Source-linked AI summary
Heterogeneity-aware Twitter Bot Detection with Relational Graph Transformers
Shangbin Feng, Zhaoxuan Tan, Rui Li, Minnan Luo
TL;DR
Twitter bot detection must address malicious automated accounts while overcoming topology-aware methods’ limited treatment of heterogeneous relations and influence. The paper constructs heterogeneous information networks and combines relational graph transformers with semantic attention networks for detection, consistently outperforming state-of-the-art baselines on a comprehensive benchmark. Additional studies support the effectiveness of the graph-learning strategy, including under limited training data.
Problem
Twitter bot detection is important because automated accounts can harm online discourse, while existing topology-aware methods neglect heterogeneous relations and influence among users.
Method
The framework represents users and diversified relations in heterogeneous information networks, learns representations with relational graph transformers, and aggregates cross-relation messages with semantic attention networks.
Results
The method consistently outperforms state-of-the-art baselines on a comprehensive benchmark, while ablations support the effectiveness of its graph-learning components and heterogeneity-aware design.
Takeaways & Limitations
Graph-based modeling of Twitter topology together with relation and influence heterogeneity is supported as an effective approach to bot detection, including with as little as 40% training data.
Takeaways & Limitations
The framework follows the state-of-the-art user-information encoding procedure for fairness, assuming user feature representations supplied through that procedure.
Abstract
from arXiv · showhide
Twitter bot detection has become an important and challenging task to combat misinformation and protect the integrity of the online discourse. State-of-the-art approaches generally leverage the topological structure of the Twittersphere, while they neglect the heterogeneity of relations and influence among users. In this paper, we propose a novel bot detection framework to alleviate this problem, which leverages the topological structure of user-formed heterogeneous graphs and models varying influence intensity between users. Specifically, we construct a heterogeneous information network with users as nodes and diversified relations as edges. We then propose relational graph transformers to model heterogeneous influence between users and learn node representations. Finally, we use semantic attention networks to aggregate messages across users and relations and conduct heterogeneity-aware Twitter bot detection. Extensive experiments demonstrate that our proposal outperforms state-of-the-art methods on a comprehensive Twitter bot detection benchmark. Additional studies also bear out the effectiveness of our proposed relational graph transformers, semantic attention networks and the graph-based approach in general.
Introduction
Twitter bot detection addresses automated accounts that threaten online discourse, but existing topology-aware methods overlook heterogeneous relations and influence. The paper proposes a graph-based framework that models this heterogeneity for detection.
- Motivation: Automated Twitter bots can pursue malicious goals, including election interference, misinformation, and extreme-ideology promotion, threatening online communities.The paper motivates effective detection as necessary to protect online discourse integrity.
- Prior approaches: Earlier detectors used engineered tweet and metadata features, neural models, and increasingly graph structure to represent the Twittersphere.Graph neural networks enabled topology-aware detection, but did not fully model Twitter’s heterogeneous interactions and influence.
- Research gap: Existing topology-aware methods overlook relation heterogeneity and influence heterogeneity, limiting their ability to recognize subtle differences between genuine users and novel bots.Users interact through distinct actions such as liking, commenting, retweeting, and blocking, while their influence varies in range and intensity.
- Proposed framework: The proposed framework constructs heterogeneous information networks with users as nodes and diversified relations as edges, then applies relational graph transformers and semantic attention networks.The transformers model influence intensity and learn node representations, while semantic attention aggregates messages across users and relations.
- Contributions: Extensive experiments on a comprehensive benchmark show that the proposal consistently outperforms baseline methods, with further studies supporting its graph-based and heterogeneity-aware design.The reported evaluation covers both the overall detector and its graph-learning strategy.
Related Work
Related work progresses from manually engineered user information and deep models toward graph-based bot detectors and heterogeneous information networks. Existing graph methods motivate architectures that more directly exploit relational structure.
- Feature-based and deep-learning methods: Early Twitter bot detectors combined manually designed features from tweets, metadata, or both with traditional classifiers.Later deep-learning systems used connected networks, recurrent neural networks, generative adversarial networks, and multimodal architectures.
- Graph-based detection: SATAR jointly leverages multimodal user information, but uses Twitter’s graph structure through feature engineering rather than state-of-the-art graph neural network architectures.This distinction positions graph neural networks as a separate direction for exploiting network structure.
- Graph-based detection: Graph-based detectors model Twitter as a user network or heterogeneous information network and apply graph convolutional or relational graph neural networks for bot detection.Relational graph methods extend graph convolutional approaches to networks containing diversified entity interactions and relations.
- Heterogeneous information networks: Heterogeneous information networks represent diversified and interactive entities and support analysis in social networks, graph mining, and natural-language processing.Relational graph convolutional networks were proposed to analyze such networks by extending standard graph convolutional networks.
Methodology
The framework represents Twitter as a heterogeneous information network, models relation-specific influence with relational graph transformers, and aggregates relation-aware representations with semantic attention before classification.
- Graph construction: The model constructs a heterogeneous information network with users as nodes and diversified Twitter interactions as relation-specific edges.
- Feature initialization: User features are transformed through a fully connected layer into initial GNN representations, using the same encoding procedure as prior work for fair comparison.
- Relational graph transformers: Relational graph transformers compute attention-based influence weights between nodes and aggregate relation-specific neighborhoods across multiple attention heads.
- Representation aggregation: A gate mechanism smooths the learned relation-specific representations before semantic attention aggregates information across relations.
- Prediction and training: After L GNN layers, an output and softmax layer predicts bot labels, while supervised annotations and a regularization term train the detector.
Experiments
The experiments use TwiBot-20, a large benchmark providing user-follow relationships, and compare the proposed approach with several established bot-detection methods.
- Dataset: TwiBot-20 contains 229,573 users, 33,488,192 tweets, 8,723,736 user property items, and 455,958 follow relationships.
- Compared methods: The benchmark comparison includes feature-engineering, traditional-classifier, sequence-based, and neural bot-detection methods.
Implementation
The implementation uses established PyTorch-based graph and transformer libraries, reports reproducibility settings, and documents model characteristics and hyperparameters in tables.
- Software: The framework is implemented with PyTorch, PyTorch Lightning, Torch Geometric, and the Transformers library.
- Method characteristics: Table 1 records whether methods use deep learning, interactions, user representations, graphs, and Twitter heterogeneity, alongside their performance.
- Hyperparameters: Table 2 reports the model hyperparameters and uses follower and following information as relations.
- Reproducibility: The implementation is trained on a Titan X GPU with 12GB memory, and the code is publicly available on GitHub.
Experiment Results
On TwiBot-20, the proposal is benchmarked against bot-detection methods and consistently outperforms all baselines, including BotRGCN. The experiments also examine graph structure and heterogeneity-related effectiveness.
- The proposal consistently outperforms all baselines, including the state-of-the-art BotRGCN, on TwiBot-20.
- The ablation study removes different parts of the constructed Twitter HINs to assess graph-structure contributions.
- Graph-based methods generally outperform traditional approaches that do not model Twitter as graphs or networks.
- The study examines the roles of graphs and heterogeneity, as well as data efficiency and representation learning capability.
Graph Learning Study
The graph-learning study evaluates both the heterogeneous graph construction and the proposed GNN architecture. Complete follower-following connectivity and the architecture’s main components are reported as important to performance.
- Graph construction: The model constructs a HIN from user follower and following relationships, treating them as two edge types between user nodes.
- Graph construction: The complete graph with both follower and following edges outperforms reduced edge settings in the ablation study.
- GNN architecture: Ablations show that relational graph transformers, the gate mechanism, and semantic attention networks are all essential architecture components.
- Overall contribution: Both the constructed HIN and the proposed GNN architecture contribute to the model’s performance, supporting the graph-based approach.
Heterogeneity Study
The study finds that modeling both relation and influence heterogeneity improves Twitter bot detection, with most heterogeneous relations contributing to performance and attention helping identify bot groups.
- Relation Heterogeneity: Most heterogeneous relation settings outperform homogeneous ones, supporting relation heterogeneity for Twitter bot detection.
- Relation Heterogeneity: Most heterogeneous relations contribute similarly to performance, whereas user interest domain information is less effective.
- Relation Heterogeneity: Incorporating relation heterogeneity improves bot detection performance, with most relations significant in model decisions.
- Influence Heterogeneity: Attention mechanisms outperform their absence, and multi-head attention generally outperforms single-head variants.The comparison uses C > 0, D > 0 versus C = 0, D = 0, and C > 1, D > 1 versus C = 1, D = 1.
- Influence Heterogeneity: Attention weights between bots are generally larger, helping the method identify bots that act in mutually influential groups.
- Influence Heterogeneity: Leveraging influence heterogeneity improves detection performance while exposing information about the model’s decisions.
Data Efficiency Study
The data-efficiency study evaluates reduced training data, interactions, and user features, finding that the method remains competitive with limited annotations but depends on comprehensive user information.
- The method still outperforms BotRGCN with as little as 40% training data and remains robust to changes in user interactions.
- Performance drops significantly when user features are reduced, indicating that detection relies on comprehensive user information alongside graph structure.
Representation Learning Study
The representation study compares user embeddings learned by the proposed method and baselines, finding stronger clustering of genuine users and bots for the proposed representations.
- The proposed method produces higher collocation for groups of genuine users and Twitter bots than the baselines.
- The stronger grouping indicates that the method learns high-quality user representations.
Conclusion and Future Work
The paper concludes that its graph-based, heterogeneity-aware framework consistently outperforms state-of-the-art baselines and remains effective across additional studies. Future work will explore more diversified graph representations and extend graph-based bot detection.
- The framework constructs HINs and combines relational graph transformers with semantic attention networks for Twitter bot detection.
- Extensive experiments on a comprehensive benchmark show that the method consistently outperforms state-of-the-art baselines.
- Further studies support the effectiveness of the graph-learning strategy and Twitter heterogeneity, including performance with limited data and high-quality user representations.
- Future work will investigate more diversified ways to model the Twittersphere as graphs and extend the graph-based bot detection approach.