Unsupervised Graph-Based Anomaly Detection Using Variational Autoencoder

Anomaly detection is recognized as a critical task in machine learning, particularly when applied to high-dimensional and graph structured data. Traditional clustering based methods for anomaly detection are often challenged by such settings. In this paper, by leveraging Variational Graph Autoencoders (VGAEs) a new method is proposed to embed graph-structured data into a low-dimensional latent space, thereby enhancing anomaly detection through unsupervised learning. Synthetic labels, generated via spectral clustering, are incorporated to enable the model to learn from the intrinsic structure of the graph while improving cluster separability in transductive learning settings. Additionally, a dynamic early stopping mechanism based on Exponential Moving Average (EMA) is introduced to optimize training duration. Extensive experimental results demonstrate that VGAEs outperform earlier works in terms of accuracy and efficiency, particularly when applied to complex, noisy, and high-dimensional datasets. By combining deep learning with clustering, the proposed approach offers potential applications in areas such as healthcare and social network analysis, where the detection of anomalies and clustering of complex data are essential. By embedding graph data into meaningful clusters, this work addresses the challenges posed by high-dimensional and sparse data, contributing to the advancement of deep clustering models.