yadda yadda yadda

Right now we use label encoding by default (and backend-specific native categorial support). But still we need to make some discovery on which encoding type is better? * OHE introduces...

Expose the train/split ratio in the model config

2

Now it's hardcoded on 80%, which is fine for most of the cases, but not nice.

So a user can replicate the complete demo locally. Useful for a quickstart guide.

It can be useful for the ranklens demo, when you can start metarank with config file URL instead of messing with docker volumes. The S3/HTTP support should be transparent for...

As there is a ton of basic grammar mistakes

Why not LambdaMART? Pairwise-logreg is: * explainable, so weights can be either trained from historical data, or entered manually with no training. * manual weights definition allows user to play...

Geolocation feature extractor

3

If there's something like IP address, then we can map it to a country and city with GeoLite2 database. But there are some problems: * the database itself is free,...

two options: 1. call the `/train` rest endpoint 2. run the `metarank train` from cli/docker/k8s-cronjob

Typical approaches to writing a feature extractor configuration.

What is a good NDCG, what looks suspicious.