Comparison of rank-based vs. score-based aggregation for ensemble gene selection

Gene selection is an essential step in much bioinformatics research in order to handle the thousands or tens of thousands of gene expression levels generated by gene microarrays. It is especially important that this gene selection is robust and will produce consistent results even in the face of changes to the dataset. Ensemble gene selection can help improve robustness, by combining gene rankings from multiple gene selection techniques into a single gene subset. Typically this is performed by performing multiple runs of feature (gene) selection, finding each gene's rank within the different runs, and aggregating these ranks into a final ranked list. However, another option exists: instead of performing the ranking on each list and then aggregating, the raw scores produced by the gene ranking algorithms (which would normally be compared to generate a ranking) are aggregated directly, and these aggregate scores are used to create a final ranking. This potentially results in a different final ranking, since adjacent genes (e.g., those with no genes in between them) which are particularly close to or far from one another will be treated as such. Also, score aggregation can help reduce computation time due to the ranking step only taking place once, rather than separately for each list being aggregated. In this experiment, we use eleven DNA microarray datasets and nine univariate feature selection techniques, along with twelve feature subset sizes, to demonstrate these two approaches on a commonly used aggregation technique: mean aggregation. The results show that for seven of the nine feature selection techniques, we see strong similarity between the two approaches, but the feature subsets are not identical. However, two of the techniques do show high levels of diversity between the two approaches. This allows us to state that further research is required in order to determine the abilities of the two approaches.