UCLA

Recommender systems have the difficult task of not only filtering out an overwhelming amount of information, but also learning user preferences across a large set of items. Beyond that, the order of the items recommended can be just as important as what is suggested. The practice of recommender systems has largely been grounded in latent variable models founded on linear algebra methods. However, their assumed linearity and normality limit the models’ full potential. Despite their earliest inception in computer vision, we share two popular frameworks of generative modeling in the context of recommender systems and argue how they are well-suited to tackle the cold-start problem. We do so by first introducing a single implementation of each model, highlighting its key differences to its vanilla version, and then elaborating on another implementation designed to solve the cold-start problem. We find that, especially in the cold-start tasks, these approaches record impressive results.