The Physics of News, Rumors, and Opinions

TL;DR

It is argued that statistical physics provides a suitable and necessary framework for analyzing the unfolding of these complex dynamics on socio-technological systems and highlighted the valuable insights obtained from physics-based efforts to investigate these phenomena of high societal impact.

Abstract

The boundaries between physical and social networks have narrowed with the advent of the Internet and its pervasive platforms. This has given rise to a complex adaptive information ecosystem where individuals and machines compete for attention, leading to emergent collective phenomena. The flow of information in this ecosystem is often non-trivial and involves complex user strategies from the forging or strategic amplification of manipulative content to large-scale coordinated behavior that trigger misinformation cascades, echo-chamber reinforcement, and opinion polarization. We argue that statistical physics provides a suitable and necessary framework for analyzing the unfolding of these complex dynamics on socio-technological systems. This review systematically covers the foundational and applied aspects of this framework. The review is structured to first establish the theoretical foundation for analyzing these complex systems, examining both structural models of complex networks and physical models of social dynamics (e.g., epidemic and spin models). We then ground these concepts by describing the modern media ecosystem where these dynamics currently unfold, including a comparative analysis of platforms and the challenge of information disorders. The central sections proceed to apply this framework to two central phenomena: first, by analyzing the collective dynamics of information spreading, with a dedicated focus on the models, the main empirical insights, and the unique traits characterizing misinformation; and second, by reviewing current models of opinion dynamics, spanning discrete, continuous, and coevolutionary approaches. In summary, we review both empirical findings based on massive data analytics and theoretical advances, highlighting the valuable insights obtained from physics-based efforts to investigate these phenomena of high societal impact.

Figures (8)

• Figure 1: Models of cultural diffusion---an illustration adapted from Jordan's 1982 The Human Mosaicjordan1982human. (Slightly modified for consistency.) Note that "diffusion" in this figure is used in the sense of the social sciences and computer science: I.e., it does not preclude a change in the total mass of whatever spreads. Note also that the order of spreading is no longer thought to be determined by socioeconomic status rosen2002anatomy.
• Figure 2: A visualization from Ref. liben-nowell2008tracing showing the propagation of chain-letter e-mails petitioning against the 2002--2003 Iraq war. The generation number of the email increases downwards.
• Figure 3: A figure from Ref. ugander2012structural (studying the early growth of Facebook---the time when the platform primarily grew by email invitations) showing the paper's main finding. In (A), the black nodes are the focal node's friends who have sent the focal node an email invitation. This defines the "contact neighborhood" which the paper claims is the strongest determinant of the conversion rate (successful recruitment per time). Panels (B–D) show the relative conversion rates for two-node, three-node, and four-node contact neighborhood graphs. Shading indicates differences in component count. Invitation conversion rates are reported on a relative scale, where 1.0 signifies the conversion rate of one-node neighborhoods. Error bars represent 95% confidence intervals and implicitly reveal the relative frequency of the different topologies.
• Figure 4: Mesoscale structures in online social networks. Bow-tie structure in a complex network highlights the organization into a core, identified by a strongly connected component (SCC) where all users can reach each other through information diffusion and peripheral components. The IN component consists of users who can reach the SCC but not vice versa, whereas the OUT component consists of users from whom the SCC can reach but cannot. Tendrils attach to these sub-systems without connecting to the core. See the text for further details. Figure reproduced from Ref. han2024modelling.
• Figure 5: In (a), sketch representing the logic behind the physically-inspired metrics introduced in the main text to characterize opinion dynamics. In the two regular networks, nodes (agents) can be in any of two states $S_i=\pm 1$, here represented by the two different colors. Both networks have the same value for the magnetization $m = \sum_i S_i / N$, since the proportion of the two opinions is the same. On the contrary, the density of active links $\rho$ measures the level of order in the system, and it is lower in the above network because opinion clusters are larger there. In (b), sketch representing a single update in the voter model. The central agent is chosen to attempt an opinion update and adopts the opinion of a neighbor selected uniformly at random. In this case, it becomes green and orange with probability $3/4$ and $1/4$, respectively. In (c), we display how $\rho(t)$ captures different ways to reach the consensus state, i.e., $|m| = 1$ and $\rho = 0$: either through a coarsening process ($\rho(t) \sim t^{-1/2}$; voter model in a one-dimensional lattice) or through a finite-size fluctuation after the stabilization in a metastable plateau $\rho^{\text{st}} = \frac{1}{2}\frac{\langle k \rangle - 2}{\langle k \rangle - 1}$ for a timescale that grows with $N$, where $\langle k \rangle$ is the mean degree of the complex network on top of which the voter model evolves. In (d), we sketch the functional form of the probability for an agent to change state depending on the fraction of her neighbors in the opposite state, for different opinion dynamics models explained in the main text. In (e), we show individual trajectories of the magnetization. For the standard voter model ($a=0$), 50 independent trajectories are presented, all departing from $m(0) = 1/3$. We can appreciate how the average magnetization is conserved over the ensemble of trajectories (solid line). In the other panels, we show a single trajectory of the noisy voter model in the bimodal ($a < a_c$) and unimodal ($a > a_c$) phase, and one at the critical value of the noise ($a = a_c$). This panel has been adapted from artime2019herding. In (f), stationary probability distribution for the magnetization $P(m)$ for the nonlinear noisy voter model (with exponent $\alpha = 6$) on the left and for the noisy voter model on the right. In the panels of the top row, the system is kept symmetric and, as noise is increased, we observe a transition from a bimodal to a trimodal distribution and from a trimodal to a unimodal one. In the bottom panels, the noise is kept constant and the asymmetry is varied. We identify a transition from a trimodal to a bimodal distribution and a transition from bimodal to unimodal one. Markers come from numerical simulations, while solid lines are the analytical approximations. This panel has been adapted from peralta2018analytical. The rightmost plot shows the modality transition for the noisy voter model. Lines corresponds to the noise values used in the magnetization trajectories in (e).