Recommender Systems Reading Guide

Purpose

These notes are organized around the shape that most production recommenders eventually take. First define the objective. Then retrieve a tractable candidate set. Then rank those candidates with a richer model. Then measure what happened without lying to yourself about feedback loops. The section is built around that order.

Reading Path

If I wanted to rebuild the area from scratch, I would read it in this order:

1. Recommender Systems
2. Ranking Objectives and Implicit Feedback
3. Retrieval and Ranking
4. Two-Tower Retrieval
5. Wide and Deep
6. Sequential and Graph Recommenders
7. Bias, Marketplace Effects, and Counterfactual Evaluation

After that, the paper notes land better:

• Deep Neural Networks for YouTube Recommendations
• Practical Lessons from Predicting Clicks on Ads at Facebook

Notes

• Foundations: recommender systems
• Labels and losses: ranking objectives and implicit feedback
• Pipeline shape: retrieval and ranking
• Retrieval architecture: two-tower retrieval
• Retrieval architecture and experiment: two-tower retrieval with MovieLens 100K
• Ranking architecture: wide and deep
• Temporal and structural signals: sequential and graph recommenders
• Evaluation and ecosystem effects: bias, marketplace effects, and counterfactual evaluation
• Data infrastructure: MapReduce and Spark
• Production case studies:
  • YouTube recommendations
  • Facebook ads CTR

Related notes

• Two-Tower Retrieval and the MovieLens 100K experiment

Sources

• Covington, Adams, and Sargin (2016), Deep Neural Networks for YouTube Recommendations
• Cheng et al. (2016), Wide & Deep Learning for Recommender Systems
• Hu, Koren, and Volinsky (2008), Collaborative Filtering for Implicit Feedback Datasets