Multi-label Classification Based On Feature Screening And Marker Correlation

In addition to the traditional single classification,there are many data with polysemy,which need to be represented by multiple labeling information,and these data are called multi-labeling data.The topic of this paper is multi-label classification.There are a large number of features in the multi-label data set,which not only makes the calculation process more complicated,but also produces redundancy,resulting in the decrease of the accuracy of the experiment.At present,scholars have adopted some methods for feature screening,but seldom take into account the correlation between markers,which often has a certain correlation.If the correlation between markers can be used on the basis of feature screening,the result of classification can be improved to a certain extent.Therefore,this paper first proposes a multi-label classification algorithm(MIR)based on mutual information feature screening and association rule marker correlation,which performs feature screening based on global feature set.At the same time,through existing studies,it is found that the characteristics of screening are similar when the marker correlation is similar,and the marker correlation exists in local instances.Therefore,this paper carries out feature screening on the basis of MIR algorithm,finds feature subsets in the class,and proposes a feature screening algorithm based on global and local mutual information(MIFSGL).The research content of this paper is as follows:(1)Multi-label classification algorithm based on mutual information and association rules(MIR): In multi-label classification,most studies only focus on feature screening to improve classification results,and ignore the correlation between markers.Therefore,this paper selects important features by setting thresholds based on the correlation between markers and features,and then updates the tag set by using association rules between markers to achieve the correlation between markers.Finally,comparison experiments with other algorithms are carried out on five data sets.The results show that the algorithm is better than the comparison algorithms.(2)Global and local mutual information feature screening algorithm(MIFSGL)and its improved algorithm(IMIFSGL): In order to make better use of marker correlation screening features,this paper firstly uses DPC clustering algorithm to achieve local clustering of marker sets,and filters out local features according to the corresponding table with the highest correlation between features and markers in the class.Because this method screened fewer features,in the case of sparse marker space,the classification results would decline.Therefore,an improved method is proposed in this paper to filter features by establishing a corresponding table between features and multiple tags with high correlation.Considering the influence of global important features on classification,this paper adopts global and local feature fusion.Experiments on six data sets show that the algorithm has some improvement in One error and Average precision.