Abstract
Interaction detection in genetic studies and many other fields is a popular research topic. Detecting interactions especially high-order interactions in genome-wide data remains a major challenge due to high dimensionality. We developed a random forest based algorithm, high order interaction hunting (HIH) to efficiently discover interaction signals. The random forest framework allows our method to deal with different types of outcomes including categorical, continuous and survival outcomes. In our method, information from pairwise minimal depth (PMD) matrix is used to select potential interactive features. We proposed and evaluated two variable importance measures, pairwise minimal depth variable importance (PMD VIMP) and normalized high order interaction variable importance (nmHIVIMP). The PMD VIMP serves as a valid and fast filter for potential interaction candidates, and nmHIVIMP is used for ranking the high-order interactions. Our findings through various simulation studies revealed that HIH exhibited good and consistent performance in ranking interactions compared to other interaction detection methods.