RecSys 2020: Takeaways and Notable Papers
Eugene Yan
RecSys 2020 conference (September 22-26) featured research emphasizing ethics, bias correction through inverse propensity scoring, and increased use of sequence models and reinforcement learning in recommender systems. Netflix research showed that recommendation effectiveness varies by placement type and user effort level, with 1:1 recommendations requiring 18% consecutive genre matching versus flexible context-dependent expectations. Offline evaluation methodology emerged as critical, with papers demonstrating that validation set creation and data splitting strategies significantly affect relative model performance, leading organizations to combine offline metrics with human evaluation before A/B testing.
Why it matters
Emphasis on bias, more sequential models & bandits, robust offline evaluation, and recsys in the wild.
