New Insights and Methods For Predicting Face-To-Face Contacts

Christoph Scholz, Martin Atzmueller, Alain Barrat, Ciro Cattuto, Gerd Stumme

Abstract

The prediction of new links in social networks is a challenging task. In this paper, we focus on predicting links in networks of face-to-face spatial proximity by using information from online social networks, such as co-authorship networks in DBLP, and a number of node level attributes. First, we analyze influence factors for the link prediction task. Then, we propose a novel method that combines information from different networks and node level attributes for the prediction task: We introduce an unsupervised link prediction method based on rooted random walks, and show that it outperforms state-of-the-art unsupervised link prediction methods. We present an evaluation using three real-world datasets. Furthermore, we discuss the impact of our results and of the insights we glean in the field of link prediction and human contact behavior.

Introduction

Link prediction—the task of estimating the likelihood of a future connection between nodes in a network—is a fundamental problem in social network analysis. While extensively studied in digital environments like co-authorship or online social networks, predicting physical, face-to-face (F2F) interactions presents unique challenges. Human contact networks are characterized by high sparsity and extreme dynamism.

A significant hurdle is the "cold-start" problem: recommending connections when initial interaction data is minimal. In environments like professional conferences, where many participants are strangers, traditional supervised models often lack sufficient training data. This research explores how auxiliary "online" digital footprints, such as DBLP co-authorship records and paper similarity, can be leveraged to predict "offline" physical interactions.

Methodology: Hybrid Rooted PageRank

The primary contribution of this work is the Hybrid Rooted PageRank algorithm. Traditional Rooted PageRank identifies proximity by performing a random walk on a single graph. The hybrid variant extends this by operating across multiple network layers simultaneously, allowing for the integration of diverse data sources.

The algorithm follows a specific random walk logic:

  1. Restart: With probability $\alpha$, the walker jumps back to the "root" node (the individual for whom predictions are being made).
  2. Network Selection: With probability $1 - \alpha$, the walker selects a specific network $N_i$ (e.g., the physical contact network or the DBLP co-authorship network) based on a predefined probability distribution $P$.
  3. Traversal: The walker then moves to a neighbor within the chosen network, proportional to the edge weights.
The Hybrid Rooted PageRank algorithm logic for cross-network link prediction.

This approach allows the model to "cross-pollinate" insights, using digital history to inform physical proximity predictions. Additionally, the researchers integrated "encounter" data (spatial co-location via RFID readers) and "paper profile similarity" (TF-IDF analysis of published works) as influence factors.

Experimental Setup

The study evaluates the proposed method using three real-world datasets collected via the SocioPatterns platform at academic conferences:

  • LWA 2010: 77 participants, 1,004 initial contacts.
  • Hypertext 2011 (HT2011): 68 participants, 698 initial contacts.
  • LWA 2012: 42 participants, 478 initial contacts.

These datasets are particularly rich, combining high-resolution RFID proximity data (detecting F2F contact within 1–1.5 meters) with crawled DBLP publication records and full-text manuscript data.

Key Results and Discussion

The evaluation demonstrates that digital proximity is a strong predictor of physical proximity. Specifically, co-authorship and shared research interests (paper similarity) are significant indicators of future contact duration.

Performance and Synergy

The Hybrid Rooted PageRank consistently outperformed state-of-the-art unsupervised baselines, such as Weighted Resource Allocation (WRA) and Weighted Jaccard Coefficient (WJC). A key finding is the synergy achieved by combining data sources. As shown in the results for the HT2011 dataset, integrating DBLP co-authorship data with physical contact data ($\beta = 0.5$) yields a higher Area Under the Curve (AUC) than using either network in isolation.

AUC performance on the HT2011 dataset across different time thresholds, comparing single-network and hybrid approaches.

The models exhibit higher accuracy when predicting "strong" links—contacts with longer durations—compared to "weak," brief interactions. This is reflected in the rising AUC scores as the duration threshold for defining a "contact" increases. Furthermore, paper similarity (content-based) often proved more predictive than raw co-authorship networks, as it captures latent shared interests between potential future collaborators.

Conclusion

This research highlights the utility of unsupervised methods for addressing the cold-start problem in dynamic social environments. By blending physical sensor data with digital footprints, the Hybrid Rooted PageRank provides a practical, "out-of-the-box" solution for link prediction at professional gatherings. Future work will focus on automating parameter estimation and testing the model's generalizability in less structured social settings like public spaces or offices.

Summary updated

Abstract

The prediction of new links in social networks is a challeng- ing task. In this paper, we focus on predicting links in net- works of face-to-face spatial proximity by using information from online social networks, such as co-authorship networks in DBLP, and a number of node level attributes. First, we analyze influence factors for the link prediction task. Then, we propose a novel method that combines information from different networks and node level attributes for the pre- diction task: We introduce an unsupervised link prediction method based on rooted random walks, and show that it out- performs state-of-the-art unsupervised link prediction meth- ods. We present an evaluation using three real-world datasets. Furthermore, we discuss the impact of our results and of the insights we glean in the field of link prediction and human contact behavior.
BibTeX
@inproceedings{scholz2021new,
  author = {Christoph Scholz and Martin Atzmueller and Alain Barrat and Ciro Cattuto and Gerd Stumme},
  title = {{New Insights and Methods For Predicting Face-To-Face Contacts}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/icwsm.v7i1.14415},
}
Figure