The goal of this study was to evaluate the performance of traditional gradient boosting (GB) and neural network models on diverse tabular datasets that differ in scale, class balance, and feature composition (numerical, categorical, or mixed). We focused on six representative datasets: adult census income, bank marketing, credit card fraud, breast cancer diagnosis, diabetes, and in-vehicle coupon recommendation, each with distinct challenges related to dimensionality, sample size, and heterogeneity. We benchmark the predictive performance of XGBoost and LightGBM (gradient boosting models) against Multilayer Perceptrons (MLP), Tabular Transformers, and tabular prior-data fitted network (TabPFN), using metrics such as accuracy, F1 score, ROC-AUC, and log loss. To ensure transparency and interpretability, we applied SHapley Additive exPlanations (SHAP) and Local Interpretable Model-Agnostic Explanation (LIME) to all models and evaluated the explanation quality using stability, fidelity, and consistency criteria. Our findings confirm that gradient boosting models consistently achieve the best balance of performance, calibration, and interpretability across heterogeneous and imbalanced datasets. SHAP-based insights show that gradient boosting (GB) models provide more stable and interpretable feature attributions, making them well suited for high-stakes domains such as finance and healthcare. These results emphasize the practical advantages of gradient boosting methods for structured data tasks and highlight the interpretability limitations of deep learning models when applied to tabular datasets. Future work will explore hybrid architectures and pretraining strategies to close this performance gap.
Lazar et al. (Wed,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: