Key Takeaways from RecSys 2020: Conference Organization and Notable Research Highlights

Conference Organization Highlights

The RecSys 2020 conference, originally planned for Brazil, moved online due to the pandemic, which unexpectedly improved accessibility for participants worldwide. Organizers worked around the clock, offering multiple live sessions per paper to accommodate different time zones, using the Whova platform for scheduling, live streaming, Q&A, and community interaction, and employing Gather.town to host virtual poster sessions and informal discussions.

Industrial Research Highlights

PURS: Personalized Unexpected Recommender System for Improving User Satisfaction – Developed by a NYU doctoral student in collaboration with Alibaba, this system addresses filter‑bubble issues, showed a 4.6% increase in average watch time in A/B tests on the Youku app, and incorporates personalized exploration.

Behavior‑based Popularity Ranking on Amazon Video – Presented by an Amazon Video applied scientist, this work uses machine‑learning to predict popularity, mitigating cold‑start problems by leveraging content text, historical popularity, and user interaction features such as the “Age” (time on platform) feature.

Query as Context for Item‑to‑Item Recommendation – From ESTY.COM, this approach builds candidate sets using Word2Vec embeddings of user‑clicked queries and retrieves similar items with Faiss, followed by LightGBM re‑ranking, while exploiting contextual cues (e.g., seasonal themes) to personalize recommendations.

Academic Research Highlights

Counterfactual Learning for Recommender System – A Huawei Noah’s Ark Lab principal researcher introduced Uniform Unbiased Data collected from 1% random traffic, enabling unbiased offline evaluation and achieving a 3% lift in key metrics.

Debiasing Item‑to‑Item Recommendations with Small Annotated Datasets – A Microsoft Research scientist demonstrated that a modest amount of labeled data can effectively correct bias in item‑to‑item recommendation models, releasing both data and source code.

Large‑scale Open Dataset for Bandit Algorithms – Presented at the REVEAL 2020 workshop, this dataset provides uniform‑rank and Bernoulli‑rank generated shopping behavior logs for evaluating offline policy evaluation methods, accompanied by the Open Bandits Pipeline codebase.

Conclusion

The virtual RecSys 2020 experience exceeded expectations, offering high‑quality presentations and valuable engineering insights despite initial concerns about online delivery. The highlighted works demonstrate practical advances in recommendation systems, spanning both industry deployments and cutting‑edge academic research, and set the stage for future innovations.