Supervised incremental feature selection using regularization vector for dynamic multi-scale interval valued datasets

Abstract

Feature selection is pivotal for enhancing machine learning and data mining models, where its accuracy directly affects model performance and applicability. Traditional methods often overlook the dynamic nature of data and the multi-scale aspect of high-dimensional datasets, leading to limitations in real-world applications. This paper introduces a novel incremental feature selection method using a regularization vector $\left(RV\right)$ tailored for dynamic multi-scale interval valued fuzzy decision systems $\left(D\text{-}MIvFD\right)$. The paper first establishes the concepts of object affiliation relation and class, providing a theoretical basis for integrating replay and regularization. It then introduces the affiliation contradictory state $\left(ACS\right)$ and $RV$, broadening the application of contradictory state $\left(CS\right)$ in dynamic settings and enabling efficient feature selection. The integration of regularization and replay strategies is realized through four algorithms designed for different update patterns. Empirical results across various datasets show that the proposed method significantly outperforms multiple conventional techniques, highlighting its practical potential for real-world deployments.