Abstract
Whenever many people interact, as on social media, their behaviour can be described as a network. While existing research often focuses on network topology or the pursuit of social capital, the relationship between network position and individual behavioural choices remains under-explored. This study demonstrates that an individual's behaviour correlates significantly with their number of connections. By analysing a large-scale author network and reconstructing behaviour using a Graphon Mean Field Game (GMFG), we show that the explicitness of published content is linked to network connectedness. Our model indicates that this correlation arises not just from the popularity of explicit content, but from the influence of an author’s specific neighbourhood. These results suggest that to accurately model individual behaviour in crowds, network topology must be integrated into the mathematical framework.
Introduction
The study of crowd psychology has long debated whether collective environments enrich individuals through social cohesion or lead to a loss of identity through conformity. In the digital age, this debate manifests in how users navigate social networks to gain "social capital." While it is established that individuals strive for connections, mathematically modelling how these connections alter behaviour at scale is challenging. Standard models often treat agents in isolation or assume uniform interaction with the entire crowd. This research addresses whether factoring in specific network connectedness improves the prediction of individual behaviour within a large-scale digital ecosystem.
Methodology: Graphon Mean Field Games
To model massive populations tractably, we employ Mean Field Games (MFGs), which abstract large groups of agents into a probability distribution. This allows for the analysis of a representative agent's interaction with the "mean field" of the crowd. To incorporate specific network structures, we extend this to a Graphon Mean Field Game (GMFG).
A "graphon" serves as a continuous limit for dense networks, providing a smooth function to represent connection probabilities between node types. This avoids the computational burden of massive adjacency matrices. We compared two distinct models:
- Reward-only model (Standard MFG): Predicts behaviour based solely on the expected reward of an action.
- Network-aware model (GMFG): Predicts behaviour based on both expected rewards and the state distribution of the author’s specific neighbourhood, derived from the estimated graphon.
Equilibrium states were learned using an Online Mirror Descent (OMD) algorithm, with empirical graphons estimated via penalized B-splines.
Dataset and Behavioral Mapping
The study utilized the Darmstadt Fanfiction Corpus (2020-2023), comprising 67,456 texts from 22,127 authors and approximately 1.1 million reviews.
- Behavioral State: Defined by the age restriction of published stories (P6, P12, P16, P18), serving as a proxy for content explicitness.
- Reward Metric: Defined as the number of reviews a story received.
- Sampling: A subset of 1,000 users was used to computationally estimate the network structure.
Structural Insights: The "Hollow" Center
Analysis of the fanfiction network revealed a non-canonical structure. Unlike platforms like Facebook, where highly connected users form a dense core, this interest-based network is "hollow." Highly connected authors form isolated peripheral clusters, while the center is populated by weakly connected authors who act as "weak ties," bridging these clusters and facilitating information flow.
Results
The data revealed a stark correlation: highly connected authors overwhelmingly produce highly explicit (P18) content. This aligns with reward patterns, as P18 texts receive an average of 31.65 reviews, compared to 4.09 for P6 texts.
The GMFG model significantly outperformed the standard MFG in reproducing empirical behaviour:
- L1 Error Reduction: Decreased from 14.25% (reward-only) to 12.32% (network-aware).
- L2 Error Reduction: Decreased from 18.65% to 15.65%.
- Statistical Significance: Wilcoxon signed-rank tests confirmed the improvement ($p < 0.001$) with a moderate effect size (Cohen’s $d = 0.47$).
Discussion
Our findings support the notion that as individuals become more embedded in a network, they tend to conform to the "will of the crowd"—in this case, producing explicit content to maximize engagement. This suggests a degree of identity loss in favour of network-driven incentives, supporting classical theories of crowd psychology.
While the current model is limited by its cross-sectional nature (preventing definitive causal claims) and the simplification of literary behaviour, it opens new avenues for quantitative sociology. Future work will explore specialized estimators for non-canonical networks and the use of hypergraphs to model complex group interactions.
Summary updated