GNNRI: detecting anomalous social network users through heterogeneous information networks and user relevance exploration

Detecting anomalous users in social networks is an imperative but challenging task. The increasing complexity of inter-personal behaviors and interactions further complicates the development of effective user anomaly detection techniques. Current state-of-the-art methods heavily rely on static personal features, making it difficult to quantify the hidden relevance of user behaviors through traditional feature engineering. This loss of accuracy is exacerbated by the rise of sophisticated camouflage and disguising techniques, which blur the distinction between anomalous and regular users. In this paper, we present GNNRI, an innovative framework for detecting anomalous users in social networks. Our approach leverages a network representation learning model and a heterogeneous information network (Hin) to explore hidden semantic connections from user metadata, tweets, and interaction information. We extract both user metadata and behavioral features to construct a Hin and introduce two distinct learning layers to explore explicit and implicit user relevance. First, we employ a relation-based self-attention layer to aggregate neighbor node closeness under specific relations and across different relationships. Subsequently, we apply graph convolution network-based convolutional learning layers, which enhance embedding effectiveness by capturing graph-wide node similarity. We evaluate GNNRI using real-world datasets, and our results demonstrate that it outperforms all other comparative baselines, achieving approximately 90% accuracy for user classification, with a 5–15% improvement over other GNN variants. Notably, even when using only 20% of the data for training, GNNRI achieves 87.8%, 86.57%, and 87.1% accuracy for detecting zombies, spammers, and bots, respectively.