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

Area of Science:

• Machine Learning
• Statistics
• Computational Neuroscience

Background:

• Interpretable machine learning often focuses on individual feature importance.
• The collective impact of related features, or feature groups, is frequently overlooked.
• This is particularly relevant for multimodal datasets where natural feature groupings exist.

Purpose of the Study:

• To develop a novel, efficient method for assessing the importance of feature groups in Generalized Additive Models (GAMs).
• To provide a more holistic view of predictor importance beyond individual features.

Main Methods:

• Introduced a new approach to calculate group feature importance for GAMs.
• The method is efficient, requires no model retraining, and allows for posthoc and overlapping group definitions.
• Validated on synthetic datasets and applied to real-world multimodal neuroscience and healthcare data.

Main Results:

• Demonstrated the method's effectiveness across various data regimes using synthetic experiments.
• Identified important feature groups for predicting depressive symptoms from multimodal neuroscience data.
• Highlighted the significance of social determinants of health in post-total hip arthroplasty outcomes.

Conclusions:

• Analyzing feature group importance offers a more accurate and holistic perspective than single-feature analysis.
• This approach is crucial for understanding complex medical conditions and health outcomes.
• The proposed method is efficient and versatile for high-dimensional data.