Find in Library

The objective function of canonical correlation analysis (CCA) is equivalent to minimising an L(2)-norm distance of the paired data. Owing to the characteristic of L(2)-norm, CCA is highly sensitive to noise and irrelevant features. To alleviate such problem, this study incorporates robust feature extraction and group sparse feature selection into the framework of CCA, and proposes a feature fusion method named robust and sparse CCA (RSCCA). In RSCCA, L(2,p)-norm is adopted as the distance measurement of paired data, which can alleviate the effect of noise and irrelevant features and achieve robust performance. The experimental results show that our method outperforms CCA and its variants for feature fusion.