Aiding classification of gene expression data with feature selection: A comparative study

This paper presents an application of supervised machine learning approaches to the classification of the yeast S. cerevisiae gene expression data. Established feature selection techniques based on information gain ranking and principal component analysis are, for the first time, applied to this data set to support learning and classification. Different classifiers are implemented to investigate the impact of combining feature selection and classification methods. Learning classifiers implemented include K-Nearest Neighbours (KNN), Naive Bayes and Decision Trees. Results of comparative studies are provided, demonstrating that effective feature selection is essential to the development of classifiers intended for use in high-dimension domains. In particular, amongst a large corpus of systematic experiments carried out, best classification performance is achieved using a subset of features chosen via information gain ranking for KNN and Naive Bayes classifiers. Naive Bayes may also perform accurately with a relatively small set of linearly transformed principal features in classifying this difficult data set. This research also shows that feature selection helps increase computational efficiency while improving classification accuracy.