A Cascade GCN-LightGBM Framework for Abnormal Account Detection in Social Networks

Abstract

The detection of abnormal accounts in social networks (such as zombie accounts and Spam accounts) faces challenges such as dynamic changes in features, cyberspace complexity, and low efficiency of massive data processing. This paper proposes a cascade model (GCN-LightGBM) that fuses a graph convolutional network (GCN) and LightGBM. GCN is used to capture topological relationships between accounts to generate structural features, and then input into LightGBM for efficient classification. Experiments on public data sets show that the F1 value of the model reaches 88.5%, which is significantly higher than that of single GCN (87.9%) and XGBoost (85.5%), and the training speed is increased by about 20 times, making it a lightweight anomaly detection in dynamic environments. Provide new solutions.