Abstract:

In order to accelerate the computation of kernel density estimation （KDE） and to reduce the complexity of KDE model, a fast KDE algorithm based on sparse Bayesian regression is proposed. The algorithm takes the jittered approximation of the distribution function as the input and obtains the very sparse representation of KDE. Experimental results indicate that, as compared with the conventional KDE algorithm, the proposed algorithm results in a much smoother density estimation when training with a small sample set, and it remarkably improves the space-time efficiency with a comparative computational precision and with a reduced model error in most cases. Moreover, the applications of independent component analysis and Gaussianization to the proposed algorithm allevi- ate the curse of dimensionality to some extent.

Key words: machine learning, kernel density estimation, Bayesian regression, ill-posed inverse problem, jittering regularization, Gaussianization