Inferring feature importance with uncertainties with application to large genotype data.

Area of Science:

• Machine Learning
• Statistical Modeling
• Bioinformatics

Background:

• Estimating feature importance is crucial for understanding data-driven models and the underlying data generation process.
• Existing methods like Shapley additive global importance (SAGE) can be computationally intensive.
• There's a need for efficient and uncertainty-aware feature importance estimation.

Purpose of the Study:

• To present a Shapley-value-based framework, Sub-SAGE, for inferring individual feature importance with uncertainty.
• To develop a computationally efficient method for tree-based models, avoiding resampling.
• To demonstrate the framework's applicability on synthetic and large-scale genotype data.

Main Methods:

• Developed Sub-SAGE, a novel feature importance estimator building on SAGE.
• Utilized bootstrapping for estimating uncertainty in the Sub-SAGE estimator across model types.
• Applied the framework to tree ensemble methods and large genotype datasets.

Main Results:

• Sub-SAGE provides efficient feature importance estimation for tree-based models without resampling.
• Bootstrapping effectively estimates uncertainty in Sub-SAGE across various model types.
• Demonstrated successful application in predicting feature importance for obesity using genotype data.

Conclusions:

• Sub-SAGE offers a robust and computationally efficient method for feature importance and uncertainty estimation.
• The framework is valuable for interpreting complex models and understanding biological data.
• This approach enhances the explainability of machine learning models in scientific research.