jrb46
- jrb46d
predictions are from two tree based models (XGBoost), one trained to predict the target another trained to predict the meta model.jrb46
is just the target model and jrb46d
is just a model of the meta model (meta-model model?). And everything in between is interpolating between various proportions of meta-model model subtraction. The base model hasn’t been updated in over a year now, and the meta model model was trained in January. IMO, it’s a failed experiment, nothing of value there, except the negative result. Surprising how well you can do on corr by just predicting the averages.
jrb19
OTOH is a mixture of hypernetworks jointly trained on all the targets in the v4
dataset with a custom loss to maximise fncv3
. Another relic from the 3x TC era.
EDIT: I just took a look at the code and it turns out that jrb46
is a LightGBM model and not an XGBoost model. The other model (meta-model model) is XGBoost.