DataEngBytes 2023 - PER-T1-06 - Kamal Hossain
summary
Built a realtime recommendation pipeline that addresses listings changing by the minute and users whose preferences shift in seconds, because embedding models take 3–5s model inference while the site requires <500ms frontend SLA. The system precomputes candidates, prefetches and caches results in a real-time DB cache (10–15ms reads), filters out expired items with an Elasticsearch active-listing filter, and balances exploitation with exploration by injecting one or two random items into 4–5 card recommendations, yielding a notable lift in click-through.
AI-written summary · about
chapters (9)
- 0:00Real-time recommendation systems
- 2:09Personalized property recommendations
- 5:16Long tail problem
- 7:22Real-time property challenges
- 10:10Sequence modeling for preferences
- 13:01Implementation challenges
- 16:53Real-time response and expiration
- 18:36Exploitation vs exploration
- 19:41Conclusion
related talks
12mQuickUpdate: a Real-Time Personalization System for Large-Scale Recommendation Models
24mLarge Scale On-Demand Low-Latency Near Real-Time Predictions
27mDataEngBytes 2023 - SYD-T1-10 - Kanishka Mohaia
33mPowering Real-time Recommendations with Machine Learning on GCP
41mBig Data Serving: The Last Frontier. Processing and Inference at Scale in Real Time
27mEnhancing Personalization Using Recommendation Systems & Multi-Armed Bandits
34mThe Billion-Request Content Recommendation System Challenge
38mEmbeddings! Embeddings everywhere!