Tilted Neural Additive Models

Neural additive models (NAMs) have recently attracted increasing attention due to their strong predictive performance and desirable interpretability. However, most existing approaches rely on empirical risk minimization (ERM), which often underperform in challenging scenarios where average performance is not a suitable criterion for the problems of interest, such as complex noise perturbation. In this article, we address these challenges by learning sparse NAMs under the tilted ERM (TERM) framework. The proposed approach retains feature-level interpretability while exhibiting robustness to a broad class of complex noise perturbations. On the theoretical side, we develop oracle-type function approximation analyses for the proposed method under both independent identically distributed (i.i.d.) and stationary $\beta $ -mixing settings, yielding explicit error bounds that characterize how noise structure, model complexity, and the tilting mechanism of TERM jointly affect robustness performance. Empirical evaluations on synthetic and real-world coronal mass ejections (CMEs) datasets demonstrate the effectiveness and competitiveness of our approach.