Abstract: To remedy the class imbalance and the lack of necessary connections between fraud nodes in financial fraud detection, we propose a financial fraud detection method combined with graph contrastive learning (FFD-GCL). According to the scarcity of labels, the pseudo-labels of unlabeled nodes were generated by improving the label propagation algorithm. On this basis, nodes and edges on the original subgraph of a target node were picked with a label-consistency sampler to construct the purification subgraph. A graph attention network that combined time and position encoding was designed to encode target node on the purification and original subgraph, and contrastive learning was applied to revise the node representation on the original subgraph which was used to identify whether the node was fraudulent. Experimental results on two financial anti-fraud datasets demonstrate that FFD-GCL outperforms state-of-the-art baselines.