Neural Additive and Basis Models with Feature Selection and Interactions

Researchers propose enhanced neural additive and basis models (NAM/NBM) that incorporate feature selection mechanisms to improve computational efficiency and interpretability of deep neural networks. The advancement enables these models to handle high-dimensional datasets and capture feature interactions while reducing training costs and model sizes compared to traditional approaches.

This research addresses a fundamental challenge in machine learning: balancing model interpretability with computational efficiency. Neural additive models have gained traction because they decompose predictions into individual feature contributions, making them suitable for regulated industries and applications requiring transparency. However, their scalability limitations have restricted their practical deployment in complex, high-dimensional environments. The proposed enhancement introduces dynamic feature selection during training, automatically identifying relevant features and eliminating computational overhead from irrelevant dimensions. This approach is particularly significant because it enables two-input neural networks to capture feature interactions—a capability previously impractical for large datasets. The feature selection layer operates as a lightweight filtering mechanism, reducing dimensionality without sacrificing model performance. The research demonstrates computational efficiency gains while maintaining or exceeding state-of-the-art performance metrics. For practitioners, this development opens pathways to deploy interpretable models in production environments where both accuracy and explainability matter, such as financial services, healthcare, and regulatory-heavy sectors. The method's simplicity means adoption barriers remain low for existing NAM/NBM practitioners. Looking forward, the research suggests feature selection mechanisms could become standard components in interpretable machine learning architectures, potentially spurring broader adoption of additive models in industry applications where black-box deep learning currently dominates but explainability is increasingly demanded.

• →Feature selection mechanisms reduce computational costs and model sizes for neural additive models by automatically filtering irrelevant features during training.
• →The enhanced models now support two-input neural networks for capturing feature interactions even in high-dimensional datasets, previously computationally intractable.
• →Proposed NAM and NBM variants achieve comparable or superior performance to state-of-the-art generalized additive models with improved efficiency.
• →The approach maintains interpretability advantages of additive models while addressing scalability bottlenecks that limited practical deployment.
• →Simple implementation suggests broad adoption potential for practitioners requiring both accuracy and explainability in production environments.