Cascades on clique-based graphs

We present an analytical approach to determining the expected cascade size in a broad range of dynamical models on the class of highly-clustered random graphs introduced by Gleeson [J. P. Gleeson, Phys. Rev. E 80, 036107 (2009)]. A condition for the existence of global cascades is also derived. Applications of this approach include analyses of percolation, and Watts’s model. We show how our techniques can be used to study the effects of in-group bias in cascades on social networks.

💡 Research Summary

The paper tackles the long‑standing problem of predicting cascade dynamics on networks that exhibit strong clustering, a feature that standard random‑graph models (Erdős–Rényi, configuration models) fail to capture. The authors adopt the “clique‑based graph” framework introduced by Gleeson (2009), in which the network is constructed from a mixture of complete subgraphs (cliques) of varying sizes. Each vertex may belong to several cliques, so that intra‑clique connections are essentially dense (clustering coefficient close to one) while inter‑clique links remain sparse. This construction reproduces the high clustering and modularity observed in many empirical social and biological networks.

The core contribution is an analytical method that yields the expected cascade size for a broad class of threshold‑type dynamical processes on such graphs. The authors first formalize the graph generation: a prescribed distribution (P(k)) determines the probability that a randomly chosen clique has size (k); vertices are assigned to cliques accordingly, producing a joint degree distribution that reflects both intra‑clique and inter‑clique edges.

Next, they introduce a “clique activation function” (g_k(m)), which gives the probability that a clique of size (k) becomes fully active when (m) of its members are already active. This function is derived by combining the underlying node‑level activation rule—expressed as a generic threshold function (f(\theta))—with the binomial statistics of neighbor states inside a clique. By treating each clique as a super‑node (or “hyper‑node”), the cascade dynamics on the original graph can be mapped onto a dynamics on a reduced network of hyper‑nodes. The interaction between hyper‑nodes is captured by a cascade matrix (M), whose entry (M_{ij}) represents the conditional probability that hyper‑node (j) activates given that hyper‑node (i) is active.

The expected fraction of active vertices after any number of update steps is then obtained by iteratively applying (M) to an initial activation vector. In the thermodynamic limit (infinite network size), the condition for a global cascade reduces to a spectral criterion: the largest eigenvalue (\lambda_{\max}) of (M) must exceed unity. This mirrors the classic branching‑process condition for unclustered networks, but here (\lambda_{\max}) is a function of the clique‑size distribution and the intra‑clique activation probabilities, making it highly sensitive to clustering.

To demonstrate the utility of the framework, the authors apply it to two canonical models.

Percolation: By interpreting node or edge removal as a cascade with a deterministic activation rule (a node becomes “inactive” if it loses all its connections), they derive the percolation threshold for clique‑based graphs. The analysis shows that the threshold is largely governed by the density of inter‑clique links; dense intra‑clique connectivity makes cliques robust, so the network disintegrates only when enough inter‑clique edges are removed.
Watts’s Threshold Model: Here each node adopts a new state if a fraction (\theta) of its neighbors are already active. The authors compute the cascade condition as a function of (\theta) and the clique‑size distribution. Small (\theta) values lead to global cascades even from a tiny seed, especially when large cliques are abundant, because the dense intra‑clique neighborhoods quickly satisfy the threshold. Conversely, larger (\theta) values require a higher proportion of large cliques to sustain propagation across the sparse inter‑clique backbone.

Finally, the paper explores “in‑group bias” by assigning different activation probabilities to intra‑clique versus inter‑clique edges (parameters (\alpha) and (\beta)). Increasing (\alpha) (stronger intra‑group influence) confines cascades within cliques, suppressing global spread, while decreasing (\beta) (weaker inter‑group resistance) facilitates cross‑clique propagation. This quantitative treatment provides a mechanistic explanation for empirical observations that homophily and group loyalty can either hinder or accelerate information diffusion, depending on the balance of internal and external influence.

The authors validate their analytical predictions with extensive Monte‑Carlo simulations, finding excellent agreement across a range of parameter settings. Their results not only extend cascade theory to highly clustered topologies but also offer a versatile toolbox for studying contagion, opinion dynamics, and systemic risk in realistic networked systems where community structure is pronounced. The paper thus bridges a critical gap between abstract random‑graph theory and the nuanced behavior of cascades on real‑world, modular networks.