Dealing with categorical variables to predict new values in the input feature about sklearn2pmml

[vruusmann]

Cat codes transformation is able to deal with new values while predicting

Could you tell me more about this fabulous "Cat codes" transformation? Which Python class, in which Python package? Perhaps it would be appropriate to incorporate it into JPMML-SkLearn and SkLearn2PMML projects.

Can you suggest a way how to deal with the PMML with the encoding, either in python or in JPMML side.

When dealing with categorical features, then the first transformation of a pipeline should be sklearn2pmml.decoration.CategoricalDomain, which lets you customize missing value and invalid value treatments according to the PMML specification.

You could try the following parametrizations:

• CategoricalDomain(invalid_value_treatment = "as_is") - Invalid values are dispatched to the model. It is assumed that the model contains logic for detecting and handling these.
• CategoricalDomain(invalid_value_treatment = "as_missing") - Invalid values are converted to missing values, and then dispatched to the model.
• CategoricalDomain(invalid_value_treatment = "as_missing", missing_value_replacement = "1.0"). Convert invalid values to missing values, and then perform missing value replacement. The total effect is that the model sees invalid values as "1.0".