1
$\begingroup$

I am asking if there already exist approaches and researches on the following topic.

Imagine there are 10 stores and in 3 stores labeled training data was available, so I built 3 classification models based on those datasets (e.g. to classify my customers into A, B, C customers).

All stores basically have similar characteristics, but Model A will work best with Store A, Model B best with Store B, etc. because of geographical or cultural differences.

If training data is not available for my other 7 stores, is there any approach to find out if I should use Model A, B or C for my Store D data?

I assume the data distribution of Store D must be compared with those of A, B and C to estimate which model will work best. Any ideas or recommendations of existing papers about that topic?

Thanks!

$\endgroup$
3
  • $\begingroup$ Isn't a single model that accounts for the differences between stores possible? From your short description it sounds like it is not about individual stores but rather the differences between them. $\endgroup$
    – Tim
    Commented Feb 7, 2015 at 18:10
  • $\begingroup$ I agree with @Tim, a single model should be able to handle this $\endgroup$
    – bdeonovic
    Commented Feb 7, 2015 at 18:14
  • $\begingroup$ My question is, given that store A is in USA, store B in Australia and store C in Russia and each model works best in its country because there are country-specific characteristics in the data. Now store D opens in Italy and all you have is unlabeled data, so you can't train a new model for store D and you can't validate your existing models because you don't know the labels. Is there an approach to decide which model (A, B or C) works best for store D by comparing the data distributions from store D to stores A, B and C? $\endgroup$
    – HansHupe
    Commented Feb 8, 2015 at 10:11

1 Answer 1

1
$\begingroup$

Please check the paper "A survey on Transfer Learning" by SJ Pan. Your problem is related to domain adaptation and I think the survey will help you to find related methods to your problem. (I see that the reply is a bit too late though :))

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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