Things I Learned at the 2022 ACM Recommender Systems Conference in Seattle

Things I Learned at the 2022 ACM Recommender Systems Conference in Seattle

Last week I attended the 2022 ACM Recommender Systems (RecSys) conference in Seattle. My manager had reminded our team that Meta has a policy of being able to attend one conference per year (paid for by the company), and I was eager to learn more about recommender systems since I work on a recommendations infrastructure team for Facebook Marketplace. I was also very excited to have the opportunity to meet more people in industry after not having attended business conferences for 3 years.

Let's start with some photos.

There were around 1200 conference attendees this year, which is the most RecSys has ever had. 500 out of the 1200 were remotely attending, and around 30-40% of the speakers were remote and gave presentations over Zoom.

Here are some of the things that I learned this year:

[Nvidia] Building and Deploying a Multi-Stage Recommender System with Merlin

  • Nvidia has open-sourced a library for end-to-end GPU-accelerated recommender systems. Github: https://github.com/NVIDIA-Merlin/Merlin
  • “Each stage of the Merlin pipeline is optimized to support hundreds of terabytes of data.”
  • I wasn’t sure how similar their recommender system design is to our system, but from a high level I see how they also have retrieval, filtering, scoring, and ordering (a.k.a. ranking).
image.png

[Eindhoven University of Technology, Netherlands] Exploring the longitudinal effect of nudging on users’ genre exploration behavior and listening preference

  • This was more of a psychology / consumer behavior study, and I found the results very interesting!
  • Users of a music genre exploration tool are presented with an initial setting to have more personalized recommendations (similar to their listening history) or more exploratory recommendations (further from their listening history).
  • The users who had the initial setting nudged towards more exploratory recommendations influenced their preferred personalization level.
  • Perceived helpfulness of the recommendation system increased when users explored further away from their current preferences.
  • I found this very intriguing because Marketplace also has some type of balance between recommending items that users have recently viewed, and Marketplace also shows some exploratory recommendations. This paper might have some findings for our product teams.