2
$\begingroup$

So basically I have a large set of features corresponding to a metric - like many ML problems.

What I want to know is: can we correlate the variance of metric with the variation in each feature.

ex:

I have features x, y, z that produce an output say 10. When I vary x, no matter how much I vary it, the output stays relatively close to 10. However, when I vary y the output is heavily influenced.

Is there a good technique to be able to assign a value correlating x and/or y to the metric?

I'm mostly looking for direction here.. i.e. techniques or relevant papers. In my experience I haven't really come across this problem. I don't have a good solution in my toolbelt.

Thanks!

$\endgroup$
1
$\begingroup$

If the variation in values of feature then should not you remove the column? Your model will not learn from it if the variance is low.

you can Q-Q plots when you vary the features to check how close are the 2 datasets comparing their distribution

| improve this answer | |
$\endgroup$
  • $\begingroup$ I'm not actually building a model. The entire purpose is actually exactly what you presumed. That is: to see which features can be removed (due to no variance) and which should be tested more (due to high variance). Thanks, I'll look into Q-Q plots. $\endgroup$ – uMdRupert Dec 20 '19 at 3:40

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.