Neural partially linear additive model

Abstract

Interpretability has drawn increasing attention in machine learning. Most works focus on post-hoc explanations rather than building a self-explaining model. So, we propose a Neural Partially Linear Additive Model (NPLAM), which automatically distinguishes insignificant, linear, and nonlinear features in neural networks. On the one hand, neural network construction fits data better than spline function under the same parameter amount; on the other hand, learnable gate design and sparsity regular-term maintain the ability of feature selection and structure discovery. We theoretically establish the generalization error bounds of the proposed method with Rademacher complexity. Experiments based on both simulations and real-world datasets verify its good performance and interpretability.