On component interactions in two-stage recommender systems

2021/06/28 by Jiri Hron, Hron, Jiri, Karl Krauth +5 · 2 citations
Computer Science · Decision Sciences · Mathematics · #Advanced Bandit Algorithms Research #FOS: Computer and information sciences #Information Retrieval (cs.IR) #Machine Learning (cs.LG) #Machine Learning (stat.ML) #Machine Learning and Algorithms #Recommender Systems and Techniques #cs.IR #cs.LG #stat.ML

paper · pdf · doi:10.48550/arxiv.2106.14979

Appears in the proceedings of the NeurIPS 2021 conference

openalex publication_date 2021/06/28 · arxiv created 2022/01/12 · arxiv updated 2022/01/14 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28

Abstract

Thanks to their scalability, two-stage recommenders are used by many of today's largest online platforms, including YouTube, LinkedIn, and Pinterest. These systems produce recommendations in two steps: (i) multiple nominators, tuned for low prediction latency, preselect a small subset of candidates from the whole item pool; (ii) a slower but more accurate ranker further narrows down the nominated items, and serves to the user. Despite their popularity, the literature on two-stage recommenders is relatively scarce, and the algorithms are often treated as mere sums of their parts. Such treatment presupposes that the two-stage performance is explained by the behavior of the individual components in isolation. This is not the case: using synthetic and real-world data, we demonstrate that interactions between the ranker and the nominators substantially affect the overall performance. Motivated by these findings, we derive a generalization lower bound which shows that independent nominator training can lead to performance on par with uniformly random recommendations. We find that careful design of item pools, each assigned to a different nominator, alleviates these issues. As manual search for a good pool allocation is difficult, we propose to learn one instead using a Mixture-of-Experts based approach. This significantly improves both precision and recall at K.

Citations

• Doubly Robust Policy Evaluation and Learning
• Evaluating collaborative filtering recommender systems
• Matplotlib: A 2D Graphics Environment
• Matrix Factorization Techniques for Recommender Systems
• Stochastic gradient boosting
• The Nonstochastic Multiarmed Bandit Problem
• Taming the Monster: A Fast and Simple Algorithm for Contextual Bandits
• Adaptive Mixtures of Local Experts
• Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer
• Hybrid Recommender Systems: A Systematic Literature Review
• Learning Tree-based Deep Model for Recommender Systems
• On the Difficulty of Evaluating Baselines: A Study on Recommender Systems
• PyTorch: An Imperative Style, High-Performance Deep Learning Library
• Offline Reinforcement Learning: Tutorial, Review, and Perspectives on Open Problems
• Adapting to Misspecification in Contextual Bandits
• Array programming with NumPy

Cited by

• Online Two-Stage Submodular Maximization
• LongRetriever: Towards Ultra-Long Sequence based Candidate Retrieval for Recommendation

Related

• Exploration in two-stage recommender systems
• Towards a Theoretical Understanding of Two-Stage Recommender Systems
• Learning over No-Preferred and Preferred Sequence of Items for Robust Recommendation (Extended Abstract)
• Learning over no-Preferred and Preferred Sequence of items for Robust Recommendation
• On the Generalizability and Predictability of Recommender Systems