🔬 physics

Extracting behavioural properties from face-to-face interactions temporal networks: a measure of egonet persistency

This paper introduces the Neighbourhood Persistency Criterion (NPC), a statistically grounded framework for quantifying egonet persistence in temporal networks that reveals a common exploration–exploitation trade-off in face-to-face interactions across conferences, showing that these interaction strategies are driven more by contextual factors than by socio-demographic traits.

Original authors: Gabriel Maurial, Elisa Klüger, Mathieu Génois

Published 2026-06-29

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Gabriel Maurial, Elisa Klüger, Mathieu Génois

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are at a big, busy conference. You have a few hours to mingle, but the room is full of hundreds of people. You face a constant choice: Do I stick with the people I already know and have a deep conversation, or do I wander around and meet as many new faces as possible?

This paper is about figuring out how people make that choice over time, and how to measure it without getting tricked by the math.

The Problem: The "Popular Person" Trap

Researchers have tried to measure how "sticky" a person's social circle is. They ask: Does this person keep talking to the same group, or do they constantly switch?

But there's a catch. If you are a very popular person who talks to everyone in the room, your social circle looks "sticky" by accident, even if you are just randomly bumping into people. It's like saying a person who owns a giant bucket is "holding more water" than someone with a cup, even if the cup is full to the brim and the bucket is half-empty. The old math couldn't tell the difference between genuine loyalty and just being busy.

The Solution: The "Neighbourhood Persistency Criterion" (NPC)

The authors created a new tool called NPC (Neighbourhood Persistency Criterion). Think of this as a social "lie detector" that filters out the noise.

It compares two things:
- Who you talked to (Identity).
- How long you talked to them (Intensity).
It uses a "Control Group" (The Null Model):
Before judging a person, the computer simulates what would happen if everyone in the room just shuffled around randomly, keeping their same number of conversations but swapping partners.
The Score:
- If a person talks to the same people more than the random shuffle would predict, they get a high score (The "Explorer" who found their tribe).
- If they talk to the same people less than random chance (actively avoiding old friends), they get a low score (The "Avoider").
- If they are just like the random shuffle, they get a zero (The "Random Mixer").

What They Found: The "Exploration vs. Exploitation" Dance

The researchers applied this tool to data from four different academic conferences, tracking face-to-face interactions down to the second. Here is what they discovered:

The "Steady Eddies": Some people have a very consistent style. They either stick to a small, trusted group all day (high persistence) or they constantly hunt for new people (low persistence).
The "Switchers": Many people don't fit into one box. They might spend the morning exploring (meeting new people) and the afternoon exploiting (deepening bonds with the new people they met). Their behavior changes depending on the break or the session.
The "Avoiders": A tiny group actively avoids people they've already met, constantly seeking fresh faces.

The Big Surprise:
The authors looked at the attendees' ages, genders, and backgrounds to see if these behaviors were "personality traits."

Result: Almost nothing.
Meaning: Being a "Steady Eddie" or a "Switcher" isn't about who you are (your personality or age). It's about where you are and what you are doing right now. A 25-year-old might be a "Steady Eddie" at one conference and a "Switcher" at another, depending on the schedule, the coffee breaks, or their specific goals for that day.

The "Bucket vs. Cup" Analogy for Time

The study also found a subtle difference between who you talk to and how long you talk.

It is easier to keep talking to the same group of people (keeping the bucket full of the same friends).
It is harder to keep the amount of time you spend with each friend exactly the same.
Analogy: You might stick with your three best friends all day (high "Who" score), but you might spend 20 minutes with one, 5 minutes with another, and 40 minutes with the third (low "Time" score). The paper shows these two behaviors don't always match up perfectly.

The Bottom Line

This paper gives us a better ruler to measure social behavior. It proves that at big events, people aren't just acting on their "personality." Instead, they are constantly balancing a trade-off: Do I dig deep with the people I know, or do I cast a wide net for new connections?

The tool they built (NPC) helps us see this dance clearly, showing us that our social strategies are often a reaction to the moment, not a fixed part of who we are.

Technical Summary: Extracting Behavioural Properties from Face-to-Face Interactions via Temporal Networks

Problem Statement
The analysis of temporal networks requires robust methods to understand how individuals repeat social interactions over time. While repeated interactions are known to shape information diffusion and social structure, existing measures of egonet persistence often conflate genuine behavioural regularities with structural effects, such as node degree. High-degree nodes inherently exhibit high similarity in their neighbourhoods across time periods, creating a bias that makes it difficult to distinguish meaningful temporal correlations from random mixing. Furthermore, there is a lack of interpretable observables that directly quantify egocentric interaction persistence while accounting for both contact identity and interaction intensity.

Methodology
The authors introduce the Neighbourhood Persistency Criterion (NPC), a statistically grounded framework designed to quantify egonet persistence across time. The methodology proceeds in three stages:

Similarity Measures: The framework first computes two classical similarity metrics for an individual's egocentric network between two time periods ( $T$ and $T'$ ):
- Jaccard Index ( $\lambda_J$ ): Measures the overlap in contact identities, independent of interaction strength.
- Cosine Similarity ( $\lambda_{cos}$ ): Incorporates interaction weights (total duration), capturing whether the intensity of interactions is preserved alongside identity.
Statistical Normalization via Null Models: To correct for the bias introduced by node degree and network topology, the authors apply two specific null models via randomization:
- $P[k]$ : Rewires the network while preserving node degrees. This serves as the null model for the Jaccard index to control for topological correlations.
- $P[L, p(w)]$ : Retains the network topology but randomly redistributes interaction weights according to the empirical distribution. This serves as the null model for cosine similarity to control for correlations between topology and interaction intensity.
The NPC Observable ( $\alpha$ ): The authors define a standardized observable, $\alpha$ , based on the deviation of the observed similarity from the median of the null distribution ( $\mu_l$ ).
- $\alpha$ ranges from $[-1, 1]$ .
- $\alpha = 1$ indicates a strong preference for re-engaging with the same individuals.
- $\alpha = -1$ indicates a strong avoidance of repeated interactions.
- $\alpha \approx 0$ indicates no deviation from random interaction patterns.
- This formulation allows for continuous, probabilistic interpretation of behavioural deviation, distinguishing between positive and negative deviations.
Dimension Reduction and Behavioral Categorization: To characterize individual dynamics across multiple time periods, the authors analyze the temporal consistency of $\alpha$ values. Recognizing that standard deviation is sensitive to outliers and that individuals may alternate strategies, they employ the Trimmed Mean Dispersion using Tukey's method (TMDT) with a percentile $q=12\%$ . This metric ( $TMDT_{12}$ ) quantifies the intra-individual consistency of behaviour, distinguishing between stable patterns and high variability (alternating strategies).

Data Application
The framework was applied to high-resolution face-to-face interaction networks collected at four Computational Social Science conferences (WS16, ICCSS17, ECSS18, ECIR19) using the SocioPatterns platform. Data was aggregated into contiguous high-activity windows (e.g., coffee breaks, lunch, poster sessions) to capture socially meaningful events while filtering out spurious short-duration contacts.

Key Results

Exploration–Exploitation Trade-off: Analysis of $\alpha$ $α$ versus $TMDT_{12}$ $T M D T_{12}$ reveals a common behavioural structure across all events. Attendees generally fall into distinct categories:
- Stable Exploiters: Individuals with low dispersion and high $\alpha$ (near 1), maintaining a consistent set of contacts.
- Stable Explorers: Individuals with low dispersion and $\alpha$ near 0, frequently changing partners without a clear structure.
- Avoiders: A small group with low attendance and systematic avoidance of previous contacts ( $\alpha \approx -1$ ).
- Switchers: Individuals with high dispersion, indicating an alternation between exploration and exploitation strategies across breaks.
Identity vs. Intensity: There is a strong correlation between $\alpha_J$ and $\alpha_{cos}$ , but attendees tend to be slightly more persistent in maintaining contact identity than in maintaining the specific duration of those interactions.
Lack of Socio-Demographic Determinism: Crucially, the study found no significant systematic association between these behavioural dynamics (consistency or variability) and stable socio-demographic attributes (age, gender) or self-reported traits. While weak trends exist (e.g., younger attendees showing slightly more persistence in some contexts), the results suggest that interaction strategies are shaped primarily by contextual factors (event structure, scheduling, immediate objectives) rather than stable individual traits.

Significance and Claims
The paper claims that the NPC framework provides a flexible and interpretable tool for studying egonet persistence in temporal networks. Its primary significance lies in:

Decoupling Behaviour from Structure: By utilizing tailored null models, NPC successfully separates genuine behavioural regularities from structural artifacts like node degree.
Nuanced Behavioral Typology: It enables the identification of distinct behavioural archetypes (exploration vs. exploitation) and the quantification of their temporal stability, moving beyond simple aggregate measures.
Contextual Insight: The application to conference data challenges the assumption that interaction strategies are fixed individual traits, highlighting instead the role of situational context in shaping social dynamics.

The authors conclude that while NPC offers a robust method for analyzing temporal networks, future work should explore daily-level persistency patterns and extend the framework to link-level persistency to capture finer-grained temporal structures.