Abstract: Kernel-based attribute reduction methods have shown great advantages for removing redundant information and adjusting nonlinear structure of input data. But in real applications, it is difficult for kernel-based attribute reduction methods to select optimal parameters. To do this, a new feature extraction method based on the optimization learning of adaptive kernel function was proposed in this paper. By use of improved Fisher kernel matrix measure criterion, an optimization framework of adaptive kernel function was established to deal with multi-classification task. Combining with optimization results, eigenvectors which make greater contribution to Renyi entropy estimation of input data were selected. New features were extracted based on selected eigenvectors in KECA feature subspace. Experimental results show that presented method can not only enhance the classification accuracy, but also restrain noise interference.