A Simulation Framework for Studying Recommendation-Network Co-evolution in Social Platforms

TL;DR

This work presents a closed-loop agent-based simulator that jointly models content production, tie formation, and a learnable GAT recommender to study how recommendations co-evolve with social networks. Calibrated on Mastodon and validated against Bluesky, the framework enables counterfactual experiments, scale-up analyses, and sensitivity checks across activation timing, user composition, and exploration. Key findings show that delaying recommendation activation reduces transitivity and increases content diversity, effects that persist but attenuate with larger scale, and that recommender architecture has less impact than the timing of activation. The framework offers reproducible schemas and scripts to stress-test design choices for federated platforms, informing policy and platform design before live deployment. Collectively, the work advances understanding of causal mechanisms in recommendation-driven network evolution and provides a practical tool for pre-deployment experimentation and governance.

Abstract

Studying how recommendation systems reshape social networks is difficult on live platforms: confounds abound, and controlled experiments risk user harm. We present an agent-based simulator where content production, tie formation, and a graph attention network (GAT) recommender co-evolve in a closed loop. We calibrate parameters using Mastodon data and validate out-of-sample against Bluesky (4--6\% error on structural metrics; 10--15\% on held-out temporal splits). Across 18 configurations at 100 agents, we find that \emph{activation timing} affects outcomes: introducing recommendations at $t=10$ vs.\ $t=40$ decreases transitivity by 10\% while engagement differs by $<$8\%. Delaying activation increases content diversity by 9\% while reducing modularity by 4\%. Scaling experiments ($n$ up to 5,000) show the effect persists but attenuates. Jacobian analysis confirms local stability under bounded reactance parameters. We release configuration schemas and reproduction scripts.

Figures (6)

• Figure 1: System architecture showing drivers and outcomes. Drivers (homophily, activity, exploration) shape agent actions and discovery; mechanisms (content dynamics, recommendation) mediate exposure and learning; outcomes are measured by network and content metrics (e.g., $\rho$, $C$, $Q$, $\ell$, diversity). Arrows indicate information flow.
• Figure 2: Impact of recommendation activation timing on (a) content diversity, (b) user retention, and (c) engagement rate for early (t=10), standard (t=25), and late (t=40) activation. Lines show mean with 95% confidence intervals (error bars).
• Figure 3: Impact of exploration rate on (a) content spread, (b) diversity, and (c) user engagement. Points and lines show mean with 95% confidence intervals; consistent axes improve cross-panel comparison.
• Figure 4: Diachronic evolution of key network metrics in simulation: density ($\rho$), clustering ($C$), and average path length ($\ell$) over 50 steps. Lines show mean across runs with 95% confidence intervals (shaded). When observed time series are available, we overlay them (red) for direct comparison.
• Figure 5: Network evolution split by experiment type. Rows: activation, composition, scale, exploration. Columns: density ($\rho$), clustering ($C$), average path length ($\ell$). Lines show mean with 95% confidence intervals for with (blue) and without (red) recommendations. Late-run gaps are larger than early-run gaps across all types.