Adaptive Graph Attention Networks with Incremental Learning for Robust Financial Fraud Detection in Dynamic Transaction Networks

Abstract

The problem of financial fraud detection in dynamic transaction networks is a daunting task due to the dynamism of fraudulent activities in networks and the high level of class imbalance of real-world data. We propose an adaptive graph attention network framework to model the financial ecosystem as a heterogeneous graph with nodes representing different entities such as customers, accounts, and merchants and directed financial interactions among them with rich attribute information. We use a central graph attention network as our approach, and in the process of passing messages, it allocates varying importance to the neighboring nodes of a graph, enabling us to selectively attend to suspicious subgraphs and ignore irrelevant links. One of the main innovations of our method is the adaptive learning module that constantly analyses the statistical characteristics of the data flow (such as the confidence of prediction or the change in the distribution of transactions) to detect the concept drift. When significant discrepancy with the training distribution is detected, the system starts an incremental update process, which optimizes model parameters without necessarily re-training the entire network with all the data once more. The use of this mechanism will prevent the learned embeddings from becoming stale, and will not be prohibitively expensive to compute.