…his post and all the others related to it. Posts in time-series clasfication or regression are scarse. From what I could understand, from all these methods. only the time series forest classifier is interpretable right? For example, rocket, it presents amazing performances but as it transforms the features, one cannot interpret the linear classifier, in case of a multivariate dataset.
The Shapelet Transform Classifier is interpretable — you can extract and plot shapelets with little trouble. See the documentation here: Shapelets and the Shapelet Transform with sktime — sktime documentation
In concept, it might be possible to extract features from ROCKET by extracting meaning from the most significant kernels, but Ihave not seen this done.
I’ll look into it further and might put together a blog post around interpretability.