Gene expression data are increasingly used in the area of cancer classification, for tissue annotation, cancer diagnosis and disease outcome
prediction. Several classic machine learning approaches have been proposed for these supervised classification problems. However, features of gene
expression data complicate the task. The major constraints are the low number of samples available, the high number of genes monitored, and of course
the high level of noise present in the data.
Here we propose to use a simple statistical method to detect outlying measurements. An outlier is a gene expression measurement which differs
surprisingly from the other measurements obtained for the same gene on other samples. We integrate this procedure in a Naive Bayes classifier and
show experimentally on several data sets that the detection and deletion of outlying measurements can substantially reduce the misclassification rate.