Abstract
The problem of variable selection from a large number of candidate predictors has recently been addressed in the machine learning of bioinformatics field. This is due to advances in high-throughput micro array techniques such as Affymetrix Gene Chips, and Illumina micro arrays that allow for studying thousands of genes in a single experiment. However, the resultant data from such genomic tools suffers from an p >> n problem, where the number of genes (p) to be examined is much larger than the number of samples (n). In such a model selection, the learning is considered hard, and the goal is to achieve accurate predictions from the inferred models alongside with their interpretability. Towards this goal, this work will experiment with feature selection methods and show how to improve the choice of the tuning parameter (s) in the lasso estimate feature selection method by adding an extra layer of filter feature selection methods to the lasso estimate path when learn from p n gene expression datasets. The results show that when the lasso estimate is ensemble with filter feature selection methods, the prediction accuracy for the chosen predictors for each targeted variable has improved.