Unsupervised discriminant canonical correlation analysis based on spectral clustering

Canonical correlation analysis (CCA) has been widely applied to information fusion. However, it only considers the correlated information between the paired data and ignores the correlated information between the samples in the same class. Furthermore, class information is helpful for CCA to extract the discriminant feature, but there is no class information available in application of clustering. Thus, it is difficult to utilize the correlated information between the samples in the same class. In order to utilize this correlated information, we propose a method named Unsupervised Discriminant Canonical Correlation Analysis based on Spectral Clustering (UDCCASC). Class membership of the samples is calculated using the normalized spectral clustering, while the mappings for feature fusion are computed by using the generalized eigenvalue method. These two algorithms are executed alternately before the desired result is obtained. Two extensions of UDCCASC are proposed also to deal with multi-view data and nonlinear data. The experimental results on MFD dataset, ORL dataset, MSRC-v1 dataset show that our methods outperform traditional CCA and part of state-of-art methods for feature fusion.