Research On Multi-label Learning Method Based On Local Label And Feature Correlations

In practice,an instance object is labeled with multiple category labels,which is called multi-label learning.Multi-label learning is dedicated to learning a classifier based on existing labeled data and then using the learned classifier to predict all relevant category labels for new data.Label correlations play an important role in multi-label learning and the modeling of label correlation can help reduce the category labels that need to be predicted in multi-label learning.At present,global label correlation has been widely used in multilabel classification,but under some circumstances,different instances correspond to different label correlations,so the label correlation may be locally applicable.In addition,most of existing methods are only focused on using label correlation to improve the performance of the model,while ignoring the influence of the relationship between features on multi-label learning.Similar to label correlation,feature correlation may also be locally applicable.And learning local feature correlation can more fully mine the relationship between features,so as to better remove the redundancy in the feature space and provide a compact feature space for multi-label learning.In this thesis,local correlations of labels and features are integrated into a unified framework,and a multi-label learning method based on label and feature local correlation is proposed.The method first clusters the training data into multiple data subsets by clustering method,and corresponding label subsets of data subsets are obtained according to the different data subsets.Then local label correlation and local feature correlation are obtained by label manifold and feature manifold on these data subsets and label subsets,and a corresponding classifier is learned for each data subset to further improve the performance of the multi-label classification method.In addition,a new regular term is constructed to obtain the relationship between different subsets by controlling the consistency of model coefficients.For the sake of improve the accuracy of prediction,the arithmetic mean of the prediction results of multiple classifiers is taken as the final test result of the predicted data.On twelve common multi-label data sets,seven evaluation indexes commonly used in classification performance evaluation are selected for experimental analysis,and the effectiveness of the proposed multi-label classification method is verified.