Graph neural network applied to predict corporate financial distress

Two financial distress prediction frameworks integrating information from the time and spatial domains are proposed based on graph neural network (GNN)． A class imbalance handling module is designed after fusion of financial data with patent information to enhance prediction accuracy． To evaluate this model, effectiveness of class imbalance handling module, synergistic values for incremental information of patent data, technological innovation information, and graph structure modeling are each verified． It is found that multi-graph convolutional recurrent network with the “spatial-first, temporal-later” information aggregation strategy delivers superior performance． The present work provides new strategies for financial distress prediction．