Threshold-limited spreading in social networks with multiple initiators

A classical model for social-influence-driven opinion change is the threshold model. Here we study cascades of opinion change driven by threshold model dynamics in the case where multiple {\it initiators} trigger the cascade, and where all nodes possess the same adoption threshold $\phi$. Specifically, using empirical and stylized models of social networks, we study cascade size as a function of the initiator fraction $p$. We find that even for arbitrarily high value of $\phi$, there exists a critical initiator fraction $p_c(\phi)$ beyond which the cascade becomes global. Network structure, in particular clustering, plays a significant role in this scenario. Similarly to the case of single-node or single-clique initiators studied previously, we observe that community structure within the network facilitates opinion spread to a larger extent than a homogeneous random network. Finally, we study the efficacy of different initiator selection strategies on the size of the cascade and the cascade window.

💡 Research Summary

The paper investigates opinion spreading in social networks using the classic threshold model, extending the analysis from the well‑studied single‑seed or small‑clique initiator cases to scenarios where a finite fraction p of nodes act as initiators. All nodes share the same adoption threshold φ, and a node becomes active (adopts the opinion) permanently once at least a fraction φ of its neighbors are active. The authors explore how the cascade size S depends on the initiator fraction p, the threshold φ, and the underlying network structure.

First, the authors confirm that for any φ < 1 there exists a critical initiator fraction p_c(φ) such that when p > p_c a discontinuous transition to a global cascade occurs. This “tipping point” is independent of system size and can be identified by the maximum of the derivative d \tilde{S}/dp, where \tilde{S}= (S−p)/(1−p) measures the fraction of non‑initiators that eventually adopt. As φ increases, p_c rises smoothly, and for dense networks (large average degree ⟨k⟩) the relationship approaches the trivial limit p_c = φ of a complete graph.

The study then examines the role of average degree. In Erdős‑Rényi (ER) graphs, cascades are suppressed when ⟨k⟩ is too low (the network fragments into small components) or too high (each node requires many active neighbors to satisfy the threshold). An intermediate “cascade window” of ⟨k⟩ values maximizes the range of p that yields global cascades. Within this window, increasing p widens the window, allowing cascades even at higher ⟨k⟩.

Three heuristic initiator‑selection strategies are compared: (i) random selection, (ii) descending degree order, and (iii) descending k‑shell index. Simulations on ER graphs (N = 10⁴, φ = 0.18) show that degree‑based selection consistently produces the largest cascade size and the fastest propagation. The advantage stems from high‑degree nodes being hard to influence when they are not seeds, yet capable of activating many low‑degree neighbors once seeded. The k‑shell method, while also targeting high‑degree nodes, tends to pick nodes embedded in dense high‑degree neighborhoods, reducing the number of easily influenceable neighbors and thus performing worse than pure degree selection. Random selection performs the poorest because it often picks low‑degree nodes that quickly encounter high‑degree bottlenecks.

To provide an analytical baseline, the authors adapt a tree‑approximation method previously used for single‑seed cascades. This approach treats the network as locally tree‑like, computes the cascade progression from the periphery toward the root, and yields an estimate of p_c(φ). The approximation matches simulation results for low ⟨k⟩ and small p but deviates for larger degrees or larger initiator fractions, where loops and clustering invalidate the tree assumption.

The impact of network topology beyond average degree is investigated using an empirical high‑school friendship network (Add Health data, N ≈ 921, ⟨k⟩ ≈ 5.96) that exhibits strong community structure and high clustering. The authors generate two ensembles of degree‑preserving randomized networks: (1) an x‑swap procedure that rewires edges while preserving degree sequence but reduces clustering, and (2) a fully random configuration model that eliminates both clustering and community structure. Keeping p = 0.01, they find that the original network yields significantly larger cascade sizes and higher probabilities of global cascades than either randomized counterpart. This demonstrates that local clustering (triadic closure) and modular community organization facilitate the simultaneous activation of multiple neighbors, thereby easing the satisfaction of the threshold condition.

Overall, the paper delivers several key insights: (1) a finite initiator fraction can overcome high adoption thresholds that would otherwise block cascades; (2) the critical initiator fraction p_c is a smooth function of φ and converges to φ in the dense‑graph limit; (3) the cascade window is governed by average degree, with an optimal intermediate range; (4) selecting initiators by descending degree maximizes both cascade size and speed, outperforming k‑shell and random strategies; (5) clustering and community structure markedly enhance cascade likelihood, confirming earlier findings for single‑seed cascades; and (6) analytical tree‑based approximations are useful but limited to sparse, low‑p regimes.

These results have practical implications for viral marketing, public‑health campaigns, and any scenario where a coordinated set of early adopters is deployed to induce widespread behavioral change. Targeting high‑degree individuals, especially those situated at the interface of densely clustered communities, can dramatically reduce the required seed size and accelerate diffusion, even when individuals are relatively resistant (high φ). The work thus bridges a gap between theoretical cascade models and real‑world strategies for influencing large social systems.