Multi-label learning aims at assigning more than a class label to each instance. Due to the development of digital technologies, in many real-world applications that are high-dimensional. Feature selection methods are widely used in multi-label learning to reduce the dimensionality of the data. In this paper, multivariate and filter feature selection method based on the ant colony optimization and mutual information is proposed. The proposed method employs the ant colony optimization to rank the features based on their importance. To this end, first the search space is mapped to a graph and then each ant moves through the graph to select a predefined number of features. Moreover, a novel information-theoretic measure is proposed to evaluate the features selected by each ant. This measure uses the concept of mutual information to calculate the relevancy of each feature with a set of labels. This measure is also employed to assess the redundancy between selected features. The pheromones of features are assigned based on the quality of solutions founded by the ants. Finally, the features are sorted based on their assigned pheromones, and then those of top features are chosen as the final feature set. The proposed method is performed on some real-world datasets and the results show the superiority of the proposed method in comparison with some well-known and state-of-the-art methods.
