Kernel-based Informative Feature Extraction via Gradient Learning

We consider the problem of feature extraction for kernel machines. One of the key challenges in this problem is how to detect discriminative features while mapping features into kernel spaces. In this paper, we propose a novel strategy to quantify the importance of features. Firstly, we derive an informative energy model to quantification of feature difference. Secondly, we move the features in the same class closer and push away those belong to different classes according to the model and derivate its objective function. Finally, gradient learning is employed to maximize this function. Experimental results on real data sets have shown the efficient and effective in dealing with projection and classification.