MGTAB-Live: A Production-Ready Multi-Relational Graph-Based Framework for Real-Time Twitter Bot Detection Using Relational Graph Convolutional Networks

The proliferation of sophisticated automated accounts on social media platforms poses a critical threat to the integrity of online discourse, democratic processes, and public opinion formation. Existing bot detection methodologies, which predominantly rely on metadata heuristics or isolated textual analysis, exhibit diminishing efficacy against modern adversarial bots that leverage large language models for content generation and forge realistic social connections. In this paper, we present MGTAB-Live, a production-grade, full-stack framework that operationalizes the Multi-relational Graph-based Twitter Account Bot (MGTAB) detection benchmark as a real-time, deployable web application. Our system constructs dynamic multi-relational ego-graphs comprising seven semantically distinct edge types - follower, friend, mention, reply, quote, URL co-occurrence, and hashtag co-occurrence-and encodes each node with a 788-dimensional feature vector amalgamating 20 normalized profile attributes with 768-dimensional Language-Agnostic BERT Sentence Embeddings. A two-layer Relational Graph Convolutional Network performs heterogeneous message passing over these typed edges, achieving 88.23% test accuracy and 90.29% bot recall on the MGTAB dataset of 10,199 expert-annotated Twitter accounts-outperforming Graph Convolutional Network (79.21%), Graph Attention Networks (81.67%), and GraphSAGE (87.16%) baselines. We detail the novel engineering contributions required for production deployment, including a rate-limit-resilient scraper architecture, a Server-Sent Events streaming pipeline for real-time progress reporting, and a graceful degradation strategy for partial ego-graphs. The system is deployed on Hugging Face Spaces with a React 19 frontend, demonstrating the feasibility of graph neural network-based bot detection as a practical, user-facing service.