2026-09-10 –, Unconference
New items in a marketplace have no behavioral signals. Ranking models ignore them. They never get booked. They never get signals. The loop never breaks.
This talk is a three-year, experiment-driven journey to break that loop: what we tried, what failed, and what ultimately worked. You'll leave with a reusable blueprint for solving cold start in any ranking or recommendation system.
Cold start is a structural trap in two-sided marketplaces. New items lack behavioral signals, so ranking models under-expose them, which delays the very signals needed to rank them well. Left unaddressed, this feedback loop suppresses new inventory, weakens supplier trust, and degrades long-term marketplace health.
This talk presents our journey to break that loop in a large-scale travel marketplace with 200k+ activities. Over three years, we evolved from a brittle fixed-slot exposure system to a fully integrated ranking framework. The path was not linear: several of our most intuitive ideas failed.
We'll walk through three years of experiments: what we tried, what failed, and what worked. Along the way, we share how we framed trade-offs between short-term revenue and long-term marketplace health, and how reframing the core question unlocked the right metric, the right business case, and the roadmap.
You'll leave with three transferable lessons:
- Explore fast: on new problems, small experiments beat big designs
- Constraints are hypotheses: giving the model more freedom consistently outperformed restricting it
- Change the question, change the outcome: the metric you optimise for determines what solutions become visible
Theodore Meynard is a data science manager at GetYourGuide.He leads the evolution of their ranking algorithm, helping customers to find the best activities to book and locations to explore. Beyond work, he is one of the co-organizers of the Pydata Berlin meetup and the conference. When he is not programming, he loves riding his bike and looking for the best bakery-patisserie in town.