Interpretable Transfer Learning for Small Sample Coal and Gas Outburst Risk Identification Using TabNet

Abstract

The identification of coal and gas outburst risks is crucial for the safe production of coal mines. The application of deep learning techniques in this domain shows significant promise, particularly in small sample scenarios. This paper investigates the use of transfer learning and self-supervised learning strategies in static outburst risk identification models under small sample data scenarios. A TabNet-based model was utilized, focusing on performance improvements achieved through pretraining, particularly with respect to recall rate and false negative rate. The model was pretrained using a combination of self-supervised and supervised learning to enhance adaptability and generalization capabilities for small sample data scenarios, followed by evaluation with stratified fivefold cross-validation. Experimental results demonstrated that the pretrained TabNet model significantly outperformed the non-pretrained model as well as traditional machine learning models, including random forest, XGBoost, LightGBM, SVM, and MLP, in terms of accuracy and stability. Furthermore, removing features with weak correlations to the target variable further improved model performance, emphasizing the importance of integrating various learning strategies during data preprocessing and model training, particularly in limited data contexts. Model interpretability was also analyzed using SHAP and TabNet's inherent interpretability, confirming consistent feature importance rankings and highlighting the model's robustness and reliability.