Model Instability in Microarray Gene Expression Class Prediction Studies

摘要

This work is devoted to the problem of building a sample classifier based on data from microarray gene expression experiments. Two specific issues related to this are tackled in this paper: (a) selection of parameters of a classification model to ensure best generalization power, and (b) variability of expected prediction error (EPE) for new data as a function of the model parameters. A method is presented for selection of model parameters minimizing the EPE in studies where the number of samples (n) is much smaller then the number of attributes (d). Due to very unstable behaviour of the EPE in the space of model parameters, it seems essential that microarray studies involve systematic search for the right model parameters, as shown in this work.