Bayesian Learning with Local Support Vector Machines for Cancer Classification with Gene Expression Data

International audienceThis paper describes a novel method for improving classification of support vector machines (SVM) with recursive feature selection (SVM-RFE) when applied to cancer classification with gene expression data. The method employs pairs of support vectors of a linear SVM- RFE classifier for generating a sequence of new SVM classifiers, called local support classifiers. This sequence is used in two Bayesian learning techniques: as ensemble of classifiers in Optimal Bayes, and as attributes in Naive Bayes. The resulting classifiers are applied to four publically available gene expression datasets from leukemia, ovarian, lymphoma, and colon cancer data, respectively. The results indicate that the proposed approach improves significantly the predictive performance of the baseline SVM classifier, its stability and robustness, with satisfactory results on all datasets. In particular, perfect classification is achieved on the leukemia and ovarian cancer datasets