Support for Bayesian Personalized Ranking

[jerry-rubikloud]

Hi there, I'm thinking of contributing codes for BPR. What is the best way to extend the code to handle this optimization in your opinion?

[geffy]

Hi @jerry-rubikloud ! I think that we need to implement 2 components:

1. BPRSampler. It will generate pairs from source data and put them into batch. Example: for N pairs we will have a batch with N*2 samples (2 * i-th are winners, 2 * i + 1-th are losers)
2. BRPLoss. It will take expanded batch output ([2*N, 1]), convert it to pairs ([N, 2]) and calculate resulting loss. This design seems to be easy for implementation and allows to keep FM core routine unchanged