Add Adult Census dataset description

[ArturoAmorQ]

Fixes #657.

A potentially controversial PR, since adding a new dependency, passing through correctness of the wording and ending with the veracity of the interpretation given.

All feedback is welcomed!

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[ogrisel]

I think it's problematic to discuss and draw conclusions from the distributions of (gender or age based) sub-groups without taking fnlwgt properly into account. Furthermore I am not certain how we would do this: I have tried to see if you can use fnlwgt to recover the expected approximately 50%/50% relative representation between Male and Female adults in the US population but I failed.

We would have to dig the origin of this dataset to find out how it was built and how to properly use fnlwgt to draw conclusions on such sub group distributions but I think this goes beyond what we want to achieve with this MOOC.

[ogrisel]

For the record, here is the quick check I made:

>>> from sklearn.datasets import fetch_openml
>>> X, y = fetch_openml("adult", return_X_y=True)
>>> (X["sex"] == "Male").mean()
0.6684820441423365
>>> (X["sex"] != "Male").mean()
0.33151795585766347
>>> ((X["sex"] == "Male") * X['fnlwgt']).sum() / X['fnlwgt'].sum()
0.6757528069510574
>>> ((X["sex"] != "Male") * X['fnlwgt']).sum() / X['fnlwgt'].sum()
0.32424719304894256

[ogrisel]

I have the feeling that we can close this PR once #663 is accepted and merged if others agree.