How to replicate StructuredDataClassifier and Regressor functionality in 2.0?

[adunega]

Hi there, huge fan of AutoKeras. It's been a tremendous tool for my financial research. I just noticed in the new release of 2.0 that the primary functions I was using were dropped from the code (StrucutredDataClassifier and StructuredDataRegressor).

I'm curious on the reason for dropping those, but more importantly do you have any recommendations/advice on another tool I should use for replacing that functionality? Can I rebuild it myself using AutoModel or should I just fork AK 1.1 if I want to keep using those functions?

Thanks in advance for your help.

[haifeng-jin]

Thanks, @adunega !

Yes, it is removed.
The main reason is the search space is limited to neural networks, which may not be as good as other models for structured data.
It is also adding a lot of maintenance overhead.

The feature encoding part is also not good enough since AutoML cannot understand the features but only brute-forcely search it.

I suggest you to do the feature encoding for all categorical features by yourself and use it with a normal Input() and RegressionHead() or ClassificationHead() with the IO API in AutoKeras.

Or you can just use XGBoost, which may be better for structured data tasks.

[adunega]

Thanks for the explanation and suggestions!