Abstract：To explore the specific features of each label in multilabel data sets and effectively utilize the specific features of the label, the multilabel learning based on boosting clustering trees (MLLBCT) was proposed. The overall framework of BCTMLL was built, and the correlations among data samples were explored to deal with multilabel learning problems. The clustering feature tree was introduced to find the correlation between data samples, and the randomly selected subsets of training data were learned to construct label specific features, which could generate proper features for each label and improve the overall performance of multilabel learning. The proposed method was compared with the classical multilabel learning methods of LIFT, LLSF, REEL and LLSFDL in specific feature fields on 11 data sets such as flag and emotions. The results show that the new method can effectively improve the prediction performance in all evaluation indicators of Hamming Loss, Oneerror, Ranking Loss, Average Precision, Microaveraged FMeasure, and the proposed method is simple and flexible.
