Dual Graph Neural Networks for Social Bot Detection: Social Relation with an Auxiliary Behavior Similarity Graph

Abstract

Social bots threaten the integrity of online social platforms by spreading misinformation, manipulating public opinion, and carrying out fraudulent activities. Detecting social bots is therefore not just a classification problem: looking ahead, a detector can also act as a discriminative component in agent-based social network generation, supplying realism-assessment signals for a “Generate–Detect–Explain–Refine” loop. Existing graph-based detection methods rely mainly on social relation graphs and do not explicitly treat behavioral similarity as a separate parallel graph branch. This paper proposes a dual-branch graph neural network for Twitter bot detection in which a heterogeneous social relation graph is accompanied by a sparse, conditionally active behavior-similarity auxiliary view; the paper’s main contributions are a competitive detector and a diagnostic protocol for attributing and repairing the value of that auxiliary view. The relation branch uses a Relational Graph Convolution Network to process a heterogeneous social relation graph, while the auxiliary branch applies a Graph Attention Network to a behavior similarity graph that links users with similar follower–following ratios, tweet type distributions, and post influence metrics. Experiments on TwiBot-22 show that the proposed model reaches 89.48 ± 0.07% accuracy and 78.11 ± 0.14% macro-F1 over five random seeds at a per-seed validation-calibrated decision threshold (89.92 ± 0.07% accuracy and 77.28 ± 0.31% macro-F1 at the default 0.5 threshold). Per-class and threshold analyses further show that the default operating point is precision-oriented on the minority bot class (bot recall 54.83 ± 1.60%), while operating near the validation-calibrated threshold (τ = 0.40) raises bot recall to 64.12 ± 1.31%. Ablation studies show that the semantic text embedding provides the largest single performance contribution and that the relation branch supplies the bulk of the graph signal. A degree-stratified analysis further shows that the behavior branch is functionally active on its ≈1% covered users—2.9× more prediction changes and 4.0× larger probability shifts than on uncovered nodes—where a validation-only subgroup threshold lifts macro-F1 from 51.85 to 67.17 without retraining, so that it can be turned into a deployable calibrated component.