Robustness of the Random Forest-based gene selection methods

Gene selection is an important aim of a microarray data analysis, as it can reveal information leading to a better understanding of the mechanisms of the investigated phenomenon. On the same time it is a very hard task due to the noisy nature of this data, becoming extreme if one wants to identify non-obviously relevant genes which subtle impact lies near the level of noise. To this end, it is often approached through machine learning; in particular with the Random Forest method which has several features crucial for this purpose. In this work, four state-of-art Random Forest-based feature selection methods are compared; the analysis is focused on the stability of selection, as it is a necessity for the significance of the results, yet it is often ignored in similar studies. The comparison of accuracy obtained by classifiers trained on a selected subsets of genes revealed that none of the investigated methods significantly improved the classification and majority of them did not degraded it either. Out of all analysed methods, the Boruta algorithm proved to find most possibly important genes on the same time achieving highest signal-to-noise ratio; those results were also fairly stable with respect to the used importance source. Although it was also the most computationally intensive method, its demand could be reduced to the level comparable with other analysed method by replacing the Random Forest importance with this produced by Random Ferns, a similar but simplified algorithm.