iiinova — where your digital self keeps evolving

statmills.bsky.social

@statmills.bsky.social

37 followers 150 following 11 posts

Data Science, CFB, Abundance Agenda www.statmills.com

Posts Replies Media

statmills.bsky.social @statmills.bsky.social · 07/04/2026

The result is smooth curves that can learn high dimensional interaction effects that you can fit at scale!

000

statmills.bsky.social @statmills.bsky.social · 07/04/2026

Then fitting our Gradient Boosting Spline model is as simple as calling {DecisionTreeRegressor} in a for loop to build up our coefficient predictions.

100

statmills.bsky.social @statmills.bsky.social · 07/04/2026

{JAX} makes this super easy because all we need to do as practitioners is to define our loss function as a function of our parameters and let autodiff handle calculating the partial derivatives

100

statmills.bsky.social @statmills.bsky.social · 07/04/2026

Gradient Boosting Machines will generate predictions for each observation by iteratively learning to predict the gradient of the loss function at each iteration. In this blog post I show you can use the same methodology to fit entire parameter sets

100

statmills.bsky.social @statmills.bsky.social · 05/02/2025

There is an additional layer of smoothing you can do with low-rank smoothers, which not only smooths the data but speeds up the processing because we use less parameters!

000

statmills.bsky.social @statmills.bsky.social · 05/02/2025

The smoothing you can do with location data is really interesting! Similar to penalizing neighboring coefficients in a GAM, you can penalize the difference between neighboring census tracts

100

statmills.bsky.social @statmills.bsky.social · 05/02/2025

This allows us to look at some cool maps like where first time home buyers are buying, and where the most expensive homes are

100

statmills.bsky.social @statmills.bsky.social · 05/02/2025

I hate how it's never the actual data analysis that trips me up when switching between R and Python, it's the silly base stuff like `int` and `len` that takes me multiple tries to switch over 🤬

030