Local multiple kernel learning method could learn a specific combination kernel function for various samples according to the local space characteristics

therefore it has better discriminant ability. In this paper

we propose a local variable sparsity based multiple kernel learning method. In our method

the samples are divided into a few groups with a soft grouping method and the sparsity of kernel weights in various local spaces is determined by the similarity of kernels. We use an alternative optimization method to solve this problem. The experiment on synthetic dataset indicates that our method has a strong advantage in discriminative feature learning and against noise. Finally we apply our method into image scene classification and the accuracy is improved obviously.