Sparse Generalized Canonical Correlation Analysis (DSGCCA)

Multi-view learning (MVL) is a strategy for fusing multi-view data, which has better generalization performance than single-view learning algorithms. Canonical correlation analysis (CCA) is a representative multi-view subspace learning approach, which plays an important role in MVL classification and information retrieval. Traditional CCA can only be used to calculate the correlation of two views, and the learned features are usually dense. Moreover, it is unsupervised, and hence wastes label information in supervised learning. To overcome these limitations, this paper proposes discriminative sparse generalized CCA (DSGCCA), which integrates generalized CCA to handle more than two views, and supervised discriminative sparse principal component analysis to make use of the label information. DSGCCA can handle small multi-view datasets with high feature dimensionality and any number of views. Experiments on four classification datasets demonstrated that DSGCCA outperformed several other representative CCA-based MVL approaches.