Tutorial on Theoretical Foundation for Recommender Systems

In the past several years, I spent most of my time productionizing recommender systems and researching the various gaps between what we observe in practice and what existing theory tells. Modern recommender system, in my experience, has evolved a long way from the initial Netflix challenge and collaborative filtering approaches. The same story happens for Information Retrieval (IR) as well, though it seems the boudary between IR and Recsys is gradually fading these days – many shared ideas and solutions are binding these two fields, and I believe this trend will continue in the future. In general, the most exciting advancement in the past decade has to do with deep learning (DL), and in one way or another, Recsys seems to be adopting it more than IR. Please don’t get me wrong, embracing DL does not mean significant progress has been made in the field, to me it just means more questions and challenges ahead. (Not finished…)