Sparsity Preserving Canonical Correlation Analysis

Canonical correlation analysis (CCA) acts as a well-known tool to analyze the underlying dependency between the observed samples in multiple views of data. Recently, a locality-preserving CCA, called LPCCA, has been developed to incorporate the neighborhood information into CCA. However, both CCA and LPCCA are unsupervised methods which do not take class label information into account. In this paper, we propose an alternative formulation for integrating both the neighborhood information and the discriminative information into CCA and derive a new method called Sparsity Preserving Canonical Correlation Analysis (SPCCA). In SPCCA, besides considering the correlation between two views from the same sample, the cross correlations between two views respectively from different within-class samples, which are automatically determined by performing sparse representation, are also used to achieve good performance. The experimental results on a series of data sets validate the effectiveness of the proposed method.