To best understand how we made our Uber Eats recommendations more accurate, it helps to know the basics of how graph learning works. Many machine learning tasks can be performed on data structured as graphs by learning representations of the nodes. The representations that we learn from graphs can encode properties of the structure of the graph and be easily used for the above-mentioned machine learning tasks. For example, to represent an eater in our Uber Eats model we don't only use order history to inform order suggestions, but also information about what food items are connected to past Uber Eats orders and insights about similar users.

For eaters, our system offers personalized restaurant recommendations, but ultimately eaters are looking for specific dishes to order. So, we are working on taking our recommendations to the dish level, creating more tailored eater experiences. This is analogous to the music industry's shift from selling albums to selling songs, and we believe it will be a huge leap forward in terms of the experience we can provide. In addition, for new eaters that are checking out the platform, we are working on methods to bootstrap our recommendations and solve the cold start problem often seen in recommender systems. For restaurant-partners, we are working to balance the surfacing of promotions and deals offered to eaters, as these short-term initiatives create interesting effects on the system by changing the behaviors of eaters who respond to them.