Later this month, a group of researchers from the Drexel Howley College School of Computer and Information Sciences (SCIS) will attend the 20th ACM Conference on Recommender Systems (RecSys 2026). The Impact, Novation, Effectiveness, and Responsibility of Technology for Information Access Laboratory (INERTIAL) team, led by Michael Ekstrand, PhD, assistant professor of information science, will gather in Minneapolis with fellow researchers, take in the latest developments while presenting their own and make or renew connections in the field between September 28 and October 2.
Recommender systems, an application of machine learning, help users find relevant content or products, applying data gathered from a user’s previous engagement with a platform to display new suggestions without requiring a search for them.
Drexel’s team will present three papers, a poster and a tutorial at the sessions while also filling organizing and volunteer roles (with SCIS associates in bold on first mention):
- Main-track full paper, “On the Convergent Validity of Offline Evaluation Designs for Recommender Systems”, by SCIS PhD students Sushobhan Parajuli and Samira Vaez Barenji, with Ekstrand.
- Research and Practice Note (poster), “Monte Carlo Power Analysis for Small-System A/B Trials”, by Ekstrand (with external collaborators from POPROX, an experimental platform for recommendation research).
- Tutorial, “Continuous Automated Reproducibility: Practical Tools and Workflows for Reproducible Research Workflows” by Ekstrand.
- Paper at the RecTemp Workshop on Temporal Reasoning in Recommender Systems, “Book Readership During Movie Releases: An Exploratory Analysis”, by Parajuli, Vittoria Vineis (Sapienza University PhD student, currently a visiting scholar at Drexel), Vaez Barenji, and Ekstrand.
- Mallika Mainali (adviser: Rosina Weber, PhD), a SCIS PhD student unaffiliated with INERTIAL, is presenting “Toward Human-Aligned Recommender Systems via Cognitive and Behavioral User Modeling” in the Doctoral Symposium.
- Ekstrand is co-organizing the Second AltRecSys Workshop on Reimagining Recommender Systems Research & Practice.
- Ekstrand is the Reproducibility Track co-chair while also serving on the ACM RecSys Executive Committee.
- Parajuli and Vaez Barenji are student volunteers.
The team’s research produced new findings in recommender system evaluation and potential improvements for book recommendations, among others.
“Our full paper in the main track finds that widely used strategies for evaluating recommender systems with widely available data are at best weakly correlated with evaluations using higher quality, but more expensive data, and that the consistency between different strategies is not the same in different domains and datasets,” said Ekstrand. “This adds to the growing body of knowledge about the limitations of common evaluation practice.”
Additionally, the RecTemp workshop paper found “when a movie based on a book is released, the book sees corresponding increase in reader interest, but this pattern is not necessarily reflected in algorithmic book recommendations,” he said.
Work for the conference began with initial research and writing, followed by the team preparing and practicing to present. According to Ekstrand, additional duties for the committees varied depending on their purpose.
"In my role as Reproducibility co-chair, my co-chair and I were responsible for the Reproducibility Track: drafting the call for reproducibility and resource papers, recruiting the program committee to review them, overseeing the review process, and making the final accept/reject decisions for that track,” he said.
Whereas for the executive and steering committee, the work is mostly meetings about the current conference and decisions about future ones, Ekstrand said.
When the INERTIAL team returns to Drexel’s Howley College of Engineering and Computing from RecSys 2026, they already have plans to continue projects on “valid and reliable measurements of recommender systems, personalized news recommendation, and understanding how to make the best possible use of available data for both recommendation and evaluation,” Ekstrand said.