Introduction
Scheduling a large, multi-track academic conference is a logistical nightmare of conflicting preferences. For events like ACM CHI, which involve hundreds of papers and thousands of attendees across dozens of parallel tracks, the goal is to create thematically coherent sessions while ensuring that papers an attendee wants to see aren't scheduled simultaneously. Traditionally, this was a manual process prone to error and limited by the expertise of a few chairs.
The Challenge of Traditional Scheduling
A previous innovation, "author-sourcing" (via the Cobi system), engaged authors to identify papers similar to theirs. However, author-sourcing suffers from a "searchlight" problem: authors are typically only shown a small, pre-selected list of potentially similar papers. This approach misses connections across different sub-fields and ignores the interests of the thousands of non-author attendees who constitute the majority of the audience.
Methodology: Attendee-Sourcing via Confer
This paper introduces attendee-sourcing, a method that leverages the collective interest of the entire conference community to discover hidden paper affinities and minimize scheduling conflicts. To implement this, the researchers built Confer, a paper recommendation and personal scheduling tool.
Accepted papers were released on Confer 45 days before the official schedule announcement. Attendees were invited to explore the catalog and "star" papers they intended to see. Unlike the static lists in author-sourcing, Confer used a recommendation engine (collaborative filtering and TF-IDF) to help users discover relevant work. As users starred more papers, the system suggested more relevant ones, creating a feedback loop of discovery. These "stars" were then converted into affinity scores; if many people starred both Paper A and Paper B, it suggested a high affinity, indicating they should be in the same session or at least not scheduled concurrently.
Comparing Attendee and Author Sourcing
The study compared data from CHI 2014, involving 634 authors and 347 Confer participants. While author-sourcing provides "high precision" data—authors know exactly what is technically related to their work—attendee-sourcing provides "high recall." Attendees found 150 pairs of papers with strong affinity that authors missed entirely because the papers belonged to different sub-communities.
Attendee-sourcing also provides a crucial "popularity" metric. If a paper is starred by hundreds of people, organizers can proactively move it to a larger room to avoid overcrowding.
Key Results
Despite having fewer participants than the author-sourcing group (347 vs. 634), attendee-sourcing generated nearly double the total preferences (7,228 vs. 3,869). On average, an attendee marked approximately 21 papers of interest, whereas authors only indicated about 6 related papers. The attendee-sourcing method covered 98.98% of all accepted papers, compared to 91.14% for author-sourcing.
Conclusion and Future Work
Attendee-sourcing produces data that provides more alternatives in scheduling and complements existing methods for creating coherent sessions. While attendee data can be noisy—influenced by "Best Paper" awards or social media promotion—it offers a much broader and denser map of paper relationships. Future work aims to move toward "Automated Session Creation," where algorithms use attendee-sourcing data as an optimization objective to generate conflict-free schedules entirely from scratch.
Summary updated