Does anyone know for when Rocket is applied in the multivariate setting how random kernels are generated?
Namely is a 1-D kernel randomly generated and applied to a randomly selected feature?
Or is a 2-D kernel generated randomly and this kernel is applied using all features?
Neither the original paper nor the documentation of packages that implement the algorithm seem to mention anything about it.
Looking at the source code, the sktime implementation randomly selects the number of features to use for each kernel and then randomly chooses the input features.
For example, if your dataset contains three features, sktime's ROCKET will first sample a number between 1 and 3 as the number of input channels in the kernel. As a next step, it will then draw the channels to consider.
Although this approach to multivariate timeseries is not mentioned in the original paper, the authors refer to this approach as a naive facility to apply ROCKET to multivariate timeseries in the follow-up "MiniRocket: A Very Fast (Almost) Deterministic Transform for Time Series Classification"
