Building a predictive model from data in high dimensions with application to analysis of microarray experiments

This work presents a comparative study of methods for building predictive models from data in high dimensionality spaces, i.e. where the number of features describing items to be classified is high as compared with the available number of items used to build the model and test its predictive performance. Application of such methods may be quite diverse, ranging from data analysis in life sciences (e.g., analysis of data from experiments generating thousands of feature-numbers per tested case, such a microarray or RT-PCR techniques), to analysis of monitoring data from a complex, highly reliable technical system, where the relationship is being sought between the monitoring data and the relatively infrequent occurrences of some event (such as a faulty or somewhat untypical state of the system). This latter case may be of special interest in early prediction of abnormal conditions in systems focused on dependability. The multidimensional data analysis challenges and generic methods are described in this paper using a very problem specific language of life sciences, namely classification of samples based on gene expression profiles obtained using DNA microarrays. We concentrate on feature selection methods (which in the context are gene selection methods). We also propose a method to evaluate performance of feature (gene) selection methods by looking at predictive power of classifiers based on selected features.