Integrating User’s Emotional Behavior for Community Detection in

Social Networks

Andreas Kanavos, Isidoros Perikos, Ioannis Hatzilygeroudis and Athanasios Tsakalidis

Computer Engineering and Informatics Department, University of Patras, Patras, Hellas

Keywords:

Community Analysis, Graph Mining, Inﬂuential Community Detection, Sentiment Analysis, Tweet Emotion

Recognition, User Inﬂuence.

Abstract:

The analysis of social networks is a very challenging research area. A fundamental aspect concerns the de-

tection of user communities, i.e. the organization of vertices in clusters, with many edges joining vertices of

the same cluster and comparatively few edges joining vertices of different clusters. Detecting communities

is of great importance in sociology, biology as well as computer science where systems are often represented

as graphs. In this paper we present a novel methodology for community detection based on users’ emotional

behavior. The methodology analyzes user’s tweets in order to determine their emotional behavior in Ekman

emotional scale. We deﬁne two different metrics to count the inﬂuence of produced communities. Moreover,

the weighted version of a modularity community detection algorithm is utilized. Our results show that our

proposed methodology creates inﬂuential enough communities.

1 INTRODUCTION

The increasing popularity of social media, including

Twitter which we consider in the present manuscript,

has gained in recent years, huge research interest as

well as new opportunities for studying the interac-

tions of different groups of people. Two popular top-

ics in the investigation and better understanding of so-

cial networks are community detection and sentiment

analysis. Community detection on the one hand, tries

to analyze a social network with the aim of ﬁnding

groups of associated individuals in it, while sentiment

analysis attempts to determine user emotional behav-

ior and consequently specify their stance and opinion

on various topics, i.e. recognize how individuals feel.

Analyzing the way that users formulate social

communities, the determination of user behavior in

each one of the communities as well as in the whole

social network are fundamental aspects of social net-

work analysis. More speciﬁcally, studying the com-

munity structure of a network leads in explaining so-

cial dynamics of interaction among groups of indi-

viduals and several research works point to this direc-

tion (Deitrick et al., 2013). The accurate discovery

and analysis of communities is a topic of extremely

high research interest with wide range applications.

The economical and marketing implications of com-

munity detection approaches can also be considered

as of extreme importance. The main interest of the

discovery of structured communities and as a matter

of fact, the analysis of each one produced by the com-

munity detection approaches, could improve the ad-

vertising performance of marketing industry by iden-

tifying and targeting the proper group of users in a

speciﬁc network.

The structure of Twitter is formulated and comes

in terms of “follow” and “following” relationships be-

tween the users. Twitter platform gives user the abil-

ity to follow other users they want. In such a case,

each speciﬁc user can receive notiﬁcations regarding

the public posts of the users they follow in real time.

The indication of adding another user as a friend also

results in receiving post notiﬁcations but simultane-

ously indicates a closer relationship between these

two speciﬁc users (Java et al., 2007).

Emotions are essential to all aspects of human

lives and as a matter of fact, social networks can in-

ﬂuence people’s decisions as well as their social rela-

tionships (Wang et al., 2012). Analyzing tweets and

in following recognizing their emotional content, is

a very interesting and challenging topic in the mi-

croblogging area (Choudhury et al., 2012). Hence, it

is necessary for deeper understanding people’s behav-

ior and for providing at the same time a number of in-

dicative factors regarding the public attitude towards

different events and topics. This emotional content

Kanavos, A., Perikos, I., Hatzilygeroudis, I. and Tsakalidis, A.

Integrating User’s Emotional Behavior for Community Detection in Social Networks.

In Proceedings of the 12th International Conference on Web Information Systems and Technologies (WEBIST 2016) - Volume 1, pages 355-362

ISBN: 978-989-758-186-1

355

understanding can describe the emotional status of a

community, a group of people, a city or even a coun-

try (Quercia et al., 2012).

However, most of the existing methodologies for

determining structured communities in a network do

not take into account the aspects regarding users’ be-

havior. Users’ emotional behavior can be considered

as an important parameter that can assist in detecting

better (in terms of density) and more structured com-

munities.

As a matter of fact, in the present manuscript, a

novel methodology for analyzing Twitter social net-

work and in following determining communities in it,

is introduced. This speciﬁc methodology takes into

account each user’s emotional personality and their

activity in the whole network. The methodology ini-

tially analyzes users’ tweets with the aim of deter-

mining each user’s emotional behavior. The emo-

tional behavior of a user is then modeled and speciﬁed

on Ekman’s emotional scale (Ekman, 1999). Ekman

emotion model is a popular categorical model, which

assumes that there is a ﬁnite number of basic and dis-

crete emotions and speciﬁes the following six human

emotions: anger, disgust, fear, happiness, sadness and

surprise. Futhermore, users’ tweets are further ana-

lyzed in order to calculate the inﬂuence metric of each

user in a speciﬁc network, as we introduce a number

of temporal and non temporal characteristics concern-

ing users’ behavior in this speciﬁc network. The aim

of the developed system is to provide a scalable and

distributed approach that allows accurate analysis of

the extracted network and its emerging user commu-

nities in real-time.

The rest of the paper is structured as follows. Sec-

tion 2 presents background topics in sentiment analy-

sis and community detection. Section 3 presents our

methodology followed and the system developed. In

Sections 4 and 5, details of the implementation of

the system as well as the evaluation study conducted

and the results gathered on both the sentiment anal-

ysis topic and the community detection topic are re-

spectively presented. Finally, Section 6 concludes our

work and presents directions for future research.

2 RELATED WORK

Community analysis in social networks has a long

history, which is related to graph clustering algo-

rithms, web searching algorithms, as well as biblio-

metrics. In general, a community is a group of net-

work nodes within which the links connecting nodes

are dense but between which they are space (Yang

et al., 2010). It corresponds to groups of nodes on

a graph or a network that share common properties or

have a common role in the organization and the oper-

ation of the system.

Over the last years, community detection in so-

cial networks has attracted a lot of interest and several

works examine the way users formulate communities

for developing algorithms with structured user com-

munities. A complete overview of approaches and

wide used techniques can be found in (Papadopoulos

et al., 2012), (Plantie and Crampes, 2013).

Concerning communities, the problem that is

known in bibliography, is related to graph partition-

ing. The algorithm proposed in (Girvan and New-

man, 2002) for identifying the edges lying between

communities and their successive removal can be con-

sidered as a breakthrough in this area; a procedure

that after some iterations, leads to the isolation of the

communities. One should also mention techniques

that use modularity, a metric that designates the den-

sity of links inside communities against the density of

links outside communities (Fortunato, 2010), (New-

man, 2004), with the most popular being the algo-

rithm proposed by (Blondel et al., 2008).

Recently, sentiment analysis methods and tech-

niques for recognition of emotions and opinions in

social networks has attracted a lot of interest. An

overview of approaches and methodologies can be

found in (Liu and Zhang, 2012). Several studies point

out the important role that they can play in the anal-

ysis of users’ state as well as in the recognition of

public stance towards speciﬁc topics. There are many

emerging works and applications so as to identify

whether a text is subjective or objective (Barbosa and

Feng, 2010) and also whether an opinion expressed is

positive or negative (Pak and Paroubek, 2010).

In (Xu et al., 2011), authors introduced two meth-

ods for identifying communities with similar senti-

ments with the aim of helping companies in market

segmentation and in the design of marketing strate-

gies. The ﬁrst method assumes that sentiment can

be either positive or negative, whereas in the second

method, the range of sentiment is divided into inter-

vals and in following users are categorized into groups

according to the speciﬁc differences in the ranges of

sentiment values. In (Deitrick et al., 2013), once

community structures have been discovered, authors

use Naive Bayes sentiment classiﬁers trained with

the Sanders dataset towards improving the modular-

ity values.

Despite the increasing signiﬁcance of social me-

dia analysis and the proliferation of methods for de-

tecting communities, most of the techniques rely on

node connectivity, assuming all nodes to be equal and

neglect special characteristics of the nodes. However,

SRIS 2016 - Special Session on Social Recommendation in Information Systems

356

we believe that in social networks, like Twitter, users’

characteristics such as their emotional behavior, are of

predominant importance and could provide vital and

meaningful information regarding the users/nodes of

the network. According to our knowledge, there has

been no previous effort to enhance community detec-

tion techniques with users’ emotional behavior; this is

the ﬁrst work that tries to assist to with user emotional

behavior as defended on Ekman’s psychometric scale.

In this manuscript, the contributions of our work

rely on the following areas: Firstly, an approach for

the automatic analysis of tweets and in following the

determination of each user’s emotional behavior is

presented. Later, we introduce a method for investi-

gating the user’s actions in the social network and cal-

culate their inﬂuence based on their behavior. Finally,

we present an approach to identify the most inﬂuen-

tial communities based on user’s emotional behavior

as well as their analytics proﬁle. Very similar, to the

current manuscript, works are the ones in (Kanavos

et al., 2014b) and (Kanavos and Perikos, 2015).

3 PROPOSED METHOD

In this section, we present the methodology followed

and the system developed to analyze and model con-

versations on speciﬁc topics in Twitter. The method-

ology initially analyzes tweets and determines their

place on Ekman emotional scale (Ekman, 1999).

Then, it estimates user’s inﬂuence in the network

(with the use of user proﬁle) and detects the more in-

ﬂuential communities in the corresponding network.

The produced inﬂuential communities can be seen

as the representation of the emotional interactions in

this network and are utilized based on the emotional

content and tweets as well as the user’s inﬂuence.

The overall architecture of the proposed system is de-

picted in Figure 1.

Figure 1: System Architecture.

3.1 User Emotional Proﬁle Regarding

Tweets

In this subsection, the emotional analysis of the tweets

based on the tool presented in (Perikos and Hatzi-

lygeroudis, 2013) is described. The analysis and the

emotional content of a tweet is conducted on the sen-

tence level as depicted in Figure 2. That is, a tweet

is split into sentences and thereafter each sentence is

handled separately by the tool. The tool recognizes

the existence of the six basic emotions proposed by

Ekman (Ekman, 1999) in natural language sentences.

Figure 2: Architecture of Emotion Recognition Tool.

Initially, on a sentence level, the structure of the

sentence is analyzed using Part-of-Speech (POS) tag-

ging and parsing processes. More speciﬁcally, the

ﬁrst level of analysis concerns the morphosyntactic

analysis conducted by the tree tagger; this tagger is

used in order to specify each word’s grammatical role

in the sentence as well as to determine its base form

(lemma). Then, the Stanford Parser deeper analyzes

the structure of the sentence, speciﬁes the relation-

ships between the words of the sentences and ﬁnally

creates the corresponding dependency tree. The de-

pendency tree represents the grammatical relations

between the words of the sentences in a tree based ap-

proach. Relationships are presented as triplets which

consist of the names of the relation, the governor

and the dependent respectively. When the sentence

morphosyntactic analysis is completed and the depen-

dency tree is created, special parts of the sentence as

well as speciﬁc words are deeper analyzed based on

the knowledge base of the tool.

Moreover, the knowledge base developed, uti-

lizes lexical resources and stores information regard-

ing emotional words that are known to convey spe-

ciﬁc emotional content. These emotional words are

spotted based on WordNet Affect source (Strappar-

ava and Valitutti, 2004), which we extended by man-

ually adding emotional words. For each emotional

word, information regarding its grammatical role as

Integrating User’s Emotional Behavior for Community Detection in Social Networks

357

well as the word’s emotional category are stored. Fur-

thermore, knowledge base handles quantiﬁcation and

negation words and thus can modify the valence and

the strength of emotional words in a sentence when

interacting with them. Examples of such words are

all, none, very, quite, rather, etc. Then, all the emo-

tional words spotted in the sentence are further ana-

lyzed and the relations that denote the exact way they

interact with other words in the sentence are deter-

mined. Speciﬁc types of interactions with quantiﬁca-

tion and negation words are also used so as to estimate

their emotional strength in the sentence. Finally, the

overall emotional status of a tweet is speciﬁed based

on the emotion content of each sentence’s emotional

parts. The output of the tool contains the emotions

that are detected to be conveyed in the user’s speciﬁc

tweet.

3.2 User Analytics Proﬁle and Inﬂuence

Metric

In this subsection we describe the methodology (and

extend the ones from (Kafeza et al., 2014a), (Kanavos

et al., 2014b), (Kanavos et al., 2014a)) for estimating

the importance as well as the inﬂuence of a user in

a Twitter Network. Assuming that our methodology

could be utilized as a graph, then Twitter users would

be represented by nodes. As a matter of fact, the edges

which connect these nodes are the relations of “Fol-

lower to Following”, introduced by Twitter.

Initially, that inﬂuence metric should not depend

merely on the number of “Followers” for each spe-

ciﬁc user, even if that number is big enough and thus

corresponding user’s tweets are received by a large

number of other users (more speciﬁcally by their fol-

lowers). However, this ratio is also not sufﬁcient. An-

other important measurement is the actual number of

posts (Tweets) that a user has addressed. For speci-

fying the importance of this factor, let us see the case

where two Twitter users have nearly the same FtF ra-

tio. Furthermore, as another important factor (with

similar features as the latter one) we have utilized the

Frequency of user’s Tweets, which depicts whether a

user enjoys talking on a regular basis and thus posting

more frequent than other users.

Furthermore, we take into consideration some fea-

tures that deal with interaction between users, i.e. the

number of Retweets and Replies as well as the Clicks,

Favorites and Mentions received. Precisely, concern-

ing the number of Retweets and Replies, they show

that a speciﬁc user enjoys to take part in conversations

either by republishing other users’s posts or comment-

ing on them. Moreover, Retweets can be very helpful

in identifying web trends and content that interests

other users or their followers or simply Tweets that

have the capacity to go viral. In addition, favoriting

is becoming an increasingly popular way to engage

on Twitter. In fact, with a single click, one can show

their appreciation or simply let the author know their

Tweet has been seen.

In our proposed approach, for calculating the

above rates, the latest k tweets of the user are pro-

cessed, according to the Twitter API (e.g. for our

experiments see the following section for values of

k). The proposed Inﬂuence Metric depends on all of

the aforementioned features/metrics of the examined

user, as deﬁned in following Equation 2. Thus, the

Inﬂuence of a user, based on the above parameters is

calculated as follows:

PostImpact = ((Retweets + 1) ∗ (Replies + 1)∗

(Favorites + 1) ∗ (Mentions + 1) ∗ (Clicks + 1))/

DirectTweets (1)

where PostImpact deals with posts met-

rics/characteristics.

In f Metric = log(FtF +1)∗Freq.∗PostImpact

(2)

The above Inﬂuence Metric depends on all of the

aforementioned characteristics of each user. The FtF

ratio is placed inside a base-10 log for avoiding out-

lier values. In addition, the ratio is added by 1 so as

to avoid the metric being equal to 0 in cases that the

value of “Followers” is equal to “Following”. What is

more, we have added the ratio of Retweets, Replies,

Favorites, Mentions and Clicks Received divided by

the absolute number of Direct Tweets. The 5 proposed

ratios are also added by 1 so as to avoid the metric

being equal to 0 in cases that Retweets, Replies, Fa-

vorites, Mentions or Clicks Received, are 0.

3.3 Determining User’s Emotional

Behavior

The accurate assessment of the user’s emotional be-

havior poses two interlinked and mutually related

questions. The ﬁrst concerns the quantity and conse-

quently the frequency of user’s tweets in order to ana-

lyze and determine the user’s emotional status. In ad-

dition, the second deals with the combination as well

as the speciﬁcation of the user’s overall emotional sta-

tus based on each emotional content for every post (in

cases a post can be characterized as emotional).

Initially, in order to answer the ﬁrst question, the

methodology introduced uses a time window of 3

weeks for this process. Speciﬁcally, the methodology

analyzes user’s tweets in the last 3 weeks in order to

SRIS 2016 - Special Session on Social Recommendation in Information Systems

358

determine their emotional status in that period. The

time window of the 3 weeks has been set based on

empirical estimations and evaluation results as well

as on the principle that user’s emotional status can

dynamically change during the passage of time. Set-

ting a too narrow time window, a decent and balanced

amount of the user’s post activity would fail to be pro-

vided. Moreover, a narrow time window could not

be emotionally stable and could rapidly alter in vari-

ous emotional directions. On the other hand, a wider

time window would fail to specify accurately and also

represent meaningfully the alternation of each user’s

emotional status.

User’s tweets in the last 3 weeks are speciﬁed and

in following retrieved by the crawler. Then, all the

tweets are analyzed and emotionally annotated by the

aforementioned process and the tool developed. Af-

ter user’s tweets are analyzed and emotionally anno-

tated, the user’s recent emotional status can be deter-

mined. For each tweet, we can measure and specify

whether it conveys each one of the 6 basic emotions

deﬁned by Ekman emotional tool. In following, the

overall user’s emotional status is calculated based on

the emotional annotation of each one of their tweets

in the speciﬁc time window in a quantity approach.

That is, initially it is determined whether the user has

a vivid emotional status or whether their statuses are

emotionally neutral. More speciﬁcally, a user is char-

acterized to have emotional status/cue if at least 10%

of their posts are recognized as emotional and con-

vey one or more emotions; otherwise, their emotional

status is set to be neutral.

The threshold of the 10% is set based on experi-

ments employed on different Twitter datasets. In gen-

eral, emotions in Twitter posts can vary and show a

highly skewed distribution. In most cases, 10 − 15%

of the posts in the following mentioned datasets, were

recognized so as to convey emotions. Furthermore,

the analysis of the emotional tweets revealed that

emotions such as joy (happiness) can be present in

up to 50% of the emotional tweets while emotions

such as disgust and surprise can be present in less

than 10%. So, the threshold of 10% seems to be

a good choice, thus giving a balanced ratio regard-

ing emotional and neutral users’ annotations/statuses.

Speciﬁcally, emotions such as joy and anger are very

strong ones and are expressed more often by users and

almost always explicitly with the use of emotional

words. In contrast, emotions like surprise and dis-

gust may be expressed more rarely by users and what

is more, they can be implicitly expressed in a user’s

post.

3.4 Communities Decomposition

In our approach, we aim to identify the most inﬂuen-

tial communities in the produced users graph; where

each user proﬁle can be considered as the union of the

two above characteristics, i.e. emotional and analyt-

ics proﬁle. Though several algorithms with modular-

ity based community detection are considered, here

we utilize the one in (Blondel et al., 2008) by adding

an additional transformation as a pre-processing step.

Our inﬂuential community detection approach

combines the modularity optimization of network

community structure with the emotional state of each

user’s retrieved tweets in the graph. We introduce this

information by transforming the retrieved graph to its

dual graph, which is known as line graph. In follow-

ing, the weighted version of modularity community

detection algorithm of (Blondel et al., 2008) is uti-

lized so as to extract the inﬂuential communities in a

ranked list. Finally, we transform again the line graph

to its dual so as to understand the extracted communi-

ties based on the initial retrieved social graph.

Concretely, the methodology is modulated in the

following steps:

1. Transformation to Line Graph, where line graph is

the dual of an initial graph; a dual is the inverted

nodes-edges graph. This transformation is pre-

sented in following Figure 3. Users (e.g. nodes)

are represented by the vector of their tweets emo-

tional scale based on Ekman model (with 1 an

emotion is present while with 0 is not present)

and the edges between them represent the “Fol-

lowing” relationship (they have different labels as

they connect different nodes and it is also neces-

sary for creating the line graph).

Figure 3: Transformation to Line Graph.

2. Utilization of weighted community detection al-

gorithm, that is a method described in (Blondel

et al., 2008) so as to identify communities within

the Twitter network and is based on modularity

optimization.

3. Transformation to initial dual node graph.

Integrating User’s Emotional Behavior for Community Detection in Social Networks

359

4 IMPLEMENTATION

The experimental procedure was based on the Twit-

ter API so as to gather data from Twitter which is

appropriate for our analysis and methodology. The

Twitter4J

constitutes of a Java API, used for col-

lecting tweets which are published in various peri-

ods of time for a variety of topics using correspond-

ing various keywords. Our Twitter data consists of

the following 5 topics (from 4 emotional categories),

where each topic consists of at least 15.000 posts

and the corresponding list of hashtags was compiled

accordingly. The topics studied are Malaysia Air-

lines Flight 370 disappearance, Spectre, Stock mar-

ket, Obamacare and SyrianRefugees.

In order to get an insight regarding users emo-

tional attitude, we calculate the number of Tweets that

express speciﬁc emotional dimensions versus the to-

tal number of Tweets. In Table 1 we observe that ap-

proximately 30% of the posts contain emotional in-

formation. Moreover, in Table 2, the topics studied as

well as their corresponding Ekman emotional scales

are presented.

Table 1: Distribution of Tweets.

Topic Emotional Neutral

Malaysia Airlines Flight

370 disappearance 67% 33%

Spectre 27% 73%

Stock market 35% 65%

Obamacare 31% 67%

SyrianRefugees 43% 57%

In the context of this study, the topics examined

were selected based on the principle to possess diver-

sity in their emotional content. The ﬁve topics are

quite rich in emotions and demonstrate a diversiﬁca-

tion in their distribution. Concretely, the emotional

analysis of the tweets indicates that the happiness is

the predominating emotion in two out of the ﬁve top-

ics, the fear in one topic and the sadness in the remain-

ing two. Indeed, regarding Malaysia Airlines Flight

370 disappearance topic, sadness emotional content is

express in almost 65% of the emotional tweets, while

in happiness has approximately the half percentage in

Spectre and Obamacare topics.

Due to space considerations, the following evalua-

tion (including the corresponding ﬁgures) refers only

to the #Spectre graph. More speciﬁcally, the graph

utilized consists of 1000 nodes, where each user/node

has addressed a post in the above topic.

Twitter4J library: http://twitter4j.org/en/index.html

5 EVALUATION

In the following Figures 4, 5 and 6, we present the

performance of each of our algorithms in determin-

ing the inﬂuential communities. Namely, we rank the

inﬂuence of a community using different metrics for

different application scenarios (see previous work as

well (Kafeza et al., 2014b)).

Figure 4: Comparison of Inﬂuential Community Detection

Approaches based on the percentage of Tweets.

Figure 5: Comparison of Inﬂuential Community Detection

Approaches based on the percentage of Followers.

Figure 6: Comparison of Inﬂuential Community Detection

Approaches based on the percentage of Community Size.

The extracted communities in each case are

ranked based on the Inﬂuence Metric which has been

described above (see Equation 2). Since our moti-

vation stems from the fact that we are interested in

identifying the more inﬂuential communities and not

just the ﬁrst one, our examination is focused on the

ﬁrst 5 ranked communities. In the above Figures 4, 5

and 6, the 14 ranked communities are presented in

SRIS 2016 - Special Session on Social Recommendation in Information Systems

360

Table 2: Topics and corresponding percentages for Ekman emotional scales.

Topic Anger Disgust Fear Happiness Sadness Surprise

Malaysia Airlines Flight 370

disappearance 9% 2% 9% 5% 65% 10%

Spectre 3% 4% 6% 43% 32% 12%

Stock market 5% 9% 42% 22% 19% 3%

Obamacare 6% 8% 15% 49% 9% 13%

SyrianRefugees 23% 4% 19% 2% 47% 5%

which the percentage of Tweets, Followers and Size

for each Community is examined respectively.

More respectively, in Figure 4 we can observe that

inﬂuential communities (regarding a topic or a spe-

ciﬁc time period or an event) based on the emotional

factors produce more Tweets than inﬂuential com-

munities detected only from social network structure

(Blondel et al., 2008).

Figure 5 depicts that the proposed emotional ap-

proach slightly decreases the percentage of Follow-

ers in the top 5 communities as compared to (Blondel

et al., 2008). This occurs due to the Inﬂuential Met-

ric that is more generic and deals with an overall es-

timation of the impact of each user in the produced

community such as the number of Retweets, Replies,

Clicks received, Mentions etc.

As it is obvious in Figure 6, the top 5 communities

in our method requires fewer nodes that the simple

approach. This happens due to the inequality of the

weights distribution in the connected nodes which ef-

fects modularity optimization community detection as

well as to the density of links inside communities as

compared to links between communities. It is noted

that this factor can be useful when cost is associated

with the size of the communities and thus smaller

communities but with larger impact are required.

Table 3: Normalized Metric for Rating Inﬂuential Commu-

nities.

Inﬂuential Communities

Detection

Tweets

/ Size

Followers

/ Size

Simple Community

Detection 1,326 1,640

Emotional Community

Detection 1,587 1,549

The above results are totally consistent with the

Metric/Size metric as Table 3 shows. More speciﬁ-

cally, the results indicate that the detection of com-

munities based on users’ emotional behavior results

in a higher average number of Tweets per Commu-

nity Size. The communities determined are denser

and the higher number of Tweets per Community de-

notes that the formulation of the communities is more

structured and achieved with a ﬁner and more sophis-

ticated approach. In addition, these results support the

rational that users’ emotional behavior can be helpful

and provide meaningful data towards the detection of

inﬂuential communities in Social Networks.

On the other hand, as previously mentioned,

the average number of Followers per Community is

slightly lower when the emotional methodology is

followed. This is mainly a result of the way that In-

ﬂuential Metric is deﬁned as it deals with an overall

estimation of the impact of each user in the produced

community.

6 CONCLUSIONS AND FUTURE

WORK

In this paper, we propose a novel method on identi-

fying inﬂuential communities in a network with the

utilization of the users emotional behavior as well

as users inﬂuence in a speciﬁc timeframe. We ini-

tially present an approach for the automatic analy-

sis of users tweets, then analyze each user comment

and also estimate their emotional behavior. Thereafter

since all users are modelled as emotional or neutral

and are assigned with a speciﬁc inﬂuence metric, our

system ﬁnally identiﬁes the most inﬂuential commu-

nities based on user’s emotional behavior and analyt-

ics proﬁle. The method is based on the emotional con-

tent of each post as well as on an inﬂuence metric of

each user that interacts in a speciﬁc topic. With use of

the Ekman emotional model, we can identify whether

one or more out of the 6 basic human emotions exist

or not.

As future work, it is in our keen interest to in-

vestigate the scalability problems that emerge when

considering bigger graphs. Also, we aim to make

more experiments using several subjects and identify

the parameters that inﬂuence the results of our algo-

rithm in a ﬁner granularity level. Another key as-

pect of our future work will be the extension of the

recognized emotions in our methodology and in fol-

lowing the use of different emotional models such as

the ones in (Ortony et al., 1988). Moreover, this ap-

proach can be introduced in a tool for viral marketing

Integrating User’s Emotional Behavior for Community Detection in Social Networks

361

or for branches’ advertising purposes. In conclusion,

we will examine the evolution of inﬂuential commu-

nities in time, i.e. temporal networks.

REFERENCES

Barbosa, L. and Feng, J. (2010). Robust sentiment detection

on twitter from biased and noisy data. In Proceed-

ings of the 23rd International Conference on Compu-

tational Linguistics: Posters (COLING), pages 36–44.

Blondel, V. D., Guillaume, J., Lambiotte, R., and Lefebvre,

E. (2008). Fast unfolding of community hierarchies

in large networks. Journal of Statistical Mechanics:

Theory and Experiment, P10008.

Choudhury, M. D., Gamon, M., and Counts, S. (2012).

Happy, nervous or surprised? classiﬁcation of human

affective states in social media. In Proceedings of the

6th International AAAI Conference on Weblogs and

Social Media (ICWSM).

Deitrick, W., Valyou, B., Jones, W., Timian, J., and Hu, W.

(2013). Enhancing sentiment analysis on twitter using

community detection. Communications and Network,

5(3):192–197.

Ekman, P. (1999). Basic Emotions. John Wiley & Sons Ltd,

New York.

Ferrara, E. (2012). A large-scale community structure anal-

ysis in facebook. EPJ Data Science, 1(9).

Fortunato, S. (2010). Community detection in graphs.

Physics Reports, 486(3–5):75–174.

Girvan, M. and Newman, M. E. J. (2002). Com-

munity structure in social and biological networks.

Proceedings of the National Academy of Sciences,

99(12):7821–7826.

Gurini, D. F., Gasparetti, F., Micarelli, A., and Sansonetti,

G. (2014). iscur: Interest and sentiment-based com-

munity detection for user recommendation on twitter.

In 22nd International Conference on User Modeling,

Adaptation, and Personalization (UMAP), pages 314–

319.

Java, A., Song, X., Finin, T., and Tseng, B. (2007). Why

we twitter: Understanding microblogging usage and

communities. In Proceedings of the 9th WebKDD and

1st SNA-KDD 2007 Workshop on Web Mining and

Social Network Analysis, WebKDD/SNA-KDD ’07,

pages 56–65.

Kafeza, E., Kanavos, A., Makris, C., and Chiu, D. (2014a).

Identifying personality-based communities in social

networks. In Advances in Conceptual Modeling,

pages 7–13.

Kafeza, E., Kanavos, A., Makris, C., and Vikatos, P.

(2014b). T-pice: Twitter personality based inﬂuen-

tial communities extraction system. In Proceedings of

the 2014 IEEE International Congress on Big Data,

pages 212–219.

Kanavos, A. and Perikos, I. (2015). Towards detecting emo-

tional communities in twitter. In 9th IEEE Interna-

tional Conference on Research Challenges in Infor-

mation Science RCIS, pages 524–525.

Kanavos, A., Perikos, I., Vikatos, P., Hatzilygeroudis, I.,

Makris, C., and Tsakalidis, A. (2014a). Conversation

emotional modeling in social networks. In 24th IEEE

International Conference on Tools with Artiﬁcial In-

telligence (ICTAI), pages 478–484.

Kanavos, A., Perikos, I., Vikatos, P., Hatzilygeroudis, I.,

Makris, C., and Tsakalidis, A. (2014b). Modeling

retweet diffusion using emotional content. In Artiﬁ-

cial Intelligence Applications and Innovations AIAI,

pages 101–110.

Liu, B. and Zhang, L. (2012). A survey of opinion mining

and sentiment analysis. In Mining Text Data, pages

415–463.

Newman, M. E. J. (2004). Fast algorithm for detecting

community structure in networks. Physical Review E,

69(6):066133+.

Ortony, A., Clore, G. L., and Collins, A. (1988). The Cog-

nitive Structure of Emotions. Cambridge University

Press.

Pak, A. and Paroubek, P. (2010). Twitter as a corpus for sen-

timent analysis and opinion mining. In Proceedings of

the International Conference on Language Resources

and Evaluation (LREC).

Papadopoulos, S., Kompatsiaris, Y., Vakali, A., and Spyri-

donos, P. (2012). Community detection in social

media. Data Mining and Knowledge Discovery,

24(3):515–554.

Perikos, I. and Hatzilygeroudis, I. (2013). Recognizing

emotion presence in natural language sentences. In

14th International Conference on Engineering Appli-

cations of Neural Networks (EANN), pages 30–39.

Plantie, M. and Crampes, M. (2013). Survey on social com-

munity detection. In Social Media Retrieval, Com-

puter Communications and Networks, pages 65–85.

Springer.

Quercia, D., Ellis, J., Capra, L., and Crowcroft, J. (2012).

Tracking ”gross community happiness” from tweets.

In Proceedings of the ACM 2012 Conference on Com-

puter Supported Cooperative Work, CSCW ’12, pages

965–968.

Strapparava, C. and Valitutti, A. (2004). Wordnet-affect:

an affective extension of wordnet. In Proceedings of

International Conference on Language Resources and

Evaluation (LREC), volume 4, pages 1083–1086.

Wang, W., Chen, L., Thirunarayan, K., and Sheth, A. P.

(2012). Harnessing twitter ”big data” for automatic

emotion identiﬁcation. In Conference on Privacy, Se-

curity, Risk and Trust (PASSAT) and Confernece on

Social Computing (SocialCom), pages 587–592.

Xu, K., Li, J., and Liao, S. S. (2011). Sentiment community

detection in social networks. In Proceedings of the

iConference, iConference ’11, pages 804–805.

Yang, B., Liu, D., and Liu, J. (2010). Discovering commu-

nities from social networks: Methodologies and appli-

cations. In Handbook of Social Network Technologies

and Applications, pages 331–346.

SRIS 2016 - Special Session on Social Recommendation in Information Systems

362