Feature importance ranking
Different options.
- SHAPley, which was previously used by umami
- Integrated gradients
- DeepTaylor
- Layerwise relevancy propagation
Once we have this we should evaluate which variables we can remove from the training.
cc @avaitkus