Leaf Values as Coordinates: Exact Contrastive Explanation for Gradient-Boosted Ensembles

Recent research introduces a novel method for exact contrastive explanation for gradient-boosted ensembles by reinterpreting leaf values as coordinates. This approach allows for precise identification of differences between model predictions by treating each instance as a point in a multi-dimensional space where the model acts linearly. This method facilitates direct traceability of predictive disparities to specific decision splits within the ensemble without requiring further fitting or assumptions of feature additivity.

Why it matters

This development enhances the explainability of complex machine learning models, which is crucial for building trust, meeting regulatory requirements, and facilitating model auditing. By providing exact and traceable explanations, organizations can better understand model decisions, diagnose issues, and ensure ethical and fair application of AI systems, particularly in sensitive domains.

Key insights

• Gradient-boosted ensemble leaf values can be viewed as coordinates, transforming model predictions into a linear operation on these coordinates.
• This perspective enables exact contrastive explanation, identifying specific coordinates (leaf values) responsible for differences in predictions between instances.
• The method pinpoints differences to actual splits within the decision trees of the ensemble.
• It avoids reliance on fitting, sampling, or assuming feature additivity, as additivity is inherent in the coordinate space.
• A recourse method has been developed and evaluated based on this new representation.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2608.19127