Uncertainty-Aware Feature Selection Framework Based on Three-Way Decision Theory for Explainable Machine Learning

Abstract

Feature selection is a critical stage in machine learning because it improves predictive performance, reduces computational complexity, and enhances model interpretability. However, many existing feature selection methods rely on binary selection strategies that either retain or discard features based on fixed relevance thresholds. This approach can be inadequate when features contain uncertainty, instability, redundancy, or borderline predictive value. To address this limitation, this article proposes an uncertainty-aware feature selection framework based on three-way decision theory for explainable machine learning. The proposed framework classifies features into three decision regions: accepted features, rejected features, and deferred features. Accepted features are considered highly relevant and reliable, rejected features are removed due to low relevance or high redundancy, while deferred features are subjected to further evaluation because of uncertain or inconsistent importance. By integrating uncertainty measurement with feature relevance scoring, the framework provides a more flexible and transparent alternative to conventional two-way feature selection. The framework also supports explainability by providing clear justification for why each feature is selected, rejected, or deferred. This makes it particularly useful for high-dimensional and high-stakes machine learning applications where model transparency, reliability, and decision accountability are important. The proposed approach is expected to improve feature selection stability, reduce uncertainty in selected feature subsets, and enhance the interpretability of machine learning models.