Research And Application Of Multi-view Multi-label Learning

As one of the most potential technologies in machine learning,multi-view multi-label learning has been attaching great attention recently and was widely applied in many fields,such as heterogeneous media data analysis and bioinformatics.The main idea of this technology is to depict the objects represented with multiple feature sets(i.e.views)from different perspectives.In multi-view multi-label learning,each sample is characterized by heterogeneous feature views,and is associated with several non-exclusive semantic labels.Traditional multi-view multi-label learning algorithms aim to obtain the labels of such objects by integrating their multiple representations.Most of these multi-view data fusion strategies are supervised,which need large number of labeled samples as inputs for model training.However,in the real-world,labeling multi-view multi-label objects is very expensive,which requires lots of manpower and money.Besides,the information communication of the same object among multiple views may greatly affect its learning performance.Furthermore,to better solve the semantic ambiguity problem of multiple subsamples of a sample across different views,a more complex framework named multi-view multi-instance multi-label learning paradigm is spurred.The objective of this framework is to acquire the label set of a sample(bag)by learning the features of its subsamples(instances)in multiple views.Conventional multi-view multi-instance multi-label learning approaches are supervised,which ignore the performance influence of multi-type relations between bags,instances and labels.In addition,all of these algorithms are difficult to generalize to a more complicated weakly-supervised scenario where the bag mapping information across multiple views are partially unknown and the labels of bags are incomplete.To handle the above challenging issues,this thesis combines multi-view multi-labellearning,co-training,matrix factorization,weakly-supervised learning and active learning for the above mentioned multi-view complicated objects modeling and analysis.This thesis is mainly made of the following contents:1.A novel algorithm named multi-label co-training(MLCT)is presented in this thesis,which aims to better address the label communication problem among multiple feature views of complicated multi-view multi-label objects by using a large number of unlabeled samples.MLCT first leverages information concerning the co-occurrence of pairwise labels to handle the class-imbalance challenge during the combination of co-training and multi-label frameworks.Then a new predictive reliability measure is introduced to select samples and corresponding labels with the most confidence for communication among various co-training classifiers.Experimental studies on multiple multi-view multi-label benchmark datasets evualate the effectiveness of our model.2.In order to explore the performance influences of multi-type relations between various typed objects,(i.e.bags,instances and labels)and the difference of intrinsic structures of these objects in multi-view multi-instance multi-label learning,this thesis introduces a new algorithm based on collaborative matrix factorization(M3Lcmf).The proposed M3 Lcmf first uses a heterogeneous network composed of nodes of bags,instances,and labels,to encode different types of relations by exploiting multiple relational data matrices.It then collaboratively factorizes these relations to explore the latent relations of bags,instances and labels,and acquires bag-level(or instance-level)prediction by selectively merging those multi-type relation data matrices.Experimental results on multiple multi-instance multi-label benchmark datasets show the effectiveness of M3 Lcmf on predictions both at instance-level and bag-level.3.This thesis extends existing multi-view multi-instance multi-label learning framework(WSM3L)to a more flexible and open wealy-supervised scenario,to address the challenges of weakly-paired bags across multiple views and incomplete bag labels.WSM3 L firstly adapts multimodal dictionary learning to learn a shared dictionary across multiple views and individual encoding vectors of bags for each view.It then jointly uses the label similarity and encoded feature similarity of bags to match bags ac-ross views.WSM3 L further introduces a dispatch and aggregation term for label replenishment of bags by using their neighbourhoods' label information and instance predictions.Experimental results on multiple benchmark and real datasets have validated the effectiveness of the proposed WSM3 L model.4.To reduce the labeling costs of complicated multi-view multi-instance multi-label objects,a novel model named Multi-view Multi-instance Multi-label Active Learning(M3AL)is introduced to address this problem by combining active learning and multi-view multi-instance multi-label learning.The presented M3 AL firstly adapts the multi-view self-representation learning to evacuate the shared and individual information of bags across/within views.Next,it develops a new query strategy that can simultaneously leverage the shared and individual information,and the diverse instance distribution of bags across multiple views to select the most informative bag-label pairs for classifier performance improvement,and simultaneously reduce the annotation price of unlabeled complex objects.Experimental studies on multiple public datasets show that using the presented M3 AL can indeed and effectively save the labeling costs of unlabeled multi-view multi-instance multi-label objects,and simultaneously improve their classification accuracy.