I have a dataset with 600 variables and 5000 observations which I want to use for a classification problem (sick/not sick) but I'm worried about multicollinearity so Im trying to figure out a good way to do some feature selection before running my neural network. Things I've considered are PCA and stacked auto-encoders. I'm kind of a n00b when it comes to ML and Deep learning though so I would love some advice on what methods I could/should apply?

  • 1
    $\begingroup$ "Multicollinearity" is not really what you have to worry about. The issue is that with more variables and less points, the chances of finding random connections (as a result of say, measurment noise) between your input and output is higher, i.e. the model will not be that meaningful. $\endgroup$ Feb 25, 2018 at 18:41
  • $\begingroup$ Interesting, two question where can I read more about this phenomena? And what precautions can I take to mitigate the risk of finding random connections? $\endgroup$ Feb 26, 2018 at 9:37