Interplay between media and social inﬂuencein the collective behavior of opinion dynamics
Francesca Colaiori
1,2
and Claudio Castellano
3,2
1
Istituto dei Sistemi Complessi (ISCCNR), UOS Sapienza, c/o Dipartimento di Fisica,Sapienza Universit`a di Roma, Piazzale Aldo Moro 5, 00185 Rome, Italy
2
Dipartimento di Fisica, Sapienza Universit`a di Roma, Piazzale Aldo Moro 5, 00185 Rome, Italy
3
Istituto dei Sistemi Complessi (ISCCNR), Via dei Taurini 19, 00185 Roma, Italy
(Dated: October 14, 2015)Messages conveyed by media act as a major drive in shaping attitudes and inducing opinion shift.On the other hand, individuals are strongly aﬀected by peers’ pressure while forming their own judgment. We solve a general model of opinion dynamics where individuals either hold one out of twoalternative opinions or are undecided, and interact pairwise while exposed to an external inﬂuence.As media pressure increases, the system moves from pluralism to global consensus; four distinctclasses of collective behavior emerge, crucially depending on the outcome of direct interactions amongindividuals holding opposite opinions. Observed nontrivial behaviors include hysteretic phenomenaand resilience of minority opinions. Notably, consensus could be unachievable even when media andmicroscopic interactions are biased in favour of the same opinion: the unfavoured opinion mighteven gain the support of the majority.
PACS numbers: 87.23.Ge, 89.65.Ef, 02.50.Le, 05.45.a
I. INTRODUCTION
Public opinion formation is a complex mechanism:people constantly process a huge amount of informationto make their own judgment. Individuals form, reconsider and possibly change their opinions under a steadyﬂow of news and a constant interchange with others. Social interactions have a strong eﬀect on opinion formation [1]: the pressure to conform to the opinion of othersmay in some cases even overcome empirical evidence [2]or undermine the wisdom of the crowd [3]. On the otherhand media also play a central role: they can inﬂuencepublic knowledge, attitudes and behavior by choosing theslant of a particular news story, or just by selecting whatto report. An intense activity has recently investigatedmacroscopic eﬀects of these fundamental mechanisms exploiting the connection between opinion dynamics andsimple non–equilibrium statistical physics models [4, 5].Media inﬂuence has been considered in the literature,in particular within the framework of cultural dynamics [6–8] and continuous opinion dynamics with heterogeneous conﬁdence bounds [9–12]. The fundamental caseof binary opinions has received less attention for whatconcerns the role of media [13–15], although the similarcase of proselytism by committed agents (zealots) has attracted considerable interest [16, 17]. The case of binaryopinions is clearly relevant to yes/no questions, but notlimited to that: when people are prompted with important questions that admit many possible answers theirattitudes tend to be polarized, most people sharing oneout of two opposite opinions [18].In this paper we address the problem of how peopleform their opinions based on the message they receive,both from traditional media and their social network,by studying a general class of opinion dynamics models. At any given time, each individual in the population is either supporting one of two alternative opinions or “undecided”. The possibility to be in a thirdstate may have crucial eﬀects in the case of a binarychoice [19–22]. Individuals are exposed to some externalsource of information biased towards one opinion (mainstream) and exchange information upon pairwise interaction: both factors may cause them to change their state.The undecided state accounts for individuals being uninterested, uninformed, or generally confused on the givenissue. Carrying no opinion, they are assumed to havea passive role in the interactions. We account for theeﬀect of media in the simplest possible way by assuming that people have some general tendency to conformto the media recommendation. On the other hand, weconsider totally general rules for the interactions amongpairs of agents. Our general model encompasses therefore a large class of speciﬁc models, each one identiﬁed bya given set of parameters deﬁning the interaction rules.We study in meanﬁeld this general model, determiningthe stationary solutions and their stability as a functionof the external bias and of the parameters specifying pairinteractions. We uncover the emergence of four distinctclasses of collective behavior, characterized by diﬀerentresponses to the media exposure. Only two linear combinations of the parameters deﬁning the general dynamicsare relevant in determining which class a speciﬁc modelbelongs to. The results of numerical simulations, performed in systems with interaction patterns described bycomplex networks, support the general validity of the MFpicture.
II. GENERAL MODEL
We now deﬁne in detail the general model. Each agentcan be in one of three states: holding opinion
A
, holding
2
A,A B,B U,U A,U B,U A,BA
−
A
1 0 0 0 0 0
B
−
B
0 1 0 0 0 0
U
−
U
0 0 1 0 0 0
A
−
U ϕ
2
0 0 1
−
ϕ
2
0 0
B
−
U
0
γ
2
0 0 1
−
γ
2
0
A
−
B α
1
α
2
α
3
α
4
α
5
α
6
TABLE I. Each row in the table corresponds to an interaction,and each column to a possible outcome. Elements in thetable indicate the probabilities of each possible outcome forthe given interaction. In the last row
6
i
=1
α
i
= 1.
an opposing opinion
B
, or being undecided (
U
). Agentstend to conform to the media recommendation by adopting opinion
A
at a constant rate
r
, independently of theircurrent state, and interact pairwise at rate
f
. In the following we set
f
= 1 to ﬁx the time scale. Individualsin the same state are unaltered by their mutual interaction. In
A
−
U
(
B
−
U
) interactions undecided agentsmay adopt, with given probability, their partner’s opinion, that they cannot alter: an agent holding opinion
A
(
B
) has a probability
ϕ
2
(
γ
2
) to convince the
U
agent.We assume in general
ϕ
2
=
γ
2
, allowing for
A
and
B
to have unequal eﬃcacy in persuading others. Interactions among agents holding opposite opinions (
A
−
B
)may have any outcome: each of the two agents may either keep her opinion, change it to match her partner’sopinion, or get confused and turn to the undecided state,in any combination. We indicate with
{
α
i
}
6
i
=1
the probabilities of the six possible outcomes. Seven independentparameters ﬁx the probabilities of each possible outcomefor any interacting pair, as summarized in Table I. Whilewe consider for simplicity symmetric roles for the two interacting partners, our mean–ﬁeld (MF) analysis holdsmore generally, also including models that assign distinct roles (e.g. speaker/listener) to the two partners.This more general case is discussed in Appendix A. Notethat asymmetric models are always equivalent, in meanﬁeld, to their symmetrized version, the outcome of a symmetric interaction being the average result of asymmetricinteractions with exchanged roles.
III. MEAN FIELD EQUATIONS
The evolution of the system is described in MF by thedynamical equations for the density of agents in eachstate:
˙
n
A
=
r
(1
−
n
A
) + 2
ϕ
1
n
A
n
B
+ 2
ϕ
2
n
A
(1
−
n
A
−
n
B
)˙
n
B
=
−
rn
B
+ 2
γ
1
n
A
n
B
+ 2
γ
2
n
B
(1
−
n
A
−
n
B
) (1)where
n
A
(
n
B
) denote the density of agents in the
A
(
B
) state, and
ϕ
1
=
α
1
−
α
2
−
α
3
−
α
5
,
γ
1
=
−
α
1
+
α
2
−
α
3
−
α
4
. The density
n
U
of undecided agents isobtained by normalization:
n
U
= 1
−
n
A
−
n
B
. Thedensities must belong to the
physical region
of the plane
(a)
ϕ
1
γ
1
01 1011
FCM VCM ZCM TC
(b)
01 1
γ
∗

ϕ
1
γ
1
r
*
r
c
(b)
01 1
γ
∗

ϕ
1
γ
1
r
*
r
c
FIG. 1. (color online)(a) Phase diagram in the (
ϕ
1
,
γ
1
) plane.The curve
γ
∗
1
separating regions
VCM
and
ZCM
depends on
ϕ
2
and
γ
2
(in the ﬁgure
ϕ
2
= 0
.
1 and
γ
2
= 0
.
5). (b) Plot of
r
c
(saddle–node bifurcation curve) and
r
∗
(transcritical bifurcation line) as functions of
γ
1
, and for
ϕ
1
=
−
0
.
8.
(
n
A
,n
B
), deﬁned by the three constraints
n
A
≥
0,
n
B
≥
0,
n
A
+
n
B
≤
1. The coeﬃcients
α
i
appear in Eq. (1)only in two linear combinations
ϕ
1
and
γ
1
that representthe total average variation in
A
(
B
) states due to an
A
−
B
interaction. This reduces the number of eﬀectiveparameters from seven to four plus the external bias
r
,acting as a control parameter. The parameters
ϕ
1
and
γ
1
are bounded by
−
1
≤
γ
1
≤
1,
−
1
≤
ϕ
1
≤
1,
γ
1
+
ϕ
1
≤
0(the sum
γ
1
+
ϕ
1
is minus the average net productionof undecided in an
A
−
B
interaction, that has to benon negative). Stationary solutions of Eq. (1) are theintersections of the two conic sections:
n
A
[(
ϕ
2
−
ϕ
1
)
n
B
+
ϕ
2
n
A
−
ϕ
2
+
r/
2] =
r/
2
C
1
n
B
[(
γ
1
−
γ
2
)
n
A
−
γ
2
n
B
+
γ
2
−
r/
2] = 0
C
2
(2)The curve
C
1
is an hyperbola. One of its asymptotesis the axis
n
A
= 0, so that only the upper branch
C
+1
(
n
A
>
0) is physically relevant. The curve
C
2
factorizesinto the product of two lines,
R
1
, and
R
2
. The line
R
1
(
n
B
= 0) always intercepts
C
+1
in
P
1
≡
[
n
B
= 0
,n
A
= 1],corresponding to the absorbing state of total consensuson opinion
A
. Depending on the parameters,
R
2
may intersect
C
+1
in one, two (possibly coincident) points, eitherinside or outside the physical region, or none. Varyingthe control parameter
r
, diﬀerent ﬁxed points arise andmove entering and exiting the physical region. Their ﬂowand stability determine the collective response to the external bias. By studying them we ﬁnd the phasediagramrepresented in Fig. 1(a), which constitutes the main result of our paper. As a function of the external bias
r
there are four distinct classes of collective behavior, associated with diﬀerent regions of the parameter space. Notably, the emergent behavior is essentially ruled by onlytwo of the parameters,
ϕ
1
and
γ
1
. In the next sectionwe discuss the qualitative features of each class, referringto Appendix B for analytical details on the derivation of the phase diagram. Fig. 2 represents the shapes of thecurves in Eq. (2), allowing to understand the existence
3
P
1
Q
1
P
2
Q
2
n
A
n
B
r=0ClassClassClassClass I II III IV P
1
Q
2
Q
1
n
A
n
B
0<r<r
c
P
1
Q
1
=Q
2
n
A
n
B
r=r
c
P
1
n
A
n
B
r>r
c
P
1
Q
1
P
2
n
A
n
B
r=0P
1
Q
1
n
A
n
B
0<r<r
*
P
1
=Q
2
Q
1
n
A
n
B
r=r
*
P
1
Q
2
Q
1
n
A
n
B
r
*
<r<r
c
P
1
Q
1
=Q
2
n
A
n
B
r=r
c
P
1
n
A
n
B
r>r
c
P
1
Q
1
P
2
n
A
n
B
r=0P
1
Q
1
n
A
n
B
0<r<r
*
P
1
=Q
1
n
A
n
B
r=r
*
P
1
n
A
n
B
r>r
*
P
1
Q
1
P
2
n
A
n
B
r=0P
1
n
A
n
B
r>0
FIG. 2. Plots of isoclines and ﬁxed points for each class of models and each range of
r
. Red solid lines are
R
1
and
R
2
, theblu dashed curve is the hyperbola
C
1
. Red stars denote stable ﬁxed points, black circles denote saddle points, blue trianglesdenote repulsive ﬁxed points. The shaded triangle represents the physical region
n
A
≥
0,
n
B
≥
0,
n
A
+
n
B
≤
1. Plots in theﬁrst row correspond to
ϕ
1
=
−
0
.
6,
ϕ
2
= 0
.
1,
γ
1
=
−
0
.
3,
γ
2
= 0
.
5, and
r
= 0,
r
=
r
c
/
2,
r
=
r
c
,
r
=
r
c
+ 0
.
05. Plots in thesecond row correspond to
ϕ
1
=
−
0
.
6,
ϕ
2
= 0
.
05,
γ
1
= 0
.
1,
γ
2
= 0
.
3, and
r
= 0,
r
=
r
∗
/
2 =
γ
1
,
r
=
r
∗
=
γ
1
,
r
= 0
.
2
r
∗
+ 0
.
8
r
c
,
r
=
r
c
,
r
=
r
c
+ 0
.
05. Plots in the third row correspond to
ϕ
1
=
−
0
.
5,
ϕ
2
= 0
.
7,
γ
1
= 0
.
3,
γ
2
= 0
.
5, and
r
= 0,
r
=
r
∗
/
2 =
γ
1
,
r
=
r
∗
=
γ
1
,
r
=
r
∗
+ 0
.
1. Plots in the fourth row correspond to
ϕ
1
= 0
.
1,
ϕ
2
= 0
.
7,
γ
1
=
−
0
.
2,
γ
2
= 0
.
5, and
r
= 0,
r
= 0
.
2.
(a)P
1
n
A
n
B
(b)P
1
n
A
n
B
P
1
n
A
n
B
(c)
FIG. 3. (color online) Phase portrait relative to Eq. (1): ﬂows in the vector ﬁeld indicate the time evolution of the system.Fixed points are the intersections between either of the two lines
R
1
,
R
2
(red solid lines), and the curve
C
1
(blue dashed line).Black circles are saddle points, red stars are attractive ﬁxed points. (a)
FCM
model in the region
r < r
c
(
ϕ
1
=
−
0
.
6,
ϕ
2
= 0
.
1,
γ
1
=
−
0
.
3,
γ
2
= 0
.
5,
r
= 0
.
1); (b)
VCM
model in the range
r < r
∗
(
ϕ
1
=
−
0
.
6
,ϕ
2
= 0
.
4,
γ
1
= 0
.
1,
γ
2
= 0
.
3,
r
= 0
.
13); (c)
TC
model (
ϕ
1
= 0
.
1,
ϕ
2
= 0
.
7,
γ
1
=
−
0
.
2,
γ
2
= 0
.
5,
r
= 0
.
2). The ﬂow is qualitatively similar to (a) for
VCM
models with
r
∗
< r < r
c
, to (b) for
ZCM
models with
r < r
c
, to (c) for any class of models above
r
c
. See Appendix B for further details.
4
0 0.2 0.4 0.6 0.8 1
Class I: FCM
0
r
c
0.2
r n
A
n
B
n
U
(a)Class II: VCM
0
r r
c
0.5
r
*
n
A
n
B
n
U
(b)Class III: ZCM
0
r
1
r
*
n
A
n
B
n
U
(c)Class IV: TC
01
r n
A
n
B
n
U
(d)
FIG. 4. (color online) Theoretical results for the densities of agents for realizations of each of the four classes of models: (a)
Class I
model (FCM) (
ϕ
1
=
−
0
.
5,
ϕ
2
= 0
.
1,
γ
1
=
−
0
.
4, and
γ
2
= 0
.
5), (b)
Class II
model (VCM) (
ϕ
1
=
−
0
.
5,
ϕ
2
= 0
.
1,
γ
1
= 0
.
1, and
γ
2
= 0
.
5), (c)
Class III
model (ZCM) (
ϕ
1
=
−
0
.
5,
ϕ
2
= 0
.
1,
γ
1
= 0
.
3, and
γ
2
= 0
.
5), (d)
Class IV
model (TC)(
ϕ
1
= 0
.
3,
ϕ
2
= 0
.
1,
γ
1
=
−
0
.
5, and
γ
2
= 0
.
5). Solid (dashed) lines represent stable (unstable) lines.
and positions of stationary solutions. Fig. 3 provides information on ﬂows and the stability of solutions. Fig. 4depicts the resulting behavior of the densities of agentsin the diﬀerent states as a function of
r
.
IV. CLASSES OF COLLECTIVE BEHAVIORA. Class I: Finite Critical Mass (FCM) Models
When
ϕ
1
<
0,
γ
1
≤
0, i.e. in models where
A
−
B
interactions produce on average an increase in undecidedindividuals with no net gain in
A
nor in
B
states, thesystem undergoes a ﬁrst order transition at a ﬁnite value
r
=
r
c
of the external bias (see Fig. 2, ﬁrst row). Forlarge enough
r P
1
is the only ﬁxed point and the system ﬂows into the absorbing state of total consensus onopinion
A
for any initial condition. At
r
=
r
c
the systemundergoes a saddle–node bifurcation [23]: two additionalcoinciding ﬁxed points appear. As
r
is decreased below
r
c
they split (one, with larger
n
B
, stable, the otherunstable, see Fig. 3(a)). In the nontrivial stable ﬁxedpoint the two opinions
A
and
B
coexist in the population, together with a fraction of undecided (we call thisstate “pluralism”), see Fig. 4(a). The initial conditionsdetermine whether the pluralistic state (
n
B
>
0) or theconsensus state (
n
A
= 1) is asymptotically reached. Thevalue of
n
B
at the unstable ﬁxed point stays ﬁnite in thelimit
r
→
0 (see Fig. 2, ﬁrst row), implying that a ﬁnite“critical mass” [24] of dissenters is always needed to reachthe pluralistic state, no matter how small is the externalbias.
B. Class II: Vanishing Critical Mass (VCM)Models
This class is identiﬁed by
ϕ
1
<
0, 0
< γ
1
< γ
∗
1
, andcorresponds to models where
A
−
B
interactions causeon average a small increase of
B
states at the expenseof
A
states. As in
Class I
, lowering
r
below
r
c
thesystem undergoes a ﬁrst order transition separating aregime (
r > r
c
), where consensus on opinion
A
is theonly stable state from a regime (
r < r
c
) where a stableand an unstable additional ﬁxed points appear through asaddle–node bifurcation (Fig. 2, second row). However,in this case, further reducing
r
, the unstable ﬁxed pointcollides with the point
P
1
(consensus on
A
) at a ﬁnitevalue
r
∗
(0
< r
∗
< r
c
), and then exits the physical region. This is a transcritical bifurcation [23]: when thetwo ﬁxed points cross each other, they exchange stability(Fig. 3(b));
P
1
becomes unstable, so that below
r
∗
thesystem, unless started with
n
B
≡
0, always ﬂows to thepluralistic state (Fig. 4(b)). The initial presence of even afew dissenters suﬃces for opinion
B
to survive. The curve
γ
∗
1
(
ϕ
1
) =
ϕ
1
γ
2
/
(
ϕ
1
−
ϕ
2
−
γ
2
) separating
Class II
and
III
depends on the parameters
ϕ
2
and
γ
2
(see Appendix B).The transcritical bifurcation line is
r
=
r
∗
= 2
γ
1
(seeAppendix B), and always lies below the saddle–node bifurcation line (see Fig. 1(b)).
C. Class III: Zero Critical Mass (ZCM) Models
When
ϕ
1
<
0,
γ
1
≥
γ
∗
1
, corresponding to models where
A
−
B
interactions give an increase of undecided and alarge increase in
B
at the expense of
A
, the system undergoes at
r
=
r
∗
= 2
γ
1
a continuous transition (transcritical bifurcation) between total consensus on opinion
A
(
r > r
∗
) and pluralism (
r < r
∗
), see Fig. 2, third row,and Fig. 4(c). Initial conditions do not play any role.
D. Class IV: Total Consensus (TC) Models
The region
ϕ
1
≥
0 corresponds to models where
A
−
B
interactions result in a net increase of individuals holdingopinion
A
. The behavior is trivial (see Fig. 3, fourth row):irrespectively of the value of all other parameters, for anyinitial condition, and no matter how small the external
5forcing is, the system always converges to the consensusstate (
P
1
) (see Fig. 3(c) and 4(d)).
V. BEYOND MEAN–FIELD
Before discussing the analytical results obtained in theprevious section within MF and examining some of theirconsequences we test their validity beyond MF by numerical simulations on synthetic and real–word networks.The simulations are performed on four microscopic dynamical models, each belonging to one of the universalityclasses derived above.The single event of the dynamics occurs as follows [25].We select randomly a node
i
(speaker) and, with probability
r/
(1 +
r
), we set his state to
A
, as eﬀect of theexternal bias. Instead, with complementary probability1
/
(1 +
r
) an interaction process takes place: we select alistener
j
among the neighbors of
i
, and modify the stateof the pair (
i
,
j
) according to Table II in Appendix Bwith probabilities
ψ
2
=
ω
2
=
ϕ
2
,
δ
2
=
ǫ
2
=
γ
2
and each
λ
i
=
µ
i
=
α
i
for any
i
. Each single event occurs duringa temporal interval 1
/N
(where
N
is the total number of nodes in the network), so that
N
updates are attemptedin a time unit. Starting from the initial conﬁguration, foreach value of
r
we let the system evolve during 5000 timesteps to reach the stationary conﬁguration and determinethe densities
n
A
,
n
B
and
n
U
by performing averages over5000 additional time steps. In order to characterize thepossible presence of discontinuous phase transitions andthe associated hystereric eﬀects, for each set of data weconsider two diﬀerent initial conditions: either all nodesin state
B
(
n
B
= 1), or all nodes in state A except fora very small fraction of nodes in state
B
(
n
A
= 0
.
99,
n
B
= 0
.
01). We always keep the values
ϕ
2
= 0
.
1 and
γ
2
= 0
.
5 ﬁxed and vary
γ
1
and
ϕ
1
in order to encompass all four quantitatively distinct behaviors found inthe analytic approach.
A. Simulations on synthetic networks
In Fig. 5 (upper panel) we plot the stationary valueof the densities as a function of
r
when the interaction pattern is given by three synthetic networks of size
N
= 20000: a Random Regular Graph (RRG) whereeach node has 10 neighbors; an Erd¨osR´enyi(ER) graphof average degree 10; a network built using the Uncorrelated Conﬁguration Model (UCM) with minimum degree3 and degree distribution decaying as
k
−
2
.
5
. For all thesecases, predictions are very well matched by numericalsimulations.
B. Simulations on real networks
A tougher test of the MF results is provided by simulations performed on real–world networks, incorporating additional topological features such as clustering andcorrelations. In Fig. 5 (lower panel) we report resultsfor: a network of size
N
= 81860 representing movieactor collaborations obtained from the Internet MovieDatabase (MOVIES) (average degree
k
= 89
.
53, ﬂuctuations
k
2
/
k
= 594
.
91); a network of size
N
= 24608representing connections of Internet Autonomous Systems in 2006 (AS2006) (average degree
k
= 4
.
05, ﬂuctuations
k
2
/
k
= 259
.
94) ; the largest connected component (size
N
= 33696) of the Enron email exchangenetwork (ENRON) (average degree
k
= 10
.
02, ﬂuctuations
k
2
/
k
= 140
.
07) [26].Also in these cases numerical simulations agree wellwith the outcome of the MF approach: the behavior ineach of the classes qualitatively reproduces the analytical predictions, with only (expected) variations in theposition of transition points. The only variation in thisrespect concerns class I. In this case it turns out that forvery small values of the rate
r
the state with overwhelming majority of
B
is dynamically reached even startingfrom
n
B
as low as 0
.
01, at odds with the MF prediction.
VI. DISCUSSION AND CONCLUSIONS
We ﬁnally discuss some interesting and nontrivial consequences that follow from the classiﬁcation scheme derived within the MF approach.Although the parameters regulating the interactionswith undecided agents have a marginal role in determining the collective behavior, the presence itself of the
U
state is crucial in several respects. In the absence of the third, undecided state, the system would either converge to total consensus on opinion
A
(
Class IV
)– whenasymmetric interactions favor
A
, or exhibit a continuoustransition between consensus and pluralism (
Class III
) –when asymmetric interactions favor
B
[22].Several models in the literature that allow for a thirdstate require
U
to be a necessary intermediate step whenchanging opinion [21, 27]. Our results imply that suchsystems undergo a discontinuous transition by varyingthe media exposure:
U
being a necessary intermediatestep requires
α
1
=
α
2
= 0, giving
ϕ
1
≤
0,
γ
1
≤
0; therefore these models always fall in
Class I
. The condition fora model to be in
Class I
is however more general, onlyexcluding
average
gain of
A
or
B
in
A
−
B
interactions.Within our framework consensus is always achieved,whatever the interactions, for strong enough media exposure. Then a natural question arises: is consensus stableupon removal of the media pressure? The answer to thisquestion is diﬀerent for diﬀerent classes. In FCM class,once the consensus is reached, it is kept also when themedia exposure is removed. In ZCM class dissenters nucleate as soon as the media exposure is lowered below thethreshold needed to reach consensus (
r < r
c
). VCM classshows interesting hysteretic behavior: once consensus isreached with a suﬃciently high media exposure (
r > r
c
),it is kept when lowering the media pressure that however