Personalized ranking is the right goal for restaurant discovery, and the two-tower embedding model described here is a practical, elegant way to get there. When popularity ranking fails, and it often does, especially for users who want something beyond what everyone else is eating, a lightweight approach like this can restore the human element to a system that had lost it.
The core insight is straightforward: most recommendation engines lean heavily on aggregate signals. The most-reviewed, the highest-rated, the most-booked. That works well enough for the mainstream, but it creates a blind spot for anyone looking for a quiet neighborhood spot, a specific cuisine, or a place that fits their unique preferences. The two-tower model solves this by building separate embedding spaces, one for the user, one for the restaurant, and learning to map them together. It does not need massive compute resources or a complete overhaul of your data pipeline. It makes personalization accessible to teams that are not operating at the scale of a major tech platform.
What this means for practitioners is that you can improve discovery without chasing complexity. The model described is a variant of a well-known architecture, adapted for a domain where the interaction signal is sparse. Restaurants are not movies. People do not rate them as often, and their preferences shift with context, mood, time of day, company. The two-tower approach handles this by focusing on the relationship between user behavior and item attributes, rather than trying to predict a universal score. It is a reminder that smarter embeddings do not require bigger models; they require better framing of the problem.
The practical takeaway is this: if your current ranking feels like it is serving the same results to everyone, look at how you are representing your users and items. A two-tower model gives you a path to personalization that is both interpretable and deployable on modest infrastructure. It does not promise to know you better than you know yourself. It promises to stop guessing what everyone wants and start learning what you actually do. For any team building a discovery experience, that is a trade worth making.