Proxy-Based Approximation of Shapley and Banzhaf Interactions

2026-05-21 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, quick

Summary

ProxySHAP is a novel method designed to approximate Shapley and Banzhaf interactions in machine learning, addressing the common trade-off between speed and accuracy in existing estimators. This approach combines the high sample efficiency of tree-based proxy models with a principled residual correction for consistency. Theoretically, ProxySHAP introduces a polynomial-time generalization of interventional TreeSHAP, enabling exact interaction index computation for tree ensembles and avoiding the exponential tree-depth dependencies found in prior methods. The system also includes a formal analysis of its residual adjustment strategy, detailing conditions under which Maximum Sample Reuse (MSR) corrects proxy bias without exponentially scaling variance with interaction size. Benchmarking demonstrates ProxySHAP establishes a new benchmark for approximation quality, outperforming ProxySPEX and KernelSHAP-IQ across small- and large-budget scenarios, including applications with thousands of features, and delivering better performance on downstream explainability tasks.

Key takeaway

For Machine Learning Engineers focused on model explainability, ProxySHAP offers a significant advancement for accurately estimating Shapley and Banzhaf interactions. If you are currently struggling with the speed-accuracy trade-off in existing methods like KernelSHAP-IQ or ProxySPEX, consider integrating ProxySHAP. Its superior performance in both small- and large-budget scenarios, even with thousands of features, can enhance the precision and efficiency of your interpretability efforts.

Key insights

ProxySHAP improves Shapley and Banzhaf interaction estimation by combining tree-based proxies with residual correction for speed and accuracy.

Principles

• Tree-based proxies offer high sample efficiency.
• Residual correction ensures principled consistency.
• Polynomial-time computation bypasses exponential dependencies.

Method

ProxySHAP generalizes interventional TreeSHAP to polynomial time for exact interaction indices in tree ensembles, using Maximum Sample Reuse (MSR) for residual bias correction.

In practice

• Apply for large-scale ML with thousands of features.
• Improve explainability task performance.
• Achieve lower error in small- and large-budget regimes.

Topics

• Shapley Interactions
• Banzhaf Interactions
• Machine Learning Explainability
• Tree Ensembles
• ProxySHAP
• Feature Importance

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Data Scientist

Related on AIssential

• Identifying Interactions at Scale for LLMs — SPEX and ProxySPEX efficiently identify influential interactions in ML models by exploiting sparsity and hierarchy.
• TN-SHAP-G: Graph-Structured Tensor Network Surrogates for Shapley Values and Interactions — TN-SHAP-G efficiently computes Shapley values for graph-structured inputs using a graph-aligned tensor network surrogate.
• Metamorphic Testing with the Rashomon Set: Explanation Faithfulness in Machine Learning — Metamorphic testing with the Rashomon Set assesses ML explanation faithfulness without ground truth.
• Explainable AI based ensemble model for the identification of Schizophrenia prodromal phase — An XAI-integrated ensemble ML model accurately predicts Schizophrenia prodromal symptoms for early clinical intervention.
• What Is the Prediction Actually Made Of? — SHAP alone is insufficient for explaining retrieval-based model predictions, which require an "inner explanation" of neighbor influence.
• The Representational Limit of Scalar Interactions: An Interventional Decomposition — Scalar interaction scores inherently conflate uniqueness, redundancy, and synergy, necessitating advanced decomposition for accurate feature attribution.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.