Research On Classification Algorithm Based On Multi-label Learning

Multi-label classification has a more general form than traditional classification,and has many applications in text annotation,image recognition,and biological function prediction.However,multi-label classification algorithms generally face problems such as high complexity,high sample data dimensions,and class imbalance.In this thesis,the performance of multi-label classification algorithm is improved by making full use of label correlation and dimension reduction of label space.The main research contents are as follows:Aiming at the problem that the multi-label classification algorithm can not fully utilize the label correlation,a multi-lable algorithm based on gravity model is proposed.The algorithm mines the different correlations between labels by establishing the positive and negative correlation coefficients of the labels.First,based on the label distribution of all neighbor samples in the neighbor set of each sample,a positive and negative correlation matrix is established for each training sample to obtain the correlation between the labels;then,the neighbor set of each training sample is calculated.The neighbor density and the neighbor weight are used as the parameters of gravity calculation in the gravity model.Finally,the multi-label classification function is constructed by judging the gravity between the data particles.The results of experiment show that the proposed algorithm can effectively use the different correlations between labels,and improve the performance of the algorithm.The average value of each index is increased by 5.66%.Aiming at the problem of class imbalance of multi-label data and classification algorithm dealing with data with large-scale labels,a multi-label classification algorithm based on equalization local linear embedding is proposed.Firstly,an index that measures the imbalance of categories is proposed and applied to the clustering of data clusters to reduce the impact of class imbalance on the classification results.Then,the improved local linear embedding is used to reduce the dimension of the label part of the multi-label data,so as to avoid the extremely high complexity brought by the high-dimensional label.Finally,the classification of the data is done by using a simpler classifier in a low-dimensional space.The experimental results show that the proposed algorithm can effectively compress the label space,improve the classification efficiency,and obtain better classification results on the data with large-scale labels.The average value of each index is increased by 8.06%.