Jeremy Schiff, Senior Manager, Data Science, OpenTable
summary
OpenTable’s recommender combines an IR search stack with collaborative filtering, serializes CF models for real‑time scoring and uses Spark for batch tuning and hyperparameter search; the network covers 32,000 restaurants and ~ $30B annual partner spend with ~16M seatings/month and ~450k reviews/month. Iteration is driven by A/B tests that take weeks for significance, offline training that takes days and analytics that take hours. Practical details include diversity constraints (avoid top‑10 all‑sushi), learning‑to‑rank, and text topics via non‑negative matrix factorization / LDA.
AI-written summary · about
chapters (17)
- 0:00Intro
- 2:44Building Recommendation Systems
- 3:01What's the Goal
- 4:27Measuring & The Iteration Loop
- 5:28Ranking Objectives
- 6:24Training, Generalization, and Online Error • Training Train on your specific dataset
- 7:35Fundamental Differences in Usage
- 8:46Recommendation Stack
- 9:24Using Context, Frequency & Sentiment
- 12:11Training Example
- 13:17Training DataFlow
- 14:16Compelling Recommendations
- 14:53Recommendation Explanations
- 16:18Summarizing Content
- 16:58Summarizing Restaurant Attributes
- 17:54Analyzing Review Content
- 18:50Topics with NMF using TF-IDF
related talks
35mData Science for Dinner: Recommending Recipes
27mOur experiments with food recommendations
39mRecommendations at XING: From theory to production
41mIntegrating Realtime User Behavioral Embeddings Into OpenSearch
31mCreating recommender systems with Spark, Scala and Prediction.Io
01h 48mRecommendation engines explained
18mAdvancing Search at Uber with OpenSearch
32mLeveraging recommender systems to personalise search results