[ENH] extending reducers to hierarchical data

[danbartl]

This is the final step to enable global forecasting via an efficient application of make_reduction using the new argument transformers.

I will do some refactoring, introduce checks etc., but already posted to discuss implementation strategy, see below.

Example use case

# -*- coding: utf-8 -*-
"""Test extraction of features across (shifted) windows."""
__author__ = ["danbartl"]

import pandas as pd
from sklearn.ensemble import RandomForestRegressor
from sklearn.pipeline import make_pipeline

from sktime.datasets import load_airline
from sktime.datatypes import get_examples
from sktime.forecasting.compose import make_reduction
from sktime.forecasting.model_selection import temporal_train_test_split
from sktime.transformations.series.summarize import WindowSummarizer

# Load data that will be the basis of tests
y = load_airline()
y_pd = get_examples(mtype="pd.DataFrame", as_scitype="Series")[0]
y_series = get_examples(mtype="pd.Series", as_scitype="Series")[0]
y_multi = get_examples(mtype="pd-multiindex", as_scitype="Panel")[0]
# y Train will be univariate data set
y_train, y_test = temporal_train_test_split(y)

# Create Panel sample data
mi = pd.MultiIndex.from_product([[0], y.index], names=["instances", "timepoints"])
y_group1 = pd.DataFrame(y.values, index=mi, columns=["y"])

mi = pd.MultiIndex.from_product([[1], y.index], names=["instances", "timepoints"])
y_group2 = pd.DataFrame(y.values, index=mi, columns=["y"])

y_grouped = pd.concat([y_group1, y_group2])

# Get different WindowSummarizer functions
kwargs = WindowSummarizer.get_test_params()[0]
kwargs_alternames = WindowSummarizer.get_test_params()[1]
kwargs_variant = WindowSummarizer.get_test_params()[2]

regressor = make_pipeline(
    RandomForestRegressor(),
)

forecaster = make_reduction(
    regressor,
    scitype="tabular-regressor",
    transformers=[WindowSummarizer(**kwargs)],
    window_length=10,
)

forecaster.fit(y_grouped, fh=1)

y_pred = forecaster.predict(fh=1)

Open questions:

What kind of arguments do we want to use in transformers? Currently I can only think about WindowSummarizer, but we could of course apply any kind of function where we want a grouped application. Currently only the first transformer in the list is applied (I will extend this after we resolved the open implementation question)

What is currently best practice to apply an Imputer across the different y time series grouped together in pd-multiindex? Should that also be covered here?

[fkiraly]

notebook is failing - related to the bug in get_time_index fixed here?
#2380