Fuzzy label rate-guided feature selection with three-way mixed entropy in adaptive fuzzy neighborhood rough sets

Feature selection based on label consistency is crucial for data analysis, yet many methods lack robustness to ambiguous or conflicting samples due to rigid evaluation criteria and limited modeling of local structures and uncertainty. To address this, we propose FLR-MEFS, a structure-aware feature selection method within adaptive fuzzy neighborhood rough sets. FLR-MEFS introduces a fuzzy label rate (FLR) to quantify label consistency continuously and to adapt neighborhood radii to local variations. Guided by FLR, samples are semantically partitioned using three-way decision theory to represent different levels of local uncertainty. Based on this partition, we design a three-way mixed entropy criterion to evaluate feature subsets by distinguishing informative boundary ambiguity from detrimental inter-class conflict. Finally, a greedy entropy-minimization strategy is employed to obtain compact and discriminative feature subsets. Experiments on 18 public datasets show consistent gains over nine state-of-the-art methods in both robustness and effectiveness.