OptSHAP: Explaining Dimensionality Reduction-based Models for Tabular Data via Optimization

ICLR 2026 Conference Submission18573 Authors

Keywords: Explainable Artificial Intelligence, Dimensionality Reduction, Tabular Data, Optimization

TL;DR: A novel attribution explanation method for Dimensionality Reduction-based Models

Abstract: Dimensionality Reduction-based Models (DRbMs), which couple a dimensionality reduction technique with a predictive model, are commonly used to mitigate overfitting and reduce computational complexity regarding high-dimensional tabular data. However, their two-stage architecture presents considerable challenges for explainability, as the projection obscures the original feature space, thus making the model output difficult to interpret in terms of the input features. Model-agnostic explanation methods are applicable to DRbMs but typically rely on sampling-based approximations, leading to instability and low-faithfulness explanations. To address these limitations, we introduce OptSHAP, the first optimization-based attribution specifically designed for DRbMs. Our method leverages reduced-space attributions and then redistributes them back to the original feature space through a transformation that satisfies the principle of efficiency. Additionally, we propose a novel evaluation metric, the $k$-Local Stability Score (LSS), which quantifies the stability of feature attribution methods by averaging their distances to local explanations. Extensive empirical evaluations across high-dimensional datasets, various dimensionality reduction techniques, and multiple machine learning models demonstrate that OptSHAP outperforms state-of-the-art attribution methods, achieving up to $24\times$ stability and $2\times$ fidelity on key benchmarks.

Primary Area: interpretability and explainable AI

Submission Number: 18573