Building Recommendation Systems in Python and JAX

Chapter 17. Sequential Recommenders

In our journey so far, you’ve learned about a variety of features that appear as explicit or as latent components in the recommendation problem. One kind of feature, which has appeared implicitly, is the history of previous recommendations and interactions. You may wish to protest here: “All of the work we’ve done so far considers the previous recommendations and interactions! We’ve even learned about prequential training data.”

That is true, but it fails to account for more explicit relationships between the sequence of recommendations leading up to the inference request. Let’s look at an example to distinguish the two. Your video-streaming website knows that you’ve previously seen all of Darren Aronofsky’s films, so when The Whale is released, the website is very likely to recommend it. But this type of recommendation is different from one you might receive after finishing episode 10 of Succession. You may have been watching Aronofsky films over a long time period—Pi many years ago and Black Swan earlier this year. But you have been watching an episode of Succession each night this week, and your entire recent history is made up of Logan Roy. This latter example is a sequential recommendation problem: using the most recent ordered list of interactions to predict what you’ll enjoy next.

In terms of the modeling objective, the recommenders we’ve seen use pairwise relationships between potential recommendations and historical interactions. Sequential ...