0
$\begingroup$

I have around 5000-6000 observations of nearly 8-10 variables (of which 2 are discrete, categorical) and a single numerical target parameter. As per initial evaluation, random forest regression might be a good algorithm for the current case.

Is the current observations/variables count adequate for the proposed method? If other regression algorithms are recommended as things are described currently, kindly let me know.

$\endgroup$
  • $\begingroup$ RF works perfectly well even with the iris data, with only 150 samples in total for 3 classes (and actually less used for training, after train-test split or cross-validation). $\endgroup$ – desertnaut Apr 21 '20 at 12:10
0
$\begingroup$

The important is not the number of observation but the quality of this observations. If you have a look at toy datasets of sklearn they are way smaller than that.

Random forest is a good algorithm when there is small data since it is a bagging of decision trees with bootstrap. Each decision tree is feed with a sample of data with replacement, in this way even if the data is small there are bigger chances of making a good model.

In a high level, yes it seems a good way to go, but with out knowing more at the data is hard to tell.

I would suggest to give it a try with a Generalized Linear Model, a support vector machine and a gradient boosting. Since your data is small you will not need much computation time for it.

$\endgroup$

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.