Uber Eats Next-Gen Restaurant Recommendation: Transformer-Based Sequence Model with Real-Time Features

Uber Eats overhauled its homefeed recommendation model architecture, moving from static statistics-based features to a hybrid system combining DLRM/DCNv2 with Transformer-based sequence modeling. The system processes real-time user behavioral sequences through multi-head self-attention layers, capturing fine-grained temporal dependencies and evolving user intent. It leverages target-aware sequence modeling where the candidate merchant is appended to the user's action sequence, allowing the Transformer to compute direct relationships between past behavior and candidate merchants. The Near-Real-Time feature system, built on Uber's Next Personalization Platform, uses UserContext event-sourced architecture to compute features on the fly from a user's sequence of past actions, replacing the previous 24-hour+ batch processing lag. The system synthesizes billions of signals from real-time behavioral cues to geographic context to rank optimal options for every session. This powers the primary gateway for millions of Uber Eats users worldwide, reducing cognitive load for users while providing a critical visibility and growth platform for restaurants, grocery, alcohol, and retail merchants.