Decomposition of Shapley Additive Explanations for a Stacking Model

Ensemble learning methods such as stacking models have shown significant improvements in predictive performance by combining multiple heterogeneous base learners. However, the complexity of the layered architecture and the diversity of base learners in stacking models often come at the cost of interpretability. Shapley Additive exPlanations (SHAP) provide an axiomatic framework, derived from cooperative game theory, to quantify feature contributions. This paper introduces the concept of partial Shapley values for stacking models and proposes a method to decompose the Shapley values of stacking models into a sum of partial Shapley values propagated through base learners. We illustrate the approach using a stacked generalization example on a pharmacogenomics dataset and visualize both local and global partial Shapley values.