Research On Hierarchical Robust Multi-view Learning

As the development of the information technology and computer science,many prac-tical applications can produce multi-view data that has multiple feature representations Multi-view learning is a research direction that aims to improve the learning performance by simultaneously using the multiple views of multi-view data.In real applications,un-certainty is unavoidable in the process of collecting and transmitting multi-view data.For example,the white noise and outliers in the within-view level and different noise degrees in the between-view level.Focusing on the uncertainty of multi-view data,this disserta-tion studies on robust multi-view learning,that is,studying how to handle the performance degeneration caused by uncertainty.Studying on the robustness of multi-view learning can not only enrich the research theory of multi-view learning,but also promote the prac-tical application of multi-view learning methods.The main results and contributions of this dissertation are summarized as follows1.By analyzing the uncertainty of multi-view data,it is found that multi-view data has two levels of uncertainty.That is,the uncertainty within one view and the uncertain-ty across multiple views.Considering the hierarchical uncertainty of multi-view data,a hierarchical robust multi-view learning framework is proposed.Within this framework,the content of robust multi-view learning is discussed,and the methods to deal with the within-view and between-view uncertainty are presented.This framework provides guide-lines for designing robust multi-view methods in the rest of this dissertation2.Within the hierarchical robust multi-view learning framework,we respectively utilize the robust l2,1 loss and adaptive weighting to handle the within-view and between-view uncertainty,and propose the supervised and semi-supervised robust multi-view clas-sification methods using linear regression models with the l2,1 norm.In the supervised case,a label-adjusting term is introduced into the regression model to increase the discrim-inability of different classes and the cohesion within the same class,thereby enhancing the robustness against noise and outliers.Meanwhile,we explicitly promote the classifiers of different views to be diverse,so that more complementary information can be captured to improve the classification performance.In the semi-supervised case,sample weight is used to reduce the possible negative effects brought by the uncertainty of unlabeled samples.Adaptively optimized view weights balance the contributions of different views automatically,which makes the performance robust against the existence of low-quality views.Experimental results shows that the above two methods have advantages over the compared methods in handwritten digit recognition and scene recognition,respectively.Thus,they are more practical.3.Considering the case that multi-view data could contain random sparse corrup-tions,based on the latent representation assumption and subspace assumption,using the low-rank recovery technique of the hierarchical robust multi-view learning framework,the Multi-view Latent Row Space Pursuit(MLRSP)method is proposed to recover both the row space of the latent representation and the possible corruptions in multi-view da-ta.When there are no corruptions in the multi-view data,it can be proved theoretically that MLRSP can exactly recover the latent row space,which determines the clustering structure.The proposed method has good subspace recovery ability and clustering per-formance,and is applicable to multi-camera video surveillance for effective moving object detection.4.In the case that the multi-view data is incomplete,especially when there are both missing views and missing variables,we propose the incomplete multi-view learn-ing framework based on low-rank embedding.This framework contains several exist-ing incomplete multi-view learning methods as special cases,and can be used to adapt some full-view methods to be applicable to incomplete multi-view data.This not only improves the efficiency of dealing with multi-view data,but also establishes the relation-ship between traditional full-view learning and incomplete multi-view learning.Within this framework,based on the assumption that samples have approximate linear subspace structure,we propose an incomplete multi-view learning method with block diagonal rep-resentation,which increases the robustness against missing entries and improves the per-formance of incomplete multi-view learning.5.Recall that the original intention of multi-view learning is to obtain better learn-ing performance than single-view learning,thus,the reliability of multi-view learning relative to single-view learning is deemed to be the basic meaning of robust multi-view learning.For clustering,there is no label information available for judging whether multi-view clustering results outperform the single-view results.Within the hierarchical robust multi-view learning framework,using the late fusion strategy to deal with the between-view uncertainty,we propose the Reliable Multi-View Clustering(RMVC)method,such that better clustering performance can be obtained,compared with the given best single-view clustering.Specifically,given the benchmark single-view clustering results of each view and several candidate multi-view clustering results,RMVC maximizes the worst-case performance gain against the best single-view clustering.Under certain condition,the reliability of RMVC can be proved theoretically.This is the first study on the reliability of multi-view learning over single-view learning.