Research On Label Distribution Learning Based On Granular Computation And Label Correlation

Label distribution learning has a wider application scenario than multi-label learning,multi-label learning can only describe whether an instance is related to a label,and label distribution learning can further describe the degree of correlation between instances and labels.There are at least the following problems in the existing label distribution learning:(1)Label correlation is an important factor affecting label distribution learning performance,however,most of the existing studies on label distribution assume that all samples have the same label correlation rules,this assumption is not entirely true.Therefore,this paper uses clustering algorithm to dig into local label correlation.(2)Most of the existing label distribution learning methods analyze the correlation among all labels simultaneously,this method tends to cause high time complexity.Therefore,this paper designs a label granulation algorithm based on the idea of granular computing,group the most relevant labels into a category,correlations between labels in the same class are then analyzed.(3)Most of the existing label distribution learning assumes that the label space is determined by the complete feature space,all features are considered equally important and all features are given the same weight.So this paper introduces the idea of feature selection,learn a specific feature set for the label space,the maximum information coefficient MIC was introduced to assign different weights to different features.After the above discussion,the research work in this paper includes:(1)A label granulation algorithm based on neighborhood mutual information NMI and granular computing theory is designed,group highly relevant labels into the same category,label dependencies are then analyzed in each label class.The correlation analysis is not carried out between labels with less correlation and the number of labels in the label class is small,therefore,analyzing label correlations in label classes is much less complex than analyzing correlations among all labels at the same time.(2)To mine local label dependencies and global label dependencies,in this paper,a clustering algorithm is used to cluster training sets in the label space,then learn its unique label dependencies for each class cluster.In order not to lose global label dependencies,simulated annealing algorithm is introduced to combine global label correlation with local label correlation,a label distribution algorithm based on particle computation and label correlation(LDL-PCALC)was proposed.(3)Given the advantages of integrated learning,in this paper,two feature selection strategies based on integration correlation are designed based on maximum correlation and minimum redundancy.A feature selection strategy is used to learn a specific feature set for the label space,different weights were assigned to different features based on the maximum information coefficient MIC.(4)In order to dig the mapping relationship between feature space and label space,in this paper,fuzzy C-means clustering algorithm is introduced to cluster training sets in feature space,then the mapping relationship between the feature and the label is used to predict the label,a label distribution algorithm based on feature selection and fuzzy clustering(LDL-FSAFC)is proposed.(5)The results show that LDL-PCALC and LDL-FSAFC have good performanc e and stability.