A general framework for inference on algorithm-agnostic variable importance.

Area of Science:

• Statistical modeling
• Machine learning
• Bioinformatics

Background:

• Assessing feature importance is crucial for prediction tasks.
• Current methods often depend on specific algorithms, potentially misrepresenting intrinsic feature value.
• There's a need for algorithm-agnostic variable importance measures.

Purpose of the Study:

• To propose a general framework for nonparametric, interpretable, and algorithm-agnostic variable importance inference.
• To define variable importance based on population-level prediction contrasts.
• To develop methods for valid confidence intervals and hypothesis testing.

Main Methods:

• Developed a general framework for nonparametric inference on variable importance.
• Defined variable importance as a contrast between oracle predictiveness with and without specific features.
• Proposed an efficient nonparametric estimation procedure for confidence intervals.
• Outlined a strategy for testing the null importance hypothesis.

Main Results:

• The proposed framework provides interpretable and algorithm-agnostic variable importance.
• The estimation procedure allows for valid confidence intervals, even with machine learning.
• Simulations demonstrate good operating characteristics of the proposed method.
• The approach was illustrated using data from an HIV-1 antibody study.

Conclusions:

• The proposed framework offers a robust method for assessing intrinsic feature value.
• This approach overcomes limitations of algorithm-dependent importance measures.
• The method facilitates reliable variable importance inference in diverse applications, including biomedical research.