MedGraph: malicious edge detection in temporal reciprocal graph via multi-head attention-based GNN

With the popularity of various online dating applications, it has become a crucial task to detect anomalous or malicious users from a large number of reciprocal users. Essentially, this task could be converted to a malicious edge detection problem, which is an important yet challenging task due to the following difficulties. First, malicious users may fake their profiles to avoid being detected by the service platform. Second, malicious behaviors, i.e., malicious edges, might vary from time to time which greatly challenges most existing approaches. To address the aforementioned issues, this paper proposes the multi-head attention-based GNN approach to detect malicious edges from a temporal reciprocal graph, called MedGraph. Particularly, the proposed MedGraph approach employs a transformer component to first capture both long-term and short-term behavior characteristics of malicious users from their historical interaction data. Then, a co-attention component is designed to differentiate important features that account for the prediction of malicious edges. We evaluate the proposed approach on two public datasets and one real-world dataset collected by ourselves. The promising results demonstrate that our approach could achieve state-of-the-art performance against a number of both baseline and SOTA approaches with respect to the widely adopted evaluation criteria.