Replies: 3 comments
-
(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: but the main problem with this is that it is slow for reasons not fully understood (maybe from Contributions that put the right column names internally and/or fix the speed issue are appreciated. This preserves the names |
Beta Was this translation helpful? Give feedback.
-
You can do this by using the 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 |
Beta Was this translation helpful? Give feedback.
-
This should work via the On the best estimator, find the feature importances as you normally would within the reducer. |
Beta Was this translation helpful? Give feedback.
-
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....
Beta Was this translation helpful? Give feedback.
All reactions