OTPCL: Optimal Transport Driven Pseudo-Labeling with Contrastive Learning for Social Bot Detection

Abstract

Social bot detection is vital for protecting online platforms from misinformation and manipulation. In recent research, graph neural networks (GNNs) have emerged as a powerful approach, since they leverage relational patterns and social interactions to identify coordinated bot behaviors. However, two key challenges arise from the nature of real-world social networks: First, bots often actively interact with human users through follows, replies, and mentions, creating numerous ''heterophilous'' edges, i.e., connections between different classes. These cross-class ties disrupt the homophily assumption underlying many GNNs, causing messagepassing to propagate and amplify errors. Second, due to the high cost and time required for manual annotation, social media platforms typically contain a large proportion of unlabeled data, with only a small fraction labeled for bot detection. Unlabeled data are often underutilized, making supervision sparse. To address this, we propose OTPCL (Optimal Transport Driven Pseudo-Labeling with Contrastive Learning), a plug-in framework for GNN-based social bot detection. OTPCL first employs contrastive learning to obtain well-separated node representations. It then formulates pseudolabel assignment as an optimal transport problem, which simultaneously generates pseudo-labels and quantifies their reliability via transport scores. These scores guide two key mechanisms: selective removal of unreliable heterophilous edges to purify the graph structure, and reducing the influence of pseudo-labels with transport scores below the dynamic threshold. Extensive experiments on three widely used benchmark datasets demonstrate that OTPCL consistently improves the detection performance across six different GNN backbones, showing strong robustness and generalization in both low-labeled and fully-labeled scenarios.