Function to make top-k recommendations to all users in a fast way for ALS model.

[Tych0n]

Function to make top-k recommendations to all users in a fast way for ALS model. Main feature is custom C++ argsort for numpy 2d ndarray that runs in parallel.

For users_items == ~11kk x 130k it requires ~45 minutes to evaluate on 80-CPU server and is 10-20 times faster than looping model.recommend. I think it closes #29 and #72 . May be it can be improved by ideas from #126.

Please feel free to criticise. I can move it anywhere in code to be more consistent from your point of view with the library structure.

[ita9naiwa]

I'll check the code soon...

[ita9naiwa]

I briefly checked the code(I didn't understand the detail of cython code) but

(It's totally my opinion)

(1) This function is very helpful for both actual recommendation and evaluation.

I think it should be a method of MatrixFactorizationBase because BPR model also generates exactly the same way as ALS does.

implicit/implicit/recommender_base.py

Line 128 in 94e4e7c

class MatrixFactorizationBase(RecommenderBase):

[ita9naiwa]

Oh, I've just seen that the owner wants it to be in evaluation.py at #72
hmm...

[Tych0n]

We don't have evaluation.py yet, but we have evaluation.pyx instead)

[ita9naiwa]

That's not the point btw. Anyway, could you please consider moving that function into the method of MatrixFactorizationBase?

[ita9naiwa]

one last thing...
its great if you add some unittests for this.

[ita9naiwa]

Thanks!!