Guided rewiring of social networks reduces polarization and accelerates collective action

TL;DR

This work investigates how guided link rewiring influences polarization and cooperative action in social networks using a principled agent-based model. It compares heuristic rewiring strategies (random, local, bridge) with topology-based recommenders (Who to Follow and node2vec) across synthetic and empirical networks, accounting for backfire dynamics and an external pro-cooperation field. Key findings show heterophilic (opposite) rewiring often accelerates consensus and can sustain cooperation under higher levels of backfiring, while random rewiring frequently outperforms many complex algorithms and certain recommender methods; WTF can reduce cooperation and inflate inequality depending on network structure. The results reveal a nuanced interplay between topology, rewiring strategy, and polarization dynamics, with implications for designing social platforms to promote depolarization and collective action despite backfire effects.

Abstract

Global social and ecological challenges represent collective action problems requiring rapid and sufficient cooperation with pro-mitigation norms. Sociopolitical polarization hinders such cooperation. Prior agent-based models showed polarization emerges naturally in structured social networks and polarized cluster dissolution rate limits consensus formation rate. Here we study how guided link rewiring affects depolarization dynamics across synthetic and empirical (Facebook, Twitter) network topologies. We compare heuristic rewiring algorithms representing random meetings, mutual acquaintance introductions, and community bridging, alongside topology-based link recommender algorithms (Who to Follow and node2vec). Our heuristic algorithms all outperform Who to Follow in generating cooperative consensus. Homophilic rewiring generates cooperative consensus when agents can easily change opinions. However, heterophilic rewiring achieves this over broader conditions and can accelerate cooperative consensus formation by ~20%, including where up to 33% of the population experiences backfiring interactions. Heterophilic rewiring also vastly outperforms topology-based recommender algorithms. Random rewiring performed consistently well, achieving higher steady-state cooperation than seven out of eight more complex algorithms. Large disparities in steady-state cooperation for topology-based recommender systems highlight their volatility across network structures. Overall, our work reveals a subtle interplay between topology, rewiring algorithm and social depolarization, suggesting strong potential for carefully redesigning social networking technologies for social good.

Figures (4)

• Figure 1: Rewiring algorithms. The colour of a node shows the sign of its opinion (blue = cooperator, orange = defector). We focus on individual $i$ and mark available rewiring targets with dashed circles. A and B show local rewiring, in which $i$ can choose among agents within 2 network steps (friends of friends). Dashed lines show an exemplary new link. In A (B) $i$ only links to an agent whose opinion is similarly (oppositely) signed. In D and E, $i$ can establish links to agents outside its own topological cluster and the opinion constraint (similar, opposite) is the same as above. C illustrates the Who to Follow algorithm, which calculates a circle of trust (COT, green dashed lines) for $i$ and suggests a connection to an agent commonly followed by the COT. F shows node2vec, which calculates and compares embeddings for all nodes, then suggests that $i$ should link to the agent with the highest structural similarity to itself. On directed networks, we assume an agent $i$ chooses others to follow by selecting from outgoing links.
• Figure 2: Comparing the evolution of network structure and cooperation across all rewiring scenarios. (a) A dynamic network topology does not guarantee higher cooperation or lower polarization than a static baseline (orange line). Of those that do, $x(\text{similar})$ and random rewiring algorithms consistently support the emergence of positive cooperation in the steady state (solid arrows). (b) Representative ($N=100)$ network states for selected rewiring algorithms, given a shared, fixed initial ($t=0$) network state. The Who to Follow (WTF) algorithm dramatically increases in-degree inequality and clustering producing the visible core-periphery structure, with the high in-degree nodes on the outside, and followers clustered around. This is less pronounced in the local(similar) and bridge(opposite) algorithms. Local$(y)$ algorithms cluster nodes together more tightly than the $\text{bridge}(y)$ algorithms, which have a more diffuse core, for example in bridge(opposite) (B-opp). This allows nodes with contrasting opinions to meet more frequently, and drives cooperation, especially when subject to small network sizes (five initial cooperators). Final network states shown in (b) do not necessarily correspond to network evolutions with default $N$ i.e., $N=800$. For example the special case is shown for B-sim where given $N<150$, the network separates into two opinion clusters, showing strong structural polarization. Otherwise, all results were simulated with default parameters (See \ref{['sec:methods']}).
• Figure 3: Convergence rates and steady state cooperation, $\langle a^* \rangle$ for all rewiring algorithms under default parameters. The grey reference line $x+y=1$ represents a perfect trade off between convergence rate and final cooperation, i.e., where improvements in convergence rate necessarily reduce final cooperation. Variance is highest in the mid point along the reference and collapses toward a convergence rate of $1$. The red dashed line estimates the Pareto front by indicating scenarios that have no better alternative in terms of final cooperation and convergence rate. Random rewiring and static scenarios represent balanced solutions, but are outperformed by bridge(similar) on the Twitter and FB topologies, which represent the most optimal solutions. Local(similar) on CSF, and random on FB were also optimal solutions, but skewed to favour final cooperation and convergence rate respectively.
• Figure 4: (a) Effect of varying stubbornness, $w$, and diverger fraction, $\rho$ on equilibrium cooperation, $\langle a^* \rangle$ and polarization $\langle P^* \rangle$. Results on the twitter topology are qualitatively similar to other topologies. Cooperation is impossible in the region $\rho > 0.33$ for all algorithm-topology pairs except Who to Follow (WTF) on the directed preferential attachment (DPA) topology (not shown) ($\rho > 0.53$). Our $x(\text{opposite})$ algorithms (heterophilic) resulted in a greater number of cooperative steady states ($\langle a^* \rangle > 0$), and higher ensemble final cooperation, but also higher polarization. Stubbornness mitigated the otherwise strictly negative effect of divergers on cooperation. (b) Change in final cooperation and polarization over stubbornness values ($w$). Results are grouped by opinion constraint, i.e., opposite and similar. WTF and $x(\text{similar})$ algorithms produces severely non-monotonic outcomes with increasing stubbornness values. WTF leads to poor cooperation in the low ($\rho \leq 0.33$) and medium ($\rho \leq 0.66$) stubbornness regimes, which represent plausible interaction conditions in online social media exchanges.