Group mixing drives inequality in face-to-face gatherings

The study examines how inequalities in social connectivity emerge from face-to-face interactions, where segregation and homophily can produce unequal opportunities in education, employment, and health. Recent sensing technologies have revealed universal properties of face-to-face interaction, but models focusing solely on intrinsic individual attractiveness neglect pairwise group dynamics. The authors hypothesize that group mixing preferences, coupled with group-size imbalance, drive degree inequality between social groups in gatherings. They aim to quantify these effects in empirical data and propose a mechanistic model integrating individual and pairwise (group) preferences to explain and predict observed disparities.

Prior work in sociology and anthropology has long emphasized the importance of face-to-face interaction for social group formation and segregation. Sensor-based studies in schools, workplaces, and conferences have uncovered regularities such as a characteristic average number of interaction partners per gathering and broad distributions of interaction durations and inter-event times. Mechanistic models based on intrinsic attractiveness and activity can replicate several micro-dynamics but fall short of explaining group-level mixing patterns. Homophily theory indicates that individuals often prefer same-group interactions, which can alter network structure and ranking outcomes for minorities. Existing network research has shown homophily’s impact on minority ranking and perception biases, but a quantitative, mechanistic understanding of how group mixing and size imbalances generate degree inequality in real face-to-face gatherings has been lacking.

Data: Six sensor-based face-to-face interaction datasets (four schools and two academic conferences) were analyzed. Each node is an individual; an undirected edge exists if two individuals had at least one face-to-face contact during the study. Gender defines two groups (minority labeled 0, majority 1). Networks display unimodal degree distributions. Dataset statistics include N in [115, 327], minority fraction f0 in [0.27, 0.46], and varying average degrees.

Mixing assessment: Inter- and intragroup edges were compared to configuration-model nulls (degree-preserving randomizations) via z-scores using 500 random instances per dataset to assess over-/under-representation of intra/intergroup ties.

Model: The attractiveness-mixing model augments intrinsic individual attributes with pairwise group mixing. Each agent i has: group label b_i ∈ {0,…,B−1} (B=2 here), intrinsic attractiveness η_i ∈ [0,1], and activation probability r_i ∈ [0,1]. Group-level pairwise tendencies are encoded in a B×B mixing matrix H, where row b gives the probabilities of interacting with each group. Agents perform random walks on a 2D L×L periodic space. At each time step, an active agent moves with probability α_i(t) that depends on the local neighborhood within radius d, otherwise remains and attempts interaction with neighbors based on mixing likelihood via β_i(t). Movement step length is v with random direction; inactive agents become active with probability r_i; isolated active agents become inactive with probability 1−r_i. Intrinsic attractiveness η_i and activation r_i are drawn from Uniform[0,1]. Together, intrinsic and mixing effects constitute social attractiveness.

Analytical derivation: Under a dilute-system assumption (low spatial density so interactions are mostly pairwise), the authors derive closed-form expressions for the normalized shares of intra- and intergroup edges e_rs as functions of group fractions f_r and mixing matrix entries. For B=2, expressions for e_00 and e_11 are given in terms of f_0, f_1, h_00, h_11 and intergroup terms; analogous expressions for e_01 and e_10 are provided. These enable estimating H from empirical edge composition.

Estimation and simulation: From empirical networks, the shares of intra/intergroup edges were computed, and h_00 and h_11 were estimated via numerical optimization to match the analytical expressions. Simulations then used the empirical N, group sizes, and estimated H to test whether the model reproduces group degree inequality and temporal properties. Generic simulation parameters included L=100, d=1, v=1, N=200 for parametric analyses; for empirical comparisons, simulations ran until the number of edges matched the data. Results were averaged over 50 runs.

Ranking adjustment: To assess representation in rankings, individuals were ranked by degree and by intrinsic attractiveness (in simulations). To reduce misrepresentation, a group-adjusted ranking was defined by computing each node’s degree z-score within its own group, then ranking all individuals by this adjusted score.

• Empirical inequality: Across six datasets, systematic degree inequality exists between gender-defined groups. Examples: School 1 average degree minority 64.85 vs majority 72.08; School 2: 24.71 vs 28.32; School 3: 23.46 vs 25.11; School 4: 36.48 (minority) vs 34.87 (majority); Conference 1: 95.35 vs 96.12 (near equal); Conference 2: 118.61 (minority) vs 110.56 (majority).
• Homophily: Intragroup ties are more frequent than expected by the configuration-model null in most datasets (positive z-scores), indicating significant homophily in face-to-face interaction.
• Intrinsic-only models fail: A model with only intrinsic attractiveness cannot simultaneously explain both degree inequality and observed intragroup mixing patterns.
• Model validation: Using mixing matrices H estimated from data, the attractiveness-mixing model reproduces the observed group average-degree disparities and temporal properties (interaction duration, inter-interaction times, weight distributions), supported by Kolmogorov–Smirnov tests.
• Regimes: With two groups (minority fraction f0), the system exhibits regimes: heterophilic (h<0.5) where minority has degree advantage, and homophilic (h>0.5) where minority has a degree disadvantage due to reduced cross-group opportunities.
• Asymmetry: Majority mixing h_11 explains much of the variance in minority degree deviation from the network average; changing h_11 can shift minorities from advantage to disadvantage, while minority mixing h_00 only modestly attenuates inequality.
• Critical size: The response of minority average degree to changes in h_00 depends on minority size. An approximate critical fraction f0^c ≈ h_11 / (2(h_00 + h_11)) separates regimes where increasing minority homophily increases (for f0 > f0^c) or decreases (for f0 < f0^c) minority degree. An upper-limit critical size f0* (exact expression provided) indicates when raising minority homophily is always beneficial regardless of majority mixing (e.g., for h_00=0.5, f0*>36.7%).
• Ranking misrepresentation: Degree rankings misrepresent group potential relative to intrinsic attractiveness. In homophilic regimes, minorities are underrepresented in top-k degree ranks; in heterophilic regimes, they are overrepresented. A group-adjusted degree ranking (degree z-score within group) markedly reduces misrepresentation and correlates better with intrinsic attractiveness except at extreme mixing values (h=0 or h=1).

The findings demonstrate that degree inequality in face-to-face social networks emerges from the interplay of group-size imbalance and mixing preferences, not solely from individual intrinsic attractiveness. By integrating pairwise group mixing into a spatiotemporal interaction model, the study explains when minorities become less connected (homophily) or more connected (heterophily). Spatiotemporal constraints inherent to gatherings create limited interaction opportunities; majority homophily consumes opportunities within the majority, reducing cross-group exposure and thereby minority visibility and degree. The analysis shows that majority mixing chiefly drives inequality, while the efficacy of minority strategies depends on minority size. A critical mass enables minorities to benefit from increased cohesion; below that, higher minority homophily exacerbates inequality. These insights have practical implications for interpreting minority positions in networks and for algorithmic systems that rank or recommend individuals, where raw degree may reflect mixing biases rather than individual potential.

This work introduces and validates an attractiveness-mixing model that unifies intrinsic individual preferences with group mixing dynamics to explain and predict group degree inequality in face-to-face networks. The model reproduces empirical disparities across six real-world gatherings, analytically characterizes how mixing and group sizes shape inter- and intragroup connectivity, identifies heterophilic and homophilic regimes, and reveals critical minority sizes guiding effective strategies. It also uncovers misrepresentation of minorities in degree-based rankings and proposes a simple group-adjusted ranking to reduce bias. Future research directions include: empirically estimating individual-level parameters (activation, intrinsic attractiveness), extending to multiple and continuous attributes, incorporating richer spatial data to refine spatiotemporal constraints, exploring dynamic processes (e.g., diffusion) under different mixing regimes, and detecting latent group structure influencing mixing.

• Analytical results rely on a dilute-system assumption (low spatial density where interactions are mainly pairwise), which may not hold in all settings.
• The study focuses on a binary attribute (gender-defined groups); while the model can be extended, results may differ with multiple or continuous attributes.
• Spatial aspects are modeled abstractly; richer spatial data are needed to validate and refine spatial constraints in real gatherings.
• The adjusted ranking correlates better with intrinsic attractiveness except at extreme mixing values (h=0 or h=1), indicating limits under extreme homophily/heterophily.
• Parameter estimation for activation probabilities and intrinsic attractiveness was not performed empirically; these were assumed uniformly distributed.