An Improved Multi-label Classification Based On Label Relationship

Multi-label classification problem is a branch of traditional classification problem and widely used in the real world concluding data mining,scene classification,protein function classification,music categorization,,and text categorization.The concept of multi-label classification problem come from the comparison with tradition binary classification problem.In traditional classification,normally we have two labels and each instance would belong to one of them,while in the multi-label classification,we may have more than two labels and each instance would belong to uncertain number of labels.This kind of difference make multi-label classification more difficult to solve.In fact,Overlapped and related labels will cause mis-classification which also happened in multi-label problem.According to solve this problem,this paper proposed an label relationship and classification confidence based algorithm and try to extract the label relationship from training data and use it combine with the SVM part by the dynamic weight strategy and try to improve the performance of multi-label classification problem.Firstly,BR decomposition technique is used to divide a multi-label classification task into multiple independent binary classification sub-problems.By using the probability output SVM,we can get the classification probability.Secondly,based on training data,label relationship could be obtained.And by the cross-validation and KNN method,classification confidence could be obtained.Thirdly,combine the result of binary classification,classification confidence and label relationship,and we could obtain the final classification probability.Then get the instance ranking and use the threshold got from the cross-validation to decide the label of each instance.Experiment results on three well-known multi-label benchmark datasets show that the proposed method outperforms some conventional multi-label classification methods.