Growing user engagement with RL-driven personalisation
summary
Tesco builds a real‑time personalization pipeline using contextual bandits over autoencoder user embeddings, implementing a LinUCB policy with 15-minute batch updates and a random baseline. A six‑week A/B test on one grocery homepage tile shows a reproducible 30–40% uplift in clicks versus random after the model shifts from exploration to exploitation. Experiments sweep alpha (0.01 to 5), reveal a weaving artifact from batched updates, and highlight cold‑start, placement and context limitations that drive further simulation and self‑tuning work.
AI-written summary · about
chapters (13)
- 0:00RL personalisation at Tesco
- 1:07Tesco’s data science evolution
- 3:36Content orchestration importance
- 6:48Problem statement and objectives
- 7:50Contextual Bandits for personalisation
- 9:56User context with autoencoder embeddings
- 12:08LinUCB algorithm mechanics
- 15:39Real-time personalisation infrastructure
- 16:01A/B test setup and baseline
- 18:50Results and engagement improvements
- 20:15Alpha impact on exploration
- 23:47Next steps and research
- 25:52Q&A
related talks
41mKeynote: Personalization, Bandits, and Causal inference
20mWinning the Game: Leveraging Contextual Bandits for Creative Optimization
27mTransform online shopping experiences with personalized recommendations
27mEnhancing Personalization Using Recommendation Systems & Multi-Armed Bandits
22mPersonalizing Explainable Recommendations with Multi-objective Contextual Bandits
12mQuickUpdate: a Real-Time Personalization System for Large-Scale Recommendation Models
21mDemystifying Machine Learning for Programmatic UA & Retargeting
27mHow we personalized onet.pl with multi-armed bandits