A quick glance through the
scikit-learndocumentation on linear models, or the CRAN task view on Mixed, Multilevel, and Hierarchical Models in R reveals a number of different procedures for fitting models with linear structure. Each of these procedures meet different needs and constraints, and some of them can be computationally intensive to compute. But in the end, they all have the same underlying structure: outcome is modelled as a linear combination of input features.But the existence of so many different algorithms, and their associated software, can obscure the fact that just because two models were fit differently, they don’t have to be run differently. The fitting implementation and the deployment implementation can be distinct. In this note, we’ll talk about transferring the coefficients of a linear model to a fresh model, without a full retraining.
I had a similar problem about 18 months ago, though much easier than the one Nina describes, as I did have access to the original data and simply needed to build a linear regression in Python that matched exactly the one they developed in R. Turns out that’s not as easy to do as you might think: the different languages have different default assumptions that make the results similar but not the same, and piecing all of this together took a bit of sleuthing.