A Hybrid Two-Stage Framework for Twitter Bot Detection: Decoupling Graph Representation Learning and Classification

Abstract

Graph-based Twitter bot detection models such as BotRGCN rely on end-to-end training that jointly optimizes graph representation learning and linear classification. This coupling creates a bottleneck: the GNN is constrained to produce linearly separable representations, and joint optimization restricts the classifier to differentiable models. This study proposes a hybrid two-stage framework that decouples these two components. In Stage 1, BotRGCN serves as a graph-aware feature extractor. In Stage 2, XGBoost is trained independently on the frozen embeddings, a non-differentiable classifier only accessible through decoupling. Across 5 random seeds on TwiBot-22, where severe class imbalance and million-scale heterogeneous structure make each F1 score reflect substantial discriminative gain, the proposed method achieves F1 of 60.41 (+2.81 over the BotRGCN baseline) while reducing cross-seed variance from 0.78 to 0.26. The improvement is consistent across GCN, GAT, and BotRGCN architectures, suggesting that coupling, rather than architecture choice, may be a significant bottleneck. Classifier ablation further shows that even Logistic Regression on frozen embeddings outperforms end-to-end training, indicating that decoupling serves as a key mechanism.