Do little interactions get lost in dark random forests?

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in genomics

Background:

• Random forests are frequently cited for their ability to identify interaction effects.
• The distinction between capturing and detecting interaction effects in random forests remains ambiguous.
• This study addresses the capability of random forest variable importance measures to discern gene-gene interactions.

Purpose of the Study:

• To investigate if random forest variable importance measures can capture or detect gene-gene interactions.
• To differentiate between capturing (identifying a variable involved in an interaction) and detecting (identifying the interaction effect itself).

Main Methods:

• Extensive simulation studies were conducted.
• Evaluated various random forest variable importance measures, including Gini importance, permutation importance, and pairwise importance methods.
• Assessed the ability of these measures to capture and detect gene-gene interactions under different scenarios.

Main Results:

• Gini importance captured interactions in most simulations but was often masked by marginal effects.
• Permutation importance showed a lower proportion of captured interactions.
• Pairwise measures performed similarly, with joint variable importance showing a slight advantage.
• The overall detection of interactions was low, with models containing only marginal effects often showing higher detection rates than those with interaction effects.

Conclusions:

• Random forests can capture gene-gene interactions, but current variable importance measures fail to detect them as distinct interactions.
• Marginal effects frequently mask interaction effects, preventing differentiation.
• Researchers should exercise caution when attributing the discovery of interactions solely to random forests.