Where Will They Click Next? A Social Foraging Model for Collaborating Teams

Paper Title

Where Will They Click Next? A Social Foraging Model for Collaborating Teams

Publication Info

• Topic area: Predictive modeling of collaborative navigation in information-rich environments.
• Keywords: Social foraging, information scent, collaborative debugging, predictive modeling, PFIS-T, Information Foraging Theory, team coordination, software engineering, social cues, navigation prediction.

Background and Problem

• Problem / challenge: Existing computational models of information foraging are individual-centric and fail to account for social dynamics in collaborative tasks, such as teammates' actions and communication. This limits their utility in team-based environments.
• Significance: Collaborative problem-solving is increasingly central in domains like software engineering, crisis response, and scientific discovery. Addressing the gap in predictive models for team navigation can improve coordination and efficiency in these contexts.
• Motivation and related work: Prior work on Information Foraging Theory (IFT) and its computational models (e.g., PFIS family) has successfully predicted individual navigation but lacks integration of social cues. Social Information Foraging Theory (SIFT) provides a theoretical framework for group foraging but has not been operationalized into predictive models.

Solution

• Proposed approach: PFIS-T, a predictive computational model of social information foraging, integrates teammates' navigation history (implicit cues) and conversational data (explicit cues) to predict collaborative navigation steps.
• Novelty:
  1. Extends Information Foraging Theory to synchronous collaboration by modeling implicit and explicit social cues as sources of information scent.
  2. Introduces PFIS-T, the first operational model of social foraging, incorporating mechanisms like social cue weighting, temporal decay, and dynamic re-ranking.
  3. Demonstrates empirical improvements in predictive accuracy and coverage over individual-centric baselines.
  4. Provides a publicly available dataset and algorithm for reproducibility and further research.
• Procedure and key techniques:
  1. Incorporates teammates' recent navigation (implicit cues) and conversational references (explicit cues) into a dynamic graph-based model.
  2. Applies a "social gate" mechanism to balance individual and team influences on navigation predictions.
  3. Evaluates the model using a controlled lab study with ten three-person teams engaged in a collaborative debugging task.

Results

• Concrete findings:
  • PFIS-T predicted 81.5% of team navigations, reducing unknown predictions by 12.57% on average compared to the baseline PFIS3′.
  • Achieved a Hit@10 accuracy of 62.4%, improving by 2.7 percentage points over the baseline.
  • Implicit cues provided steady gains, while explicit cues offered variable but strong improvements at higher ranking thresholds.
• Advantage over baselines:
  • Reduced unknown predictions for 9 out of 10 teams, with statistically significant reductions for 4 teams.
  • Improved ranking accuracy (Hit@k) at larger thresholds, with combined implicit and explicit cues yielding the best results.
• Experiments / evaluation:
  • Conducted a virtual lab study with 30 participants (10 three-person teams) debugging a large open-source project.
  • Collected navigation logs and conversation transcripts to evaluate predictive performance using metrics like unknown rate and Hit@k.
  • Analyzed the impact of social cues and model parameters on prediction quality.
• Limitations and future work:
  • Limited to implicit cues from navigation history and explicit cues from speech/text; future work could incorporate richer signals like gaze or annotations.
  • Evaluated in a controlled lab setting with student and early-career participants; additional studies in professional environments are needed.
  • Focused on debugging tasks; broader validation across other collaborative domains is required.

Summary

PFIS-T extends Information Foraging Theory to collaborative settings by incorporating teammates' navigation history and conversational data as social cues. It significantly improves predictive coverage and ranking accuracy over individual-centric models, reducing unknown predictions by up to 57.1% for some teams. The model demonstrates the utility of social cues in guiding team navigation and offers practical design implications for collaborative tools. While evaluated in software engineering, PFIS-T’s mechanisms are generalizable to other domains requiring shared exploration of complex information spaces.