Beyond network centrality: Individual-level behavioral traits for predicting information superspreaders in social media

TL;DR

A nonlinear network-based algorithm is derived to quantify individuals’ influence and susceptibility from multiple spreading event data and enable predictions of future superspreaders above and beyond network centrality, and reveal new insights into the network positions of the superspreaders.

Abstract

Understanding the heterogeneous role of individuals in large-scale information spreading is essential to manage online behavior as well as its potential offline consequences. To this end, most existing studies from diverse research domains focus on the disproportionate role played by highly-connected ``hub" individuals. However, we demonstrate here that information superspreaders in online social media are best understood and predicted by simultaneously considering two individual-level behavioral traits: influence and susceptibility. Specifically, we derive a nonlinear network-based algorithm to quantify individuals' influence and susceptibility from multiple spreading event data. By applying the algorithm to large-scale data from Twitter and Weibo, we demonstrate that individuals' estimated influence and susceptibility scores enable predictions of future superspreaders above and beyond network centrality, and reveal new insights on the network position of the superspreaders.

Figures (4)

• Figure 1: Quantifying individual influence and susceptibility from observational data. To illustrate the intuition behind the proposed influence-susceptibility algorithm, in panel a, we show how two nodes (A, B) with the same number of outgoing links and the same number of propagation events can achieve a widely different influence score $I$. The size of orange (blue) nodes denotes their susceptibility score $S$ (influence score $I$); the thickness of the arrows represents the number of propagation events. Both A and B have three outgoing links, meaning that they influence three nodes. However, A's neighbors have a lower susceptibility than node B's neighbors, because they are influenced fewer times by their other neighbors (the blue nodes in the outer-most shell). As A influences less susceptible individuals than B does, according to the IS algorithm, A is more influential than B. With a similar argument, in panel b, node C is more susceptible than D.
• Figure 2: Empirical correlations between individual-level properties. We divide the spreading events into $6$ consecutive non-overlapping periods. The date on the $x$-axis represents the starting date of each period. In each period, we reconstruct individual influence and susceptibility via the IS algorithm, and measure the Spearman's correlation coefficients, $\rho$, between indegree ($k^{in}$), outdegree ($k^{out}$), influence ($\hat{I}$) and susceptibility ($\hat{S}$). Only those value pairs where both values are not equal to $0$ are conserved. Filled markers denote correlation values that are significantly larger or smaller than the correlation values calculated on randomized networks ($P<0.05$, see Supplementary Note $4$ for details); empty markers denote correlation values that do not significantly differ from the correlation values calculated on randomized networks ($P>0.05$). The strongest consistent empirical correlation is the one between outdegree ($k^{out}$) and influence ($\hat{I}$); the other correlations are weak or non-significant.
• Figure 3: Empirical assortativity properties between degree, influence, and susceptibility. We divide the spreading events into 6 consecutive non-overlapping periods. The date on the $x$-axis represents the starting date of each period. In each period, we reconstruct individual influence and susceptibility via the IS algorithm, and we measure Spearman's correlation coefficients, $\rho$, between an individual's properties and her neighbors' properties. The only consistent correlations are the positive one between an individual's susceptibility ($\hat{S}$) and the influence of the individuals she retweets ($\hat{I}_{nn}^{in}$), and the negative one between an individual's outdegree ($k^{out}$) and the susceptibility of the individuals who retweet from her ($\hat{S}_{nn}^{out}$).
• Figure 4: Predicting superspreaders. We rely on the Random Forest classification algorithm and use different features as input to predict superspreaders, where the superspreaders are defined as individuals with top $5\%$ spreading capacity. Panels a, b, d, e show the superspreaders predicting performance on Weibo COVID and Twitter COVID. Two metrics, AURPC, and precision, are adopted to measure the performance of models. Panels c, f show the feature importance resulting from training the combined model. Across all windows, the best-performing model is either the combined model or the Behavior-based (IS) model, which points to the essential role of the IS scores for the superspreader prediction. Panels c, f show the feature importance obtained from training the combined model. IS-based features tend to be more important than centrality-based features.