Retraining an existing XGBoost model

[ya...@gochange.co]

I have a pretty accurate XGBoost model which I wish to deploy.

This Model should be retrained on a daily basis with a new dataset (sliding window)

Is there a way to retrain (not rebuild) an existing XGBoost model with the new DataSet before I run a prediction?

Thanks!

[Darren Cook]

Sliding window means you don't just want to have it use new data, but
also forget about some older data? I'd say, no, you have to rebuild each
time.

If you have enough data, you could build one model per day, using only
that day's data. And then ensemble a sliding window of models. That
would need some careful evaluation though, to see if it is as accurate.
(It could even be more accurate.)

I've had a system in production using a similar idea of an ensemble of
models built on different sliding windows of data. So a new model is
built each week, based on the last N weeks of data, and the oldest model
is dropped at the same time.

Darren



--
Darren Cook, Software Researcher/Developer