An Empirical Study of GNN-Based Anomaly Detection in Smart Home Environments

Abstract

This thesis presents a systematic comparative study of Graph Neural Network (GNN)-based anomaly detection architectures for smart home multivariate time-series data. The study focuses on two representative architectures: Graph Deviation Network (GDN), a sparse forecasting-based graph model, and MTAD-GAT, a dense attention-based hybrid architecture combining forecasting and reconstruction objectives. Experiments are conducted using the BRE and CU smart home datasets as well as a merged heterogeneous dataset configuration under a unified experimental framework with consistent preprocessing, training, threshold￾ing, and evaluation protocols. The results show that reconstruction-enhanced anomaly scoring in MTAD-GAT provides greater robustness and statistical separability in heterogeneous smart home environments, while GDN produces sharper but more localized anomaly responses through sparse depen￾dency modeling. Threshold sensitivity analysis further demonstrates that dense attention￾based architectures maintain more stable operating regions across varying detection thresh- olds. In addition, graph and device-level analyses reveal distinct interpretability behaviors between sparse localized graph structures and distributed attention-based dependency mod eling.