KERNEL REGRESSION SYSTEM, METHOD, AND PROGRAM

摘要

PROBLEM TO BE SOLVED: To provide a multiple kernel function learning technique that can handle a problem of a practical size with the reasonable amount of computation.SOLUTION: Similarity matrixes are created for each piece of data corresponding to different kernels in training data, and individual graph Laplacians are formed from the similarity matrixes. The whole graph Laplacian is defined as a linear combination based on coupling constants of the individual graph Laplacians, and normal distribution is assumed for observation variables and latent variables attached thereto; while gamma distribution is assumed for the coupling constants. The distribution of the observation variables and the coupling constants are thereby obtained with the reasonable amount of computation, based on a variation Bayesian method. When the distribution of the observation variables and the coupling constants are obtained, predictive distribution to optional input data can be obtained by Laplace approximation.