Fitting a dataset into another dataset

Question

I had heard that there is a body of literature devoted to the following problem:

You have a dataset and you produced a good predictive model for it. Now you have a different dataset, derived from different instruments, or different data sources, but similar enough that you can hope to scale one to the other so that you can use the predictive model on it.

Alternately, you do not have enough data to produce a model and hope that combining different datasets by scaling one to the other will help you reach the quantity of data you need.

First off what is this problem called so that I can search it more effectively? And also does anyone know of a good recent survey of techniques for this, either in book or paper form? I currently have access to most academic journals so those links work for me as well.

Say I have dataset A that I have a model for, dataset B occupies the same database schema but is from a difference source with different factors that are not included in my feature set.

Initially my intuition was to construct a QQplot and fit lines (or curves) to features that I thought should be similar. If the difference in the way feature 1 from A increases is similar to the way feature 1 from B increases but with a constant factor then fitting a line can reveal this factor. If the difference was exponential or logarithmic then I could scale using a fitted function. In this way I could constrain the way one variable increased to fit how another variable increased.

However this is just my intuition. I can certainly test it for overfitting but when I had heard that there was a lot of literature devoted to this subject then it seemed as though I should learn a few ways in which I could question my assumptions. It would probably be good for me to review the literature.

Does anyone know what the tag for this literature might be?

Answer 1

My original source finally got back to me. The topic I want is called transfer learning. I'll share any survey resources I find here.

EDIT: As I promised, if I find a literature survey I'll include it here. I found this one.

---

Reference

PAN, S. J.; YANG, Q. A Survey on Transfer Learning. IEEE Transactions on Knowledge and Data Engineering, v. 22, n. 10, p. 1345–1359, out. 2010.

Answer 2

Since you have a different dataset, which is derived from a different data source, but there are some commonalities among them, you can try to adapt the model that you have produced for the first dataset (where we assume that we have a sufficient number of data available) to solve the new problem of the second dataset. In the literature this topic of machine learning is called Transfer Learning or in some cases you can find it as Domain Adaptation. The most important in this kind of adaptation/transfer is the kind of difference among the two datasets. They can have different marginal distributions or different feature spaces. In the latter case, which is called heterogeneous Transfer Learning, you need to find a way to align the two different feature spaces.

If we assume that the source domain is the old dataset and the target domain is the newer one, then the Domain Adaptation process attempts to alter the source domain in a way to bring the distribution of the source closer to that of the target domain.

Transfer Learning in comparison with Domain Adaptation, is a broader topic. Apart from the above explained difference in the marginal distributions or in feature spaces, Transfer Learning can also include the case when you want to transfer knowledge from an old task to a new and different one.

Here are some surveys that analyzing the topic of Transfer Learning,

PAN, S. J.; YANG, Q. A Survey on Transfer Learning. IEEE Transactions on Knowledge and Data Engineering, v. 22, n. 10, p. 1345–1359, out. 2010.

Weiss, K., Khoshgoftaar, T.M. & Wang: A Survey of Transfer Learning. D. J Big Data, 2016.