Attribute Reduction For Multi-label Classification Based On Chain Decomposition

In a multi-label classification,each sample is associated with more than one label.And the number of related labels is not certain,can be any combination of all labels.Therefore,the key to multi-label classification is to use the correlation between labels to improve the computational efficiency of the algorithm.In this paper,we solve the multi-label problem by considering the chain correlation between labels.First,the labels are reordered and the number of samples related to each label is counted.When there is more certain labels in the neighborhood of the sample,it is considered that the label is related to the sample.Arrange labels in descending order according to the number of the relevant samples.The pre-order label is added to the attribute set and the post-order label is used as the category label to construct a chain of single label classification problems.Finally,two attribute reduction methods based on chain decomposition are proposed with rough set theory.In chapter 2,two kinds of label sorting methods are proposed,which are local sample counting method and global sample counting method.The sorted labels are successively added to the attribute set to establish multi-label chain decomposition.For each subproblem,the similarity class,positive domain and dependency were redefined,and the attribute reduction algorithm was designed.The experimental results show that the proposed method has better reduction effect and classification accuracy on multi-label datasets.The third chapter mainly combines the concept of k-nearest neighborhood with the chain decomposition of multi-label classification.A new attribute reduction algorithm is constructed,Where k-nearest neighbor refers to the k samples that are closest to the sample.k-nearest neighborhood add a radius restriction to k-nearest neighbor,and only the k-nearest neighbors whose distance from the sample is less than the specified radius are selected to construct similar classes.Different from the past,when were calculated the k-nearest neighborhood of the sample,we should not only calculate the distance between the original attributes,but also calculate the distance between the pre-order labels.Then we regard the k-nearest neighborhood as a similar class,and give the definitions of the lower approximation,positive region and dependency of each chain subproblem.Finally,text classification,scene recognition and other data sets are selected for verification experiments.The experimental results show that the k-nearest neighborho-d rough set reduction algorithm based on chain decomposition can remove most of the redundant attributes without reducing the classification accuracy.