Prediction via Shapley Value Regression

2025-05-07

AI Summary

To address the high computational overhead of post-hoc Shapley value estimation in black-box model explanation, this paper proposes ViaSHAP: an end-to-end learning paradigm that directly models the Shapley value function. Its core innovation unifies prediction and explanation—each feature’s Shapley value is predicted explicitly by the model, and the final prediction is obtained as their sum, thereby eliminating inference-time explanation overhead. Built upon the Kolmogorov–Arnold representation theorem, ViaSHAP employs a purpose-designed neural architecture that jointly ensures universal approximation capability, predictive accuracy, and explanation fidelity. Experiments across diverse tabular datasets demonstrate state-of-the-art (SOTA) predictive performance; moreover, ViaSHAP yields significantly more accurate Shapley value estimates than FastSHAP. Its generalizability is further validated on both tabular and image classification tasks.

Technology Category

• Computer Vision: Large Vision Models
• Machine Learning: Feature Construction/Reformulation
• Humans and AI: Learning Human Values and Preferences

Application Category

• User Modeling, Personalization and Recommendation: Explainable and interpretable methods for personalization
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Shapley values have several desirable, theoretically well-supported, properties for explaining black-box model predictions. Traditionally, Shapley values are computed post-hoc, leading to additional computational cost at inference time. To overcome this, a novel method, called ViaSHAP, is proposed, that learns a function to compute Shapley values, from which the predictions can be derived directly by summation. Two approaches to implement the proposed method are explored; one based on the universal approximation theorem and the other on the Kolmogorov-Arnold representation theorem. Results from a large-scale empirical investigation are presented, showing that ViaSHAP using Kolmogorov-Arnold Networks performs on par with state-of-the-art algorithms for tabular data. It is also shown that the explanations of ViaSHAP are significantly more accurate than the popular approximator FastSHAP on both tabular data and images.

Problem

• Overcome computational cost of post-hoc Shapley value computation
• Propose ViaSHAP to directly derive predictions from Shapley values
• Improve explanation accuracy compared to FastSHAP on tabular and image data

Innovation

• Learns function to compute Shapley values directly
• Uses universal approximation and Kolmogorov-Arnold theorems
• Performs on par with state-of-the-art tabular algorithms