I´m asking this question based on the following discussion: https://github.com/sktime/sktime/issues/4336
I´m using the RocketClassifier (https://www.sktime.net/en/latest/api_reference/auto_generated/sktime.classification.kernel_based.RocketClassifier.html) from the sktime module as classifier for multivariate time series. My time series data is in sktime mtype format ="pd-multiindex" (intervals resp. instances containing a fixed number of timespoints, the corresponding time points and the features as columns), like in this example data:
from sktime.datasets import load_unit_test
test_data = get_examples(mtype="pd-multiindex", as_scitype="Panel")[0])
test_data
Regarding to the discussion (https://github.com/sktime/sktime/issues/4336), a workaround for using the SHAP approach (https://shap-lrjball.readthedocs.io/en/latest/index.html) as kind of a feature importance extractor for Rocket multivariate time series classifier is to convert the input data into a 2D array. I´ve tryed this with my data and it worked so far. However, when transforming to 2D array, the information on the instances gets lost (only time points and feature values per time point are transformed into the 2D array.)
Does anyone have an idea, how to use this workaround for a panel pd-multiindex like input data like in the given example?
One code example given in the discussion thread from one of the contributers was this one: https://gist.github.com/varelasaraiva/fd16a3108a5bb1ae47b27443ca20933f
However, I´m not sure about the input format of the data in this code example and if it deals correct with the multi-index information in my case.
