Using Supervised Complexity Measures in the Analysis of Cancer Gene Expression Data Sets

Supervised Machine Learning methods have been successfully applied for performing gene expression based cancer diagnosis. Characteristics intrinsic to cancer gene expression data sets, such as high dimensionality, low number of samples and presence of noise makes the classification task very difficult. Furthermore, limitations in the classifier performance may often be attributed to characteristics intrinsic to a particular data set. This paper presents an analysis of gene expression data sets for cancer diagnosis using classification complexity measures. Such measures consider data geometry, distribution and linear separability as indications of complexity of the classification task. The results obtained indicate that the cancer data sets investigated are formed by mostly linearly separable non-overlapping classes, supporting the good predictive performance of robust linear classifiers, such as SVMs, on the given data sets. Furthermore, we found two complexity indices, which were good indicators for the difficulty of gene expression based cancer diagnosis.