Excluding some negatives from training

Hi!

First of all, thank you for this great library!

I'm currently using it for extending and comparing some matrix factorisation models for recommendation. I am working on the [LFM-2b dataset](http://www.cp.jku.at/datasets/LFM-2b/).

For design reasons, part of my interaction matrix (or rather an extended version of it) consists of a block of zeros, which however should NOT be considered while training. In other words, if we imagine to be using either BPR or WARP, the negative samples should never belong to this block.

To explain things better, I want to train on a matrix like this:
with **A, B, C** being sparse matrices and **0** being an empty matrix. The model should NOT use the entries of the bottom-right block for training.

Is there a way to do this?

Thanks :)
- Marta

Start by examining the BPR and WARP negative-sampling paths and how the interaction matrix is passed to training. Define how the excluded bottom-right block is represented and verify that sampled negatives never come from it; the issue provides no named files or tests.

Written by the indexing model from the issue text.