Building Recommendation Systems in Python and JAX

Chapter 1. Introduction

Recommendation systems are integral to the development of the internet that we know today and are a central function of emerging technology companies. Beyond the search ranking that opened the web’s breadth to everyone, the new and exciting movies all your friends are watching, or the most relevant ads that companies pay top dollar to show you lie more applications of recommendation systems every year. The addictive For You page from TikTok, the Discover Weekly playlist by Spotify, board suggestions on Pinterest, and Apple’s App Store are all hot technologies enabled by the recommendation systems. These days, sequential transformer models, multimodal representations, and graph neural nets are among the brightest areas of R&D in machine learning (ML)—all being put to use in recommendation systems.

Ubiquity of any technology often prompts questions of how the technology works, why it has become so common, and if we can get in on the action. For recommendation systems, the how is quite complicated. We’ll need to understand the geometry of taste, and how only a little bit of interaction from a user can provide us a GPS signal in that abstract space. You’ll see how to quickly gather a great set of candidates and how to refine them to a cohesive set of recommendations. Finally, you’ll learn how to evaluate your recommender, build the endpoint that serves inference, and log about its behavior.

We will formulate variants of the core problem to be solved by recommendation ...