Non-Intrusive Graph-Based Bot Detection for E-Commerce Using Inductive Graph Neural Networks

Malicious bots abuse e-commerce services while evading conventional defenses. IP/rule blocking is brittle under proxy rotation, and CAPTCHAs add friction yet are often bypassed. We propose a non-intrusive framework that models session-URL interactions as a bipartite graph and uses an inductive GNN (GraphSAGE) to classify session nodes. Combining topology with lightweight behavioral and URL semantics enables detection of “feature-normal” automation. On real-world traffic with high-confidence bot labels, GraphSAGE outperforms a session-feature MLP baseline in AUC and F1, and remains robust under mild adversarial edge perturbations and in cold-start inductive evaluation-supporting real-time deployment without client-side instrumentation.