Advantages of MRMR

norita

I chose MRMR selection for the first selection and then forward selection as further selection for my model. Now I want to justify my choice.

So far, I would say MRMR is not expensive in compution time and therefor good for a first selection (150 variables). And it is better as other filter methods (also computional inexpensive methods) as e.g select by weight by correlation /information gain because it adresses the interferance of variables amongst each other.

What do you think of my reasoning? Is there an advantage of the MRMR algorithm I missed? Are there any recommondations where it is likely to use?

MartinLiebig

Hi @norita ,

for me there is a tradeoff between the runtime/complexity of the algorithm. I quickly drew how i think the power/rumtime tradeoff looks like

So MRMR is actually a nice tradeoff here. But you are right it is not perfect. I also often use it as a pre-filter to go down to a handable sample size first.

Cheers,

Martin