Home // DBKDA 2025, The Seventeenth International Conference on Advances in Databases, Knowledge, and Data Applications // View article

Exploring Latent Concepts in SHAP Values -A New Approach Using Singular Value Decomposition -

Authors:
Yukari Shirota
Tamaki Sakura

Keywords: XAI; Shapley values; SHAP; Singular Value Decomposition; India automakers

Abstract:
In this paper, we introduce a novel explainable AI method called “SHAP_SVD” for regression analysis. The Shapley value, originally developed by Lloyd Shapley, has gained prominence as a key tool in explainable AI (XAI) through its adaptation as SHAP by Lundberg. In regression analysis, SHAP values are computed using characteristic functions of the data, representing the contribution of each explanatory variable to the target value. Our proposed SHAP_SVD method applies Singular Value Decomposition (SVD), a dimensionality reduction technique, to the SHAP value matrix. The eigenvalues and eigenvectors extracted via SVD capture the core structure of the SHAP matrix, revealing "concepts" or "latent semantic concepts." In SVD, these concepts are represented by two sets of eigenvectors, one from the matrix UΣ and the other from ΣV. As a case study, we demonstrate the regression analysis of stock price growth rates for Indian and Japanese automakers, where two principal concepts were identified, consistently reflected across both sets of eigenvectors.

Pages: 1 to 6

Copyright: Copyright (c) IARIA, 2025

Publication date: March 9, 2025

Published in: conference

ISSN: 2308-4332

ISBN: 978-1-68558-244-9

Location: Lisbon, Portugal

Dates: from March 9, 2025 to March 13, 2025