Sparse kernel entropy component analysis for dimensionality reduction of biomedical data

Dimensionality reduction is ubiquitous in biomedical applications. A newly proposed spectral dimensionality reduction method, named kernel entropy component analysis (KECA), can reveal the structure related to Renyi entropy of an input space data set. However, each principal component in the Hilbert space depends on all training samples in KECA, causing degraded performance. To overcome this drawback, a sparse KECA (SKECA) algorithm based on a recursive divide-and-conquer (DC) method is proposed in this work. The original large and complex problem of KECA is decomposed into a series of small and simple sub-problems, and then they are solved recursively. The performance of SKECA is evaluated on four biomedical datasets, and compared with KECA, principal component analysis (PCA), kernel PCA (KPCA), sparse PCA and sparse KPCA. Experimental results indicate that the SKECA outperforms conventional dimensionality reduction algorithms, even for high order dimensional features. It suggests that SKECA is potentially applicable to biomedical data processing.