This is a simple implementation for paper Learning Attribute-to-Feature Mappings for Cold-Start Recommendations. Zeno Gantner, Lucas Drumond, Christoph Freudenthaler, Steffen Rendle, Lars Schmidt-Thieme. ICDM 2010. It can be applied to cold-start situations in recommendation systems.
In detail, Bayesian Personalized Ranking (BPR) model is a matrix factorization model for item recommendation. However, when new item comes with no rating information (i.e. cold-start), it's better to use the private information (i.e. attributes) to infer probable ratings of the items, or to infer feature of the underlying BPR model for better prediction for new items.
The paper offers four simple mapping functions (and a random prediction as baseline), which make it possible to predict under cold-start situation with underlying BPR model. Please refer to the paper for further detail.
Codes of BPR model are modified and refined from This repository. Many thanks to gamboviol.