Fast Calculation of Feature Contributions in Boosting Trees

Area of Science:

• Machine Learning
• Explainable AI
• Computational Statistics

Background:

• Fast algorithms exist for Shapley value decomposition in tree models, enabling local feature attribution.
• Global evaluation of feature contributions is needed, but individualizing coefficients of determination (R^2) is difficult due to quadratic losses.

Purpose of the Study:

• To propose Q-SHAP, an efficient algorithm for calculating Shapley values under quadratic losses.
• To improve computational efficiency and accuracy in feature-specific R^2 estimation.

Main Methods:

• Developed Q-SHAP, an algorithm reducing Shapley value computation for quadratic losses to polynomial time.
• Conducted simulations to evaluate Q-SHAP's performance.

Main Results:

• Q-SHAP significantly improves computational efficiency.
• Q-SHAP enhances the accuracy of feature-specific R^2 estimates.

Conclusions:

• Q-SHAP provides an efficient and accurate method for global feature contribution analysis in tree models.
• The algorithm addresses the challenges of individualizing R^2 with quadratic losses.