Segregation Before Polarization: How Recommendation Strategies Shape Echo Chamber Pathways

TL;DR

This work investigates how recommendation strategies shape echo chambers by extending the dynamic bounded confidence model to jointly evolve agent opinions and follower networks under content-based and link-based recommendations. It demonstrates a segregation-before-polarization (SbP) pathway induced by content-based curation, where structural segregation precedes opinion divergence and accelerates collective polarization, while link-based strategies tend toward a polarization-before-segregation (PbS) trajectory. A paradoxical effect of reposting is shown: while increasing network connections, reposts amplify latent opinion differences and deepen echo chambers, especially in SbP regimes. To quantify these dynamics, the authors introduce indices for structural homophily ($I_h$), polarization ($I_p$, $I_s$), and the pathway ($I_w$), enabling stage-aware interventions that transition from content-centric to structure-centric strategies as the network evolves, with practical implications for mitigating polarization on real platforms.

Abstract

Social media platforms facilitate echo chambers through feedback loops between user preferences and recommendation algorithms. While algorithmic homogeneity is well-documented, the distinct evolutionary pathways driven by content-based versus link-based recommendations remain unclear. Using an extended dynamic Bounded Confidence Model (BCM), we show that content-based algorithms--unlike their link-based counterparts--steer social networks toward a segregation-before-polarization (SbP) pathway. Along this trajectory, structural segregation precedes opinion divergence, accelerating individual isolation while delaying but ultimately intensifying collective polarization. Furthermore, we reveal a paradox in information sharing: Reposting increases the number of connections in the network, yet it simultaneously reinforces echo chambers because it amplifies small, latent opinion differences that would otherwise remain inconsequential. These findings suggest that mitigating polarization requires stage-dependent algorithmic interventions, shifting from content-centric to structure-centric strategies as networks evolve.

Figures (11)

• Figure 1: (a), (c) Evolution of a random network (500 agents) under random recommendation. (a) PbS pathway ($\alpha = 0.05, q=0.025$); (c) SbP pathway ($\alpha = 0.005, q=0.025$). (b), (d) Microscopic dynamics showing Normalized Opinion Difference (NOD) vs. current opinion. (b'), (d') Effective potential landscapes $V(x)$ derived from (b) and (d). Color gradient indicates normalized time $t_n$, showing the formation and flattening of the double-well potential.
• Figure 2: (a) Typical evolutionary pathways of the society model. From a (1) random state, the model evolves to (4) consensus or (5) segregation via (2) SbP or (3) PbS pathways. (b) Trajectories corresponding to these different pathways plotted on the $I_p$-$I_h$ diagram.
• Figure 3: Heatmaps of simulation results. (a--c) Pathway index $I_w$. (d--f) Polarization trajectory index $I^T_p(t_a)$ for consensual scenarios. Columns: (a, d) structure-based, (b, e) opinion-based, (c, f) differences. $I^T_p(t_a)$ are omitted for segregated cases as well as $I^T_h(t_a)$, as they consistently remain near 1.
• Figure 4: Impact of Reposting on Evolutionary Outcomes. (a--c) Probability Density Functions (PDF) of the pathway index $I_w$, estimated via Gaussian KDE. Colors indicate the final state: Consensual (1 peak) or Polarized ($\ge 2$ peaks). (a) shows the aggregate bimodal distribution, while (b) and (c) compare scenarios without reposting ($p=0$) and with reposting ($p>0$). (d, e) Temporal evolution of Social Force for (d) PbS and (e) SbP cases. (f) Aggregated view for SbP cases, showing that reposting lowers the mean opinion shift but drastically increases the variance.
• Figure 5: Comparison of feature indices across different algorithmic and societal characteristics. Panels show the relationship between simulation conditions and: (a, e) the total count of rewiring events; (b, f) the normalized occurrence time of rewiring events; (c, g) the AUC of the subjective polarization index $I_s$; and (d, h) the final number of closed triads. Top row (a--d) aggregates results by recommendation algorithms (Structure- vs. Opinion-based), while bottom row (e--h) aggregates by reposting behavior (!R: no repost; R: repost). Error bars indicate standard deviations.