A multiple kernel support vector machine (MK-SVM) scheme, consisting of feature selection, rule extraction and prediction modeling is proposed to improve the explanation capacity of SVM. In this scheme, we show that the feature selection problem can be translated into an ordinary multiple parameters learning problem. And a shrinkage approach: 1-norm based linear programming is proposed to obtain the sparse parameters and the corresponding selected features. We propose a novel rule extraction approach using the information provided by the separating hyperplane and support vectors to improve the generalization capacity and comprehensibility of rules and reduce the computational complexity.
Two public gene expression datasets: leukemia dataset and colon tumor dataset are used to demonstrate the performance of this approach. Using the small number of selected genes, MK-SVM achieves encouraging classification accuracy: more than 90%for both two datasets. Moreover, very simple rules with linguist labels are extracted. The rule sets have high diagnostic power because of their good classification performance.