Tuning implicite hyperparameter of tranformers inside TransformedTargetForecaster

Is your feature request related to a problem? Please describe. I have locally created a class TuningTransformer inspired by this TabularToSeriesAdaptor in order to tune the implicite hyperparameter of using a tranformers or not inside TransformedTargetForecaster in order to tune this all together in a ForecastingGridSearchCV. Therefore, the class TuningTransformer has an arg use: bool to turn on and off the use of a tranformer.

Describe the solution you’d like All together looks as follows:

TuningTransformer(BoxCoxTransformer(method="pearsonr"))

Or more complicated with TabularToSeriesAdaptor:

TuningTransformer(TabularToSeriesAdaptor(MinMaxScaler()))

So in the end I will put each tranformer inside a TuningTransformer and all together with GS looks like this:

pipe = TransformedTargetForecaster(steps=[
    ("deseasonalise", TuningTransformer(transformer=Deseasonalizer())),
    ("minmax", TuningTransformer(transformer=TabularToSeriesAdaptor(MinMaxScaler()))),
    ("power", TuningTransformer(transformer=TabularToSeriesAdaptor(PowerTransformer()))),
    ("model", TBATS()),
    ]
)

cv = SlidingWindowSplitter(initial_window=int(len(y_train) * 0.5), window_length=24, start_with_window=True, step_length=24)
param_grid = {
    "deseasonalise__transformer__model" : ["multiplicative", "additive"],
    "deseasonalise__transformer__sp" : [12],
    "deseasonalise__use" : [True, False],

    "power__transformer__transformer__method" : ["yeo-johnson", "box-cox"],
    "power__transformer__transformer__standardize" : [True, False],
    "power__use" : [True, False],

    "minmax__transformer__transformer__copy": [(0,1), (0,2)],
    "minmax__use" : [True, False],

    'model__sp': [12],
}
gscv = ForecastingGridSearchCV(forecaster=pipe, param_grid=param_grid, cv=cv, n_jobs=-1)
gscv.fit(y)

It can look a bit confusing to have grid arguments like minmax__transformer__transformer__copy, however this is all consistent with existing stuff in sktime and sklearn. And it is very handy for tuning 😃

Additional context I might contribute this to sktime in case it is appreciated, otherwise I dont mind to keep it in our private repo. I am open to adjust my implementations in case there are other ideas how to solve the implicit hyperparameter tuning in this case. Maybe the class can have a better name…

---

Martin Walter [e-mail], Mercedes-Benz AG on behalf of Daimler TSS GmbH. https://github.com/Daimler/daimler-foss/blob/master/LEGAL_IMPRINT.md

[fkiraly]

Quick question, @aiwalter, what does this do? Does it simply default to the identity transformer if use=False, and to the wrapped transformer if use=True?

Seems like a really good idea to me!

The name though seems sub-optimal, because you would expect a TuningTransformer to tune (which it doesn’t). How about Optional, or OptionalTunable or similar?

PS: beware overfitting by overtuning…

[fkiraly]

Regarding implementation, it’s probably a good idea to have an OptionalPassthrough transformer.

Agreed, seem my point above:

Might be worth having the OptionalPassthrough as a shorthand, but that would imply that it should perhaps be wherever the multiplexer is, and it should perhaps even be a contraction of the multiplexer?