A Bayesian Approach to Improving Decision Making in Support Vector Machine and its Application in Bioinformatics

Conventional support vector machine (SVM) utilizes the sign function to classify test data into different classes, which has demonstrated some limitations that hinder its performance. This paper explores the feasibility of using Bayesian statistics to support decision making in the SVM and demonstrated its application in Bioinformatics. The proposed methodology was tested on two real biological problems: identification of photoreceptor-enriched genes and classification of dilated cardiomyopathy patients based on gene expression data. The results attained indicated that by incorporating the Bayesian statistic into SVM decision making process, a significant improvement was achieved (p < 0.005). The proposed methodology not only can improve the overall prediction performance but also can make the classification with the SVM less sensitive to the selection of input parameters. In particular, the approach can significantly improve the sensitivity to the minority class when using the SVM-based model to deal with imbalanced data.