Problem
Lyft offers rideshare, bikes, scooters and rentals. In large markets a rider can see 8–10 options at once and book the wrong one by accident. New products such as Wait & Save and Shared got almost no visibility above the fold, because there was no data to rank them. Conversion is not the only goal. Ride reliability and the balance between rider demand and driver supply matter too.
Example from the article: in 2019 the mode selector preselected the user's last mode, in 2020 the most frequently used mode, and in 2022 the mode with the highest predicted propensity.
Solution in brief
Machine learning propensity models predict the chance that a rider converts into each mode and rank the list by it. The top 3–4 modes are shown above the fold, and the top one is preselected. A post-processing layer adjusts the model output to promote new modes and to serve marketplace goals.
Questions this case answers
- How to rank many product options per user by predicted conversion when the options differ in business value
- How to surface a newly launched option in a ranking when the model has no training data for it
- How to balance a conversion model with supply and demand goals in a two-sided marketplace
- How to rank a long list of product options for each user so that the right one is visible above the fold?
Tools and methods
- LightGBM — gradient-boosted model, each mode is a class, weights come from mode-specific financial metrics
- Learning to rank (lambda rank) or multi-class classification — objective, chosen per use case
- Propensity modeling — probability of converting into each mode, per rider and context
- Post-processor — rule layer on top of model scores for cold start and business goals
Google or a one-time email code. No password.