Scaling up Lyft’s Growth via Causal ML and Bandits
summary
Lyft builds a growth stack that predicts and optimizes incremental impact: HTE models estimate per-user lift from coupons, a linear program allocates a budget to maximize incremental rides under cost constraints, and contextual bandits handle exploration (epsilon‑greedy and Thompson sampling). Deployment of a Dragonet‑style net and bandits yields ~20–50% efficiency gains versus random allocation, but deep models show high variance (AA tests differ ~10%), so off‑policy evaluation shortens feedback from months to days.
AI-written summary · about
chapters (7)
- 0:00Lyft’s growth challenges
- 3:16Supply and demand balancing
- 8:15Growth platforms and incentives
- 14:12Heterogeneous treatment effects
- 17:18Bandits for exploration
- 24:03Scaling causal ML lessons
- 26:14Q&A
related talks
32mWhy start using uplift models for more efficient marketing campaigns
32mAccelerating autonomous vehicle solutions with SageMaker, featuring Lyft
24mUnlocking revenue: newcomer classification for boosted profits
21mRise of the Machines: Removing the Human-in-the-Loop
39mGrowing user engagement with RL-driven personalisation
21m"Capacity Prediction" instead of "Capacity Planning":
21mLeveraging Multi-Armed Bandit Algorithms for Dynamic Decision Making
28mEnabling Machine Learning with Apache Flink