Attribute Reduction For Multi-label Classification Based On Identifiable Label Pairs Decomposition

With the development of multi-label classification technology,different multi-label classification algorithm is applied to various fields.The attribute reduction is an important part of multi-label classification.This paper considers the correlation between labels by comparing identifiable label pairs.The multi-label classification problem is transformed into a series of binary classification problems.Three attribute reduction algorithms are proposed with rough set theory.In the second part,the samples are divided by pairwise comparison of labels.For label pairs(l_j,l_k),the samples related to labell_j and unrelated to labell_k are defined as a positive sample set.Otherwise,they are defined as negative sample set.According to rough set theory,the lower approximation of the positive samples is redefined.The corresponding algorithm were designed.Multi-label data set was selected for numerical experiment.The algorithm is compared with the traditional multi-label positive region reduction algorithm and other attribute reduction algorithms to verify the effectiveness of the proposed algorithm.In the second part,we only considering the positive sample and ignore the negative samples.In actual classification task,in same circumstances,we will pay attention to the sample without the labels.In order to solve the different labels classification task,in the third part,we propose an attribute reduction algorithm based on negative sample set,and redefined the positive region and dependence reduction.In addition,in the classification task with high classification accuracy requirements,we will pay attention to the positive and negative samples.In the fourth part,we propose an attribute reduction algorithm for the positive and negative samples.The experiments result show that the propose a algorithm can not only on complete attribute reduction reducing classification accuracy but also can effectively enhance the performance for the multi-label classification.