From Abstract to Actionable: Pairwise Shapley Values for Explainable AI

TL;DR

Pairwise Shapley Values address key interpretability and scalability issues in Shapley-based explanations by grounding attributions in explicit comparisons between a target instance and a proximal reference. The method combines a flexible pair-selection strategy with single-value imputation and model-agnostic estimation (KernelSHAP) to produce intuitive, locally faithful explanations while reducing computational cost. Through extensive evaluation on real estate, polymer, and drug datasets, the approach demonstrates improved interpretability via normalized and directionally consistent attributions, robust locality properties like dummy pairs, and faster runtimes than traditional distribution-based methods. The work suggests a practical path toward more transparent AI in high-stakes domains and outlines extensions to other data modalities and extensions beyond tabular data.

Abstract

Explainable AI (XAI) is critical for ensuring transparency, accountability, and trust in machine learning systems as black-box models are increasingly deployed within high-stakes domains. Among XAI methods, Shapley values are widely used for their fairness and consistency axioms. However, prevalent Shapley value approximation methods commonly rely on abstract baselines or computationally intensive calculations, which can limit their interpretability and scalability. To address such challenges, we propose Pairwise Shapley Values, a novel framework that grounds feature attributions in explicit, human-relatable comparisons between pairs of data instances proximal in feature space. Our method introduces pairwise reference selection combined with single-value imputation to deliver intuitive, model-agnostic explanations while significantly reducing computational overhead. Here, we demonstrate that Pairwise Shapley Values enhance interpretability across diverse regression and classification scenarios--including real estate pricing, polymer property prediction, and drug discovery datasets. We conclude that the proposed methods enable more transparent AI systems and advance the real-world applicability of XAI.

Figures (14)

• Figure 1: Shapley value attribution comparison between traditional implicit methods using empirical feature distributions and the proposed explicit pair selection method.
• Figure 2: Feature importance scores (mean absolute Shapley values) for features: (a) grade, (b) city_BELLEVUE, and (c) sqft, derived from the base model on the test data of the Home dataset. The methods include single-value imputation using zeros (B0) and mean values (BM), uniform distribution imputation (UF), marginal distribution imputation using all training data (MA) or a K-means summary of the training data (MK), conditional distribution imputation using all training data (CA), and a model-specific method, TreeShap (TS).
• Figure 3: Waterfall plot of feature attributions for an explicand from the Home dataset using different methods: (a) MA, (b) CA, and (c) PC. The y-axis lists the top nine features with their values for the explicand (relative values in (c)); remaining features are aggregated as "Other Features." The x-axis shows feature attributions in dollars. $f(x)$ is the model's predicted value; $E[f(X)]$ is the expected prediction based on the background data.
• Figure 4: Distribution of normalized Shapley values for feature sqft in the Home dataset for different methods. (a) Distribution plot for seven existing non-pairwise methods; (b) Distribution plot for three selected existing non-pairwise methods (UF, MA, TS) and three pairwise methods with different similarity algorithms (PR, PC, PS).
• Figure 5: Monotonicity measure (matched percentage) of Shapley value estimates across different methods for three features: sqft, grade, and noise_traffic.