Global Joker · Delivery Hero · 2019–2024
Making Joker offers more relevant
The problem
Not every offer fits every user
Joker could show users multiple restaurant offers, but not every restaurant was equally relevant to every user. A generic ranking meant users could spend their limited offer window browsing options they were unlikely to choose.
I was the product manager, working with data science, engineering and local market teams.
The product question
How can we show each user the restaurants they are most likely to order from?
Before ML
How ranking worked before
Before machine learning, restaurants were ranked with simple rules called sorters. Each sorter ranked restaurants by one thing: a user’s favorite cuisines, the most popular restaurants, what people usually order at that time of day, or how often a restaurant’s offers turned into orders. Markets could choose which sorters to use and how much weight each one got.
The problem: each rule only looked at one signal at a time, and finding the right mix took a lot of manual testing.
From rules to personalization
Starting with signals we already had
We started with behavioral signals already available in the platform. A user’s historical orders were grouped by cuisine over a six-month period, so restaurants matching their favorite cuisines could rank higher. Restaurant popularity and order volume were later added as extra ranking signals.
Moving to ML
From one signal to many
As Joker scaled, ranking evolved from predefined sorting rules into a machine learning approach. Instead of relying on a single signal, the ML ranker combined many signals to estimate how relevant each restaurant was for each offer.
Going beyond ranking
Deciding what to show at all
ML wasn’t only used to change the order of restaurants. Joker also explored using the predicted likelihood of accepting an offer to decide which restaurants should be shown at all.
A user with a low likelihood of accepting a Joker offer would only see restaurants with a high predicted acceptance probability, while a highly engaged user could see a broader selection.
Experiment, learn, iterate
A product hypothesis, not just a model
ML ranking was treated as a product hypothesis rather than only a technical implementation. Different combinations of ranking signals were tested against existing approaches and measured on conversion and revenue.
The ML ranker kept evolving through several versions, with later versions adding more contextual and behavioral signals. Initial A/B tests of the latest version reported up to 6% revenue uplift.