On Anomaly Detection in Graphs as Node Classification

Abstract

Graphs have been exercised as appealing candidates for modeling relational datasets in different domains such as cryptocurrency transaction networks, social networks, rating platforms, and many more. Recently, different powerful methods have emerged to analyze datasets where the complex underlying data connectivity can be modeled by graphs. These methods have demonstrated promising performance on graph common tasks including node classification. In this paper, we explore the impact of graph-based techniques for detecting anomalous entities on real-world networks. We focus on modeling the problem of detecting anomalous entities on a network as a node classification task, and inspect the role of different approaches together with the evaluation setup and metrics to provide several useful recommendations for practical applications. We investigate different ways of handling the imbalance issue of the datasets which is a common problem when dealing with datasets containing anomalies, and demonstrate how a method that is agnostic to the dataset imbalance may show misleading performance. Through extensive experiments on six real-world datasets in balanced and unbalanced setting for a node classification task, we provide several recommendations that can shed more lights on challenges of selecting the appropriate methods, settings, and performance metrics that better align with the intrinsic attributes of a specific dataset and task.