融合时序文本与高阶交互拓扑的在线抗议预测

针对在线抗议预测技术中忽视用户文本时序差异性及用户间高阶交互拓扑的问题,提出融合时序文本与高阶交互拓扑的在线抗议预测方法.基于自注意力机制建模用户不同时刻交互文本信息对其抗议倾向的影响,构建用户文本表示向量;同时利用邻域节点的相似性,构建二阶相似性保持的用户交互拓扑表示向量;融合用户文本表示向量和交互表示向量预测用户抗议倾向.推特数据集结果表明本方法准确率可达到93.9%,为抗议活动预测提供技术支撑.

Online Protest Prediction with Time-Series Text and High-Order Interactive Topology

Aiming at the problem of neglecting user text timing differences and high-level interactive topologies among users in online protest prediction, combining temporal text and high-order interactive topology, an online protest prediction method was proposed. Modeling the influence of the text information published by users at different moments on their current protest tendency based on a self-attention mechanism, the user text representation vector was constructed. At the same time, the similarity of the neighbor nodes was used to construct the user interaction topology representation vector, maintaining the second-order similarity. Synthesizing the user text representation vector and the interactive representation vector, the user protest tendency was predicted. The results of the Twitter dataset show that the accuracy of the method can reach 93.9%, providing technical support for protest prediction.