Implement returning already liked items

[vhfmag]

As asked in issue #40

[benfred]

Thanks for the PR!

There are a couple minor issues highlighted by the failing travis ci build though : the annoy_als.py version fails to run the unittests (best should be zip(ids, dist)) and there are some style issues from flake8.

For the API, I think the parameter should be filter_liked=True - rather than return_liked=False

Also I think in this case we could be more efficient in how we're returning the items. Each 'recommend' method is querying for count = N + len(liked) items (like here), but in this case we're only need the N items. Maybe we should pull out the code that builds the 'liked' set to a common member on the base class - but only if the filter_liked flag is set. What do you think?

[vhfmag]

I'll work at it this weekend. I thought of implementing as much as possible of the recommendation function in the base class and creating a private abstract method for the algorithm-dependent logic, what do you think?

I'm not familiar with flake8, so I'll read about it before proceding :)

[vhfmag]

Hello! Did it :)

I tried to put all the recommendation logic in the RecommenderBase by creating two abstract methods: liked and best_recommendations.

Also, there were some style issues with lines being too long, specially function calls or definitions, so I put each parameter in a separated line, I don't know if it's the best practice, but it passed the test.

[vhfmag]

Any updates?

[ghvg1313]

@benfred Is the only issue being the conflicts here? @vhfmag We are trying to achieve the exact same thing to validate the model. I might be able to help resolve the conflicts.

[lautjy]

IMO this should be closed.
This PR implements desired functionality and is in master:
fd984bc