BGR-FS: A bandit-guided redundancy-aware framework for high-dimensional multi-objective feature selection

High-dimensional data usually contain numerous irrelevant and redundant features, which makes multi-objective feature selection challenging in terms of search-space reduction, redundancy control, and exploration–exploitation balance. To address these issues, this paper proposes Bandit-Guided Redundancy-Aware Feature Selection (BGR-FS), a hybrid evolutionary framework built on NSGA-II. BGR-FS first uses multiclass ANOVA F-score to construct a reduced candidate subspace, and then exploits F-score and ReliefF information to guide the evolutionary search. During offspring generation, a Pearson-correlation-based redundancy control mechanism is introduced to suppress the repeated selection of highly correlated features. Meanwhile, an $$\varepsilon $$-greedy multi-armed bandit strategy adaptively adjusts the crossover intensity according to online search feedback. The resulting framework simultaneously optimizes feature subset size and cross-validation error. Experiments on 16 high-dimensional benchmark datasets show that BGR-FS generally obtains competitive Pareto solution sets and compact feature subsets with favorable classification performance. Additional analyses based on external SVM validation, AUC evaluation, runtime comparison, ablation study, and parameter sensitivity further support the effectiveness and practical competitiveness of the proposed framework under the adopted experimental protocol.