[Paper Review] BotMoE: Twitter Bot Detection with Community-Aware Mixtures of Modal-Specific Experts

Yuhan Liu, Zhaoxuan Tan+4 more|arXiv (Cornell University)|Apr 13, 2023

Misinformation and Its Impacts4 citations

TL;DR

BotMoE proposes a community-aware mixture-of-experts framework that jointly models metadata, textual content, and graph structure to detect deceptive Twitter bots. By using modal-specific encoders and a community-aware MoE layer, it improves generalization and robustness, achieving state-of-the-art performance on three benchmarks while better handling manipulated features and diverse communities.

ABSTRACT

Twitter bot detection has become a crucial task in efforts to combat online misinformation, mitigate election interference, and curb malicious propaganda. However, advanced Twitter bots often attempt to mimic the characteristics of genuine users through feature manipulation and disguise themselves to fit in diverse user communities, posing challenges for existing Twitter bot detection models. To this end, we propose BotMoE, a Twitter bot detection framework that jointly utilizes multiple user information modalities (metadata, textual content, network structure) to improve the detection of deceptive bots. Furthermore, BotMoE incorporates a community-aware Mixture-of-Experts (MoE) layer to improve domain generalization and adapt to different Twitter communities. Specifically, BotMoE constructs modal-specific encoders for metadata features, textual content, and graphical structure, which jointly model Twitter users from three modal-specific perspectives. We then employ a community-aware MoE layer to automatically assign users to different communities and leverage the corresponding expert networks. Finally, user representations from metadata, text, and graph perspectives are fused with an expert fusion layer, combining all three modalities while measuring the consistency of user information. Extensive experiments demonstrate that BotMoE significantly advances the state-of-the-art on three Twitter bot detection benchmarks. Studies also confirm that BotMoE captures advanced and evasive bots, alleviates the reliance on training data, and better generalizes to new and previously unseen user communities.

Motivation & Objective

• To address the challenge of detecting advanced Twitter bots that manipulate metadata, text, and network features to evade detection.
• To improve model generalization across diverse Twitter user communities where bot behavior varies significantly.
• To reduce reliance on large-scale labeled training data by enhancing data efficiency through multi-modal fusion.
• To develop a more robust detection framework that can adapt to unseen communities and evade sophisticated bot disguises.

Proposed method

• BotMoE employs three modal-specific encoders to process user metadata, textual content, and graph structure independently.
• A community-aware Mixture-of-Experts (MoE) layer dynamically routes users to community-specific expert networks based on their structural and behavioral features.
• Each expert network specializes in detecting bot patterns within a specific community, improving local detection accuracy and domain adaptation.
• An expert fusion layer combines representations from all modalities and measures cross-modal consistency to enhance detection reliability.
• The model uses a routing mechanism that learns to assign users to the most relevant experts based on community embeddings derived from graph structure.
• The framework is trained end-to-end with a contrastive loss to encourage alignment between multi-modal representations and improve generalization.

Experimental results

Research questions

• RQ1 Can a multi-modal, community-aware model outperform existing single- or two-modality approaches in detecting advanced Twitter bots?
• RQ2 How effectively can the community-aware MoE layer adapt to previously unseen user communities during inference?
• RQ3 To what extent does the model reduce reliance on large-scale labeled training data while maintaining high detection accuracy?
• RQ4 How do different modalities (text, metadata, graph) contribute to detection, and can the model identify the most informative features?
• RQ5 What is the impact of feature manipulation on model performance, and how well does BotMoE resist evasion through multi-modal consistency modeling?

Key findings

• BotMoE achieves state-of-the-art performance on three major Twitter bot detection benchmarks, significantly outperforming prior methods in F1-score and AUC.
• The model demonstrates superior generalization to new, previously unseen user communities, indicating strong domain adaptation capabilities.
• Extensive ablation studies confirm that the community-aware MoE layer and expert fusion mechanism are critical for performance, especially in handling feature manipulation.
• Attention weights from the fusion layer show that metadata and graph features contribute more reliably than text, highlighting the importance of multi-modal consistency.
• The model reduces false positives by leveraging structural and metadata cues even when textual content is misleading or stolen from genuine users.
• The framework shows improved data efficiency, maintaining high performance with limited labeled data, which is crucial for real-world deployment.