Research On Multi-view Feature Extraction Techniques Based On Canonical Correlation Analysis

Learning from multi-view data has attracted much attention in recent years. Most existing multi-view based methods train learners independently and iteratively from different views. The correlationinformation can not be utilized in those methods. Canonical correlation analysis (CCA) is a standardmethod to reveal linear relationships between two views in statistics. In machine learning, CCA is oftenused to extract features that can relate different views. However, classical CCA is an unsupervisedmethod and does not take class label information into account. Classifiers trained on features extractedby CCA may have low accuracies and weak generalization abilities. In this thesis, we attempt to extendthe classical CCA model to increase its generalization ability.Firstly, we introduce the idea of cross correlation into CCA and propose a new method, called Ran-dom Correlation Analysis (RCA). In the method, both the correlations between different views from thesame point and the cross correlations between different views from the same classes are analyzed. So,by taking the advantages of CCA and cross correlation, RCA can not only extract related features fromtwo views, but also preserve as much discriminant information as possible. Experiments on handwrittendigits data set and several faces databases show the effectiveness of the proposed method.Secondly, we apply the ensemble paradigm on RCA to further improve its generalization ability.A novel ensemble method is proposed, called Random Correlation Ensemble algorithm (RCE). Onekey for ensemble learning is to train an ensemble of both accurate and diverse component classifiers.In RCE, the component classifiers are trained based on related features extracted from different viewsusing RCA, so the correlation information can be utilized. Furthermore, due to the randomness ofRCA, multiple run of RCA can produce diverse features, so the component classifiers in RCE can beboth accurate and diverse. Ensemble of them may help increasing the generalization ability of singleclassifier. Extensive experimental results on several multi-view data sets validate the effectiveness ofRCE.