Study Of Supervised And Semi-supervised Multi-view Feature Learning Methods

Multi-view data can depict one object from multiple aspects.More useful information can usually be obtained from multi-view data than from single-view data of objects.Therefore,multi-view feature learning technique has received extensive attention from many researchers.There exist both complementarity and redundancy in multi-view data.How to fully exploit the discriminative features from multi-view data is the key issue when we apply multi-view feature learning technique in classification tasks.Among the huge amount of multi-view data,there are only a small number of labeled multi-view data while a large number of multi-view data are unlabeled multi-view data.How to utilize valuable labeled multi-view data and plentiful unlabeled multi-view data to effectively perform multi-view feature learning is an issue that is worth studying.A in-depth and systematic research for multi-view feature learning technique on the above topics is carried out in this dissertation.The main research achievements are summarized as follows:?1?For the problem that how to effectively utilize the labels of samples for promoting the extraction of discriminative features in multi-view feature learning,two supervised multi-view discriminant feature learning approaches are proposed,including multi-view discriminating uncorrelated projection analysis?MDUPA?and multi-view intact discriminant space learning?MIDSL?.MDUPA defines a supervised correlation,i.e.,discriminating correlation,which represents the adverse feature's correlation between different-class samples from different views.MDUPA performs discriminant analysis for the data of each view,and utilizes class label information in correlation analysis for further exploiting discriminative features.MIDSL learns intact feature representations which can fully describe the characteristics of objects by employing the multi-view data.And it employs Fisher discriminant criterion to use class label information for further improving the discriminability of the intact feature representations.Moreover,MIDSL also utilizes Cauchy loss to construct the feature learning model for enhancing the robustness of the model to outliers and noises.Experimental results on public datasets demonstrate that the two approaches can fully discover discriminative features with the help of class label information.?2?For the problem that how to effectively employ labeled and unlabeled samples for facilitating the extraction of discriminative features in multi-view feature learning,two semi-supervised multi-view discriminant feature learning approaches are proposed,including semi-supervised multi-view correlation discriminant analysis?SMCDA?and semi-supervised multiple kernel intact discriminant space learning?SMKIDSL?.SMCDA aims to fully exploit discriminative features by conducting intra-view and inter-view discriminant correlation analysis for the multi-view data of labeled and unlabeled training samples.SMCDA transforms the matrix-variable based nonconvex objective function into a convex quadratic programming problem with one real variable,and can achieve a global optimal solution without iterative calculation.SMKIDSL can learn intact feature representations for sample objects by collaborately utilizing the multi-view data of labeled and unlabeled training samples.And SMKIDSL employs multiple kernel learning strategy,discriminant correlation analysis and l_2,1-norm regularized regression strategy for further enhancing the discriminability of the intact feature representations.Besides,SMKIDSL designs a multi-view collaboration learning scheme for semi-supervised multi-view learning to make different views contribute optimal useful information.Experimental results on public datasets indicate that SMCDA and SMKIDSL approaches can effectively exploit discriminative features from the multi-view data of labeled and unlabeled training samples.?3?For the problem that how to effectively exploit discriminative features from generalized multi-view data,a supervised deep multi-view discriminant feature learning approach is proposed,i.e.,robust supervised deep discrete hashing?RSDDH?approach.RSDDH can learn deep learning features which are more compatible with cross-view retrieval tasks by utilizing deep learning technique.The l_2,1-norm based feature selection strategy is used to select superior features for generating binary hash codes.Furthermore,RSDDH employs inter-view and intra-view consistency preservation strategies to reduce the gap of different views and enhance the discriminant power of binary hash codes.RSDDH also employs our proposed singular value decomposition based discrete hashing algorithm to effectively solve the unknown discrete hash variable.Experimental results on public datasets show that RSDDH can outperform several state-of-the-art shallow cross-view hashing retrieval methods and deep cross-view hashing retrieval methods.And the effectiveness of RSDDH is also confirmed by the experimental results.