Study Of Multi-view Feature Learning Methods

Multi-view data can provide characterizations about a specific object from multiple viewpoints and usually contains more useful information for classification than the single view data.Therefore,in recent years,multi-view learning technique has been a research hotspot.There usually exists redundancy information in data of diffetent views.How to fully extract useful features from multi-view data and remove their redundancy is a key problem for multi-view learning technique.This dissertation carries out a systematic research for multi-view learning technique on this topic.The main research achievements are summarized as follows:(1)Two multi-view discriminant analysis approaches are proposed,including group recursive discriminant subspace learning(GRDSL)and uncorrelated locality sensitive multi-view discriminant analysis(ULSMDA).GRDSL fuses multi-view data on the data level,recursively decomposes the fused data set into multiple approximate sets and the corresponding difference sets,and learns a discriminant transformation from the difference image set in each recursion.GRDSL designs a terminating criterion and a projective vector selection rule.And it can guarantee the orthogonality between discriminant transformations theoretically.With the adaptive recursive learning process,GRDSL can learn sufficient useful features from the multi-view data.ULSMDA jointly learns multiple view-specific transformations,such that in the projected subspace,the nearby original samples with the same class label are close to each other while the nearby original samples with different labels are far apart.It considers the data consistency across views and designs an uncorrelated constraint to reduce the redundancy among transformations.ULSMDA can make full use of the local structure information of multi-view data for learning uncorrelated discriminant features.Experimental results on four datasets indicate the effectivenesses of these two approaches.(2)Three multi-view dictionary learning approaches are proposed,including uncorrelated multi-view discrimination dictionary learning(UMD2L),multi-view low-rank dictionary learning(MLDL)and multi-view low-rank shared structured dictionary learning(MLS2DL).By making dictionary atoms correspond to the class labels,UMD2 L jointly learns multiple structured dictionaries from multi-view data.It designs the uncorrelated constraint to reduce the redundancy among dictionaries of different views.From the aspects of enhancing discriminative power of dictionary and removing the redundancy,UMD2 L improves the ability of multi-view dictionary learning technique for learning useful features.MLDL introduces the low-rank learning technique into multi-view dictionary learning,and adopts the low-rank matrix recovery theory to address the multi-view dictionary learning problem when noise exists in data.It designs a structural incoherence constraint,and provides an efficient classification scheme for multi-view dictionary learning technique,which is based on collaborative representation.MLDL provides a solution for multi-view dictionary learning technique,such that the useful features can be effectively learned from the noise contaminated multi-view data.MLS2 DL focuses on extracting the shared information among multiple views,and proposes to learn a view-shared low-rank structured dictionary at the same time of learning multiple view-specific low-rank structured dictionaries.MLS2 DL provides a solution for multi-view dictionary learning technique,that is,making use of favorable correlation among multiple views while removing the redundancy among views.Experiments demonstrate that: as compared with representative multi-view subspace learning methods and multi-view dictionary learning methods,and the proposed GRDSL and ULSMDA,these three approaches can obtain better classification effects.(3)An semi-supervised multi-view discriminant analysis approache is proposed,i.e.,uncorrelated semi-supervised intra-view and inter-view manifold discriminant learning(USI2MD).It provides a semi-supervised intra-view and inter-view manifold discriminant learning scheme,which can utilize the local neighborhood structures of unlabeled samples and the intra-view and inter-view discriminant information of labeled samples to extract features from multi-view data.USI2 MD designs a semi-supervised uncorrelation constraint to reduce the adverse multi-view correlation of features in the semi-supervised scenario.USI2 MD can sufficiently utilize the intra-view and inter-view useful information to learn uncorrelated discriminant features under the semi-supervised setting.Experiments demonstrate the effectiveness of this approach as compared with representative semi-supervised multi-view subspace learning methods.