[Objective] This paper provides graph convolutional neural networks with spatial and neighborhood adaptability for credit fraud detection.

[Methods] This paper presents Hyperbolic Jumping Connection Graph Convolutional Neural Networks. In the aspect of spatial adaptive, the node attribute is expressed as the trainable curvature of hyperbolic space, so as to complete the low-distortion embedding representation of the fraud network. In the aspect of neighborhood adaptive, the Hyperbolic Jumping Knowledge Networks framework was defined, and the neighborhood representation results were fused by the hyperbolic interlayer aggregation mechanism. Thus, the graph representation results of fusion space and neighborhood adaptability are provided for the relational network to complete the task of credit fraud detection.

[Results] By deploying the experiment in a large social network that is public and derived from actual business scenarios, the model achieved an 83.5% effect on the index AUC, which was 7.65% higher than the baseline model represented by GraphSAGE(NS).

[Limitations] The shallow networks have some restrictions on the advantage of neighborhood adaptability, thus our model has a more obvious advantage on large complex deep network structures.

[Conclusions] Spatial adaptive provides a more accurate description of node attribute correlation, and neighborhood adaptive selects the optimal neighborhood aggregation range for graph representation learning. The spatial and neighborhood adaptive model has a better identification effect in a large fraud network.