Knowledge-Guided Explanations for Tree Ensemble Models via Levels-Structure Values

Abstract Tree ensemble models, such as gradient boosting machines and random forests, have achieved strong predictive performance on tabular data by capturing complex nonlinear relationships and feature interactions. However, their flexibility often comes at the cost of interpretability, especially in scientific applications where features are organized by domain knowledge into a priori multi-level structures. Existing game-theoretic explanation methods, including TreeSHAP and TreeOwen, provide principled attribution frameworks based on the Shapley and Owen values; yet, they address either individual features or a single grouping level. In this paper, I introduce Tree Levels-Structure Value (TLSV), a knowledge-guided, principled, and efficient framework for computing the levels-structure value—which generalizes the Owen value from a single a priori partition to a chain of nested partitions—for tree ensemble models. TLSV therefore accommodates multi-level coalition structures, such as features within groups and groups within higher-level supergroups, while preserving coherent attributions across all resolutions, with single-level Owen-value attribution recovered as a special case. Algorithmically, TLSV combines hierarchy-aware recursion with tree-aware dynamic programming to avoid exhaustive subset enumeration within terminal feature groups. Its exact algorithm returns levels-structure values for the tree-aware cooperative game induced by the fitted ensemble. For high-branching hierarchies, I also develop a recursive Monte Carlo estimator that provides scalable approximation with unbiasedness and consistency guarantees. Through simulation studies, I assess attribution accuracy, finite-budget Monte Carlo behavior, numerical coherence, and computational scalability. Finally, I illustrate TLSV using microbiome data from an immunotherapy study, showing how taxonomically nested microbial features contribute coherently to predictions of treatment response.