A Bayesian Approach to the Cold Start Problem
summary
Bayesian binomial modeling estimates CTR for cold‑start listings so the ranking system gets enough signal before the reranker takes over. It treats impressions as trials, clicks as successes, uses weak priors because experiences are not comparable, and updates posteriors until confidence intervals or information value trigger a stop. Implementation runs on preaggregated daily tables in S3, processed with PySpark UDFs, Databricks, Airflow and dbt across 35,000 suppliers and 150,000 experiences.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 1:32New experiences on travel platforms
- 8:29Cold start problem in recommendations
- 12:16Framing the problem correctly
- 15:50Bayesian vs ML modeling
- 21:34Why Bayesian modeling fits
- 22:16Binomial Bayesian approach for CTR
- 27:08Key parameters in Bayesian modeling
- 31:03Bayesian statistics at scale
- 33:51Three takeaways
- 34:54Q&A
related talks
35mLearning to Judge
25mScaling Bayesian Experimentation
33mTwo Years of Bayesian Bandits for E Commerce
17mByron Galbraith, Chief Data Scientist, Talla, NYC 2017
37mRealtime Bayesian A-B testing with Spark Streaming
44mRobust AI Search Ranking for Radical C2C Marketplace Growth
31mLarge Scale CTR Prediction Lessons Learned
37mSupercharging CTR (Click-Through Rate) with Plug-and-Play AI Capabilities