Add inverse_transform() to random projection classes

Describe the workflow you want to enable

GaussianRandomProjection and SparseRandomProjection should have an inverse_transform() method.

Describe your proposed solution

Simply add the inverse_transform() method, with the appropriate input validation and tests, and return X @ np.linalg.pinv(self.components_.T).

Describe alternatives you’ve considered, if relevant

N/A

Additional context

My understanding is that every transformer that performs a reversible transformation should have an inverse_transform() method, unless there’s a really good reason not to (e.g., if it’s terribly inefficient, or there’s a better alternative). Was the inverse_transform() left out because it’s too inefficient?

[adrinjalali]

true, you could open a PR with that approach.

[ageron]

Done. I named the parameter fit_inverse_transform, like in KernelPCA. I also named the learned pseudo-inverse of the components components_pinv_, for lack of a better idea. I propose to continue the discussion in the PR.