Neural Additive and Basis Models with Feature Selection and Interactions

What Happened

A new preprint (arXiv:2606.19850v1) introduces enhancements to Neural Additive Models (NAMs) and Neural Basis Models (NBMs), two architectures designed to improve interpretability in deep learning. The core innovation involves integrating feature selection mechanisms and explicit interaction terms directly into these models. Traditional NAMs decompose predictions into additive shape functions for each feature, making them inherently interpretable but unable to capture feature interactions. The proposed extensions address this limitation by allowing the model to learn which features interact and how, while maintaining the additive structure that enables human understanding. The work also incorporates basis function expansions to model complex non-linear relationships more efficiently than standard neural network layers.

Why It Matters

This research tackles a fundamental tension in modern AI: the trade-off between performance and interpretability. Deep neural networks achieve state-of-the-art results but operate as black boxes, which is problematic in regulated industries like healthcare, finance, and legal analytics. NAMs and NBMs already offered a middle ground—high accuracy with interpretable additive components—but their inability to model interactions limited their applicability to real-world data where features often combine non-additively. By adding controlled interaction detection and feature selection, this work bridges the gap between fully transparent linear models and opaque deep networks. The feature selection aspect is particularly valuable: it reduces model complexity and helps practitioners identify which variables truly drive predictions, a critical need for debugging and trust.

Implications for AI Practitioners

For data scientists and ML engineers, this development has several practical consequences:

First, practitioners can now deploy interpretable models in high-stakes settings without sacrificing the ability to capture complex relationships. For example, in credit scoring, a bank could use this enhanced NAM to understand not just how income affects risk, but how income interacts with employment history—all while maintaining a transparent decision process. Second, the built-in feature selection reduces the need for separate preprocessing pipelines. This simplifies model maintenance and deployment, as the model itself learns which features are relevant, potentially lowering computational costs and improving generalization. Third, debugging and model auditing become more straightforward. When a model makes an unexpected prediction, practitioners can inspect the additive contributions and interaction terms directly, rather than resorting to post-hoc explanation methods like SHAP or LIME, which can be unreliable. Fourth, this approach may be especially useful for tabular data problems, where interpretability is often prioritized over the marginal gains from deep learning. It offers a viable alternative to gradient-boosted trees, which are currently dominant in tabular domains but lack the same level of built-in interpretability.

Key Takeaways

• The enhanced NAM/NBM framework adds interaction detection and feature selection to inherently interpretable additive models, closing a key gap in explainable AI.
• This development is most impactful for regulated industries where both accuracy and transparency are mandatory, such as healthcare, finance, and insurance.
• Practitioners gain a practical tool that reduces reliance on post-hoc explanation methods, enabling direct inspection of feature contributions and interactions.
• The approach is particularly well-suited for tabular data, offering a interpretable alternative to gradient-boosted trees and black-box neural networks.