Enhancing Personalization Using Recommendation Systems & Multi-Armed Bandits
summary
Wargaming builds an offline recommendation pipeline that retrieves candidates with a lightweight recall model then re-ranks with a heavier ranker, uses contextual multi-armed bandits for quick starter‑pack choices under 30–60s latency, and separates upstream user embeddings from downstream models. Data ingestion switches to incremental delta processing giving 100x faster data processing. Deployed on AWS with Snowflake and message brokers, A/B tests show 45% revenue uplift for high‑tier players, ~20% lift in repeat purchases and ~15% overall. Implementation fits into two‑ to three‑week sprints.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 1:08WarGaming personalization goals
- 3:01Recommendation system types
- 7:50Recommendation workflow design
- 13:41Data filtering constraints
- 15:50Multi-armed bandits for conversions
- 19:28Bandit system design
- 20:57Results revenue uplift
- 22:26Q&A
related talks
32mMulti-Armed Bandits in Recommendations
12mQuickUpdate: a Real-Time Personalization System for Large-Scale Recommendation Models
28mThe Power of Personalized Customer Experiences Through AI Innovations
41mRecommender Systems: Crafting Effective Personalised Experiences
38mRecommendations at Rappi: from MVP to Data Product
29mThe Challenges of Building a Personalized Recommendation System in E-Commerce
27mTransform online shopping experiences with personalized recommendations
40mNext Best Action: Maximizing Customer Value