A curated roundup of notable papers from RecSys 2021, covering topics such as higher-order collaborative filtering extensions to EASE^R, a revisited comparison of matrix factorization vs. neural collaborative filtering (MF still wins on accuracy), a serverless open-source recommender deployment stack on AWS, the Transformers4Rec library for session-based recommendations, best practices for operating large-scale recommender systems (RecSysOps), semi-supervised fashion compatibility modeling, and cold-start handling via shared item embeddings. Key takeaways include the continued competitiveness of simple baselines over deep learning models and the value of data quality over model complexity.
