Social Influence and Unfollowing Accelerate The Emergence of Echo Chambers

Journal of Computational Social Science
https://doi.org/10.1007/s42001-020-00084-7
RESEARCH ARTICLE
Social influence and unfollowing accelerate the emergence

of echo chambers
Kazutoshi Sasahara1,2,3 · Wen Chen3,4 · Hao Peng4,5 ·

Giovanni Luca Ciampaglia6,7 · Alessandro Flammini3,4,6 · Filippo Menczer3,4,6
Received: 8 December 2019 / Accepted: 20 August 2020

© The Author(s) 2020
Abstract
While social media make it easy to connect with and access information from any‑
one, they also facilitate basic influence and unfriending mechanisms that may lead
to segregated and polarized clusters known as “echo chambers.” Here we study the
conditions in which such echo chambers emerge by introducing a simple model of
information sharing in online social networks with the two ingredients of influence
and unfriending. Users can change both their opinions and social connections based
on the information to which they are exposed through sharing. The model dynam‑
ics show that even with minimal amounts of influence and unfriending, the social
network rapidly devolves into segregated, homogeneous communities. These predic‑
tions are consistent with empirical data from Twitter. Although our findings suggest
that echo chambers are somewhat inevitable given the mechanisms at play in online
social media, they also provide insights into possible mitigation strategies.
Keywords Echo Chamber · Opinion dynamics · Social media · Social network
* Kazutoshi Sasahara
sasahara@nagoya‑u.jp
1
Graduate School of Informatics, Nagoya University, Nagoya, Japan
2
JST PRESTO, Saitama, Japan
3
Center for Complex Networks and Systems Research, Indiana University Bloomington,
Bloomington, USA
4
Luddy School of Informatics, Computing, and Engineering, Indiana University, Bloomington,
USA
5
School of Information, University of Michigan, Ann Arbor, USA
6
Network Science Institute, Indiana University, Bloomington, USA
7
Department of Computer Science and Engineering, University of South Florida, Tampa, USA
13
Vol.:(0123456789)
Introduction
The rise of social media has led to unprecedented changes in the scale and speed
at which people share information. Social media feeds are key tools for accessing
high volumes of news, opinions, and public information. However, just by foster‑
ing such a proliferation of information to which people are exposed, social media
may interfere with cognitive selection biases, amplifying undesirable phenomena
such as extremism and the spread of misinformation [1]. Further, they may intro‑
duce new biases in the way people consume information and form beliefs, which
are not well understood yet.
Theories about group decision-making and problem-solving suggest that
aggregating knowledge, insights, or expertise from a diverse group of people is
an effective strategy to solve complex problems, a notion called collective intel‑
ligence [2, 3]. While the Web and social media have often been hailed as striking
examples of this principle in action [4, 5], some of the assumptions upon which
these systems are predicated may harm the very diversity that makes them pre‑
cious sources of collective intelligence [6]. Social media mechanisms, in particu‑
lar, tend to use popularity signals as proxies of quality or personal preference,
despite evidence that rewarding engagement may come at the expense of view‑
point diversity and quality [7]. Worse, the community structure of information
flow networks can distort decisions and increase vulnerability to malicious actors
such as social bots [8].
There is increasing empirical evidence of these phenomena: polarization is
observed in social media conversations [9–11] and low diversity is found in online
news consumption [12–15]. These observations have in common two features: net‑
work segregation (the splitting of the social network in two or more disconnected or
poorly connected groups) and opinion polarization (the high homogeneity of opin‑
ions within such groups). Figure 1 shows what an information diffusion network
with those two features looks like. Human factors such as homophily [16–18] (the
tendency to form ties with similar people) and social influence [19] (the tendency
of becoming more similar to somebody as a result of social interaction) are often
thought to drive the emergence of polarization and segregation.
Some of the consequences of socio-cognitive biases have been explored in
social dynamics models [21] and in the social psychology literature [1]. Yet, the
interplay between these and additional biases introduced by social media mecha‑
nisms is not clear. The algorithms at the heart of social media make a number
of assumptions to deliver their recommendations. For example, news feed algo‑
rithms favor stories with which users are more likely to engage in the future,
based on past engagement [22]. Friendship recommendation engines suggest new
ties based on common interests, beliefs, and friends, often resulting in the closure
of open triads [23–26]. And finally, social media empower users to dissolve ties
that, although not by design, often tend to be the ones connecting them with those
with whom they disagree [27]. Indeed, estimates suggest that tie decay in social
networks is a relatively common occurrence [28, 29]. The Appendix provides fur‑
ther empirical evidence of the dissolution of social connections on Twitter.
13
Fig. 1 Example of a polarized and segregated network on Twitter. The network visualizes retweets of
political hashtags from the 2010 US midterm elections. The nodes represent Twitter users and there is a
directed edge from node i to node j if user j retweeted user i. Colors represent political preference: red for
conservatives and blue for progressives [20]. For illustration purposes, only the nodes in the k = 3 core
are visualized. See Methods for more details
By curating the information to which users are exposed and by facilitating their
management of social ties, social media platforms may enhance homophily and con‑
firmation bias, the tendency to pay attention only to information that aligns with
prior beliefs [30]. This would have the net effect of leading social media users to
connect preferentially with like-minded individuals, which would then result in
selective exposure to only that information which conforms to their pre-existing
beliefs, as opposed to more diverse points of view [31]. Ultimately, these dynamics
would drive users of social media toward polarization and segregation [32], more so
than users of legacy media like TV, radio, or newspapers, where social sharing and
link management mechanisms are not at play [33].
The risk that online communication networks could splinter into different ‘tribes’
was recognized at the dawn of cyberspace, and given the name of cyber-balkani‑
zation [34, 35] as an analogy to the well-known phenomena of cultural, racial, and
ethnic segregation [36]. With the rise of modern personalization technologies, there
has been renewed interest in understanding whether algorithmic bias is accelerating
the fragmentation of society. The terms filter bubble [37] and echo chamber [38]
have been coined to refer to two different algorithmic pathways to opinion fragmen‑
tation, both related to the way algorithms filter and rank information. The first refers
to search engines [39], the second to social media feeds [40, 41].
The literature about the effects of technology on echo chambers presents a com‑
plicated picture, with inconsistent and somewhat contradictory evidence [42]. This
is not entirely surprising when one considers the enormous heterogeneity in news
sources, social media characteristics and usage, and human information-consumption
13
behaviors. Lab experiments show that people consistently exhibit a preference for
congenial information over uncongenial information, especially in the domain of
politics [43]. Surveys find mixed evidence of partisan selectivity [44, 45]. On social
media, opinion-reinforcing information promotes news story exposure [46]. Analy‑
sis of online consumption of political news by Facebook users shows that exposure
to cross-cutting ideological information is reduced by comparable amounts when
one considers platform algorithms and individual click choices [12]. Behavioral data
on online consumption of political news suggest that selective exposure is mostly
concentrated among a minority of the population—those who drive most of the traf‑
fic to partisan sources [47, 48]. Even if polarized media consumption may not be the
norm for most people, it is common among the most politically active, knowledge‑
able, engaged, and influential individuals [49]. This can favor the spread of misin‑
formation from partisan media and increase animosity within the population. It is
therefore important to understand how specific social media mechanisms may facili‑
tate the formation of ideological echo chambers.
Here, we study the emergence of joint opinion polarization and network segre‑
gation in online social media, specifically focusing on the interplay between the
mechanisms of influence and unfriending. One novel contribution of our approach
is to model how these mechanisms are mediated by the basic activity of information
sharing in social media. Furthermore, we explore how biases in recommendation
algorithms may exacerbate the dynamics of echo chambers. Although our model is
idealized, it captures some key features of social media sharing—limited attention,
social influence, and social tie curation. In particular we investigate the role of selec‑
tive unfollowing, which has not been studied in the literature. Through a series of
simulations that compare different scenarios by exploring two key parameters, we
find conditions under which opinion polarization and network segregation coevolve.
This provides a generative mechanism to interpret the formation of echo chambers
in social media. We also check the consistency of our model against empirical data
from Twitter. This allows us to test the micro-level assumptions underpinning our
model as well as its macro-level predictions [50].
Model
Let us introduce a model of opinion dynamics in an evolving social network. We

incorporate various ingredients from models in the literature: information diffusion
via social sharing [51], opinion influence based on bounded confidence [52], and
selective rewiring of social ties [53].
The model begins with a random directed graph with N nodes and E directed
edges, representing an online social network over which messages spread. Nodes
represent social media users and edges represent follower ties, as in Twitter and Ins‑
tagram. In the initial step, each user’s opinion (o) is randomly assigned a value in
the interval [−1, +1] . Each user has a screen that shows the most recent l messages
posted (or reposted) by friends being followed. A message conveys the identity and
opinion value m of the user who originated the post, together with the information
about who reposted it. Users can see this information. A message’s opinion m is
13
concordant with an opinion o if they are within a bounded confidence distance 𝜖

(that is, when |o − m| < 𝜖 ). It is discordant otherwise. In addition, each user can
unfollow a friend by rewiring the connection, thus following someone else in their
place. These mechanisms allow us to capture two common ingredients of social
media platforms: the possibility to share information with one’s friends, and the pos‑
sibility to form a new connections.
At every time step t, user i is selected at random, and sees messages on the screen
that are posted or reposted by friends. The opinion of user i then changes based on
the concordant messages on the screen:
∑l
I (o (t), mj )(mj − oi (t))
j=1 𝜖 i
oi (t + 1) = oi (t) + 𝜇 ∑l , (1)
I (o (t), mj )
j=1 𝜖 i
where 𝜇 is an influence strength parameter, the sum runs over the messages in i’s
screen, and I𝜖 is an indicator function for concordant opinions based on the confi‑
dence bound 𝜖:
{
1 if |o − m| < 𝜖
I𝜖 (o, m) =
0 otherwise. (2)
Equation 1 provides a simple mechanism for modeling social influence based on an

individual’s tendency to favor information that is similar to their pre-existing opin‑
ions ( oi ± 𝜖 ), such as the confirmation bias and selective exposure mentioned ear‑
lier. This is referred to as bounded social influence and its breadth and strength are
controlled by the parameters 𝜖 and 𝜇 . Larger 𝜖 indicates broader-minded users, and
larger 𝜇 indicates stronger social influence.
Two more actions may be taken by user i at each time step t. First, with prob‑
ability p, the user reposts a concordant message from the screen, if any are avail‑
able; otherwise, with probability 1 − p , they post a new message reflecting their own
opinion. Second, with probability q, the user selects a random discordant message
from the screen, if one exists, and unfollows the friend who (re)posted the message,
following a new friend in their place. We explore three different rewiring strategies
for selecting the new friend:
– random: a user is selected at random among all nodes in the network that are not
already friends of i’s;
– repost: a user is selected at random among the originators of reposts, if any are
on i’s screen;
– recommendation: a user is selected at random among non-friends who recently
posted concordant messages.
Note that the size, density, and out-degree sequence of the network stay the same
throughout each simulation, while the in-degree distribution can change over time.
As we mentioned earlier, our model incorporates various elements that have
been explored in the literature. Both opinion dynamics and the rewiring of social
ties (unfriending) are notably present in the model proposed by Holme and
13
Newman [53], which first explored the roles of the two mechanisms. There are how‑
ever a few key differences between the model presented here and previous models.
One is that in our model, opinions take continuous values and unfriending is based
on bounded confidence. More importantly, when links are rewired, they do not nec‑
essarily select nodes with concordant opinion (this is only one of the three rewiring
strategies we consider); rather, the targets of the selection are the links to be bro‑
ken—those outside the opinion confidence bound. Finally, our model aims to cap‑
ture the crucial features of information diffusion in social sharing platforms, where
influence may take place indirectly. Consider for example the following scenario:
user A posts a piece of information that reflects A’s opinion; user B reshares the
message to their followers, which include user C. Now user C may be influenced
by A’s post, even though A and C are not directly connected, and irrespective of
whether C’s opinion was concordant with or influenced by B’s opinion. This indirect
influence mechanism is asymmetric: the opinion of the consumer of a post changes,
while the opinion of the originator of the post does not. The average opinion is
therefore not conserved, unlike in the model proposed by Deffuant [52].
The code to simulate our model and reproduce our findings is available online at
http://www.github.com/soramame0518/echo_chamber_model.
To facilitate the exploration of our model, we developed an interactive demo
allowing one to run Web-based simulations with different realizations of the model
parameters. The demo makes certain simplifications to be accessible to a broad
audience: it is based on an undirected network, nodes can see all messages from
their neighbors, and unfriending only occurs by random rewiring. Figure 2 provides
a screenshot of the demo, which is available online at http://www.osome.iu.edu/
demos/echo/.
Fig. 2 Screenshot of the echo-chamber model demo
13
Results
Emergence of echo chambers
To illustrate the dynamics of the model, in Fig. 3 we show one simulation run.
Over time, due to social influence and unfriending, each user is increasingly
exposed to similar messages endorsed by friends (Fig. 3a), and the system reaches
a steady state characterized by two distinctive features often observed in reality:
opinion polarization (Fig. 3b) and network segregation (Fig. 3c). Note that by
“polarization” we mean a division of opinions into distinct homogeneous groups,
which are not necessarily at the extremes of the opinion range.
We wish to examine the conditions under which opinion polarization and
network segregation coevolve. Recall that our model has two mechanisms that
appear to be relevant to this process: social influence (regulated by parameters 𝜖
and 𝜇 ) and rewiring (regulated by q). Let us first examine the role of the confi‑
dence bound parameter 𝜖 . This threshold affects the number of final opinion clus‑
ters and the diversity of surviving opinions, in a manner consistent with the origi‑
nal bounded confidence model [52] and some of its extensions [55]. As shown
in Fig. 4a, b, the smaller 𝜖 , the more opinion clusters with more heterogeneous
opinions. If 𝜖 is large enough, the network converges to a single homogeneous
opinion cluster.
A B
Message diversity (bit)
Opinion
C
t=0 0 0 0 1 = t 0 0 0 2 = t
1
Opinion
-1
Fig. 3 Coevolution of opinion polarization and network segregation. a Average diversity of messages on

the screen, measured using Shannon entropy with the opinion range divided into 10 bins. b Temporal
changes in population opinions. c Temporal changes in the social network structure. We use parameters
N = 100 , E = 400 , 𝜖 = 0.4 , 𝜇 = 0.5 , p = 0.5 , q = 0.5 , and l = 10 . A random rewiring strategy is applied,
but similar dynamics are observed with different strategies
13
Fig. 4 Dependence of stationary opinions on bounded confidence parameter 𝜖 : a number of opinion

peaks and b maximum distance between opinions [54]. The plots consider opinions at the steady state
and show averages and standard deviations across 20 simulation runs with N = 100 , E = 400 , 𝜇 = 0.5 ,
p = 0.5 , and q = 0.5
Role of influence and rewiring
Next, let us explore the joint effects of influence and rewiring. Here we limit our
attention to the case 𝜖 = 0.4 , which yields, on average, two segregated opinion
groups as illustrated in Fig. 3. In the presence of social influence alone without
rewiring (Fig. 5a), the network structure is unaffected, but opinions may become
polarized after a long time. In the presence of rewiring alone (Fig. 5b), no opin‑
ion change can happen but like-minded users cluster together making polarization
Fig. 5 Conditions for the coevolution of opinion polarization (top) and network segregation (bottom).
Left: 𝜇 = 0.1 and q = 0 . Center: 𝜇 = 0 and q = 0.1 . Right: 𝜇 = 0.1 and q = 0.1
13
inevitable; the network may become segregated after a very long time. The joint
effect of social influence and rewiring accelerates the joint emergence of both polar‑
ization and segregation (Fig. 5c).
To further explore how influence and rewiring jointly affect the speed of emer‑
gence of echo chambers, we plot in Fig. 6a the time until two conditions are both
satisfied: (i) the network clusters are segregated and (ii) opinions are homogeneous,
i.e., any two nodes within the same cluster have opinions that differ by less than
the bounded confidence 𝜖 . In some cases, a cluster may form that is smaller than
the out-degree of one or more of its nodes, so that links from these nodes cannot
be rewired to their own cluster; these cases are excluded because segregation can
never occur. With these exceptions, segregated clusters always form when q > 0 and
𝜇 > 0 . This may take a long time if q and 𝜇 are exceedingly small. However, even
relatively small amounts of influence and rewiring greatly accelerate the emergence
of segregated and polarized echo chambers. When both the rewiring rate q and the
influence strength 𝜇 are above 0.1, echo chambers appear in a fraction of the time.
We therefore observe a synergistic effect in which influence and unfollowing rein‑
force each other in leading to the formation of echo chambers. Time until the emer‑
gence of echo chambers also shows an expected linear dependence on network size,
as shown in Fig. 6b.
Effects of rewiring strategies
All three rewiring strategies used in the model (random, repost, recommendation)
produce comparable stable states in terms of the number and diversity of sta‑
tionary opinion clusters. In other words, what leads to an echo chamber state is
selective unfollowing and not the specific mechanism by which one selects a new
friend to follow. However, the emergence of echo chambers is greatly acceler‑
ated by the rewiring strategy based on recommendations of users with concordant
Fig. 6 Time until emergence of echo chambers as a function of a influence strength 𝜇 and rewiring rate
q and b network size (N). In a, we use a logarithmic scale to explore small parameter values. For each
(𝜇, q) parameter configuration we ran 200 simulations with N = 100 , E = 400 , 𝜖 = 0.4 , and p = 0.5 . A
few simulations were excluded (see text), so that the median number of runs is 197. Colors represent
averages across these simulation runs. The simulations were stopped after tmax = 105 steps in cases when
segregation and convergence have not both occurred yet. In b, we use the same network density as in a
and plot means and standard deviations across 10 simulation runs
13
opinions. The speed of convergence to the steady state is more than doubled com‑
pared to the other rewiring strategies.
The rewiring strategy also affects the development of closed social tri‑
ads. A closed social triad is a network motif with three nodes A, B, C and links
A → B, A → C, C → B . It can be thought as the smallest unit of an echo cham‑
ber network [56], since it enables the same information to be transmitted from a
source A to a recipient B through different paths, directly and via an intermediary
C. As shown in Fig. 7a, rewiring strategies based on recommendations of users
with concordant opinions and on exposure via reposts—both common mecha‑
nisms in social media—result in significantly more closed triads than following
users at random. Repost-based rewiring, in particular, leads to doubling the num‑
ber of directed closed triads, making it much more likely that users are exposed to
the same opinions from multiple sources. The number of users posting/reposting
a message can affect its ranking and be displayed to the user through platform-
dependent interface elements, boosting user attention and exposure.
Finally, the rewiring strategy affects the in-degree distribution of the network
in the stable state. Compared to random rewiring, the other two methods yield
more skewed in-degree distributions, indicating the spontaneous emergence of
popular users with many followers, whose messages have potentially broader
reach (Fig. 7b). Again, the effect on hubs is stronger in the case of repost-based
rewiring. This is consistent with the copy model for network growth, which
approximates preferential attachment [57]. However, unlike the copy model, the
number of nodes and links is fixed in our model; only the patterns of connec‑
tion change. Thus, the skewed in-degree distribution arises due to the spread of
information. Since recipients can see who originally posted each message, the
originators of popular messages have the best chance of receiving new followers
and becoming hubs.
Fig. 7 Effects of different rewiring strategies on evolved social networks. a The number of closed triads
is averaged across 20 simulations with N = 100 and E = 400 ; all differences are statistically significant
( P < 0.01 ). b Cumulative in-degree distributions with N = 104 and E = 105
13
Empirical validation
It is tempting to use our model to reproduce a few stylized facts about empirical echo
chambers. To this end, let us consider data about an empirical retweet network of US
political conversations (see Methods A). To fit the model against this data, we plug
in values of known parameters estimated in prior work, and then perform a sweep of
the remaining parameters (Methods B). We simulate the resulting calibrated model
to see if the synthetic network snapshot generated at the end of the simulation is in
agreement with the observed snapshot of the empirical network (Methods C). As
a stopping criterion for the simulations, we check that the simulated network has
reached the same level of segregation as the empirical one (Methods D).
Figure 8 shows the results. The segregation is one feature of the network structure
that the model can reproduce. Aside from this, the random rewiring scheme used in
the model cannot produce networks with heterogeneous degree. However we can
draw a comparison between the empirical data and our simulations based on two
other metrics. The first is the fraction of closed triads in the network. To compute
the number of triads, we record each time a user reposts something in our model as
a ‘retweet,’ and build a simulated retweet network. We then count all instances of
closed directed triangles in both networks (Methods E). The central panel in Fig. 8
shows that the fraction of triads in the synthetic network is consistent with that
observed in the empirical data.
The second metric is the distribution of opinion distances. We infer the latent
opinions of the Twitter users in our data from their hashtag usage (Methods F), and
define a distance dt in hashtag binary vector space. In the model, we simply consider
the distance do (oi , oj ) = |oi − oj | between two users in the opinion space. Figure 8
shows that the average distance between neighbor opinions decreases, but we also
want to compare the distributions. Figure 9 shows that both distance distributions
have peaks around low values of distance for users in the same cluster and around
high values for users in different clusters. While the ways in which the distances are
Fig. 8 Comparison between model and empirical retweet networks. The solid blue lines represent the
evolution of three metrics as a function of simulated time (epochs). The dashed line represents the empir‑
ical value for the segregation index (left) and fraction of closed directed triads (center), defined in Meth‑
ods D and E respectively. Diversity (right) is the average distance between neighbor opinions
13
Fig. 9 Distribution of pairwise opinion distances. The calibrated model was simulated until the segrega‑
tion of the synthetic retweet network matched the observed one (see Methods E and Fig. 8). The main
plot shows the distribution of opinion distances do across pairs of users in the simulated network. The
inset shows the distribution of opinion distances dt among Twitter users from the empirical data
measured and consequently the distributions are necessarily different, the qualita‑
tively similar bimodal behaviors suggest that the calibrated model attains an analo‑
gous level of opinion polarization in correspondence of the level of network segre‑
gation observed in the empirical data.
To confirm that the proposed model yields echo chambers when applied to larger
networks with realistic degree distributions, we conducted a simulation using an
empirical Twitter follower network with N = 14,818 nodes and E = 428,557 edges
(Methods A). As shown in Fig. 10, the simulation results confirm that both opinion
polarization and network segregation emerge from this empirical network.
Discussion
In studying an idealized social media platform using an agent-based model, we fol‑

lowed the rich tradition of several models of opinion dynamics under social influ‑
ence [21], in which agents adjust their opinions based on those of the ones with
whom they are connected (social influence), and rewire their ties with peers based
on their shared opinions (social selection). Other models have explored the essential
tension between social influence and social selection [53, 58–61]. The effect of the
interaction between these two mechanisms on the emergence of opinion clusters has
also been studied within a bounded confidence framework similar to the one pre‑
sented here [62, 63].
All of these models present extremely simplified versions of real social networks.
While the model presented here is no exception, it seeks to capture more closely
the key components of social media by focusing on indirect interactions enabled
by information diffusion, in addition to disagreement-driven dissolution of ties via
13
Fig. 10 Simulation of model on a large empirical network. a Heavy-tailed in-degree distribution in the

initial network. b Initial ( t = 0 ) and final ( t = 106 ) opinion distributions, showing the emergence of
polarization. c Visualization of the follower network after t = 106 steps, showing strong segregation
unfollowing/unfriending. Furthermore, our model combines social influence and

selection with the competition for limited attention, which has been shown to explain
the empirical scale-free distribution of content popularity in social media [51, 64].
The results presented here suggest that the proliferation of online echo chambers
may be an inevitable outcome of basic cognitive and social processes facilitated by
social media: namely, the human tendency to be influenced by information and opin‑
ions to which one is exposed, and that of disliking disagreeable social ties. Social
influence and rewiring appear to provide synergistic conditions for the rapid for‑
mation of completely segregated and polarized echo chambers; this phenomenon is
accelerated by an order of magnitude in the presence of both relatively strong influ‑
ence and relatively common unfollowing, compared to cases when either mecha‑
nism is weak.
A social network that is both segregated and polarized can be also generated with
a variant of the Schelling model [36] on networks, proposed by Henry et al. [65].
This model is based on aversion-driven rewiring, but starts from a bimodal distribu‑
tion of opinions. Our approach shows how both polarization and segregation emerge
without assuming that opinions are already polarized.
The literature has explored other factors and mechanisms that foster the emer‑
gence of isolated cultural or political subnetworks as well as polarization of
opinions. Network transitivity, which is also incorporated in our model, has been
shown to lead to the formation of groups when combined with reciprocity [66] or
13
with one-to-many communication [67]. Pressures toward stronger opinions [68]

or more radical opinions [69] are not included in our model, nor are repulsion
effects driving opinions farther apart [13, 55]. Echo chambers can also emerge
from cognitive mechanisms, such as confirmation bias, when information propa‑
gates through centralized channels reaching a large portion of the population [70].
Finally, geographic segregation in urban areas may promote polarization in both
physical and online spaces by fracturing the social space of mutual exposure [71].
Our model focuses exclusively on information spreading mechanisms that are
characteristic of online social media platforms.
Focusing on the rewiring of social ties, we tested three different selection
mechanisms: two inspired by triadic closure and social recommendation—
intended to model the ways in which social media work in practice—and one
based on purely random choice. All variations yield qualitatively similar steady
states, suggesting that disagreement-driven unfollowing is a sufficient rewiring
condition for echo chamber emergence. The number of groups does depend on
the tolerance to different opinions, as predicted by bounded confidence mod‑
els [52]. This suggests that the extent to which social media allow users to exer‑
cise their preferences in determining their connections will either integrate or
fragment interaction [72].
The more realistic selection mechanisms help explain two additional features
of online social networks. First, the presence of users with many followers. These
hub nodes affect the dissemination of the same messages in many cases, but not
always [73]. Second, the large number of closed triads [26]. Triadic closure con‑
nects individuals to friends of their friends, facilitating repeated exposure to the
same opinion. Such “echoes” are a powerful reinforcing mechanism for the adoption
of beliefs and behaviors [74].
Although our model does not account for the adoption of false information, it
has been speculated that echo chambers may make social media users more vulner‑
able to this kind of manipulation [75, 76]. There are multiple ways in which echo-
chamber structure may contribute to the spread of misinformation. First, because
people are repeatedly exposed to homogeneous information inside an echo chamber,
the selection of belief-consistent information and the avoidance of belief-inconsist‑
ent information are facilitated, reinforcing confidence in minority opinions, such as
conspiracy theories and fabricated news, even in the presence of preponderant con‑
trary evidence [1]. Second, echo chambers foster herding, which may lead to quick
and premature convergence to suboptimal solutions of complex problems and sim‑
plistic interpretations of complex issues [1, 7]. Third, the threshold for perceiving
a piece of content as novel may be lower within echo chambers by virtue of the
reduced diversity of viewpoints to which people are exposed. The crafting of false
news with perceived novelty may thus be promoted, leading to faster and broader
consumption of misinformation [77] and triggering the production of more informa‑
tion about similar topics [78]. Finally, echo chambers may reinforce the influence
bias of personalized filtering algorithms toward a user’s current opinions [79]. On
the other hand, recent experimental results suggest that social information exchange
in homogeneous networks increases accuracy and reduces polarization [80], casting
doubt on theories that political echo chambers reduce belief accuracy. More work is
13
certainly needed to understand the relationship between online echo chambers and
misinformation.
Our results suggest possible mitigation strategies against the emergence of echo
chambers on social media. Often-recommended solutions involve exposure to con‑
tent that increases a user’s social distance from their preferences. However, such
strategies must be consistent with current understanding of cognitive biases [1]. For
example, it does not help to promote content that will be disregarded [76]. A more
neutral possibility suggested by our findings (Fig. 7a) is to discourage triadic closure
when recommending the formation of new social ties. By exposing the sources of
information reposted by our friends and recommending similar users, social media
platforms encourage linking choices that maximize triadic closure. While exact
knowledge about recommendation algorithms is not available, we know that triadic
closure plays a key role. But even among the many friends of our friends, an algo‑
rithm could optimize for diversity of opinions. Moreover, the complete dissolution
of ties with those users with whom one disagrees should be discouraged, for exam‑
ple by providing alternative mechanisms—some social media platforms are experi‑
menting with solutions like snooze buttons [81]—or the possibility to block only
certain types of information, but not others. Another approach is to alert users who
are about to unfollow their only conduits to certain types or sources of information.
As we better understand the unintended consequences of social media mecha‑
nisms, ethical and transparent user studies are needed to carefully evaluate coun‑
termeasures before they are deployed by platforms. We must not only ensure that
new mechanisms mitigate undesired outcomes, but also that they do not create new
vulnerabilities.
Methods
A. Data
To evaluate the model’s prediction, we use empirical data from Conover et al. [9],
who studied the political polarization of Twitter users. The data comprises a sample
of public tweets about US politics, collected during the 6 weeks prior to the 2010
US midterm elections. The tweets were obtained from a 10% random sample of all
public tweets. The dataset is available online at http://www.cnets.indiana.edu/group
s/nan/nan-datasets-and-data-tools/#icswm2011_2.
Tweets with hashtags about US politics were included in the dataset. The hashtags
were drawn from a list, which was obtained by expanding a manually curated seed
set of then-popular political hashtags, such as #TCOT (‘Top conservatives on Twit‑
ter’) and #P2 (‘Progressives 2.0’). This initial set was recursively expanded with co-
occurring hashtags above a minimum frequency, until no additional hashtag could
be found. Finally, the list was manually checked and any hashtag that was not about
US politics was expunged. The final list included 6372 hashtags about US politics.
Three networks are provided in the dataset: retweets, mentions, and retweets
plus mentions combined. We used the largest strongly connected component of the
retweet network ( N = 18,470 and E = 48,365 ), which is known to be polarized in
13
two groups, roughly corresponding to the two main US political factions—conserva‑

tives and progressives. The network is the same shown in Fig. 1.
For the larger empirical network, we started from a follower network with
N = 41,652,230 nodes and E = 1,468,364,884 edges [82]. To produce a more man‑
ageable dataset, we first randomly sampled edges from the graph and focused on the
network spanned by these edges, and then applied k-core decomposition [83] with
k = 30 . This yielded a reduced network with N = 14,818 nodes and E = 428,557
edges, which was used to initialize the simulation.
B. Parameter fitting
Our model includes several parameters that need to be estimated. The rate of repost‑
ing was set to p = 0.25 based on empirical results from Twitter [84]. Similarly, the
screen length was set to l = 10 , which is close to the average number of times that
a user stops on a post during a scrolling session on a social blogging platform [84].
The number of nodes in the simulations was taken to be the same as the num‑
ber of Twitter users in the empirical retweet network. Edges were drawn at random
between any two nodes with probability chosen so that the density of the follower
graph is d = 1.8 × 10−4 . This value is within the range observed in the literature [85,
86].
We performed a parameter scan for the rest of the parameters, finding the follow‑
ing values: influence strength 𝜇 = 0.015 and tolerance 𝜖 = 0.65 . Note that the toler‑
ance value to reproduce the two opinion clusters in the US-based online political
conversations is larger than the value 𝜖 = 0.4 found for smaller networks (Fig. 4).
Finally, for simplicity, we use the random rewiring rule.
C. Model evaluation
Our goal is to compare model predictions about the emergence of echo chambers
with empirical data from social media. Unfortunately, lacking a probability distribu‑
tion over the data, our model does not allow us to compute the likelihood of a given
network or distribution of opinions. Thus we need to devise a method to evaluate
our approach. This has become a common challenge, especially with the rise of
agent-based modeling in the social sciences [87]. There is a vast literature devoted
to developing rigorous methods to test simulation models on empirical data of social
phenomena [88, 89]. Although no single universal recipe exists, we adopt the com‑
mon approach of generating synthetic data from our agent-based model and compar‑
ing them to the empirical data under appropriate distance measures.
Our main hypothesis is that both social influence and rewiring are required to
reproduce patterns consistent with the empirical data. Under those conditions, the
system will always reach a state in which there will be no ties connecting users with
discordant opinions (see Fig. 5). However, the empirical network is not completely
disconnected in two communities. Therefore, we simulate our model until the sys‑
tem reaches the same level of segregation observed in the empirical data, and com‑
pare the two networks.
13
The empirical and model networks are however different. The former is a network of
retweets, while the latter is a network of ‘follower’ ties. Therefore we cannot compare
these two networks directly, but instead we generate a synthetic retweet network from
the simulated data. Every time a user performs a ‘repost’ action in our simulations, we
count it as a retweet, and add the corresponding edge in the simulated retweet network.
The plots in Figs. 8 and 9 are based on snapshots of this synthetic retweet network,
taken at evenly spaced time intervals of 10 epochs each. Each epoch consists of N steps
of the model, so that each user performs one post and/or rewiring action per epoch
on average. At the end of each tenth epoch we consider the latest E distinct retweet
edges, so that each simulated network snapshot is guaranteed to have the same number
of edges as the empirical one. We then consider the largest strongly connected compo‑
nent of each simulated network snapshot. Therefore, the two networks do not generally
match in the number of nodes.
D. Segregation
To measure the segregation in both the simulated and empirical networks, we group
users into two clusters (C− , C+ ) . In the simulated network, C− is defined as the set of
users having opinion o < 0 and C+ as the set with o ≥ 0 . In the empirical network, the
two clusters correspond to the labels obtained via label propagation [20]. Let Eb denote
the set of edges connecting nodes in different clusters. We define the segregation index
as:
|Eb |
s=1− . (3)
2d |C+ | |C− |
The segregation index compares the actual number of edges across the two clusters
with the number we would observe in a random network with the same density d.
When the network is completely segregated, s = 1.
E. Closed triads
Let us denote by i → j a directed edge from node i to node j. A closed directed triangle
or closed triad is any node triplet i, j, k ∈ V such that {i → j, j → k, i → k} ⊆ E . An
open directed triangle or open triad is any node triplet for whom only a proper subset of
those edges exists in E. Let us denote by T the set of closed triads and by Tu the set of
open triads. We compute the frequency of closed triads as the ratio
|T| |T|
fT = = ,
|T ∪ Tu | NT !∕(NT − 3)!
( )
N
where NT = is the number of all node triplets.
3
13
F. Latent opinion inference
Our model generates opinions in the [−1, +1] range, while the empirical network has
binary labels (‘liberal’ or ‘conservative’) inferred from a training set and propaga‑
tion through the network [20]. Comparing the opinions predicted by the model to
these labels would not be meaningful, since the labels are trivially correlated with
the network structure, by construction.
A more meaningful comparison is between pairwise opinion distances, which
we can generate for the Twitter users using a criterion that is not induced by the
network’s community structure. Since hashtag usage is also polarized [9], we infer
opinions distances from adopted hashtags. We say that a hashtag is adopted by a user
if it is found either in tweets retweeted by the user (incoming edges), or in tweets by
the user that were retweeted by someone else (outgoing edges). Let us consider the
i-th user and the k-th hashtag. We define an empirical opinion vector ô i ∈ {0, 1}D
where ô ik = 1 if user i adopted hashtag k, and 0 otherwise. We define the empirical
opinion distance between two opinion vectors based on shared tags:
� � ô i ⋅ ô j
dt ô i , ô j = 1 − � �, (4)
min ‖̂oi ‖, ‖̂oj ‖
where ‖ ⋅ ‖ is the L1 norm, or number of shared tags. To mitigate the effects of spar‑
sity and noise, we use only the D = 20 most popular hashtags to define the vectors.
The selected hashtags were adopted by 93% of the users. We restrict the retweet net‑
work to the subgraph spanned by those users, but the overall results do not change
significantly if we select enough hashtags to cover 100% of the users.
Acknowledgements We are grateful to Thomas Hills and Marco Minici for helpful feedback. K.S. was
supported by JST PRESTO grant no. JPMJPR16D6, JST CREST grant no. JPMJCR17A4, and JSPS/
MEXT KAKENHI grant numbers JP19H04217 and JP17H06383 in #4903. G.L.C. was supported by the
Indiana University Network Science Institute, where he carried out the work presented here. F.M. and
A.F. were supported in part by DARPA contract no. W911NF-17-C-0094. Any opinions, findings, and
conclusions or recommendations expressed in this material are those of the authors and do not necessar‑
ily reflect the views of the funding agencies.
Author Contributions K.S., G.L.C., A.F., and F.M. designed the research. K.S. and W.C. performed sim‑
ulations, W.C. and F.M. analyzed the data. H.P. developed the online demo. K.S., G.L.C., F.M., and A.F.
drafted the manuscript. All authors reviewed and approved the manuscript.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License,
which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long
as you give appropriate credit to the original author(s) and the source, provide a link to the Creative
Commons licence, and indicate if changes were made. The images or other third party material in this
article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line
to the material. If material is not included in the article’s Creative Commons licence and your intended
use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permis‑
sion directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/
licenses/by/4.0/.
13
Fig. 11 Distribution of average change in the number of user friends per tweet (left) and per day (right),
based on a sample of active Twitter users
Appendix: Prevalence of unfollow events
Our model features edge rewiring to represent the dissolution of a social tie. This
is based on the assumption that users cut social links by occasionally unfollowing
friends. To get a more quantitative estimate of empirical unfollow frequencies, we
collected 451,844 tweets from April 25, 2020 to May 25, 2020 from a 10% random
sample of tweets. In this sample, 16,317 users had at least two tweets. We sorted
the tweets chronologically, and for each pair of consecutive tweets by the same user
we measured the change Δf in the number of friends of the user (i.e., the number of
people followed by the user). By comparing the timestamps of the two tweets we
could estimate the change in the number of friends per day. Similarly, by comparing
the total numbers of tweets, we estimated the change in the number of friends per
tweet. In the latter case, we removed pairs where the number of tweets decreased
due to tweet deletions. Finally, we averaged across all consecutive pairs from the
same user to get a single average change estimate for each user.
Figure 11 plots the cumulative distributions of the average changes in num‑
bers of followers across the active Twitter users in the sample. Approximately
18.5% of users have negative changes, meaning that the number of friends has
decreased—more unfollows than new friends, on average. This number can be
understood as a lower bound on the actual probability of unfollowing.
References
1. Hills, T. T. (2019). The Dark Side of Information Proliferation. Perspectives on Psychological
Science, 14(3), 323. https://doi.org/10.1177/1745691618803647.
2. Bonabeau, E. (2009). Decisions 2.0: The Power of Collective Intelligence. MIT Sloan Manage-
ment Review, 50(2), 45.
3. Page, S. (2008). The Difference: How the Power of Diversity Creates Better Groups, Firms,
Schools, and Societies. Princeton: Princeton University Press.
4. Surowiecki, J. (2005). The Wisdom Of Crowds. New York: Anchor Books.
5. Benkler, Y. (2006). The Wealth of Networks: How Social Production Transforms Markets and
Freedom. New Haven: Yale University Press.
6. Salganik, M. J., Dodds, P. S., & Watts, D. J. (2006). Experimental study of inequality and unpre‑
dictability in an artificial cultural market. Science, 311(5762), 854.
13
7. Ciampaglia, G. L., Nematzadeh, A., Menczer, F., & Flammini, A. (2018). The production of
information in the attention economy. Scientific Reports, 8, 15951. https://doi.org/10.1038/s4159
8-018-34203-2.
8. Stewart, A. J., Mosleh, M., Diakonova, M., Arechar, A. A., Rand, D. G., & Plotkin, J. B. (2019).
Information gerrymandering and undemocratic decisions. Nature, 573(7772), 117. https://doi.
org/10.1038/s41586-019-1507-6.
9. Conover, M., Ratkiewicz, J., Francisco, M., Gonçalves, B., Menczer, F., Flammini, A. (2011). in
International AAAI Conference on Web and Social Media (AAAI, Palo Alto, CA, USA, 2011),
ICWSM ’11, pp. 89–96. https://www.aaai.org/ocs/index.php/ICWSM/ICWSM11/paper/view/2847
10. Conover, M. D., Gonçalves, B., Flammini, A., & Menczer, F. (2012). Partisan asymmetries in online
political activity. EPJ Data Science, 1(1), 6.
11. J. Bright, arXiv:1609.05003v2 (2017)
12. Bakshy, E., Messing, S., & Adamic, L. A. (2015). Exposure to ideologically diverse news and opin‑
ion on Facebook. Science, 348(6239), 1130.
13. Del Vicario, M., Bessi, A., Zollo, F., Petroni, F., Scala, A., Caldarelli, G., et al. (2016). The spread‑
ing of misinformation online. Proceedings of the National Academy of Science U.S.A., 113(3), 554.
14. Flaxman, S., Goel, S., & Rao, J. M. (2016). Filter Bubbles, Echo Chambers, and Online News Con‑
sumption. Public Opinion Quarterly, 80(S1), 298. https://doi.org/10.1093/poq/nfw006.
15. Schmidt, A. L., Zollo, F., Del Vicario, M., Bessi, A., Scala, A., Caldarelli, G., et al. (2017). Anat‑
omy of news consumption on Facebook. Proceedings of the National Academy of Science U.S.A.,
114(12), 3035. https://doi.org/10.1073/pnas.1617052114.
16. Marsden, P. V. (1988). Homogeneity in Confiding Relations. Social Networks, 10(1), 57. https://doi.
org/10.1016/0378-8733(88)90010-X.
17. McPherson, M., Smith-Lovin, L., & Cook, J. M. (2001). Birds of a Feather: Homophily in Social
Networks. Annual Review of Sociology, 27(1), 415. https://doi.org/10.1146/annurev.soc.27.1.415.
18. Centola, D. (2011). Science, 334(6060), 1269. https://doi.org/10.1126/science.1207055.
19. Friedkin, N. E. (2006). A structural theory of social influence (Vol. 13). Cambridge: Cambridge
University Press.
20. Conover, M., Gonçalves, B., Ratkiewicz, J., Flammini, A., Menczer, F. (2011). in Proceedings of
3rd IEEE Conference on Social Computing (SocialCom), pp. 192–199. https://doi.org/10.1109/
PASSAT/SocialCom.2011.34
21. Castellano, C., Fortunato, S., & Loreto, V. (2009). Review of Modern Physics, 81, 591. https://doi.
org/10.1103/RevModPhys.81.591.
22. Backstrom, L. (2016). In Proceedings of the Ninth ACM International Conference on Web

Search and Data Mining (ACM, New York, NY, USA), WSDM ’16, pp. 469–469. doi
10.1145/2835776.2835848
23. Adamic, L. A., & Adar, E. (2003). Social Networks, 25(3), 211. https://doi.org/10.1016/S0378
-8733(03)00009-1.
24. Backstrom, L., Huttenlocher, D., Kleinberg, J., Lan, X. (2006). In Proceedings of the 12th ACM
SIGKDD International Conference on Knowledge Discovery and Data Mining. (ACM, New York,
NY, USA), KDD ’06, pp. 44–54. doi 10.1145/1150402.1150412
25. Leskovec, J., Backstrom, L., Kumar, R., Tomkins, A. (2008). In Proceedings of the 14th ACM SIG‑
KDD International Conference on Knowledge Discovery and Data Mining (ACM, New York, NY,
USA, 2008), KDD ’08, pp. 462–470. https://doi.org/10.1145/1401890.1401948
26. Weng, L., Ratkiewicz, J., Perra, N., Gonçalves, B., Castillo, C., Bonchi, F., Schifanella, R., Menczer,
F., Flammini, A. (2013). In Proceedings of the 19th ACM SIGKDD International Conference on
Knowledge Discovery and Data Mining (ACM, New York, NY, USA), KDD ’13, pp. 356–364. doi
https://doi.org/10.1145/2487575.2487607.
27. Sibona, C., Walczak, S. (2011). In 2011 44th Hawaii International Conference on System Sciences ,
pp. 1–10. https://doi.org/10.1109/HICSS.2011.467.
28. Raeder, T., Lizardo, O., Hachen, D., & Chawla, N. V. (2011). Social Networks, 33(4), 245. 10.1016/j.
socnet.2011.07.002. http://www.sciencedirect.com/science/article/pii/s0378873311000463.
29. H. Kwak, H. Chun, S. Moon, in Proceedings of the SIGCHI Conference on Human Factors in Com-
puting Systems (Association for Computing Machinery, New York, NY, USA, 2011), CHI ’11, p.
1091–1100. https://doi.org/10.1145/1978942.1979104.
30. Nickerson, R. S. (1998). Review of General Psychology, 2(2), 175.
31. Sears, D. O., & Freedman, J. L. (1967). Public Opinion Quarterly, 31(2), 194. https://doi.

org/10.1086/267513.
13
32. Sunstein, C. R. (2017). Republic: Divided Democracy in the Age of Social Media. Princeton Univer‑
sity Press: Princeton.
33. Gentzkow, M., & Shapiro, J. M. (2011). The Quarterly Journal of Economics, 126(4), 1799. https://
doi.org/10.1093/qje/qjr044.
34. Katz, J. E. (1998). The ANNALS of the American Academy of Political and Social Science, 560(1),
194. https://doi.org/10.1177/0002716298560001015.
35. C.R. Sunstein, Republic.com, reprint edn. (Princeton University Press, Princeton, NJ, USA, 2002)
36. Schelling, T. C. (1971). The Journal of Mathematical Sociology, 1(2), 143. https://doi.

org/10.1080/0022250X.1971.9989794.
37. E. Pariser, The filter bubble: What the Internet is hiding from you (Penguin UK, 2011)
38. Jamieson, K., & Cappella, J. (2008). Echo Chamber: Rush Limbaugh and the Conservative Media
Establishment. Oxford: Oxford University Press.
39. Fortunato, S., Flammini, A., Menczer, F., & Vespignani, A. (2006). Proceedings of the National
Academy of Sciences, 103(34), 12684. https://doi.org/10.1073/pnas.0605525103.
40. Nikolov, D., Oliveira, D. F., Flammini, A., & Menczer, F. (2015). PeerJ Computer Science, 1, e38.
https://doi.org/10.7717/peerj-cs.38.
41. Nikolov, D., Lalmas, M., Flammini, A., & Menczer, F. (2019). Journal of the Association for. Infor-
mation Science and Technology, 70(3), 218. https://doi.org/10.1002/asi.24121.
42. A. Guess, B. Lyons, B. Nyhan, J. Reifler. Avoiding the echo chamber about echo chambers: Why
selective exposure to like-minded political news is less prevalent than you think. Knight Foundation
(2018). https://kf-site-production.s3.amazonaws.com/media_elements/files/000/000/133/original/
Topos_KF_White-Paper_Nyhan_V1.pdf
43. Hart, W., Albarracín, D., Eagly, A. H., Brechan, I., Lindberg, M. J., & Merrill, L. (2009). Psycho-
logical Bulletin, 135(4), 555.
44. Stroud, N. J. (2010). Journal of Communication, 60(3), 556.
45. E. Dubois, G. Blank, Information, Communication & Society 21(5), 729 (2018). https://www.tandf
online.com/doi/full/10.1080/1369118X.2018.1428656
46. Garrett, R. K. (2009). Journal of Computer-Mediated Communication, 14(2), 265.
47. J. Shore, J. Baek, C. Dellarocas, MIS Quarterly 42(3), 849 (2016). https://osf.io/preprints/socarxiv/32urj/
48. A.M. Guess. (Almost) Everything in Moderation: New Evidence on Americans’ Online Media Diets.
Unpublished (2018). https://webspace.princeton.edu/users/aguess/Guess_OnlineMediaDiets.pdf
49. K. Garimella, G. De Francisci Morales, A. Gionis, M. Mathioudakis, in Proceedings of the 2018
World Wide Web Conference (2018), pp. 913–922
50. Flache, A., Mäs, M., Feliciani, T., Chattoe-Brown, E., Deffuant, G., Huet, S., et al. (2017). Journal
of Artificial Societies and Social Simulation, 20, 4.
51. Weng, L., Flammini, A., Vespignani, A., & Menczer, F. (2012). Scientific Reports, 2, 335.
52. G. Deffuant, D. Neau, F. Amblard, Advances in Complex Systems 03(01n04), 87 (2000)
53. Holme, P., & Newman, M. E. J. (2006). Physical Review E, 74, 056108. https://doi.org/10.1103/
PhysRevE.74.056108.
54. L.H. Negri, C. Vestri. lucashn/peakutils: v1.1.0. Zenodo (2017). doi 10.5281/zenodo.887917.
55. Del Vicario, M., Scala, A., Caldarelli, G., Stanley, H. E., & Quattrociocchi, W. (2017). Scientific
Reports, 7, 40391.
56. Jasny, L., Waggle, J., & Fisher, D. R. (2015). Nature Climate Change, 5(8), 782.
57. J.M. Kleinberg, R. Kumar, P. Raghavan, S. Rajagopalan, A.S. Tomkins, in Computing and Combi-
natorics (Springer, 1999), pp. 1–17
58. D. Crandall, D. Cosley, D. Huttenlocher, J. Kleinberg, S. Suri, in Proceedings of the 14th ACM
SIGKDD International Conference on Knowledge Discovery and Data Mining
59. Durrett, R., Gleeson, J. P., Lloyd, A. L., Mucha, P. J., Shi, F., Sivakoff, D., et al. (2012). Proceedings
of the National Academy of Sciences, 109(10), 3682. https://doi.org/10.1073/pnas.1200709109.
60. Y. Yu, G. Xiao, G. Li, W.P. Tay, H.F. Teoh, Chaos: An Interdisciplinary Journal of Nonlinear Sci‑
ence 27(10), 103115 (2017). https://doi.org/10.1063/1.4989668
61. Teza, G., Suweis, S., Gherardi, M., Maritan, A., & Lagomarsino, M. C. (2019). Physics Review E,
99(3), 032310.
62. Kozma, B., & Barrat, A. (2008). Physics Review E, 77, 016102. https://doi.org/10.1103/PhysR
evE.77.016102.
63. Kozma, B., & Barrat, A. (2008). Journal of Physics A: Mathematical and Theoretical, 41(22),
224020. https://doi.org/10.1088/1751-8113/41/22/224020.
13
64. Gleeson, J. P., Ward, J. A., O’Sullivan, K. P., & Lee, W. T. (2014). Physics Review Letters, 112,
048701. https://doi.org/10.1103/PhysRevLett.112.048701.
65. Henry, A. D., Prałat, P., & Zhang, C. Q. (2011). Proceedings of National Academy of Science
U.S.A., 108(21), 8605.
66. Gray, K., Rand, D. G., Ert, E., Lewis, K., Hershman, S., & Norton, M. I. (2014). Psychological Sci-
ence, 25(4), 982. https://doi.org/10.1177/0956797614521816.
67. Keijzer, M. A., Mäs, M., & Flache, A. (2018). Complexity, 2018, 9502872. https://doi.

org/10.1155/2018/9502872.
68. Evans, T., & Fu, F. (2018). Royal Society Open Science, 5(10), 181122. https://doi.org/10.1098/
rsos.181122.
69. Baumann, F., Lorenz-Spreen, P., Sokolov, I. M., & Starnini, M. (2020). Physics Review Letters, 124,
048301. https://doi.org/10.1103/PhysRevLett.124.048301.
70. Geschke, D., Lorenz, J., & Holtz, P. (2019). British Journal of Social Psychology, 58(1), 129. https
://doi.org/10.1111/bjso.12286.
71. A.J. Morales, X. Dong, Y. Bar-Yam, A. ‘Sandy’ Pentland, Royal Society Open Science 6(10),
190573 (2019). doi 10.1098/rsos.190573.
72. Van Alstyne, M., & Brynjolfsson, E. (2005). Management Science, 51(6), 851. https://doi.

org/10.1287/mnsc.1050.0363.
73. E. Bakshy, J.M. Hofman, W.A. Mason, D.J. Watts, in Proceedings of the Fourth ACM International
Conference on Web Search and Data Mining (ACM, New York, NY, USA, 2011), WSDM ’11, pp.
65–74. doi 10.1145/1935826.1935845.
74. Centola, D., & Macy, M. (2007). American Journal of Sociology, 113(3), 702. https://doi.

org/10.1086/521848.
75. Lazer, D. M. J., Baum, M. A., Benkler, Y., Berinsky, A. J., Greenhill, K. M., Menczer, F., et al.
(2018). Science, 359(6380), 1094. https://doi.org/10.1126/science.aao2998.
76. Lewandowsky, S., Ecker, U. K., & Cook, J. (2017). Journal of Applied Research in Memory and
Cognition, 6(4), 353.
77. Vosoughi, S., Roy, D., & Aral, S. (2018). Science, 359(6380), 1146. https://doi.org/10.1126/scien
ce.aap9559.
78. Ciampaglia, G. L., Flammini, A., & Menczer, F. (2015). Scientific Reports, 5, 9452. https://doi.
org/10.1038/srep09452.
79. Perra, N., & Rocha, L. E. C. (2019). Scientific Reports, 9, 7261. https://doi.org/10.1038/s41598-019-
43830-2.
80. Becker, J., Porter, E., & Centola, D. (2019). Proceedings of the National Academy of Sciences,.
https://doi.org/10.1073/pnas.1817195116.
81. S. Muraleedharan. Introducing snooze to give you more control of your news feed. https://newsr
oom.fb.com/news/2017/12/news-feed-fyi-snooze/ (2017). Last accessed 27 February 2019
82. J. Leskovec, A. Krevl. SNAP Datasets: Stanford large network dataset collection. http://snap.stanf
ord.edu/data (2014)
83. Alvarez-Hamelin, J. I., Dall’Asta, L., Barrat, A., & Vespignani, A. (2006). Advances in Neural
Information Processing Systems, 41–50.
84. X. Qiu, D. F. M. Oliveira, A. Sahami Shirazi, A. Flammini, F. Menczer, Nature Human Behaviour
1, 0132 (2017). doi 10.1038/s41562-017-0132
85. M. Cha, H. Haddadi, F. Benevenuto, K. Gummadi, in International AAAI Conference on Web and
Social Media (AAAI, Palo Alto, CA, USA, 2010), ICWSM ’10, pp. 10–17. https://www.aaai.org/
ocs/index.php/ICWSM/ICWSM10/paper/view/1538
86. Bollen, J., Gonçalves, B., Ruan, G., & Mao, H. (2011). Artificial Life, 17(3), 237. https://doi.
org/10.1162/artl_a_00034.
87. Epstein, J., & Axtell, R. (1996). Project, Growing Artificial Societies: Social Science from the Bot-
tom Up. USA: Complex Adaptive Systems. Cambridge: MIT Press.
88. P. Windrum, G. Fagiolo, A. Moneta, Journal of Artificial Societies and Social Simulation 10(2), 8
(2007). http://jasss.soc.surrey.ac.uk/10/2/8.html
89. Ciampaglia, G. L. (2013). Advances in Complex Systems, 16, 1350030. https://doi.org/10.1142/
S0219525913500306.
Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published
maps and institutional affiliations.
13

Social Influence and Unfollowing Accelerate The Emergence of Echo Chambers

Uploaded by

Copyright:

Available Formats

Social Influence and Unfollowing Accelerate The Emergence of Echo Chambers

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Social Influence and Unfollowing Accelerate The Emergence of Echo Chambers

Uploaded by

Copyright:

Available Formats

Journal of Computational Social Science

Social influence and unfollowing accelerate the emergence

Kazutoshi Sasahara1,2,3 · Wen Chen3,4 · Hao Peng4,5 ·

Received: 8 December 2019 / Accepted: 20 August 2020

Keywords Echo Chamber · Opinion dynamics · Social media · Social network

Let us introduce a model of opinion dynamics in an evolving social network. We

concordant with an opinion o if they are within a bounded confidence distance 𝜖

Equation 1 provides a simple mechanism for modeling social influence based on an

Fig. 2 Screenshot of the echo-chamber model demo

Emergence of echo chambers

Fig. 3 Coevolution of opinion polarization and network segregation. a Average diversity of messages on

Fig. 4 Dependence of stationary opinions on bounded confidence parameter 𝜖 : a number of opinion

Role of influence and rewiring

Effects of rewiring strategies

In studying an idealized social media platform using an agent-based model, we fol‑

Fig. 10 Simulation of model on a large empirical network. a Heavy-tailed in-degree distribution in the

unfollowing/unfriending. Furthermore, our model combines social influence and

with one-to-many communication [67]. Pressures toward stronger opinions [68]

two groups, roughly corresponding to the two main US political factions—conserva‑

F. Latent opinion inference

Appendix: Prevalence of unfollow events

You might also like

Fig. 2 Screenshot of the echo-chamber model demo

Fig. 3 Coevolution of opinion polarization and network segregation. a Average diversity of messages on

Fig. 4 Dependence of stationary opinions on bounded confidence parameter 𝜖 : a number of opinion

Fig. 10 Simulation of model on a large empirical network. a Heavy-tailed in-degree distribution in the