Mobile Order Click-Through Rate (CTR) Recommendation with Ray on Apache Spark at Burger King
summary
Burger King deploys a Transformer Cross Transformer that runs a sequence Transformer (current-session token plus averaged purchase-history token) and a context Transformer (location, time, weather, device) whose outputs are combined by an elementwise latent cross then fed to a classifier. Item embeddings are initialized from pretrained BERT and ResNet, compressed with a VAE, and training runs on a unified pipeline using Ray on Apache Spark with MXNet (in-memory plasma transfers, Orca/Analytics Zoo). Production A/B shows +7.5% conversion and +4.7% add-on sales; model is slightly slower but operationally simpler than a separate GPU cluster.
AI-written summary · about
chapters (11)
- 0:00Introduction and overview
- 1:03Mobile order CTR challenges
- 7:04Transformer Cross Transformer model
- 10:32Item embedding initialization
- 13:21Model comparisons and results
- 15:07Online inference performance
- 16:52Training pipeline with Ray on Spark
- 17:13Intel's AI on big data work
- 21:47Rayon Spark implementation
- 26:38Burger King's production use
- 28:02Wrap-up and Q&A
related talks
37mSupercharging CTR (Click-Through Rate) with Plug-and-Play AI Capabilities
49mUnderstanding Customer Journeys with Deep Learning
32mHow Uber Optimize Marketplaces with Ray
39mGrowing user engagement with RL-driven personalisation
31mLarge Scale CTR Prediction Lessons Learned
39mAI-Powered Search in Modern E-Commerce
41mText to Context: How We Introduced Hybrid Search
27mTransform online shopping experiences with personalized recommendations