Interpretable Graph Neural Networks for Heterogeneous Tabular Data

The chapter introduces the challenge of applying interpretable machine learning models to sensitive domains like medicine and finance, where trustworthiness is crucial. It highlights the limitations of black-box models and post-hoc explanation methods, emphasizing the need for inherently interpretable models. The proposed approach, IGNH, leverages graph neural networks to capture interactions within heterogeneous tabular data, seamlessly handling both categorical and numerical features. The method provides exact feature attributions alongside predictions, ensuring transparency and trustworthiness. The chapter includes a large-scale empirical investigation demonstrating that IGNH's explanations align with true Shapley values and that its predictive performance is comparable to state-of-the-art models like XGBoost. The study concludes with directions for future work, including exploring alternative interaction modeling techniques and extending the approach to non-tabular datasets.