Source-linked AI summary
BotRGCN: Twitter Bot Detection with Relational Graph Convolutional Networks
Shangbin Feng, Herun Wan, Ningnan Wang, Minnan Luo
TL;DR
Twitter bot detectors must recognize both disguised bots and coordinated bot communities. BotRGCN combines multimodal user information with a heterogeneous follow graph and relational graph convolutional networks. On TwiBot-20, it achieves state-of-the-art performance and outperforms baselines using follow relationships.
Problem
Twitter bot detection must handle bots that resemble genuine users and bots that act collectively in communities.
Method
BotRGCN jointly encodes multimodal user information and applies relational graph convolutional networks to a heterogeneous graph constructed from Twitter follow relationships.
Results
BotRGCN achieves state-of-the-art performance on TwiBot-20 and outperforms baselines that also use user follow relationships.
Takeaways & Limitations
The experiments support the effectiveness of BotRGCN’s user-information encoding and graph-learning approach for Twitter bot detection.
Abstract
from arXiv · showhide
Twitter bot detection is an important and challenging task. Existing bot detection measures fail to address the challenge of community and disguise, falling short of detecting bots that disguise as genuine users and attack collectively. To address these two challenges of Twitter bot detection, we propose BotRGCN, which is short for Bot detection with Relational Graph Convolutional Networks. BotRGCN addresses the challenge of community by constructing a heterogeneous graph from follow relationships and applies relational graph convolutional networks. Apart from that, BotRGCN makes use of multi-modal user semantic and property information to avoid feature engineering and augment its ability to capture bots with diversified disguise. Extensive experiments demonstrate that BotRGCN outperforms competitive baselines on a comprehensive benchmark TwiBot-20 which provides follow relationships.
I. INTRODUCTION
Twitter bot detection must handle bots that disguise themselves as genuine users and bots that act collectively through communities. BotRGCN addresses these challenges by combining multimodal user information with a heterogeneous graph built from follow relationships.
- Bots can induce undesirable social effects, creating an urgent need for robust Twitter bot detectors.
- Existing bot detectors use feature engineering or deep learning based on user, tweet, property, neighborhood, recurrent, or adversarial features.
- Disguise involves bots resembling genuine users, while community involves individually genuine-looking bots acting together toward malicious goals.
- BotRGCN uses numerical and categorical user properties, pretrained language models for tweets, and a heterogeneous Twitter graph with relational graph convolutional networks.
- Twitter bot detection identifies bots among users using descriptions, tweets, numerical and categorical properties, and neighborhood information.
B. User Feature Encoding
BotRGCN jointly encodes user descriptions, tweets, and numerical and categorical properties into a unified user representation. It uses pretrained language modeling, normalization, one-hot encoding, and learned transformations rather than manual feature engineering.
- BotRGCN jointly encodes user descriptions, tweets, numerical properties, and categorical properties as user features.
- User descriptions are encoded with pretrained RoBERTa and transformed into fixed-dimensional representation vectors.
- User tweets are encoded with RoBERTa, and their representations are averaged to obtain a user-tweet representation.
- Numerical properties are taken directly from the Twitter API, z-score normalized, and represented using MLPs and graph neural networks.
- Categorical properties are one-hot encoded, concatenated, and transformed with a fully connected layer and leaky-ReLU.
C. GNNs Architecture
BotRGCN represents Twitter users in a heterogeneous follow graph and applies relational graph convolutional networks to learn user representations.
- Graph construction: Twitter users are modeled as nodes connected by distinct following and follower relations.These two relation types define relational neighborhoods and can be extended when additional relation data are available.
- BotRGCN architecture: R-GCN layers are applied to the heterogeneous graph after transforming user features into initial node representations.The learned representations are subsequently transformed with an MLP.
- BotRGCN architecture: The architecture uses an MLP after R-GCN layers to produce each user’s final representation.The resulting representation is denoted h_i.
D. Learning and Optimization
BotRGCN performs bot detection from R-GCN-derived user representations using a softmax output and optimizes a regularized classification loss.
- Prediction: A softmax layer conducts Twitter bot detection from user representations derived by R-GCN.The output layer contains learnable parameters W_O and b_O.
- Optimization: BotRGCN’s loss combines supervised binary cross-entropy over annotated users with L2 regularization on learnable parameters.Y denotes annotated users, y_i the ground-truth label, and θ all learnable parameters.
A. Experiment Settings
Experiments use TwiBot-20 and compare BotRGCN with feature-based, sequence-based, graph-based, and self-supervised bot detectors, including architectural ablations.
- Dataset: TwiBot-20 provides follow relationships for evaluating Twitter bot detection with BotRGCN.The resulting heterogeneous graph contains 229,580 nodes and 227,979 edges.
- Baseline methods: The baseline set includes random-forest methods using account metadata, tweet content, or engineered user features.It also includes a BiLSTM detector, anomaly detection, and Botometer’s feature-based service.
- Baseline methods: Graph-oriented comparisons include Cresci et al.’s bot-group analysis, graph convolutional networks, and SATAR’s self-supervised user representations.These methods provide comparisons involving online-action sequences, graph structure, or follow relationships.
- Ablation settings: Figures 2 and 3 evaluate reduced user feature sets and alternative graph neural network types, respectively.The studies examine the contributions of user information and graph-learning choices.
B. Bot Detection Performance
On TwiBot-20, BotRGCN achieves state-of-the-art performance and outperforms baselines that also use user follow relationships.
- Overall performance: BotRGCN achieves state-of-the-art performance among all evaluated methods on TwiBot-20.The authors describe this result as demonstrating general effectiveness for Twitter bot detection.
- Overall performance: BotRGCN outperforms Alhosseini et al. and SATAR, which also leverage user follow relationships.The authors attribute this comparison to better utilization of follow relationships in users’ social context.
C. User Feature Study
The ablation study shows that every aspect of user information contributes to BotRGCN’s performance, with categorical properties contributing most.
- Every aspect of user information is essential to BotRGCN’s performance.
- User categorical properties contribute most to BotRGCN’s performance.
D. GNN Study
The GNN study examines alternative graph models and R-GCN depths, finding that R-GCN supports performance and that two layers offer a favorable configuration.
- Replacing R-GCN with GAT, GCN, or MLP indicates that R-GCN contributes to BotRGCN’s performance.
- BotRGCN with 2 R-GCN layers achieves better bot detection performance with fewer learnable parameters and less training complexity.
IV. CONCLUSION
BotRGCN jointly encodes multi-modal user information, represents Twitter with a heterogeneous graph, and applies relational graph convolutional networks to address bot disguise and communities. Experiments demonstrate its efficacy, while further studies identify its information encoding and graph learning approaches as essential to performance.
- BotRGCN jointly encodes multi-modal user information, constructs a heterogeneous Twitter graph, and applies relational graph convolutional networks.
- BotRGCN is designed to tackle bot disguise and bot communities.
- Extensive experiments demonstrate BotRGCN’s efficacy compared with state-of-the-art baseline methods.
- Further explorations show that BotRGCN’s user information encoding strategy and graph learning approach are essential to performance.