Multi-Scale Contrastive Learning Networks for Graph Anomaly Detection

The detection of anomalies in attributed networks has attracted significant attention because of the versatile applications of utilizing networks to simulate intricate systems. Emerging developments in deep learning-driven anomaly detection techniques exhibit potential, particularly in networks characterized by attributes of high dimensionality and intricate formations. Nevertheless, existing methods, which primarily rely on graph autoencoders, might not fully utilize the wealth of data present in the network, leading to less than optimal performance. Moreover, these approaches often lack a direct focus on detecting anomalies during their learning objectives as well as may struggle to expand effectively to magnify networks because of the exhaustive educating process involving the entire graph. To address these limitations, we introduce the Multi-scale Contrastive Learning Networks for, Graph Anomaly Detection (MCLGAD for abbreviation) framework. This framework leverages a self-supervised learning with contrastive methods approach to enhance anomaly detection on networks with attributes through focusing on contrasting two sub graphs to effectively capture local information and relationships within the network. By sampling carefully designed instance pairs and training the contrastive learning model, MCLGAD aims to uncover various anomalies present within networks by learning matching patterns and diverse mismatching scenarios. The framework is specifically designed to provide discriminative scores for ranking anomalies, making it well-suited for analyzing large-scale network data. Experimental results demonstrate the effectiveness and superiority of MCLGAD compared to various baseline methods in graph anomaly detection tasks. Our implementations are available at https://github.com/jza123/MCLGAD/tree/5052d48d7cddd8cf530926d12ba64da3ea5ace46/MCLGAD.