Abstract

In recent decades, anomaly detection has undoubtedly become one of the most important areas of research. This is because applications such as financial transactions, medical fraud, and anomaly detection can be used to solve a wide range of real-life problems. Data from these applications can be modeled using large graphs of many different nodes and edges. Because of the size and heterogeneity of the data contained in the graph, it is a very difficult task to detect abnormal patterns. In this paper, we proposed a method for detecting abnormal patterns in a large homogeneous graph. The proposed method consisted of two steps. In the first step, the graph was transformed into a vector using a semi-supervised graph neural network (GCN). The second step was based on DBSCAN, an unsupervised clustering method. Various performance evaluations were performed to show the superiority of the proposed method. Experimental results showed that the proposed method could detect abnormal nodes with high accuracy in homogeneous static graphs.