A Bayesian Multiple Kernel Learning Method Based on Relevant Vector Machine

摘要

Kernel methods have been widely used in large amounts of learning tasks. A good kernel function incorporating application background knowledge or induced from historic data may affect positively to learning machines. Multiple Kernel Learning (MKL) adopts a linear weighted combination mechanism to learn an application specific kernel function. We propose a novel MKL method based on Bayesian theory, which finds a weight vector to combine a set of kernel functions achieving maximal posterior probability given a training data set. The method adopts sparse Bayesian learning to derive a weight vector for the learned kernel through a relevant vector machine like method. The proposed method works effectively with various basic kernel function classes. We report an application of the proposed method on music informatics to build a classifier that can tell whether a paragraph of some music is classical or popular. The proposed method achieves the best result against two current state-of-the-art methods, which shows the effectiveness of our method.