through convolution operations. In the problem of maximizing influence, graph convolutional neural networks can extract complex features from user behavior data, social media content, and image information to capture the potential influence relationships between users.

[0005] Therefore, the existing technology requires a method that can effectively capture the relationship between nodes and influence in social networks and dynamically track influential nodes to better solve the problem of maximizing influence in dynamic social networks. Summary of the invention

[0006] In order to solve the above problems, the present invention provides a method for maximizing the spread of influence in dynamic social networks based on deep learning and structural evolution agent, which aims to select a group of the most influential seed users to maximize the spread of influence in a dynamically evolving social network.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] A method for maximizing the spread of influence in a dynamic social network based on deep learning and structural evolution agent, characterized by comprising the following steps:

[0009] Step 1: Collect social network data and model it into a dynamic social network graph through user relationships. Divide the network into several time network snapshots according to different time intervals.

[0010] Step 2: Design a structure evolution agent method, select some nodes as structure evolution agent nodes according to the changes in network topology and node selection strategy;

[0011] Step 3: Use random diffusion and convolutional neural network to extract the local topological structure of the agent node and form node embedding;

[0012] Step 4: Through the graph convolutional neural network, the information increment of the proxy node is used to update the historical node embedding and predict the node influence value. According to the influence value predicted by the model, a group of nodes with the greatest influence are selected as the seed set of the current network snapshot;

[0013] Step 5: Repeat steps 2, 3 and 4 to select a set of optimal seed sets for the entire dynamic network.

[0014] Preferably, in step 1, users are modeled as nodes, the relationships between users are modeled as edges between nodes, and the network is divided into several network snapshots according to time intervals, so as to construct a dynamic social network graph: G = {G 1 ,G 2 ,...,G T}, where T represents the number of network snapshots. Each network snapshot G t <V t ,E t >Contains|V t | nodes and |E t | edge, such as Figure 1 shown.

[0015] Preferably, in step 2, in order to improve the utilization rate of network topology historical information, reduce repeated calculations, and improve algorithm efficiency, a structural evolution agent method is proposed; according to the changes in the network, some nodes are selected as structural evolution agent nodes to recalculate local structural features as recorders of network evolution information.

[0016] Preferably, in step three, the structural evolution proxy nodes obtained in step two are processed. In order to locally characterize the characteristics of the nodes, the propagation characteristics are characterized by a random diffusion method and the local feature matrix of the nodes is constructed in combination with the node degree. Then, a convolutional neural network is used to extract the deep local information of these nodes and form a node embedding for each proxy node.

[0017] Prioritize, in step 4, in order to correct the historical node embedding, a graph neural network is used to fuse the proxy node embedding with the historical node embedding to achieve incremental update and correct the representation error of the historical node embedding. After that, the model is used to predict the influence of each node, and by obtaining the predicted node influence ranking (from large to small), the top k nodes are selected as the most influential seed nodes for the current network snapshot.

[0018] Compared with the prior art, the present invention has the following beneficial effects:

[0019] 1. The present invention defines a method for maximizing the influence of dynamic social networks based on deep learning and structural evolution agents. The method of structural evolution agents is used to avoid computational redundancy and to make full use of deep learning to deeply mine node features and maintain node embedding. This enables the present invention to effectively reduce repeated calculations in dynamic networks, achieve incremental updates, predict the influence values ​​of nodes in network snapshots, and select a group of the most influential seed user sets, thereby maximizing the spread of influence in dynamic social networks.

[0020] 2. The present invention designs a structural evolution proxy method, which reduces the number of repeated calculation nodes and unnecessary calculations, thereby improving the efficiency of the algorithm.

[0021] 3. The present invention introduces a random diffusion method to characterize the propagation characteristics of influence and incorporates propagation dynamics into node characteristics.

[0022] 4. The logical structure of the present invention is easy to understand and implement, and has good generalization ability, and can be well applied to solving the influence maximization problem in most types of dynamic social networks.

[0023] 5. The present invention can improve the utilization rate of historical information of network topology through the incremental update method of dynamic structural evolution, and only calculate some key nodes in network evolution, effectively reducing unnecessary redundant calculations. Deep learning describes node information locally and updates historical information using network topology changes. The dynamic update method based on deep learning is expected to overcome the shortcomings of existing methods, better utilize the local topological information of nodes, incrementally update, and correct the historical embedding of nodes, reduce computational redundancy, and thus obtain more competitive and influential seed users. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 A schematic diagram of a dynamic social network provided by the present invention.

[0025] Figure 2 The present invention provides a deep learning-based incremental update dynamic social network influence maximization framework.

[0026] Figure 3 This is a schematic diagram of the proxy node selection provided by the present invention.

[0027] Figure 4 This is a schematic diagram of generating a node local matrix provided by the present invention. DETAILED DESCRIPTION

[0028] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0029] Example 1

[0030] A method for maximizing the spread of influence in a dynamic social network based on deep learning and structural evolution agent, comprising the following steps:

[0031] Step 1: Collect social network data and construct a dynamic social network graph based on user relationships and time intervals;

[0032] According to different needs, the network is divided at different time intervals to construct a dynamic social network graph: G = {G 1 ,G 2 ,...,G T Among them, the main changes in the dynamic network include four operations: adding nodes, adding edges, deleting nodes, and deleting edges. Figure 1 shown.

[0033] Step 2: Design a structure evolution proxy method. According to the changes in network topology and node selection strategy, only some nodes in the snapshot are selected as structure evolution proxy nodes for recalculation. The remaining nodes do not need to repeatedly calculate the network topology structure. Figure 2 As shown in (b), according to the changes of each network snapshot, the structural evolution agent node required by the current snapshot is queried. First, the network needs to be processed according to four network change situations, and the change frequency of the nodes involved in the change within the current time interval is counted: the frequency of the new node corresponding to the adding node operation increases by 1; the frequency of the first-order neighbors of the node corresponding to the deleting node operation increases by 1; the frequency of the two nodes corresponding to the deleting or adding edge operation increases by 1. For each network snapshot, the obtained frequency dictionary F = {node:frequency|node∈V t} Sort from large to small. Then start looping the frequency dictionary F until the dictionary F is empty and end the selection of the structure evolution proxy node. In the loop, select the node v with the highest update frequency to add to the proxy node set U. Then, delete the node v and its first-order neighbors from the dictionary F, and delete the edges between them from the current network snapshot. Finally, after the loop, the proxy node set U that needs to be updated is obtained. Figure 3 Yes Select Figure 1 China Network Snapshot G 3 Schematic diagram of proxy nodes. According to the network evolution, the dictionary F = {5:3, 1:2, 2:2, 9:2, 3:1, 6:1, 10:1} is obtained, and then the node 5 with the highest update frequency is selected to join the proxy node set U. The corresponding fields of node 5 and the first-order neighbor N(5) = {6, 7, 8, 9, 10} in the dictionary F are deleted from the snapshot G. 3 Delete the edge between node 5 and N(5). Then, select nodes 2 and 1 with the highest frequency in the dictionary F. At this time, dictionary F is empty, and the proxy node selection process is completed. The pseudo code for generating proxy nodes is shown in Algorithm 1.

[0034]

[0035]

[0036] Step 3: Use random diffusion and convolutional neural network to extract the local topological structure of the agent node and form node embedding;

[0037] like Figure 2 As shown in (c), according to the update set generated in step 2, a node local matrix is ​​generated for each proxy node. For each proxy node v, a certain number of random diffusion paths starting from v are first generated. Then, the L-1 nodes with the largest diffusion intensity in the path are selected and form a local subgraph with the proxy node v, and each node in the subgraph is renumbered from 0. According to the local subgraph, its adjacent matrix A is obtained v . Each element Indicates whether there is an edge between nodes i and j in the subgraph in v's subgraph, as described in the formula:

[0038]

[0039] Then according to the obtained subgraph adjacency matrix A u , generate the local feature matrix of the proxy node v Each element Calculated by formula, the description is as shown in the formula:

[0040]

[0041] in, is the degree of node j in the subgraph in the original graph. According to the above formula, each proxy node can generate a feature matrix B v , which is used to describe the local topological characteristics of the node. Figure 4 As shown, this is a snapshot G 3 Node 5 in generates the local feature matrix B 5 .

[0042] like Figure 2 As shown in (d), the obtained proxy node feature matrix is ​​input into the convolutional neural network model, and the feature vector e of each proxy node is obtained after being processed by the convolutional neural network model. 1*d , where d is the node feature dimension.

[0043] Step 4: Use the graph convolutional neural network to update the historical node information using the proxy node embedding obtained in step 3 and predict the node influence value. According to the influence value predicted by the model, select a group of nodes with the greatest influence as the seed set of the current network snapshot.

[0044] like Figure 2 As shown in (e), the node embeddings of historical nodes and proxy nodes are merged to form all the node embeddings of the new snapshot network. The snapshot embedding Input into the graph neural network model. Through the message passing and information aggregation of the graph neural network, the proxy node embedding is used to correct the embedding of the historical nodes, and finally the influence values ​​of all nodes in the new snapshot are predicted.

[0045] like Figure 2 As shown in (f), a group of nodes with the greatest influence is selected as the seed set of the current network snapshot according to the predicted influence values ​​obtained in (e). The seed set of each network snapshot constitutes the influence seed set of the dynamic network G.

[0046] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for maximizing the spread of influence in dynamic social networks based on deep learning and structural evolution agents, characterized in that: The steps include: Step 1: Collect social network data and model it into a dynamic social network graph through user relationships. Divide the network into several time network snapshots according to different time intervals. Step 2: Design a structure evolution agent method, select some nodes as structure evolution agent nodes according to the changes in network topology and node selection strategy; Step 3: Use random diffusion and convolutional neural networks to extract the local topological structure of the agent node and form node embedding; Step 4: Through the graph convolutional neural network, the information of the proxy node is used to update the historical node embedding and predict the expected value of the node influence propagation. According to the influence value predicted by the model, a group of nodes with the greatest influence are selected as the seed set of the current network snapshot; Step 5: Repeat steps 2, 3 and 4 to select a set of seeds for the entire dynamic network.

2. The method for maximizing the spread of influence in a dynamic social network based on deep learning and structural evolution agent according to claim 1, characterized in that: In step 1, users are modeled as nodes, and the relationships between users are modeled as edges between nodes. The network is divided into several network snapshots according to time intervals, thereby constructing a dynamic social network graph: G = {G1, G2, ..., G T }, where T represents the number of network snapshots.

3. The method for maximizing the influence propagation of dynamic social networks based on deep learning and structural evolution agent according to claim 1, characterized in that: In step 2, in order to improve the utilization rate of network topology historical information, reduce repeated calculations, and improve algorithm efficiency, a structural evolution agent method is proposed; according to the changes in the network, some nodes are selected as structural evolution agent nodes to recalculate local structural features as recorders of network evolution information.

4. The method for maximizing the influence propagation of dynamic social networks based on deep learning and structural evolution agent according to claim 1, characterized in that: In step three, the structural evolution proxy nodes obtained in step two are processed. In order to characterize the characteristics of the nodes locally, the propagation characteristics are characterized by the random diffusion method and the local feature matrix of the nodes is constructed in combination with the node degree. Then, the convolutional neural network is used to extract the deep local information of these nodes and form a node embedding for each proxy node.

5. The method for maximizing the spread of influence in a dynamic social network based on deep learning and structural evolution agent according to claim 1, characterized in that: In step 4, in order to correct the historical node embedding, a graph neural network is used to fuse the proxy node embedding with the historical node embedding, perform incremental updates, and correct the representation errors of the historical node embedding. After that, the model is used to predict the influence of each node, and by obtaining the predicted node influence ranking, the top k most influential nodes in the current network snapshot are selected as seed users.

6. A computer-readable storage medium comprising a computer program, which, when processed by a GPU, can implement the method steps provided in claims 1-5.

Images