Effective Pruning for Top-k Feature Search on the Basis of SHAP Values

Area of Science:

• Machine Learning
• Explainable AI (XAI)

Background:

• Machine learning models are increasingly influential, necessitating methods to explain their predictions.
• The SHAP framework quantifies feature contributions to model predictions.
• Existing SHAP methods can be computationally expensive, especially for model-agnostic approaches.

Purpose of the Study:

• To propose TopShap, a novel model-agnostic algorithm for efficient approximation of top-k SHAP values.
• To reduce the computational cost associated with calculating SHAP values for important features.

Main Methods:

• TopShap approximates SHAP values for the top-k most important features.
• It utilizes confidence interval bounds to dynamically identify and remove features that cannot be part of the top-k set.
• This pruning strategy optimizes computational resource allocation.

Main Results:

• TopShap demonstrates efficient pruning of the feature search space.
• Substantial reductions in execution time were observed compared to Kernel SHAP, a leading model-agnostic method.
• Evaluations across diverse datasets and models confirm TopShap's efficiency and model-agnosticism.

Conclusions:

• TopShap offers a computationally efficient solution for approximating top-k SHAP values in a model-agnostic manner.
• The algorithm's ability to reduce execution time makes it suitable for large-scale machine learning interpretability tasks.