When you hit the road and open Waze, a lot is happening behind the scenes. This lecture focuses on the ads served in the online geo space based on routes, map, and venue data using deep learning models in production. Serving ads in the online geo space along routes is a unique problem to online navigators. Today, Waze is the only navigator to serve these types of ads with unique targeting options such as destination types and proximity to routes in different locations. The main challenge is how to predict where the user will be, and what they will see on their screen, while at the same time accounting for different advertisers’ goals and user relevance.

This lecture explains how the map and the route data is structured and used while considering advertisers’ budgets and user relevance.

Starting with structuring the data by breaking down the routes and the map into small segments and storing it in a data mesh architecture, a deep learning model then predicted the display probability of a venue given the remaining route, considering the visible screen boundaries at each segment. Finally, this prediction was combined with budget pacing and user relevancy to form one score.
Testing such logic changes in the serving mechanism is another challenge. The common user split doesn’t hold, because it would split advertisers’ budgets that are shared across control and test groups, creating data leakage. This lecture will provide the in-house solution that was designed to overcome this challenge.

Attendees will get the opportunity to explore the unique maps and routing data at Waze and learn how such complex data is structured and used in ML models in production and how it drives significant impact to Waze’s business metrics