Feature selection from an interaction perspective: a review of methods, advances, and challenges

Feature selection plays a vital role in machine learning and data mining by identifying a representative subset to improve model performance, reduce computational cost, and prevent overfitting. Although traditional methods aim to eliminate redundant and irrelevant features, research into feature interactions has not been sufficiently studied. Feature interaction refers to the joint contribution of multiple features to predictive performance; effectively capturing such interactions can substantially enhance model accuracy. This article reviews the evolution of feature selection methods over the past three decades, highlighting their strengths and limitations. In addition, it investigates efficient strategies for identifying interactive features, taking into account relevance, redundancy, interactivity, and complementarity. To provide a broader and up-to-date perspective, recent advances in interaction-aware, explainability-driven, and deep learning-based feature selection methods are also discussed. Finally, the article summarizes the open issues in the search for feature subsets and outlines key challenges for future research.