MLOperation

[eduard93]

A lot of models conform to the same interface:

• Fit(X, Y, Modeltype, Hyperparameters) → trained model
• Predict(X) →Y
• Optimize(Model, HyperparameterBoundaries, SearchStrategy) → Optimized model (via Hyperparameter optimization)

I propose adding a new BO (or extending existing PythonOperation) to automatically train/use the models, by providing these high-level requests:

• FitRequest
• PredictRequest
• ValidateRequest

Questions:

1. Passing X, Y is trivial but what about hyperparams? I can write/extract a list of params for each supported model but how to make it convenient for user?
2. This is also a departure from our current approach of Python code being written by user into a low-code territory. What do you think about it?
3. How do we do pipelines?

Ideas? @lukyanchikov, @bdeboe, @tom-dyar

[eduard93]

Also add some Hyperparameter optimization.