GIN-LDP: A Graph Isomorphism Network-based Model for Loan Default Prediction

Loan default prediction is a key task in financial risk management, which directly affects the risk and return of financial institutions. To enhance the prediction accuracy, this study proposes a loan default prediction model GIN-LDP based on graph isomorphism network(GIN). First, based on the multi-source data, the lender’s historical credit behavior features, recent credit trend features and personal wealth soft information features are extracted respectively. Then, based on the above three kinds of extracted features, the graph structure is constructed with the lender as the node of each graph. In the multi-view graph structure, the random walk positional encoding(RWPE) is introduced to capture the global dependency features of the nodes, which are fused with the node features to perform a comprehensive characterization and through the GIN module, which aggregates adjacent and higher-order neighbor features. Finally, a fully connected layer is used for node classification prediction to discriminate whether the lender’s current loan application is at risk of default. Experiments conducted on home credit default risk dataset and lending club dataset validate the effectiveness of the GIN-LDP model.