Towards Multi-label Classification

Objects in real world are complicated and diverse,a single label cannot describe the whole semantic information.In order to describe the rich semantic information of objects,multi-label learning has emerged and applied in various fields,such as text classification,scene analysis and so on.The task of multi-label classification is to learn a model that predicts a set of related labels for an unseen instance.Correlations existing among labels make the task of multi-label classification harder than multi-class classification.In order to make the model have high accuracy and generalization,this paper studies the multi-label classification from two aspects: label independence and label relationship enhancement.The main research works are as follows:We propose a new ensemble method based on LIFT method,which is a label independence method.The main limitation of LIFT is to adopt k-means in training stage.K-means is a heuristic method that has the risk of getting stuck in local optima and is unstable for high-dimensional data.For this issue,we propose to mitigate the limitations for high classification accuracy and generalization with ensemble way.Specifically,our approach firstly constructs multiple LIFT classifiers under Bagging framework.Furthermore,different classifiers are weighted automatically according to the loss of each classifier.Finally,for each instance,the predicted label is obtained by weighted ensemble classifiers learned.Experimental results demonstrate the label independence method adopting ensemble way achieves better performance.We propose a new joint learning method based on label independence and label correlation enhancement method.In order to improve the performance of classification,the proposed method introduces two virtual labels,namely virtual relevant label and virtual irrelevant label,and claims there is a large margin between two virtual labels according to the thought of label ranking,where the virtual relevant label is related with all relevant labels and the virtual irrelevant label is related with partial irrelevant labels that are hard classified.Experimental results demonstrate the proposed method achieves promising high classification and generalization.