AMFSA: Adaptive fuzzy neighborhood-based multilabel feature selection with ant colony optimization

For multilabel classification, the correlations among labels of samples are always ignored by existing feature selection models, which results in inefficient predictions. In addition, the neighborhood radius of samples needs to be manually set, which requires much computation. To effectively conquer these limitations, this article proposes a novel adaptive fuzzy neighborhood-based multilabel feature subset selection approach with ant colony optimization (ACO) for multilabel classification. First, the feature cosine similarity and the label Jaccard similarity between samples are constructed from the feature space and label space. By combining the two abovementioned similarities, the entire similarity between samples is proposed to effectively reflect the similarity between samples in the overall space, and a dynamic adjustment coefficient is developed to control the importance of label similarity. The discriminant relation is proposed to judge the homogeneous or heterogeneous relationship between samples. Second, to address the problem that the fuzzy neighborhood radius is usually chosen artificially, we calculate the average distance between the target sample and all heterogeneous or homogeneous samples. The difference between these two average distances can be used as the adaptive fuzzy neighborhood radius of the target samples. Then, novel adaptive fuzzy neighborhood rough sets are presented, and the fuzzy neighborhood dependency degree is studied to evaluate this distinguishing ability between features and samples under a fuzzy background. Furthermore, neighborhood entropy measures and adaptive fuzzy neighborhood mutual information are investigated. Finally, by integrating fuzzy neighborhood mutual information and dependency degree into the state transition rules of the ACO, a novel ACO-based feature selection algorithm is constructed to achieve this optimal feature set for multilabel classification. Experiments applied to 15 multilabel datasets prove that our developed algorithm is effective in achieving an excellent feature subset with great classification efficiency.