计算机科学与探索

• 学术研究 •    下一篇

基于强化学习的社交网络影响力最小化模型GCNNs-DDQN

陈梓彦,袁得嵛,孙泽宇,程佳琳   

  1. 1. 中国人民公安大学 信息网络安全学院,北京 100038
    2. 安全防范与风险评估公安部重点实验室,北京 102623

A social network influence minimization model GCNNs-DDQN based on reinforcement learning

CHEN Ziyan, YUAN Deyu, SUN Zeyu, CHENG Jialin   

  1. 1.School of Information Network Security, People's Security University of China, Beijing 100038, China
    2.Key Laboratory of Public Security Ministry for Security and Risk Assessment, Beijing 102623, China

摘要: 谣言的传播迅速且危害较大,因此,利用辟谣信息对谣言传播进行抑制对于社会稳定有重要意义。为了使已经传播的谣言的影响力尽快减小并消失,提出一种结合图神经网络GCN、GNN、Double DQN的动态辟谣GCNNs-DDQN模型。首先使用GCN获得节点作为源节点和目标节点的初始节点嵌入,结合节点状态向量,将其作为输入传入4个耦合的GNN以获得复杂节点嵌入,最后进行Q值计算,同时使用Double DQN结合Q值与奖励,优化4个耦合的GNN和Double DQN模型的权重参数,从而实现根据谣言传播的情况,选择当前情况下影响力最大的节点发布辟谣信息。使用10个不同大小的真实数据集来评估模型的辟谣效果,并对各参数对辟谣效果的影响进行分析,最后进行消融实验。实验结果表明GCNNs-DDQN模型具有更强的泛化能力,能够应用于不同的社交网络中,相较于传统算法,辟谣时间最高可缩短2个单位时间;对谣言有利的参数值的增加,会加大辟谣难度,增加阻断时间,而对辟谣信息有利的因素,则不一定会减少阻断时间,反而可能阻碍辟谣。

关键词: 谣言影响力最小化, 图神经网络, Double DQN

Abstract: The spread of rumors is rapid and poses significant harm, thus, utilizing disinformation to curb the dissemination of rumors is of great importance for social stability. To minimize and eliminate the influence of already disseminated rumors as soon as possible, a dynamic rumor refutation model named GCNNs-DDQN, combining Graph Convolutional Networks (GCN), Graph Neural Networks (GNN), and Double Deep Q-Network (DDQN), is proposed. First, GCN is used to obtain the initial node embeddings for both source nodes and target nodes. These embeddings, combined with node state vectors, serve as inputs to four coupled GNNs to derive complex node embeddings. Subsequently, Q-value computation is performed, and Double DQN is employed to optimize the weight parameters of the four coupled GNNs and the Double DQN model using Q-values and rewards. This approach enables the selection of the most influential nodes to release rumor-correcting information based on the current state of rumor propagation. Ten real-world datasets of varying sizes were used to evaluate the rumor refutation effectiveness of the model, and an analysis was conducted on the impact of various parameters on the rumor refutation performance. An ablation experiment was also conducted. The experimental results demonstrate that the GCNNs-DDQN model possesses stronger generalization capabilities and can be applied to different social networks. Compared to traditional algorithms, it can reduce the rumor refutation time by up to 2 unit times. An increase in parameter values that favor rumors will make it more difficult to counteract the rumors and prolong the time needed for their suppression. Conversely, factors that favor disinformation may not necessarily reduce the suppression time; instead, they could potentially hinder the effort to counteract rumors.

Key words: Rumor influence minimization, Graph neural network, Double DQN