Research On Multi-label Classification Based On Classifier Chains

Along with the rapid development of science and technology,multi-label learning has become a hotspot of machine learning,and in real life there are widely various examples of multi-label learning.Unlike traditional classifiers problem,multi-label classifiers usually assign each instance to one or multiple labels.In the research of multi-label classification algorithm,classifier chain is one of the most common methods.Classifier chain can not only model label correlations on the basis of binary relevance method and maintain acceptable computational complexity.However,on the one hand,although it takes into account label correlations,it ignores the potential label redundant information and there is no specific measure of label correlations.On the other hand,the order of labels affects final classification results in classifier chain and it is easy to lead to propagation of information errors.Therefore,in this paper we research two existing problems in classifier chain,respectively is:Firstly,for the expansion attribute process of each base classifier in the classifier chain,this paper presents a selective classifier chain,it is mainly used for selective expansion attributes and finally has obtained the good effect.First,for each base classifier chooses labels as its extended attributes,it ignores the redundancies between labels and there is no specific measure of label correlations.Because the expansion process of attribute is actually a feature selection process and it also considers the correlations and redundancies between attributes,this paper combines maximum-relevanceminimum-redundancy feature selection algorithm with attribute expansion process.Secondly,the algorithm is mainly used to measure and select the maximum relevance minimum redundancy attributes,which not only considers attribute relevances and redundancies but also reduces the number of expansion attributes,thus a selective classifier chain with simple structure and good classification performance is obtained.Finally,the experiment is carried out by using multiple multi-label datasets and some multi-label classification algorithms,the results show that the selective classifier chain is effective in this paper.Secondly,for the determination of class ordering problem in the classifier chain,this paper proposes an ordered classifier chain,which uses the correlations and redundancies of labels to find the appropriate label ordering for the classifier chain.First,for the determination of class ordering problem in the classifier chain,it seriously affects final classification results and it is easy to lead to propagation of information errors.With the increasing number of class variables,there are more and more random sequences of chains,learning the complexity of classifier chain will increase,and there are very few ways to overall measure the relationship between class on this issue,so this article uses maximum-relevance-minimum-redundancy feature selection algorithm to measure concretely class relevances and redundancies,each class variable is selected with distinct subsets of the maximum relevance minimum redundant label.Secondly,based on the selected results,each class variable selects an optimal label subset from different label subsets,then using optimal label subset information of each class variable to determine the appropriate chain sequence for classifier chain.Finally,the experiment is carried out by using the multi-label datasets,the results show that the ordered classifier chain has excellent classification performance.Finally,this paper summarizes the research contents and results,which lays a foundation for the further research classifier chain.