Accuracy decay. The top panels show the plots for the average accuracy of Factor 3 alternatives ranging 1 to 10, and n−9 to n. Average predictive accuracy tends to fall as lower-ranking gene sets are used to constitute features (see text for details). The trend lines shown in the top panels are the ones minimizing the residual least squares. The bottom panel gives the accuracy boxplot for the batch experiments. 10 highest-ranking and the 10 lowest-ranking (respectively) gene sets are combined to produce a feature set for learning. Again, the dominance of the former in terms of accuracy is obvious. Each point in the top panels and each box plot in the bottom panel follows from 16,000 learning experiments.
Holec et al. BMC Bioinformatics 2012 13(Suppl 10):S15 doi:10.1186/1471-2105-13-S10-S15