Social information network key node identification method based on reinforcement learning

Abstract

The invention belongs to the technical field of network key node identification, and particularly relates to a social information network key node identification method based on reinforcement learning. Comprising the following steps: acquiring original graph data of a social information network to be identified, introducing a virtual node to be connected with all nodes and representing the virtual node as an undirected graph; generating an initial node feature matrix based on the adjacent matrix of the undirected graph; carrying out topological relation enhancement processing according to the adjacent matrix and the initial node feature matrix to obtain a weighted topological incidence matrix; performing channel-level dynamic calibration on the initial node feature matrix to obtain a weighted node feature matrix; performing hierarchical topological aggregation according to the weighted topological incidence matrix and the weighted node feature matrix to obtain three aggregation features; fusing the three aggregation features to obtain weighted fusion features; the weighted fusion features serve as agent input, a reinforcement learning method is adopted to train a key node recognition network, and a trained key node recognition network is obtained; according to the invention, the node identification accuracy is improved.

Description

Technical Field

[0001] This invention belongs to the field of network key node identification technology, specifically relating to a method for identifying key nodes in social information networks based on reinforcement learning. Background Technology

[0002] With the rapid development of internet technology and information infrastructure, complex networks, represented by social networks, transportation networks, and information systems, have formed large-scale systems with over 5 billion node interactions daily. These networks generally possess topological characteristics such as the small-world effect and scale-free properties. Their information propagation and energy flow exhibit highly dynamic and nonlinear coupling patterns, forming a complex system structure of "borderless interactive propagation." In these systems, the efficiency of information or matter propagation is closely related to the network topology. A few core nodes with high connectivity and strong structural vulnerabilities act as hubs for network functions, often having a decisive impact on the overall system performance. Currently, the mainstream methods for identifying key nodes include the following:

[0003] 1. Local Attributes: Local attributes of a network refer to the inherent properties of nodes and the interaction or connection methods between nodes and their neighbors. These key node identification methods primarily summarize information about nodes and their neighbors to identify key nodes in the network. The advantage of these methods is their computational simplicity, making them suitable for large-scale, complex networks. Among them, the degree centrality algorithm is the simplest method for analyzing the criticality of nodes in complex networks. It posits that the more neighbors a node has, the greater its influence. The higher the degree of a node in the network, the stronger its information propagation ability; a high-degree node has a greater probability of propagating information in all directions and also a greater chance of receiving information from surrounding nodes. For directed networks, researchers have proposed the Cluster Rank algorithm based on its properties. This algorithm can effectively analyze the position of a node in the network and its relationship with surrounding nodes, and estimate the impact of clustering coefficients on information propagation in the network. The accuracy of the Cluster Rank algorithm is superior to that of local centrality and degree centrality algorithms. However, methods based on local attributes typically rely only on the node's direct neighbor information (such as degree centrality) or its neighborhood structure (such as clustering coefficients). This limitation prevents them from fully reflecting the location and role of nodes within the entire network, especially in applications that require consideration of global impact. Many important nodes may be underestimated because they are located on the periphery of the network or in relatively isolated parts.

[0004] 2. Global Attributes: Global attributes of a network consider the overall behavioral and structural characteristics of the network. These key node identification algorithms primarily assess the characteristics and influence of nodes within the entire network. Betweenness centrality and closeness centrality are two classic algorithms based on global network attributes. Betweenness centrality describes the importance and busyness of nodes during information propagation in a network; it is defined as the number of times the shortest path between any two nodes passes through that node. By calculating the number of shortest paths from a node to other nodes, betweenness centrality reflects the control a node's position in the network has over information transmission. Specifically, the betweenness centrality algorithm calculates the number of times each node appears as an intermediate node in the shortest path and normalizes these counts to obtain a betweenness centrality value, used to measure the node's importance in information propagation. Closeness centrality measures the closeness between nodes in the network, i.e., the speed and ease with which a node reaches other nodes. Global attribute methods typically require traversing the entire network to calculate the shortest paths between node pairs or the connectivity between nodes. This results in very high computational complexity, especially in large-scale networks, where computation time and memory overhead increase dramatically. While this method considers the overall network structure, it pays less attention to the local structural information of nodes. It may overestimate the influence of nodes indirectly connected through multiple nodes, while neglecting nodes that are crucial in local communities or dense subnetworks.

[0005] 3. Location Attributes: Based on the location attributes of a network, it's possible to determine whether a node occupies a critical position in a complex network, thus assessing its criticality. A node in a critical position has a greater influence, while a node at the edge has significantly limited influence. The most representative algorithm for identifying critical nodes based on location attributes is the K-Shell algorithm. The K-Shell algorithm partitions the network hierarchically, considering nodes in the inner layers to be the most critical. The network partitioning is based on the degree of the nodes; according to the rules of the K-Shell algorithm, nodes with a degree less than or equal to K are removed layer by layer until the kernel nodes are identified. Some researchers use the Eulerian distance formula to quantify the degree centrality, K-Shell index, and betweenness centrality of nodes in complex networks to assess the accuracy of the node criticality calculated by these methods. Mining algorithms based on network location attributes have certain requirements for the network structure. For example, the K-Shell algorithm is not applicable to star-shaped networks and BA scale-free networks, and it is difficult to determine the optimal weight factors for each index. However, the location attribute method relies on the assumption of uniformity in network topology, which fails in special structures such as star networks and scale-free networks with BA, and it lacks the ability to quantify the coupling effect of multidimensional indicators.

[0006] 4. Random Walk: Key node identification algorithms based on random walks are commonly used for webpage ranking. This algorithm treats links between different webpages as interconnected and mutually supportive, simulating random clicks by users across webpages to determine the importance and criticality of each page. This algorithm is widely used in search engines and other fields. Typical methods include HITS (Hypertext-Induced Topic Search), Google's PageRank algorithm, and Leader Rank. However, random walk algorithms are susceptible to interference from isolated nodes, and while traditional self-loop processing methods maintain algorithm continuity, they can lead to a higher false positive rate in sparse networks.

[0007] In existing complex networks, there is an urgent need for a critical node identification method that takes into account both network structure and node characteristics. This method needs to have high identification accuracy, wide applicability, and be able to identify critical nodes that have a significant impact on the network better and faster. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention proposes a method for identifying key nodes in social information networks based on reinforcement learning. The method includes: acquiring the social information network to be identified and inputting it into a trained key node identification network for processing to obtain the key node identification result.

[0009] The training process of the key node identification network includes:

[0010] S1: Obtain the original graph data of the social information network to be identified, introduce virtual nodes to connect with all nodes, and represent the social information network as an undirected graph;

[0011] S2: Generate the initial node feature matrix based on the adjacency matrix of the undirected graph;

[0012] S3: Perform topological relationship enhancement processing based on the adjacency matrix and initial node feature matrix of the undirected graph to obtain the weighted topological association matrix;

[0013] S4: Perform channel-level dynamic calibration on the initial node feature matrix to obtain the weighted node feature matrix;

[0014] S5: Perform hierarchical topological aggregation based on the weighted topological association matrix and the weighted node feature matrix to obtain three types of aggregation features: first-order aggregation features, second-order aggregation features, and global context aggregation features.

[0015] S6: Combine the three aggregation features to obtain a weighted fusion feature;

[0016] S7: Using weighted fusion features as input to the agent, a reinforcement learning method is used to train the key node recognition network, resulting in a well-trained key node recognition network.

[0017] Preferably, the initial node feature matrix is ​​represented as follows:

[0018]

[0019] in, This represents the element in the i-th row and k-th column of the initial node feature matrix. Let i represent the set of k-order neighbors of node i. Let V represent the set of k-order neighbors of node j, and let V represent the set of nodes.

[0020] Preferably, the process of obtaining the weighted topological correlation matrix includes:

[0021] Calculate the node similarity matrix based on the initial node feature matrix;

[0022] The node similarity matrix is ​​subjected to exponential normalization. Based on the node similarity matrix after exponential normalization, the adjacency matrix is ​​weighted to obtain the weighted topological association matrix.

[0023] Furthermore, the formula for calculating the node similarity matrix is:

[0024]

[0025] in, Represents the node similarity matrix. express Activation function Represents the initial node feature matrix. Let represent the first learnable parameter matrix.

[0026] Preferably, the process of obtaining the weighted node feature matrix includes:

[0027] The initial node feature matrix is ​​subjected to global average pooling to obtain channel statistics;

[0028] The channel statistics are input into two fully connected layers to learn the channel importance coefficients, thus obtaining the compression ratio;

[0029] The initial node feature matrix is ​​weighted channel by channel according to the compression ratio to obtain the weighted node feature matrix.

[0030] Preferably, the process of hierarchical topological aggregation based on the weighted topological association matrix and the weighted node feature matrix includes:

[0031] The weighted topological incidence matrix is ​​used as a first-order adjacency matrix. The first-order adjacency matrix is ​​then squared to obtain a second-order adjacency matrix.

[0032] The first-order adjacency matrix and the weighted node feature matrix are aggregated to obtain the first-order aggregated features;

[0033] The second-order adjacency matrix and the weighted node feature matrix are aggregated to obtain the second-order aggregated features;

[0034] The weighted node feature matrix is ​​aggregated to obtain the global context aggregated features.

[0035] Furthermore, the first-order adjacency matrix and the initial node feature matrix are aggregated and represented as follows:

[0036]

[0037] The aggregation of the second-order adjacency matrix and the initial node feature matrix is ​​expressed as follows:

[0038]

[0039] The initial node feature matrix is ​​aggregated and represented as follows:

[0040]

[0041] in, Represents first-order aggregate features. Indicates second-order aggregation features, Represents global context aggregated features. express Activation function This represents the eigenvector of node j in the weighted node feature matrix. This represents the element in the i-th row and j-th column of a first-order adjacency matrix. This represents the element in the i-th row and j-th column of a second-order adjacency matrix. This represents the second learnable parameter matrix. This represents the third learnable parameter matrix. This represents the fourth learnable parameter matrix. Indicates the number of nodes. Let i represent the set of second-order neighbor nodes of node i.

[0042] Preferably, the process of fusing the three aggregation features includes:

[0043] By concatenating the first-order aggregated features, the second-order aggregated features, and the global context aggregated features, a combined vector is obtained.

[0044] The combined vectors are subjected to nonlinear transformation to obtain the intermediate feature vector u;

[0045] The intermediate feature vector u is sliced ​​dimensionally based on the dimensions of the first-order, second-order, and global context aggregated features to obtain sub-vectors. Attention scores for the three aggregated features are calculated based on the three sub-vectors.

[0046] The weighted fusion feature is obtained by weighting and summing the first-order aggregated feature, the second-order aggregated feature, and the global context aggregated feature based on the attention scores of the three aggregated features.

[0047] Furthermore, the formula for calculating the attention score of the aggregated feature is:

[0048]

[0049] in, Subvector The attention score of the corresponding aggregated feature, where S represents the multi-scale set. ; Subvector , Subvector , This represents the fifth learnable parameter matrix.

[0050] The beneficial effects of this invention are as follows:

[0051] The key node identification technology proposed in this invention explicitly models the dependencies between convolutional feature channels, performs hierarchical topological aggregation based on a weighted topological correlation matrix and node features, generates first-order, second-order, and global context aggregation features, and improves the reward design by introducing the number of K-kernel nodes in the decoding module. This significantly improves the accuracy, efficiency, and robustness of key node identification, effectively overcoming the shortcomings of existing technologies such as ignoring global position with local attributes, high computational complexity of global attributes, reliance on topological uniformity assumptions for positional attributes, and susceptibility to interference from isolated nodes in random walk algorithms. This invention considers both network structure and node characteristics, achieving high identification accuracy and wide applicability, and can better and faster identify key nodes that have a significant impact on the network. Attached Figure Description

[0052] Figure 1 This is a flowchart illustrating the training process of the key node identification network in this invention.

[0053] Figure 2 This is a diagram of the DQN deep reinforcement learning architecture in this invention. Detailed Implementation

[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] This invention proposes a method for identifying key nodes in social information networks based on reinforcement learning, such as... Figure 1 As shown, the method includes the following:

[0056] The social information network to be identified is obtained and input into a trained key node identification network for processing to obtain the key node identification results.

[0057] The training process of the key node identification network includes:

[0058] S1: Obtain the original graph data of the social information network to be identified, introduce virtual nodes to connect with all nodes, and represent the social information network as an undirected graph.

[0059] Obtain the original graph data of the social information network to be identified, introduce virtual nodes and connect them to all nodes to ensure connectivity between nodes; represent the social information network as an undirected graph G=(V,E).

[0060] S2: Generate the initial node feature matrix based on the adjacency matrix of the undirected graph.

[0061] This invention designs a novel encoding module. This module first generates an initial node feature matrix X based on the adjacency matrix of an undirected graph. It then constructs an initial feature vector by statistically analyzing the degree centrality of each node. The generated initial node feature matrix is ​​represented as follows:

[0062]

[0063] in, This represents the element in the i-th row and k-th column of the initial node feature matrix, which is the feature vector for node i. The k-th component, Let i represent the set of k-order neighbors of node i. Let V represent the set of k-order neighbors of node j (k=1,2), and let V represent the set of nodes.

[0064] This calculation is achieved through exponentiation of the adjacency matrix: The non-zero element positions correspond to k-order neighbor relationships, and the number of k-order neighbors for each node is obtained by summing the rows. The final generated feature matrix The normalized number features, including first-order and second-order neighbors, form the initial node feature representation driven by the topology structure.

[0065] S3: Perform topological relationship enhancement processing based on the adjacency matrix and initial node feature matrix of the undirected graph to obtain the weighted topological association matrix.

[0066] Calculate the node similarity matrix based on the initial node feature matrix:

[0067]

[0068] in, Represents the node similarity matrix. express Activation function Represents the initial node feature matrix. Let represent the first learnable parameter matrix.

[0069] The binary connectivity relationships in the original adjacency matrix are transformed into a weighted topological incidence matrix. This transformation is achieved through regularization: the node similarity matrix is ​​exponentially normalized, and the adjacency matrix is ​​weighted based on the exponentially normalized node similarity matrix. The weighted topological incidence matrix is ​​obtained; where, Let represent the set of neighboring nodes of node i, and m represent the set of neighboring nodes of node i. This represents the element in the i-th row and j-th column of the node similarity matrix. This represents the element in the i-th row and m-th column of the node similarity matrix; this processing enables the connection between nodes to have a differentiable quantified association strength, providing a topological weight basis for subsequent hierarchical topological aggregation.

[0070] S4: Perform channel-level dynamic calibration on the initial node feature matrix to obtain the weighted node feature matrix.

[0071] Global average pooling is performed on the initial node feature matrix to capture the feature distribution information of the entire image and obtain channel statistics. :

[0072]

[0073] in, This indicates the number of nodes in a social information network.

[0074] The channel statistics are input into two fully connected layers to learn the channel importance coefficients, thus obtaining the compression ratio. :

[0075]

[0076] in, This represents the learnable weight matrix of the first fully connected layer. This represents the learnable weight matrix of the second fully connected layer. This represents the Sigmoid function. This represents a non-linear activation function, namely the ReLU function.

[0077] The calibration process performs a channel-by-channel weighting operation on the initial node feature matrix according to the compression ratio, thereby improving the response intensity of key feature channels and suppressing interference from noise channels.

[0078]

[0079] in, This represents the eigenvector of node i in the weighted node eigenma matrix.

[0080] This process provides optimized feature inputs for subsequent hierarchical topology aggregation.

[0081] S5: Perform hierarchical topological aggregation based on the weighted topological association matrix and the weighted node feature matrix to obtain three types of aggregation features: first-order aggregation features, second-order aggregation features, and global context aggregation features.

[0082] Constructing a third-order topology:

[0083] The first-order adjacency matrix is ​​directly adopted using the weighted topological incidence matrix. The second-order adjacency matrix is ​​obtained by squaring the first-order adjacency matrix. Capture extended neighborhood.

[0084] Perform feature transformation and weighted aggregation operations at each scale, specifically:

[0085] This is achieved by aggregating direct neighbors, i.e., the first-order adjacency matrix, with the weighted node feature matrix, to obtain the first-order aggregated features:

[0086]

[0087] By fusing indirect relationships, i.e., aggregating the second-order adjacency matrix and the weighted node feature matrix, we obtain the second-order aggregated features:

[0088]

[0089] By injecting graph-level information, i.e., aggregating the weighted node feature matrix, we obtain the global context aggregated features:

[0090]

[0091] in, Represents first-order aggregate features. Indicates second-order aggregation features, Represents global context aggregated features. express Activation function This represents the eigenvector of node j in the weighted node feature matrix. This represents the element in the i-th row and j-th column of a first-order adjacency matrix. This represents the element in the i-th row and j-th column of a second-order adjacency matrix. This represents the second learnable parameter matrix. This represents the third learnable parameter matrix. This represents the fourth learnable parameter matrix. Let i represent the set of second-order neighbor nodes of node i.

[0092] The encoding module designed in this invention achieves excellent scalability and adaptability through a hierarchical topology aggregation strategy. Its core innovation lies in constructing a three-level processing architecture of "feature calibration - topology enhancement - adaptive fusion". First, channel-level dynamic calibration is implemented in the feature preprocessing stage. Feature statistics are obtained through global average pooling, and channel importance coefficients are learned through two fully connected layers to achieve channel-by-channel weighted optimization, effectively improving the response of key feature channels and suppressing noise interference. Subsequently, hierarchical topology aggregation is implemented based on the weighted adjacency matrix and calibration features, innovatively constructing a three-order topology structure. Each topology level performs feature transformation and weighted aggregation through an independent parameter matrix to generate multi-scale representations.

[0093] S6: Combine the three aggregation features to obtain the weighted fusion feature.

[0094] To adaptively fuse multi-scale features, a dynamic receptive field selection mechanism is designed, specifically:

[0095] The first-order aggregated features, second-order aggregated features, and global context aggregated features are concatenated into a combined vector. [ h i 1 || h i 2 || h i g ] .

[0096] A nonlinear transformation is performed on the combined vector to generate weighting coefficients:

[0097] u=δ( W a [ h i 1 || h i 2 || h i g ])

[0098] in, It is a nonlinear transformation learnable weight matrix. It is a non-linear activation function ReLU, and the final result u is still a vector.

[0099] The intermediate feature vector u is sliced ​​dimensionally based on the dimensions of the first-order, second-order, and global context aggregated features to obtain sub-vectors. Specifically: , , It is a local-scale enhancement feature. It is their overall container, here The dimension is , , The sum of dimensions, where the first-order aggregate features The dimension is Second-order aggregation features The dimension is Global context aggregation features The dimension is Split u according to its dimensions, and extract sub-vectors using array slicing operations. Before taking u One element, , After taking u Each element.

[0100] Attention scores for the three aggregated features are calculated based on the three sub-vectors:

[0101]

[0102] in, Subvector The attention score of the corresponding aggregated feature, where S represents the multi-scale set. , Subvector , Subvector , This represents the fifth learnable parameter matrix.

[0103] The weighted fusion feature is obtained by weighting and summing the first-order aggregated feature, the second-order aggregated feature, and the global context aggregated feature based on the attention scores of the three aggregated features. :

[0104]

[0105] By integrating three aggregation features, the model can automatically adjust the contribution ratio of neighbors of different orders based on the local topological complexity of a node.

[0106] The dynamic receptive field selection mechanism designed in this invention involves splicing multi-scale representations, performing nonlinear transformations to generate hidden vectors, calculating attention scores through learnable parameters, and ultimately achieving adaptive weighted fusion. This design enables the model to automatically adjust the contribution ratio of neighbors at each order based on the local topological complexity of nodes. While inheriting the architectural advantages of parameter-independent graph size, it significantly improves the adaptability to dynamic graphs and the mathematical representation ability of graph structural features.

[0107] S7: Use the weighted fusion features calculated in step S6 as input to the agent, and train the key node recognition network using reinforcement learning to obtain the trained key node recognition network.

[0108] This invention designs a decoding module to solve for the key node recognition result. This module consists of input, an agent, and an environment. The feature vector calculated by the encoding module, i.e., the weighted fused feature, is used as the input to the decoding module. The decoding model uses a reinforcement learning method to train the key node recognition network, including:

[0109] The system defines a state, action, and reward function. The state represents the remaining subnet structure after removing the selected key node. The action is to remove the selected key node. This invention introduces a penalty term in the reward function for the number of 2-core nodes within the largest connected component of the remaining network. The total reward formula is as follows:

[0110]

[0111]

[0112]

[0113] in, It is the total reward. It is the original reward. It is a penalty item. This represents the number of nodes with a K-core of 2 in the remaining maximum connected component (LCC) of the graph. This indicates the number (size) of nodes in the network's largest connected component (GCC). This represents the number of nodes in the largest connected component (GCC) of the original network G. This indicates the removal of a node from network G. The remaining subnets.

[0114] The core architecture of the intelligent agent employs a Graph Convolutional Network (GCN). The state is used as input to the GCN, undergoing multiple convolutional operations combined with the ReLU activation function to generate feature representations of the state and different actions. These representations are then passed through fully connected layers and a softmax output layer to output the Q(S,A) value. Based on the Q value, the corresponding action is selected—that is, the removal of the selected key node. The environment (the network being analyzed) provides a reward and the next state based on this action. The model iteratively optimizes the loss function with the goal of minimizing cumulative connectivity, thereby achieving key node detection.

[0115] The cumulative connectivity of the network after removing a node is:

[0116] like Figure 2 As shown, a deep reinforcement learning architecture based on DQN is used to train the model, and the training data is sampled through an experience replay mechanism: from the experience pool... Uniformly draw state-action-reward-transition state tuples Each trajectory The complete node removal sequence is represented. A dual-network architecture—the main network Q and the target network Q—is used for stable Q-value estimation. The latter parameter is updated synchronously with the main network every C training steps, remaining fixed during this period to avoid oscillations in the value function. The training process is based on a dynamically generated synthetic network, and its termination condition is set when the network's largest connected component (GCC) is completely decomposed.

[0117] Each training episode consists of a state-action chain ( It consists of ) components, in which the exploration strategy adopts Greedy algorithm, initial exploration rate The threshold value decreases linearly to 0.01 over 10,000 rounds to balance the explore-exploitation tradeoff. During inference, a deterministic strategy is implemented, selecting nodes to be removed based on the maximum value of the Q-function until the termination state. Parameter optimization employs stochastic gradient descent, sampling mini-batch data from the experience pool after each round and updating network weights by minimizing the temporal difference loss function.

[0118] The model was trained for a total of 50,000 episodes, with a replay buffer of 20,000 units storing the most recent state transitions. Every 300 episodes, the agent's performance was evaluated using 100 synthetic graphs of the same size as the training graph, and the average was taken. After each episode, a mini-batch of state transitions was randomly sampled from the replay buffer, and stochastic gradient descent was performed to update the loss function. The key node identification algorithm based on deep reinforcement learning is as follows:

[0119] 1) Initialize model settings, including experience replay pool size, number of iterations, and various weights;

[0120] 2) Establish a Q-target network and a Q-estimation network. The two networks have the same structure but different parameter updates.

[0121] 3) For each n from 1 to N, perform the following operation:

[0122] 4) Generate a scale-free BA network G;

[0123] 5) Obtain the 2-kernel graph of G ;

[0124] 6) Initialize the state sequence (empty sequence);

[0125] 7) For each t from 1 to T, perform the following operation:

[0126] 8) If the random number is less than Choose a random action Otherwise, select an action. ;

[0127] 9) Perform the action (from Add or remove nodes and observe the rewards. ;

[0128] 10) Update the partial solution to ;

[0129] 11) If t > n, then convert... Stored in the experience replay pool;

[0130] 12) Randomly sample a small batch of transformations from D. ;

[0131] 13) Randomly select a set of Mini Batch samples from the experience pool M, and optimize the parameters using the gradient descent method according to the reward and loss formulas;

[0132] 14) Update the target network parameters after each C-iteration iteration;

[0133] 15) Once the maximum number of iterations is reached, save the optimal model;

[0134] 16) Use the optimal model for testing.

[0135] The social information network to be identified is obtained and input into a pre-trained key node identification network for processing. This yields the importance values ​​of each node in the social information network, i.e., the key node identification results, which are then sorted from highest to lowest value. The higher the node's ranking, the more important it is in the network.

[0136] This invention demonstrates significant value across multiple fields in practical applications: In social network analysis, it can accurately identify influential core users (opinion leaders), significantly optimize information dissemination paths, and provide technical support for curbing misinformation and accelerating public welfare campaigns; in the field of critical infrastructure operation and maintenance (such as transportation networks and power grids), it can accurately locate the nodes in the system most prone to cascading failures, thereby formulating targeted protection strategies and systematically improving the resilience and stability of infrastructure; in the field of biomedical research, this technology can efficiently identify key pivot proteins in protein-protein interaction networks, providing a powerful tool for accelerating the discovery of potential drug targets and understanding complex life processes.

[0137] The above-described embodiments further illustrate the purpose, technical solution, and advantages of the present invention. It should be understood that the above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made to the present invention within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for identifying key nodes in a social information network based on reinforcement learning, characterized in that, include: The social information network to be identified is obtained and input into a trained key node identification network for processing to obtain the key node identification result. The training process of the key node identification network includes: S1: Obtain the original graph data of the social information network to be identified, introduce virtual nodes to connect with all nodes, and represent the social information network as an undirected graph; S2: Generate the initial node feature matrix based on the adjacency matrix of the undirected graph; S3: Perform topological relationship enhancement processing based on the adjacency matrix and initial node feature matrix of the undirected graph to obtain the weighted topological association matrix; S4: Perform channel-level dynamic calibration on the initial node feature matrix to obtain the weighted node feature matrix; S5: Perform hierarchical topological aggregation based on the weighted topological association matrix and the weighted node feature matrix to obtain three types of aggregation features: first-order aggregation features, second-order aggregation features, and global context aggregation features. S6: Combine the three aggregation features to obtain a weighted fusion feature; S7: Using weighted fusion features as input to the agent, a reinforcement learning method is used to train the key node recognition network, resulting in a well-trained key node recognition network.

2. The method for identifying key nodes in a social information network based on reinforcement learning according to claim 1, characterized in that, The initial node feature matrix is ​​represented as follows: ， in, This represents the element in the i-th row and k-th column of the initial node feature matrix. Let i represent the set of k-order neighbors of node i. Let V represent the set of k-order neighbors of node j, and let V represent the set of nodes.

3. The method for identifying key nodes in a social information network based on reinforcement learning according to claim 1, characterized in that, The process of obtaining the weighted topological incidence matrix includes: Calculate the node similarity matrix based on the initial node feature matrix; The node similarity matrix is ​​subjected to exponential normalization. Based on the node similarity matrix after exponential normalization, the adjacency matrix is ​​weighted to obtain the weighted topological association matrix.

4. The method for identifying key nodes in a social information network based on reinforcement learning according to claim 3, characterized in that, The formula for calculating the node similarity matrix is: ， in, Represents the node similarity matrix. express Activation function Represents the initial node feature matrix. Let represent the first learnable parameter matrix.

5. The method for identifying key nodes in a social information network based on reinforcement learning according to claim 1, characterized in that, The process of obtaining the weighted node feature matrix includes: The initial node feature matrix is ​​subjected to global average pooling to obtain channel statistics; The channel statistics are input into two fully connected layers to learn the channel importance coefficients, thus obtaining the compression ratio; The initial node feature matrix is ​​weighted channel by channel according to the compression ratio to obtain the weighted node feature matrix.

6. The method for identifying key nodes in a social information network based on reinforcement learning according to claim 1, characterized in that, The process of hierarchical topological aggregation based on the weighted topological association matrix and the weighted node feature matrix includes: The weighted topological incidence matrix is ​​used as a first-order adjacency matrix. The first-order adjacency matrix is ​​then squared to obtain a second-order adjacency matrix. The first-order adjacency matrix and the weighted node feature matrix are aggregated to obtain the first-order aggregated features; The second-order adjacency matrix and the weighted node feature matrix are aggregated to obtain the second-order aggregated features; The weighted node feature matrix is ​​aggregated to obtain the global context aggregated features.

7. The method for identifying key nodes in a social information network based on reinforcement learning according to claim 6, characterized in that, The aggregation of the first-order adjacency matrix and the initial node feature matrix is ​​represented as follows: ， The aggregation of the second-order adjacency matrix and the initial node feature matrix is ​​expressed as follows: ， The initial node feature matrix is ​​aggregated and represented as follows: ， in, Represents first-order aggregate features. Indicates second-order aggregation features, Represents global context aggregated features. express Activation function This represents the eigenvector of node j in the weighted node feature matrix. This represents the element in the i-th row and j-th column of a first-order adjacency matrix. This represents the element in the i-th row and j-th column of a second-order adjacency matrix. This represents the second learnable parameter matrix. This represents the third learnable parameter matrix. This represents the fourth learnable parameter matrix. Indicates the number of nodes. Let i represent the set of second-order neighbor nodes of node i.

8. The method for identifying key nodes in a social information network based on reinforcement learning according to claim 1, characterized in that, The process of integrating the three aggregation features includes: By concatenating the first-order aggregated features, the second-order aggregated features, and the global context aggregated features, a combined vector is obtained. The combined vectors are subjected to nonlinear transformation to obtain the intermediate feature vector u; The intermediate feature vector u is sliced ​​dimensionally based on the dimensions of the first-order, second-order, and global context aggregated features to obtain sub-vectors. Attention scores for the three aggregated features are calculated based on the three sub-vectors. The weighted fusion feature is obtained by weighting and summing the first-order aggregated feature, the second-order aggregated feature, and the global context aggregated feature based on the attention scores of the three aggregated features.

9. The method for identifying key nodes in a social information network based on reinforcement learning according to claim 8, characterized in that, The formula for calculating the attention score of aggregated features is: ， in, Subvector The attention score of the corresponding aggregated feature, where S represents the multi-scale set. ; Subvector , Subvector , This represents the fifth learnable parameter matrix.

Images