Research On Multi-label Active Learning Based On Label Correlation

With the rapid development of Internet information technology,the quantity of data is exploding.Classification is major way to handle these data.Multi-label active learning techniques can effectively reduce the workload of sample labeling in constructing the classifier model.Previous research had focused on the sampling part and neglected the labeling part.In order to reduce the labeling cost of obtaining the image data,the thesis focuses on training sets selection technology in multi-label image classification,and the research scheme of active learning-based multi-label image classification method is proposed.The whole research work is as follows:(1)Research the sampling strategies of multi-label active learning for data classification.Research and analysis two methods: example based and example-label based.Analysis the superioror shortcomings of typical algorithms,it provides a theoretical basis for subsequent research.(2)Most of the previous studies in active learning for multi-label classification have two shortcomings.One is they didn't pay enough attention on label correlations.The other shortcoming is that the selections of existing example-label methods are centralized on few particular labels.We proposed a chi-square statistics multi-label active learning(CSMAL)to use chi-square statistics to accurately evaluate the correlations between all labels.CSMAL not only considers the positive relationships but also the negative ones.It further uses the average correlation between the rest unlabeled labels and the selected label as the label information for each sample-label pair.CSMAL integrates uncertainty and label information to selectexample-label pairs for each label.(3)The studies of multi-label active learning did not consider the labeling procedure.We propose a semi-supervised multi-label active learning(SSMAL)method to automatically label example-label pairs.Firstly,we select the most informative example-label pairs with uncertainty and label correlations.Then we combine the predicted value of the classifier,the value of the nearest neighbor example-label pair and the value of the most correlated label to automatically label the selected example-label pair.In this thesis,the proposed method is demonstrated.Our empirical results demonstrate that our proposed method CSMAL and SSMAL outperforms the state-of-the-art active learning methods for multi-label classification.It significantly reduces the labeling workloads and improves the performance of a classifier built.