People often find themselves unable to decide what to eat. It’s a one-time, one-off thing like a business opportunity, unless you’re willing to order from three different restaurants at the same time. Like, right now, you’re craving pizza, but then, you also like to try spicy tacos. The dilemma will soon be over.
Uber Eats has recently introduced brand-new updates to its recommendation system, incorporating user signals and a listwise ranking approach to enhance restaurant discovery.
It’s like browsing your socials now, wherein the recommendations you’ll see will be based upon the things you are interested in, while also enhancing ranking efficiency across candidate restaurants you’re more likely to select. It is deployed within the Uber Eats platform to support homepage feeds and discovery surfaces.
Yicheng Chen, engineer at Uber Eats, chatted with Ridesharing Forum, saying, “Leveraging near real-time user sequence features and a Generative Recommender-style model to power Uber Eats Home Feed recommendations and evolved the homefeed ranking from hand-crafted statistical features to transformer-based sequence modeling, cut feature freshness from 24 hours to seconds.”
This updated system replaces the earlier batch-oriented feature pipelines with a real-time signal processing layer, which continuously ingests the interaction of users, adopting them, such as user clicks, searches, and order history to maintain the proper representation of user behavior.
Through moving to this almost-real-time feature updates, the system reduces or takes away the latency between the actions of consumers and personalization outcomes, enabling recommendations to more quickly adapt to changing preferences within a session.
And, when it comes to the listwise ranking, several restaurant candidates will be evaluated together within a singular interference step, instead of individually. Under this approach, the technical model will optimize relative ordering across a range of options, instead of delegating independent scoring to each restaurant. In other words, this way, the technology can ascertain what’s really on your mind. When it’s pizza, but what you’re exactly thinking is pasta. Quite stunning!
Moreove, this system also builds upon a unified representation of user behavior, combining short-term session activity with longer-term historical signals, which are processed through a shared feature extraction layer, guaranteeing consistency between offline training and online serving. This generates training data by replaying historical user sessions to simulate production environments, reducing discrepancies between model training and live inference. Interesting.
“Personalizing a marketplace at this scale isn’t just about showing ‘good food’ – it’s about balancing real-time intent, diverse merchant ecosystems, and complex ranking objectives to create a seamless discovery experience,” Brinda Panchal, product specialist at Uber, added.
Uber Eats is the second-biggest food and grocery delivery app that lets consumers order meals from local restaurants, as well as items from convenience stores and supermarkets, having them taken and delivered directly to your doorstep by couriers. If this story sounds too technical for you, ask away by creating that account and starting a discussion. This is Ridesharing Forum.