Scheduling a massive, multi-track conference is a logistical nightmare often prone to "thematic conflicts"—where two talks an attendee wishes to see happen at the exact same time. This research introduces attendee-sourcing, a method that leverages a paper-recommendation tool called Confer to gather thousands of attendee preferences months before the event. This data allows organizers to build schedules that maximize satisfaction and minimize room-size miscalculations.
The Challenge of Large-Scale Scheduling
For flagship conferences like ACM CHI, which features hundreds of papers across 16 parallel tracks, the scheduling process is traditionally a manual, "paper-on-the-wall" exercise for a few organizers. Even with digital tools, organizers struggle to account for the diverse interests of thousands of attendees.
Previous attempts to solve this, such as the Cobi system, relied on "author-sourcing"—asking authors to identify papers similar to their own. However, authors are often biased or limited by a narrow view of the field. The gap addressed here is the "unvoiced opinion" of the general attendee. Without their input, organizers often end up with incoherent sessions or "rockstar" talks scheduled in tiny rooms while niche sessions sit in half-empty auditoriums.
Methodology: The Confer Tool
The authors developed and deployed Confer, a social recommendation and scheduling tool designed to capture attendee preferences through:
- Early Access: Releasing the full list of accepted papers to attendees four months before the conference (and 45 days before the schedule is finalized).
- Natural Exploration: Unlike authors, who judge a fixed list of papers, attendees explore the entire dataset via search, keywords, and collaborative filtering recommendations.
- Preference Encoding: Every time an attendee "bookmarks" a paper they want to see, it is recorded as a preference.
- Affinity Mapping: The system calculates "affinity scores" between pairs of papers based on how many people wanted to see both. High affinity suggests papers should be in the same session or at least not scheduled concurrently.
Key Insights and Data Comparison
The study was conducted using data from CHI 2014, comparing 3,869 preferences from 634 authors against 7,228 preferences from 347 attendees.
- Recall vs. Precision: Author-sourcing provides high precision (technical relatedness) but low recall (missing connections outside immediate sub-fields). Attendee-sourcing provides high recall, identifying "hidden" affinities. For example, attendees linked papers on "Body Pose" and "Mid-Air Keyboards" via the concept of "in-air interaction," a connection missed by automated text-based tools and authors.
- Predicting Popularity: Attendee-sourcing acts as a census for room capacity. By tracking which papers are bookmarked most, organizers can move popular sessions to larger rooms, preventing the common frustration of being turned away from a full session.
- Complementary Strengths: Author data helps create thematically coherent sessions, while attendee data helps resolve conflicts between those sessions.
Results and Future Directions
Despite having fewer participants than the author group, attendees were significantly more active, marking an average of 20.8 papers each (compared to 6.1 for authors). This resulted in better coverage, capturing preferences for 98.98% of all papers.
While promising, attendee-sourcing faces challenges regarding incentives and potential social bias. Future work aims to move from providing data to organizers to fully automated session creation, treating scheduling as an optimization problem where attendee preferences serve as the primary constraints.
Summary updated