APPLICATIONS OF GRAPH NEURAL NETWORKS IN OUTLIER DETECTION

Abstract

Graph Neural Networks (GNNs) have become a tool, in detecting outliers within graphs. When designing GNNs a key aspect is choosing a filter that suits the task. This research delves into outlier analysis by examining the graph spectrum and presents a finding; the presence of outlier leads to a ’right shift’ effect, where the energy distribution in the spectrum moves towards frequencies. This revelation carries implications for GNN design suggesting that con ventional low pass filters may not be ideal, for outlier detection. To address this challenge, we propose the Beta Wavelet Graph Neural Network (BWGNN), which incorporates spectral and spatial localized band-pass filters. These filters are specifically designed to handle the ‘right shift’ phenomenon, providing a more effective approach to outlier detection. We evaluate the performance of BWGNN on four large scale outlier detection datasets and demonstrate its su periority over existing methods. Our findings not only shed light on the spectral properties of graph outliers but also pave the way for more sophisticated GNN architectures that can better capture the nuances of anomalous behavior in graph data.