2
$\begingroup$

So there's a domain of interest where the machine learning models are all specific to one entity. Let's call it a building. So there's a model made for every building. The literature in the domain all revolves around studies showing how they tailor a NN or other ML model to make a prediction for the building in question, be it an SVM, NN, or linear regression.

Now assume there are many buildings with each one having a model and each building having the same set of features. Running models trained on one building (say A) on another building (say B) sometimes produces a better result that the same type of model that was trained on the original building (B).

Is this an example of transfer learning?

What would this be called when training models on many differing datasets, all with the same features but vastly different due to geographic changes, then running the trained models against all of the different buildings? For example, 1000 models from 1000 different regions, one model for each region. Then run each of the 1000 models against the 1000 different regions (1000 x 1000 = 1,000,000 predictions). Some models will best predict their own region, while others will best predict regions that were unseen during training.

Further, which training datasets helped a model to perform better, and why?

Perhaps this is domain adaptation? According to this article (en.wikipedia.org/wiki/Domain_adaptation): "In domain adaptation, the source and target domains all have the same feature space (but different distributions); in contrast, transfer learning includes cases where the target domain's feature space is different from the source feature space or spaces." Elsewhere it's noted DA is a subset of TL. So if we can categorize the e.g. 1000 datasets based on their distribution, then could this be a start to domain adaptation and transfer learning?

Note: This question was previously asked on DataScience but moved here due to lack of feedback.

$\endgroup$

1 Answer 1

3
$\begingroup$

Typically, in transfer learning, we have one "building" in your terminology that we're particularly interested in, called the Target Domain, that we don't have much information on. And then, there are one or more other buildings that we have much more information on (the Source Domains). The objective is to transfer information from these well studied buildings to the new one. The case where we have a bunch of buildings we're interested in, and we have more or less the same amount of data on, perhaps can be technically classified as Transfer Learning, but doesn't capture its spirit.

This might be a job for hierarchical models instead.

$\endgroup$

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.