Describe the workflow you want to enable

Varying the seed of the FeatureHasher allows the user to control what inputs collide. This can allow for a better feature space either through experimentation (as a hyperparameter) or explicitly searching for a space that minimizes "bad" collisions

Describe your proposed solution

Add an optional "seed" parameter to the init of FeatureHasher which defaults to 0 (the current behavior, see the underlying hashing function). The seed would be thread through to _hashing_transform

Ditto for HashingVectorizer. Only difference here is that the seed would be passed to the FeatureHasher instance

This seems straightforward so I can implement this solution if it makes sense

Describe alternatives you've considered, if relevant

No response

Additional context

No response