In recent years, label distribution learning (LDL) has gained extensive application in actual classification endeavors. Nevertheless, most LDL datasets contain superfluous features that significantly impede the efficacy of classification algorithms and prolong their execution time. In addition, existing feature selection algorithms are usually limited to a specific label assignment paradigm or ignore the overall distribution information of the label significance in LDL. To address the aforementioned issues, this study constructs a novel uncertainty model known as neighborhood fuzzy rough sets, and proposes a LDL feature selection algorithm (LDNF) for better utilization of the overall distribution information of label significance. Specifically, the fuzzy equivalence relation for samples is defined using the correlation of label distributions, and the probabilities of classifying the same category are evaluated. To exploit the overall distribution information, the definitions of the upper and lower approximations are provided with the fuzzy equivalence relation of the label distribution. The experimental results indicate that the performance of the LDNF algorithm is superior to that of five leading-edge comparative algorithms on 10 publicly available real-world datasets. Neighborhood fuzzy rough sets and their applications will be further investigated in the future.
Read full abstract