Research And Application Of Multi-label Classification Methods

In order to accurately model real-world multi-semantic objects,the multi-label learning framework assumes that each object is simultaneously associated with multiple class labels,and the goal is to train a predictive model which can assign all relevant labels for the unseen instance.In contrast to the concept that labels are independent in the traditional classification framework,labels have correlations with each other in the multi-label learning problem,and a lot of work shows that classifier performance can be effectively improved by mining label correlation.Therefore,based on the consideration of label correlation,this paper studies multi-label classification algorithms based on label correlation.The main research work is as follows:1.For the problems that Multi-Label Learning with Label Specific Features does not consider label correlation,this thesis proposes a new algorithm called Label Correlation based Multi-Label Learning with Label Specific Features.The related labels are divided according to the correlation between label pairs measured by the label distance.The extension of the original feature space and the introduction of label correlation are completed by attaching related labels to the label specific feature space.Experiments were performed on 7 multi-label benchmark data sets and compare the proposed algorithm with other representative multi-label algorithms.The experimental results show that the performance of the proposed algorithm on multiple evaluation metrics is improved by 7.46%on average,which verifies the effectiveness of the proposed algorithm.2.For the problems that Multi-Label classification algorithm derived from Nearest neighbor rule with label dependencies ignores the examples and label diversity while considering label correlation,this thesis proposes a new algorithm called Local Label Correlation based k-Nearest Neighbor Multi-Label Learning algorithm.Local label correlation is introduced according to the label distribution of each unseen example neighbors,and is used to optimize the setting of the margin vector value in the algorithm.Experimental verification was performed on 7 multi-label benchmark data sets,and the comparative experimental results of 5 multi-label evaluation metrics were analyzed.The experimental results show that the classification performance of the proposed algorithm has been improved on metrics such as One-error,Ranking loss,and Average precision,which verifies the effectiveness of the proposed algorithm.3.The two proposed multi-label classification algorithms based on label correlation are applied to news text classification problems.Experiments were performed on 6 multi-label news benchmark data sets and 1 news data set collected from Sohu News.The experimental results show that the proposed algorithm has higher classification accuracy on the news dataset.