Existing Twitter bot detection methods often overlook the structural and temporal information of users’ dynamic social history, as well as the noise accumulation resulting from feature fusion. In order to address these limitations, this paper constructs STKG (social temporal knowledge graph) and proposes a Twitter bot detection method STKGBot (STKG for Twitter bot detection). In the STKG, STKGBot uses RE-GAT (heterogeneity-enhanced graph attention network) to learn the static social relationship feature, TE-GCN (temporal-enhanced graph convolutional network) to learn the dynamic social history feature, and a bilinear model for the feature fusion. In addition, STKGBot employs contrastive learning to alleviate the noise aggravation in the process of feature fusion. Experimental results on two public datasets demonstrate that STKGBot outperforms state-of-the-art models.

为了解决现有推特机器人检测方法对用户动态社交历史结构性和时序性的忽视以及特征融合引起的噪声累积问题, 构建社交时序知识图谱STKG, 并提出一种推特机器人检测方法STKGBot。在STKG中, STKGBot使用关系增强图注意网络RE-GAT来学习静态社交关系特征, 使用时序增强图卷积网络TE-GCN来学习动态社交历史特征, 使用双线性模型进行特征融合。此外, STKGBot使用对比学习来缓解特征融合引起的噪声累积。在两个公开数据集上的实验结果表明, STKGBot的检测结果均优于当前最先进的模型。