SE-HG-GNN: a novel explainable deep learning approach for financial distress prediction using optimized graph neural networks

Financial Distress (FD) prediction is a critical task of study for decision-makers. The accurate prediction of financial distress supports a company and its investors in avoiding major losses. The current FD prediction models have failed to achieve higher accuracy in prediction for various factors. These models are based on static feature selection, class imbalance, poor hyperparameter tuning, and a lack of explainability. To solve these issues, in this work, a novel deep learning model called Squeeze-and-Excitation Heterogeneous Graph Neural Network (SE-HG-GNN) is proposed for robust classification in imbalanced datasets. The core of the model uses Graph Neural Networks and Gated Recurrent Units (GRUs) for temporal encoding and message aggregation. The final representation is derived via an attention mechanism that dynamically weighs the importance of the self versus the neighbourhood context. To maximise predictive power, the hyperparameters of the model are tuned using Modified Spider Wasp Optimisation (MSWO). In addition, the training is stabilised against imbalance using the Focal Loss function. Experimental results and SHapley Additive exPlanations (SHAP) analysis show that the MSWO-tuned SE-HG-GNN achieves superior accuracy and provides enhanced interpretability compared to default configurations.