Presentation: Time Predictions in Uber Eats

I’m currently leading the time prediction area for UberEats. As you can imagine, precise predictions are keys for the system’s efficiency and reliability. The features I’m working on are like predicting how long it will take to deliver the food to the eater, how long will the restaurant spend to prepare the food, etc. Before this, I spent most of my time working on our intelligent dispatch system. Those two things are coupled tightly together, for example, we need to know when the food will be ready before sending the delivery partner to the restaurant to pick it up.

I’m very excited to share how we are using ML to tackle problems, especially in O2O business model since we are pioneers in that industry. We have a three-sided marketplace including delivery partners, restaurants, and eaters. In every decision we make we have to consider all of them in terms of how we are doing the tradeoffs and how we are using ML to figure out the optimal solution. For example, the hardest thing for us to make precise food preparation time prediction is that we don't have the ground truth. A restaurant does not have the incentive or responsibility to tell us how long an order will take to be prepared. But that's the most critical thing for us to figure out the perfect dispatch timing.

People who are interested in solving similar problems with ML in O2O business since it’s not commonly shared across the industry and we figured out a lot the solutions all from scratch.