Physica A 493 (2018) 267–275

Contents lists available at ScienceDirect

Physica A
journal homepage: www.elsevier.com/locate/physa

Structural diversity effect on hashtag adoption in Twitter

Aihua Zhang a, *, Mingxing Zheng a , Bowen Pang b
a
Economics & Management School, Beijing University of Posts and Telecommunications, 100876, PR China
b
International school, Beijing University of Posts and Telecommunications, 100876, PR China

highlights

• A network containing the following relationship of Twitter users is built.

• The GN algorithm is used to analyze the structural diversity of user’s ego network.
• The structural diversity effect on user’s behavior in Twitter is revealed.
• The effect of the number of influencers, the strength of relationship is examined.

article info a b s t r a c t
Article history: With online social network developing rapidly these years, user’ behavior in online social
Received 12 November 2013 network has attracted a lot of attentions to it. In this paper, we study Twitter user’s behavior
Received in revised form 4 December 2014 of hashtag adoption from the perspective of social contagion and focus on ‘‘structure
Available online 28 October 2017
diversity’’ effect on individual’s behavior in Twitter. We achieve data through Twitter’s
Keywords:
API by crawling and build a users’ network to carry on empirical research. The Girvan–
Social network Newman (G–N) algorithm is used to analyze the structural diversity of user’s ego network,
Twitter and Logistic regression model is adopted to examine the hypothesis. The findings of our
Structural diversity empirical study indicate that user’ behavior in online social network is indeed influenced
Hashtag adoption by his friends and his decision is significantly affected by the number of groups that these
friends belong to, which we call structural diversity.
© 2017 Elsevier B.V. All rights reserved.

1. Introduction

In recent years, online social networks, such as Facebook and Twitter, have been developing rapidly. According to
eMarketers’ report, increasing users log on to online social networks at least once a month, occupying 63.4% of the world’s
Internet users and 20.4% of the world population already as of 2012. Through these online social networks, users can share
news, videos, music and photos, discuss hot topics, and join discussion groups to maintain and develop interpersonal
relationships. Impacting on human social life increasingly, online social network has become an inseparable part many
people’s life.
Many types of social behavior can be thought of as contagious: participation in a social party, smoking, voting, migration,
the spread of innovations, fashions, rumors and so on [1–4]. With the popularity of online social networks, the behaviors of
users have gained wide attention from many scholars. Recently, many researchers have proven that online social network
behaviors, such as forwarding messages, linking to articles, joining groups, adopting hashtags or becoming fans of pages,
are influenced by friends whom users have followed in online social networks [5–8]. Users are likely to perform the same
behaviors that their friends have already exhibited.

* Corresponding author.
E-mail address: aihuazhang@bupt.edu.cn (A. Zhang).

https://doi.org/10.1016/j.physa.2017.09.075
0378-4371/© 2017 Elsevier B.V. All rights reserved.
268 A. Zhang et al. / Physica A 493 (2018) 267–275

The researches on users’ online social network behavior have thus far shown some factors that may affect the decision of
the user, such as the number of friends who have performed the same behavior, the strength of relationships between users
and their friends and the popularity of their friends. However, most studies of online social network structures concentrating
on their influence on diffusion effects, instead of their influence on individual decisions and behaviors. What attracts us is
that, Ugander et al. studied the influence of friends on people’s decisions to join Facebook, and they found that a person’s
decision to accept an invitation to join Facebook is tightly controlled by the number of connected components in his contact
neighborhood, which is called ‘‘structural diversity’’ in social contagion [9]. However, what influences users offline is not
always what influences users online. So far, few studies on user behavior are from the perspective of friends’ structure. As
a result, whether this structural diversity of users’ ego networks will affect their behavior in online social networks is still
unknown, which is the purpose of our study. In the next page, we try to explore the ‘‘structural diversity’’ effect on the
adoption of hashtags in Twitter.
This paper begins with a review of related work (Section 2), followed by descriptions of the study design (Section 3) and
experiment setup (Section 4). We then present our results (Section 5) and discuss our findings (Section 6).

2. Related work

A hashtag is a word or a phrase prefixed with the symbol # and is a form of metadata tag, which was invented to group
messages when users search for the hashtag and get the set of messages that contain it. However, hashtags are mostly used
as unmoderated ad-hoc discussion forums in Twitter; any combination of characters led by a hash sign is a hashtag, and any
hashtag, if promoted by enough individuals, can ‘‘trend’’ and attract more individual users to discussion. It is a popular way
to use hashtags to create live, searchable dialogs at conferences and other live events. People also use them to track shared
interests, such as #skydiving, or designate members of a group that are having a conversation when they develop a group
chat. Usually, people use hashtags in tweets to represent that they are interested in and talking about a specific topic. In this
way, a hashtag enables users to identify and participate in online chats designated by the tag [8].
Owing to its availableness, data of tweet hashtags was usually be used to study various problems. For example, Romero,
D.M., B. Meeder and J. Kleinberg studied different types of information spread on-line by analyzing the ways hashtags spread
on a network defined by the interactions among Twitter users, finding significant variation in the ways that widely used
hashtags on different topics spread [6]. Furthermore, Hsia-Ching Chang used hashtags to study the Diffusion of Innovation
Theory [10]. The use of a hashtag is a typical online social network behavior, and whether users participate in the discussion
of a specific topic may be affected by their friends’ participation in that topic thus our study takes a user’s adoption of a
specific hashtag as a research object.
As to the factors the affect the behavior of online social networks, more and more scholars are participating in the study
since 2006. Backstrom L. et al. found that an online social network user’s decision to join a discussion group is influenced
not only by the number of friends who have been members of the group but also by the relationships between the friends
in the group and the level of activeness of the group [5]. Sen S. et al. studied the factors affecting users’ selection of tags, and
they found that users’ selections were dependent on their past behavior; additionally, the choices of a user’s friends also had
a great impact on the user’s selection [7]. Romero D.M., B. Meeder and J. Kleinberg studied some widely used hashtags on
different topics covered on Twitter, and their results showed that variation in the diffusion of different topics was attributable
not only to differences in ‘‘stickiness’’, the probability of adoption based on one or more exposures, but also to a quantity that
could be viewed as a type of ‘‘persistence’’ - the relative extent to which repeated exposure to a hashtag continues to have
significant marginal effects [6]. Macskassy S A and M. Michelson studied the reason that made people spread information in
tweets through the use of a retweet. They found that models that took homophily, or similarity, into account fitted the
observed retweet behaviors much better than other general models that did not include homophily [11]. Yang L. et al.
presented a formal empirical analysis to test how the dual role of hashtags in Twitter affects hashtag adoption [8], which
concluded that hashtags served two roles: a tag of content and a symbol of membership of a community. For the role of a tag
of content, what primarily influences users’ selection would be the popularity of the tag and the relevance to the content and
(3) the closeness to the user’s personal preferences. For the role of a symbol of membership of a community, they extracted
these factors: the prestige of community members designated by a hashtag and social influence of users’ friends. The results
of a correlation analysis, a regression analysis, and a prediction analysis all suggested that a hashtag served as both a tag of
content and a symbol of membership of a community. Besides, Harrigan N., P. Achananuparp and E.P. Lim explored the effect
of local social structure on the retweeting behavior of a community of Twitter users over a five month period. They found
that community structures, particularly reciprocal ties, substantially increased social contagion [12]. Bakshy E. et al. studied
URL sharing on Facebook and found that subjects who were exposed to signals about friends’ sharing behavior were several
times more likely to share that same information and to share sooner than those who were not exposed [13].
Though the studies mentioned above, we can roughly divided the factors influencing users’ behaviors into two categories.
One is personal factors, including personal interest and personal behavior frequency, and the other one is social contagion
factors, including the number of friends who have the same behavior, the popularity of these friends, and the strength of
relationships between the user and these friends.
 A. Zhang et al. / Physica A 493 (2018) 267–275 269

Fig. 1. An individual’s friends may come from different aspects of one’s lives.

Fig. 2. Node A’s ego network.

3. Study design

Our study’s purpose is to explore the ‘‘structural diversity’’ effect on users’ adoption of Hashtag in Twitter. The ‘‘structural
diversity’’ effect was first found by Ugander, Backstrom, Marlow and Kleinberg [9]. Easily affected by their neighbors or
friends, People are always trying to have the same behavior as their neighbors/friends do. It is widely accepted that the
probability that an individual is affected by contagion grows monotonically with the number of his or her neighbors/friends
who have the same behavior. We simply call these friends ‘‘influencers’’. ‘‘Structural diversity’’ shows that for social
contagion, not only is the number of one’s influencers the crucial parameter, but also the number of distinct social contexts
that these influencers represent has an impact on it. Generally speaking, an individual’s friends may come from different
aspects of one’s lives and can be divided into several groups that reflect distinct social contexts within an individual’s life or
life history [14–16]. In Fig. 1, the person in the center has some friends, which can clearly be divided into two groups. Some
of them may be relatives of him, and some may be classmates or colleagues. According to ‘‘structural diversity’’, if influencers
are all in the same group, their effect on social contagion will not be as large as when they are from different groups. We use
community structure to reflect the different social context between an individual’s friends. Community structure divides
network naturally into groups of nodes with dense connections internally and sparser connections between groups, which
can represent different community groups in real networks.
We focus on ‘‘structural diversity’’ effect in Twitter, which is a microblogging service and is becoming more and
more popular. Twitter users usually have followers and follow others meanwhile. The following relationship requires no
reciprocation agreement. One can follow any other user, and the user being followed need not follow back [17]. In this paper
we call user who is being followed by user A as A’s followee, and all of one’s followees constituted his ego network. Taking
Fig. 2 as an example, node B, C, D, E, F, G are all the followees of A, BCDEFG and they constituted the ego network of A with
the relationships between them, and B’s activities is visible to A and then will influence A’s behavior.
In this study, we focus the behavior of Hashtag adoption. For a specific hashtag, followees who have adopted it before will
influence user’s behavior of adoption. Take Fig. 2 for instance, B, F, G have adopted a specific hashtag and this is visible to A,
so they will have an effect on A’s behavior. A’s influencers-B, F, G, have adopted a specific hashtag which is visible to A, then
B, F, G will have an effect on A’s adoption behavior. Next, we will focus on the structural diversity effect on this behavior,
which means to examine whether the number of group that these influences come from will have a relationship with the
behavior of user’s behavior.
270 A. Zhang et al. / Physica A 493 (2018) 267–275

3.1. Model and measurement

We built a model to explore the structural diversity effect in Twitter. In this model, besides the structural diversity, we
can also consider other factors that were used frequently in the study we mentioned in ‘‘Related Work’’ as variables in the
model.

Y = β0 + β1 ∗ stru + β2 ∗ exposure + β3 ∗ influencernum + β4 ∗ influencerpopularity + β5 ∗ influencerrela

+ β6 ∗ freqt weet + β7 ∗ freqhashtag .
The definition and measurement of each variable is as follows:
Y : the dependent variable is a binary attribute variable representing whether a user adopts a special hashtag (1) or not
(0). Y is recorded as 1 if the user also adopts this hashtag after his followees adopts a new hashtag within a specified period
of time, and Y is recorded as 0 if the user does not.
stru: the independent variable ‘‘stru’’ means ‘‘structural diversity’’. We try to divide users’ followees into different groups
to represent the distinct social contexts of them. Considering most users not using Twitter’s own function for grouping, our
study uses an algorithm from community detection problems to divide a user’s followees into groups. In view of the fact that
followees who come from different aspects of a user’s life may have little connection with each other, we use the connection
to divide groups and we use the number of groups that influencers come from measures ‘‘structural diversity’’.
exposure: this variable refers to the repeated exposure of a user to a hashtag before he adopts it. We use the total number
of times a hashtag is adopted by a user’s followees before the user finally adopts it to measure exposure.
influencernum : this variable is the number of followees using the special hashtag prior to the adoption of the hashtag by
the user. In another words which means the number of influencers.
influencerpopularity : in this study, we use the number of followers of a user to represent his popularity in the entire social
network. The variable influencer popularity, is measured by the average number of popularity of influencers.
influencerrela : a strong relationship will enhance the social contagion effect. Our study uses the number of mentions (@) of
a followee in a user’s tweets to represent the relation strength of a relationship between a user and a followee. The variable
influencer rela , is measured by the average strength levels of the relationships between a user and his influencers.
freqt weet : this variable refers to the activity level of a user on Twitter. It is measured by the number of a user’s tweet in a
given period.
freqhashtag : this variable refers to the hashtag activity of a user. It is measured by the number of the user’s tweets that
contain at least one hashtag in a given period.

3.2. Data

We use a dataset crawled from Twitter.com to proceed the study, containing 119,892 users, their Twitter relationships
and almost all of their tweets generated from September 28, 2012 to October 28, 2012. Our relationship data are based on
Twitter users’ following relation owing to our study’s purpose, and users can be affected only by the followees (in online social
networks and can see only tweets of followees). The relationship data were crawled from September 18, 2012 to September
28, 2012 (just before the tweets data were generated). Because of the directional relationship, we build a directed network,
connecting user X to user Y if X follows Y.
Adopting snow-balling sampling, we establish the sample network. Several users were chosen manually as the ‘‘seed’’
users. These users have a common feature, which is that they use hashtags frequently. The depth of our snow-balling
sampling is three, which means that our dataset contains the ‘‘seed’’ users, the users that ‘‘seed’’ users are following, the
users who are following ‘‘seed’’ users’ following users. However, of these 119,892 users, only 497 users (users whose depth
= 1 or = 2 in the snow-balling sampling) have followees all in our dataset. They are the users that we study to observe
whether they will adopt a hashtag or not under the influence of their friends. We call them ‘‘observers’’ in the study. To
the contrary, we cannot guarantee the others’ followees all belong to our dataset, thus we cannot analyze the effect of their
followees’ influence on their behaviors, though they may influence the behavior of the observers.
In the process of data achievement, we get each user’s data by crawling, including almost all the tweets generated by
these 119,892 users from September 28, 2012 to October 28, 2012. However, due to limitations of the API, if a user had
more than 3200 tweets, we could recover only the last 3200 tweets; all tweets of any users with fewer than 3200 tweets are
available (fortunately there is no user in our dataset reach this limit). As a result, Our dataset contains 15,326,497 tweets,
and 5,395,453 of them have at least one hashtag (see Table 1).

4. Experiment setup

4.1. Structural diversity analysis

The network sample contains 119,892 users and all their following relationships, on which our structural diversity
analysis based The structural diversity analysis aims to figure out how many groups an observer’s followees can be divided
into, and to distinguish to which group each one of them belongs.
 A. Zhang et al. / Physica A 493 (2018) 267–275 271

Table 1
Description of data.
Dataset Users Tweets Tweets with hashtags
Twitter 119,892 15,326,497 5,395,453

The Girvan–Newman algorithm we use to divide an observer’s followees into groups, is one of the methods to detect
communities in complex systems. In recent years, it has been widely used in recent years, especially in the social network
area [18]. It tries to identify the most central edge to communities by calculating ‘‘edge betweenness’’ of each edge, and
focuses on those edges that are of top ‘‘between’’ [19], and then the communities are detected by progressively removing
edges from the original graph. ‘‘Edge betweenness’’ of an edge is the number of shortest paths between pairs of nodes that run
along it. If a network contains communities or groups that are only loosely connected by a few intergroup edges, all shortest
paths between different communities must go along one of these few edges. Thus, the edges connecting communities will
have high edge betweenness (at least one of them). Removing these edges separates the groups from one another and thus
reveals the underlying community structure of the network. This algorithm can divide original network into any number of
groups as you need, but worth noting that it has two shortcomings: it does not give a clear standard to indicate when to stop
the iteration, and it has a high computational complexity.
Our dataset does not cover a particularly large range, and the calculated time of the G-N algorithm is still within an
acceptable range. To overcome the first disadvantage of G-N algorithm, we use the ‘‘Modularity’’ criterion, which was
originally introduced into the G-N algorithm to define a stopping criterion by Newman and Girvan [20]. Being used as a
standard to measure the quality of community detection, the modularity is, up to a multiplicative constant, the number of
edges falling within groups minus the expected number in an equivalent network with edges placed at random [21]. The
modularity can be either positive or negative, with positive values indicating the possible presence of community structure.
Thus, one can search for community structure precisely by looking for the divisions of a network that have positive, and
preferably large, values of modularity [22]. In our algorithm, we use the Girvan–Newman (G–N) algorithm to divide the
original network and use the modularity to measure the quality of division after each iteration. If the increment of modularity,
∆Q , is greater than or equal to 0, the division will be continued; if it is less than 0, the algorithm will stop, and the result of
the last round of division will be the final result because this round of iteration has declined the degree of modularity.
Due to G-N algorithm adaptive to undirected graphs only, it is necessary to transform our directed network into an
undirected network. In our study of structural diversity analysis, we retained only the mutual following relationships and
transformed them into an undirected edge. We did not consider the relationships in which, for example, A is following B,
but B is not following A for we believed that containing this type of relationship would deduce the accuracy of the result of
group division. For example, a normal user may follow some of his classmates and also follow some famous persons, and his
classmates may follow each other and celebrities. Though his classmates may also follow these famous persons, but there is
less possibility that these famous persons follow his classmates. Thus, the relationships between the user’s classmates and
between famous persons are mutual, but the relationship between the user’s classmates and the celebrities are one-way. If
we considered one-way relationships in our analysis, there would be many relations between user’s classmates and famous
persons, which should be divided into two different groups. As a result of the existence of these relations, it would be very
difficult for the algorithm to distinguish between them.
According to the theories above, the procedure of our analysis is:
Step 1: select an ‘‘observer’’, pick up all of his followees, and build this user’s ego network according to the relationships
between them.
Step 2: Delete all one-way relationships and remain the mutual relationships; transform the mutual relationships into
an undirected edge in the network.
Step 3: Use the G-N algorithm to divide the network into groups, calculate the modularity value and the increment of
modularity, ∆Q .
Step 4: If ∆ Q is negative, stop iteration and take the division result of the previous round as the final classification results,
then select another ‘‘observer’’ and restart Step 1; if ∆Q is positive, continue iteration (Step 3).

4.2. Samples generated

In our model, the dependent variable represents whether a user will adopt a hashtag after the user’s followee(s) adopt
in a period, and the independent ones represent particular instantiations of the measures in the ‘‘Study design’’ section. We
do not consider the users who did not tweet or use a hashtag. To generate the samples for validating our model, we follow
the method below:
(1) For each observer u , we collect all hashtags posted by u and his followees, and we sort them by the time they were
posted. If u’s followee(s) use a hashtag h , which has never been used by u before, pair ⟨u, h⟩ is formatted as a data point with
the dependent variable set to 1 if u has adopted this hashtag in a period and set to 0 if he has not used it.
(2) For data points whose dependent variable is 1, we calculate the independent variables within the period that started
from the time the user’s followees first used hashtag h to the time the user finally adopted this hashtag. Observer u’s
followees who used hashtag h in this period will be marked to calculate influencer num and exposure. It is easy to calculate stru
272 A. Zhang et al. / Physica A 493 (2018) 267–275

Fig. 3. Relationship of structural diversity and hashtag adoption probability.

Fig. 4. Distribution of samples.

by counting the number of groups these ‘‘influencers’’ come from, according to the results of group division. influencer rela ,
influencer popularity , freqhashtag , freq t weet are calculated with the use of the whole dataset with the measurement we mentioned
above.
(3) For data points whose dependent variables is 0, we deleted data points if user u’s followee(s) still used this hashtag
h in the last ∆t period of our data crawled time. Because we obtain a dataset of only 30 days, if the user’s followees still
adopted this hashtag in the last ∆t days, we would never know whether user u adopted this hashtag or not beyond our data
crawled time. If all of the influence was generated before ∆t, we still had ∆t to observer the user’s reaction. In our study, ∆t
was set to 10 days. The calculation of the independent variable is the same as data points whose dependent variable is 1.

5. Results

To exhibit the result, we use scatter plot to demonstrate the relationship between structural diversity and users’ behavior.
Since the dependent variable in this study is a 0–1 variable, we categorize our samples by variable ‘‘structural diversity’’ and
calculate the proportion of samples whose dependent variable is 1 in every category. This proportion is the hashtag adoption
probability when given a specific ‘‘structural diversity’’ value. Fig. 3 shows the positive relationship of structural diversity
and hashtag adoption probability. From this figure, we can find that probability that Twitter users are affected to adopt a
hashtag grows as the structural diversity increase.
We choose to adopt the Logistic model to verify the model we build above. Because of the great amount of data points
in accordance with the method described, we sample again in these data points to avoid unrealistically significant results
caused by large samples. And we use stratified sampling. In addition, for the proportion of data points whose dependent
variables were 0 is much larger than that of data points whose dependent variables were 1. We randomly selected 250 data
points from each subgroup to form the final samples tested. Fig. 4 shows some feature of our samples.
For variable ‘‘Structural diversity’’, when we divided random selected samples into two parts by dependent variable
(Hashtag adoption: user adopt Hashtag or not?), the values of variable ‘‘Structural diversity’’ are obviously different. In
Samples with dependent variable equal to zero, none of these samples’ variable ‘‘Structural diversity’’ is larger than 2, and
almost all of them are equal to 1. However, in samples with dependent variable equal to one, variable ‘‘Structural diversity’’
are from a wider range. This indicates the relationship between ‘‘Structural diversity’’ and dependent variable and is also
consistent with our hypothesis (see Fig. 5).
 A. Zhang et al. / Physica A 493 (2018) 267–275 273

Fig. 5. Distribution of samples (with dependent variable equal to 0 or 1).

Table 2
Result of forward stepwise regression.
B S.E, Wals Sig. Exp (B)
freqhashtag .004 .001 30.983 .000 1.004
Step 1
β0 −.507 .124 16.693 .000 .602
freqhashtag .003 .001 18.595 .000 1.003
Step 2 influencernum 1.334 .167 64.130 .000 3.795
β0 −2.913 .283 105.678 .000 .054
freqhashtag .003 .001 12.290 .000 1.003
influencernum 1.331 .170 61.191 .000 3.784
Step 3
influencerrela .329 .078 18.002 .000 1.390
β0 −3.182 .300 112.572 .000 .042
freqhashtag .003 .001 12.404 .000 1.003
influencernum 1.119 .168 44.204 .000 3.063
Step 4 influencerrela .340 .079 18.575 .000 1.404
Stru 1.319 .473 7.761 .005 3.740
β0 −4.349 .540 64.979 .000 .013

Table 3
Statistic of final model.
McFadden R-squared 0.488
LR statistic 358.386
Prob (LR statistic) 0.000

We use Logistic model for the reason that the dependent variable of our model is a binary attribute variable. In order to
identify critical factors, a stepwise regression procedure is used to make the model reasonable. Table 2-show the result of
forward stepwise regression in the logistic model.
The stepwise procedure choose variable freqhashtag in first step. In Step 2, the procedure control variable freqhashtag , and
influencernum are selected into the model since it can improves the model the most. In Step 3, after controlling freqhashtag and
influencernum , the stepwise procedure select influencerrela as the most significant one in the rest of variables. Analogously,
in the next step we control variable freqhashtag , influencernum , and influencerrela , to find out that variable Stru can significant
influence dependent variable. But after Step 4, none of the variables left can improve the model significantly. Therefore,
variables: influencernum , influencerrela , freqhashtag , and Stru are selected into the model. Since we adopt the stepwise procedure,
we can ensure that both of these four independent variables are significantly influence user’s adoption. The summary of the
statistic of final model is shown in Table 3.
The p value of LR statistic is smaller than 0.001, which means our model is significant at the 1% significance level. The
variables influencer num and stru are all significant (p < 0.001), and the coefficient of stru is the largest of all the coefficients
of variables. The results suggest that ‘‘structural diversity’’ has a significant effect on a user’s behavior of adopting hashtag
in Twitter, and in online social networks, the probability that a user is affected by the contagion grows with the number of
his followees who have the same behavior increasing. Moreover, what is more important is that the larger the number of
groups these followees belong to, the greater the possibility that the user will adopt the same behavior as their followees.
Additionally, the variables influencer rela and freqnum are also significant at the 5% significance level, whereas exposure,
influencer popularity , and freqt weet are not significant, which indicates that the behaviors of users in Twitter are indeed affected
by his or her followees, but if the repeated behaviors, such like using one hashtag to talk about something, are from the same
followee, it will not increase the probability that a user adopt it; a user’s level of activity on an online social network has
274 A. Zhang et al. / Physica A 493 (2018) 267–275

nothing to do with his or her online-behaviors, but the activity of the user’s special behavior will eventually have an impact;
a user is more vulnerable to the impact of influencers who have strong relationships with him, but popularity of influencers
has no obvious impact on him.

6. Discussion

This paper takes users’ adoption of hashtags as an example to study social contagion in Twitter. We focus on the variable
structural diversity and its impact on a user’s final decision to adopt a hashtag or not. The empirical results show that a
user’s behavior in Twitter is indeed influenced by his or her followees. If the followees who have this behavior are from
multi-groups, which may mean that they are heterogeneity, a user will be more likely to have this behavior.
Our study contributes to reveal the structural diversity effect on user’s behavior in Twitter, an online social network,
which is different from previous study, such like Ugander’s. Firstly, this study reveals the structural diversity effect of
community structure on user’s behavior. The structure diversity this research focuses on is the community structure of
network, using G-N community detection algorithm to divide user’s ego network into groups. This is different from previous
studies, which study the structure diversity of connected components. Connected components of user’s ego network can
represent the connectivity and interrelated between user’s friends, but community structure of user’s ego network can
reflect distinct social contexts within an individual’s life, such like common location, interests, occupation, etc. Secondly,
our research focuses on Twitter, an online social network. Twitter users always have a lot of followees that they have not
meet each other. This kind of network is different form an offline network since it has more strangers and weak connections
in it. Besides structural diversity, our research also has some interesting discoveries of implication meanings:
(1) When proving that the number of influencers and structural diversity has a significant impact on users’ behavior
in Twitter, we also find that multi-exposure (variable exposure) of a specific behavior/activity has no significant effect on
user meanwhile. This means that when controlling the number of influencers and the number of groups of influencers,
the repeated behavior of followees will not affect a user’s decision. This may be caused by that users would feel annoying
and boring when they see some of their followees tweet something again and again, and they may choose to ignore such
ones. Nowadays, more and more organizations and companies have their Twitter accounts to promote some activities and
products. To have a good promotion result, many of these accounts just post something repeatedly. However, according to
our study, we suggest them ensure their activities or promotions to be attractive enough to draw more users from more
groups to participate in it instead, which would help them to receive a good result.
(2) Our research also finds that the popularity of influencers has not significant effect on user’s behavior, but effect of the
strength of relationship on user’s behavior is significant. This also has its meaning when doing promotion. Many Twitter
account only focus on increasing the number of followers so that they can have a satisfactory consequence when they
promoting. But according to our result, to keep a good relationship with our target users and have more communication
with them would be more effective.
Besides, the study can also be applied in predicting the behavior of users, understanding the potential needs of users,
improving recommendation systems in online social networks and so on.
Our study also has some limitations. The network we built is a static network, but during the time we crawled tweet data,
the network may have been in the dynamic changes. These changes cannot be observed in this study, which may impact the
results.

Acknowledgment

This work was supported by a grant from the National Social Science Foundation of China (No. 17AGL026).

References

[1] M. Granovetter, Threshold models of collective behavior, Am. J. Sociol. (1978) 1420–1443.
[2] J.S. MacDonald, L.D. MacDonald, Chain migration, ethnic neighborhood formation, and social networks, An Urban World (1974) 226–236.
[3] T.C. Schelling, Dynamic models of segregation†, J. Math. Sociol. 1 (1971) 143–186.
[4] P.S. Dodds, D.J. Watts, Universal behavior in a generalized model of contagion, Phys. Rev. Lett. 92 (2004) 218701.
[5] L. Backstrom, D. Huttenlocher, J. Kleinberg, X. Lan, Group formation in large social networks: membership, growth, and evolution, in: Proceedings of
the 12th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, ACM, 2006, pp. 44–54.
[6] D.M. Romero, B. Meeder, J. Kleinberg, Differences in the mechanics of information diffusion across topics: idioms, political hashtags, and complex
contagion on twitter, in: Proceedings of the 20th International Conference on World Wide Web, ACM, 2011, pp. 695–704.
[7] S. Sen, S.K. Lam, A.M. Rashid, D. Cosley, D. Frankowski, J. Osterhouse, F.M. Harper, J. Riedl, Tagging, communities, vocabulary, evolution, in: Proceedings
of the 2006 20th Anniversary Conference on Computer Supported Cooperative Work, ACM, 2006, pp. 181–190.
[8] L. Yang, T. Sun, M. Zhang, Q. Mei, We know what@ you# tag: does the dual role affect hashtag adoption?in: Proceedings of the 21st International
Conference on World Wide Web, ACM, 2012, pp. 261–270.
[9] J. Ugander, L. Backstrom, C. Marlow, J. Kleinberg, Structural diversity in social contagion, Proc. Natl. Acad. Sci. (2012) 5962–5966.
[10] H.C. Chang, A new perspective on Twitter hashtag use: Diffusion of innovation theory, Proc. Amer. Soc. Inf. Sci. Technol. 47 (2010) 1–4.
[11] S.A. Macskassy, M. Michelson, Why Do People Retweet? Anti-Homophily Wins the Day!, in: Proceedings of the Fifth International AAAI Conference
on Weblogs and Social Media, 2011, pp. 209–216.
[12] N. Harrigan, P. Achananuparp, E.P. Lim, Influentials, novelty, and social contagion: The viral power of average friends, close communities, and old news,
Social Networks (2012).
 A. Zhang et al. / Physica A 493 (2018) 267–275 275

[13] E. Bakshy, I. Rosenn, C. Marlow, L. Adamic, The role of social networks in information diffusion, in: Proceedings of the 21st International Conference
on World Wide Web, ACM, 2012, pp. 519–528.
[14] M.S. Granovetter, The strength of weak ties, Am. J. Sociol. (1973) 1360–1380.
[15] G. Simmel, Conflict and the Web of Group Affiliations, Free Press, 2010.
[16] R.S. Burt, Structural Holes: The Social Structure of Competition, Harvard University Press, 2009.
[17] H. Kwak, C. Lee, H. Park, S. Moon, What is Twitter, A Social Network or a News Media? ACM, 2010, pp. 591–600.
[18] D. Easley, J. Kleinberg, Networks, Crowds, and Markets, Cambridge Univ. Press, 2010.
[19] M. Girvan, M.E. Newman, Community structure in social and biological networks, Proc. Natl. Acad. Sci. 99 (2002) 7821–7826.
[20] M.E. Newman, M. Girvan, Finding and evaluating community structure in networks, Phys. Rev. E 69 (2004) 026113.
[21] M.E. Newman, Modularity and community structure in networks, Proc. Natl. Acad. Sci. 103 (2006) 8577–8582.
[22] M.E. Newman, Fast algorithm for detecting community structure in networks, Phys. Rev. E 69 (2004) 066133.