Add NCE-PLRec

[deklanw]

From "Noise Contrastive Estimation for One-Class Collaborative Filtering" http://www.cs.toronto.edu/~mvolkovs/fp046-wuA.pdf

Repo:
https://github.com/wuga214/NCE_Projected_LRec

Essentially: reweights the interaction matrix with respect to item popularity, does randomized SVD factorization, and then uses the user representations from that for L2-regularized linear regression with a closed-form representation (like EASE) to predict item interactions.

There is a variant in the paper which allows for "non-uniform weighting of users and items". They don't test that method nor implement it in their code, so neither did I.

Performs very well on ml-100k in my testing

[2017pxy]

This PR has been removed to #1154