SF-KCCA: Sample Factoring Induced Kernel Canonical Correlation Analysis

The Canonical Correlation analysis (CCA), such as linear CCA and Kernel Canonical Correlation Analysis (KCCA) are efficient methods for dimensionality reduction (DR). In this paper, a method of sample factoring induced KCCA is proposed. Different from traditional KCCA method, sample factors are introduced to impose penalties on the sample spaces to suppress the effect of corrupt data samples. By using a sample factoring strategies: cosine similarity metrics, the relationships between data samples and the principal projections are iteratively learned in order to obtain better correlation projections. By this way, the authentic and corrupt data samples can be discriminated and the impact of the corrupt data samples can be suppressed. Extensive experiments conducted on face image datasets, such as Yale, AR, show our approach has better classification and DR performance than that of linear CCA and KCCA, especially in noisy datasets.