ParGCN: Abnormal Transaction Detection based on Graph Neural Networks

This paper improves GraphSAGE from two aspects: 1) performing a sampling compensation before the training to avoid the possible information losses due to the sampling; and 2) adding a hopping connection with the initial inputs in the aggregating phase to avert the potential loss of the initial features of nodes. The empirical study shows that FastGCN can obtain a relatively higher recall of detection but with a lower precision due to its randomness of Monte-Carlo methods and ignoring the special impacts of neighbors; while the improved GraphSAGE gets a relatively higher precision of detection but with a lower recall due to only focusing on neighbors. This paper proposes a graph-based approach to improve both precision and recall of the abnormal transaction detection by hybridizing the improved GraphSAGE with FastGCN, called ParGCN (Precision and recall), describes the mathematical formulas of the hybrid model, and analyzes the time complexity. A set of experiments on the two data-sets with significant differences of the numbers of features are performed to compare and evaluate the proposed approach to demonstrate the validity in terms of the precision and recall.