Multi-label classification aims at considering multiple labels for each sample. The main issue of multi-label classifiers is reducing their performances while tackling with high dimensional tasks. Feature selection is a general approach to overcome this issue. This paper proposes a multi-label feature selection approach using ant colony optimization (ACO). The proposed method is a multivariate and filter approach, which integrates the mutual information with ACO to rank features. In the early step of the proposed method, the solution space is converted to a graph where nodes represent features, and each shows the relevancy between its corresponding features. In this paper, the relevancy between features is calculated by using the mutual information theory. Then, a graph of features is generated, and then ants are moved through the graph to weight the features based on their importance. In this paper, a novel measure based on mutual information is used to evaluate solutions chosen by the ants. This means that in this paper there is no need to use any learning model to evaluate the solutions and thus the proposed method is classified as a filter approach. The quality of each solution is added to its corresponding features as their pheromone values. Finally, the final feature set including those features which take high pheromone values will be created. The performance of our method is evaluated by performing a set of experiments on five datasets. The obtained results show the better performance of our method in comparison with the other methods.
