C-MTAD-GAT: Unsupervised Anomaly Detection for Mobile Networks

Researchers propose C-MTAD-GAT (Context-aware Multivariate Time-series Anomaly Detection with Graph Attention), an unsupervised framework for detecting anomalies in large-scale mobile networks. The model combines temporal and feature-wise graph attention with static and dynamic context conditioning, plus a dual-head decoder for reconstruction and forecasting. It produces per-element, per-feature anomaly scores and uses unsupervised thresholds from validation residuals. The approach addresses the impracticality of supervised learning due to labeling costs and handles context shifts and nonstationarity. The framework is designed as a single shared model across heterogeneous network elements in the radio access network and packet core, monitoring high-dimensional KPI time series. The TELCO dataset is used for evaluation.