Anomaly detection is a critical task in domains such as finance, healthcare, cybersecurity, and the Internet of Things, where identifying rare and irregular patterns is essential. Graph Autoencoders (GAEs) have emerged as a powerful tool by leveraging neighborhood information to detect hidden anomalies. However, a single GAE is highly sensitive to initialization, architecture, and loss functions, which limits its stability and generalization. To overcome these limitations, we introduce the Ensemble of Graph Autoencoder (EGA) for Unsupervised Anomaly Detection, a framework that introduces diversity through four ensemble mechanisms: RandNet, Bagging, Boosting, and the Random Subspace Method. We further explore the impact of ensemble size, graph similarity measures (Cosine, Euclidean, Pearson), and reconstruction losses (MSE, BCE, Pearson correlation) on ensemble performance. Experiments on several real-world datasets demonstrate that EGA consistently outperforms single GAEs and classical anomaly detection methods. Furthermore, Wilcoxon signed-rank tests confirmed that the improvements achieved by EGA are statistically significant across datasets. Notably, ensembles of size 10 models provide better results, highlighting the effectiveness of ensemble design for graph-based anomaly detection.
Nawaz et al. (Sun,) studied this question.