Ground control to Major Tom… to MARS and earth and back…

I’ve finished Matloff’s excellent R book, and I’m now wading through some books on using R for Bayesian inference and trying to get handle on machine learning with R in a catch-as-catch can kind of a way (with some help from Conway’s O’Reilly text:

One of the machine learning methods that has really captured my interest as of late–which¬†unfortunately is not expicitly addressed in Conway’s book–is MARS (which I’ve come to learn is now a term “owned” by Salford Systems–but stands for Multivariate adaptive regression splines As someone who has used quite a bit of MLR, MARS represents a really cool extension in that it creates inflection points (splines) in linear functions to optimize model fit.

In medicine, where heterogineity of patient populations can manifest as different treatment effects (based on underlying demographic, behavioral, or genetic characteristics), this method of modeling data seems really appaealing. I can see how knot placment, in-and-of-itself, may lend some interesting insights as to how subpopulations may gain differential relative benefit from medications or other kinds of interventions!.. which may lead to a better unerstanding of what interventions bring the highest value for target sub-populations!

That said, I’m looking forward to applying MARS to a future project.

As far as I can tell the R package that implements MARS is called earth, and I have began to browse through the manuals available on CRAN ( Fun stuff indeed….