Multi-label Feature Selection based on Label-specific Features

Multi-label learning algorithm handles cases in which each sample is related with several labels synchronously. As is known to all, each label might possess its own peculiarities, such as LIFT algorithm, i.e. multi-label learning with Label-specific Features. It constructs feature by performing cluster techniques based on negative and positive training samples of each label. However, the main drawback of this kind of algorithm is the large amounts of irrelevant features or redundant features in its feature space. To solve this problem, this paper puts forward an effective algorithm named LEFS, i.e. multi-label Feature Selection based on Label-specific features with fuzzy Entropy. The approaches proposed are examined on the two realistic multi-label benchmark data sets, which are compared with several multi-label learning approaches. A few features are selected from original features to fed classifier, but they remain the same or even slightly improve accuracy from 91.82% to 92.49% on data set- Medical. Results of another data sets are similar to that of the Medical. Experiment results show that these approaches can not only decrease the dimension of the construct features, but also gain an effective classification performance compared with three well-established multi-label learning approaches.