A Multi-Latent Transition model for evolving preferences in recommender systems.

•We capture user preferences dynamics based on a multi-latent analysis.•We design a joint objective function and we propose an efficient optimization algorithm.•We evaluate our method on 2 extended benchmark datasets that span 3 and 4.5 years.•Our model outperforms baselines for users with stable and dynamic preferences.•The extended datasets of MovieLens-1M and Last.fm-1K are made publicly available.