yadda yadda yadda
yadda yadda yadda
technically it is fixed in https://github.com/metarank/metarank/pull/432, but we never ran it on jvm 17.
Seems to be an issue related to the `pred_contrib=True` handling in the C library itself: in the `lightgbm4j` library we have an issue with concurrent prediction crashing the whole JVM...
I guess there's a more generic ticket exists: https://github.com/metarank/metarank/issues/664 Currently, metarank just randomly samples click-through sessions into either train or validation set (with some hardcoded probability, like 80%). This is...
We have a plan for a State Store API #571 so you can nuke/request everything Metarank knows about a specific user with an API call. So: * you receive a...
AFAIK decoders generated with `io.circe.generic.auto._` do not take any `Configuration` from the `generic.extras` into the account. But if you change the import to `io.circe.generic.extras.auto._`, the code should work.
On mapdb 3.0.8, scala 2.12.11 and jdk11 the issue is not reproducible anymore.
AFAIK all three libs doing LambdaMART, LightGBM, XGBoost and CatBoost can handle categorial features natively the way you've described with TargetEncoder. The problem is that LightGBM is the only one...
@mullerhai We already did quite a lot of prototyping for the recommendations in Metarank, and using PyTorch models is indeed possible. But there are still some questions left before implementing...
Spent some time playing with javacpp-torch, and I guess it can be worth digging a bit more. In simple words, we wanted to call python libs as then we don't...