reducers: custom lags, feature importance

[fkiraly]

From discord Apr 12, 2023, @bavquant:

I am new to this channel and have a question regarding forecasting with SkTime. First of all thanks for the great tool!!! I was able to use XGBoost with additional regressors, and was also able to get the feature relevance when just using reduction for a univariat forecast. Where I am a bit lost are the following topics: a) Allocating the feature importance to the correct features when using additional regressors with lag, b) selectively limiting the lags to be created for the original series (e.g. do a lag of 3 and 7, but not all lags 1...7), and c) accessing the feature importance (or at least an average of it) when doing a backtesting / cross validation with given step size and forecasting horizon using 'SlidingWindowSplitter 'and 'evaluate' or 'SlidingWindowSplitter' and 'ForecasterRandomizedSearch', the latter for parameter optimization. Would be fantastic if anyone could give me a hint....

[fkiraly]

(a) Allocating the feature importance to the correct features when using additional regressors with lag

this is unfortunately not easy, as the names get lost internally afaik. FYI @danbartl if you know a trick.

Related to this is a rework of the reducer we have been working on:
#3224
(you can already import the classes DirectReductionForecaster and RecursiveReductionForecaster from `sktime.forecasting.compose._reduce)

but the main problem with this is that it is slow for reasons not fully understood (maybe from combine_first and sorting)

Contributions that put the right column names internally and/or fix the speed issue are appreciated.

This preserves the names

[fkiraly]

(b) selectively limiting the lags to be created for the original series

You can do this by using the transformers argument with a WindowSummarizer or a Lag transformer.

There should be, hopefully, an example in the notebook 01c_forecasting_hierarchical_global.ipynb, see here on binder: https://mybinder.org/v2/gh/sktime/sktime/main?filepath=examples

An alternative is using YToX and Lag transformers and then using a window length of 1, to do this manually.

[fkiraly]

c) accessing the feature importance

This should work via the get_fitted_params method, and/or by accessing the best_estimator_ parameter. If using get_fitted_params the same fitted estimator is under the key "best_estimator".

On the best estimator, find the feature importances as you normally would within the reducer.