MSSL: a memetic-based sparse subspace learning algorithm for multi-label classification

Researchers have considered multi-label learning because of its presence in various real-world applications, in which each entity is associated with more than one class label. Since multi-label data suffers from the curse of high-dimensionality, providing effective feature selection methods is necessary to enhance the learning process. Various multi-label feature selection methods have been proposed so far. However, the existing methods have not yet reached acceptable performance in this research field due to the existence of datasets with various dimensions. This paper proposes a new feature selection algorithm based on subspace learning and a memetic algorithm to provide global and local search in multi-label data. This is the first try that uses a filter-based memetic algorithm for multi-label feature selection. The objective function consists of two conflicting objectives: reconstruction error and sparsity regularization. Finally, nine filter-based multi-label feature selection methods are compared with the proposed method. The comparisons are conducted based on the famous performance evaluation criteria for multi-label classification, such as classification accuracy, hamming-loss, average precision, and one-error. Based on the results obtained in eight real-world datasets, the proposed method is superior to comparing methods according to all evaluation criteria.