Wedding Cocktail Hour Contact Webs: Temporal Proximity Network of a Privately Hosted Social Event
Source: arXiv:2605.30291 · Published 2026-05-28 · By Joshua Z. Stadlan, Richard B. Kahn, Michelle Birkett
TL;DR
This study addresses a notable gap in face-to-face interaction research by collecting and publicly releasing the first fine-grained temporal proximity network dataset from an informally structured, privately hosted social event—a wedding cocktail hour. Previous datasets typically come from institutionally structured environments like schools, workplaces, or hospitals, where formal roles and schedules influence interactions and may obscure organic social dynamics. By contrast, this dataset captures 95 participants freely mingling outdoors for about an hour, with rich meta-data on relationship-based social groups relative to the wedding couple. The dataset contains 7,213 contact events over 2,760 dyads recorded every 10 seconds using ultra-wideband proximity sensors with approximately 1.5 meter range. Participants self-reported their social group membership, enabling novel analyses of social mixing grounded in meaningful relationship categories instead of proxies like demographic features or organizational roles.
The authors validate the sensors via bench experiments comparing badge contacts to video, adjust temporal resolution to account for intermittent missed detections, and provide a clean event window with at least 75% badge activity. This dataset enables testing whether well-known temporal network signatures such as heterogeneous contact rates, clustering, bursty dynamics, and group mixing observed in institutional settings generalize to informal socializing. Importantly, it supports applied modeling of contagion risk, social-space design, and contact tracing policies in realistic social events. The open release further advances the network science community's efforts to study natural human interaction patterns with contextually grounded group labels.
Key findings
- The dataset contains 7,213 contact events spanning 2,760 unique dyads among 95 participants during a 58-minute outdoor cocktail hour.
- Proximity events were recorded using Ultra-Wideband (UWB) badges with ~±0.1 m accuracy and a 1.5 m detection threshold, coarsened to 10-second temporal bins.
- Bench-validation against video confirmed high contact detection rates with no false positives but showed occasional missed 5-second intervals, prompting temporal aggregation.
- Participants self-reported relationship categories relative to the wedding couple allowing analysis based on meaningful social groupings rather than demographic proxies.
- About 40% of guests wore badges; data does not capture interactions with non-badge wearers, a notable limitation for group mixing completeness.
- Contact events end when distance exceeds 2 meters, capturing dynamic social proximity relevant for infection risk modeling.
- The dataset preserves k-anonymity by collapsing smaller groups and non-responses into an “other” category to protect privacy.
- Physical environment (outdoor patio, buffet lines) and participant behavior introduce contact scenarios where proximity does not always imply interaction (e.g., standing back-to-back).
Threat model
n/a — The study is purely observational and data-collection focused without consideration of adversarial actors or attempts to compromise sensor data integrity.
Methodology — deep read
Threat model & assumptions: The study is observational without adversarial components, focusing on organic social interactions at a private wedding cocktail hour. Participants consented to data collection; no adversarial manipulations or attempts to spoof proximity badges were considered.
Data provenance & size: Data collected from 95 adult guests out of ~220 total attendees at an outdoor cocktail hour lasting ~65 minutes. After filtering for 75% badge activity, final observation window is 3,510 seconds (58.5 minutes). Participants wore UWB badges detecting other badges within ~1.5 meters with ±0.1 m accuracy and 5-second scanning intervals. Total 7,213 contact events recorded, spanning 2,760 unique dyads.
Architecture/algorithm: No predictive model was developed. Instead, raw proximity events were logged as discrete-time temporal edges with 10-second time bins, indicating undirected contact between pairs. Participant metadata labels indicate self-reported relationship categories to the wedding couple, e.g., spouse A’s friends, mutual friends, relatives, etc., collapsed for privacy.
Training regime: Not applicable as no ML model training occurred. Data processing included temporal aggregation of raw 5-second contact reports into 10-second bins to mitigate missed detections, and trimming to stable badge activity intervals.
Evaluation protocol: The authors performed bench-validation comparing badge contact detections to timestamped video recordings in controlled scenarios to assess measurement accuracy, responsiveness, and symmetry. They verified no false positives but identified some missed contact records leading to data coarsening. The dataset enables further analyses of temporal network properties, group mixing patterns, and could be used for epidemiological modeling.
Reproducibility: The entire proximity network and node attribute datasets are publicly available on Zenodo under DOI:10.5281/zenodo.20430824 along with a detailed README. Complete raw sensor data is aggregated and anonymized. Data collection protocols and validation procedures are documented.
Example end-to-end: During the 58-minute interval, each participant’s badge recorded other badges within 1.5 m every 5 seconds, which were then aggregated into 10-second temporal edges. For example, if two badges detected each other during any 5-second interval within a 10-second window, an undirected temporal edge is recorded for that time bin. These contacts are linked with participant relationship labels, facilitating analyses such as group-wise contact rates, homophily, and contact network topology in an informal social setting.
Technical innovations
- First public temporal proximity network dataset from an informally structured private social event with participants freely mixing outdoors.
- Inclusion of contextually meaningful self-reported relationship-based group labels rather than demographic proxies or institutional roles.
- Validated use of Ultra-Wideband proximity badges with high spatial resolution (±0.1 m) for fine-grained social interaction capture outside institutional settings.
- Methodology of temporal aggregation and data trimming to handle sensor missed contacts and variable badge activity in a free-moving social environment.
Datasets
- Wedding Cocktail Hour Temporal Proximity Network — 7,213 contact events, 95 participants — Publicly available at Zenodo DOI: 10.5281/zenodo.20430824
Limitations
- Only ~40% of wedding guests wore badges; interactions involving non-badge wearers (~60%) are unobserved, potentially biasing social mixing analyses.
- Proximity does not always equal social interaction, e.g., people standing back-to-back or in buffet/bar lines may be proximate without interaction.
- Venue geometry, participant outfit constraints, and badge placement affect distance detection consistency; the effective detection range may vary.
- 5-second badge sensing cycles are not globally synchronized, leading to occasional missed contact intervals and asymmetries between badge reports.
- No detailed logs on badge distribution timing; contact records may conflate genuine social interactions with badge pickup/drop-off events.
- Relationship group labels reflect connections to the wedding couple and may omit other salient social ties (e.g., romantic partners) or demographic heterogeneity.
- Privacy considerations limited release of richer participant attributes; group sizes are unbalanced which may affect statistical power in subgroup analyses.
Open questions / follow-ons
- How do informal social gathering network signatures (e.g., contact rate distributions, clustering) quantitatively differ from institutional settings when adjusting for group structure?
- Can richer metadata capturing multiple overlapping social group memberships or dynamic affiliations improve modeling of social mixing in informal events?
- What are best practices to overcome partial participation when only a subset of attendees wear sensors, to infer whole-event interaction patterns?
- How do variations in spatial layout and event organization influence contact dynamics and can these insights guide safer social-space design?
Why it matters for bot defense
This dataset and study provide bot-defense and CAPTCHA practitioners insight into the fine-grained organic interaction patterns of humans in informal social settings, differing from structured institutional contexts commonly used for modeling human behavior. Understanding realistic face-to-face contact dynamics including burstiness, clustering, and group mixing with socially meaningful labels can inform the design of more human-like conduct that bots should emulate or be distinguished against. For instance, temporal proximity networks elucidate plausible natural patterns of user co-presence and interaction timing that advanced behavioral CAPTCHAs or anomaly detection systems might leverage to differentiate automated from human agents. Additionally, the limitations highlight challenges in interpreting proximity as interaction, important when using sensor or behavioral data for authentication or bot-detection heuristics. Overall, this work contributes to a more grounded understanding of human social contact networks relevant for sophisticated bot-defense modeling in real-world social-event scenarios.
Cite
@article{arxiv2605_30291,
title={ Wedding Cocktail Hour Contact Webs: Temporal Proximity Network of a Privately Hosted Social Event },
author={ Joshua Z. Stadlan and Richard B. Kahn and Michelle Birkett },
journal={arXiv preprint arXiv:2605.30291},
year={ 2026 },
url={https://arxiv.org/abs/2605.30291}
}