Interpretability TabNet: Enhancing Tabular Deep Learning with Entmax Stability Masking

Deep models for tabular data are useful because they can learn task-specific representations, but they are often compared against very strong tree-based alternatives. In settings where model outputs influence decisions, it is also important that an explanation is not an artifact of a particular random initialization. This work examines that issue for TabNet by changing the sparse normalization used in the attentive masks from Sparsemax to Entmax with α=1.5. Rather than treating this replacement as a new architecture, we evaluate how it changes the repeatability of TabNet’s intrinsic explanations under repeated training. On Adult Income, the Entmax-based configuration gives slightly higher accuracy and AUC and a lower average variation in feature importance. We further evaluate Breast Cancer, Digits, Wine, and Iris using two explanation-stability summaries, five random seeds, predictive baselines, paired tests, and an α ablation. The results are not uniformly favorable to Entmax; instead, they reveal dataset-dependent trade-offs among prediction, sparsity, and explanation consistency. The study highlights the value of reporting explanation stability across seeds for intrinsically interpretable tabular models.