Individual variable priority: a model-independent local gradient method for variable importance

Area of Science:

• Statistics
• Machine Learning
• Bioinformatics

Background:

• Traditional variable importance metrics often fail to capture individual-level variations.
• Existing methods for addressing heterogeneity can be model-dependent and introduce bias.

Purpose of the Study:

• To introduce individual variable priority (iVarPro), an extension of the Variable Priority (VarPro) framework.
• To provide a more precise and interpretable measure of variable importance that accounts for individual heterogeneity.

Main Methods:

• iVarPro utilizes rule-based, data-driven partitioning to estimate the gradient of the conditional mean function.
• The method focuses on gradients to assess the impact of small variable perturbations on individual outcomes.

Main Results:

• Simulations and analysis of a real-world survival dataset demonstrate iVarPro's advantages.
• iVarPro more accurately captures true functional relationships by effectively utilizing local sample information.

Conclusions:

• iVarPro offers a superior approach to variable importance assessment compared to traditional methods.
• The framework provides enhanced interpretability and precision for understanding individual-level feature effects.