Closed dfsnow closed 5 months ago
I tested this locally using two different sets of hyperparameters. Both performed well, better than a linear model but worse than the current main model. Unfortunately, the Cubist implementation here has two major drawbacks:
The combination of these two things made it difficult/impossible to do a full grid search, as a sequential search would take days and a parallel search exhausts 250GB of memory.
Would like to play around with this in the future, but for now it's more trouble than it's worth.
Closes #37.
This PR tests using a Cubist model as an alternative to LightGBM and other GBDT models. It uses a simplified CV loop to just test the Cubist results, rather than a full pipeline refactor.