Feature Selection

[JakeColtman]

Resolution for #26

[JakeColtman]

The current approach to selecting feature importance requires generating lots of permutations of the model to produce a null distribution of feature inclusions. At the moment, there's no way to run models in parallel because joblib can't be nested. It would be nice to expose the chains to an external runner, so that all chains from all models can be run at once

[JakeColtman]

Having played around with the current implementation, two things strike:

• There's definitely a need for a light weight version for bigger data sets
• Restricting to run models sequentially is punishingly slow - the Sklearn model needs to expose a List[Delayed] for the different chains that other processes can use

[JakeColtman]

Still to do - wrap the logic up into an Sklearn BaseFilter

[JakeColtman]

Added to the PR:

• Sklearn wrapper for null distribution based feature selection
• Sklearn wrapper for simple proportion of splits feature selection for bigger data sets
• Ability for external clients to directly access the delayed chain methods to run multiple permutations of the model at once. I think this would allow trivial scaling across a dask cluster, but at the very least it allows big speep up from using chunkier machines