The most important features in generalized additive models might be groups of features

2026-03-21 · Source: Machine learning : nature.com subject feeds · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Advanced, long

Summary

A new method has been developed to determine the importance of feature groups in Generalized Additive Models (GAMs), addressing the common oversight of joint signals from related features in interpretable machine learning. This approach is efficient, does not require model retraining, allows for post-hoc and overlapping group definitions, and remains effective in high-dimensional settings. The method's properties were demonstrated through three synthetic experiments. Furthermore, its practical utility was showcased in two medical case studies: identifying depressive symptoms from a multimodal neuroscience dataset (NCANDA, NIH funding AA021681, AA021690, AA021691, AA021692, AA021695, AA021696, AA021697) and analyzing social determinants of health after total hip arthroplasty (PHC4 data, accessed June 24, 2025). These applications revealed that group importance analysis provides a more accurate and holistic understanding of complex medical issues than single-feature analysis.

Key takeaway

For research scientists developing interpretable machine learning models, you should integrate group feature importance analysis, especially when dealing with multimodal or naturally grouped datasets. This approach provides a more comprehensive understanding of underlying factors, moving beyond isolated feature effects to reveal critical joint signals, which can lead to more robust and clinically relevant insights in fields like neuroscience and public health.

Key insights

Analyzing feature groups, not just individual features, offers critical insights in interpretable machine learning.

Principles

• Joint feature signals are often overlooked.
• Group importance is meaningful in high dimensions.
• Overlapping groups can be defined post-hoc.

Method

The proposed method for GAMs efficiently determines group importance without retraining, supports post-hoc and overlapping group definitions, and is effective in high-dimensional contexts.

In practice

• Apply group importance to multimodal datasets.
• Use for medical diagnosis with complex factors.
• Analyze social determinants in healthcare outcomes.

Topics

• Generalized Additive Models
• Feature Group Importance
• Interpretable Machine Learning
• Multimodal Data Analysis
• Healthcare Applications

Best for: Research Scientist, AI Researcher, AI Scientist, Data Scientist

Related on AIssential

• Aggregate Models, Not Explanations: Improving Feature Importance Estimation — Ensembling models directly improves feature importance estimation by reducing excess risk, especially for complex ML models.
• “Features” aren’t always the true computational primitives of a model, but that might be fine anyways — Features in mechanistic interpretability exist on a spectrum from memorization to computational primitives.
• Identifying Interactions at Scale for LLMs — SPEX and ProxySPEX efficiently identify influential interactions in ML models by exploiting sparsity and hierarchy.
• Conditional Local Importance by Quantile Expectations — CLIQUE provides model-agnostic local variable importance, accurately capturing conditional relationships and reducing bias.
• Learning and Naming Subgroups with Exceptional Survival Characteristics — Sysurv identifies exceptional survival subgroups using differentiable, non-parametric methods and interpretable rules.
• Alike Parts: A Feature-Informed Approach to Local and Global Prototype Explanations — Integrating feature importance into prototype selection enhances both local explanations and global prototype diversity without compromising model fidelity.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine learning : nature.com subject feeds.