Home // International Journal On Advances in Software, volume 18, numbers 3 and 4, 2025 // View article
Authors:
Yukari Shirota
Tamaki Sakura
Keywords: Explainable AI (XAI); SHAP values; Shapley theory; Singular Value Decomposition (SVD); Dimensionality reduction; Regression analysis; Latent semantic structure; Stock price analysis; Feature attributions.
Abstract:
In this paper, we propose a novel explainable AI (XAI) method for regression analysis, named SHAP_SVD, which integrates SHAP (SHapley Additive exPlanations) values with Singular Value Decomposition (SVD) to uncover latent structures in model interpretations. The Shapley value, a concept initially introduced by Lloyd Shapley in the field of cooperative game theory, has recently gained substantial attention in the AI community, particularly through its adaptation into the SHAP framework by Scott Lundberg. SHAP values enable us to interpret the contribution of each explanatory variable to a specific prediction by treating the prediction process as a game in which variables are players, and their marginal contributions are evaluated across all possible coalitions. In regression analysis, SHAP values can be seen as a matrix of attributions: for each observation, the contribution of each feature is calculated relative to a baseline. Our SHAP_SVD method applies SVD to this SHAP value matrix, thereby reducing dimensionality while preserving key information. The eigenvalues and corresponding eigenvectors obtained from SVD allow us to identify "concepts" or "latent semantic structures" that govern the interaction between features and the target variable. These concepts are encoded in both the left singular vectors. As a case study, we conducted a regression analysis of stock price growth rates for leading Indian and Japanese automobile manufacturers. The SHAP values were computed using a tree-based ensemble regression model, and our SHAP_SVD method was applied to reveal underlying structures. Two principal components emerged from the decomposition. In the extended analysis, we present a more detailed examination of the SHAP value distribution and structure. Specifically, we analyze how the SHAP captures nonlinear interactions that are invisible in traditional raw data correlation matrices. By comparing the performance of models evaluated using raw input features with those interpreted through SHAP, we demonstrate that SHAP-based interpretation yields greater stability, clarity, and interpretability, particularly in cases where multicollinearity or redundant variables obscure the true contribution of each feature. Our results demonstrate that SHAP analysis not only enhances the transparency of complex models but also, when combined with SVD, offers a powerful tool for discovering and visualizing conceptual dimensions underlying the data. The SHAP_SVD approach thus serves both as a diagnostic tool for regression models and a framework for semantic exploration in high-dimensional datasets.
Pages: 181 to 190
Copyright: Copyright (c) to authors, 2025. Used with permission.
Publication date: December 30, 2025
Published in: journal
ISSN: 1942-2628