1887
banner image
No data available.
Please log in to see this content.
You have no subscription access to this content.
No metrics data to plot.
The attempt to load metrics for this article has failed.
The attempt to plot a graph for these metrics has failed.
Multi-stage complex contagions
Rent:
Rent this article for
USD
10.1063/1.4790836
/content/aip/journal/chaos/23/1/10.1063/1.4790836
http://aip.metastore.ingenta.com/content/aip/journal/chaos/23/1/10.1063/1.4790836
View: Figures

Figures

Image of FIG. 1.
FIG. 1.

(a) Schematic of a single-stage complex contagion. All nodes can be either inactive (S 0) or active (S 1). Nodes that are barely above the S 1-threshold have the same level of influence as nodes that are strongly above that threshold. (b) Multi-stage complex contagion. A subset of active nodes (called “S 1-active”) can become hyper-active (called “S 2-active”) and have additional influence. Note that S 2-active nodes are necessarily also S 1-active.

Image of FIG. 2.
FIG. 2.

Comparison of single-stage [panel (a)] and multi-stage [panel (b)] cascade dynamics on the Oklahoma Facebook network. 56 Time t is on the horizontal axis, and we indicate the fractions of nodes in each state on the vertical axis. Light blue, blue (and purple), and red regions represent S 0-, S 1-, and S 2-active nodes, respectively. In panel (a), ; and in panel (b), . The threshold parameter values are and . We use 348 S 1-active seed nodes (corresponding to ) and zero S 2-active seeds ( ). There is no cascade in panel (a), but some of the nodes (colored in purple) are well above the S 1-threshold. The bonus influence of S 2-active nodes drives a cascade in panel (b).

Image of FIG. 3.
FIG. 3.

Comparison of single-stage [panel (a)] and multi-stage [panel (b)] cascade dynamics on the Oklahoma Facebook network. 56 As in Fig. 2 , the time t is on the horizontal axis, and we indicate the fractions of nodes in each state on the vertical axis. Light blue, blue, and red regions represent S 0-, S 1-, and S 2-active nodes, respectively. We use 348 S 2-active seed nodes (corresponding to ) and . In panel (a), , so the S 1 dynamics are slaved to the S 2 dynamics. A small change in the threshold parameter ( ) in panel (b) yields a small number of additional S 1-active nodes, which are nevertheless enough to trigger a cascade.

Image of FIG. 4.
FIG. 4.

Demonstration of dynamics when an S 2-cascade drives an S 1-cascade. Panels (a) and (b) show the aggregate fractions of S 1- and S 2-active nodes, panels (c) and (d) show these fractions for each degree class separately, and panel (e) shows the fractions of nodes in each degree class that are S 1-active but not S 2-active. We show the numerical results (averaged over 100 realizations) using symbols and the analytical results given by Eqs. (2) and (3) using curves. The timescales are independent of network size N, which we take to be . The values of the other parameters are , and . (We choose seed nodes uniformly at random.) We use an upper threshold of to model the single-stage case.

Image of FIG. 5.
FIG. 5.

Demonstration of dynamics when an S 1-cascade drives an S 2-cascade. Panels (a) and (b) show the aggregate fractions of S 1- and S 2-active nodes, panels (c) and (d) show these fractions for each degree class separately, and panel (e) shows the fractions of nodes in each degree class that are S 1-active but not S 2-active. We show the numerical results (averaged over 100 realizations) using symbols and the analytical results given by Eqs. (2) and (3) using curves. The timescales are independent of network size N, which we take to be . We use the values for the single-stage case [panels (a) and (c)] and the values and for the multi-stage case [panels (b), (d), and (e)]. The values of the other parameters are and (where we choose the seed nodes uniformly at random).

Image of FIG. 6.
FIG. 6.

Comparison of numerical computations (symbols) with analytical predictions of Eqs. (2) and (3) (curves) for (a) the final fractions of S 2-active nodes as functions of the bonus influence β for ensembles of (4,5)-uncorrelated random networks (blue dashed-dotted curve) and Erdős-Rényi random graphs with mean degree z = 5 (black dashed curve) and (b,c) the temporal evolution of active fraction of nodes in each degree class in (4,5)-uncorrelated random networks for . In panel (a), we use the response functions of Eq. (1) with and uniform thresholds and . For the ER graphs, we also use a solid red curve to show the case in which the R 2 thresholds are Gaussian-distributed with mean and standard deviation . The total number of nodes in each network is . For the numerical simulations, we initially S 1-activate a fraction of nodes chosen uniformly at random, and we average over 100 realizations of networks and initial conditions.

Image of FIG. 7.
FIG. 7.

Two-parameter bifurcation diagram for (whose value is indicated by color) calculated from Eqs. (2) and (3) for Erdős-Rényi random graphs. The mean degree z is on the horizontal axis, and the bonus influence β is on the vertical axis. The first threshold is , and the second threshold R 2 is Gaussian-distributed with mean and standard deviation . The initial seed fractions are and . The dashed curve gives the boundary of the cascade condition, and the solid curve is a numerical continuation of the saddle-node bifurcation.

Image of FIG. 8.
FIG. 8.

Two-parameter bifurcation diagram for (whose value is indicated by color) calculated from Eqs. (2) and (3) for Erdős-Rényi random graphs with mean degree z = 4. We plot the uniform thresholds R 1 and R 2 on the horizontal and vertical axes, respectively. The bonus influence is , and the initial seed fractions are and . The dashed red line gives the boundary of the cascade condition. The labeled regions (which are separated by dashed white lines) indicate cascades that are driven by (a) low influencers, (b) low and high influencers, and (c) high influencers. See the description in the text.

Image of FIG. 9.
FIG. 9.

Tree-like structure of a network near node A, which we treat as the root of the tree. For every two nodes connected by an edge (e.g., nodes B and C), the node that is closer to A is called the parent; thus, node B is the parent of node C, and node C is the child of node B. Only the influence from nodes within a distance n of node A can reach A in n time steps.

Image of FIG. 10.
FIG. 10.

This figure is analogous to Fig. 4 of the main text, but here we use a different value of the upper threshold R 2 for the multi-stage case [panels (b), (d), and (e)]. In this example, we use the threshold value ; in Fig. 4 , we used . As discussed in the text, the state-segregation effect is more pronounced for , which implies that all of the neighbors of degree-4 nodes must be S 1-active for such a node to become S 2-active (whereas only three S 1-active neighbors are needed for such S 2-activation when ). Our analytical approximation does not properly account for state segregation, and the consequences of that can be seen by comparing panels (d) in the two figures. The match with the numerical simulations in this figure is clearly worse than that in Fig. 4 .

Loading

Article metrics loading...

/content/aip/journal/chaos/23/1/10.1063/1.4790836
2013-02-21
2014-04-19
Loading

Full text loading...

This is a required field
Please enter a valid email address
752b84549af89a08dbdd7fdb8b9568b5 journal.articlezxybnytfddd
Scitation: Multi-stage complex contagions
http://aip.metastore.ingenta.com/content/aip/journal/chaos/23/1/10.1063/1.4790836
10.1063/1.4790836
SEARCH_EXPAND_ITEM