Dynamics of social network emergence explain network evolution

Numerous systems across social, biological, and technological domains evolve as networks. While network evolution models are well-established, the initial emergence phase remains less understood. Existing models focus on processes like transitivity, reciprocity, homophily, and preferential attachment, which explain network features and outcomes like inequality and cooperation. However, these models don't fully capture the initial formation of networks in new social environments (e.g., a new job, university). In these settings, unfamiliar individuals form relationships for the first time. The network could emerge gradually, consistent with existing growth models, or initial interactions could drastically differ from later, more stable periods. This paper focuses on the latter scenario, where network emergence poses a greater challenge to current theory. The authors hypothesize that many emerging networks undergo a 'churn' process – an initial expansion and contraction of ties before stabilization. The distribution of ties from this churn could significantly influence subsequent network evolution and individual outcomes (academic success, voting behavior, cognitive health). The study aims to empirically investigate this churn concept and its implications using a unique dataset.

The paper reviews existing literature on network evolution models, highlighting the processes governing tie formation and dissolution, such as transitivity, reciprocity, fitness, homophily, and preferential attachment. These processes generate crucial structural features like transitive triplets, motif distributions, and skewed degree distributions. The existing literature explains patterns of inequality, health outcomes, and cooperative stability in human populations. However, the origins of social networks, particularly in novel social settings, remain less explored. The authors note previous work on large-scale online game networks that showed differences in tie formation and dissolution probabilities between early and later evolutionary periods but lacked the granularity of their study. They also cite research demonstrating the impact of social network integration on academic performance and other outcomes, providing a basis for investigating the long-term effects of early network formation.

To study churn, the researchers conducted a daily survey in a four-week training program attended by 76 individuals. The program provided a controlled environment for observing network emergence from inception, with minimal pre-existing relationships (only 1.1% of dyads knew each other beforehand). Participants daily reported interactions with others, including duration. Data was collected from day 1 to 26. Geweke's diagnostic was used to identify a significant change in network density after day 6, dividing the data into 'churn' (days 1-6) and 'post-churn' (days 7-26) periods. Statistical tests (t-tests, Mood's median test, Kolmogorov-Smirnov test) compared tie counts, mutuality, clustering, and interaction frequency between periods. To model network evolution, a temporal extension of the exponential random graph model (TERGM) was employed. Two TERGMs were fitted to the post-churn period: one including churn ties as a predictor and another omitting them. The models included structural and homophily-based variables (in-two-stars, out-two-stars, four-cycles, GWESP, homophily terms based on gender, nationality, and ethnicity) and a memory term accounting for tie autoregression. Maximum pseudolikelihood estimation with bootstrapped confidence intervals was used. Finally, a stylized model of the churn process was developed, incorporating parameters for social bandwidth and time cost to simulate network density and structural features (two-stars and triads). The model was optimized against observed network density.

The analysis revealed significant differences in tie formation behavior between the churn and post-churn periods. The churn period showed higher mean, median, and variance in reported tie counts, greater reciprocity and clustering, and shorter interaction durations compared to the post-churn period. Interaction frequency increased over time, suggesting relationships became more meaningful. Dyads who met during churn reported more subsequent interactions and spent more time together post-churn. The TERGM analysis revealed that churn ties significantly predicted post-churn ties, independently of other factors, indicating churn ties explain variance beyond simple tie autoregression. The impact of churn ties was disproportionately large compared to ties formed in any other period of the program. This held true regardless of whether raw interaction counts or normalized counts were used. The churn process also influenced scientific collaboration. Individuals who met during churn formed more durable co-authorship ties. A TERGM analysis of the scientific collaboration network showed that churn ties significantly increased the likelihood of collaboration, even controlling for other factors and network memory. Similar to interaction networks, churn ties in the scientific collaboration network also explained a disproportionate amount of variance in subsequent collaboration. The model of churn process, incorporating social bandwidth and time costs, successfully simulated the observed network density and structural features, although it under-predicted global transitivity.

The findings strongly support the concept of churn as a key factor shaping network emergence and evolution. The initial period of rapid tie expansion and contraction significantly impacts both the structure of the network and later behavioral outcomes. The results demonstrate that early interactions are crucial for predicting long-term network structure and collaboration patterns, extending well beyond the immediate timeframe of the study. This has implications across various fields, from understanding the dynamics of scientific collaborations to predicting the success of social interventions. The study highlights the limitations of solely focusing on stable network states for understanding network dynamics, emphasizing the critical role of the emergence phase.

This research introduces the concept of 'churn' in network emergence, characterized by an initial burst of tie formation followed by contraction. The empirical findings demonstrate that this churn significantly influences network evolution and subsequent behavioral outcomes. A stylized model effectively captures the observed network dynamics, emphasizing the role of social bandwidth and time costs. Future work should explore the generalizability of churn across diverse network contexts, investigate the influence of structured events, examine community formation in emerging networks, and incorporate Bayesian frameworks for modeling network evolution.

The study's empirical component relies on a single case study of network emergence, which limits the generalizability of the findings. The specific context of the summer program, with its structured events and mix of social and professional components, might influence the results. Further research on diverse network settings is necessary to confirm the broader applicability of the findings. The model, although successfully capturing key features, under-predicts global transitivity, suggesting areas for future refinement.