How to productionize a k-fold target-encoded feature?

Question

I am attempting to build a model that has many predictors which are both categorical and large in cardinality. Target encoding looks to be a good solution for including these features, but I'm unsure of how to productionize the mapping if k-fold target encoding is used to avoid overfitting as described below:

Say 5 folds are divided at random; each category in each feature will then have 5 means in the training set. How do I decide which value to map to in production? My only thought is that the folds should be divided deterministically; then in production, we have a mapping for each fold definition - then each event that runs through our production model will have its fold calculated in realtime and mapped according to that value?

Image taken from: https://towardsdatascience.com/getting-deeper-into-categorical-encodings-for-machine-learning-2312acd347c8

Answer 1

The answer is in the source code for the blog post: https://github.com/groverpr/Machine-Learning/blob/9963e59823fe0ff18cc8e1b2657b71c01f133193/catboost/utils.py#L25-L27

They re-fit the target encoding using the entire training set, before making out-of-sample predictions.

Note that they also fill missing values with the global mean of the target.

Edit: I wonder if it could be useful to average together the results of the K encodings, rather than re-training a single encoding on the entire dataset.