Building Recommendation Systems in Python and JAX

Chapter 18. What’s Next for Recs?

We find ourselves in a transitionary time for recommendation systems. However, this is quite normal for this field, as it is in many segments of the tech industry. One of the realities of a field that is so closely aligned with business objectives and with such strong capabilities for business value is that the field tends to be constantly searching for any and all opportunities to advance.

In this chapter, we’ll briefly introduce some of the modern views of where recommendation systems are going. An important point to consider is that recommendation systems as a science spread both depth first and breadth first simultaneously. Looking at the most cutting-edge research in the field means that you’re seeing deep optimization in areas that have been under study for decades or areas that seem like pure fantasy for now.

We’ve chosen three areas to focus on in this final chapter. The first you’ve seen a bit of throughout this text: multimodal recommendations. This area is increasingly important as users turn to platforms to do more things. Recall that multimodal recommendations occur when a user is represented by several latent vectors simultaneously.

Next up is graph-based recommenders. We’ve discussed co-occurrence models, which are the simplest such models for graph-based recommendation systems. They go much deeper! GNNs are becoming an incredibly powerful mechanism for encoding relations between entities and utilizing these representations, making ...