Reproduce “Interpretable Model-Aware Counterfactual Explanations for Random Forest” on OpenResearch

About this paper

Joshua S. Harvey, Guanchao Feng, Sai Anusha Meesala, Tina Zhao, Dhagash Mehta

Researchers from BlackRock and Prospect 33 developed an interpretable, model-aware framework for generating counterfactual explanations for Random Forest models, using RF-GAP proximities and an iterative trajectory search. This approach yielded explanations that were significantly sparser and more effective for actionable recourse compared to SHAP and LIME, particularly for financial credit assessment tasks.

Key insights

• Leveraging the Random Forest's internal geometry through RF-GAP proximities enables the generation of counterfactuals that are inherently faithful to the model's decision process and grounded in realistic data examples.
• Constructing counterfactuals via iterative, local steps along data-driven trajectories ensures plausibility and interpretability, mitigating the issue of generating unrealistic or out-of-distribution instances.
• The "partition tally" mechanism provides significantly sparser and more directional explanations compared to model-agnostic methods, highlighting only the most critical features and their required changes for a prediction flip.
• Explanations derived from model-aware counterfactuals are empirically more effective for algorithmic recourse, offering superior guidance for achieving desired outcome changes in a model's prediction.

Summary by alphaXiv.