Social clustering reinforces external influence on the majority opinion model

Abstract

Public opinion is subject to peer interaction via social networks and external pressures from the media, advertising, and other actors. In this paper, we study the interaction between external and peer influence on the stochastic opinion dynamics of a majority vote model. We introduce a model where agents update their opinions based on the combined influence from their local neighbourhood (peers) and from an external actor in the transition rates. In the first model, the external influence is only felt by agents non-aligned with the external actor (”push strategy”). In the second model, agents are affected by external influence, independently of their opinions (”nudging strategy”). In both cases, the external influence increases the possible macroscopic outcomes. These outcomes are determined by the chosen strategy. We also find that the social network structure affects the opinion dynamics, with high social clustering positively reinforcing the external influence while degree heterogeneity weakens it. These findings are relevant to businesses and policy making, hel** to understand how groups of individuals collectively react to external actors.

[inst1]organization=Department of Data-Analysis, Ghent University,addressline=Henri Dunantlaan 1, city=Ghent, postcode=9000, country=Belgium

[inst2]organization=Department of Physics and Astronomy, Ghent University,addressline=Proeftuinstraat 86, city=Ghent, postcode=9000, country=Belgium

[inst3]organization=Department of Economics, Ghent University,addressline=Sint-Pietersplein 5, city=Ghent, postcode=9000, country=Belgium

The interaction between peer pressure and an external field leads to more complex behaviour than reinforcing the external opinion.

The intensity and weight of an external field push the population into a greater variation of macroscopic states.

Macroscopic outcomes depend on the alignment of agents to the external field.

Social clustering reinforces the influence of an external field.

Degree heterogeneity and hubs hinder the effect of an external field.

1 Introduction

Statistical physics offers helpful tools to study social systems involving collective and emergent phenomena [1, 2, 3, 4, 5]. In particular, physical models have been used to explore and understand the collective behaviour of opinions in society [6, 7]. Supporting the idea that human behaviour is not wholly unpredictable [8]. Given the universality in order-disorder transitions, it is appropriate to use physics-based models to investigate large-scale regularities as collective effects. Models composed of adaptive agents have been effective at describing complex social behaviour [9, 10, 11]. The rules and constraints of those models, for example, threshold imitation or homophily [12], should be informed by psychology and social science. Opinion dynamics is a multidisciplinary field built on early sociological observations that people tend to conform to each other’s behaviours and opinions. This led to questions about collective behaviour and inspired early models in sociology and economics, such as the Schelling model [13] and Axelrod’s cultural dissemination model [14]. Behaviour and opinions can only spread when there is interaction, implying the importance of the structure of the underlying contact networks [15].

In public opinion formation, next to interpersonal social interaction, also mass media and other external actors play a role [16, 17, 18, 19, 20]; thus opinion dynamics is subject to the interaction of personal predispositions, external media influence and social interaction [21]. We are interested in exploring the interaction between the internal peer pressure towards conforming, and an external influence. In the last decades, many models for opinion dynamics have been created for discrete and continuous opinion spaces or a combination of both [22]. Voting problems have served as early applications. Examples include the well-known Voter Model [23, 24] and its many variations [25]. We focus here on binary opinion models, particularly complex contagions [26], where opinions require reinforcement to become more easily adapted. This idea of reinforcement has been implemented in different ways. For example, by interacting with more than one neighbour [27, 28, 29, 30, 31], via the use of memory [32], by introducing different levels of confidence in the opinion [33], or by repeated interactions [34] and algorithmic personalisation  [35, 36]. We will follow the first approach and have individuals take their entire social neighbourhood into account when updating their opinions.

Several investigations have been conducted to include an external influence, such as the presence of media, in opinion dynamics models. In the binary opinion case, for the Voter Model and some of its extensions [37, 38, 39], the Majority Rule model [40] and the Sznajd Model [41]. Often, at every time step, there is a probability $r$ with which an agent interacts with its environment, feeling only internal pressure, and a probability $1-r$ they only feel the external media influence in that time step. The model we introduce in this paper makes a different assumption that agents feel both internal and external pressure at every time step. This allows us to understand the effect of opposing and reinforcing interactions simultaneously; an agent does not feel external influence independently from their neighbourhood pressure. This paper also addresses the impact of external influence applied to complex contagion. It proposes asymmetric scenarios where the external influence is felt depending on the (opinion) alignment of the agent with the media. The concurrent influence on complex contagions and the comparison between homogeneous and alignment-based influence have been subject to little investigation in the literature [7].

This paper introduces an opinion dynamics model containing the following properties: (i) complex contagion spread, (ii) concurrent interaction between internal and external pressure, (iii) social network structure, and (iv) stochasticity. The latter is chosen to resemble real-world noise and incomplete information during decision-making. To capture the above properties, we propose a model where, at each time step, a randomly chosen agent updates their opinion with a certain transition rate that results from the interaction of internal and external influence. Internal influence contributes to the transition rate according to the fraction of nearest neighbours disagreeing with the chosen agent. The external influence adds or subtracts from that transition rate based on the disagreement or agreement between the agent and the media, respectively. The media influence will have a strength fixed in time, and the relative weight for internal and external pressures is set by a parameter $\alpha$. We test two versions of this model: an asymmetric version where the media only influences agents that disagree with it and a symmetric version where every agent, independently of its own opinion, feels the external media influence such that the media adjusts transition rates in both directions. Both versions enrich the space of possible consensus regimes over the absence of external influence, with the asymmetric version showing the most variation concerning possible final consensus states.

2 Model

2.1 Networks

A network comprises a set of $N$ nodes $i$ connected by $M$ links $(i,j)$, where we exclude self-links $(i,i)$ and multiple links between the same node pair. We will restrict ourselves to undirected networks; a link $(i,j)$ implies a symmetric link $(j,i)$. The topology is characterised by a symmetric matrix $\boldsymbol{A}$, the adjacency matrix, with entries $a_{ij}=1$ if a link exists between node $i$ and $j$, and zero otherwise. Each node $i$ has a degree $k_{i}=\sum_{j}a_{ij}$ representing the number of other nodes $j$ to which $i$ is connected. All networks used in this work are connected, i.e. $k_{i}>0,\;\forall i$, and have $N=1000$ nodes with $\langle k\rangle\approx 10$. The average shortest path length is given by $\langle l\rangle=\frac{1}{N(N-1)}\sum_{i\neq j}l_{ij}$, where $l_{ij}$ is the shortest-path between nodes $i$ and $j$, and the number of local bridges $\mu$ are the links whose endpoints have no common neighbours. The average clustering coefficient is $\langle cc\rangle=\frac{1}{N}\sum_{i}\frac{2M_{i}}{k_{i}(k_{i}-1)}$, where $M_{i}$ is the number of links between common neighbours or $i$, and the betweenness centrality of node $k$ is $B_{k}=\sum_{ij}\frac{\sigma(i,k,j)}{\sigma(i,j)}$, where $\sigma(i,j)$ is the number of shortest-paths between $i$ and $j$, and $\sigma(i,k,j)$ is the number of shortest-paths between $i$ and $j$ passing through node $k$.

We will use three theoretical network models to study the opinion dynamics and investigate the sensitivity of our results on the different structures observed in real-world social networks. We are particularly interested in studying short path length, the combination of short path length and high clustering, and the presence of high-degree nodes (i.e. hubs). The first model is the Erdős-Rényi (ER) model, where nodes are randomly connected with probability $p$. This model exhibits short path lengths $\langle d\rangle$ and low (local) clustering $\langle cc\rangle$. The second model is the Watts-Strogatz (WS) model, which combines short path lengths with high local clustering. It is constructed from a regular lattice by rewiring edges with a probability $q$. According to this probability, the WS model leads to structures intermediate between the regular lattice ($q=0$) with high clustering and long path lengths and the ER model ($q=1$) with low clustering and short path lengths. Finally, the Barabási-Albert (BA) preferential attachment growth model, where nodes are added to the network and connected preferentially to existing nodes with higher degrees, is used to obtain a structure with heterogeneous degree distribution. Table 1 shows the mean values of relevant network measures characterising those networks.

	ER	WS	BA
$\langle k\rangle$	9.926 $\pm$ 0.077	10 $\pm$ 0	9.95 $\pm$ 0
$\langle cc\rangle$	0.010 $\pm$ 0.001	0.089 $\pm$ 0.003	0.040 $\pm$ 0.003
$\langle d\rangle$	3.267 $\pm$ 0.009	3.382 $\pm$ 0.004	2.977 $\pm$ 0.012
$\langle\mu\rangle/M$	0.909 $\pm$ 0.009	0.558 $\pm$ 0.007	0.685 $\pm$ 0.014
$\langle B_{k}\rangle$	0.002 $\pm$ 0.001	0.002 $\pm$ 0.001	0.002 $\pm$ 0.007

Table 1: The mean and the standard deviation ($\pm$) of relevant network measures over ten realisations of the ER, WS, and BA network models with $N=1000$ nodes.

2.2 Opinion Dynamics

We devise a model with $N$ agents (nodes) $i=0,1,\ldots,N-1$. At time step t, each agent has a binary opinion $o_{i}(t)\in\mathcal{O}=\{A,B\}$. We assume that the opinion of agent $i$ can change as a result of two competing mechanisms: (i) pressure from its social contacts $\nu(i)$, i.e. the ratio of neighbours $j$ of $i$ having opinion $o_{j}(t)\neq o_{i}(t)$ (Eq. 1) and (ii) an external fixed source promoting opinion $A$ and affecting all agents equally $\mathcal{E}^{A(B)}\in[0,1]\subset\mathbb{R}$, i.e. an external field representing, for example, mass media, institutions, or nudging.

$$\mathcal{I}_{i}(t)=\frac{\sum_{<ij>}^{o_{j}(t)\neq o_{i}(t)}}{k_{i}}o_{j}(t).$$ (1)

Agents with opinion $o_{i}(t)=A(B)$ have a transition probability towards opinion $B(A)$ at time $t$ given by $W_{i}^{A(B)\rightarrow B(A)}(t)$. The relative importance of the external field concerning social pressure is regulated via the parameter $0\leq\alpha\leq 1$, reflecting the trust agents have in the external opinion. Larger $\alpha$ means that agents give more weight to the external field than to pressure from social contacts. We consider two stochastic versions of the model, differing in how agents are subject to the pressure of the external field.

In the first model (hereafter called asymmetric model), the external field only influences agents with an opinion opposite to the field, i.e. agents agreeing with the external field (defined to be opinion $A$) are only influenced by their social contacts (there is no reinforcement). Equations 2 give the transition rates for agents with opinions $A$ and $B$. In this case, the external field can promote the transition from $B\rightarrow A$ but not vice-versa. This corresponds to a scenario where the external field acts towards one of the opinions, e.g. an employee has to decide on adopting a method or software and is subject to both “outside” influence (employer suggests using software $A$) and “internal” peer pressure (colleagues recommend $A$ or $B$). An employee who uses the default option $A$, as set by the company, does not feel pressure from the employer but only from colleagues using $B$. Meanwhile, when using $B$, the employee feels the influence of colleagues and the employer. In the limiting cases, the model reduces to complex contagion for $\alpha=0$ whereas for $\alpha=1$, transitions from $B\rightarrow A$ only depend on the external field. We could also model the case of outside pressure affecting only those with the same opinion as the external field, but that would lead to confirmation bias and, thus, simple reinforcement.

$$\begin{split}W_{i}^{A\rightarrow B}(t)&=\mathcal{I}_{i}^{B}(t)\\ W_{i}^{B\rightarrow A}(t)&=(1-\alpha)\mathcal{I}_{i}^{A}(t)+\alpha\mathcal{E}^{A}.\end{split}$$ (2)

In the second model (hereafter called symmetric model), all agents may be influenced by the external field (Eqs. 3). This version will not only facilitate that agents with opposing opinions switch to the external opinion (i.e. through nudging opinion $A$ by controlling $\alpha$) but will also hinder agents who agree with the external field from changing their opinion (i.e. reinforcing opinion $A$ to counter-balance social pressure).

$$\begin{split}W_{i}^{A\rightarrow B}(t)&=\text{max}[0,(1-\alpha)\mathcal{I}^{B}(t)-\alpha\mathcal{E}^{A}]\\ W_{i}^{B\rightarrow A}(t)&=(1-\alpha)\mathcal{I}_{i}^{A}(t)+\alpha\mathcal{E}^{A}\end{split}$$ (3)

These two versions of the model are simulated computationally using agent-based modelling. All simulations start with an initial configuration where each agent $i$ has an opinion $A$ or $B$ chosen uniformly. At each time step $t$, an agent $i$ is chosen uniformly. The fraction of disagreeing social contacts (Eq. 1) is calculated and used to determine the opinion flip rate $W_{i}(t)$, based on social pressure, the external field, and $\alpha$. Agent $i$ changes its opinion with probability $W_{i}(t)$ and keeps it with probability $1-W_{i}(t)$. This procedure is repeated for $T$ time steps (Fig. 1), chosen long enough for the system to reach the stationary state ($T_{ER}\approx 40\times 10^{4}$; $T_{WS}\approx 80\times 10^{4}$; $T_{BA}\approx 60\times 10^{4}$). Because time-dependent measures only depend on other time-dependent measures at the same time step, the explicit time dependence will be omitted for the sake of brevity.

Figure 1: An example of the transition rates with external influence $\mathcal{E}^{A}=0.7$, where $\mathcal{I}_{i}^{B}=0.6$ ($\mathcal{I}_{i}^{A}=0.4$) and $\alpha=0.5$.

We use different measures to quantify the consensus formation in the opinion dynamics model across scales [2]. The opinions $o_{i}=A$ and $o_{i}=B$ are modeled by spin variables $\sigma_{i}=+1$ and $\sigma_{i}=-1$. A magnetization-like measure is used to quantify the level of global consensus $m=\frac{1}{N}\sum_{i}\sigma_{i}$, which is the fraction of the total population holding an opinion ($A$ or $B$). The values lie in the interval $[-1,1]$, where the extremes $m=1$ and $m=-1$ resemble the total population holding opinion $A$ and $B$ (full consensus), respectively.

The local consensus is measured using the so-called density of interfaces (Eq. 4), where $N_{p}$ is the total number of nearest neighbour pairs (i.e. ṡocial contacts $(i,j)$) and $N_{a}$ the number of such pairs with neighbours having different opinions (i.e. $o_{i}\neq o_{j}$ for $(i,j)$). The case $n_{a}=1/2$ corresponds to the fully mixed state, and $n_{a}=0$ indicates complete order.

$$n_{a}=\frac{N_{a}}{N_{p}},$$ (4)

The two measures are redundant in the case of full consensus but not in a mixed state. In the latter, $n_{a}$ can be used to gain additional information about how the opinions are locally ordered, i.e. to measure opinion clusters. These measures provide no information about how the population relates to the external opinion. An energy-like measure is used to measure the combined social stress in the population due to local disagreement and disagreement with the opinion of the external field (Eq. 5). This quantity is not conserved and thus will not be referred to as energy but social stress $S_{E}$.

$$S_{E}=-(1-\alpha)\sum_{<ij>}\sigma_{i}\sigma_{j}-\alpha\mathcal{E}\sum_{i}\sigma_{i}.$$ (5)

Social stress increases when two contacts disagree, and agents disagree with the external field opinion. An algorithm based on breadth-first search [42] is used to count distinct opinion clusters in the network. An opinion cluster ($\Omega_{o}$) is thus here defined as a group of agents with the same opinion $o_{i}$ for which there exists a path between any two agents in the group without having to go through an agent of the opposite opinion (Fig. 2).

Figure 2: An example network highlighting the different opinion clusters. Six distinct opinion clusters are identified, two for opinion “red” and four for opinion “green”.

For a given network model and set of parameters, averages are computed using 100 points (10 realizations for each of the ten starting configurations).

3 Results

3.1 Consensus Formation

We start by studying consensus formation and how the external influence affects the stationary population opinion. Without loss of generality, we assign opinion $A$ to the external field. We study the consensus formation by analysing how the strength of the external field $\mathcal{E}$ affects the prevalence of opinion $P_{A}$ for a fixed $\alpha$, which can be interpreted as the amount of trust placed in the external actor relative to one’s peers. At $t=0$, both opinions are distributed uniformly at random over the population such that the prevalence of both opinions is approximately equal, $P_{A}(0)\approx P_{B}(0)\approx 0.5$. Figure 3 shows the temporal evolution of $P_{A}(t)$ for different levels of the external field in both versions of the model.

Figure 3: The prevalence $P_{A}(t)$ of opinion $A$ on the ER, WS, and BA network models (See Methods) for the (a) asymmetric model with $\mathcal{E}^{A}=0$, (b) asymmetric model with $\mathcal{E}^{A}=0.5$, (c) asymmetric model with $\mathcal{E}^{A}=1$ (d) symmetric model with $\mathcal{E}^{A}=0$, (e) symmetric model with $\mathcal{E}^{A}=0.5$, and (f) symmetric model with $\mathcal{E}^{A}=1$. The parameter $\alpha=0.5$ is fixed. The shaded area represents the standard error.

The dynamics lead to a stationary state where the fraction of agents following the external opinion $A$ plateaus below the maximum value, except for the WS network, where the dynamics need longer times to reach stationarity. The plateau values are higher for the symmetric than for the asymmetric version. For $\mathcal{E}^{A}=0.5$ in the asymmetric model and $\mathcal{E}^{A}=0$ in the symmetric model, the prevalence of the opinions remains balanced, except by the increasing prevalence observed for opinion $A$ in the WS network, likely due to reinforcement. The prevalence of opinion $A$ on the WS networks are relatively larger, whereas the prevalence is relatively lower for the BA networks. Generally, increasing $\mathcal{E}^{A}$ leads to increasing $P_{A}$, but there is a difference between both models for low $\mathcal{E}^{A}$. The asymmetric version allows for parameter combinations where the opposing opinion becomes most prevalent on average, whereas this is not possible with the symmetric version. These differences are more significant at the stationary state for various combinations of parameters (Figs. 4-5).

Figure 4: Results of the asymmetric model for various parameter combinations. (a) global consensus $m$ and ER model, (b) local consensus $n_{a}$ and ER model, (c) social stress $S_{E}$ and ER model, (d) global consensus $m$ and WS model, (e) local consensus $n_{a}$ and WS model, (f) social stress $S_{E}$ and WS model, (g) global consensus $m$ and BA model, (h) local consensus $n_{a}$ and BA model, (i) social stress $S_{E}$ and BA model, for different values of the model parameters $\alpha$ and $\mathcal{E}^{A}$.

For the asymmetric model, the stationary state results show two regions of high (average) global consensus and order for all network structures. Agents move more towards global consensus when they are subjected to low or high $\mathcal{E}^{A}$ external field, with much trust in the external source ($\alpha>0.5$), than when they are subject to moderate levels of external pressure ($\mathcal{E}^{A}\approx 0.5$). The asymmetric model allows each opinion to become the majority opinion in the population, depending on $\alpha$-$\mathcal{E}^{A}$ combinations. The effect of strong consensus formation in the case of high trust in the external field but low field intensity is due to this model’s asymmetry in transition rates (Eq. 2). In this region with large $\alpha$, the asymmetric model is subject to two different influences: $W_{i}^{A\rightarrow B}=\mathcal{I}_{i}^{B}$ and $W_{i}^{B\rightarrow A}=\mathcal{E}^{A}$ (eq. 2). Thus, the region of high $\alpha$ and low $\mathcal{E}^{A}$ is driven by local peer pressure alone, while the region with large $\alpha$ and $\mathcal{E}^{A}$ is driven by an interaction of peer and external pressure (Fig. 4). When the agents are only subject to social pressure ($\alpha=0$), the population ends up in a mixed state. This is due to the averaging over multiple stochastic realisations where various levels of consensus are reached for both opinions. Another region with, on average, low global order is visible for values of $\alpha\neq 0$ and medium $\mathcal{E}^{A}$. The higher local disorder and social stress in this area compared to the $\alpha\approx 0$ area could indicate that this average global disorder results from actual disordered states instead of averaging over more or less ordered states in both opinion directions. This area appears as a disordered transition region between going from one majority opinion to the other by increasing or decreasing the intensity of the external field.

Figure 5: Results of various parameter combinations for the symmetric model. (a) global consensus $m$ and ER model, (b) local consensus $n_{a}$ and ER model, (c) social stress $S_{E}$ and ER model, (d) global consensus $m$ and WS model, (e) local consensus $n_{a}$ and WS model, (f) social stress $S_{E}$ and WS model, (g) global consensus $m$ and BA model, (h) local consensus $n_{a}$ and BA model, (i) social stress $S_{E}$ and BA model, for different values of the model parameters $\alpha$ and $\mathcal{E}^{A}$.

In the symmetric model, the resistance to moving away from the opinion of the external field significantly reduces the possible final states, which are qualitatively similar for every network model (Fig. 5). All measures point to high levels of consensus and order in the large $\alpha$ and large $\mathcal{E}^{A}$ upper half of the parameter space. The average fully mixed state at $\alpha=0$ is still visible and appears for $\mathcal{E}^{A}=0$, regardless of $\alpha$. For a fixed $\alpha>0$ and $\mathcal{E}^{A}>0$, there is generally a monotonous increase in consensus visible with increasing external field and trust, respectively. Thus, higher external pressure results in higher levels of consensus as well as a reduction in the variance around the stationary state (Fig. 3 (d - f)). Global consensus is now only possible towards the external field opinion, in contrast with the asymmetric model, where both forces can define consensus on average. For $\alpha>0$ and $\mathcal{E}>0$, the global consensus and local ordering levels are significantly higher than in the asymmetric model, as reflected in the low levels of social stress attained. This restraining effect of the external field is present in all network structures; the difference can be attributed to how much the structure facilitates the spread of opinion. For the WS networks, high levels of consensus are reached for a larger region of the parameter space. In contrast, BA networks show signs of stronger resistance towards opinion spread, requiring a stronger external field. In the WS model, agents reinforce each other due to the redundancy in the social structure (i.e. high clustering), leading to values of $\mathcal{I}_{i}$ larger than $0.5$. Therefore, even a small influence of the external field is amplified due to social reinforcement. On the other hand, the hubs (high-degree nodes) in the BA network act as a type of bottleneck. Hubs can affect a large part of the network, composed mainly by low degree nodes, kee** $\mathcal{I}_{i}$ close to 0.5. If a hub is chosen to update its opinion, the chance that all its neighbours agree on a specific opinion is relatively low because they are not exposed to reinforcement, leading to $\mathcal{I}_{i}\sim 0.5$ for hubs. The poor clustering of the BA network structure hinders the creation of a majority opinion and resists changing towards the external force.

3.2 Opinion Clusters

Figures 6-9 show the evolution of the number $|\Omega|$ and average size $\langle\Omega\rangle$ of the opinion clusters for different levels of the external field $\mathcal{E}^{A}$. In the asymmetric version (Figs. 6-7), the majority of opinions $B$ and $A$ at respectively low and high $\mathcal{E}^{A}$ is reflected in the number and size of their respective clusters. When an opinion reaches a majority, its opinion clusters grow and decrease in number. In contrast, for an opinion that declines in its number of followers, the clusters become more numerous but smaller. As an opinion loses ground and agents acting as bridges in their cluster switch to the opposite opinion, the clusters fracture, and agents of the losing opinion become more isolated as the opinion keeps losing ground. The opposite is visible when an opinion grows in popularity; new followers combine nearby clusters into larger ones.

Figure 6: The temporal evolution of the number of opinion clusters $|\Omega|$ with (a-c) opinion $A$ and (d-f) opinion $B$, for the asymmetric model. Different values of the external field $\mathcal{E}^{A}$ are combined with fixed $\alpha=0.5$ (equal weight to peers and external influence). Shaded areas represent the standard error. These results are averaged over 25 realisations of a given network/parameter combination.

Figure 7: The temporal evolution of the average size of opinion clusters with opinion $A$, $\langle\Omega_{A}\rangle$ (a-c), and opinion B, $\langle\Omega_{B}\rangle$ (d-f), for the asymmetric model. Different values of the external field $\mathcal{E}^{A}$ are combined with fixed $\alpha=0.5$ (equal weight to peers and external influence). Shaded areas represent the standard error. These results are averaged over 25 realisations of a given network/parameter combination.

Figure 8: The temporal evolution of the number of opinion clusters $|\Omega|$ with (a-c) opinion $A$ and (d-f) opinion $B$, for the symmetric model. Different values of the external field $\mathcal{E}^{A}$ are combined with fixed $\alpha=0.5$ (equal weight to peers and external influence). Shaded areas represent the standard error. These results are averaged over 25 realisations of a given network/parameter combination.

Figure 9: The temporal evolution of the average size of opinion clusters $\langle\Omega\rangle$ with (a-c) opinion $A$ and (d-f) opinion $B$, for the symmetric model. Different values of the external field $\mathcal{E}^{A}$ are combined with fixed $\alpha=0.5$ (equal weight to peers and external influence). Shaded areas represent the standard error. These results are averaged over 25 realisations of a given network/parameter combination.

This process is also observed in the symmetric version (Figs. 8-9), where low $|\Omega_{A}|$ for every $\mathcal{E}^{A}$ indicates that indeed the media opinion $A$ does not, on average, become the minority opinion in the symmetric version. Fragmentation of the minority opinion clusters and conglomeration of the majority opinion clusters also occur in the symmetric version. Still, this effect seems more extreme than the asymmetric version. The symmetric version shows higher $|\Omega|$ and lower $\langle\Omega\rangle$ for a minority cluster and visa versa for a majority cluster. The values for $|\Omega|$ and $\langle\Omega\rangle$ approach those at $\mathcal{E}^{A}=1$ for lower values of $\mathcal{E}^{A}$ in the symmetric version, showing that here a strong impact of the external opinion is already present for median values of $\mathcal{E}^{A}$.

For both model versions, $|\Omega|$ and $\langle\Omega\rangle$ display differences for the different network structures. These differences are most pronounced in the regimes that lead to an $|m|>0$, i.e. where, on average, we have a majority and minority opinion. We notice that the BA and ER topologies have opinion communities that evolve similarly concerning their average size but not when it comes to their numbers. When there is fragmentation of a minority opinion, this is more extreme in BA than in any other network topology. The asymmetrical degree distribution and hub structure of the BA network could lead to the many low-degree nodes being isolated in their single or small opinion clusters. For WS, cluster size and number evolution differ from ER and BA. When fragmentation occurs, $|\Omega|$ first peaks, after which it declines to settle at a lower value. This effect can be explained by the fact that much higher values of $m$ are reached for WS. So, the initial rise is fragmentation, as seen for BA and ER. Still, the following decline is because WS ends up with fewer agents that adhere to the minority opinion and, thus, fewer opinion communities. This is also seen by looking at the values of $\langle\Omega_{A}\rangle$ for WS, which are significantly more extreme than other network topologies.

4 Discussion

The asymmetric model results in three regimes that can be mapped in the $\alpha-\mathcal{E}^{A}$ parameter space (Fig. 10(a)). In the first regime B, the population consensus moves towards the opinion opposite to the external field. This regime is for high $\alpha$-low $\mathcal{E}^{A}$ combinations. The second regime A/B, showing average low consensus and order with high social stress, is at intermediate values of $\mathcal{E}^{A}$ or low values of $\alpha$. A third regime A of average high global and local consensus is formed in favour of the external field opinion. It is located in the parameter space’s high $\alpha$-high $\mathcal{E}^{A}$ part. This picture is recurrent across all studied network models with differences in the stationary state values and the sizes of and transition regions between the different regimes (Figs. 3-4). The average stationary consensus in the A/B regime is qualitatively similar to the situation where agents ignore the external influence, i.e. when $\alpha=0$. For a fixed $\alpha>0$, increasing the strength of the media opinion results in breaking up majority opinion clusters (regime B) into smaller ones, increasing $n_{a}$ or the number of disagreeing nearest neighbour pairs. This leads to fracturing of those clusters, increasing social stress at first (regime A/B), decreasing again when enough agents take the media opinion. Large clusters of opinion are formed (regime A) (Figs. 3-7). The existence of regime B for the asymmetric version can be understood from eq. 2 where $W_{i}^{B\rightarrow A}$ is small for high $\alpha$ and small $\mathcal{E}^{A}$. In contrast, $W_{i}^{A\rightarrow B}$ is unaffected, meaning that the asymmetry results in the internal pressure away from the field being felt the strongest in that region.

(a)

(b)

Figure 10: Qualitative diagram of the macroscopic population opinion of the (a) asymmetric and (b) symmetric models at the stationary state.

The symmetric model exhibits two regimes in the $\alpha-\mathcal{E}^{A}$ parameter space (Fig. 10(b)). A disordered regime A/B presents itself at low $\alpha$ and low $\mathcal{E}^{A}$ and for $\alpha=0$ or $\mathcal{E}^{A}=0$. An ordered regime A with high consensus in favour of the external opinion emerges for higher values of $\alpha$ and $\mathcal{E}$. Fixing a positive $\alpha$ and increasing $\mathcal{E}^{A}$ leads to a shift from a state where (on average) consensus towards both opinions is possible (with a significant variation on the ordering – regime A/B), to a state where holding a different opinion to the external field is unlikely, with the emergence of consensus and order (regime A). This pattern is observed for all the network structures, with the main difference being that for WS networks the regime A is larger. The WS case shows the most different patterns in both directions. This results from the larger social clustering and fewer bridges concerning the ER and BA networks. These topological characteristics facilitate the spread of complex contagions, which can amplify the effect of the external opinion. The difference in the structure also leads to the higher prevalence of the majority opinions and the more extreme fragmentation of the minority opinion clusters. The exit probability for the voter model on degree heterogeneous networks (BA model) strongly favours the opinion held by large degree vertices [43], implying that high degree nodes (i.e. hubs) play an important role in simple (i.e. without reinforcement) contagion. In complex contagion, a hub requires several other agents to change its opinion and enforce it to its neighbours. This may explain why the BA structure is more resistant to external influence than the random structure (ER model). Similarly, the difference in degree distribution explains the higher opinion fragmentation in the BA structure compared to the ER structure.

The results suggest that an external field can affect consensus formation. In the absence of external influence ($\alpha=0$), consensus (global order) ($m$) and local order ($n_{a}$) can form, and both opinions are equally likely to reach a majority. The amount depends on the underlying network structure, but on average, no preference toward any specific opinion is observed, resulting in the absence of consensus. Introducing an external field pushing a particular opinion creates a more diverse set of possibilities where different regimes appear depending on $\mathcal{E}^{A}$ and the weight $\alpha$. The external opinion tends to disrupt the equality between both opinions, going from being equally likely to achieve a majority, towards a situation where a favoured opinion is more likely than the other one to gain a majority. Without this external field, the difference between the average numbers and sizes of the resulting opinion clusters of both opinions is small, implying that both opinions can achieve a majority. In general, the symmetric model reliably results in lower levels of social stress because it restricts opinion change away from the media opinion. In contrast, the asymmetric model allows for more fluctuations.

5 Conclusions

We studied the concurrent impact of media influence (as an external field) and social interactions (peer pressure) on the opinion dynamics of a stochastic majority model. We introduced one model version in which the external field only influences those agents who disagree with the field, and another version in which all agents are subject to the external field. The external field led the population to different macroscopic regimes in which different opinions will likely become the majority. In the first model, the population could converge to one of the opinions or to a mixed state where both opinions coexist. In the mixed case, the population behaves qualitatively in the same way in the presence or absence of external influence. In the second model, either the external opinion dominated, or the population had mixed opinions. These macroscopic patterns depended on the intensity of the external influence and on how much weight of the external field compared to the weight of social contacts. If the external influence is not strong enough, the population can move towards the opposite opinion to the one being enforced. Furthermore, the external field is reinforced by social clustering, which helps align the population towards the external opinion, even in the case of low-intensity external influence. On the other hand, high-degree nodes help to prevent the population from moving towards the external field. By controlling whether agents feel the external field equally, we showed that agents being exposed to the external influence, independently of their opinions, led the population to states where one of the opinions (the one not being enforced) was never dominant. On the other hand, if only agents aligned with the external field were affected by this field, the space of possible macroscopic states was equal for both opinions. This can be relevant to business advertising, health, and other governmental intervention campaigns.

The impacts of the concurrent influence of peers and an external actor have been investigated in relatively simple social structures. More complex network topologies could reveal other underlying structures affecting the opinion dynamics in real populations. An example could be using stochastic block models to represent community structure or network models combining homophily with preferential attachment [44]. Additionally, agents could have heterogeneous weights representing differences in their susceptibility to external actors. A more in-depth study of the connection between complex opinion spread and topological characteristics should include using measures adapted to complex contagions [45]. The WS network is the sole topology where full consensus was reached. This result, of complete order, is understood as a finite-size effect for the classical voter model on WS networks [46]. Despite limitations, our study shows that introducing simultaneous external and peer influence in opinion dynamics can lead to a more diverse behaviour than a monotonous increase in consensus for increasing external influence. The emergence of different macroscopic population opinions can help our understanding of how groups with varying trust in the media and subject to certain levels of media presence, react.