Predictive discrete latent factor models for large scale dyadic data