Reproduce “A Comprehensive Survey on Retrieval Methods in Recommender Systems” on OpenResearch

About this paper

Junjie Huang, Jizheng Chen, Jianghao Lin, Jiarui Qin, Ziming Feng, Weinan Zhang, Yong Yu

Researchers from Shanghai Jiao Tong University and China Merchants Bank provide the first comprehensive survey of retrieval methods in recommender systems, establishing a unified taxonomy, empirically benchmarking diverse algorithms, and detailing real-world industrial practices through a case study. The work highlights the evolution from shallow to deep learning-based techniques and addresses challenges in large-scale candidate generation.

Key insights

• The optimal retrieval method is highly context-dependent, with no single model consistently outperforming others across diverse datasets and application domains.
• Deep learning-based Embedding-Based Retrieval (EBR) paradigms, like Two-Tower, Autoencoder, and GNN-based methods, are increasingly crucial for learning comprehensive user/item representations and handling complex interaction patterns.
• Effective negative sample construction and efficient indexing mechanisms (e.g., ANN, tree-based) are critical for optimizing training and enabling scalable, low-latency retrieval in large-scale industrial systems.

Summary by alphaXiv.