Research On Key Technologies For Multi-label Learning Under Complex Background

In multi-label learning,each object is commonly represented by a single instance,which is associated with multiple labels simultaneously.The purpose of multi-label learning is to construct a learning framework.Under this framework,relevant labels can be predicted for unseen instances.Multi-label learning has been successfully applied to a variety of domains,such as text categorization,multimedia content understanding,and bioinformatics.With the extension of the applications,the research for multi-label learning must adapt to the actual need,thereby providing the solution for more complex learning problems.Thus,based on the aforementioned standard multi-label learning framework,this paper tries to conduct the research on key technologies for multi-label learning under complex background,and the main work is summarized as follow:(1)High-dimensional data is often notorious to tackle for multi-label learning.To tackle the learning problem,previous methods for multi-label feature selection are either directly transformed from traditional single-label feature selection methods or half-baked in the labeling information exploitation.We propose an embedded multi-label feature selection method MDFS based on manifold regularization.The proposed MDFS approach generates a low-dimensional embedding based on the feature space,which can fit local and global label correlations to reduce dimensionality.Experimental results demonstrate that MDFS can obtain an optimal feature subset to improve the performance.(2)Aimed at seeking inherent characteristics of labels,i.e.label-specific features,we propose a label-specific feature learning method MLFC for multi-label classification.MLFC constructs an optimization framework.Under this framework,label-specific features are available via the sparse learning based method,in the meanwhile,additional features,mapped by label correlations information,are regarded as the candidates of label-specific features.Experimental results reveal that MLFC has the advantage in multi-label learning,especially label-specific feature learning.(3)In multi-label learning,the resource of feature representation may come from multiple sources.Based on multi-source(heterogeneous)multi-label data to modeling,we propose a multi-source aware optimization framework MLSO for multi-label consensus learning.MLSO first generates the multi-label prediction based on each data source by using neighborhood information,and then assigns the source weights to obtain the multilabel consensus classification result.Experimental results show that MLSO not only can achieve the performance improvement by combining with existing multi-label learning methods,but also can fuse multiple sources to further improve the performance.(4)For weakly supervised multi-label learning,we propose a new learning method MSWL with labeling information augmentation.Based on the data,i.e.,partially labeled data even abundant unlabeled data,MSWL digs out the side information,such as " high-order" label relationship and local structures of features,to train multi-label classifier,which is not only capable to enrich the existing incomplete label assignment,but also can learn by exploiting both labeled and unlabeled data.Experimental results show that MSWL can deal with weak-label and semi-supervised learning problems for multi-label learning.