Social Bot Detection via Spatio-Temporal Graph Self-Supervision

Abstract

Social bots on online social platforms increasingly exhibit profile camouflage, coordinated group behavior, and evolving interaction patterns over time, which makes static-graph and fully supervised detectors less effective under label scarcity and cross-community shifts. To address this issue, this paper proposes STGSLBot, a spatio-temporal graph self-supervised framework for social bot detection. The method constructs a sequence of dynamic interaction snapshots with fixed temporal granularity and initializes node features by combining account metadata with tweet semantics. In addition, temporal statistical signals are introduced to characterize coordinated structures and dynamic reciprocal behaviors. At the model level, a temporal dual-encoder is designed to learn topology-driven commonality and attribute-driven uniqueness, while a causal masked temporal self-attention module aggregates historical information under a strict history-only constraint. During pretraining, neighborhood reconstruction, self-feature reconstruction, cross-view semantic consistency learning, and community-based pseudo-label self-training are jointly optimized to enhance cluster-level discrimination and temporal consistency. Finally, the pretrained model is fine-tuned with a small labeled set for binary classification. Experiments on two temporal subsets derived from TwiBot-22 show that the proposed framework achieves strong and stable performance, especially in accuracy and label-efficient learning.