Abstract: Considering the unbalanced distribution of the training samples and the multiformity of the features. A multiple kernel SVM based on K-means cluster algorithm was proposed. Firstly, training samples was clustered into K classes, different penalty factors were used for each class in order to balance the contributions of each class. Secondly, the multiple kernel support vector machine was proposed for diversity of the features. The stabilized training sample was obtained via active feedback learning. The result show that the detection rate can be improved at least 2 percent by the proposed method, compared with the single kernel SVM and the multiple kernel SVM.

Key words: K-means cluster, Multiple kernel SVM, Microcalcification, Active feedback learning