An improved MLTSVM using label-specific features with missing labels

Multi-label twin support vector machine (MLTSVM) is an excellent multi-label classification algorithm, which has attracted much attention. Although MLTSVM can effectively solve the multi-label classification problem, it has some drawbacks. a) MLTSVM uses the same feature representation for each label, but in practice, each label has its own specific features. Therefore, MLTSVM might not obtain the optimal classification results. b) In practical applications, there are a large number of samples with missing labels and only a small number of samples with complete labels, because it is expensive to obtain all labels of samples. However, MLTSVM can only use expensive samples with complete labels, not cheap samples with missing labels. For the above drawbacks, we propose an improved MLTSVM using label-specific features with missing labels (LSFML-MLTSVM) in this paper. LSFML-MLTSVM first extracts label-specific features via using semi-supervised clustering analysis and then obtains the structure information of samples and the geometry information of the marginal distribution. Furthermore, in the label-specific feature space, the above two valuable information is introduced into MLTSVM to reconstruct the classification model. Finally, the successive overrelaxation (SOR) algorithm is used to solve the classification model efficiently. Experimental results on benchmark multi-label datasets show that LSFML-MLTSVM has better classification performance.