Feature request for sequential feature selector: random subsets at each step about mlxtend HOT 2 OPEN

Hi, Sergey,
thanks for the suggestion and I totally agree with you hear: SFS is darn expensive (O(n^2)) when it comes to high dimensional datasets. The random subsampling sounds like a neat idea, and it shouldn't be too hard to add it to the existing implementation. I am a bit busy at the moment, but I will add it to my todo list for when I am back from SciPy :)!

Btw. it reminds me a bit of Bergstra & Bengio's paper on Random hyperparameter search (Random Search for Hyper-Parameter Optimization; http://www.jmlr.org/papers/v13/bergstra12a.html). I.e., that 60 random points fall within the true maximum with 95% probability. Or if each draw has a 5% chance to be in the true-maximum interval, we have
1 - (1 - 0.05)^n > 0.95 and n >= 60.

from mlxtend.

Sounds great! I've been using this in my own mini-implementation of SFS and it seems to work well.

from mlxtend.