Interest Assortativity in Twitter
Francesco Buccafurri, Gianluca Lax, Serena Nicolazzo and Antonino Nocera
DIIES, University Mediterranea of Reggio Calabria, Via Graziella, Localit`a Feo di Vito, 89122, Reggio Calabria, Italy
Keywords:
Assortativity, Social Network Analysis, Twitter.
Abstract:
Assortativity is the preference for a person to relate to others who are someway similar. This property has
been widely studied in real-life social networks in the past and, more recently, great attention is devoted to
study various forms of assortativity also in online social networks, being aware that it does not suffice to apply
past scientific results obtained in the domain of real-life social networks. One of the aspects not yet analyzed
in online social networks is interest assortativity, that is the preference for people to share the same interest
(e.g., sport, music) with their friends. In this paper, we study this form of assortativity on Twitter, one of the
most popular online social networks. After the introduction of the background theoretical model, we analyze
Twitter, discovering that users clearly show interest assortativity. Beside the theoretical assessment, our result
leads to identify a number of interesting possible applications.
1 INTRODUCTION
Assortativity (often called assortative mixing) (New-
man, 2002) is an empirical measure describing a pos-
itive correlation in personal attributes of people so-
cially connected with each other. Traits exhibiting as-
sortative mixing can be various, like age, physical ap-
pearance, education, socio-economic status, religion,
etc. For example, we say that age is assortative in
a given set of people if the probability that two in-
dividuals with close age are friends is higher than
the probability that two randomly selected individu-
als are friends. Despite the difficulty of explaining
its cause among homophily (Lazarsfeld and Merton,
1954; McPherson et al., 2001), contagion, opportu-
nity structures and sociality mechanisms (Ackland,
2013), the empirical observation of assortative mix-
ing in real-life social networks has been considered
by sociologists of remarkable importance, as basis to
study a community under the point of view of friend-
ship formation and social influence.
From the birth of Facebook, the extraordinary
growth of online social networks (Buccafurri et al.,
2013; Buccafurri et al., 2014a) has enforced scientists
to enlarge their point of view in order to consider on-
line social networks as a specific social phenomenon.
Thus, it is important to study the social dynamics of
online social networks being aware that they cannot
be trivially understood by applying past scientific re-
sults obtained in the domain of real-life social net-
works. Assortativity is one of these phenomenons.
Indeed, it is not obvious whether assortativity, es-
pecially that of psychological states (Bollen et al.,
2011), takes place also in online social networks. In
general, the fact that social ties are only mediated by
online networking services instead of physical inter-
actions, might influence the behavior of the commu-
nity. As a matter of fact, a lot of work exist aimed
at studying various people traits from the assortativ-
ity perspective (Bliss et al., 2012; Bollen et al., 2011;
Ciotti et al., 2015; Ahn et al., 2007; Johnson et al.,
2010; Benevenuto et al., 2009).
In this paper, we consider a trait which is basilar in
social dynamics, that is people’s interests. However,
to the best of our knowledge, assortativity of this trait
has not been studied so far in online social networks.
Online social networks are a good laboratory to
study whether (and how much) interests are assorta-
tive because in real-life communities interest assorta-
tivity is mostly dominated by opportunity structures
and sociality mechanisms (for examples, the physi-
cal place attended by individuals who share an inter-
est, like a cinema club). Thus, online social networks
may help us to better understand whether friends tend
to share interests for homophily or contagion reasons.
To study interest assortativity in online social net-
works, this paper uses an approach based on public
figures of Twitter. The choice of Twitter is related to
both the goal of trying to have results little affected by
physical friendship and the fact that it has been widely
used in several heterogeneous application scenarios
(Lax et al., 2016; Buccafurri et al., 2014b). Indeed, as
Buccafurri, F., Lax, G., Nicolazzo, S. and Nocera, A.
Interest Assortativity in Twitter.
In Proceedings of the 12th Inter national Conference on Web Information Systems and Technologies (WEBIST 2016) - Volume 1, pages 239-246
ISBN: 978-989-758-186-1
Copyright
c
2016 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
239
shown in (Panek et al., 2013), differently from Twit-
ter, Facebook users tend to add as friend people they
know in real life in order to transform latent ties to
weak ties (Ellison et al., 2007). On the other hand,
(Hargittai and Litt, 2011) highlights that Twitter use
is driven primarily by interest for entertainment news,
celebrity news, and sports news. This allows us to
map the abstract concept of interest (or topic) to the
concrete entity of public figure, to the extent that a
public figure in a given field, say Gordon Ramsay,
acts as a representative of a topic, haute cuisine (i.e.,
high-level cooking), in our example. Thus, we assim-
ilate the followship of a user to the Twitter profile of
Gordon Ramsay to the fact that this user is interested
in high-level cooking.
The way to study interest assortativity is to ob-
serve, for a number of public figures, if the measured
probability that two Twitter friends share the follow-
ship to the same public figures is higher than the ran-
dom case (i.e., with no assortativity). The result ob-
tained is that interests are significantly assortative in
Twitter. Interestingly, the quantified degree of inter-
est assortativity is much more higher than other forms
of assortativity measured in social networks in the
past. The knowledge of this form of assortativity may
help to understand several aspects of social networks,
such as information propagation, identifying influen-
tial and susceptible members, network resilience. The
contributions of our paper are summarized in the fol-
lowing:
1. we define a new form of assortativity in a social
network, the interest assortativity, and we study
how to measure it;
2. we characterize the random graph of the net-
work, commonly denoted as null model (New-
man, 2002), necessary to quantify the assortativ-
ity;
3. we compare the measured value of interest assor-
tativity with other forms of assortativity discov-
ered in the past.
4. we sketch how our result impacts on a number of
possible application contexts.
The plan of this paper is as follows. Section 2 presents
our assortativity measure. Section 3 describes the ex-
perimental campaign carried out on Twitter both to
validate the new assortativity measure and to verify
whether Twitter shows an assortative/disassortative
behavior about interests. In Section 4, we show how
the result about interest assortativity reported in this
paper may support several applications in the context
of social networks. Section 5 deals with literature
about assortativity. Finally, in Section 6, we draw our
conclusions and discuss possible future work.
2 INTEREST ASSORTATIVITY
In this section, we define the interest assortativity of
a social network and how this can be measured. First,
we model a social network and introduce the notation
used in the following.
• An online social network is modelled as a directed
graph G = hN, Ei, where N is the set of nodes
(accounts), and E is the set of edges, i.e., or-
dered pairs of nodes (representing a relationship
between two accounts).
• Given an interest I and a positive integer t, we de-
note by N
t,I
the set of the top t indegree nodes
followed from other nodes interested in I
1
;
• Given a node n ∈ N, we denote by Γ
out
(n) the set
{n
′
∈ N s.t. (n, n
′
) ∈ E}, and by Γ
in
(n) the set
{n
′
∈ N s.t. (n
′
, n) ∈ E}.
Γ
out
(n)
and
Γ
in
(n)
represent outdegree and indegree of n, respec-
tively.
• The set Γ
out
(n) is said the set of friends (or fol-
lowings) of n, whereas the set Γ
in
(n) is said the
set of followers of n.
• Given d ≥ 0, we denote by N
d
⊆ N the set of nodes
with outdegree d.
We are now ready to introduce interest assortativity.
Interest assortativity occurs when the probability that
two friends share the same interest is higher than that
observed in a network in which friendship edges are
set in a random way. As common in this context
(Holme and Zhao, 2007; Bliss et al., 2012), assorta-
tivity is quantified as the difference between the mea-
sure of the studied trait in the observed network and
that computed in the corresponding random graph of
the network (Newman, 2002). Therefore, we need to
measure these two quantities. Let us start from the
observed network.
Definition 2.1. Given a network G, an interest I and a
positive integer t, we define the Interest Friend Frac-
tion towards the top t nodes of I in G as
IFF
t,I
=
L
S
n∈N
t,I
Γ
in
(n)
where L =
{a ∈ N s.t. ∃ b ∈ Γ
out
(a), n
1
, n
2
∈ N
t,I
∧
a ∈ Γ(n
1
) ∧ b ∈ Γ(n
2
)}
.
In this equation, the numerator is the cardinality
of the set L composed of the nodes a such that (1)
have at least another node b in their neighborhood
(i.e., b ∈ Γ
out
(a)), (2) are followers of a node n
1
in
1
In Section 3, we will clarify how to find this set.
WEBIST 2016 - 12th International Conference on Web Information Systems and Technologies
240
Figure 1: Interest Friend Fraction computation.
the top t indegree nodes in I (i.e., n
1
∈ N
t,I
) and (3) b
is follower of a node n
2
in the top t indegree nodes in
I; the denominator is the cardinality of the set of the
followers of any node n in the top t indegree nodes in
I. Observe that, we use the set N
t,I
to assimilate the
abstract concept of interest (or topic) to the concrete
entity of public figures, seen as representatives of in-
terests. Broadly speaking, the Interest Friend Frac-
tion measures the fraction of the nodes interested in I
having at least one friend interested in I too. In Figure
1, it is represented a network with 10 nodes in which
the arrows show the direction of the following rela-
tionships among nodes. Assuming that n
′
and n
′′
are
the top 2 nodes in the interest music (thus, they belong
to N
2,music
), then the Interest Friend Fraction towards
the top 2 nodes of music IFF
2,music
=
2
10
= 0, 2, be-
cause two nodes (i.e., a
′
and a
′′
) have as friends nodes
following nodes in N
2,music
(i.e., b
′
and b
′′
).
After measuring the phenomenon in the observed
network by IFF, we need to characterize the ran-
dom graph of the network, commonly denoted as null
model (Newman, 2002). This graph models the case
in which no assortativity occurs.
Definition 2.2. Given a network G, the null model of
G is the random graph
ˆ
G = hN,
ˆ
Ei such that for each
v ∈ N, it holds that:
1)
{(x, w) ∈ E s.t. x = v}
=
{(x, w) ∈
ˆ
E s.t. x = v}
and
2)
{(w, x) ∈ E s.t. x = v}
=
{(w, x) ∈
ˆ
E s.t. x = v}
.
In words, it is obtained by maintaining the nodes
of G and by replacing the deterministic occurrence
of edges by a random variable in such a way that in-
degree (condition 1) and outdegree (condition 2) of
nodes are maintained. Now, we define how to mea-
sure IFF in the null model.
Definition 2.3. Given a network G, an interest I, and
a positive integert, we define the Interest Friend Frac-
tion towards the top t nodes of I in
ˆ
G as:
\
IFF
t,I
= P
y ∈ F s.t. y ∈ Γ
out
(x) ∧ x ∈ F
where P (X) stands for probability of X and the set
F =
S
n∈N
t,I
Γ
in
(n).
In words,
\
IFF
t,I
is the probability that a node in-
terested in I has at least one friend interested in I too.
Clearly, nodes interested in I are modelled by the set
F, which is composed of all nodes following any of
the top t nodes for the interest I. The following theo-
rem allows us to compute this probability.
Theorem 2.1. Given a network G, an interest I, and
a positive integer t, then:
\
IFF
t,I
=
u
∑
d=2
N
d
N
·
1− γ
N
− 2, d − 1,
F
− 1
!
where:
γ(a, b, c) =
c
∏
k=0
a− b− k
a− k
and F =
[
n∈N
t,I
Γ
in
(n)
and N is the number of nodes, N
d
is the number of
nodes with outdegree d, and Γ
in
(n) is the set of fol-
lowers of n.
Proof.
\
IFF
t,I
is obtained as the sum for any degree
2 ≤ d ≤ u of the product of two factors: the probabil-
ity of having a node with outdegree d, and the proba-
bility that this node and at least one of its friend follow
a node in F.
In the theorem above, it is necessary to compute
|F|, which is the number of followers in the null
model of any node in N
t,I
. The next theorem allows
us to do this.
Theorem 2.2. Given a null model
b
G, an interest I,
and a positive integer t, then:
[
n∈N
t,I
Γ
in
(n)
=
N
·
1−
∏
n∈N
t,I
1−
Γ
in
(n)
N
!!
Proof.
|Γ
in
(n)|
|N|
is the probability to be follower of the
node n, which has indegree |Γ
in
(n)|, whereas 1 minus
this value is the probability of not following it. In
the null model, the probability to follow a node n
′
is
independent of that to follow a node n
′′
, so that the
products
∏
n∈N
t,I
1−
Γ
in
(n)
N
is the probabilityto not
follow any node in N
t,i
. Thus, 1 minus this value is
the probability to follow any node in N
t,i
. Finally, this
probability multiplied by the overall number of nodes
in the network is the searched value.
Now we are ready to give the formal definition of
Interest Assortativity.
Interest Assortativity in Twitter
241
Definition 2.4. Given a network G, an interest I, and
a positive integer t, we define the Interest Assortativ-
ity towards the top t nodes of I in G as:
IA
t,I
= IFF
t,I
−
\
IFF
t,I
This measure gives us an index of how much a so-
cial network is biased w.r.t. the null model in terms of
probability of finding friends sharing the same inter-
est. The higher the value of assortativity, the higher
the correlation in being friends and sharing the same
interest.
We conclude this section by showing how to com-
pute the interest assortativity towards the top 2 nodes
of music for the network depicted in Figure 1. We
have already computed IFF
2,music
= 0.2. Concerning
the null model of this network, by using Theorem 2.2
we obtain that |F| = 6 in the null model (i.e., there are
6 nodes following n
′
or n
′′
therein). Concerning the
degree, we have 2 nodes, n
′
and n
′′
, with outdegree
0 (thus, N
0
= 2), 3 nodes a
′
, a
′′
and b
′′
, with outde-
gree 2 (thus, N
2
= 3), and the remaining nodes with
outdegree 1 (thus, N
1
= 5). Consequently,
\
IFF
2,music
=
N
2
N
·
1− γ
8, 1, 5
!
=
=
3
10
1−
3
8
!
= 0.1875.
Finally, we have that:
IA
2,music
= IFF
2,music
−
\
IFF
2,music
=
= 0.2− 0.1875 = 0.0125.
This result indicates that the network in Figure 1
does not show interest assortativity, as the measured
value is very close to zero. From a qualitative point
of view, this means that the particular configuration
of the network, with only two nodes (i.e., a
′
and a
′′
)
whose friends (i.e., b
′
and b
′′
) are interested in music,
is a result that we could obtain with a high probability
also by randomly setting the edges of the network.
Therefore, no relation between node friendship and
interest in music exists in this network.
3 EXPERIMENTS
This section aims at describing the tests carried out
to measure interest assortativity on real-life data. For
this purpose, we choose Twitter as target social net-
work both because it is one of the most popular and
Table 1: Dataset characteristics.
# Vip indegree (Min-Max) # Checks
Music 30 18.038.914-65.389.090 6.435.561
Sport 30 6.850.604-33.686.429 8.207.264
Cinema 30 7.262.979-26.732.512 10.581.658
studied social sites (Gjoka et al., 2010; Patriquin,
2007; Kwak et al., 2010) and for some of its fea-
tures which fit well with our definitions. Indeed, one
of the major advantage of Twitter is that all accounts
are publicly available and accessible through a set of
APIs
2
. Moreover, because it is focused on the diffu-
sion of information embedded inside short messages
(Tweets), the notion of public figures (i.e., celebrities)
is very prominent as people often use Twitter to get
news from their favorite celebrities. As mentioned
above, we use the “following” relationships (see Def-
inition 2.1) between users and public figures as an ex-
plicit declaration of interest to the field to which the
celebrities belong to.
In our experiments, we considered the three cate-
gories Music, Cinema and Sport as interests and we
identified the t most followed celebrities on Twitter
for each category. The value t is chosen by apply-
ing as a criterium that the number of followers of the
top t chosen public figures is of the same magnitude
order as the number of Twitter users. We found out
that t = 30 satisfies the above criterium for the chosen
interests. Consequently, we built three sets each in-
cluding 30 celebrities, which corresponds to the sets
N
t,I
introduced in Section 2). For each celebrity, we
extracted its first- and second-level neighbors to ob-
tain information necessary for the computation of our
assortativity measure.
The characteristics of the analyzed real-life
dataset are presented in Table 1. The second column
reports the number of followers of the 30
th
and 1
st
celebrity (min and max, resp.), whereas the last col-
umn reports how many of such followers have been
randomly selected and analyzed in the computation
of the interest assortativity. It is worth noting that the
number of visited nodes (seemingly small compared
with the actual size of the Twitter network) does not
limit the validity of the results, because (1) such users
are randomly selected and (2) the value of assortativ-
ity measured for each of them is extremely high (as
we will show in the following). Moreover, the car-
dinality of our dataset is consistent with that of other
researches addressing the same topic (see, for exam-
ple, the datasets used in (Bliss et al., 2012)).
To compute our assortativity measure for each cat-
2
The API documentation is available at
https://dev.twitter.com/overview/documentation.
WEBIST 2016 - 12th International Conference on Web Information Systems and Technologies
242
egory, we need to obtain IFF
t,I
and
\
IFF
t,I
(see Def-
inition 2.4). These computations proceed as follows.
Given a public figure n in the top t nodes of a given
category, we consider the nodes who are followers of
n (i.e., the first-level neighbors of n). Then, we an-
alyze the neighbors of these nodes (i.e., the second-
level neighbors of n) and verify whether they are fol-
lowers of any of the top t nodes.
To compute
\
IFF
t,I
for each category defined
above, we need to build three instances of the null
model according to Definition 2.2. To do this, we
have to compute the parameters required by Theo-
rem 2.1 and Theorem 2.2. Specifically, the number of
users of Twitter is set to 645,750,000 according to the
annual report of 2015
3
. Concerning the degree distri-
bution of nodes, it follows a power law as proved in
(Lu and Wang, 2014). For this reason, we used this
kind of distribution to approximate the social network
characteristics measured in our sample and we found
that the average degree d of Twitter is about 63.
In our experiments, we randomly selected a
large number of Twitter accounts following a given
celebrity and checked if such an account shows in-
terest assortativity, measured according to Definition
2.4.
In Table 2, we report the values of interest assor-
tativity on Twitter measured for each category at the
end of the experiment.
Table 2: The computed values of IA for the three categories.
Category IFF
d
IFF
IA
Music 1.000 0.263 0.737
Cinema 0.988 0.177 0.811
Sport 0.973 0.197 0.776
To fully understand the results of our experiments,
it is worth recalling that values of assortativity greater
than 0.3 means that the phenomenon observed in the
social network is very prominent. Indeed, in the first
paper on assortativity (Newman, 2002), which will be
discussed in Section 5, the most assortative network
was the physics coauthorship one, with an assortativ-
ity value equal to 0.363 (see Table 3). Under this rea-
soning and by considering our results, we can state
that Twitter is highly assortative w.r.t. the considered
trait (i.e., interests). Moreover, the comparison be-
tween the three categories (i.e., the three different top-
ics) shows very little differences among them. This
means that, the assortative behavior of Twitter users
is not bound to a single and well-defined interest, but
the general trend is uniform w.r.t. different topics.
3
Available at http://twittercounter.com/pages/100.
Table 3: Degree assortativity measured on a number of dif-
ferent real-world networks.
Network Assortativity
physics coauthorship (Newman, 2002) 0.363
biology coauthorship (Newman, 2002) 0.127
mathematics coauthorship (Newman, 2002) 0.120
film actor collaborations (Newman, 2002) 0.208
company directors (Newman, 2002) 0.276
SCN (Zhang et al., 2012) 0.161
CA-HepTh (Zhang et al., 2012) 0.268
CA-GrOc (Zhang et al., 2012) 0.659
4 APPLICATIONS
In this section, we show how the result about inter-
est assortativity reported in this paper can be quali-
tatively related to some applications. A quantitative
study is left as future work. In our study, we used
public figures to characterize interests, and showed
that the probability that two friends follow the same
interest (i.e., the same public figure) is (much) higher
than the probability that a pair of users randomly cho-
sen do this. This means that interests are assortative
in the social network.
Preliminarily, observe that the applications of in-
terest assortativity may rely also on the fact that our
notion (as it usually happens for assortativity mea-
sures) is not simply on/off, but gives us a measure
of the degree of correlation between sharing of inter-
ests and being friends of a given network/subnetwork.
Moreover, observe that even though the aim of this
paper is to compute interest assortativity of Twitter,
the same method can be used to compute interest
assortativity of another social network or that of a
meaningful subnetwork of a given social network –
for example the ego-network (Leskovec and Mcauley,
2012) of a give user. The applications sketched in this
section are thought considering that the above exten-
sions are easily feasible.
Consider this first application. Suppose we would
like to propagate some information in a social net-
work (say Twitter) with the target of reaching the
largest set of people sharing a given interest I. In
principle, we could use as starters the highest possi-
ble number of public figures representing this inter-
est. A reasonable way to do this is to take the top-t
public figures (in terms of followers), where t is such
that the magnitude order of the involved followers is
that of Twitter. Observe that this is just the criterium
we have followed to measure the interest assortativ-
ity of Twitter w.r.t. I. Denote by IA
I
this value. We
have to consider that, typically, reaching an agree-
ment with a public figure for this job could be very
Interest Assortativity in Twitter
243
hard and expensive. As seen earlier, a realistic value
for t could be 30 (for the interests we have consid-
ered in this paper), which is definitely high for the
application we are considering. We should drastically
reduce the number of starters. But how to do this? Is
a blind way acceptable? Plausibly, our interest assor-
tativity measure could drive the above minimization.
Indeed, we could start by considering the top-1 pub-
lic figure and by measuring the gap existing between
IA
I
and the value computed by considering only this
public figure. We expect that the measured assorta-
tivity is lower than IA
I
. Then, we could consider the
top-2 public figures and iterate the measure, and so
on, until an acceptable closeness between the mea-
sured assortativity and IA
I
is reached. This allows us
to reach a number of followers whose friends share
the given interest with high probability, thus heuris-
tically increasing the probability of reaching a large
community sharing this interest.
A second application could concern the identifi-
cation of one (or more) expert(s) in a given field, rep-
resented by a given interest I. Suppose we have a
number of social network (say Twitter) profiles rep-
resenting possible candidates to play the role of ex-
perts on I. We can argue that the degree of exper-
tise of this user about the topic I is positively corre-
lated to the degree of interest assortativity of her/his
egonetwork. Indeed, the higher the assortativity, the
higher the probability that people in her/his egonet-
work are directly connected by a friendship relation.
Thus, the higher the exchange of information regard-
ing the topic I involving (directly or indirectly) U.
Finally, another possible application of our results
is related to the resilience of the social network w.r.t.
information diffusion. Indeed, public figures are typ-
ically source of information: news, opinions, com-
ments, links, etc. Suppose now that the public figure
P tweets a piece of information I. Obviously, the full
success of the diffusion occurs whenever all the fol-
lowers re-tweets I (and, in turn, the same happens for
their followers). Now, it is well known that the so-
cial network universe can be thought as a set of com-
munities, with inner strong correlations (strong ties)
weakly connected by means of weak ties. What is
important is that the information I reaches at least
one element (or a few elements) of the community, as
the high communication activity among members of
the community will allow the whole community to be
reached by I (eventually). Now, the results we have
here demonstrated about interest assortativity tell us
that in case a given followerU of P receives the infor-
mation I but fails in re-tweetting it, there is a consid-
erable probability that some users at distance 1 from
U are also followers of P, thus recipients of the infor-
mation I. Therefore, the probability that one of these
users re-tweets I (thus, propagating it) is higher than
the case of absence of interest assortativity. In other
words, interest assortativity increases the resilience of
the network w.r.t. its capability of propagating infor-
mation, in case of failure of some nodes.
5 RELATED WORK
Newman was the first author to introduce a formal
definition and a metric for the concept of assorta-
tivity. In particular, in (Newman, 2002) he demon-
strates that social networks are often assortatively
mixed, in the sense that the nodes in the network
having many relationships tend to be connected to
other nodes highly connected themselves. Starting
from (Newman, 2002) further studies concerning so-
cial network assortativity have been proposed, such
as (Newman and Park, 2003; Catanzaro et al., 2004;
Goh et al., 2003) Specifically, the authors of (New-
man and Park, 2003) present a deep analysis about
the relation between clustering and assortativity in the
communities composing a social network. As result
they obtain that these communities are characterized
by both high levels of clustering and assortative mix-
ing. By contrast, Catanzaro et al. (Catanzaro et al.,
2004) compare technological and biological networks
and social networks, showing that, while the former
appear, in general, to be disassortative with respect
to the degree, social networks are typically assorta-
tive. Moreover, in (Goh et al., 2003) a study on the
relationship between assortativity and betweenness
centrality correlation for scale-free networks is pre-
sented. The work proposed in (Hu and Wang, 2009)
analyzes the structural evolution of large online so-
cial networks and concludes that, with the huge in-
crease of their size, many network properties show a
non-monotone behavior. This is the case of density,
clustering, heterogeneity, and modularity. Some re-
sults focusing on degree assortativity are presented in
(Ciotti et al., 2015; Ahn et al., 2007; Johnson et al.,
2010; Benevenuto et al., 2009) Ciotti et al. (Ciotti
et al., 2015) investigate degree correlations in two on-
line social networks. The major result of their paper
is that, while subnetworks characterized by assorta-
tive mixing by degree have in general links express-
ing a positive connotation (i.e., endorsement or trust),
networks in which links have a negative connotation,
(i.e., disapproval and distrust) are described by disas-
sortative patterns. The authors of (Ahn et al., 2007)
compute the degree assortativity of Cyworld, MyS-
pace and Orkut and find that these online social net-
works do not show a degree correlation pattern similar
WEBIST 2016 - 12th International Conference on Web Information Systems and Technologies
244
to that of real-life social networks. This can be due
to the fact that they encourage activities that cannot
be copied in real life. Indeed, an opposite behavior
is observed for those online social networks handling
activities similar to real-life ones. The most relevant
and recent studies on Twitter assortativity have been
carried out by (Kwak et al., 2010; Bollen et al., 2011;
Bliss et al., 2012). The analysis of Twitter assortativ-
ity (Kwak et al., 2010) proved that users with 1,000
followers or less are likely to have the same num-
ber of followers (that is, the same popularity) of their
reciprocal-friends and also to be geographically close
to them. An attempt to define assortativity on mul-
tiple social networks instead of on single social net-
works is presented in (Buccafurri et al., 2015). In par-
ticular, the authors measure the tendency of users of
associating their Facebook and Twitter account with
others in different social sites. Moreover, they pro-
vide the behavioral and sociological interpretation of
the experimental results. Finally, the authors identify
an interesting relationship between explicit member-
ship overlap assortative mixing and implicit member-
ship overlap. This led to the discovery that assorta-
tivity may be source of private information leakage,
as it can improve the chance of disclosing implicit
membership overlap. Our perspective is quite differ-
ent from that of the works already present in litera-
ture, because it deals with assortativity in people’s in-
terests, that is a basilar trait in social dynamics. We
carry out our analysis by relying on users’ interests in
Twitter and, to the best of our knowledge, this form of
assortativity has not yet been studied so far in online
social networks.
6 CONCLUSION
In this paper, we have defined a new form of as-
sortativity, called Interest Assortativity and studied
it in Twitter. Our analysis has been carried out by
measuring the value of interest assortativity for real-
life accounts. The approach used to identify inter-
ests is based on Twitter public figures: in particular,
we have considered the “following” relationships be-
tween users and public figures as an explicit declara-
tion of interest towards the field to which the celebri-
ties belong to. The results of our study allow us to
state that Twitter is highly assortative in users inter-
ests and that there are not significant differences for
the three categories of interests we have considered.
This means that users behave uniformly w.r.t. differ-
ent topics. We showed also possible applications of
our results related to information propagation and so-
cial network resilience.
Future work could extend the analysis by con-
sidering other online social networks and by study-
ing from a quantitative point of view the relationship
between interest assortativity and social network re-
silience.
ACKNOWLEDGEMENTS
This work has been partially supported by the Pro-
gram “Programma Operativo Nazionale Ricerca e
Competitivit`a” 2007-2013, Distretto Tecnologico Cy-
berSecurity and project BA2Know (Business Analyt-
ics to Know) PON03PE
00001 1, in “Laboratorio in
Rete di Service Innovation”, both funded by the Ital-
ian Ministry of Education, University and Research.
REFERENCES
Ackland, R. (2013). Web social science: Concepts, data
and tools for social scientists in the digital age. Sage.
Ahn, Y., Han, S., Kwak, H., Moon, S., and Jeong, H. (2007).
Analysis of topological characteristics of huge online
social networking services. In Proc. of the Interna-
tional Conference on World Wide Web (WWW’07),
pages 835–844, Banff, Alberta, Canada. ACM.
Benevenuto, F., Rodrigues, T., Almeida, V., Almeida, J.,
and Gonc¸alves, M. (2009). Detecting spammers and
content promoters in online video social networks.
In Proc. of the International Conference on Research
and Development in Information Retrieval (SIGIR
’09), pages 620–627, Boston, MA, USA. ACM.
Bliss, C., Kloumann, I., Harris, K., Danforth, C., and
Dodds, P. (2012). Twitter reciprocal reply networks
exhibit assortativity with respect to happiness. Jour-
nal of Computational Science, 3(5):388–397.
Bollen, J., Gonc¸alves, B., Ruan, G., and Mao, H. (2011).
Happiness is assortative in online social networks. Ar-
tificial life, 17(3):237–251.
Buccafurri, F., Lax, G., Nicolazzo, S., and Nocera, A.
(2014a). A Model to Support Multi-Social-Network
Applications. In Proc. of the International Confer-
ence Ontologies, DataBases, and Applications of Se-
mantics (ODBASE 2014), pages 639–656, Amantea,
Italy. Springer.
Buccafurri, F., Lax, G., Nicolazzo, S., Nocera, A., and
Ursino, D. (2013). Measuring Betweennes Centrality
in Social Internetworking Scenarios. In Proc. of Inter-
national Workshop on Social and Mobile Computing
for collaborative environments (SOMOCO’13), pages
666–673, Gratz, Austria. Springer Verlag.
Buccafurri, F., Lax, G., Nicolazzo, S., Nocera, A., and
Ursino, D. (2014b). Driving Global Team Formation
in Social Networks to Obtain Diversity. In Proc. of the
International Conference on Web Engineering (ICWE
2014), pages 410–419, Toulouse, France. Springer.
Interest Assortativity in Twitter
245
Buccafurri, F., Lax, G., and Nocera, A. (2015). A new
form of assortativity in online social networks. Inter-
national Journal of Human-Computer Studies, 80:56–
65.
Catanzaro, M., Caldarelli, G., and Pietronero, L. (2004).
Social network growth with assortative mixing. Phys-
ica A: Statistical Mechanics and its Applications,
338(1):119–124.
Ciotti, V., Bianconi, G., Capocci, A., Colaiori, F., and Pan-
zarasa, P. (2015). Degree correlations in signed social
networks. Physica A: Statistical Mechanics and its
Applications, pages 25–39.
Ellison, N. B., Steinfield, C., and Lampe, C. (2007). The
benefits of facebook friends: social capital and college
students use of online social network sites. Journal
of Computer-Mediated Communication, 12(4):1143–
1168.
Gjoka, M., Kurant, M., Butts, C., and Markopoulou, A.
(2010). Walking in Facebook: A case study of un-
biased sampling of OSNs. In Proc. of the Interna-
tional Conference on Computer Communications (IN-
FOCOM’10), pages 1–9, San Diego, CA, USA. IEEE.
Goh, K., Oh, E., Kahng, B., and Kim, D. (2003). Between-
ness centrality correlation in social networks. Physical
Review E, 67(1):017101.
Hargittai, E. and Litt, E. (2011). The tweet smell of
celebrity success: Explaining variation in twitter
adoption among a diverse group of young adults. New
Media & Society, 13(5):824–842.
Holme, P. and Zhao, J. (2007). Exploring the assortativity-
clustering space of a network’s degree sequence.
Physical Review E, 75(4):046111.
Hu, H. B. and Wang, X. F. (2009). Evolution of a large on-
line social network. Physics Letters A, 373(12):1105–
1110.
Johnson, S., Torres, J., Marro, J., and Munoz, M. (2010).
Entropic origin of disassortativity in complex net-
works. Physical review letters, 104(10):108702.
Kwak, H., Lee, C., Park, H., and Moon, S. (2010). What
is Twitter, a social network or a news media? In
Proc. of the International Conference on World Wide
Web (WWW’10), pages 591–600, Raleigh, NC, USA.
ACM.
Lax, G., Buccafurri, F., Nicolazzo, S., Nocera, A., and Fo-
tia, L. (2016). A new approach for electronic signa-
ture. In Proc. of the International Conference on In-
formation Systems Security and Privacy (ICISSP 16)),
Rome, Italy.
Lazarsfeld, P. and Merton, R. (1954). Friendship as a social
process: A substantive and methodological analysis.
Freedom and control in modern society, 18(1):18–66.
Leskovec, J. and Mcauley, J. J. (2012). Learning to discover
social circles in ego networks. In Advances in neural
information processing systems, pages 539–547.
Lu, J. and Wang, H. (2014). Variance reduction in large
graph sampling. Information Processing & Manage-
ment, 50(3):476–491.
McPherson, M., Smith-Lovin, L., and Cook, J. (2001).
Birds of a feather: Homophily in social networks. An-
nual review of sociology, 27:415–444.
Newman, M. (2002). Assortative mixing in networks. Phys-
ical Review Letters, 89(20):208701.
Newman, M. and Park, J. (2003). Why social networks are
different from other types of networks. Physical Re-
view E, 68(3):036122.
Panek, E. T., Nardis, Y., and Konrath, S. (2013). Mirror or
megaphone?: How relationships between narcissism
and social networking site use differ on facebook and
twitter. Computers in Human Behavior, 29(5):2004–
2012.
Patriquin, A. (2007). Connecting the Social Graph: Mem-
ber Overlap at OpenSocial and Facebook. Compete.
com blog.
Zhang, G.-Q., Cheng, S.-Q., and Zhang, G.-Q. (2012). A
universal assortativity measure for network analysis.
arXiv preprint arXiv:1212.6456.
WEBIST 2016 - 12th International Conference on Web Information Systems and Technologies
246
