Abstract: A new feature dimensionahty reduction method called kernel covariance component analysis (KCCA) was put forward on the criterion that the transformed data best preserves the concept of the difference (denoted as D-vs-E) betwecn total densities and Renyi entropy of the input data space, induced from kernel covariance matrix. The generalized version of D-vs-E was also developed here. KCCA achieves its goal by projections onto a subset of D-vs-E preserving kernel principal component analysis (KPCA) axes and this subset does not generally need to correspond to the top eigenvalues of the corresponding kernel matrix, in contrast to KPCA. However, KCCA is rooted at the new concept of D-vs-E rather than Renyi entropy. Experimental results also show that KCCA is more robust to the choice of Gaussian kernel bandwidth when it is used in clustering.

Key words: KECA, KCCA, Clustering, Covariance matrix, Gaussian kernel bandwidth, Renyi entropy