Yuanchi Ning - UberEats Discovery:Food Recommendation

Cartel

2017/12/18 发布于 技术 分类

ArchSummit全球架构师峰会是InfoQ中国团队推出的面向高端技术管理者、架构师的技术大会,参会者中超过50%拥有8年以上的工作经验。 ArchSummit秉承“实践第一、案例为主”的原则,展示新技术在行业应用中的最新实践,技术在企业转型中的加速作用,帮助企业技术管理者、CTO、架构师做好技术选型、技术团队组建与管理,并确立技术对于产品和业务的关键作用。

文字内容
1. How AI Powers Yuanchi Ning (Uber Eats) December 2017
5. Agenda ●  Uber Eats Overview ●  AI Platform ●  AI Challenges ○  Challenges as a marketplace ○  Challenges of Uber Eats discovery ■  Restaurants ranking and recommendation ■  Guided Exploration
6. Uber Eats Overview
7. Make eating well effortless at anytime, for anyone. Uber Eats mission
9. Uber Eats Timeline ●  March, 2009: Uber founded ●  August, 2014: UberFRESH launched in LA ●  April, 2015: UberFRESH rebranded to Uber Eats ●  December, 2015: Uber Eats is spun off into a separate standalone app and launched in Toronto ●  March, 2016: Uber Eats launched in LA, Chicago, Houston, and SF ●  Today: Uber Eats launched in 200+ cities, 30+ countries, and 6 continents
11. AI Platform
13. Feature Report
14. Model Accuracy Report
15. AI Challenges with Uber Eats
16. Uber Eats as a Marketplace Make eating well effortless at any time, for anyone Eaters Restaurant Owners Pick up Delivery Partners
17. AI & Uber Eats ●  Restaurant ranking and recommendation (Homepage feed) ●  Guided exploration (Search) ●  Demand-supply forecasting ●  ... Restaurant Owners Eaters Pick up Delivery Partners
18. AI & Uber Eats ●  Restaurant ranking and recommendation (Homepage feed) ●  Guided exploration (Search) ●  Demand-supply forecasting ●  ... Eaters Restaurant Owners Pick up Delivery Partners ●  Courier positioning ●  Dispatch ●  Batching
19. AI & Uber Eats ●  Restaurant ranking and recommendation (Homepage feed) ●  Guided exploration (Search) ●  Demand-supply forecasting ●  ... Eaters ●  Time prediction (estimated time of delivery) ●  Dynamic pricing ●  Intelligent spending ●  ... Restaurant Owners Pick up Delivery Partners ●  Courier positioning ●  Dispatch ●  Batching
20. Today’s Discussion: Eats Discovery ●  Restaurant ranking and recommendation (Homepage feed) ●  Guided exploration (Search) ●  Demand-supply forecasting ●  ... Eaters ●  Time prediction (estimated time of delivery) ●  Dynamic pricing ●  Intelligent spending ●  ... Restaurant Owners Pick up Delivery Partners ●  Courier positioning ●  Dispatch ●  Batching
21. Restaurant Ranking And Recommendation
22. A Few Unique Challenges ●  Ranking to serve the marketplace ●  Relevance vs. diversity ●  Building a fair marketplace ●  ...
23. Ranking to Serve the Marketplace ●  Conventional ML Model ○  Single objective ■  Keep users or ■  Keep restaurant owners ○  GBDT, RankSVM
24. Ranking to Serve the Marketplace ●  Conventional ML Model ○  Single objective ■  Keep users or ■  Keep restaurant owners ○  GBDT, RankSVM Efficient frontier Conversion Rate ●  Solution: Multi-Objective Optimization ○  Multiple objectives ■  Keep users and ■  Keep restaurant owners ○  Linear / Quadratic Programming (LP/QP) 1 - Restaurant Churn Rate
25. MOO: Multi-Objective Optimization is the ML/AI model for the kth objective For example: is the conversion rate ML/AI model is the 1 - restaurant churn rate ML/AI model Challenge is to formulate the above problem as convex optimization problem (LP / QP) Conversion Rate Efficient frontier 1 - Restaurant Churn Rate
26. MOO Example: Relevance vs. Diversity ●  Pointwise ranking is greedy ●  Listwise ranking is costly ●  Holistic ranking ○  Estimate CTR of each restaurant with an ML Model ○  Optimize the ranking of all restaurants holistically given estimated CTR ML Model LP / QP Optimizer Final Ranking Estimated CTR Tradeoff scores
28. Explore-Exploit with Multi-Armed Bandit ●  Bayesian modeling for posterior variance ○  New /low-volume restaurant - high variance ○  Well-established restaurant - low variance ●  Multi-armed bandit ○  ML model to estimate the mode of conversion rate ○  Bandit algorithm for explore-exploit
29. Guided Exploration (Search) Item listicle Search results
30. Challenges ●  Understand user query and our food ○  Restaurant ○  Dish types ○  Cuisine types ●  No results / low results ○  Not on the platform ○  Out of delivery radius / time ●  Ranking ○  Personalized - but not so much
31. AI/ML Solutions ●  Understand user query and our food - Representation Learning ○  Restaurant ○  Dish types ○  Cuisine types ●  No results / low results - Food Knowledge Graph ○  Not on the platform ○  Out of delivery radius / time ●  Ranking - ML/AI models ○  Personalized - but not so much
32. Food Graph ●  Chipotle ○  Is it a restaurant? ○  Fast food? ○  Sells burritos? ○  Similar restaurants? ●  Poke ○  Is it a cuisine? ○  Similar cuisines? Illustration credit: Ting Chen @ Uber, GraphDB
33. Representation Learning ●  Food graph-based ●  Latent space-based ○  Word2Vec, GloVe ○  End-to-end deep neural network
34. Ranking ●  Personalized model? ●  Closed/missing restaurants ●  In-menu search and ranking
35. Takeaways ●  Uber Eats is a marketplace for eaters, restaurant owners and delivery partners. ●  AI is the underlying engine that runs this marketplace.
36. Thank you and bon appétit Q&A
37. Proprietary and confidential © 2017 Uber Technologies, Inc. All rights reserved. No part of this document may be reproduced or utilized in any form or by any means, electronic or mechanical, including photocopying, recording, or by any information storage or retrieval systems, without permission in writing from Uber. This document is intended only for the use of the individual or entity to whom it is addressed and contains information that is privileged, confidential or otherwise exempt from disclosure under applicable law. All recipients of this document are notified that the information contained herein includes proprietary and confidential information of Uber, and recipient may not make use of, disseminate, or in any way disclose this document or any of the enclosed information to any person other than employees of addressee to the extent necessary for consultations with authorized personnel of Uber.