A method for predicting user influence for rapid evaluation of social network information propagation effect

Abstract

This invention discloses a method for predicting user influence for rapid evaluation of information dissemination effects on social networks. The method includes: acquiring and preprocessing graph data of the target social network; inputting the preprocessed graph data and node features into a cascaded perceptron graph neural network, performing multiple rounds of cascaded perceptron message passing, and generating embedded representations of all candidate nodes in the target social network; inputting the embedded representations into a deep reinforcement learning policy network and evaluation network, using a Mannhausen reinforcement learning mechanism to calculate the value function of the current candidate node's action, and evaluating the influence potential of each candidate node; selecting the candidate node with the highest score based on the evaluation results and adding it to the seed set, and updating the node state mask; repeating the above steps until the seed set reaches a preset budget, and outputting the final influence-maximizing seed set. This invention significantly improves the accuracy of seed selection, decision stability, and computational efficiency.

Description

Technical Field

[0001] This invention belongs to the field of social network analysis and data mining technology, specifically relating to a method for predicting user influence for rapid evaluation of the effect of information dissemination on social networks. Background Technology

[0002] Influence Maximization (IM) is a key problem in social network analytics and data mining. Its goal is to find an initial set of nodes (i.e., a seed set) within a given social network to maximize the number of nodes that the information ultimately reaches under a specific propagation model. This problem has significant applications in scenarios such as viral marketing and public opinion monitoring.

[0003] Early classic solutions relied on the Greedy Algorithm. While this method offered some guarantee of approximation ratio, it depended on large-scale Monte Carlo simulations to estimate the sphere of influence, resulting in extremely high computational costs and making it difficult to apply to large-scale social networks. Subsequently, to improve efficiency, algorithms based on Reverse Influence Sampling (RIS) were proposed, reducing computational complexity by generating random propagation paths. However, when dealing with ultra-large-scale, dynamically changing networks, such methods still suffer from high memory consumption and long processing times, limiting their application in scenarios with high timeliness requirements.

[0004] In recent years, with the rapid development of deep learning and reinforcement learning (RL), using graph neural networks (GNNs) and deep reinforcement learning to solve the inference problem (IM) has become a new research direction. The core idea of ​​these methods is to learn the mapping relationship between network topology and diffusion dynamics through offline training, thereby achieving real-time inference on seed nodes. Although some deep reinforcement learning-based schemes (such as S2V-DQN, PIANO, etc.) have emerged, they still have the following significant drawbacks in practical applications: 1) Limitations in Feature Extraction Dimensions. Most existing learning-based methods focus only on the static topology of social networks (such as low-order features like node degree and centrality) or use general graph neural networks (such as GCN and GraphSAGE) to extract node representations. However, influence propagation is essentially a dynamic cascade effect with recursive characteristics; higher-order cascade features in the network are closely related to the potential probability of information diffusion. Existing graph neural network architectures often struggle to explicitly model this complex cascade propagation mechanism, leading to inaccurate estimations of influence potential.

[0005] 2) Stability Challenges in Reinforcement Learning Training. When dealing with large-scale social networks, RL agents face an extremely large action space (number of nodes). Traditional Q-learning algorithms are prone to policy instability and overestimation when handling such sparse rewards and complex state spaces. This not only leads to slow model convergence but also severely affects the quality of the final seed set, making learning methods often less robust when facing complex and varied real-world network topologies.

[0006] Therefore, how to capture the high-level hierarchical characteristics of influence diffusion at the feature perception level and improve the training efficiency and stability of reinforcement learning at the decision-making level are key technical problems that urgently need to be solved in the field of maximizing influence. Summary of the Invention

[0007] To address the problems in related technologies, this invention proposes a predictive user influence method for rapid evaluation of information dissemination effects on social networks. By constructing a computational framework that combines deep perception and stable decision-making, it solves the problems of incomplete feature capture and unstable training process in existing technologies when dealing with large-scale dynamic propagation networks.

[0008] The technical solution of this invention is implemented as follows: A method for predicting user influence for rapid evaluation of the dissemination effect of information on social networks, comprising the following steps: S1. Obtain the graph data of the target social network and perform preprocessing; S2. Input the preprocessed graph data and node features into the cascaded perceptual graph neural network, execute multiple rounds of cascaded perceptual message passing, and generate the embedded representation of all candidate nodes in the target social network; S3. Input the embedded representation into the policy network and evaluation network of the deep reinforcement learning, and use the Mannhausen reinforcement learning mechanism to calculate the value function of the action of the current candidate node, and evaluate the influence potential of each candidate node. S4. Based on the evaluation results, select the candidate node with the highest score to add to the seed set, and update the node state mask; S5. Determine whether the size of the current seed set has reached the preset budget: If not, repeat steps S2 to S4; if it has, output the current seed set as the final seed set for maximizing influence.

[0009] This invention provides a predictive user influence method for rapid evaluation of information dissemination effects on social networks. It utilizes a cascaded perceptual graph neural network to perform multi-round message passing on social network graph data, generating an embedded representation that reflects the node's propagation potential, thus improving the accuracy of node influence potential evaluation. Furthermore, it employs a Mannhausen reinforcement learning mechanism to calculate and evaluate the action-value function of candidate nodes, significantly enhancing the training stability and policy convergence of the reinforcement learning agent in the sequential decision-making process. Finally, through an iterative seed node selection and state mask update mechanism, it achieves adaptive construction of the seed set under a given budget, ensuring that the marginal benefit of the selected node in terms of propagation range is maximized.

[0010] As a further improvement to the above solution, the step of acquiring and preprocessing the graph data of the target social network specifically involves: S1-1. Obtain the target social network data and abstract it into a graph structure. And calculate each edge based on historical interaction frequency or a preset propagation model. activation probability ; in, It is the set of nodes in the network, where n represents the total number of nodes. It is the set of edges in the network; S1-2. Initialize the network features for each node. Constructing the initial feature vector Features include the in-degree, centrality, and state mask indicating whether the current node has been selected as a seed. S1-3, Preset seed node quantity budget And initialize an empty seed set. .

[0011] By abstracting the target social network into a graph structure with activation probabilities and initializing its features, while pre-setting a seed budget and initializing an empty set, the normalization of input data and the quantitative expression of propagation probability are achieved. This provides nodes with a feature foundation that combines static attributes and dynamic identification, laying a complete data and task foundation for subsequent cascaded sensing feature extraction and sequence decision-making.

[0012] As a further improvement to the above scheme, the step of performing multi-round cascaded perceptual message passing to generate embedded representations of all candidate nodes in the target social network specifically involves: In each iteration, the current node Aggregate its neighbor nodes The feature information, combined with the activation probability The weighting adjustment is performed using the following formula: ; in, For the first Layer nodes The embedding vector; For the first Layer nodes The embedding vector; For the first Layer nodes The embedding vector; Gather for neighbors; For feature splicing operations; It is a cascaded factor function; go through After layer aggregation, the final embedding representation matrix of all candidate nodes in the target social network is extracted. .

[0013] By using cascaded sensing message passing and activation probability weighted aggregation, a node embedding representation containing multi-level cascade effects is generated, providing a precise feature representation rich in propagation potential information for subsequent influence assessment.

[0014] As a further improvement to the above scheme, the final embedding representation is extracted after L-layer aggregation, specifically as follows: By stacking multiple cascaded sensing layers, the cascade effect within the aggregation range from layer 1 to layer L is captured. The output features of multiple cascaded sensing layers are then fused through a feature fusion layer to obtain the final embedding representation matrix of all candidate nodes in the target social network.

[0015] By capturing the cascading effects within the aggregation range of layers 1 to L and performing a feature fusion layer, the final embedding representation can integrate multi-scale propagation information, effectively improving the node embedding's ability to represent high-level cascading features.

[0016] As a further improvement to the above scheme, the step of using the Mannhausen reinforcement learning mechanism to calculate the value function of the current candidate node's action and evaluate the influence potential of each candidate node is as follows: The action of each candidate node is calculated using the Mannhausen reinforcement learning mechanism. Value function Furthermore, during the calculation process, the action value is corrected by introducing a Mannhausen regularization term: ; in, The modified action value function after introducing the Manhausen regularization term; State; This is the scaling factor; For the Manhausen item, The temperature is the entropy temperature.

[0017] By introducing Manhausen regularization to correct the original action values, the overestimation bias of the value function under ultra-large-scale node space is effectively suppressed, and the decision stability of the reinforcement learning agent in the seed node sequence selection process is improved.

[0018] As a further improvement to the above scheme, the step of selecting the candidate node with the highest score to add to the seed set based on the evaluation results and updating the node state mask is as follows: S4-1. Based on the evaluation results, select the node with the highest score from the set of candidate nodes. ; S4-2, Select the node Add to seed set ; S4-3. Update the node status mask, mark selected nodes as unselectable, and provide immediate rewards. Instant rewards According to the node The expected increase in influence under the current graph topology Sure.

[0019] By iteratively selecting the node with the highest score to add to the seed set and updating the state mask, the selected node is ensured not to be selected repeatedly. At the same time, an instant reward is designed based on the marginal influence increment, which enables the model to effectively identify the complementarity between nodes and avoid excessive overlap of seed nodes in the propagation range.

[0020] As a further improvement to the above scheme, the instant reward Defined as: ; in, For the seed set; A function for the scale of influence propagation; For node actions.

[0021] By defining an immediate reward, the model can quantify the marginal contribution of each candidate node, effectively identify the complementarity between nodes, avoid excessive overlap of selected seed nodes in the propagation range, and improve the overall coverage efficiency of the seed set.

[0022] As a further improvement to the above scheme, the final seed set is constructed using an iterative incremental approach during the method execution phase, specifically as follows: Through the seed masking mechanism, a negative bias mask is applied to the nodes that have been selected into the seed set in each round of selection, ensuring that the same node is not selected repeatedly in subsequent steps, and enabling the cascaded perceptual graph neural network to readjust the perceptual features according to the distribution of the remaining unselected nodes. Through an end-to-end inference process, a real-time snapshot of the target social network is input into the trained cascaded perceptual graph neural network during model deployment, and the node with the largest current influence increment is directly selected based on the probability output of the policy network. By generalizing across scales, after training on small-scale graph data, it can be directly applied to larger-scale target social networks to perform inference.

[0023] By using a seed masking mechanism to avoid repeated node selection and dynamically adjust perceptual features, and by enabling rapid decision-making without Monte Carlo simulation through end-to-end inference, and by supporting training on small graphs and applications on large graphs through cross-scale generalization, inference efficiency, scalability and practicality are significantly improved.

[0024] As a further improvement to the above scheme, the reinforcement learning algorithm adds the KL divergence constraint between the current policy and the target policy to the loss function.

[0025] By using KL divergence constraints to force the policy network to remain smooth during the update process, the reward fluctuations caused by superconnected nodes are effectively suppressed, thereby improving the training stability and generalization ability of reinforcement learning in complex social networks.

[0026] Compared with traditional processing methods, the present invention has the following advantages: (1) Breakthrough in Feature Perception Capability: Traditional learning-based IM methods rely solely on the static topological features of nodes or general graph neural networks, making it difficult to model the recursive cascading characteristics of information diffusion. This invention, through a cascading perceptual graph neural network, explicitly incorporates the activation probability on edges as a cascading factor function into the message passing process, enabling node embeddings to deeply encode the expected diffusion increment in higher-order propagation paths. Compared to existing deep learning methods such as S2V-DQN, this invention improves the influence propagation range by more than 14% on real social network datasets such as Douban and Soc-Epinions, significantly solving the technical bottleneck of insufficient perception of dynamic cascading features.

[0027] (2) Improved decision stability: Traditional deep reinforcement learning methods are prone to value function overestimation bias and policy oscillation when dealing with ultra-large-scale node spaces, leading to difficulty in model convergence and large fluctuations in seed selection. This invention introduces the Manhausen mechanism into the influence maximization task for the first time. By introducing a Manhausen regularization term based on the log probability of the current policy into the objective value function, the action value is smoothly corrected, effectively suppressing the value function overestimation problem; at the same time, combined with the KL divergence policy regularization constraint, the policy update is forced to remain smooth. This improvement significantly enhances the training stability and policy robustness of reinforcement learning agents in complex social network topologies.

[0028] (3) Improved computational efficiency: Traditional greedy algorithms rely on massive Monte Carlo simulations, and RIS-based methods face severe memory and sampling overhead in ultra-large-scale networks, making it difficult to meet real-time decision-making requirements. This invention adopts an end-to-end architecture of offline training and online inference. During the deployment phase, only one forward propagation of the neural network is needed to iteratively output seed nodes, without any Monte Carlo simulation or backsampling. When processing networks of tens of millions, this invention can generate high-quality seed sets in milliseconds, which is 2-3 orders of magnitude faster than traditional algorithms, providing a feasible technical solution for time-sensitive scenarios such as sudden public opinion monitoring and real-time precision marketing. Attached Figure Description

[0029] Figure 1 This is a flowchart of the algorithm of the present invention; Figure 2 This is a performance comparison chart of the present invention and existing technologies on multiple real social network datasets. Detailed Implementation

[0030] 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 a part of the embodiments of the present invention, and not all of them. 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.

[0031] Example 1 like Figure 1 As shown, this embodiment provides a method for maximizing influence based on cascaded perceptual graph neural networks and Mannhausen reinforcement learning, including the following steps: Step S1: Obtain the graph data of the target social network and perform preprocessing.

[0032] This embodiment primarily focuses on directed weighted graphs, but it is equally applicable to undirected or unweighted graphs. The target social network data is acquired and abstracted into a graph structure. ,in It is the set of nodes in the network, where n represents the total number of nodes; It is the set of edges in the network. Represents the number of edges. Each edge Associated with a diffusion probability , represents a node Successfully activated node The probability of.

[0033] This embodiment performs information diffusion under an independent cascaded model, and the diffusion process is a discrete time-step process: at the initial time... Given a selected set of seed nodes In this set, nodes are active, and the rest are inactive; at any given time... Each The node that has just been activated Each node has one and only one chance to try to activate each of its currently inactive neighboring nodes. The probability of successful activation is ;if Successfully activated ,but exist A node becomes active at a certain moment. Once a node becomes active, it will remain active until the diffusion process ends; the diffusion process terminates when no new nodes are activated in the network.

[0034] Next, network feature preprocessing and input construction are performed, taking the input social network... This is transformed into an adjacency matrix A and an edge weight matrix W, where W stores the activation probabilities of each edge. Then, feature initialization is performed on the network for each node. Constructing the initial feature vector This vector contains the node's static topological attributes (such as in-degree, out-degree, and local clustering coefficients) and a dynamic identifier reflecting the node's current state, such as whether the node has been selected into the seed set via one-hot encoding or a binary mask. Finally, the number of seed nodes is preset to k, and an empty seed set is initialized. .

[0035] Step S2: Input the preprocessed graph data and node features into the cascaded perceptual graph neural network, perform multiple rounds of cascaded perceptual message passing, and generate the embedded representations of all candidate nodes in the target social network.

[0036] In this embodiment, in each iteration, the current node Aggregate its neighbor nodes The feature information, combined with the activation probability The weighting adjustment is performed using the following formula: ; in, For the first Layer nodes The embedding vector; For the first Layer nodes The embedding vector; For the first Layer nodes The embedding vector; Gather for neighbors; For feature splicing operations; This is a cascade factor function used to reflect propagation attenuation under different paths.

[0037] In this embodiment, It is a cascaded sensing function used to adjust neighboring nodes based on the diffusion probability of edges. For the current node The degree of feature contribution makes the generated embedding representation It can explicitly include the "expected propagation increment" originating from that node. Regarding multi-scale information fusion, by stacking multiple cascaded perceptual layers, the model can capture cascading effects within a propagation range of 1 to L steps. After L layers of aggregation, a global pooling layer or a non-linear fully connected layer fuses the features from the multiple layers to extract the final embedding representation matrix of all candidate nodes in the target social network. .

[0038] Step S3: The policy network and evaluation network, which are embedded to represent inputs to the deep reinforcement learning, are used to calculate the value function of the current candidate node's action using the Manhausen reinforcement learning mechanism, and to evaluate the influence potential of each candidate node.

[0039] In this embodiment, to solve the influence maximization problem using reinforcement learning, the seed selection process is modeled as a multi-step decision-making process. At each step... In the middle, state Represents the current network topology and the distribution of selected seed nodes, actions From the current set of non-seed nodes Choose a new node to add to the seed set, and receive a reward. Representative select node The resulting increase in influence.

[0040] The node embedding matrix H is input into the policy network and evaluation network of the deep reinforcement learning, and the action of each candidate node is calculated using the Mannhausen reinforcement learning mechanism. Value function In this embodiment, to address potential Q-value estimation biases during searches in a very large node space, a Mannhausen regularization term is introduced to correct the action values ​​during the calculation process. ; in, The modified action value function after introducing the Manhausen regularization term; State; This is the scaling factor; For the Manhausen item, The temperature is the entropy temperature. It is used to reward stable actions consistent with the current policy, and improves decision stability in a large search space by increasing policy entropy to suppress overestimation bias of the value function.

[0041] In this embodiment, the algorithm incorporates a KL divergence constraint between the current policy and the target policy into the loss function. This forces the policy to remain smooth during updates, avoiding reward fluctuations caused by some "super-connection nodes" in the social network, thus enabling the model to learn more general influence propagation patterns. The immediate reward *r* generated after each round of action is defined as the marginal influence increment. ; Where σ(·) represents the influence propagation scale function.

[0042] Step S4: Select the candidate node with the highest score based on the evaluation results and add it to the seed set, and update the node state mask.

[0043] Based on the calculated revised influence potential score, select the node with the highest score from the candidate node set. Select the node Add to seed set Update the node state mask, mark selected nodes as unselectable, and provide an immediate reward r. The immediate reward r is determined based on the node's state. The expected increase in influence under the current graph topology Sure.

[0044] Step S5: Determine if the size of the current seed set S has reached the preset budget. If... If so, jump back to step S2, re-extract features based on the updated node state, and proceed to the next round of selection; if If the condition is met, the process terminates and the final set of seed nodes is output. .

[0045] In this implementation, the final seed set is constructed using an iterative, incremental approach during the method execution phase, specifically as follows: Through the seed masking mechanism, a negative bias mask is applied to the nodes that have been selected into the seed set in each round of selection, ensuring that the same node is not selected repeatedly in subsequent steps, and enabling the cascaded perceptual graph neural network to readjust the perceptual features according to the distribution of the remaining unselected nodes. Through an end-to-end inference process, a real-time snapshot of the target social network is input into the trained cascaded perceptual graph neural network during model deployment, and the node with the largest current influence increment is directly selected based on the probability output of the policy network. By generalizing across scales, after training on small-scale graph data, it can be directly applied to larger-scale target social networks to perform inference.

[0046] like Figure 2 The figure shows the performance results of this embodiment and existing technologies on multiple real social network datasets. When processing ultra-large-scale networks (such as Douban), the method proposed in this embodiment achieves a performance gain of more than 14% compared with similar deep learning methods, effectively solving the problem of insufficient perception of high-level hierarchical features by traditional methods.

[0047] Example 2 The difference between this embodiment and Embodiment 1 lies in the different implementation architecture of the reinforcement learning framework in the decision optimization module. In step S3 of Embodiment 1, a value function-based Manhausen deep Q-network architecture is used. In this embodiment, it is replaced by a policy gradient-based reinforcement learning architecture, such as the proximal policy optimization algorithm (PPO) or the soft actor-critic algorithm (SAC). Under this alternative architecture, the logarithmic policy compensation term provided by the Manhausen mechanism is transformed into entropy regularization constraints or KL divergence constraints in the policy loss function to achieve smooth processing of the seed selection policy.

[0048] The other steps in this embodiment, including graph data acquisition and preprocessing in step S1, cascaded sensing message passing and embedded representation generation in step S2, node selection and state mask update in step S4, and loop termination judgment in step S5, are exactly the same as in embodiment 1.

[0049] Based on the disclosure and teachings of the foregoing specification, those skilled in the art can make changes and modifications to the above embodiments. Therefore, the present invention is not limited to the specific embodiments disclosed and described above, and some modifications and changes to the invention should also fall within the protection scope of the claims of the present invention. Furthermore, although some specific terms are used in this specification, these terms are only for convenience of explanation and do not constitute any limitation on the present invention.

Claims

1. A method for predicting user influence for rapid evaluation of information dissemination effects on social networks, characterized in that, Includes the following steps: S1. Obtain the graph data of the target social network and perform preprocessing; S2. Input the preprocessed graph data and node features into the cascaded perceptual graph neural network, execute multiple rounds of cascaded perceptual message passing, and generate the embedded representation of all candidate nodes in the target social network; S3. Input the embedded representation into the policy network and evaluation network of the deep reinforcement learning, and use the Mannhausen reinforcement learning mechanism to calculate the value function of the action of the current candidate node, and evaluate the influence potential of each candidate node. S4. Based on the evaluation results, select the candidate node with the highest score to add to the seed set, and update the node state mask; S5. Determine whether the size of the current seed set has reached the preset budget: If not, repeat steps S2 to S4; if it has, output the current seed set as the final seed set for maximizing influence.

2. The method for predicting user influence for rapid evaluation of information dissemination effects on social networks according to claim 1, characterized in that, The process of acquiring and preprocessing the graph data of the target social network specifically includes: S1-1. Obtain the target social network data and abstract it into a graph structure. And calculate each edge based on historical interaction frequency or a preset propagation model. activation probability ; in, It is the set of nodes in the network, where n represents the total number of nodes. It is the set of edges in the network; S1-2. Initialize the network features for each node. Constructing the initial feature vector Features include the in-degree, centrality, and state mask indicating whether the current node has been selected as a seed. S1-3, Preset seed node quantity budget and initialize an empty seed set. .

3. The method for predicting user influence for rapid evaluation of information dissemination effects on social networks according to claim 1, characterized in that, The step of performing multi-round cascaded perceptual message passing to generate embedded representations of all candidate nodes in the target social network specifically involves: In each iteration, the current node Aggregate its neighbor nodes The feature information, combined with the activation probability The weighting adjustment is performed using the following formula: ； in, For the first Layer nodes The embedding vector; For the first Layer nodes The embedding vector; For the first Layer nodes The embedding vector; Gather for neighbors; For feature splicing operations; It is a cascaded factor function; go through After layer aggregation, the final embedding representation matrix of all candidate nodes in the target social network is extracted. .

4. The method for predicting user influence for rapid evaluation of information dissemination effects on social networks according to claim 3, characterized in that, After L-layer aggregation, the final embedding representation is extracted, specifically: By stacking multiple cascaded sensing layers, the cascade effect within the aggregation range from layer 1 to layer L is captured. The output features of multiple cascaded sensing layers are then fused through a feature fusion layer to obtain the final embedding representation matrix of all candidate nodes in the target social network.

5. The method for predicting user influence for rapid evaluation of information dissemination effects on social networks according to claim 1, characterized in that, The method of using the Mannhausen reinforcement learning mechanism to calculate the value function of the current candidate node's action and evaluate the influence potential of each candidate node is as follows: The action of each candidate node is calculated using the Mannhausen reinforcement learning mechanism. Value function Furthermore, during the calculation process, the action value is corrected by introducing a Mannhausen regularization term: ； in, The modified action value function after introducing the Manhausen regularization term; State; This is the scaling factor; For the Manhausen item, The temperature is the entropy temperature.

6. The method for predicting user influence for rapid evaluation of information dissemination effects on social networks according to claim 1, characterized in that, Based on the evaluation results, the candidate node with the highest score is selected and added to the seed set, and the node state mask is updated, specifically as follows: S4-1. Based on the evaluation results, select the node with the highest score from the set of candidate nodes. ; S4-2, Select the node Add to seed set ; S4-3. Update the node status mask, mark selected nodes as unselectable, and provide immediate rewards. Instant rewards According to the node The expected increase in influence under the current graph topology Sure.

7. A method for predicting user influence for rapid evaluation of information dissemination effects on social networks, as described in claim 6, is characterized in that... The instant reward Defined as: ； in, For the seed set; A function for the scale of influence propagation; For node actions.

8. The method for predicting user influence for rapid evaluation of information dissemination effects on social networks according to claim 1, characterized in that, During the method execution phase, the final seed set is constructed using an iterative, incremental approach, specifically as follows: Through the seed masking mechanism, a negative bias mask is applied to the nodes that have been selected into the seed set in each round of selection, ensuring that the same node is not selected repeatedly in subsequent steps, and enabling the cascaded perceptual graph neural network to readjust the perceptual features according to the distribution of the remaining unselected nodes. Through an end-to-end inference process, a real-time snapshot of the target social network is input into the trained cascaded perceptual graph neural network during model deployment, and the node with the largest current influence increment is directly selected based on the probability output of the policy network. By generalizing across scales, after training on small-scale graph data, it can be directly applied to larger-scale target social networks to perform inference.

9. A method for predicting user influence for rapid evaluation of information dissemination effects on social networks, as described in claim 1, is characterized in that... Reinforcement learning algorithms incorporate the KL divergence constraint between the current policy and the target policy into the loss function.

Images