Figure 8. Interaction simulation comparison of feature ranking methods for different numbers of null genes.

Comparison of optimized-k Relief-F, standard Relief-F using k = 10 nearest neighbors, edgeR, and tuned Random Forest for detecting an interaction between two genes amongst various numbers of null genes. The panels are sorted in order of increasing negative binomial parameter θ. Each point is the average of the worst ranking gene of the two simulated interacting genes across n = 100 replicate simulations and then divided by the total number of simulated genes. The number of null or background genes increases from 100 to 12,800 total genes, plotted on the log2 scale. Each simulation contains one pure interaction (no main effects) XOR model between two negative binomial genes.