2013 年 5 巻 p. 49-52
In this letter, we propose a new method of multiple kernel learning (MKL) that utilizes an adaptively weighted regularization. The proposed method controls strength of penalty for each kernel depending on its importance so that important components are amplified and unimportant components are diminished. To show the effectiveness of the proposed method, a theoretical justification is provided based on the recently developed unifying framework for the learning rate of MKL. Numerical experiments are carried out to support the usefulness of the proposed method.