Dynamics of social network emergence explain network evolution

The study investigates how new social networks originate and whether early tie formation dynamics differ from later periods of network evolution. Prior models focus on steady, lower-order processes (e.g., transitivity, reciprocity, homophily, fitness, preferential attachment) that explain structural features and downstream outcomes but rarely observe networks from inception. The authors hypothesize that emerging networks undergo a churn phase characterized by rapid expansion and contraction of ties as individuals initially meet many alters and later prune connections, leading to more stable relationships. Because network evolution depends on prior states, the distribution of ties formed during churn may shape subsequent network structure and behavior (e.g., collaboration). The purpose is to measure and model this churn process and test whether early interactions predict later interactions and scientific collaboration.

The paper situates its contribution within network evolution literature, citing models of tie formation processes (transitivity, reciprocity, homophily, preferential attachment) that generate higher-order structure and explain outcomes across social, biological, and technological systems. It references work on growth processes producing skewed degree distributions and motif structures, and studies linking networks to inequality, health, and cooperation. It notes the gap in observing networks at inception and indicates that early dynamics may differ from later steady-state evolution. It also connects to work on co-authorship networks (clustering and triadic closure), integration in emerging student networks predicting academic performance, and large-scale online game networks showing temporal variation in tie dynamics.

Empirical data collection: A daily survey was administered during a four-week residential scientific training program in summer 2018 with 76 participants (graduate students, faculty, researchers from at least 22 countries). Participants reported daily interactions with other participants and the duration of each interaction for days 1–26. Only 1.1% of dyads preexisted, enabling observation from network inception. The program also included structured events for forming research collaborations: a research speed dating session on day 2, two group meetings in weeks 2 and 3, and optional co-authored paper submissions within one year, producing a co-authorship network.

Churn identification: Applied Geweke's diagnostic to daily network density to test equality of means between early and later segments. A significant change after day 6 led to defining days 1–6 as the churn period and days 7–26 as the post-churn period.

Descriptive analyses: Compared tie counts, mutuality, closed triads, interaction durations across periods using t-tests, Mood's median test, F-tests, and conditional uniform graph tests for mutuality and transitivity. Assessed whether dyads meeting during churn exhibited greater post-churn interaction frequency and time spent together.

Inferential modeling: Used temporal exponential random graph models (TERGMs) for the interaction network (days 7–26) and for the scientific collaboration network over its waves. Models included structural effects (in/out two-stars, four-cycles, GWESP for triadic closure), homophily (gender, nationality, ethnicity), and a memory term capturing prior-period ties (presence/absence). Estimation employed maximum pseudolikelihood with bootstrapped confidence intervals; coefficients interpreted as log-odds.

Key predictor: Inclusion of an indicator for whether a dyad had at least one churn-period interaction. Compared models with and without churn ties to assess added explanatory power and independence from other predictors.

Disproportionate variance tests: In supplementary analyses, compared the predictive weight of churn ties to ties observed in other periods, using both raw counts and counts normalized by number of waves.

Theoretical model and simulations: Proposed a stylized churn model incorporating fixed sociality (probability of tie gain), a time cost parameter (α) driving probability of tie loss decreasing with time via p_t = (T_t)^(-α), and a social bandwidth parameter (β) limiting maintainable degree β(V−1). Algorithm: (i) set V nodes; (ii) choose P_gain; (iii) set α to define P_loss; (iv) set β to cap degree; below cap, nodes may gain and lose ties; above cap, only lose ties. Calibrated parameters to observed density (V = 76, α = 0.45, P_gain = 0.07, β = 0.21). Ran 1000 simulations to compare density, two-stars, and global transitivity with observed data.

• Churn vs post-churn differences: Days 1–6 had significantly higher mean, median, and variance in reported tie counts than days 7–26 (t-test: X_1:6 = 6.21 vs X_7:26 = 2.47, p < .01, t = 24.46; Mood's median test: medians 6 vs 2, p < .01, χ² = 469.63; F-test for variances: F_1:6 = 9.78 vs X_7:26 = 2.98, p < .01, F = 3.28). Churn-period networks had higher counts of reciprocated ties and closed triads at statistically significant rates (p < .05), except mutuality on days 23 and 26.
• Interaction strength: Mean time spent together increased from churn to post-churn (24 mins vs 61 mins; p < .01, t = 3.65). Dyads that met during churn had more post-churn interactions and spent more time together than dyads that first met after churn (counts: 3.20 vs 2.08, p < .01, t = 8.91; KS p < .01, D = 0.20; Mood's medians 2 vs 1, p < .01, χ² = 40.83; time: 153 mins vs 85 mins, p < .01, t = 6.10; KS p < .01, D = 0.14; Mood's medians 62 vs 41 mins, p < .01, χ² = 11.77).
• Interaction network TERGM (days 7–26): Churn ties increased odds of a post-churn tie by a factor of 3.06 (log-odds 1.12). Other coefficients and the memory term were stable with/without churn predictor, indicating added independent explanatory power beyond autoregression and standard structural/homophily effects.
• Disproportionate variance in interaction evolution: With raw counts, each churn tie had a larger coefficient than ties from any single preceding period (0.54, 95% CI [0.48, 0.58] vs 0.33, 95% CI [0.30, 0.37]; z = 6.67). With normalized counts, the all-ties coefficient exceeded the churn coefficient (5.49 vs 3.22), implying each churn tie accounts for ≈59% (3.22/5.49) of subsequent tie formation.
• Scientific collaboration evolution: Majority of collaborators had met during churn (percent churn ties across waves: 75.0%, 64.1%, 72.9%, 75.3%). From first group meeting to final papers, churn-based collaboration ties dissolved less than non-churn ties (36.4% vs 62.6%; χ² = 20.84, p < .001). TERGM showed churn ties increased odds of collaboration by 1.38 (log-odds 0.32), independent of memory and clustering effects.
• Disproportionate variance in collaboration: With raw counts, churn ties had larger coefficients than any single prior-period ties (0.20, 95% CI [0.18, 0.23] vs 0.11, 95% CI [0.06, 0.21]; z = 2.27). With normalized counts, each churn tie explained ≈54% (1.20/2.22) of subsequent collaboration tie formation.
• Simulations: The churn model reproduced observed density distributions and two-star counts closely but under-predicted global transitivity while matching its temporal trend. Higher sociality, higher bandwidth, and lower time costs yielded higher densities in simulations.

Findings support the hypothesis that emerging social networks undergo an initial churn phase with rapid tie expansion and contraction, distinct from later steady evolution. Early ties during churn disproportionately predict later interaction frequency, durability, and strength, and also predict scientific collaboration up to a year later, beyond what is explained by autoregressive stability, homophily, or clustering. This demonstrates that the earliest interactions set trajectories for subsequent network structure and behavioral outcomes, highlighting the importance of observing and modeling networks from inception. The simulation model grounded in social bandwidth and time costs captures key features of the empirical dynamics, suggesting general mechanisms by which churn can produce observed network evolution. The under-prediction of transitivity indicates additional mechanisms (e.g., stronger triadic closure) may be at play beyond the stylized model.

The paper introduces and empirically validates the concept of churn in network emergence, showing that an initial cycle of tie expansion and contraction is central to explaining later network structure and behavior. Using fine-grained temporal data from a newly formed social setting and TERGMs, the study demonstrates that churn ties strongly predict subsequent interactions and scientific collaboration, providing explanatory leverage unavailable if observation begins later. A stylized model incorporating social bandwidth limits and time costs reproduces key empirical patterns, offering a framework for studying network emergence. Future research directions include: developing a taxonomy of "new" networks; assessing generalizability across contexts; examining community formation dynamics and meso-structural evolution; exploring implications for cooperation evolution; employing Bayesian learning and predictive frameworks; implementing randomized designs to isolate mechanisms; considering node entry/exit; and extending the model to weighted or signed ties.

The empirical analysis centers on a single case of network emergence in a unique, structured environment combining social and professional components, limiting generalizability. Program-organized events may have influenced interactions beyond self-organized behavior. The theoretical model, while reproducing densities and two-stars, under-predicts global transitivity, indicating omitted mechanisms. The study assumes a fixed population without node turnover. Additional contexts, randomized designs, and model extensions are needed to validate and refine the churn framework.