Intelligent social bots uncover the link between user preference and diversity of news consumption - arXiv

Page created by Ernest Allen
 
CONTINUE READING
Intelligent social bots uncover the link between user preference and diversity of news consumption - arXiv
Intelligent social bots uncover the link between user preference
                                                        and diversity of news consumption
                                                               Yong Min∗                                Tingjun Jiang                               Xiaogang Jin
                                             Zhejiang University of Technology               Zhejiang University of Technology                   Zhejiang University
                                                   Hangzhou, Zhejiang                              Hangzhou, Zhejiang                           Hangzhou, Zhejiang
                                                    myong@zjut.edu.cn                                jtj9678@163.com                            xiaogangj@zju.edu.cn

                                                                                  Cheng Jin                                       Qu Li
                                                                             Tencent Corporation                  Zhejiang University of Technology
                                                                            Shenzhen, Guangdong                         Hangzhou, Zhejiang
                                                                            moonjin@tencent.com                           liqu@zjut.edu.cn
arXiv:1907.02703v1 [cs.SI] 5 Jul 2019

                                        ABSTRACT                                                                  1   INTRODUCTION
                                        The boom of online social media and microblogging platforms has           With the growing popularity of social media and microblogging
                                        rapidly alter the way we consume news and exchange opinions.              platforms, the way people obtain information and form opinions
                                        Even though considerable efforts try to recommend various con-            has undergone substantial change. A recent study found that social
                                        tents to users, loss of information diversity and the polarization of     media has become the primary source for over 60% of users to obtain
                                        interest groups are still an enormous challenge for industry and          news [2]. These users are exposed to more personalized information,
                                        academia. Here, we take advantage of benign social bots to de-            which is considered to limit the diversity of content they consume
                                        sign a controlled experiment on Weibo (the largest microblogging          and even cause filter bubbles [3–5]. In fact, news consumption on
                                        platform in China). These software bots can exhibit human-like            social media has been extensively studied to determine what factors
                                        behavior (e.g., preferring particular content) and simulate the for-      lead to polarization [6, 7]. Recent works suggest that confirmation
                                        mation of personal social networks and news consumption under             bias or selective exposure plays a significant role in online social
                                        two well-accepted sociological hypotheses (i.e., homophily and tri-       dynamics [6, 8]. That is, online users tend to select messages or
                                        adic closure). We deployed 68 bots to Weibo, and each bot ran for         information sources supporting their existing beliefs or cohering
                                        at least 2 months and followed 100 to 120 accounts. In total, we          with their preferences and hence to form filter bubbles [5]. However,
                                        observed 5,318 users and recorded about 630,000 messages exposed          the mechanism of polarization remains an open question that needs
                                        to these bots. Our results show, even with the same selection be-         further clarification.
                                        haviors, bots preferring entertainment content are more likely to            In this paper, we conduct a controlled experiment to explore the
                                        form polarized communities with their peers, in which about 80%           pattern of news consumption and the factors leading to the emer-
                                        of the information they consume is of the same type, which is a           gence of polarization on Weibo (NASDAQ: WB), the most popular
                                        significant difference for bots preferring sci-tech content. The result   Chinese microblogging service. Weibo is the most famous Chinese
                                        suggests that users'preference played a more crucial role in limiting     social networking and microblogging platform where users can
                                        themselves access to diverse content by compared with the two             post messages for instant sharing. It has more than 500 million
                                        well-known drivers (self-selection and pre-selection). Furthermore,       registered users and 313 million active monthly users, including
                                        our results reveal an ingenious connection between specific content       government agencies, celebrities, enterprises, and ordinary neti-
                                        and its propagating sub-structures in the same social network. In         zens. The information dissemination across Weibo is chiefly based
                                        the Weibo network, entertainment news favors a unidirectional             on the “attention” relationship between users [9]. User A can fol-
                                        star-like sub-structure, while sci-tech news spreads on a bidirec-        low user B (i.e., becoming a “follower” of B, and B then becomes a
                                        tional clustering sub-structure. This connection can amplify the          “following” of A) and then receive all posts by user B on his page.
                                        diversity effect of user preference. The discovery may have impor-        Therefore, the social network of Weibo is a directed graph. Repost-
                                        tant implications for diffusion dynamics study and recommendation         ing or forwarding followings’ messages are the primary methods
                                        system design.                                                            of diffusing information in Weibo. One microblog can be reposted
                                                                                                                  multiple times. In a reposted microblog, the usernames of the orig-
                                        KEYWORDS                                                                  inal author and the last re-poster are visible to the current user
                                        Social network, Information polarization, Controlled experiment,          (Figure 1). We focus on two types of user preference (entertainment
                                        Social bot, Text classification, Microblogging                            and sci-tech news) and analyze their effects on news consumption
                                                                                                                  behaviors as well as the local network structure.
                                                                                                                     In order to explore the extent and pathway of preferences af-
                                        “…forms of media favor particular kinds of content and therefore are      fecting news consumption, we propose an experimental approach
                                        capable of taking command of a culture.”                                  by using intelligent social bots [10]. Social bots are generally con-
                                                                                          Neil Postman [1]        sidered to be harmful, although some of them are benign and, in
                                                                                                                  principle, innocuous or even helpful. Therefore, social bots are
                                        ∗ Corresponding   author
Intelligent social bots uncover the link between user preference and diversity of news consumption - arXiv
Composing area
                                                                                           especially entertainment [16]. When the overwhelming entertain-
                                                                                           ment news obscures the scientific, financial and social news that
                                                                                           reaches their home page daily, users are likely to become addicted
                                                                                           to the “entertainment gossip” and consume more gossip-related
                                                                                           news [17]. A high proportion of these users are so-called fanboys
 Weibo type &
 search area

                                                                                           or fangirls who desperately support their favorite celebrities, and
                                                                                           some others are anti-fans of those celebrities [18]. Fans and anti-

                                                                     Originial weibo
                                                                                           fans have similar features and behavior in that they are usually
                                                                                           active in polarized clusters of like-minded people, maintaining and
                                                                                           reinforcing their current standpoints and even attacking each other
                                                                                           [19]. These polarized users could be the important force in making
                                                                                           and spreading ridiculous rumors about celebrities because they
                                                                                           want to look only for news that supports their views and does not
  Weibo Operations

                                                                     Reposted weibo
                                                                                           care about the facts. Besides, most social networks are usually
                                                                                           flooded with malicious attacks and profanities between fans and
                                                                                           anti-fans, making the network environment increasingly worse
                                                                                           [20]. Moreover, these extreme behaviors gradually extend into real
                                                                                           life. For instance, some netizens unearth the real identity of people
                                                                                           who hold opposing views by launching a search and then exposing
                                                                                           personal information (including photos, cell phone number, home
                     Original author    Reposter     Direct source                         address, and other details) to disturb their opponent’s real life [21].
                      of the weibo                   of reposting                          These phenomena reveal the potential interaction between informa-
                                                                                           tion polarization in online social networks and offline behavioral
                            Figure 1: Weibo user interface.                                polarization.
                                                                                               The formation mechanism of polarization and the method of
                                                                                           restraining polarization are the most pressing issues in the field of
                                                                                           information polarization [3, 5, 8, 22]. Many studies have suggested
often the subject of research that needs to be eliminated [11], but                        that personalization, both self-selection, and pre-selection, limit
researchers have yet to recognize their potential value as a powerful                      people's exposure to diverse content and increases polarization
tool in social network analysis [10, 12]. Our social bots can use text                     [23]. Self-selection is the tendency for users to consume content
classification algorithms to simulate the selection behavior of users                      or build new relationships that confirm their existing beliefs and
on content and information sources, thereby simulating the evolu-                          preferences. Pre-selection depends on computer algorithms to per-
tion of personal social networks of users with specific preferences.                       sonalize content for users without any conscious user choice. The
By analyzing the personal social networks of these robots and the                          present research generally agreed that self-selection is the primary
information obtained from them, we can clearly show the route of                           cause of information polarization [8].
user preferences affecting information polarization, thus revealing
the complex interactions behind information polarization.

2 RELATED WORK                                                                             2.2    Controlled experiment on online social
2.1 Polarization of news consumption                                                              networks
The advent and growing popularity of social media and personal-                            At present, research about online social networks relies mainly on
ization services have, to a large extent, changed the way people                           observational methods [7, 8], which means that researchers cannot
consume information, which is regarded as a factor in the fostering                        intervene in the study object but can only process and analyze the
of the polarization of information diffusion and public opinions [8].                      passively acquired data. However, the big open data contains noise
In collective and individual levels, the term “polarization” has two                       that can affect the results [24]. To overcome the shortcomings of
possible meanings. One is that like-minded people form exclusive                           a purely observational study, some researchers have ingeniously
clubs sharing similar opinions and ignoring dissenting views, i.e.,                        adopted the method of natural experiments or quasi-experiments to
opinion polarization [6]; the other is that an individual is exposed                       make comparative analyses and causal inferences from the available
to less diverse content and is limited by a narrow set of informa-                         datasets [25–27]. For example, Phan et al. (2015) analyzed the
tion, i.e., information polarization [7]. For this paper, we focus on                      dynamic development of social relationships after natural disasters
information polarization on social networks and media.                                     in the context of Hurricane Ike in 2008 [26]. Nevertheless, the
   Many scholars have investigated the polarization of political                           uncontrolled approach limits the freedom of research. In contrast,
news consumption [7, 13–15]. For example, Schmidt et al. (2017)                            the controlled experiment is more rigorous and can provide a more
found that users prefer to pay attention to a narrow set of news                           solid foundation for causal analysis [24].
outlets [7]. However, the phenomenon of information polarization                              As online social media and networks have expanded to serve
is present not only in political content but also in other content,                        hundreds of millions of users, and their functions have become
                                                                                       2
more accessible and intelligent, the ability to test sociological the-            The social bot is based on two well-known sociological hypothe-
ories and behavioral phenomena has increased. Application pro-                ses. The first is triadic closure [39]. Triadic closure suggests that
gramming interfaces from operators and innovative third-party                 among three people, A, B, and C, if a link exists between A and B,
toolkits allow researchers to execute novel experimental designs.             and A and C, then there is a high probability of a link between B
For example, Facebook provides interfaces to customize application            and C. The hypothesis reflects the tendency of a friend of a friend
features for particular users and enables the design of controlled            to become a friend and is a useful simplification of reality that
experiments [28]. Amazon Mechanical Turk lets experimenters                   can be used to understand and predict the evolution of social net-
engaging worldwide users in a precisely defined experimental en-              works. Because the relationships embedded in Weibo should be
vironment [29, 30]. These powerful tools are enabling a revolution            modelled by a directed graph, the bots use directed triadic closure
of experimental social science and social network analysis. As a              to expand their followings (Figure 2) [40]. The second hypothesis
result, the controlled experiment approach has begun to reveal                is homophily [6], i.e., people who are similar have a higher chance
the nuanced causal relationships among the design and operation               of becoming connected. In online social networks, for example, an
of the social network and human behavior. For example, recent                 individual creating her new homepage tends to link it to sites that
large-scale experiments have deepened our understanding about                 are close to her interests (i.e., preferences).
information sharing and diffusion, behavior spreading, voting and                 Based on the two hypotheses, the workflow of the bots includes
political mobilization, and cooperation [24, 31–35]. The large scale          five steps (Figure 2). (1) Initially, each bot is assigned 2 or 3 de-
of modern online social networks also enables researchers to con-             fault followings, who mostly post or repost messages consistent
duct experiments involved millions of people to explore the variety           with the preference of the bot. (2) A bot will periodically awaken
of behaviors of diverse users. For example, the experiment con-               from idleness at a uniformly random time interval. When the bot
ducted by Aral and Walker (2012) estimates the heterogeneity of               awakened, it can view the latest messages posted or reposted by
the impact of influence-mediating messages on different types of              its followings. As well known, not all unreaded information can be
consumers [28]; Muchnik et al. (2013) examine in depth whether                exposed to users of social networks. A bot can assess only the latest
opinion change or selective turnout creates the social influence bias         50 messages (i.e., the maximum amount in the first page on Weibo)
that they estimate exists in online ratings [36].                             after waking up. Please note that we exclude the influence of algo-
   Large-scale experiments on social networks, while actively pro-            rithmic ranking and recommendation systems (i.e., pre-selection)
moting innovation of sociology and communication studies, also                on the information exposure by re-ranking all possible messages
bring about some serious problems [37]. For instance, the politi-             according to the descending order of posting time. (3) After viewing
cal voter mobilization experiment conducted by Robert et al. on               the exposed messages, the bot selects only the messages consistent
Facebook was sharply criticized by a considerable number of schol-            with its preferences. The step depends on the FastText text classifi-
ars, who argued that scientific research should not interfere with            cation algorithm, and all classification results from the algorithm
national politics or users'voting behavior [34]. Therefore, inappro-          are further verified by the experimenters to ensure correctness. (4)
priate design on actual large-scale social networks is likely to cause        If there are reposted messages in selected messages, according to
social injustice or infringe on the privacy of users, which violates          directed triadic closure, the bot randomly selects a reposted mes-
the ethics of scientific research and may even be illegal. All these          sage and follows the direct source of the reposted message. Please
issues are due to improper intervention in the behavior of social             note that Weibo limits direct access to the information about the
network users and the collection of their private data.                       followings and followers of a user; thus, bots must find new fol-
   In 2018, the Deep Mind team published a breakthrough study.                lowings through reposting behavior. (5) If the following number
By training the recurrent neural network, researchers simulated               reaches the upper limit, the bot stops running; otherwise, the bot
artificial agents with the function of autonomous mobility and                becomes idle and waits to wake up again. To avoid legal and moral
judgment of spatial location in two-dimensional virtual space [38].           hazards, the bots in the experiment do not produce or repost any
These artificial agents produced a structure similar to that of the           information.
neural network cells of humans. The study provides a theoretical                  Technically, the bot is built using the Selenium WebDriver API
basis for using artificial intelligence technology to simulate human          (http://www.seleniumhq.org/projects/webdriver) to drive Google
social behavior. If artificial agents can be used to replace real users       Chrome Browser as a human to navigate the Weibo website. To
to conduct experiments on large-scale social networks, potential              train the FastText classifier for identifying preferred contents [41],
moral and legal risks can be avoided, and the cost and threshold of           we use the THUCNews dataset (http://thuctc.thunlp.org), which
the experiments can be reduced.                                               contains approximately 740 thousand Chinese news texts from 2005
                                                                              to 2011 and has been marked for 14 classes. We further combine
3 PROPOSED METHOD                                                             and filter the original classes in the dataset to meet our requirement
                                                                              for recognizing Weibo text. All codes and data for building bots are
3.1 Social bot design                                                         freely accessible in GitHub (anonymous for review).
A social bot is a computer algorithm that automatically interacts
with humans on social networks, trying to emulate and possibly al-
ter their behavior [10]. The social bot in our experiment is designed         3.2    Experiment design
to imitate similarity-based relationship formation, which reflects            We are interested in exploring the following overarching question:
the human behavior of selecting information and relationships                 what role do the individual preferences of social media users play
depending on one's preferences (i.e., self-selection).                        in their news consumption? Based on the above intelligent social
                                                                          3
5 Manual checking interface                                                                                       Exposed news 1
                                                                         Start       Main-thread
                                                                                                                        4

                                                                                                                                   2      3
                                                                      Initializing                                             1
                                                                      social bot
                                                                                                                                   1
                                                                 1
                                                                       Reading                                                 Preferred news 2
                                                                     exposed news
                                                                                                                        4

                                                                 2                                                                 2      3
                                                                       Selecting
                                                                       preferred                                               1
                                                                         news
                                                                                                                                   1
                                                                 3     Selecting
                                                                      a reposted
                                                                         news                                                  Reposted news 3
               Sub-thread
                                                                                                                        4
                  5                                                                    Yes
                       Manually                                                                  Waiting for                       2      3
                       checking                                        Empty?                   random time
                                                                                                                               1
                                                                             No
                                                                 4                                                                 1
                         Legal?        Yes       Follow               Adding
                                                  tree            a new following
                                                                                                                                New following 4
                              No
                                                                                                   No
                        Removing                                  Followings are
                      illegal follow                                    full
                                                                             Yes
                                                                                                                              Directed
                                                                                                                               triadic
                                                                         End                                                  closure

                                                          Figure 2: Design of social bot.

bots, we designed a controlled experiment to help us observe the                     is a reversed arc < v j , vi >, is a reciprocal edge (vi , v j ) is defined
news consumption and personal social network evolution of the                        between two nodes. As a result, the reciprocal ratio can be defined
bots with a specific preference. For each bot, we mainly measured                    as:
the following response variables:                                                                                       n
                                                                                                                       Õ   ne (vi )
   V1: The probability that the bot is exposed to the preferred con-                                              r=                                          (1)
                                                                                                                           n
                                                                                                                       i=1 a i
                                                                                                                              (v )
tent. For a given time t, the probability can be quantified by the
preferred content ratio (PCR), which is the ratio of the amount of                   where ne is the number of edges linking to node i, and na is the total
information matching its preference to the total amount of infor-                    number of arcs linking to i. We can also measure the correlation
mation from 0 to t. Since the running of each bot is not completely                  (called reciprocal correlation) between ne /na and na to reflect the
synchronized, we normalize the time scale from 0 to 1, where t = 0                   structural features of personal social networks.
represents the initial state after given the initial user, and t = 1                    For V1 and V2, we can quantify the extent to which the pref-
represents the time when the bot ends the running.                                   erence for particular kinds of content affects the diversity of new
   V2: The distribution of word frequency in the exposed content.                    consumption. For V3 and V4, we can reveal the impact of users’
For a word, the word frequency is defined as the ratio of the number                 preference on their personal social network structure and explore
of messages that contain the word to all exposed messages of a bot.                  the relationship between the social network structure and content
   V3: The attributes of all followings of a bot. These attributes                   consumption.
include the proportion of followings with the same preference (the                      We designed two experimental treatments and evaluated the
preference of followers is judged according to their screen name as                  difference in the above response variables. The two treatments
well as the tags and content of their microblogs) and the number                     were designed to reflect the variance in content preference on
of the followings'followers and followings.                                          the Weibo. The two treatments are an entertainment group (EG)
   V4: The structure of the followings network of a bot (i.e., the                   and a science and technology (sci-tech) group (STG). Bots in EG
personal social network of the bot), including clustering coefficient,               prefer entertainment messages, including celebrity gossip, fashion,
in- and out-degree distribution, in- and out-degree correlation, and                 movies, TV shows, music, and such, and bots in STG prefer sci-
reciprocal ratio (r ) [42, 43]. For a directed arc < vi , v j >, if there            tech news, including nature, science, engineering, technological
                                                                               4
Table 1: Standard for text classification
                                                                                                                **      (A)                      **    (B)
                                                                                                 1                                          **

                                                                                Average ratio
 Standard      Entertainment                Sci-Tech                                            0.8                                    **

 Accept        (1) celebrity gossip,        (1) nature, science, engi-                          0.6
               fashion, movies, TV          neering, technological                              0.4
                                                                                                                                                 *
               shows, and pop music;        advances, digital prod-
                                                                                                0.2
               (2) Explicitly contain-      ucts, and Internet; (2)
               ing the name, account        technical company and                                0
                                                                                                          EG         STG         EG STG1STG2STG3
               or abbreviation of           university.
                                                                                                 1
               entertainers.
                                                                                                                           EG                          (C)
 Common        (1) commercial advertisement; (2) less than 5                                    0.8

                                                                                Average ratio
 reject        Chinese characters.                                                              0.6
 Specific      (1) ACG content (i.e., an-   (1) financial or business
                                                                                                0.4              STG
 reject        imation, comic, and dig-     report of technical or
               ital game); (2) art, lit-    Internet company; (2)                               0.2
               erature, and personal        price of digital products;
                                                                                                 0
               feeling; (3) simple lyrics   (3) military equipment;                                   0        0.2      0.4      0.6             0.8         1
               and lines; (4) personal      (4) daily skills; (5) con-                                                 Normalized time
               leisure activities.          stellations and divina-
                                            tion; (6) weather fore-          Figure 3: Preference drives the polarization of exposed con-
                                            cast; (7) environmental          tent. (A) In the initial state, there is an approximately 50%
                                            conservation; (8) docu-          higher preferred content ratio (PCR) in the entertainment
                                            mentary with irrelevant          (EG) than in the sci-tech (STG) treatment (∗ ∗ p < 0.01, two-
                                            content.                         sided Mann-Whitney U test, n = 34 bots). (B) In the final
                                                                             state, the PCR of the EG is 4.4 times that of the STG treat-
                                                                             ment (∗ ∗ p < 0.01, two-sided Mann-Whitney U Test, n = 34
                                                                             bots). Furthermore, we divided bots in STG into three sub-
advance, digital products, Internet, and so on. Each treatment               groups according to their initial PCR: STG1 (0.6 ≤ PCR,
contained 34 bots who operated on the Weibo platform between                 n 1 = 14), STG2 (0.3 ≤ PCR < 0.6, n 2 = 11), and STG3 (PCR < 0.3,
13 March and 28 September 2018. The initial followings come from             n 3 = 9). Only STG1 and STG2 display a significant difference
a large enough user pool, which contained at least 100 candidates            (∗p < 0.05, Kruskal-Wallis with Mann-Whitney post hoc test).
for each preference. The idle period of all bots was 2 4 hours. The          (C) PCR changes with time. The filling areas indicate the
maximum number of followings for each bot was 120, according to              range between the 25th and 75th percentiles of the ratios in
the Dunbar number.                                                           the two treatments.
   Compared with the conventional method, our design is character-
ized by using intelligent bots as the agents to conduct experiments
in a real online social environment and indirectly observe real
human behavior. These bots act entirely on predefined behavior hy-
potheses and intelligent algorithms, needing manual operation by             4 RESULTS
the experimenters only in the judgment of preference classification.         4.1 User preference drives polarization of
Therefore, our experiment can be regarded as a type of randomized                content
assignment.
                                                                             In social media, users'self-selection behavior of information source
                                                                             and information has been considered the primary mechanism of po-
3.3    Standard for preference classification                                larization. In this experiment, social bots were designed to expand
The most significant feature of the social bots is that they can             personal social networks based on the two hypotheses, which aimed
autonomously identify whether information matches their pref-                to simulate users'self-selection. We aimed to validate whether the
erences, benefiting from automatic and manual text classification.           effect of self-selection is identical under difference preferences.
Therefore, the criteria of text classification are fundamental. We de-          For all 68 social bots, since the initial users have been set as the
fined a content classification standard for the two kinds of content         user who mainly reposts the corresponding preferred content, the
in our experiment (Table 1).                                                 PCR of EG and STG in the initial state can be as high as 76.77%
   The classification criteria shown in Table 1 were used to se-             and 50.0% on average (Figure 3(A)). That is, most of the messages
lect initial followings, prepare the training corpus of the FastText         exposed to both treatments are concentrated in the corresponding
classifier, manually verify, and classify exposed texts. The unified         preference. Even so, preference has a significant impact on infor-
standard can ensure consistency in the judgment of preference and            mation diversity. The PCR of EG was significantly higher than that
the classification of text.                                                  of STG (p < 0.01, two-sided Mann-Whitney U test). This initial
                                                                         5
difference between EG and STG provides a baseline for our further              observed the personal social network of each bot and found that
analysis.                                                                      the bots present different structures according to their preferences.
   Compared to the initial state, we are more concerned about the
                                                                                  4.2.1 The attributes of followings. The followings of a bot are
diversity of messages consumed by social bots after personal social
                                                                               the nodes of its personal social network. Our results showed that
networks have formed (Figure 3(B)). As the PCR of STG in the initial
                                                                               in the EG treatment, the average proportion of nodes with the
state is significantly less than that of EG and has a large variety
                                                                               entertainment preference is 85.32% (up to 92.14% and down to
(Figure 3A), we further divide STG into three subgroups according
                                                                               73.47%) (Figure 5(A)). The high proportion suggests that, with the
to their initial PCR to explain the effect of initial deviations: STG1
                                                                               entertainment preference, users'self-selection can effectively filter
(initial PCR is more than 0.6, 14 bots), STG2 (initial PCR is between
                                                                               the information source, making most of the information sources
0.3 and 0.6, 11 bots), and STG3 (initial PCR is less than 0.3, 9 bots).
                                                                               consistent with their preference and thus resulting in a filter bubble.
The results show that when the personal social network reaches 100
                                                                               In the STG treatment, however, the proportion is only 35.33% on
to 120 followings, EG still has the higher PCR with an average of
                                                                               average (up to 47.86%, down to 23.12%) (Figure 5(B)). Based on
63.78%, while the average PCR of the three STG subgroups is lower
                                                                               the information diffusion mechanism of Weibo and the design of
than 20%. Moreover, Only STG1 and STG2 display a significant
                                                                               social bots, a low proportion indicates that users with the sci-tech
difference (∗p < 0.05, Kruskal-Wallis with Mann-Whitney post hoc
                                                                               preference have a high possibility of following users with other
test). The results suggest that regardless of the initial state, bots in
                                                                               preferences, and users with various preferences could forward the
the STG treatment can receive diverse information rather than sci-
                                                                               sci-tech content. The interaction between multiple preferences
tech content. From the changes of PCR with the growth of personal
                                                                               effectively enhances the diversity of news consumption. For the
social networks, we can also observe the effect of user preference
                                                                               sci-tech preference, as a result, the content-based self-selection
on information polarization (Figure 3(C)). With the social network
                                                                               behavior is inefficient in developing information polarization.
grows, the PCR of EG declines only slightly in the early stage and
                                                                                  Not only the composition of followings but also the features of
remains unchanged after the further increase of followings. In STG,
                                                                               followings in the two treatments are significantly different. Fig-
new followings can continuously reduce the PCR; that is, they can
                                                                               ure 5(C) and (D) show that, compared with STG, EG has 53.78%
bring more nonpreferred content. Therefore, there is a correlation
                                                                               fewer followings (i.e., an indirect information source for bots) but
between the users’ preferences and the polarization of their news
                                                                               23.43% more followers on average (p < 0.01, two-sided Mann-
consumption in Weibo.
                                                                               Whitney U Test). In other words, users with entertainment prefer-
   The effect of preference on content consumption is reflected
                                                                               ence tend to interact with fewer news sources while diffusing infor-
not only on the type of content but also on the semantic level. In
                                                                               mation to a large number of audiences. The “trumpet-like” feature
both EG and STG treatments, the word frequency of the preferred
                                                                               can incite information polarization. Therefore, users’ preferences
content exposed to the social bots presents a power-law distribution
                                                                               not only directly affect news consumption through self-selection
(Figure 4(A)). The power-law distribution (y ∼ x −α ) indicates that
                                                                               but also change their social networks to reinforce this effect.
there are some dominant words with high word frequency in the
preferred content. Moreover, the dominance of a few words is more                 4.2.2 The connection of personal social networks. In addition to
severe in EG than in STG (α = 2.23 for EG and α = 4.74 for STG).               the followings’ attributes, the structure of the personal social net-
In the EG treatment, the maximum frequency of a word can be up                 work is also modified by the user preference. By roughly comparing
to 0.4 on average, which means that the same word can be found in              the in- and out-degree distributions of EG and STG's personal so-
40% of messages on average. However, in the STG treatment, the                 cial networks, we find that both present power-law distributions
maximum word frequency is no more than 0.15 (Figure 4(B)). Even                (α = 1.766 for in-degrees and α = 1.847 for out-degrees) and are
among the top 10 words of EG, the first and second words are more              almost identical (Figure 6(A) and (B)). Furthermore, the visualiza-
dominant than other words. By analyzing the top 10 high-frequency              tion of network evolution also shows that EG and STG have similar
words of the two treatments, we find that the high-frequency words             personal social networks (Figure 7). However, additional structural
of EG are usually the names of certain entertainment celebrities               features reveal the opposite result.
and related specific words, while the high-frequency words of STG                 Figure 6(C) shows that the average clustering coefficient of STG
are some common words, such as China, America, technology,                     networks is 35% higher than that of EG networks (p < 0.01, two-
market, and company (Figure 4(C)). From this, it can be seen that              sided Mann-Whitney U Test). Given the low level of information
the messages received by bots preferring entertainment content not             polarization in STG, this is an unexpected result, as a dense com-
only concentrate on a single type but also show a univocal trend in            munity spreads diverse information. This counterintuitive phenom-
semantics.                                                                     enon can be attributed to differences in the connection pattern of
                                                                               personal social networks. Figure 6(D) shows that there is a high
                                                                               positive correlation (0.6) between the in- and out-degrees of STG
                                                                               networks, while the degree correlation of EG networks is very weak
4.2    The personal social network                                             (0.2). This correlation implies a difference between the two types of
The dissemination mechanism of Weibo makes the followings of                   networks on reciprocal edges. We found that an average of 60% of
a bot become the primary and direct information sources for the                the arcs in STG networks belong to the reciprocal edge, while the
bot. Therefore, the personal social network of a social bot (i.e., the         proportion is only 40% in EG (Figure 6(E)). Besides, in STG, the de-
network consisting of its followings) is its media environment and             gree of reciprocal edge presented a weak positive correlation (0.25)
reflects a local structure to diffuse specific types of information. We        with nodes’ total degree; that is, the higher-degree nodes might
                                                                           6
100                                             (A)
                                                                                              EG03 top10 words                            (C)
                                                                                              01) Wu, Yifan (name)
                                      10-2

                       Distribution
                                                                                              02) Mr                       音乐
                                                        EG                                    03) Hip hop
                                                        y~x-2.23                              04) Music
                                      10-4                                                    05) Lu, Han (name)     鹿晗                 说唱
                                                        STG                                   06) China
                                                        y~x-4.74                              07) SKR
                                                                                                                     电影
                                                                                                                                吴亦凡
                                                                                              08) Diss
                                      10-6                                                    09) Internet
                                                                                              10) Movie
                                             10-4       10-3        10-2       10-1     0.8                               中国
                                                            Word rank                                                           网络
                                       0.8
                                                                                      (B)               科技    智能
                                                                                                                         STG07 top10 words
                                       0.6                                                             手机            公司
                          Frequency

                                                                                                            中国                04) Sci-Tech
                                                                                                  美元                          05) Market
                                       0.4                                                           美国    市场 全球
                                                                                                                              06) Mobile phone
                                                                                                        技术                    07) US dollar
                                                                                                              01) China       08) Global
                                       0.2                                                                    02) Corporation 09) Technology
                                                                                                              03) US          10) Intelligence
                                        0
                                              1     2   3   4   5    6     7   8   9 10

Figure 4: Semantic analysis of exposed contents. (A) Inverse cumulative distribution of word frequency (log-log plot). The EG
treatment has a longer tail than the STG treatment; i.e., the portion of the EG distribution has more higher-frequency words.
(B) Box plots of the frequency of top 10 words in both the EG and STG treatments. (C) Demonstration of top 10 words in
exposing messages from EG 03 bot and STG 07 bot. The size of the bubble indicates the word frequency.

have more reciprocal edges (Figure 6(F)). In EG, the degree of the                                 Moreover, different news consumption patterns are accompanied by
reciprocal edge is negatively correlated with the total degree (-0.18);                            different compositions and structures of personal social networks.
that is, the higher-degree nodes prefer the one-way relationships
in Weibo (Figure 6(F)).
   The visualization analysis also shows the same results. If all
                                                                                                   5    DISCUSSION
nonreciprocal arcs are removed, we find that the STG networks
still have a high connection density, and the high-degree nodes                                    In this paper, we create a new controlled experiment approach
are still maintained in the centrality, while EG networks are the                                  to reveal the link between user preferences and patterns of news
opposite, with the sparse connection, and the high-degree nodes                                    consumption on a Chinese microblogging platform (Weibo). This
are marginalized (Figure 7). Based on the above results, an EG                                     method can be seen as a real-time user behavioral simulation on
network tends to be a unidirectional star-like structure with a few                                real-world social networks using the computer and artificial in-
high-degree nodes, while an STG network tends to be a bidirec-                                     telligence technology. (1) The method regards real-world social
tional clustering structure. The unidirectional star-like structure                                networks as an objective media environment and records and ana-
means selecting one or several central nodes to play the role of                                   lyzes the feedback of the environment to specific user behaviors. (2)
broadcasters to send information to all other nodes with few inter-                                This method uses artificial intelligence technologies (e.g., natural
actions between them. However, the diffusion path of this structure                                language processing) to simulate specific user behaviors instead
is mainly dependent on the selected central node. If the central                                   of employing volunteers or collecting users’ private data. The
nodes in the star-like structure produce only a narrow set of news,                                approach not only improves the freedom of the experimenters in
the others will be limited to a lower diversity of content. This                                   designing and controlling the experiment but also avoids the ethical
selection effect of central nodes tends to enlarge the polarization                                risk of violating users’ privacy. (3) Compared with the experimental
of the content [44]. In the bidirectional clustering structure, all                                method of directly altering the real-world social network [34], this
nodes are placed in a clustering group, and most connections are                                   method gives the experimenter more flexibility and allows more
bidirectional. Those nodes can efficiently exchange information                                    researchers to conduct their own experiments without internal
and realize a complementary effect with their friends with distinct                                assistance from social network operators. (4) Compared with the
preferences. The complementary effect can promote the diversity                                    experimental method based on artificial social networks [31, 32],
of information [44].                                                                               this method can work directly on real-world social networks and
   Although all social bots select potential followings based on                                   can control user behaviors more directly with lower cost. How-
preferred content to ensure that these users post or repost the                                    ever, limited by the current technical conditions, this method can
content they like, whether the self-selection behavior can narrow                                  intelligently simulate only relatively simple user behaviors. How
users’ news consumption is dependent on the users’ preferences.                                    to simulate complex user behaviors (especially initiative feedback
                                                                                                   behaviors such as comments and chats) is still a challenging task.
                                                                                              7
(A)                                                (B)                                                         (A)                                                            (B)
                                               92%                                                                            10-1                                                                 10-1

                                                                                                           Distribution

                                                                                                                                                                         Distribution
                                                                                     35%
                                                     73%
                                                                                                                              10-2                                                                 10-2
                                                                           23%
                                                                                                                                                  EG
                                       85%
                                                                                                                                                  STG
                                                                                                                              10      -3
                                                                                                                                                                                                   10-3
                                                                                                                                                  y~x-1.77                                                       y~x-1.85
                                                                                           48%

                                                                                106                                                    100            101          102                                100            101          102
                     1000                                                  3                                                                 In-degree                                                          Out-degree
                            (C)          **                                    (D)                                                   0.5                                                            0.8
Average followings

                                                                                           **
                                                       Average followers

                      800                                                                                                                (C)     **                                                       (D)

                                                                                                                                                                              In-Out correlation
                                                                           2                                                         0.4                                                                                    **
                      600                                                                                                                                                                           0.6

                                                                                                              Clustering
                                                                                                                                     0.3
                      400                                                                                                                                                                           0.4
                                                                           1                                                         0.2                                                                        **
                      200
                                                                                                                                                                                                    0.2
                                                                                                                                     0.1
                        0                                                  0
                                  EG          STG                                    EG          STG                                  0                                                              0
                                                                                                                                                 EG          STG                                                EG          STG
Figure 5: Demographic features of the followings of bots.
                                                                                                                                     0.8                                                            0.3

                                                                                                                                                                          Reciprocal correlation
(A) and (B) Percentage of followings with the same prefer-                                                                                 (E)                                                            (F)
                                                                                                                                                       **                                                                   **
ence with bots in the EG and STG treatments. The white and                                                    Reciprocal arc ratio                                                                  0.2
colored dotted lines indicate the minimum and maximum                                                                                0.6
                                                                                                                                                                                                    0.1
percentages, respectively. (C) The average number of follow-
                                                                                                                                     0.4                                                             0
ings of followings (∗ ∗ p < 0.01, two-sided Mann-Whitney U
Test, 2,675 followings in EG and 1,622 followings in STG). (D)                                                                                                                                     -0.1
                                                                                                                                     0.2
The average number of followers of followings (∗ ∗ p < 0.01,                                                                                                                                       -0.2
two-sided Mann-Whitney U Test, 2,675 followings in EG and                                                                                                                                                       **
                                                                                                                                      0                                                            -0.3
1,622 followings in STG). Due to the high variance, we do not                                                                                    EG          STG                                                EG          STG
show error lines in this plot.
                                                                                                            Figure 6: Structural features of personal social networks. (A)
                                                                                                            Mean of inverse cumulative distribution of in-degree (log-
                                                                                                            log plot). (B) Mean of inverse cumulative distribution of
   Taking advantage of our experimental method, we compare the                                              out-degree (log-log plot). (C) Personal social networks in
news consumption of users with different preferences. Previous                                              STG have a higher clustering coefficient than networks in
studies on the polarization mechanism focus on the effect of users’                                         EG (∗ ∗ p < 0.01, two-sided Mann-Whitney U Test, n = 34
self-selection behaviors [23]. First, a large number of studies focus                                       networks). (D) Compared with EG, nodes in STG present a
on the dissemination of different opinions in response to particular                                        stronger correlation between in- and out-degrees (∗∗p < 0.01,
content, such as political news [7, 13–15]. In this case, users are                                         Pearson correlation coefficient, 54,823 nodes for EG, 51,838
actually in the same media environment, and the differences in                                              nodes for STG). (E) Personal social networks in STG obtain
content preferences can be ignored. Second, the rise of pre-selection                                       more reciprocal edges than networks in EG (∗ ∗p < 0.01, two-
technologies (e.g., recommendation systems) makes people more                                               sided Mann-Whitney U Test, n = 34 networks). (F) Nodes
attentive to the comparison of pre-selection and self-selection than                                        in EG show a weak negative correlation between reciprocal
to the impact of the media environment [8]. However, our research                                           edge degree (> 0) and total degree. In contrast, nodes in STG
shows that user preferences might be the primary factor affecting                                           show a positive reciprocal degree correlation (∗ ∗ p < 0.01,
the diversity of news consumption. Distinct preferences can lead to                                         Pearson correlation coefficient, 3,546 nodes for EG, 3,368
entirely different news consumption patterns even with the same                                             nodes for STG).
self-selection behavior. Such results deepen our understanding
of the mechanism of information polarization. Furthermore, in
this experiment, we make the ideal assumption that each user has                                            bot, the messages that a bot can receive is limited to the follow-
only a single preference. However, people usually have multiple                                             ings'posts and reposts. Therefore, information diversity depends
preferences. How the interaction of multiple preferences shapes                                             on followings'self-selection and the structure of personal social
the diversity of news consumption remains an open question.                                                 networks. The impact of network structure on information dif-
   Our results also show that user preference not only affects in-                                          fusion is well known [42, 45, 46], but people often overlook the
formation diversity but also shapes personal social networks. De-                                           role of content in it. Until now, only a few works have noticed
pending on the mechanism of Weibo and the design of a social                                                the link between content and structure [7, 47]. In communication
                                                                                                       8
EG05

   STG02

                 Initialized (0%)               20%                         50%                   End (100%)            Reciprocal edges only

Figure 7: Visualization of the evolution of personal social networks. In this demonstration, Social bot 05 in EG (EG05) and bot
02 in STG (STG02) display a similar growth process from initialized two followings to the end of running. However, STG02
obtains more reciprocal edges than EG05. The triangles represent the nodes with high in-degree, that is, the corresponding
users have a large number of followers. The visualization is based on the radial layout with in-degree centrality; thus the node
with higher in-degree is closer to the center of the plot.

studies, the media ecology theory has revealed the connection be-               Our work is a preliminary experimental analysis of the relation-
tween the form and structure of media and content [48]. They even            ship between user preference and news consumption, so there are
think that “the medium is the message.” [49]. Our results show               still many limitations. First, the impact of initial users requires
that even in the same media, different content should spread in              further evaluation. In the experiment, it was difficult for us to find
distinct sub-structures. For example, we found that personal so-             users who post merely sci-tech content (RCP ≥ 0.8) as the initial fol-
cial networks of users with entertainment preferences tend to be a           lowings of bots. Although we find that the weak selection intensity
unidirectional star-like structure, which is more powerful for the           does not affect the result, the effect of the initial values still requires
broadcasting of narrow content. It can be seen that user preference          a more detailed analysis. Second, limited by the accuracy of the text
drives the coevolution of content consumption and personal social            classification algorithm, the robot currently only selects potential
network structure, and the coevolution strengthens information               followings through the received real-time messages. The selection
polarization.                                                                is relatively loose. If the robot can examine all the information
   Our work provides valuable insights into the theoretical analysis         about potential followers (historical posts, tags, avatars, etc.), the
of diffusion dynamics. We found that, even in the same network,              effect of rigorous selection requires further study. Third, the two
information about different contents should be propagated through            types of preferences are not enough to explain all the problems.
different sub-structures. This phenomenon has led us to re-examine           The study of more preferences (e.g., sport, finance, politics, etc.)
the results of previous studies based on a unified macroscopic struc-        will help us to understand the complexity of the triad of preference,
ture, such as small-world networks or scale-free networks [42].              content, and structure in news consumption.
Fortunately, researchers have begun to use multilayer network                   In sum, we have introduced a new experimental approach using
models to describe the differences in sub-structures. However,               social bots and revealed the collective behavior behind information
the multilayer network model for content differences and the re-             polarization, that is, a group of users with the same preference can
lated propagation mechanisms remain to be studied. Furthermore,              form a media environment facilitating polarization by specialized
our work is also instructive for the design of the recommendation            structure and selection behaviors.
system [50]. Users with different preferences need different rec-
ommendation strategies. For example, our results show that users
preferring entertainment content need more diverse recommenda-
tions to avoid polarization, while users preferring sci-tech content         ACKNOWLEDGMENTS
need more relevant content to maintain their activity in social net-         This work was supported by the National Natural Science Foun-
works [51]. As a result, the recommendation system not only needs            dation of China (Grant Nos. 71303217, 31370354 and 31270377)
to pay attention to individual preferences and behaviors but also            and the Zhejiang Provincial Natural Science Foundation of China
needs to understand the collective behavior of the user groups with          (Grant No. LY17G030030, LGF18D010001, LGF18D010002).
the same preferences.                                                          We thank Professor Jie Chang and Professor Ying Ge (Zhejiang
                                                                             University) for their guidance on the diversity theory.
                                                                        9
We thank Aizhu Liu, Chenyi Fang, Conger Yuan, Fan Li, Hao                                         April 2014.
Li, Hao Wu, Haochen Hou, Hengji Wang, Jian Zhou, Jiaye Zhang,                                   [25] Yubo Chen, Qi Wang, and Jinhong Xie. Online social interactions: A natural
                                                                                                     experiment on word of mouth versus observational learning. Journal of Marketing
Jinmeng Wang, Junhao Xu, Lingjian Jin, Longzhong Lu, Lu Chen,                                        Research, 48(2):238–254, April 2011.
Luchen Zhang, Qiuhai Zheng, Qiuya Ji, Renyuan Yao, Ruonan                                       [26] Tuan Q. Phan and Edoardo M. Airoldi. A natural experiment of social net-
                                                                                                     work formation and dynamics. Proceedings of the National Academy of Sciences,
Zhang, Shang Gao, Shicong Han, Songyi Huang, Ting Xu, Wei Fang,                                      112(21):6595–6600, May 2015.
Wei Zhang, Xingfan Zhang, Yijing Wang, Yingjie Feng, Yinting                                    [27] Soroush Vosoughi, Deb Roy, and Sinan Aral. The spread of true and false news
Chen, Yiqi Ning, Yujie Bao, Yuying Zhou, Zheyu Li, Ziyu Liu for                                      online. Science, 359(6380):1146–1151, March 2018.
                                                                                                [28] Sinan Aral and Dylan Walker. Identifying influential and susceptible members
their contribution to the manual text classification.                                                of social networks. Science, 337(6092):337–341, July 2012.
   We thank Haidan Yang for her introduction to the theory of                                   [29] David G. Rand and Martin A. Nowak. The evolution of antisocial punishment in
communication.                                                                                       optional public goods games. Nature Communications, 2:434, August 2011.
                                                                                                [30] Winter Mason and Duncan J. Watts. Collaborative learning in networks. Pro-
                                                                                                     ceedings of the National Academy of Sciences, 109(3):764–769, January 2012.
                                                                                                [31] Damon Centola. The spread of behavior in an online social network experiment.
REFERENCES                                                                                           Science, 329(5996):1194–1197, September 2010.
 [1] Neil Postman. Amusing ourselves to death: public discourse in the age of show              [32] Damon Centola. An experimental study of homophily in the adoption of health
     business. Penguin, 2006.                                                                        behavior. Science, 334(6060):1269–1272, December 2011.
 [2] Nic Newman, David Levy, and Rasmus Kleis Nielsen. Reuters institute digital                [33] David G. Rand, Samuel Arbesman, and Nicholas A. Christakis. Dynamic social
     news report 2015. SSRN Scholarly Paper ID 2619576, Social Science Research                      networks promote cooperation in experiments with humans. Proceedings of the
     Network, Rochester, NY, June 2015.                                                              National Academy of Sciences, 108(48):19193–19198, November 2011.
 [3] Eli Pariser. The filter Bubble: How the new personalized web is changing what we           [34] Robert M. Bond, Adam D. I. Kramer, Cameron Marlow, Christopher J. Fariss,
     read and how we think. Penguin, May 2011.                                                       Jaime E. Settle, James H. Fowler, and Jason J. Jones. A 61-million-person experi-
 [4] Seth Flaxman, Sharad Goel, and Justin M. Rao. Filter bubbles, echo chambers,                    ment in social influence and political mobilization. Nature, 489(7415):295–298,
     and online news consumption. Public Opinion Quarterly, 80(S1):298–320, January                  September 2012.
     2016.                                                                                      [35] Adam D. I. Kramer, Jamie E. Guillory, and Jeffrey T. Hancock. Experimental evi-
 [5] F.J. Zuiderveen Borgesius, D. Trilling, J. Moller, B. Bodo, C.H. de Vreese, and                 dence of massive-scale emotional contagion through social networks. Proceedings
     N. Helberger. Should we worry about filter bubbles? Internet Policy Review, 5(1),               of the National Academy of Sciences, page 201320040, May 2014.
     March 2016.                                                                                [36] Lev Muchnik, Sinan Aral, and Sean J. Taylor. Social influence bias: A randomized
 [6] P. Dandekar, A. Goel, and D. T. Lee. Biased assimilation, homophily, and the                    experiment. Science, 341(6146):647–651, August 2013.
     dynamics of polarization. Proceedings of the National Academy of Sciences,                 [37] Ulf-Dietrich Reips. Standards for internet-based experimenting. Experimental
     110(15):5791–5796, April 2013.                                                                  Psychology, 49(4):243–256, October 2002.
 [7] Ana Luca Schmidt, Fabiana Zollo, Michela Del Vicario, Alessandro Bessi, Antonio            [38] Andrea Banino, Caswell Barry, Benigno Uria, Charles Blundell, Timothy Lillicrap,
     Scala, Guido Caldarelli, H. Eugene Stanley, and Walter Quattrociocchi. Anatomy                  Piotr Mirowski, Alexander Pritzel, Martin J. Chadwick, Thomas Degris, Joseph
     of news consumption on Facebook. Proceedings of the National Academy of                         Modayil, Greg Wayne, Hubert Soyer, Fabio Viola, Brian Zhang, Ross Goroshin,
     Sciences, 114(12):3035–3039, March 2017.                                                        Neil Rabinowitz, Razvan Pascanu, Charlie Beattie, Stig Petersen, Amir Sadik,
 [8] Eytan Bakshy, Solomon Messing, and Lada A. Adamic. Exposure to ideologically                    Stephen Gaffney, Helen King, Koray Kavukcuoglu, Demis Hassabis, Raia Hadsell,
     diverse news and opinion on Facebook. Science, 348(6239):1130–1132, June 2015.                  and Dharshan Kumaran. Vector-based navigation using grid-like representations
 [9] Weibo, October 2018. https://en.wikipedia.org/wiki/Sina Weibo.                                  in artificial agents. Nature, 557(7705):429–433, May 2018.
[10] Emilio Ferrara, Onur Varol, Clayton Davis, Filippo Menczer, and Alessandro                 [39] Ginestra Bianconi, Richard K. Darst, Jacopo Iacovacci, and Santo Fortunato.
     Flammini. The rise of social bots. Communications of the ACM, 59(7):96–104,                     Triadic closure as a basic generating mechanism of communities in complex
     June 2016.                                                                                      networks. Physical Review E, 90(4):042806, October 2014.
[11] Abigail Paradise, Rami Puzis, and Asaf Shabtai. Anti-reconnaissance tools:                 [40] Daniel Mauricio Romero and Jon M Kleinberg. The directed closure process
     Detecting targeted socialbots. IEEE Internet Computing, 18(5):11–19, 2014.                      in hybrid social-information networks, with an analysis of link formation on
[12] Norah Abokhodair, Daisy Yoo, and David W McDonald. Dissecting a social                          twitter. In Proceedings of the Fourth International AAAI Conference on Weblogs
     botnet: Growth, content and influence in twitter. In Proceedings of the 18th ACM                and Social Media, 2010.
     Conference on Computer Supported Cooperative Work & Social Computing, pages                [41] Armand Joulin, Edouard Grave, Piotr Bojanowski, and Tomas Mikolov. Bag of
     839–851. ACM, 2015.                                                                             tricks for efficient text classification. arXiv preprint arXiv:1607.01759, 2016.
[13] Michael D. Conover, Jacob Ratkiewicz, Matthew Francisco, Bruno Goncalves,                  [42] Stefano Boccaletti, Vito Latora, Yamir Moreno, Martin Chavez, and D-U Hwang.
     Filippo Menczer, and Alessandro Flammini. Political polarization on Twitter. In                 Complex networks: structure and dynamics. Physics reports, 424(4-5):175–308,
     Fifth International AAAI Conference on Weblogs and Social Media, July 2011.                     2006.
[14] Markus Prior. Media and political polarization. Annual Review of Political Science,        [43] Giorgio Fagiolo. Clustering in complex directed networks. Physical Review E,
     16(1):101–127, 2013.                                                                            76(2):026107, 2007.
[15] Vidya Narayanan, Vlad Barash, John Kelly, Bence Kollanyi, Lisa-Maria Neudert,              [44] Michel Loreau and Andy Hector. Partitioning selection and complementarity in
     and Philip N. Howard. Polarization, partisanship and junk news consumption                      biodiversity experiments. Nature, 412(6842):72, 2001.
     over social media in the US. arXiv:1803.01845 [cs], March 2018. arXiv: 1803.01845.         [45] Yong Min, Xiaogang Jin, Ying Ge, and Jie Chang. The role of community mix-
[16] James G. Webster. Beneath the veneer of fragmentation: television audience                      ing styles in shaping epidemic behaviors in weighted networks. PLOS ONE,
     polarization in a multichannel world. Journal of Communication, 55(2):366–382,                  8(2):e57100, February 2013. WOS:000315184200214.
     June 2005.                                                                                 [46] Xiao-gang Jin and Yong Min. Modeling dual-scale epidemic dynamics on complex
[17] Su Jung Kim and James G. Webster. The impact of a multichannel environment                      networks with reaction diffusion processes. Journal of Zhejiang University
     on television news viewing: a longitudinal study of news audience polarization                  SCIENCE C, 15(4):265–274, April 2014.
     in South Korea. International Journal of Communication, 6(0):19, April 2012.               [47] Chun-Yuen Teng, Liuling Gong, Avishay Livne Eecs, Celso Brunetti, and Lada
[18] Mark Duffett. Understanding fandom: An introduction to the study of media fan                   Adamic. Coevolution of network structure and content. In Proceedings of the 4th
     culture. Bloomsbury Academic, New York, 1 edition edition, August 2013.                         Annual ACM Web Science Conference, WebSci ’12, pages 288–297, New York, NY,
[19] Lucy Bennett. Researching online fandom. Cinema Journal, 52(4):129–134, 2013.                   USA, 2012. ACM.
[20] Huishan Huang. Exo “vs” tfboys: Causal analysis of network wars between idol               [48] Carlos A. Scolari. Media ecology: exploring the metaphor to expand the theory.
     fans from an perspective of psychology. Journal of Henan Institute of Education,                Communication Theory, 22(2):204–225, May 2012.
     25(1):48–52, 2016.                                                                         [49] Marshall McLuhan and Quentin Fiore. The medium is the message. New York,
[21] Dong Han. Search boundaries: human flesh search, privacy law, and internet                      123:126–128, 1967.
     regulation in china. Asian Journal of Communication, 28(4):434–447, 2018.                  [50] Tao Zhou, Zoltan Kuscsik, Jian-Guo Liu, Matuvs Medo, Joseph Rushton Wake-
[22] Natalie Jomini Stroud. Polarization and partisan selective exposure. Journal of                 ling, and Yi-Cheng Zhang. Solving the apparent diversity-accuracy dilemma
     Communication, 60(3):556–576, September 2010.                                                   of recommender systems. Proceedings of the National Academy of Sciences,
[23] Kota Kakiuchi, Fujio Toriumi, Masanori Takano, Kazuya Wada, and Ichiro                          107(10):4511–4515, 2010.
     Fukuda. Influence of selective exposure to viewing contents diversity. Arxiv,              [51] Xiaogang Jin, Cheng Jin, Jiaxuan Huang, and Yong Min. Coupling effect of
     page 1807.08744, July 2018.                                                                     nodes popularity and similarity on social network persistence. Scientific Reports,
[24] Sinan Aral and Dylan Walker. Tie strength, embeddedness, and social influence:                  7:42956, February 2017.
     A large-scale networked experiment. Management Science, 60(6):1352–1370,
                                                                                           10
You can also read