MultiK-MHKS: A Novel Multiple Kernel Learning Algorithm

摘要

With the newly-proposed Canonical Correlation Analysis (CCA) named NmCCA that is an alternative formulation of CCA for more than two views of the same phenomenon, we develop a new effective multiple kernel learning algorithm. First, we adopt the empirical kernels to map the input data into m different feature spaces corresponding to different kernels. Then through the incorporation of NmCCA in a learning algorithm, one single learning process based on the regularization learning is developed, where a special term called Inter-Function Similarity Loss RIFSL is introduced for the agreement of multi-view outputs. In implementation, we select the Modification of Ho-Kashyap algorithm with Squared approximation of the misclassification errors (MHKS) as the incorporated paradigm, and the experimental results on benchmark datasets demonstrate the feasibility and effectiveness of the proposed algorithm named MultiK-MHKS.