Feature selection on microarray for supervised classification

Question

I'm studying the task of feature selection on biological microarray, thats is high dimensional dataset (thousands of features) with small number of data points (lees than one hundred). This feature selection serves to support supervised classification task.

What research or algorithm you can recommend me?

For now I learn about permutation-based algorithm, but I'm open to any advice. I need real research paper to understand how it works, I'm not looking for pre-packed programs.

EDIT1: I have read 2-3 article dealing of feature selection based on correlation between feature. The central idea of this article is to remove the correlated feature because bring redundant information. For me this is not be a good idea because some classifiers could exploit the dependence between freature (Like SVM or Decsion Tree). What you think about?

EDIT2: Is there a specific term to describe a High Dimensional Dataset with a few number of osservation like microarray? because I need it for improve my search on Google Scholar.

Answer 1

The use of a LASSO penalty term as Dadong Zhang suggests (+1) is a good place to start, I investigated this approach, but integrated the regularisation parameter out analytically using a reference prior, which worked fairly well for the datasets I investigated, the paper was published here:

G. C. Cawley and N. L. C. Talbot, Gene selection in cancer classification using sparse logistic regression with Bayesian regularisation, Bioinformatics, volume 22, number 19, pages 2348-2355, October 2006. (www)

MATLAB software here. The extension to multinomial classification was published here:

G. C. Cawley, N. L. C. Talbot and M. Girolami, Sparse Multinomial Logistic Regression via Bayesian L1 Regularisation, In Advances in Neural Information Processing Systems 19, B. Schölkopf, J. C. Platt and T. Hoffmann (eds), MIT Press, Cambridge MA USA, 2007.

Software here. While this approach is unlikely to be the best for every microarray dataset, it would make a good baseline method for comparison with others as there are no hyper-parameters to tune, and so is fully automatic.

Whatever you do, make sure you read this paper first:

Christophe Ambroise and Geoffrey J. McLachlan, "Selection bias in gene extraction on the basis of microarray gene-expression data", PNAS, vol. 99, no. 10, pp. 6562–6566, 2002. (www)

As a lot of research on this topic has been misguided due to incorrect performance evaluation.

Answer 2

I'd think the lasso variable selection algorithm may be feasible. Sine you don't provide a demo data, I recommended you a PPT document and the lasso website for your review. If you use R, the package glmnet have build in functions for that.