Multi-label feature selection has become an indispensable pre-processing step to deal with possible irrelevant and redundant features, to decrease computational burdens, improve classification performance and enhance model interpretability, in multi-label learning. Mutual information (MI) between two random variables is widely used to describe feature-label relevance and feature-feature redundancy. Furthermore, multivariate mutual information (MMI) is approximated via limiting three-degree interactions to speed up its computation, and then is used to characterize relevance between selected feature subset and label subset. In this paper, we combine MMI-based relevance with MI-based redundancy to define a new max-relevance and min-redundancy feature selection criterion (simply MMI). To search for a globally optimal solution, we add an auxiliary mutation operation to existing binary particle swarm optimization with mutation to control the number of selected features strictly to form a new PSO variant: M2BPSO. Integrating MMI with M2BPSO builds a novel multi-label feature selection method: MMI-PSO. The experiments on four benchmark data sets demonstrate the effectiveness of our proposed algorithm, according to four instance-based classification evaluation metrics, compared with three state-of-the-art feature selection approaches.
Poli, R., Kennedy, J., Blackwell, T.: Particle swarm optimization: an overview. Swarm Intell. 1(1), 33–57 (2007)CrossRefGoogle Scholar
Zhang, Y., Wang, S., Phillips, P., Ji, G.: Binary PSO with mutation operator for feature selection using decision tree applied to spam detection. Knowl.-Based Syst. 64, 22–31 (2014)CrossRefGoogle Scholar
Zhang, M., Zhou, Z.: ML-kNN: A lazy approach to multi-label learning. Pattern Recognit. 40(7), 2038–2048 (2007)CrossRefGoogle Scholar