A Methodology for Identifying Influencers and their
Products Perception on Twitter
Ermelinda Oro, Clara Pizzuti and Massimo Ruffolo
National Research Council of Italy (CNR), Institute for High Performance Computing and Networking (ICAR),
Via Pietro Bucci, 4-11C, 87036 Rende (CS), Italy
Keywords:
Social Network, Social Media, Twitter, Influential User, Twitter Influencer, Product Perception, Social
Network Analysis, Multilayer Networks, Multilinear Algebra, Tensor Decomposition.
Abstract:
The massive amount of information posted by twitterers is attracting growing interest because of the several
applications fields it can be utilized, such as, for instance, e-commerce. In fact, tweets enable users to express
opinions about products and to influence other users. Thus, the identification of social network key influencers
with their products perception and preferences is crucial to enable marketers to apply effective techniques of
viral marketing and recommendation. In this paper, we propose a methodology, based on multilinear algebra,
that combines topological and contextual information to identify the most influential twitterers of specific
topics or products along with their perceptions and opinions about them. Experiments on a real use case
regarding smartphones show the ability of the proposed methodology to find users that are authoritative in the
social network in expressing their views about products and to identify the most relevant products for these
users, along with the opinions they express.
1 INTRODUCTION
Twitter is a microblogging platform that is gaining
impressive popularity because tweets enable users to
express opinions, sentiments, and preferences about
different topics and products in a very concise form.
The availability of the enormous amount of informa-
tion posted by twitterers is attracting growing inte-
rest of both research community and business com-
panies. In fact, the possibility to timely understand
motivations of the popularity of topics or products, al-
ong with opinions expressed by people on them, can
be a valuable help to design more effective promo-
tion campaigns. E-commerce websites apply social
marketing techniques to recommend their products to
customers. In this context, marketers need to better
identify users having the capability of influencing ot-
her users’choices in order to empower product recom-
mendation techniques. So, a crucial research activity
is the identification of social network key influencers
with their products perception and preferences.
Many scholars (Probst et al., 2013; Pawar et al.,
2015) have studied the usage of Twitter contents to
analyze social behavior to create smarter and more
effective real-world applications in the area of viral
marketing and recommendation systems. Existing ap-
proaches aiming at finding influential users mainly
rely on measures based on centrality indexes, com-
puted on the network representing people relations-
hips (Riquelme and Cantergiani, 2016). Several met-
hods are based on concepts of authority and hub sco-
res (Kleinberg, 1999). But, they consider neither if a
user is active on a specific topic or product of interest,
nor her opinion.
In this paper, we propose a methodology capable
to identity influential twitterers along with their per-
ceptions about specific products. The methodology
consists of four main steps: (i) downloading tweets
related to a specific set of products and extracting re-
levant data, (ii) building a multilayer network and ten-
sor model by using extracted data, (iii) identifying in-
fluencers by using the SocialAU algorithm (Oro et al.,
2017), plus (iv) identifying dominant products and re-
lated perceptions. The methodology finds influential
users whose opinion on items of interest can be ex-
ploited by business companies for promoting or mo-
difying their marketing and sales campaigns. Expe-
riments on tweets regarding the smartphone domain
show the ability of the proposed methodology to find
users that are both authoritative in the user network
built from such tweets, and active in expressing their
viewpoint about such products.
Oro, E., Pizzuti, C. and Ruffolo, M.
A Methodology for Identifying Influencers and their Products Perception on Twitter.
DOI: 10.5220/0006675405770584
In Proceedings of the 20th International Conference on Enterprise Information Systems (ICEIS 2018), pages 577-584
ISBN: 978-989-758-298-1
Copyright
c
2019 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
577
The paper is organized as follows. Next section
reports the related work. Section 3 presents the pro-
posed methodology capable to find influential Twitter
users plus their opinions. In Section 4 experiments on
a real-world use case are presented. Finally, Section
5 concludes the paper.
2 RELATED WORK
The interest in methods capable to find influential
users on Twitter is growing exponentially. The know-
ledge of the most relevant actors, in fact, is useful for
many applications, such as viral marketing, informa-
tion propagation, and recommendation systems (Ado-
mavicius and Tuzhilin, 2005). A detailed survey on
the measures proposed in the literature to detect in-
fluencers can be found in (Riquelme and Cantergiani,
2016). In this section, the most recent methods aimed
to identify influential users is reviewed.
Since social networks are represented as graphs,
where nodes represent the users and edges their in-
terpersonal connections, a high number of methods
rely on the centrality measures to detect the most im-
port users. Many methods are based on PageRank
(Brin and Page, 1998) and HIT S (Kleinberg, 1999),
originally introduced to rank web pages. Weng et
al. (Weng et al., 2010) proposed a method, named
TwitterRank, that improves the PageRank algorithm
(Brin and Page, 1998) by automatically identifying
topics that twitterers are interested in. The method
measures twitterer influence by taking into account
the link structure of followers/following of individual
users and the topical similarity between these users.
Cha et al. (Cha et al., 2010) defined three indices:
in-degree, mentions and retweets. The in-degree me-
asure corresponds to the popularity of a user, men-
tions represent the capability of a user to attract ot-
her users in a topic discussion, while retweets express
the importance of the user’s tweet content and mea-
sure the ability of that user to diffuse interesting argu-
ments. An analysis on Twitter users highlighted that
in-degree alone does not generate influence, while the
most mentioned users are celebrities and the most ret-
weeted users are news sites and businessmen. Analo-
gously to Cha et al. (Cha et al., 2010), Anger and Kittl
(Anger and Kittl, 2011) proposed three indices to me-
asure the influence based on the number of retweets
and mentions relative to a user. Thus they define the
concepts of Retweet and Mention Ratio, Interaction
Ratio, and Social Networking Potential. Almgren and
Lee (Almgren and Lee, 2015) proposed a content-
based influence measure, named CIM, that takes into
account the social interactions of users (such as ret-
weets on Twitter or ”like” on Facebook). The authors
build a weighted directed graph, where the nodes are
the users, and the weights represent the number of so-
cial interactions that a node performed on the posts of
another node. The influence is computed considering
the concept of node centrality.
All these approaches rely on graph theory where
connections express interrelationships such as follo-
wer/following or retweet/mention. The methodology
we propose builds the relationships by extracting the
information from the tweet content, thus by analyzing
not only network topology, but also opinions expres-
sed on topics of interest.
3 METHODOLOGY
In this section we present a methodology for detecting
Twitter influencers and dominant products of a dom-
ain along with their perception. It is composed by
four main steps, shown in Figure 1: (i) collection of
tweets related to products of interest and extraction of
relevant data; (ii) building of a model based on mul-
tilayer heterogeneous networks; (iii) identification of
influencers; (iv) identification of dominant products
and related perceptions. The methodology is general
and can be applied to different product types. In the
following we describe the four steps by considering a
use case of tweets talking about mobile phones.
3.1 Tweets Collection and Data
Extraction
As shown in Figure 1, the first step (represented by
a dashed rectangle) is composed by: (i) the down-
load of tweets related to products to be analyzed (e.g.,
smarphones), then (ii) the extraction of interesting in-
formation (such as, users, products, and keywords)
from collected tweets. To download tweets of interest
we use the Twitter API
1
searching for target products
names and hashtags. We download only tweets dea-
ling with the target products, and each tweet is tagged
by the mentioned products. From each tweet, we get
the user name of the author. In order to capture inte-
ractions between other users we extract from the text
re-tweets and mentions. We considered as keywords
expressing opinions both lemmatized form of adjecti-
ves and hashtags that describe products contained in
tweets.
1
Twitter Rest API: https://dev.twitter.com.
ICEIS 2018 - 20th International Conference on Enterprise Information Systems
578
Figure 1: Methodology illustration.
3.2 Multilayer Network Building
After the information from tweets has been extrac-
ted, we build a three-layer network, as shown in Fi-
gure 1 and introduced in (Oro et al., 2017). The
three layers represent users, products, and keywords.
Intra-layer interactions model the types of connecti-
ons among the actors of the same layer, while inter-
layer interactions give the information that a user u
expresses an opinion on a product p by using a ke-
yword k. Formally, the three-layer network is a pair
M = (G, T ), where G = {G
U
, G
P
, G
K
} is a set of
graphs, and where: G
U
= (X
U
, E
U
) is a directed
weighted network representing the n users that are
active on the domain and their connections. Thus
X
U
= {u
1
, . . . , u
n
}, and E
U
= {(u
i
, u
j
) | user u
i
men-
tions user u
j
or retweets u
j
s posts}. G
P
=
(X
P
, E
P
) is the product network representing the set of
m products X
P
= {p
1
, . . . , p
m
} and their similarity re-
lationship E
P
= {(p
i
, p
j
) | sim(p
i
, p
j
)}, i.e. two pro-
ducts are connected if they satisfy a similarity crite-
rion. G
K
= (X
K
, E
K
) is the keywords network repre-
senting adjectives and hashtags that describe products
X
K
= {k
1
, . . . , k
r
} and their ties. E
K
= {(k
i
, k
j
) | a
post where k
i
and k
j
co-occur}, i.e. two keywords
are connected if they both appear in the same tweet.
T is the 3rd-order tensor representing the inter-
layer connections among all the three layers. The cor-
responding n × m × r adjacency tensor X represents
that a user u G
U
expresses z times an opinion on a
product p G
P
by using a keyword k G
K
.
Therefore, we can represent tweets as triples ca-
pable to capture the opinion expressed by users about
selected products. Figure 2 shows an example of
tweet and the information extracted from it: a con-
nection from Ivan to twandroid in the USER network
because of the retweet, the triple (Ivan, Galaxy S5,
pas s
´
ecuris
´
e) in the three-layer network because Ivan
says that the fingerprint sensor of the Galaxy S5 is
vulnerable, i.e. it no safe (pas s
´
ecuris
´
e).
Figure 2: Example of tweet where users, keywords, pro-
ducts, intra- and inter-layers connections are identified. Ex-
tracted information are denoted in red font, dashed arrows
represent extracted intra (e.g., retweets) and inter-layer arcs
(e.g., the triple (Ivan, Galaxy S5, pas s
´
ecuris
´
e).
Figure 3 shows an example of three-layer network
built from the posts published by 10 twitterers re-
garding 4 smartphones by using 8 keywords. The
Figure 3: Example of three-layer network with 10
users, 4 products, and 8 keywords, with triples
(u
3
, G3, thin)=2, (u
1
, Z3, good)=3, (u
1
, S5, good)=1,
(u
2
, Z3, good)=2, (u
2
, Z3, beauti f ul)=1, (u
2
, P7, thin)=1,
(u
2
, G3, beauti f ul)=1, (u
2
, S5, beauti f ul)=2.
thickness of arcs is proportional to the number of con-
nections between the two nodes. For instance, the
tie between user u
1
and user u
2
in the USERS layer
means that u
1
mentioned u
2
or retweeted her posts
many times, while the tie (good, beauti f ul) means
that the two words appear in the same tweet. The
triple (u
1
, Z3, good) means that user u
1
sent several
tweets (note the thickness of the arcs) containing the
keyword good on the product Z3.
A Methodology for Identifying Influencers and their Products Perception on Twitter
579
3.3 Influencers Identification
The third step of the methodology is the identification
of influential users by analyzing the multilayer model
(see Figure 1). To find the most authoritative users
sending tweets regarding particular products, we ex-
ploit the algorithm SocialAU presented in (Oro et al.,
2017). It extends the TOPHITS method, proposed by
Kolda et al. (Kolda et al., 2005), based on multilayer
networks to identify topics and the associated autho-
ritative web pages, by including the scores compu-
ted on each layer by using the HITS (Hypertext Indu-
ced Topic Selection) algorithm proposed by Kleinberg
(Kleinberg, 1999).
The definitions of authority and hub introduced in
(Kleinberg, 1999) can be adapted to users by substitu-
ting the concept of web page with that of user. Thus,
if a user u
1
links to a user u
2
she has conferred autho-
rity on u
2
. In fact, if a user u
1
mentions another user
u
2
or retweets u
2
s tweets, she deems interesting the
contents issued by u
2
, thus she has conferred authority
on u
2
. If a user u
1
links to many authoritative users,
she is said a hub. A good hub is a user that points to
many good authorities; a good authority is a user that
is pointed by many good hubs. The same notions can
be applied also to the products and keywords layers.
However, in such a case the corresponding networks
are undirected, thus the concepts of authority and hub
coincide. Analogously to TOPHITS (Kolda et al.,
2005), SocialAU computes triplets (h, a, w) from the
3-mode tensor T , where h contains the hub scores of
users, a the score of items, and w contains the sco-
res of the keywords. However, differently from TOP-
HITS, the computation takes into account the role of
objects in their own layer. Thus, while computing the
hub and authority scores of a user in the 3rd-way ten-
sor, it considers if the user is also a dominant user
in the proper monolayer network. In such a way the
opinions the user expresses in her tweets are more in-
fluential if she is an authoritative users that also sends
many tweets, i.e. she is a good hub.
Considering the example of three-layer network
shown in the Figure 3, we can see that user u
2
has
many incoming edges in the USERS network, while
u
1
has only outgoing edges, thus though both u
1
and
u
2
expresses several opinions on different items, u
2
is
considered by SocialAU more influential than u
1
be-
cause of the many mentions or retweets received.
3.4 Products Perception Identification
The last step shown in Figure 1 has the objective to
identify the most relevant products for the influential
users, along with the opinion they express. By analy-
zing the sentiment keywords extracted from the tweet
texts, it is possible to elicit the influencer’s opinion on
a product. Therefore, while the user’s score gives the
information if that user is authoritative to diffuse in-
formation, the keywords which represent her opinion,
allow to understand the kind of influence this user can
generate. Positive opinions of influential users can be
exploited by E-commerce websites to improve their
marketing techniques for recommending their pro-
ducts to customers. Our approach can be a valid sup-
port to address key marketing questions like: (a) Who
is talking about our products? (b) Who are the top in-
fluencers and what characterize them? (c) Which pro-
ducts are generating most interest? (d) Which aspects
or opinion are people associating with every product?
(e) Which is the sentiment and overall trend towards
a product?
4 USE CASE
Driven by the needs of a mobile company, which was
trying to expand in France in 2015, we conducted a
study on the main smartphone brands in the French
market by aggregating discussions and social media
information from Twitter. With this real use case,
in this section, we demonstrate the effectiveness of
the presented methodology, based on SocialAU (Oro
et al., 2017), in detecting Twitter influencers and do-
minant products along with their perception.
4.1 Dataset
From May 7th to July 27th 2015, we downloaded
24843 French tweets dealing with 51 smartphones
manufactures and models reported in Table 1. Rows
show the main brands manufacturers and columns
group similar smartphones. Smartphones belong to a
given category on the base of their features like: dis-
play size, average price and photo camera Mpxl. Ca-
tegory Cat. 1 represents top level smartphones having
highest features.
From the downloaded tweets relative to 51
smartphones, we extracted 2706 keywords, which in-
clude adjectives or hashtags. We recognized and ex-
tracted 9783 users that dealt with the smartphone
products (i.e. including authors of tweets, mentio-
ned users, retweeted users). However, many of these
users, even if they posted tweets related to smartp-
hones, did not express any opinion (i.e. their tweets
did not contain adjectives or hashtags). Hence, they
have been eliminated because spawning no useful in-
formation for our goal of analysis. In addition, We
removed users that talk about smarthphones but wit-
ICEIS 2018 - 20th International Conference on Enterprise Information Systems
580
Table 1: Smartphone brands and competitive products.
BRAND CAT. 1 CAT. 2 CAT. 3 CAT. 4 CAT. 5
Alcatel Hero Idol 2s Idol S, Idol 2, Mini S,
Pop S7 Pop S3
HTC One Max One M8 One Mini2
Huawei Ascend Ascend G7 Ascend P7, Ascend Ascend
Mate 7 P8 G620s Y550
LG Flex G3S G3 L80, L90, L Bello F70, L60
Nokia 1320, 1520 830, 930 535, 735 635
Samsung Galaxy Galaxy S5 Galaxy Galaxy Galaxy Ace 4,
Note 4 Alpha, A8 Grand Core 4G
Sony T2 T3 Z3, M2, E3
Z3 compact M2 Aqua
Wiko Gateway, Rainbow 4G, Birdy,
Highway 4G Wax Kite
hout expressing opinions, and 112 users having no
followers or followees because they represent spam-
mers generating the same tweets. Therefore, we kept
4953 interesting users. The user network we gene-
rated contains 9 028 nodes with 9191 arcs. The key-
word network is composed of 2706 nodes and 29 554
arcs. The number of triples of the 3rd-order tensor
9028×51×2 706 are 26673. Statistics of the Smartp-
hone dataset are listed in Table 2. The dataset is pu-
blished for further analysis and comparisons
2
.
Table 2: Statistics of the Smartphone dataset.
#user #smartphone #keywords #tensor
nodes arcs nodes arcs nodes arcs triple arcs
9028 9191 51 268 2706 29 554 26 673
4.2 Influencers
In this section we focus on the authoritative users re-
turned by our algorithm SocialAU and compare them
with those obtained by TOPHITS with respect to well-
know influence measures (Oro et al., 2017), as sum-
marized in the Table 3. For each user, we computed
the value of each influence measure and we used the
relative order of users’ ranks as a measure of diffe-
rence. Users have been sorted by descending order
value of each measure, so that the rank 1 indicates the
most influential user, and increasing rank denotes less
influential users.
To measure the strength of the association bet-
ween two rank sets, we computed the Spearman’s
ρ (Pirie, 1988) and the Kendall’s Tau τ (Conover,
1980) correlation coefficients for both approaches So-
cialAU and TOPHITS with respect to all the defined
evaluation measures. As pointed out by Cha et al.
(Cha et al., 2010), positive correlation values between
two measures mean that users receive similar scores.
Thus, a user considered influential by one measure
because having a high score, is considered influential
2
http://www.lindaoro.com/datasets/twitter.html
Table 3: Meaning of the influence measures used to evaluate
the results.
Notation Meaning
r
F
Followers/Following Ratio compares the amount of users who
have subscribed to the updates of a user u with the number of users
that u is following.
r
ri
Retweet influence Ratio measures the fraction of retweets relative
to a user.
r
mi
Mention influence Ratio measures the fraction of mentions contai-
ning user’s name.
r
RT
Retweet and Mention Ratio enables to detect how many out of the
total tweets of a user u imply a reaction from other users.
r
nRT
Normalized Retweet and Mention Ratio measures r
RT
normalized
with respect to the fraction of tweets posted by the user.
r
I
Interaction Ratio measures how many different individual users
interact with a user.
r
nI
Normalized Interaction Ratio measures r
I
normalized with respect
to the number of followers.
SNP Social Networking Potential represents the potential of interactions
within the network of followers on Twitter.
r
nRMU
User Normalized Retweet and Mention Ratio measures the re-
action of a user u to the tweets of other users. The user’s activity
is normalized with respect to the fraction of tweets posted by u.
r
nIU
User Normalized Interaction Ratio weighs the number of retweets
and mentions posted by a user u with respect to the followees nor-
malized by the maximum number of followees.
UA User Activity measures the percentage of comments a user u sends
on the target products out of the total number of posts.
also by the other measure, i.e. the two influence mea-
sures agree on the role of influencers played by users.
The closer the value to +1 or 1, the stronger the
positive or negative correlations between two measu-
res respectively. We observed that both SocialAU and
TOPHITS have a negative correlation with the Follo-
wers/Followings ratio r
F
. Such a negative correlation
is expected since both approaches do not use the in-
formation on followers/followings to determine influ-
ential users. The normalized interaction ratio has po-
sitive coefficients, higher for SocialAU than for TOP-
HITS. All the other correlation values, except for UA,
are higher for SocialAU than for TOPHITS. In fact,
TOPHITS is based on UA, while SocialAU, besides
UA, exploits also the networks of each layer. Thus,
SocialAU indirectly exploits the number of retweets
and mentions obtained by a user, as well the number
of users that did them. This result points out that So-
cialAU effectively takes advantage of the information
coming from the User network because it gives a hig-
her rank to users that are mentioned more times than
other users, and whose tweets raise interest.
This result is confirmed by the ranking of Twitter
users. The first 50 dominant users sending tweets on
these smartphones obtained by SocialAU and TOP-
HITS are shown in Table 4 and Table 5, respectively.
Other than the relative position assigned by the com-
pared approaches, for each user u indicated by its
A Methodology for Identifying Influencers and their Products Perception on Twitter
581
name we report: Fw, the number of followers of user
u; Fg, the number of users that u follows; NRMU,
the sum of the number of retweets posted by u and of
mentions towards other users; NRM, the sum of num-
ber of retweets and mentions obtained by u; URM, the
sum of the number of users that retweeted or mentio-
ned u; T
u
, the number of three-layer connections the
user u participates.
As can be seen from the Tables 4 and 5, SocialAU
returns as the top user twandroid while TOPHITS gi-
ves kinghousse. The result of TOPHITS is obvious
since this user generated the highest number of three-
layer connections. However kinghousse is an online
sale web site for smartphones and accessories, that
sent a lot of tweets regarding smartphone promotions
that have not been considered interesting by others.
In fact, neither this user receives mentions never it is
retweeted. twandroid, instead, is a very popular An-
droid community giving information about smartpho-
nes and delivering opinions about them. Though the
number of triples it generated in the considered period
is much lower than those of kinghousse, 78 against
870, the number of retweets and mentions, as well as
the number of users performing them, is high (517
and 149 respectively). Thus twandroid can be consi-
dered much more influential than kinghousse since
it has been judged authoritative by other users. Mo-
reover, by considering the enormous number of pe-
ople (73 800 followers) it can reach when broadcas-
ting information, its opinion could sensibly bias the
choices of many users. Looking at the Table 4, which
shows the top 50 Twitter users sorted according to So-
cialAU results, same considerations can be applied to
other users, such as phonandroid, that is ranked 110
by TOPHITS and third by SocialAU, considering the
number of its retweets and mentions, orangejeux,
ranked 2 043 by TOPHITS and 8
th
by SocialAU, that
sent only one tweet generating the reaction from 475
users, or mobileactus, ranked 137 by TOPHITS and
17
th
by SocialAU, a web site giving news on smartp-
hones having 52 200 followers and receiving 109 ret-
weets or mention for its 14 posts regarding smartpho-
nes. Table 5 shows the top 50 users found by TOP-
HITS and the corresponding ranking given by Socia-
lAU. It is worth pointing out that almost all the values
of NMRU, NMR, URM of these authoritative users
are null, while the ranking positions assigned by both
approaches to the same user are close only when T
u
is high, and thus the hub score of users in the ten-
sor dominates the hub and authority scores in the user
network G
U
.
So, the tables highlights the different behavior
of TOPHITS and SocialAU assigning rather different
rankings to those users active in their own network.
Table 4: Top 50 Twitter users according to SocialAU results
and corresponding rank position given by TOPHITS.
SocialAU username F w Fg NRMU NRM URM T
u
TOPHITS
1 twandroid 73800 2444 14 517 149 78 14
2 magikstar29 77 36 231 0 0 168 10
3 phonandroid 27000 1693 3 222 103 43 110
4 kinghousse 930 682 0 0 0 870 1
5 droidtrackr fr 978 378 0 14 10 382 2
6 francois 974 748 1084 40 0 0 42 86
7 guillaumeg516 773 359 0 1 1 518 3
8 orangejeux 13 200 1052 1 479 475 1 2043
9 petitbuzz 3816 4176 23 46 3 455 4
10 frandroidforum 315 55 42 2 1 100 23
11 meilleurmobile 4 446 2 554 4 134 30 62 49
12 bonplanhightech 208 29 0 0 0 473 5
13 arkangelscrap 7371 6963 38 1 1 25 151
14 androidpit fr 2844 94 2 102 21 57 84
15 petitbuzzblog 1 583 1683 34 12 2 322 6
16 rnb 001 1337 744 43 0 0 24 67
17 mobileactus 52200 37 700 0 109 33 14 137
18 echosix8 221 201 25 0 0 20 83
19 rez0 4 185 2003 28 0 0 16 181
20 roxinofr 1751 13 7 65 34 31 81
21 prix discount 232 174 12 13 2 234 7
22 jprenaud78 160 326 27 7 3 26 87
23 wilborie80 25 76 24 0 0 22 63
24 dididogg 25 54 12 0 0 12 96
25 vlantenois 24 151 16 0 0 14 155
26 hichamb143 52 115 11 0 0 11 91
27 willou du80 68 62 18 1 1 18 226
28 creativeoctpus 966 807 0 0 0 137 9
29 prixchocamazon 151 3 0 0 0 307 8
30 mobileovernewz 5 571 4129 0 100 23 77 18
31 dojo l33t 374 563 22 0 0 17 412
32 informatiqueuh 227 29 0 0 0 211 11
33 andoriaofficiel 88 209 13 0 0 10 462
34 kevindecularo 231 66 13 0 0 9 120
35 cllia m 420 289 19 0 0 4 2 071
36 pangocase 248 476 12 0 0 14 270
37 lesbonsplansdun 521 71 0 2 1 165 13
38 stevelumia 31 112 15 0 0 16 150
39 17heures17 171 248 2 1 1 415 30
40 laveilletechno 7993 1716 15 0 0 9 228
41 amevorf32 47 182 13 0 0 10 308
42 ulrichrozier 6419 123 13 0 0 5 250
43 scl002 18 82 6 0 0 12 92
44 julien62162 146 63 23 0 0 32 145
45 figplans 31 32 11 0 0 17 126
46 mickael colin 449 510 8 0 0 7 637
47 trophyhunter35 304 1 701 14 0 0 8 613
48 oyonode 153 36 0 0 0 125 16
49 ericfitteduval 337 162 0 0 0 125 17
50 stephetheve 48 55 11 0 0 9 1 113
SocialAU takes into account, besides the inter-layer
connections, also the number of intra-layer links, that
determines the dominant users and products.
Figure 4 depicts a portion of the user network that
highlights the position of the first fifteen users obtai-
ned by SocialAU. It is important to notice the dense
connections around twandroid, magickstar,
ICEIS 2018 - 20th International Conference on Enterprise Information Systems
582
Table 5: Top 50 Twitter users according to TOPHITS results
and corresponding rank position determined by SocialAU.
TOPHITS username Fw Fg NRMU NRM U RM T
u
SocialAU
1 kinghousse 930 682 0 0 0 870 4
2 droidtrackr fr 978 378 0 14 10 382 5
3 guillaumeg516 773 359 0 1 1 518 7
4 petitbuzz 3 816 4176 23 46 3 455 9
5 bonplanhightech 208 29 0 0 0 473 12
6 petitbuzzblog 1 583 1 683 34 12 2 322 15
7 prix discount 232 174 12 13 2 234 21
8 prixchocamazon 151 3 0 0 0 307 29
9 creativeoctpus 966 807 0 0 0 137 28
10 magikstar29 77 36 231 0 0 168 2
11 informatiqueuh 227 29 0 0 0 211 32
12 soldes du net 15800 3 780 0 0 0 240 69
13 lesbonsplansdun 521 71 0 2 1 165 37
14 twandroid 73800 2444 14 517 149 78 1
15 phon android 594 1 0 0 0 160 66
16 oyonode 153 36 0 0 0 125 48
17 ericfitteduval 337 162 0 0 0 125 49
18 mobileovernewz 5571 4129 0 100 23 77 30
19 techekt 258 31 0 2 1 154 54
20 altetiafrance 26 4 0 0 0 90 61
21 arabianpages 55 500 54100 0 2 2 45 72
22 itnewsfrance 351 1 996 0 0 0 44 73
23 frandroidforum 315 55 42 2 1 100 10
24 ululafr 156 50 0 0 0 50 70
25 appsforward 57 102 0 0 0 47 92
26 bakoblal 92 241 0 0 0 45 75
27 editeuragoogle 133 152 0 0 0 49 77
28 mobilesactu 123 27 0 0 0 143 109
29 boriswapgeek 585 661 0 4 1 46 81
30 17heures17 171 248 2 1 1 415 39
31 ovnizefeed 33 1 0 0 0 41 86
32 achatsengroupes 227 5 0 1 1 418 108
33 shenmueforsony 4011 4156 7 0 0 20 74
34 bonplanlogiciel 201 29 0 0 0 239 121
35 keepmymindfree 643 515 0 4 2 220 155
36 angelapk4 494 1 236 0 0 0 22 123
37 crisapk1 494 1 277 0 0 0 22 124
38 davidapk1 548 1309 0 0 0 22 125
39 messiapk 387 1 068 0 0 0 22 126
40 nawalapk 469 1 156 0 0 0 22 127
41 rihanaapk 560 1353 0 0 0 22 128
42 monphpnet 42 141 0 0 0 32 117
43 neodymeind 1 596 512 0 0 0 39 136
44 doudou87000 27 51 0 1 1 27 99
45 webmasterfree 883 2 001 0 0 0 40 134
46 hassounak 1 386 413 0 0 0 37 135
47 bonplanovernewz 3 028 2847 0 9 5 20 141
48 alitsp 448 2 000 0 0 0 37 138
49 meilleurmobile 4446 2554 4 134 30 62 11
50 planspromos 34 5 0 0 0 139 107
phonandroid, francois 974, orangejeux,
milliermobile, arkangelscrap, androidpit fr,
scored 1
st
, 2
nd
, 3
rd
, 6
th
, 8
th
, 11
th
, 13
rd
, 14
th
and by
TOPHITS 14
th
, 10
th
, 110
th
, 86
th
, 2043
rd
, 49
th
, 151
st
,
84
th
, respectively. Moreover, it is possible to see
that kinghousse and bonplanhightech, scored first
and fifth by TOPHITS are isolated nodes in the user
network. The Figure 4 visually confirms the ability
of SocialAU in finding users well connected in their
own network sending opinions about smartphones.
Figure 4: A portion of the user network of the Smartphone
dataset showing the position of the first fifteen users obtai-
ned by SocialAU and reported in Table 4. The connections
of twandroid are colored in red.
4.3 Products Perception
While the knowledge of the user’s score gives us
the information if that user is authoritative to dif-
fuse information, the keywords used by such user,
which represent its opinion, allows us to understand
the kind of influence this user can generate. Ta-
ble 6 shows, for the first four dominant smartpho-
nes, the first five dominant users posting tweets on
them, along with the adjectives used. From the ta-
ble we can notice that the opinions of twandroid,
magikstar and doidtrack fr are mainly positive
for Sony Experia Z3, Samsung Galaxy S5 and LG
G3, while as regards Sony Experia Z3 Compact,
the sentiment words are less frequent. phonandroid
and kinghousse, for instance, post tweets containing
mainly hashtags (not reported in the table).
5 CONCLUSIONS
We presented a methodology able to identify the most
influential twitterers along with their perceptions and
opinions about specific products. We conducted ex-
periments on a real use case regarding smartphones
and we demonstrated the effectiveness of the propo-
sed methodology. This methodology is general and
can be applied to support the analysis and marketing
strategies related to any type of products. By exploi-
ting this methodology, marketers are able to disco-
ver social key factors hidden in Twitter, such as: who
A Methodology for Identifying Influencers and their Products Perception on Twitter
583
Table 6: Most popular smartphone models in the dataset
with dominant users and adjectives used to express a judg-
ment on each smartphone.
smartphoneuser keywords
Sony twandroid good, thin, comparative, immediate
Xperia Z3 magikstar29 beautiful, efficient, comparative, new, different,
the best, compact, thin, anti-overheating, advan-
tageous, good
phonandroid disponible,internationale,excessive,new,fast
kinghousse portable
droidtrackr fr beautiful, efficient, top, compact, good, top, bur-
ning, new, thin, the best, fast
Samsung twandroid good, vulnerable, attractive
Galaxy S5 magikstar29 compatible, huge, new, expensive, good, at-
tractive, superior, vulnerable
phonandroid good, superior, successful
kinghousse mini, rigid, expensive, portable, light, fine
droidtrackr fr compatible,good,vulnerable,attractive,superior
Sony twandroid good, bluetooth
Xperia Z3 magikstar29 compact, big
Compact phonandroid
kinghousse
droidtrackr fr compact, good
LG G3 twandroid good, available
magikstar29 the best, new, excellent, big, different, good
phonandroid small, good
kinghousse rigid, portable
droidtrackr fr the best, good, hard, super
are the top influencers and what characterizes them,
which products are generating more interest, which
opinions are associated with each product by the aut-
horitative users etc. These factors can help in the im-
plementation of effective techniques of viral marke-
ting and recommender systems.
The algorithms applied in the various steps of the
methodology can be extended and improved. For
instance, the product perception is based on opini-
ons expressed through simple extracted adjectives and
hashtags. In the future, we intend to apply more accu-
rate techniques of sentiment analysis and opinion mi-
ning (Medhat et al., 2014). In addition, we would
recognize accurate entity-targets of the opinions, i.e.,
not just the products but also features of them. In this
case, we could extend our model as a four-layer net-
work to analyze, for instance, which are the most re-
quired features discussed from users.
REFERENCES
Adomavicius, G. and Tuzhilin, A. (2005). Toward the next
generation of recommender systems: A survey of the
state-of-the-art and possible extensions. IEEE Trans.
on Knowl. and Data Eng., 17(6):734–749.
Almgren, K. and Lee, J. (2015). Who influences whom:
Content-based approach for predicting influential
users in social networks. In Proceedings of the Inter-
national Conference on Advances in Big Data Analy-
tics, pages 89–95.
Anger, I. and Kittl, C. (2011). Measuring influence on twit-
ter. In Proceedings of the 11th International Con-
ference on Knowledge Management and Knowledge
Technologies, pages 31–34. ACM.
Brin, S. and Page, L. (1998). The anatomy of a large-scale
hypertextual web search engine. Comput. Netw. ISDN
Syst., 30(1-7):107–117.
Cha, M., Haddadi, H., Benevenuto, F., and Gummadi, K.
(2010). Measuring user influence in twitter: The mil-
lion follower fallacy. In 4th International AAAI Confe-
rence on Weblogs and Social Media (ICWSM), pages
10–17.
Conover, W. (1980). Practical Non-Parametric Statistics,
2nd edn. John Wiley and Sons, New York.
Kleinberg, J. M. (1999). Authoritative sources in hyperlin-
ked environment. Journal of the ACM, 46(5):604–632.
Kolda, T. G., Bader, B. W., and Kenny, J. P. (2005). Higher-
order web link analysis using multilinear algebra. In
Proc. of the 5th International Conference on Data Mi-
ning (ICDM’05), pages 242–249.
Medhat, W., Hassan, A., and Korashy, H. (2014). Sentiment
analysis algorithms and applications: A survey. Ain
Shams Engineering Journal, 5(4):1093–1113.
Oro, E., Pizzuti, C., Procopio, N., and Ruffolo, M. (2017).
Detecting topic authoritative social media users: a
multilayer network approach. IEEE Transactions on
Multimedia.
Pawar, K. K., Shrishrimal, P. P., and Deshmukh, R. R.
(2015). Twitter sentiment analysis: A review. Interna-
tional Journal of Scientific and Engineering Research,
6(4):957–964.
Pirie, W. (1988). Spearman rank correlation coefficient. En-
cyclopedia of statistical sciences.
Probst, F., Grosswiele, L., and Pfleger, R. (2013). Who will
lead and who will follow: Identifying influential users
in online social networks - A critical review and future
research directions. Business & Information Systems
Engineering, 5(3):179–193.
Riquelme, F. and Cantergiani, P. G. (2016). Measuring user
influence on twitter: A survey. Inf. Process. Manage.,
52(5):949–975.
Weng, J., Lim, E.-P., Jiang, J., and He, Q. (2010). Twit-
terrank: Finding topic-sensitive influential twitterers.
In Proceedings of the Third ACM International Con-
ference on Web Search and Data Mining, WSDM ’10,
pages 261–270.
ICEIS 2018 - 20th International Conference on Enterprise Information Systems
584