Lessons Learned from Building Real­-Life Recommender Systems

推荐总结

2016Recsys会议分享

http://www.woshipm.com/pd/480396.html 

Xavier介绍了其在Quora和Netflix构建推荐系统时的10条经验：
(1) Implicit feedback is more important:有时需要将显示反馈与隐式反馈结合起来
(2) 认真考虑训练数据：定义合理的正负例
(3) 推荐的解释有时比预测准确率重要
(4) MF是最好的single approach，MF的变形包括FM，SVD++，ALS等;Quora开源了他们的MF模型QMF
(5) Ensemble is the master algorithm
(6) Feature Engineering: 需要理解领域信息，了解推荐目的和用户需求，将以上信息转化为特征；好的特征具有如下特点：reusable, transformable, interpretable, reliable; 深度学习也是构建特征的一种方法
(7) 合理的评价策略
i. Offline vs. online: a critical issue is how offline metrics correlate with A/B test results
ii. Long-term metric (member retention) vs. short-term metric 
(8) Model debuggability: to determine particular model to use; features to rely on; implementation of tools

推荐模型和技术
在推荐系统中被工业界广泛使用的有效的模型主要有：logistic regression，matrix factorization，Gradient boosting decision tree，其中LR被广泛用于CTR的预估；matrix factorization模型包括SVD++，factorization machine，ALS等方法，是工业很推崇的方法；今年RecSys challenge的前三名均使用GBDT的方法。其他被工业界使用的技术还包括：SGD，采样，哈希，L1&L2正则，实时流数据分析等；Spotify还声称他们用了MAB的方法。
YouTube Recommendation使用了深度学习技术，用于候选集的生成和排序

 值得关注的论文
1. A Scalable Approach for Periodical Personalized Recommendations
2. Adaptive, Personalized Diversity for Visual Discovery 
3. Field-aware Factorization Machines for CTR Prediction 
4. Local Item-Item Models for Top-N Recommendation  （Best paper）
5. Mechanism Design for Personalized Recommender Systems 
6. Deep Neural Networks for YouTube Recommendations 
7. Past, Present, and Future of Recommender Systems: An Industry Perspective （author：Xavier Amatriain）
8. Algorithms Aside: Recommendation as the Lens Of Life  （演讲的胶片非常艺术流）
9. Meta-Prod2Vec - Product Embeddings Using Side-Information for Recommendation
10. Are You Influenced by Others When Rating? Improve Rating Prediction by Conformity Modeling （余勇老师组做的工作）
工业界的几篇论文：
1. When Recommendation Systems Go Bad  (meetup)
2. News Recommendations at scale at Bloomberg Media: Challenges and Approaches (Bloomber)
3. Marsbot: Building a Personal Assistant (Foursqure)
4. Music Personalization at Spotify (Spotify)
5. Recommending for the World (Netflix)
6. The Exploit-Explore Dilemma in Music Recommendation  (Pandora)
7. Tutorial: Lessons Learned from Building Real-life Recommender Systems  (Xavier’ tutorial)