Interpretable Neural Additive Models Gain Feature Selection and Interactions

Deep neural networks (DNNs) often excel in performance but lack interpretability. Neural Additive Models (NAMs) and Neural Basis Models (NBMs) offer a solution by using neural networks as nonlinear shape functions within generalized additive models (GAMs), providing interpretability by visualizing each feature's contribution. However, these models face computational challenges when considering feature interactions or applied to high-dimensional datasets, as training can become intractable. This paper introduces a novel approach to integrate a feature selection mechanism into NAMs and NBMs. By adding a feature selection layer and updating selection weights during training, the proposed method significantly reduces computational costs and model sizes. This allows for the effective use of two-input neural networks to capture feature interactions even in high-dimensional scenarios, demonstrating improved efficiency and comparable or superior performance to state-of-the-art GAMs.