Regressionbased approaches

Linear models, generalized linear models, generalized additive models and the associated variable selection strategies provide standard tools for selecting useful subsets of genes and developing probabilistic classifiers. A limitation of these techniques is that they cannot generally handle more genes than there are samples. This can be circumvented using forward selection approaches that progressively add genes to the classifier. Recent, more accurate approaches are based on the so-called stochastic search methods (61), that generate a sample of plausible subsets of explanatory variables. The selected subsets are then subjected to additional scrutiny to determine the most appropriate classification algorithm. A combination of stochastic search with principal component analysis and other orthog-onalization techniques has proven effective in high-dimensional problems (62, 63), and has recently been employed in microarray data analysis (23).

0 0

Post a comment