Add default params_to_tune for decomposition transforms

[Mr-Geekman]

🚀 Feature Request

Add default params_to_tune for decomposition transforms

Proposal

Suggested grid for STLTransform:

{
    "model": CategoricalDistribution(["arima", "holt"]),
    "robust": CategoricalDistribution([False, True]),
}

It is expected that period is set by the user.

Suggested grid for LinearTrendTransform:

{
    "poly_degree": IntUniformDistribution(low=1, high=2),
}

Suggested grid for TheilSenTrendTransform:

{
    "poly_degree": IntUniformDistribution(low=1, high=2),
}

Suggested grid for ChangePointsSegmentationTransform. We should set default value for change_points_model like in ChangePointsTrendTransform. We tune parameters only if default values for change_points_model and per_interval_model are set:

{
    "change_points_model.change_points_model.model": CategoricalDistribution(["l1", "l2", "normal", "rbf", "cosine", "linear", "clinear", "ar", "mahlanobis", "rank"]),
    "change_points_model.n_bkps": IntUniformDistribution(low=5, high=30),
}

Suggested grid for ChangePointsTrendTransform. We tune parameters only if default values for change_points_model and per_interval_model are set:

{
    "change_points_model.change_points_model.model": CategoricalDistribution(["l1", "l2", "normal", "rbf", "cosine", "linear", "clinear", "ar", "mahlanobis", "rank"]),
    "change_points_model.n_bkps": IntUniformDistribution(low=5, high=30)
}

Suggested grid for ChangePointsLevelTransform. We tune parameters only if default values for change_points_model and per_interval_model are set:

{
    "change_points_model.change_points_model.model": CategoricalDistribution(["l1", "l2", "normal", "rbf", "cosine", "linear", "clinear", "ar", "mahlanobis", "rank"]),
    "change_points_model.n_bkps": IntUniformDistribution(low=5, high=30),
}

We can't tune per_interval_model here, because possible values are classes instead of serializable types.

Suggested grid for TrendTransform. We should set default value for change_points_model like in ChangePointsTrendTransform. We tune parameters only if default values for change_points_model and per_interval_model are set:

{
    "change_points_model.change_points_model.model": CategoricalDistribution(["l1", "l2", "normal", "rbf", "cosine", "linear", "clinear", "ar", "mahlanobis", "rank"]),
    "change_points_model.n_bkps": IntUniformDistribution(low=5, high=30),
}

To discuss:

• Do we want to tune more parameters of LinearTrendTransform and TheilSenTrendTransform that are related to the model itself?
  • It was decided not to.
• What should we do with change_points_model and per_interval_model? It look that there are a lot of problems with them.
• Do we want to try MeanPerIntervalModel and MedianPerIntervalModel for TrendTransform?

Test cases

Look at #1221. Check empty grid as well.

Additional context

No response

[Mr-Geekman]

Closed by #1243.