Hi @elenaSP ,
Here's my initial thought: DataRobot heavily uses tree based ensemble methods (Random forests, gradient boosted trees, etc.). These trees will split on variables to maximize information gain. If there are low frequency variables that cleanly separate different different classes, I suspect the tree based methods will preferentially split on these variables regardless of their frequency.
I may need more context though: is this a regression or classification problem? And what is the use case?