Winning the Game: Leveraging Contextual Bandits for Creative Optimization
summary
Argues that replacing slow A/B tests with contextual bandits (greedy, Thompson sampling, UCB) routes impressions to better creatives in real time, using features like predicted install/purchase, creative-object tags extracted with OpenCV, and app-store text embeddings. Shows retargeting work that cuts cost-per-reactivation and a case (Alien Invasion) that recoups spend within three months. Simulated Privacy Sandbox topics recover roughly 60% of device-ID signal for cohort targeting.
AI-written summary · about
chapters (11)
- 0:00Creative optimization with bandits
- 1:14Why traditional AB testing falls short
- 3:38Bandits solve explore-exploit
- 6:07Three bandit algorithms
- 9:17What makes bandits different
- 10:20Better features for selection
- 12:04Retargeting gaming apps case
- 13:10Alien Invasion ID case
- 13:50Privacy Sandbox prep
- 15:21Bandit platform tips
- 16:45Q&A
related talks
41mKeynote: Personalization, Bandits, and Causal inference
19mSession: Re-Engagement In 2026: How To Design & Scale Winning Retargeting Campaigns
16mThrough the looking glass – Following the rabbit down the data… Whole
15mAI-Driven Application Optimization
39mGrowing user engagement with RL-driven personalisation
21mDemystifying Machine Learning for Programmatic UA & Retargeting
22mPersonalizing Explainable Recommendations with Multi-objective Contextual Bandits
45mContextual Multi-Arm Bandit and its applications to digital experiments