Abstract: Graph contrastive learning, as an effective pre-training strategy, can address the issue of scarce high-quality labeled data in graph-based fraud detection methods. However, current approaches in this category face challenges where malicious behavioral features are either weakened within the aggregation mechanism of graph neural networks or compromised during the data augmentation process. To tackle this, this paper proposes an optimized graph contrastive learning method that integrates graph reconstruction and dynamic data augmentation techniques, aiming to enhance the effectiveness of fraud detection. This method reduces conflicts arising from neighbor feature aggregation by adjusting edge weights in the graph, thereby improving detection accuracy. Simultaneously, it dynam- ically adjusts the data augmentation process using label invariance and distribution diversity metrics to ensure that the augmented data retains critical fraud features while possessing necessary diversity. Experimental results on multiple graph fraud detection datasets demonstrate the effectiveness of this method,  with detection performance improvements ranging from 2% to 5% compared to state-of-the-art methods.

Key words: fraud detection, graph contrastive learning, graph data augmentation, graph reconstruction