Introduction
Predicting future social interactions is a cornerstone of network science, typically applied to digital platforms like Facebook or citation databases. However, physical, face-to-face social networks present unique challenges. Unlike digital links, which are often binary and static, physical interactions are highly dynamic, temporal, and vary significantly in duration. This paper investigates how high-resolution proximity data can be used to predict both new and recurring face-to-face contacts in professional settings.
Problem Statement
The authors address two primary questions in the context of scientific conferences:
- New Link Prediction: Can we predict which participants will meet for the first time on a subsequent day based on the network structure of the first day?
- Recurring Link Prediction: Can we predict if two people who have already met will interact again?
Understanding these dynamics is crucial for developing context-aware recommendation systems for physical environments, such as mobile apps that facilitate networking at large events.
Methods
The study utilized active RFID tags from the SocioPatterns project to capture face-to-face interactions within a 1.5-meter range. The researchers evaluated several classes of link prediction algorithms, categorized into neighborhood-based measures (local overlap) and path-based measures (global topology).
A significant contribution of this work is the adaptation of standard unweighted algorithms into weighted variants that incorporate the duration of contacts.
The performance of these algorithms was measured using the Area Under the Curve (AUC) metric, comparing the predictive power of structural features against a random baseline.
Datasets
The analysis was conducted on three real-world datasets:
- LWA 2010: 77 participants (Kassel, Germany).
- Hypertext (HT) 2011: 62 participants (Eindhoven, Netherlands).
- LWA 2012: 42 participants (Dortmund, Germany).
Despite the relatively small number of nodes, the high temporal resolution provides a dense view of social dynamics often missing from larger, sparser digital datasets.
Key Insights and Ideas
The researchers identified several critical patterns in physical social networks:
- The "Long Tail" of Human Contact: The distribution of contact durations follows a heavy-tailed pattern. While the vast majority of interactions are fleeting (less than two minutes), a small subset of "strong ties" involves much longer durations.
- Predictability of Strong Ties: The duration of contact is a powerful predictor. Longer interactions on the first day are significantly more likely to recur, making the "strong" part of the network much more predictable than the "weak" part.
- Weak Ties as Noise: Interestingly, very short contacts (under 200 seconds) often act as noise. Filtering out these weak ties from the training data actually improved the accuracy of prediction algorithms for new links.
Key Results
- New Links: Weighted neighborhood measures, particularly Resource Allocation, were the most effective, achieving AUC scores between 0.65 and 0.70. This indicates that local network overlap is a reliable signal for future first-time meetings.
- Recurring Links: Path-based measures like Rooted PageRank and Katz, when weighted by duration, performed exceptionally well, with AUC scores reaching 0.80.
- Duration as a Baseline: Simply using the length of the initial contact (the "LEN" predictor) proved to be a surprisingly robust baseline for predicting recurrence, often outperforming complex structural measures.
Conclusion
This research highlights that in physical social networks, "tie strength"—defined by interaction duration—is a fundamental driver of predictability. While weak ties are numerous, they are often stochastic. In contrast, stronger ties provide the structural stability necessary for accurate link prediction. Future work will look toward integrating spatial data and real-time dynamic modeling to further refine these predictive capabilities for live recommender systems.
Summary updated