Investigating negative reviews and detecting negative influencers in Yelp through a multi-dimensional social network based model

Abstract

In this paper, we propose an investigation of negative reviews and define the profile of negative influencers in Yelp. The methodology adopted to achieve this goal consists of two phases. The first one is theoretical and aims at defining a multi-dimensional social network based model of Yelp, three stereotypes of Yelp users, and a network based model to represent negative reviewers and their relationships. The second phase is experimental and consists in the definition of five hypotheses on negative reviews and reviewers in Yelp and their verification through an extensive data analysis campaign. This was performed on Yelp data represented by means of the models introduced during the first phase. Its most important result is the construction of the profile of negative influencers in Yelp. The main novelties of this paper are: (i) the definition of the two social network based models of Yelp and its users; (ii) the definition of three stereotypes of Yelp users and their characteristics; (iii) the construction of the profile of negative influencers in Yelp.

Full text

Investigating negative reviews and detecting negative influencers in
Yelp through a multi-dimensional social network based model
Enrico Corradini1, Antonino Nocera2, Domenico Ursino1∗, and Luca Virgili1
1DII, Polytechnic University of Marche,
2DIII, University of Pavia
∗Contact Author
e.corradini@pm.univpm.it; antonino.nocera@unipv.it; d.ursino@univpm.it;
l.virgili@pm.univpm.it
Abstract
In this paper, we propose an investigation of negative reviews and define the profile of negative
influencers in Yelp. The methodology adopted to achieve this goal consists of two phases. The
first one is theoretical and aims at defining a multi-dimensional social network based model of
Yelp, three stereotypes of Yelp users, and a network based model to represent negative reviewers
and their relationships. The second phase is experimental and consists in the definition of five
hypotheses on negative reviews and reviewers in Yelp and their verification through an extensive
data analysis campaign. This was performed on Yelp data represented by means of the models
introduced during the first phase. Its most important result is the construction of the profile of
negative influencers in Yelp. The main novelties of this paper are: (i) the definition of the two
social network based models of Yelp and its users; (ii) the definition of three stereotypes of Yelp
users and their characteristics; (iii) the construction of the profile of negative influencers in Yelp.
Keywords: Yelp; Multi-dimensional model; Negative influencers; Negative reviews; Social Net-
work Analysis; User stereotypes; Homophily
1 Introduction
Yelp1is a business directory service and a crowd-sourced platform designed to help users find businesses
like restaurants, hotels, pet stores, spas, and many more. It is one of the most widely used review
platforms on the Web. It ranks 9th on the RankRanger list of the top 100 leading websites by traffic2,
with approximately 800 million visits per month. In addition of being a business search and review
platform, Yelp is also a social network, because it allows its users to specify their friendships. Finally,
it is also a business directory, because it groups businesses into categories and sub-categories.
1https://www.yelp.com
2https://www.rankranger.com/top-websites
1

The success of Yelp has prompted many researchers to investigate this platform [1, 5, 48, 59, 31].
Several studies have striven to understand how rates are assigned to businesses [34, 45, 72, 71], and
many others have focused on the analysis of the content of text reviews from both a structural and
a linguistic viewpoint [60, 61, 8]. Some papers have studied Yelp reviews by adopting sentiment
analysis-based techniques [57, 4, 30]. Others have focused on identifying strategies for the detection
of fake reviews and rates [51, 56, 54, 44], or have investigated how people search for information
[32]. Furthermore, some authors have investigated Yelp through the concepts provided by Social
Network Analysis, like homophily [55], to study the social influence existing among friends [65]. Some
researchers have employed these results to outline the decision making of users on purchases [78],
while other ones have studied the possible impacts of electronic Word of Mouth (eWOM) in online
businesses [22, 23, 2]. Several studies have explored the causes leading people to publish reviews [33],
while others have analyzed reviewer strategies to improve their effectiveness [27, 69, 47]. Finally, some
studies have focused on the analysis of review usefulness [76, 67, 43], while others have investigated
the differences between positive and negative reviews [77, 42].
A phenomenon that represents a hot topic for both Yelp and all review platforms is the analysis
of negative reviews [9]. This topic is extremely important not only for the consequences it has in
practice, but also from a more theoretical point of view. In fact, it is well known that the Likert scale,
which the Yelp reviews and the corresponding scores are based on, is positively biased [3, 62, 12]. As
a consequence, the presence of negative reviews is a really important problem indicator for a business
and, consequently, a valuable piece of information [43, 47]. Indeed, negative reviews can provide much
more information, knowledge and improvement possibilities than positive ones [21]. For this reason,
many researchers have already investigated the role of ratings and reviews on businesses, along with
their social implications [73, 50].
Despite the numerous studies on Yelp that have been presented in the past literature, to the best
of our knowledge, no paper has proposed a multi-dimensional model capable of best capturing the
specificity of Yelp to be at the same time a review platform, a social network and a business directory.
Moreover, no paper has proposed a study focused entirely on negative reviews on Yelp that, starting
from a representative model of them, could define several stereotypes of users and, hence, build the
profile of negative influencers. This paper aims at filling this gap.
Specifically, we first define a multi-dimensional social network based model for Yelp and then use
this model to study negative reviews and build a profile of negative influencers in this social medium.
We decided to adopt this model because it perfectly fits the specificities of Yelp mentioned above. In
fact, our model represents Yelp as a set of 22 communities, one for each macro-category of this social
platform (modeling Yelp as a business directory). At the same time, it represents Yelp as a social
network, whose nodes indicate users and whose arcs denote the relationships between them. These
can be of different types. For example, they can denote friendships between users (modeling Yelp as a
social network), or the action of co-reviewing the same business (modeling Yelp as a review platform).
Through the concepts and techniques of Social Network Analysis applied to our multi-dimensional
model, our approach defines three stereotypes of Yelp users, namely the bridges, the double-life users
and the power users. These stereotypes can help the detection of the negative influencers in Yelp and
the definition of a profile for them. Both our model and the user stereotypes represent theoretical
contributions of our paper. These last are completed by a Negative Reviewer Network, which allows
2

us to investigate the main characteristics of the negative influencers in Yelp.
Among the possible questions that can be answered thanks to our approach, in this paper we focus
on the following ones: (i) What about the dynamics leading a Yelp user to publish a negative review?
(ii) How can the interaction of these dynamics increase the “power” of negative reviews and people
making them? (iii) Who are the negative influencers in Yelp?
The practical implications of negative reviews and influencers we find in this paper have a large
variety of applications. First of all, it was proved that negative reviews have a stronger effect on busi-
nesses than positive ones [2]. Furthermore, influencers play a crucial role for the successful placement
of products in a social network. So, it is important to know who are the negative influencers that
could damage a business, in order to strive to turn them into neutral, or even positive, influencers
[77, 78]. Finally, gaining trust through online reviews can help a business gather venture capitals for
its growth [26, 43]. As a matter of fact, reviews are consumer opinions, unfiltered by traditional media,
more sincere and imperfect [2, 22]. For this reason, a proper coverage of positive reviews can attract
more financiers [2, 23, 42]. On the other hand, negative reviews and influencers can drive potential
investors away from investing in a company [52].
The outline of this paper is as follows. In Section 2, we present related literature and highlight the
main novelties of our approach with respect to the past ones. In Section 3, we describe the theoretical
background and hypotheses development. In Section 4, we present the methodology we adopted during
the investigation activity. In Section 5, we illustrate the results obtained. In Section 6, we propose a
discussion and a synthesis of them, their implications, and possible future research directions. Finally,
in Section 7, we draw our conclusions.
2 Related Literature
Over the years, researchers have focused on Yelp as a reference platform for studying how users
interact with each other and build cooperative social groups. Their research efforts have also been
supported by the social medium itself, which has made available a complete snapshot of its data to
foster comprehensive analyses on it [24]. Many authors have used this snapshot to investigate the
role of ratings and reviews on businesses and their social implications [73, 50]. Researchers have also
analyzed how people search for information on Yelp [32] and what aspects (including uses and rewards)
lead them to employ this platform.
Several authors have investigated Yelp using Social Network Analysis (SNA, for short) [64, 65]. For
instance, the authors of [65] rely on the concept of homophily [55] to study the social influence possibly
existing between users and, in particular, between friends. Starting from the results obtained, they
propose the construction of the profile of an influencer in Yelp. The authors of [64] focus on the role of
friendship in this social medium. Specifically, they investigate the impact of social relationships from
the consumer’s side and find that these relationships exert a significant impact in those consumers
having at least one common purchase.
As for the analysis of social relationships, several studies have been conducted in both Yelp and
other social platforms to understand how users perceive their social contacts and how they influence
their acquaintances [48, 59, 31, 58, 69, 36, 81, 79]. For example, the authors of [58] propose an approach
to analyze a large set of brand associations obtained from social tags for marketing research. They
3

apply well-known text mining techniques to understand consumers’ perceptions of brands starting
from social tagging data. The authors of [22] analyze a dataset obtained from OpenRice.com, a
crowd-sourced social medium for restaurant reviews in Hong Kong and Macau. The authors of [27]
show that online community members rate reviews containing descriptive identity information more
positively. Indeed, a disclosure of personal information on an online review system leads to a greater
volume of sales. The authors of [69] aim at understanding how online reviewers compete to acquire
the attention, typically scarce, of users. They propose a theory explaining the strategies adopted by
online reviewers in choosing the right product and the right rate when posting reviews. As far as
Yelp is concerned, the authors of [48] investigate the effects of the review rate, the reviewer profile,
and the receiver familiarity with the platform, on the credibility of a review on this social medium.
Moreover, the authors of [59] find a strong correlation between the moral attitude of a community
of users and their tendency to express low rates and negative reviews in case some moral foundation
is violated. As for the investigations of social relationships in social media, another interesting topic
concerns information diffusion [6, 74, 41, 14, 49]. In the analysis of this topic, an increasing number
of researchers are studying the role not only of classic and direct relationships, such as friendship, but
also several other ones, such as co-posting or homophily of interests (i.e., having interest in the same
topics) [66, 13].
In all previous approaches, the reviews considered are general (i.e., they could be positive or
negative). However, to our end, negative reviews and reviewers are worth a special attention. The
importance of negative reviews in the analysis of social platforms has been investigated in the recent
scientific literature by highlighting their impact in social contexts, along with the mechanisms leading
users to make them [57, 26, 68, 5, 1]. In these studies, researchers point out that dealing with negative
reviews is a fundamental task in review-based platforms for business operators [43, 47]. In fact, it
was empirically shown that answers and justifications to negative rates contribute to the increase
of trust between users and businesses [26], and that users tend to perceive reviews confirming their
initial beliefs as more helpful [77]. Several studies focus on the key factors making a review helpful
[67, 26], while others show that negative reviews are more useful and can influence user opinions more
than positive ones [7, 20]. In this perspective, the authors of [78] propose a model to identify the key
elements leading customers to make their decisions; this model was empirically tested with 191 users
of an existing online review site. Furthermore, the authors of [2] use the VentureExpert database
to gain knowledge on a sample of famous businesses. The authors of [33] formalize a metric, called
disconfirmation, measuring the discrepancy between the expected evaluation of a product and the one
assigned by experts or other people. The authors of [26] study a set of variables to evaluate the users’
intention of employing Yelp, as well as their behavior in using a service or purchasing a product after
reading Yelp reviews. Finally, the authors of [5] analyze the reviews made by hospital patients in
order to identify a common language correlated with negative and positive reviews.
An important aspect to consider when using Social Network Analysis for evaluating reviews and
reviewers is the fact that user relationships in a social network are often heterogeneous [19]. For this
reason, many studies have proposed to decompose social media into different networks of relationships.
Indeed, multi-relationship networks have been extensively studied in the past [25, 75, 80]. For example,
the authors of [80] combine the analysis of the friendship network and the author-topic one, both
constructed starting from the information available in an online social network. Instead, the authors of
4

[75] focus on a co-authorship network and consider different types of relationships, i.e., co-authorship,
co-participation to the same edition of a conference, and geographic proximity.
In multi-relationship networks, the classical definition of influencer is extended because the role
of such users is not bound to communities derived from a single category of relationships. Instead,
it also includes the capability of providing information diffusion channels among different networks,
one for each type of relationships. To refer to this extended definition of influencer, the term “bridge”
is often adopted. In the past literature, several studies have been devoted to investigate the role of
bridges in the formation of social communities. For instance, the authors of [38] show that users
with a weak connection bridging heterogeneous groups have higher levels of community commitment,
civic interest, and collective attention than the other ones. Furthermore, the authors of [29] prove
that Internet users, who bridge heterogeneous online communities by means of weak ties, have a high
social engagement, use the Internet for social purposes, and are prone to become members of new
social communities. The interest towards users serving as bridges among communities has increased
over the years and, indeed, several studies have been done to analyze the behavior and peculiarities
of such users in complex networks [28, 70, 46, 10, 11].
Some studies have also analyzed the behavior of users serving as bridges among different social
networks [15, 18, 16]. Here, the concept of community is brought to the edge, because it is mapped to
a whole social network. Specifically, the authors of [15] report a complete identikit of users bridging
different social networks. The authors of [18] leverage the peculiarities of bridge users to define a new
crawling strategy to sample a multi-social network environment. Finally, the authors of [16] perform a
comparative study of users serving as bridges among two of the most famous social networks, namely
Facebook and Twitter.
From the above description, it can be seen that, in the literature, there is an impressive number of
papers dealing with issues similar to those analyzed in this paper. However, none of them proposed a
multi-dimensional social network based model for Yelp, capable of representing the specificity of this
social platform of being simultaneously a review platform, a social network and a business directory.
The presence of this model would allow us to answer the following research question: What about the
dynamics leading a Yelp user to publish a negative review? Furthermore, no paper proposed a study
focused entirely on negative reviews and reviewers in Yelp, which, starting from a social network based
model representing them, could define a set of stereotypes of users publishing negative reviews. Having
all this available would allow us to answer the following research question: How can the interaction of
the dynamics driving negative reviewers increase their “power” and the one of their reviews? Finally,
no past paper built a profile of a negative influencer in Yelp. Reaching this result would allow us to
answer the following research question: Who are the negative influencers in Yelp? This paper aims at
filling this gap and answer the three research questions mentioned above.
Our paper draws inspiration from the research strands mentioned previously. First of all, our
multi-dimensional social network based model of Yelp can be employed to handle different relationships
(e.g., friendship, co-review). In particular, it is possible to define an occurrence of the model for each
relationship. This way of proceeding falls within the context of multi-relationship networks, but in
a new way. In fact, differently from past multi-relationship models, ours does not require the prior
and static definition of the relationships to represent, but allows a dynamic choice of them, based
on the analysis to be performed. For example, in this paper, we have chosen friendship and co-
5

review between Yelp users. Furthermore, the choice of including in our model the macro-categories
in which the businesses are grouped in Yelp represents an additional feature of it. It makes possible
a definition of the bridge concept perfectly fitted on Yelp, which, in turn, allows for the definition
of three user stereotypes for this social platform. Therefore, the multi-dimensionality of our model
enables an analysis of Yelp users and their relationships from multiple orthogonal viewpoints, acting
simultaneously and influencing each other.
Our multi-dimensional social network based model makes our definition of bridge possible. Starting
from that definition, and operating on the model itself, we define three user stereotypes, namely: (i)
the k-bridge, i.e., a person who reviewed businesses belonging to kdifferent Yelp macro-categories;
(ii) the power user, i.e., a person very active in all the macro-categories in which she is interested;
(iii) the double-life user, i.e., a person exhibiting different behaviors in the different macro-categories
in which she operates. Compared to the generic stereotypes presented in the past literature [17], those
identified in this paper are tailored to Yelp and, therefore, can provide a more specific contribution in
the definition of the profile of negative influencers in this social medium.
Having the multi-dimensional model, the three stereotypes and the Negative Reviewer Network
at disposal, our approach can investigate negative reviews and reviewers and can build a profile of
negative influencers. These tasks are very important because it was shown that the effect of negative
reviews and reviewers is much greater than the one of positive reviews and reviewers [2]. Furthermore,
negative reviews and reviewers are not very common because people tend to give high ratings to
businesses [12, 63]. But for this very reason, the information they bring is extremely valuable. Indeed,
consumers and businesses are prone to rely on negative reviews and reviewers to understand the
reasons for possible dissatisfaction caused by a product, a service or a business [5, 1].
Compared to the works on negative reviews and reviewers described above, our approach is more
focused on the issue of influence, more specifically on negative influence. In this context, it offers a
first important contribution thanks to the definition of the Negative Reviewer Network. This tool
allows the exploitation of Social Network Analysis techniques to investigate the influence of a negative
reviewer on other users. We point out that the Negative Reviewer Network is general and can be used
to investigate the same issue in other review platforms. Starting from it and the multi-dimensional
model introduced in this paper, which is instead specific to Yelp, our approach provides a second
important contribution, i.e., it constructs the profile of a negative influencer in Yelp. Such a profile is
perfectly fitted on this social platform because it takes into account both the partitioning of Yelp into
macro-categories and the possibility to specify user friendships, provided by this platform.
3 Theoretical background and hypothesis development
Our multi-dimensional investigation of negative reviews and detection of negative influencers in Yelp
is possible thanks to a new multi-dimensional social network based model of Yelp. This model starts
from the observation that, in this social medium, businesses are organized according to a taxonomy
consisting of four levels. Level 0 includes 22 macro-categories. Each macro-category has one or more
child categories; therefore, level 1 includes 1002 categories. A category may have zero, one or more
sub-categories; as a consequence, level 2 comprises 532 sub-categories. Finally, level 3, has only 19
sub-sub-categories; indeed, most sub-categories are not further categorized. Our model represents
6

Yelp as a set of 22 communities, one for each macro-category:
Y={C1,C2,· · · ,C22}
Given the macro-category Ci, 1 ≤i≤22, a corresponding user network Ui=hNi, Aiican be
defined. Niis the set of the nodes of Ui; there is a node nipfor each user uipwho reviewed at least
one business of Ci.Aiis the set of the arcs of Ui; there is an arc apq = (nip, niq)∈Aiif there exists a
relationship between the users uip, corresponding to nip, and uiq, corresponding to niq.
Finally, an overall user network U=hN , Aicorresponding to Ycan be defined. There is a node
ni∈Nfor each Yelp user. There is an arc apq = (np, nq)∈Aif there exists a relationship between
the users up, corresponding to np, and uq, corresponding to nq.
In the definition of U(and, consequently, of Ui), we do not specify the kind of relationship between
upand uq. Actually, it is possible to define a specialization of Ufor each relationship we want to
investigate. In this paper, we are interested in two relationships existing between Yelp users, namely
friendship and co-review. As a consequence, we define two specializations of U, namely Ufand Ucr .Uf
is the specialization of Uwhen we consider friendship as the relationship between users, whereas Ucr
denotes the specialization of Uwhen co-review (i.e., reviewing the same business) is the relationship
between users.
Starting from this model, it is possible to define some Yelp stereotypes, namely: (i) the k-bridge,
i.e., a person operating in kcategories of Yelp; (ii) the power user, i.e., a person very active in all the
categories that she is interested in; (iii) the double-life user, i.e., a person showing different behaviors
in the different categories she attends. Her different behaviors can regard the activity level (access-
dl-user) or the severity of her reviews (score-dl-user). These stereotypes can lead to the detection of
negative influencers in Yelp. We formalize them in Section 4. We have introduced them here in that
their concepts are necessary to understand the following of this section.
Starting from this theoretical background, we aim at answering the three questions mentioned in
the Introduction. In particular, we use the above model and stereotypes to design and perform a
social network analysis-based campaign aiming at evaluating some hypotheses that we synthesize in
the following:
First of all, the review mechanism of Yelp is based on a scale from 1 to 5 stars. This is similar to
the review mechanisms encountered in several other social media. In this context, we formulate
the following:
Hypothesis 1 (H1) - The star-based review system of Yelp is positively biased.
In the scale adopted by Yelp, 1 means “absolutely bad” and 5 means “fantastic”. A review
with 2 stars is still negative, but 3 stars already denote a positive review. In other words, the
review mechanism of Yelp makes it more probable that users release positive reviews. Unless
the experience was really bad, the review will almost always be positive. This is confirmed by
how Yelp itself labels the stars (1 - “Eek! Methinks not”; 2 - “Meh. I’ve experienced better”; 3
- “A-OK”; 4 - “Yay! I’m a fan”; 5 - “Woohoo! As good as it gets!”).
7

On the other hand, if we consider this review mechanism from a more formal and theoretical
viewpoint, we can observe that it is based on a Likert scale, which was already shown to be
asymmetric and positively biased [3, 62, 12].
We think that the stereotypes introduced above can help very much in evaluating negative
reviews and influencers. As for a specific kind of stereotype, i.e., the double-life users, we
formulate the following:
Hypothesis 2 (H2) - access-dl-users and score-dl-users play a key role in negative reviews.
To understand the reasoning behind this hypothesis, consider score-dl-users. Clearly, they can
be partitioned into two sets. The former is made up of users who mainly write positive reviews
and few negative reviews. These are basically positive users who, for some reasons, had a bad
experience with some businesses. So, what drove them to write negative reviews, considering
that they are keen to write positive ones? A user assigns a 1-star score to a business when her
expectations were not satisfied. This was already investigated in literature (see, for instance,
[33]), where it was proved that a high discrepancy between the others’ opinions and the experience
of a user is the main driver for her to write a negative review.
The latter set of access-dl-users is much more peculiar. It comprises those users who generally
write negative reviews but, in some cases, release positive ones. These users have probably
developed very severe criteria for evaluating businesses, leading them to be satisfied only rarely.
We have already discussed about the multi-dimensionality of our model. One of its main di-
mensions is friendship. Actually, it is well known that this relationship plays a key role in social
networks [14, 66, 13]. Starting from these results, it is reasonable to formulate the following:
Hypothesis 3 (H3) - A user has a strong influence on her friends when doing negative
reviews.
This could seem obvious. In past literature it has been proved that users are influenced by others
when writing reviews. In particular, it has been found that users tend to have a positive opinion
of a product/service if it has been positively commented by other users [22].
In addition, people generally trust more those users sharing their personal profile on online
review platforms [27]. It was found that a personal information disclosure is crucial for the
spread of positive comments about a product/service, because the possibility of associating
information with a particular person gives a boost in the overall perceived confidence. All of
this is amplified when users share a common geographical location. This reasoning can also
be applied to relationships like friendship, because personal information is certainly disclosed
between friends.
Here, we hypothesize that the influence exerted by friends is valid not only for positive reviews
but also for negative ones, possibly leading to a phenomenon of negative influence between
friends.
8

Another stereotype introduced above that could play an important role as negative influencer is
the bridge one. As for it, we formulate the following:
Hypothesis 4 (H4) - Bridges have a much greater influence power than non-bridges.
If Yelp can be modeled as a network of different communities, each corresponding to a given
business macro-category, it is immediate to think of bridge users as special ones, capable of
facilitating information diffusion from a community to another. Bridge users have a position of
power in the network, and this power can even be measured [40]. If we look at classical centrality
measures in social network analysis, it is easy to argue that bridge users have a high betweenness
centrality value. On the other hand, if we look at reviews, it is plausible that a bridge could
expand the negative conception of a brand from a category to another which both the bridge
and the brand belong to.
The previous reasoning about the correlation between bridges and betweenness centrality paves
the way to think that centralities play a key role in the diffusion of negative reviews. In particular,
it is reasonable to make the following hypothesis:
Hypothesis 5 (H5) - There is a correlation between degree and/or eigenvector centrality
and the capability of being negative influencer.
Degree centrality tells us which nodes have the highest number of relationships in a network.
These are probably power users, if we consider our stereotypes. They certainly are important
users, because they are densely connected. On the other hand, eigenvector centrality can help
us to identify influential users, who do not like to appear as such (the so called grey eminences
or grey cardinals). Those kinds of users are often connected to few nodes, each having a high
number of relationships with the other users [53]. These two centrality measures can be useful
to find negative influencers in Yelp.
4 Methodology
As we have seen above, our methodology starts from the multi-dimensional social network based model
introduced in Section 3, formulates some hypotheses and aims at verifying them using an inferential
campaign based on social network analysis. This campaign makes use of a number of concepts,
stereotypes and definitions that we introduce in this section. Instead, the way they are exploited to
prove the hypotheses and, more in general, to extract useful knowledge is described in Section 5.
The first concept we introduce is a stereotype, namely the k-bridge. Specifically, a k-bridge is a Yelp
user who reviewed businesses belonging to exactly kdifferent macro-categories of Yelp. A user who
reviewed businesses of only one macro-category is a non-bridge. Finally, we use the generic term bridge
to denote a k-bridge such that k > 1. Given a k-bridge upof U, where Uis the overall user network
corresponding to Yelp and introduced in Section 3, there are knodes n1p, n2p,· · · , nkpassociated with
her, one for each macro-category containing at least one review performed by her.
9

After having introduced the k-bridge, we present some other stereotypes, namely the power user
and the double-life user. More specifically, let Ci∈ Y be one of the macro-categories of Yelp. Let
rnibe the average number of reviews of Ci. Let bpbe a Yelp bridge and let CSetpbe the set of the
macro-categories that received reviews from bp. Then:
bpis defined as a power user if, for each macro-category Cj∈CSetp, the number of her reviews
is greater than or equal to 2 ·rnj.
bpis defined as a (x,y) access double-life user (access-dl-user, for short) if both the following
conditions hold:
–for a subset CSetpx⊂C Setpof xmacro-categories, the number of reviews of each Cj∈
CSetpxis greater than or equal to 2 ·rnj;
–for a subset CSetpy⊂C Setpof ymacro-categories, such that CSetpx∩CSetpy=∅, the
number of reviews of each Ck∈CSetpyis less than or equal to 1
2·rnk.
Double-life users play an extremely interesting role because they are very rare. Therefore, we
deepen our investigation on them and introduce a second kind of double-life users. Specifically, let bp
be a Yelp bridge. Then bpis defined as a (x, y)score double-life user (score-dl-user, for short) if both
the following conditions hold:
for a subset CSetpx⊂C Setpof xmacro-categories, the average number of stars that bpassigned
to the corresponding businesses is higher than or equal to 4;
for a subset CSetpy⊂C Setpof ymacro-categories, such that CSetpx∩C S etpy=∅, the average
number of stars that bpassigned to the corresponding businesses is lower than or equal to 2.
In order to make our inferential campaign on negative reviews and reviewers complete, we need
to introduce a further network that we call Negative Reviewer Network U=hN , Ai.Nis the set of
nodes of U. There is a node ni∈Nfor each Yelp user who made at least one negative review. There
is an arc apq = (np, nq) if there exists a friendship relationship between the user up, corresponding to
np, and the user uq, corresponding to nq.
In the next section, we show how all the concepts presented here can be exploited to prove the
hypotheses formulated in Section 3. This allows us to extract knowledge about negative reviews and
negative influencers in Yelp.
5 Results
5.1 General characteristics of Yelp
We collected the data necessary for the activities connected with our inferential campaign from the
Yelp website at the address https://www.yelp.com/dataset. In order to extract information of
interest from available data, we had to carry out a preliminary analysis. A first result concerns the
presence of 10,289 businesses whose category did not belong to any of the Yelp macro-categories, and
10

482 businesses that did not have any category associated with them (recall that in Yelp a business
can belong to one or more categories). Since the total number of businesses was 192,609, we decided
to discard these two kinds of businesses, because the amount of data removed was insignificant while
their presence would have led to procedural problems.
At this point, we analyzed the distribution of the categories among the macro-categories. We report
the result obtained in Figure 1. As we can see from this figure, the macro-category “Restaurants” has
a much greater number of categories than the other ones.
Figure 1: Distribution of the categories inside the Yelp macro-categories
Figure 2 shows the average number of reviews per user for each macro-category. As we can see,
the three macro-categories with the highest average number of reviews are “Restaurants”, “Food” and
“Nightlife”. Furthermore, in Figure 3, we show the same distribution for bridges only. We can see
that the three macro-categories with the highest number of reviews are always the same. However,
the average number of reviews is generally higher for bridges than for normal users. Therefore, we
can conclude that bridges not only tend to review businesses of different macro-categories (and this
happens by definition of bridge itself) but also to do more reviews than non-bridges.
In Figure 4, we report the distribution of access-dl-users against k. From the analysis of this figure,
we observe that the number of access-dl-users is already very high for k= 2 and further increases for
k= 3; then, it decreases very quickly and becomes almost negligible for k > 4.
11

Figure 2: Average number of business reviews made by Yelp users for each macro-category
Figure 3: Average number of business reviews made by Yelp bridges for each macro-category
We start looking at the access-dl-users corresponding to the simplest case of bridges, namely 2-
bridges. Table 1 shows the total number of 2-bridges, the number of (1,1) access-dl-users and the
number of power users, together with their corresponding percentage of the overall number of 2-
bridges. This table shows that (1,1) access-dl-users and power users represent very small fractions of
12

Figure 4: Distribution of access-dl-users against k
the overall set of 2-bridges.
Type of users Number and percentage
2-bridges 427130 (100%)
(1,1) access-dl-users 745 (0.17%)
power users 375 (0.087%)
Table 1: Numbers and percentages of 2-bridges, access-dl-users and power users in Yelp
We continue by examining all the k-bridges as kgrows, until at least one of them is an access-
dl-user or a power user. We can observe that this condition occurs for k≤6. The corresponding
numbers and percentages are shown in Tables 2 - 5. From the analysis of these tables, we can see how
the number of k-bridges decreases as kincreases, but the decrease is not fast. On the other hand,
the number of access-dl-users decreases very rapidly, about one order of magnitude at each step. The
number of power users decreases more slowly.
Type of users Number and percentage
3-bridges 245123 (100%)
(1,2) access-dl-users 450 (0.18%)
(2,1) access-dl-users 374 (0.15%)
power users 200 (0.081%)
Table 2: Numbers and percentages of 3-bridges, access-dl-users and power users in Yelp
5.2 Investigating the correctness of the Hypothesis H1
In Section 3, we have seen that a user can assign a number of stars between 1 and 5 to a business in
Yelp. The higher the number of stars, the better her rating is. Therefore, we decided to study the
reviews of users focusing on the number of stars that they assigned to businesses.
13

Type of users Number and percentage
4-bridges 147101 (100%)
(1,3) access-dl-users 19 (0.013%)
(2,2) access-dl-users 59 (0.040%)
(3,1) access-dl-users 28 (0.019%)
power users 35 (0.023%)
Table 3: Numbers and percentages of 4-bridges, access-dl-users and power users in Yelp
Type of users Number and percentage
5-bridges 91680 (100%)
(1,4) access-dl-users 6 (0.007%)
(2,3) access-dl-users 11 (0.012 %)
(3,2) access-dl-users 3 (0.003%)
(4,1) access-dl-users 0 (0%)
power users 14 (0.015%)
Table 4: Numbers and percentages of 5-bridges, access-dl-users and power users in Yelp
Type of users Number and percentage
6-bridges 63708 (100%)
(1,5) access-dl-users 0 (0%)
(2,4) access-dl-users 0 (0%)
(3,3) access-dl-users 1 (0.002%)
(4,2) access-dl-users 2 (0.003%)
(5,1) access-dl-users 11 (0.017%)
power users 11 (0.017%)
Table 5: Numbers and percentages of 6-bridges, access-dl-users and power users in Yelp
Figure 5 shows the average number of stars that users assigned to the businesses of each macro-
category. As we can see from this figure, this number is very high as it is always greater than 3.
As previously pointed out, this is actually not very surprising because the mechanism based on stars
follows a Likert scale and, in literature, it is well known that this scale is generally positively biased
[3, 62, 12].
In Table 6, we report the mean, standard deviation and mode of the number of stars assigned
by bridges and non-bridges to all businesses. As we can see from this table, there is no substantial
difference in this type of behavior between bridges and non-bridges.
Statistical Parameter Bridges Non-bridges
Mean 3.73 3.57
Standard Deviation 1.44 1.72
Mode 5 5
Table 6: Values of mean, standard deviation and mode of the number of stars assigned by bridges and
non-bridges to all businesses
From the results of Table 6, it is clear that it makes no sense to talk about power users in the
star-based analysis, because almost all users have the same behavior and assign a high number of stars
to almost all businesses. All these tests allow us to define the following:
Implication 1: The star-based review system of Yelp is positively biased. Indeed, almost all
users assign a high number of stars to almost all businesses.
Implication 1 is clearly a confirmation of the correctness of the Hypothesis H1.
14

Figure 5: Average number of stars for each macro-category of Yelp
5.3 Investigating the correctness of the Hypothesis H2
In Figure 6, we report the distribution of score-dl-users against k. From the analysis of this figure we
note that it follows a power law. If we compare this figure with Figure 4, we observe that for k= 2,
the number of score-dl-users is much smaller than the one of access-dl-users. However, the decrease of
the number of score-dl-users when kincreases is much smaller because they are different from 0 until
to k= 14.
We continued our analysis by verifying whether score-dl-users and access-dl-users were the same
people or not. We carried out this analysis with k= 6, because we had no access-dl-users with higher
values of k. In this case, we could see that the intersection of the two sets was empty.
To better understand the main features of score-dl-users we considered those corresponding to
7-bridges. These users were 16 (see Figure 6), a number that allowed us to examine in detail each
review carried out by them. During this analysis we found several interesting knowledge patterns.
More specifically, we observed that (1,6) and (6,1) score-dl-users show a completely different behavior
from the other 7-bridges. In fact, in this case, each (1,6) score-dl-user assigned positive scores to all the
business of the only macro-category that she positively reviewed. Similarly, each (6,1) score-dl-user
assigned negative values to all the businesses of the only macro-category that she negatively reviewed.
This can be justified thinking that users have a strong interest in that macro-category and so they
developed more accurate and stable evaluation criteria for the businesses belonging to it.
As for the other 7-bridges, we found that (2,5), (3,4), (4,3) and (5,2) score-dl-users show a less
extreme behavior, in the sense that they do not tend to give always positive or always negative ratings
15

Figure 6: Distribution of score-dl-users against k
to all the businesses of a given macro-category.
We then repeated the previous analyses for the last category of access-dl-users that we had avail-
able, namely the 6-bridges, to verify if the particular behavior of score-dl-users was typical of this
kind of double-life user or if it was something common. Actually, 6-bridge access-dl-users were 13;
therefore, we were able to make a detailed analysis of each review performed by each user also in this
case. We examined (1,5), (2,4), (3,3), (4,2) and (5,1) access-dl-users and we did not find substantial
differences in the behavior of these five categories of users. This appeared as a confirmation of the
singularity of the behavior observed for (1,6) and (6,1) score-dl-users. The previous analyses suggest
the following:
Implication 2: (a) Score-dl-users play a key role in negative reviews. (b) They are very keen
on negatively judging the macro-category they mostly attend.
Implication 2(a) confirms the correctness of our Hypothesis H2. But there is much more. In fact,
Implication 2(b) was an unexpected result that prompted us to carry out a further experiment to
have a confirmation. In it, we considered k-bridges, with 3 ≤k≤8, and computed the percentage of
them who negatively reviewed the macro-category of businesses they attended the most. Afterwards,
we computed the same percentage taking into account only k-bridges that were score-dl-users. The
results obtained are shown in Table 7. They represent an extremely strong confirmation of the previous
qualitative analysis.
As we have seen, the definition and behavior of score-dl-users are based on the number of stars
assigned by a user to a business during a review. We have already said that this type of score is based
on a Likert scale and, therefore, it is positively biased [3, 62, 12]. In order to overcome this problem,
in the literature authors suggest to evaluate the text of the reviews and to make a sentiment analysis
on it [39, 37]. We carried out this activity using two well-known sentiment analysis tools. The first is
16

k Percentage of k-bridges Percentage of score-dl-users k-bridges
3 4.35% 91.5%
4 4.03% 79%
5 3.65% 61%
6 2.40% 63%
7 2.11% 56%
8 1.55% 33%
Table 7: Percentages of k-bridges and score-dl-users k-bridges who negatively reviewed the macro-
category they mostly attended
TextBlob3, which, given a text, specifies if the corresponding polarity is positive, negative or neutral.
We applied TextBlob to users’ review texts. The results obtained are reported in Table 8. From the
analysis of this table we can see that the difference between the score based on stars and the polarity
based on sentiment analysis is equal to 15%.
Parameters Value obtained by applying TextBlob
Reviews 6,685,902
Reviews with a number of stars less than or equal to 2 (negative reviews) 1,544,553
Reviews classified as negative by TextBlob 847,359
Reviews with a number of stars greater than or equal to 3 (positive reviews) 5,141,347
Reviews classified as positive by TextBlob 5,781,007
Reviews classified as neutral by TextBlob 57,536
Negative reviews classified as positive 823,414
Positive reviews classified as negative 154,176
Positive reviews classified as neutral 30,914
Negative reviews classified as neutral 26,620
Table 8: Comparison between the review score based on stars and the review polarity obtained by
applying TextBlob
The second sentiment analysis tool we considered is Vader [35]. Also in this case, we applied it to
the users’ review texts. The results obtained are shown in Table 9. The analysis of this table confirms
the very low difference between the score of the star-based reviews and the polarity of the review texts
(in fact, in this case, this difference is equal to 14%).
Parameter Value obtained by applying Vader
Reviews 6,685,902
Reviews with a number of stars less than or equal to 2 (negative reviews) 1,544,553
Reviews classified as negative by Vader 982,102
Reviews with a number of stars greater than or equal to 3 (positive reviews) 5,141,347
Reviews classified as positive by Vader 5,649,489
Reviews classified as neutral by Vader 54,311
Negative reviews classified as positive 724,241
Positive reviews classified as negative 184,557
Positive reviews classified as neutral 31,542
Negative reviews classified as neutral 22,767
Table 9: Comparison between the review score based on stars and the review polarity obtained by
applying Vader
This allows us to conclude that score-based evaluations are generally confirmed by the sentiment
analysis performed on the corresponding reviews.
3https://textblob.readthedocs.io
17

5.4 Investigating the correctness of the Hypothesis H3
At this point, we analyzed how users influence each other with regard to negative reviews. We took
into consideration the network of friendships Yfsince it is easier for a user to have characteristics
more similar to her friends than to people she does not know, due to the principle of homophily [55].
Therefore, the ability to influence someone and/or to be influenced by her is presumably greater with
friends than with others.
As a first analysis, for each macro-category, we considered the percentage of users such that they,
and at least one of their friends, reviewed the same business negatively. The results obtained are
shown in Figure 7. From the analysis of this figure we can see how the percentages are extremely low.
The macro-category with the highest percentage is “Restaurant”, followed by “Nightlife” and “Food”.
This result can be explained taking into account that a person often attends restaurants or nightclubs
with her friends. Therefore, it is not unlikely that her negative judgement of a business may lead to
(or, on the contrary, may be caused by) a negative judgement of one or more of her friends.
Figure 7: Percentages of users such that they, and at least one of their friends, reviewed the same
business negatively
We repeated the analysis by distinguishing bridges from non-bridges. The corresponding results
are shown in Figures 8 and 9. From the analysis of these figures we observe higher values for bridges
than for non-bridges. For example, the value of “Nightlife” for bridges is more than 4 times the value
for non-bridges. Similarly, “Food”, in case of bridges, has a percentage more than 7 times higher than
for non-bridges.
To prove the statistical significance of our results we adopted a null model to compare our findings
with those obtained in an unbiasedly random scenario. Specifically, we built our null model by shuffling
the negative reviews among users in our dataset. In this way, we left unaltered all the original features
18

Figure 8: Percentages of bridges such that they, and at least one of their friends, reviewed the same
business negatively
Figure 9: Percentages of non-bridges such that they, and at least one of their friends, reviewed the
same business negatively
with the exception of the distribution of negative reviews, which became unbiasedly random in the
null model. After that, we repeated our analysis on the null model. The results obtained are reported
in Figure 10. Comparing this figure with Figure 7, we can see that there is a certain similarity in the
19

distributions; indeed, many of the macro-categories that had the highest values in Figure 7 continue
to have the highest values in Figure 10. However, in this last case, the values of the percentages are
several orders of magnitude smaller. Therefore, we can conclude that the behavior observed in Figure
7 is not random but it is the result of the reference context.
Figure 10: Percentages of users in the null model such that they, and at least one of their friends,
reviewed the same business negatively
At this point, for each macro-category, for each user who reviewed a given business negatively,
we computed the percentage of her friends who, having reviewed the same business, made a negative
review. The results obtained are shown in Figure 11. As we can see from this figure, the percentage
values are very high for almost all macro-categories.
Figures 12 and 13 show the same distributions, but for bridges and non-bridges. From the analysis
of these figures, it can be observed that the phenomenon is always strong, regardless of whether or not
a user is a bridge. An interesting knowledge pattern to observe is that there is a strong polarization on
the macro-categories especially in the case of non-bridges. In fact, the percentages of friends influenced
by them are either above 90% or null.
All the results shown above allow us to deduce the following:
Implication 3: A user has a very high influence on her/his friends when doing negative reviews.
This implication represents a confirmation of the correctness of our Hypothesis H3.
5.5 Investigating the correctness of the Hypothesis H4
In order to evaluate the Hypothesis H4, we started with the computation of the average percentage
of users who, having made a negative review in a category, have at least Xof their friends who
20

Figure 11: Percentages of friends who, having reviewed the same business as a user who reviewed a
business negatively, also provided a negative review
Figure 12: Percentages of friends who, having
reviewed the same business as a bridge who re-
viewed a business negatively, also provide a neg-
ative review
Figure 13: Percentages of friends who, having re-
viewed the same business as a non-bridge who re-
viewed a business negatively, also provide a neg-
ative review
negatively reviewed a business in the same category. The values of Xthat we considered are 1, 2, 3,
5, 10 and 100. As an example, in Figure 14, we report the results obtained in the case of X= 5. As
we can see from this figure, the percentages are some orders of magnitude greater than the ones of
Figure 10. The macro-categories with the highest values are the same as before, i.e., “Restaurants”,
“Food” and “Nightlife”.
21

Figure 14: Average percentages of users who, having made a negative review in a macro-category,
have at least Xof their friends who reviewed a business in the same macro-category negatively
As in the previous case, we distinguished bridges from non-bridges. The results of the corresponding
analysis are shown in Figures 15 and 16. These figures, along with the previous ones involving bridges
and non bridges, allow us to define the following:
Implication 4: Bridges have a much greater power of influence than non-bridges.
Again, we made the comparison with the null model. The results obtained for X= 5 are reported
in Figures 17, 18 and 19. From the examination of these figures, we can see how results obtained are
not random but they are intrinsic to Yelp. Note that the non-randomness can be observed for bridges
but generally not for non-bridges; this is important because it allows us to conclude that this property
characterizes bridges against non-bridges.
Implication 4 represents a confirmation that our Hypothesis H4 was correct.
5.6 Investigating the correctness of the Hypothesis H5 and defining a profile of
negative influencers in Yelp
To investigate the correctness of the Hypothesis H5 we considered the Negative Reviewer Network
U=hN , Aiintroduced in Section 4.
The analysis of this network allowed us to focus on users who reviewed some businesses negatively,
because, as we saw in the previous analysis, they are uncommon. Firstly, we computed the number
of nodes, the number of edges, the clustering coefficient and the density of Uand we compared them
with the same parameters as U. Results are shown in Table 10.
22

Figure 15: Average percentages of bridges who, having made a negative review in a macro-category,
have at least Xof their friends who reviewed a business in the same macro-category negatively
Figure 16: Average percentages of non-bridges who, having made a negative review in a macro-
category, have at least Xof their friends who reviewed a business in the same macro-category
negatively
23

Figure 17: Average percentages of users in the null model who, having made a negative review in a
macro-category, have at least Xof their friends who reviewed a business in the same macro-category
negatively
U U
Number of nodes 1637138 743178
Number of edges 7392305 2199987
Average clustering coefficient 0.043 0.039
Density 0.00000551619 0.00000796645
Table 10: Characteristics of Uand U
From the analysis of this table we can observe that the number of users who made at least one
negative review is 45.39% of total users. As for the average clustering coefficient and the density, we
found that their values do not present significant differences between Uand U.
At this point, we computed the distribution of users for U; it is shown in Figure 20. As we can see
from this figure, it follows a power law.
After studying the basic parameters of U, we computed the degree centrality of the nodes of this
network. In particular, we focused on the users with the highest values of degree centrality. More
specifically, we considered the top X% users, X∈ {1,5,10,20}. Observe that as Xdecreases, the
corresponding top users are increasingly central, i.e., increasingly strong. In Figure 21, we show the
distributions against kfor the top X% of users with the highest degree centrality. Note that for X= 20,
the distribution follows a power law, even if it is flatter than the one of Figure 20, which referred to
all users. As Xdecreases, we can see how the distribution becomes flatter and flatter, moving to
the right and tending to a Gaussian shape. This allows us to conclude that more central users (i.e.,
those with the highest degree centrality) tend to be stronger also as k-bridges (i.e., characterized by
an increasingly higher value of k).
24

Figure 18: Average percentages of bridges in the null model who, having made a negative review in a
macro-category, have at least Xof their friends who reviewed a business in the same macro-category
negatively
Figure 19: Average percentages of non-bridges in the null model who, having made a negative review
in a macro-category, have at least Xof their friends who reviewed a business in the same macro-
category negatively
25

Figure 20: Distribution of users of Uagainst k
Figure 21: Distributions of the top X% of users with the highest degree centrality against k
Instead, in Figure 22, we show the user distributions against kfor the top X% of users with
the highest eigenvector centrality. The trend of these distributions as Xdecreases is very similar to
(although slightly less marked than) the one of the degree centrality.
Figure 23 shows the user distributions against kfor the top X% of users with the highest PageRank.
26

Figure 22: Distributions of the top X% of users with the highest eigenvector centrality against k
Also in this case, we have a similar trend, although the variations of the distributions as Xdecreases
are much more attenuated, compared to the two previous cases. The last three figures allow us to
define the following:
Implication 5: There is a correlation between k-bridges and top central users.
Implication 5 is valid especially for the top central users based on degree centrality. This result,
along with the previous ones, is extremely important because it allows us to determine which are the
main negative influencers in Yelp. In fact, we can define the following:
Implication 6: The main negative influencers in Yelp are score-dl-users who
simultaneously are top central users (according to degree and/or eigenvector
and/or PageRank centrality measures).
Implication 6 not only confirms the correctness of the Hypothesis H5, but goes much further. In
fact, it defines a profile of the negative influencers in Yelp and, consequently, provides a way to detect
them.
27

Figure 23: Distributions of the top X% of users with the highest PageRank against k
6 Discussion
6.1 Reference context of our paper
In the previous sections, we have investigated the phenomenon of negative reviews in Yelp and, then,
we have characterized negative influencers in this social medium. In the past, different research
papers have focused on the consequences that user-written reviews have on businesses and, generally,
on the market. As a first step in this scenario, it is interesting to understand what makes customer
reviews helpful to a consumer in her process of making a purchase decision. With regard to this,
in [67], the authors first collect reviews made on Amazon.com. Then, they distinguish between two
different product types, namely: (i) search goods, for which a consumer can obtain information on
their quality before purchasing them; (ii) experience goods, which are products requiring a purchase
before evaluating their quality. This product categorization plays a key role in understanding what
a consumer perceives more from a review. Indeed, moderate reviews are more helpful than extreme
(i.e., strongly positive or negative) ones for experience goods, but not for search goods. Furthermore,
longer reviews are generally perceived as more helpful than shorter ones, but this effect is greater for
search goods than for experience goods.
Another interesting contribution in this scenario is reported in [78], in which the authors introduce
several factors that can influence the decision making process of consumers about their purchases.
Indeed, the authors of [78] strive to understand what are the key elements guiding a user in the
purchase of a certain product. They propose a model taking systematic factors (e.g., the quality
of online reviews) and heuristic ones (e.g., the quantity of online reviews) into account. They test
28

this model on 191 users and obtain interesting results. In fact, they identify important factors to
care about; these are argument quality, source credibility, and perceived quantity of reviews. They
empirically prove that consumers receiving reviews from credible sources and perceiving the quantity
of reviews as large tend to perceive the topics in online reviews as more informative and persuasive.
This means that if consumers find review sources to be credible, their purchase intention is usually
higher. Finally, they also show that consumers are more likely to purchase products with many online
reviews rather than with few ones.
Several authors have investigated the impact of positive and negative reviews. For instance, the
authors of [22] examine how a positive Electronic Word of Mouth (hereafter, eWOM) can affect other
users’ purchasing decisions. Indeed, eWOM is strictly related to the online reviews phenomenon,
which can be regarded as a special case of it. Generally, eWOM is based on an analysis of costs
and benefits. The authors investigate the psychological motivations beneath the spread of positive
reviews. They take a sample dataset from the OpenRice.com platform, one of the most successful
review platforms in Hong Kong and Macau. Through a questionnaire, they asked people who wrote
reviews on this website their motivations. Starting from the received answers, they build a model
based on different features, namely the eWOM intention of consumers, the reputation, the reciprocity,
the sense of belonging, the pleasure to help, the moral obligation and the self-efficacy of knowledge.
They show that their model is capable of representing the behavior of users when they share (positive)
personal experiences on such online platforms.
The influence of positive reviews of businesses has been studied from many other points of view.
For example, in [42], the authors analyze celebrity sponsorships in the context of for-profit and non-
profit marketing. They actually find that famous people can influence the appreciation one has for
a product or service, in a positive or negative direction. This suggests that it makes sense studying
who negative influencers are, how they behave and how they can be detected in an online platform.
Not limited to celebrities, people are more incline to follow users disclosing their personal information
[27]. The members of an online community rate reviews containing descriptive identity information
more positively, and the prevalence of identity information disclosure by reviewers is associated with
increased subsequent sales of online products. In addition, the shared geographical location increases
the relationship between disclosure and product sales.
Wrapping up these important results, we can say that buyers are influenced by positive eWOM,
especially if it is performed by nearby identifiable users; even more, celebrities can change the ap-
preciation that people have for a product or a service. But the consequences are not just limited to
customers. Even internal decision-making processes of businesses can be influenced by online review
systems [2]. The diffusion of personal opinions through the Internet has radically changed the concept
of reviewing a product or a service that one has in traditional media. In fact, online review platforms
offer to users a space where they can express their unfiltered thoughts on products or services. In
particular, eWOM encourages a two-way communication between a source and a reader, thus being
more engaging. A very important result of [2] is that eWOM helps companies to obtain higher product
and service evaluations and, if necessary, higher amounts of funding; furthermore, it influences the
decision-making processes of companies, showing that its power is not limited only to buyers. The
other important result of [2] is that the effect of negative eWOM is much greater than the one of
positive eWOM.
29

Negative reviews open up many research issues. One of them is finding out what drives users to
write negative reviews. Discontent, or “disconfirmation”, with a product or service has been studied
as a cause of this phenomenon. The authors of [33] define disconfirmation as the discrepancy between
the expected evaluation of a product and the evaluation of the same product performed by experts.
In particular, they find that a person is more likely to leave a review when the disconfirmation she
encounters is great. They also find that the evaluation published by a person may not reflect her post-
purchase evaluation in a neutral manner; indeed, the direction of such polarization is in agreement
with disconfirmation.
The authors of [77] introduce a theory about the initial beliefs of a consumer when she is looking
for a product. According to this theory, a consumer forms an initial judgement about a product based
on its summary rating statistics. This initial belief plays a key role in her next evaluation of the
review. To prove their conjecture, the authors of [77] collected the application reviews from Apple
Store from July 1st to August 31st, 2013. By analyzing these reviews they show the existence of a
confirmation bias, which outlines the tendency of consumers to perceive reviews confirming (resp.,
disconfirming) their initial beliefs as more (resp., less) helpful. This tendency is moderated by the
consumer confidence in their initial beliefs. This bias also leads to a greater perceived helpfulness of
positive reviews when the average product rating is high, and of negative reviews when the average
product rating is low.
6.2 Main findings of the knowledge extraction process
In the Introduction, we specified that the main novelties of this paper concern: (i) the definition of
the two social network based models of Yelp; (ii) the definition of three Yelp user stereotypes and
their characteristics; (iii) the construction of the profile of negative influencers in Yelp. We also
pointed out that this paper aims at answering three research questions, namely: (i) What about
the dynamics leading a Yelp user to publish a negative review? (ii) How can the interaction of
these dynamics increase the “power” of negative reviews and people making them? (iii) Who are the
negative influencers in Yelp? In order to obtain these results and answer these questions, we conducted
a data analytics campaign that allowed us to formulate six implications.
The first tells that “The star-based review system of Yelp is positively biased. Indeed, almost all
users assign a high number of stars to almost all businesses.”. It can be explained by taking into
account that Yelp’s review system is based on a Likert scale, and it is well known that this scale
is positively biased [3, 62, 12]. This implication does not provide unexpected information, but still
represents an important confirmation about the correctness of our knowledge extraction process.
The second implication tells that “Score-dl-users play a key role in negative reviews. They are
very keen on negatively judging the macro-category they mostly attend.”. Unlike the first one, it was
not expected. Its explanation partially comes from the first implication. Indeed, if it is true that the
Likert scale is positively biased, then a user must be particularly motivated to give a negative rating.
Moreover, if such an evaluation is given by a double life user, then it means that it is provided by a
person potentially balanced in her evaluations (indeed, she gave both positive and negative evaluations
in the past). If a person with these characteristics gives a negative review, it is reasonable to assume
that she did so because she had “something important to say”. In that case, she probably provides
30

some well founded justifications for her dissatisfaction. In order to do this, she must be competent in
that macro-category, which explains the last part of the implication.
The third implication tells that “A user has a very high influence on her/his friends when doing
negative reviews.”. The first part of it represents an expected result, and is easily explained by the
homophily principle [55]. The second part was unexpected and can be explained by considering that
several studies in related literature show that negative reviews and reviewers are stronger than positive
ones.
The fourth implication tells that “Bridges have a much greater power of influence than non-
bridges.”. It represents a partially expected result if we consider that bridges generally have a high
betweenness centrality and, thus, have the ability to convey an idea, sentiment or opinion from one
macro-category to another.
The fifth implication tells that “There is a correlation between k-bridges and top central users.”.
At first glance, it may appear an expected result, but actually this is not the case. In fact, in some
contexts, for example in a Social Internetworking System, bridges connecting different social networks
are not necessarily power users [15]. Actually, the more the communities involved in a (multi-) network
scenario are integrated, the more likely a bridge is also a power user. Based on this reasoning, and
considering that Yelp’s macro-categories are closely related to each other, because both a user and a
business can belong to more macro-categories simultaneously, the result obtained is reasonable and
motivated.
Finally, the sixth implication tells that “The main negative influencers in Yelp are score-dl-users
who simultaneously are top central users (according to degree and/or eigenvector and/or PageRank
centrality measures).”. It is certainly unexpected and is one of the main findings of this paper. It was
obtained by appropriately integrating the previous five implications. For this reason, the justifications
underlying it are those that allowed us to explain the implications from which it derives.
6.3 Theoretical contributions and implications
This paper provides several theoretical contributions to the literature on online review systems and
eWOM. First of all, it introduces a new multi-dimensional social network based model of Yelp. This
model perfectly fits the category-based structure of this social medium. It represents Yelp as a set of
22 communities, one for each macro-category. At the same time, it models this social medium as a
user network Uwhere each node denotes a user and an arc between two nodes represents a generic
relationship between the corresponding users. Our model can be used in several different scenarios,
depending on the type of relationship one wants to represent. In our study, we have specialized it to
two different types of relationships, namely the friendship between users (i.e., Uf) and the co-review
of the same business carried out by different users (i.e., Ucr).
The usage of our model, together with a set of experiments performed on a Yelp dataset, allowed us
to show that the star-based review mechanism of Yelp is positively biased. This fact implies that a user
must have a strong motivation to write a negative review. In turn, this implies that all information
about negative reviews and negative influencers in Yelp is extremely valuable.
After that, thanks to our multi-dimensional model, we were able to define different stereotypes of
users in Yelp. In particular, we considered three different stereotypes, namely the bridges, the power
31

users and the double-life users. Bridges are users connecting different communities in Yelp. They are
crucial for the dissemination of information in this social platform. In fact, we have seen that the
influence exerted by bridges is greater than the one exerted by non-bridges. Power users are very
active in performing reviews in the categories of their interest. The amount of reviews they carry out
makes them extremely important in the identification of potential influencers. Double-life users show
different behaviors in the different categories in which they operate. They generally show a particular
attention and severity in a category in which they are extremely experienced. This means that they
can play a valuable role as influencers in this category.
We have defined our multi-dimensional model and these stereotypes with respect to Yelp. However,
our model can be easily generalized to other online review platforms, such as TripAdvisor, as well as
to other types of social platforms. In case of online review platforms, the extension of our model
is immediate. In fact, it is sufficient to know and report in our model the hierarchy of categories
underlying the online review platform. In case of other types of social media, the extension is possible
and quite simple. In fact, it is sufficient to specify a (possibly hierarchical) mechanism for dividing
users into groups, as well as to identify the types of user relationships of interest. It seems quite
obvious that friendship is a relationship of interest for any social platform. On the contrary, co-review
does not always make sense and could be replaced by other types of relationships.
As for stereotypes, we observe that those considered in this paper are not the only ones possible
for an online review platform. In the future, we plan to identify other stereotypes and study their
contribution to the extraction of useful knowledge from Yelp. At the same time, the three stereotypes
identified in this paper can be directly extended to any other online review platform. The concept
of power user can be easily extended to any social platform and any online social network too. The
concept of bridge and double-life user can be extended only to those cases where users of a social
platform can be organized into communities based on some parameters. In this case, a bridge is a user
acting as a link between two communities, while a double-life user is a user having different behaviors
in different communities.
The last theoretical contribution of this paper concerns the definition of the Negative Reviewer
Network. This model plays an extremely important role in the study of negative reviews and, above
all, in the identification of negative influencers, who correspond to nodes with high degree centrality
and/or high eigenvector centrality, as we have seen in Section 5.6. Analogously to what happens
for the other theoretical tools introduced in this paper, the extension of this model to other online
review platforms is immediate. Instead, its extension to other types of social platforms is much less
simple than the other models and concepts seen above. In fact, by its nature, the Negative Reviewer
Network is specifically designed to model negative reviews and reviewers. Therefore, its extension is
only possible by identifying other negative behaviors that one wants to study and by defining a form
of co-participation of multiple users to these behaviors.
6.4 Implications for practice
Starting from the theoretical background, the hypotheses made and the implications confirming them,
we can outline different applications of the knowledge extracted in this paper to real life scenarios. In
particular, we can identify two different perspectives, i.e., the business and the user ones.
32

The business perspective concerns all the possible actions that a company can take to expand
its customer base, to improve its brand image or to extend the products/services it offers. In this
context, the user stereotypes identified in this paper and the implications associated with them can be
extremely useful. Let us consider, for example, k-bridges. We have seen the extremely important role
that they play in disseminating information between different communities. In Section 6, we have also
seen that past literature highlights the strong impact that negative reviews can have. In this context,
a k-bridge making a negative review could have a disruptive effect on a business image.
Therefore, the possibility of detecting k-bridges provided by our approach can become a valuable
tool for a business, which can adopt a variety of policies aiming at improving their evaluation of its
products/services from negative to neutral or, even, positive. Another extremely important policy
in this sense could regard the promotion of a business to k-bridges who do not know it. This could
favor the knowledge of this business in all the communities which the k-bridges belong to. In fact, a
k-bridge belonging to a community where a business is well known and another community where this
latter is unknown could become a promoter of the business from the former community to the latter
one.
Another important application that could leverage k-bridges is the expansion of products/services
offered by a business towards new categories, or even new macro-categories, of Yelp. One way to
increase the chance of designing new products/services being of interest to users could be as follows.
A business could identify all the k-bridges belonging to the categories in which it is already known and
its products/services are highly appreciated. Then, it could determine the other categories of prod-
ucts/services where the identified k-bridges have performed revisions; in fact, the products/services of
these last categories could be of interest for the potential customers of this business. The greater the
number of k-bridges that have shown interest in these categories, the more likely customers belonging
to them will be attracted by the business if it expands its offers towards these markets.
A further application of k-bridges, collateral to the one seen above, concerns advertising campaigns.
In fact, knowing which are the most promising communities where proposing new products/services
also implies being able to carry out advertising campaigns focusing on them. In this way, the effec-
tiveness and efficiency of the advertisement activity in terms of time and costs are increased.
However, k-bridges are not the only stereotype identified in this paper having important practical
applications. In fact, both power users and double-life users are equally important. Since the latter two
stereotypes appear within the definition of negative influencers, we now see some possible applications
of this last concept that subsumes the other two ones. Negative influencers have two important
characteristics. The first concerns the high value of network centrality measures (degree centrality
and/or eigenvector centrality and/or PageRank), which makes them very influential in the communities
where they operate. The second concerns their behavior in carrying out reviews. In fact, we have seen
that a negative influencer, being a score-dl-user, tends to give positive reviews in the categories of lesser
interest, while she is very demanding and severe in the categories in which she is more experienced and
that interest her the most. This also assumes that such a user generally has a recognized leadership
exactly in the category in which she is most severe. Therefore, it becomes crucial for a business in
that category taking all possible actions to ensure that she takes a neutral, or hopefully a positive,
attitude towards the products/services it offers. On the other hand, as we have seen for k-bridges, it
is possible to think of targeted advertising and marketing actions on these users that, if successful,
33

are characterized by a high level of efficiency and effectiveness.
So far we have seen the possible exploitations of our knowledge patterns from the business view-
point. Now, we want to see how the same patterns can have practical implications for the user as
well. In particular, we want to consider what benefits a user can get by looking at other relevant users
(such as k-bridges, power users, influencers) in Yelp.
A first benefit can be obtained from the examination of the reviews of negative influencers in Yelp.
Based on the knowledge we have extracted, we can assume that these users are very experienced in a
certain category and very severe in exactly that category. Therefore, if these users have issued positive
reviews on the products/services of a business in that category, it is very likely that they are of high
quality.
A second benefit for a user concerns the knowledge of the features characterizing the profile of an
influencer in Yelp. This knowledge becomes extremely useful if she wants to become an influencer
in that social medium. In fact, based on the implications derived in our paper, the user knows that
she has a better chance to become an influencer if she becomes a k-bridge. As a consequence, she
will have to be active in making revisions in multiple categories. In addition, she should be a power
user; therefore, she must have many friendship and co-review relationships (which implies she has
a high degree centrality). Alternatively, she can have a limited number of friendship and co-review
relationships as long as the users connected to her are, in turn, power users (which implies she has a
high eigenvector centrality). Finally, she must identify one or more categories in which she wants to
be an influencer and develop a high experience in them in order to give severe, but correct, reviews.
The knowledge extracted in our paper can also be useful to define recommender systems for users
who want to discover new products/services. This can be done, for example, by leveraging k-bridges.
In fact, assume that a user follows some categories. It is possible to identify all the k-bridges of
these categories and, for these k-bridges, to consider the categories followed by them. In this way, it is
possible to identify which categories are the most followed by these k-bridges. If one of these categories
is not already followed by the user, it is possible to recommend it to her. This very general approach
could be further refined by examining the proximity, in the Yelp hierarchy, of candidate categories to
those already followed by the user. A further refinement could assign different weights to the different
k-bridges, based on the similarity of their past evaluation to those of the user of interest on the same
products/services, or based on the number of categories already followed by both them and the user
of interest.
6.5 Limitations and future research directions
The theoretical tools introduced in this paper (i.e., the multi-dimensional social network based model
of Yelp, the stereotypes and the Negative Review Network), together with the hypotheses formulated
and the implications confirming them, have allowed us to shed light on the phenomenon of negative
reviews and negative influencers in Yelp. The tools proposed and the approach followed are sufficiently
general to be extended directly to other online review platforms and, after some generalizations, to any
social platform. However, they are to be considered simply as a first step in this direction, because they
are not free from limitations, whose knowledge paves the way to new future research investigations.
The first limitation of our approach is that it is exclusively structural and does not take semantics
34

into account. Actually, a more focused study on the contents of negative reviews would be necessary
to understand the reasons that led users to formulate them. This would increase the effectiveness and
efficiency of the applications of our approach discussed in Section 6.4. In fact, given a service/product
receiving many negative reviews, we could strive to understand the main reasons for this fact and,
therefore, make the appropriate improvements aimed at satisfying as many users as possible in the
shortest time.
An in-depth semantic analysis of reviews would also be extremely useful to define one or more
taxonomies of negative influencers. This would allow us to classify them based not only on the
products/services they criticize, as in the present approach, but also on the main reasons for negativity
(which would give us several indications on where intervening first or mainly). Semantic knowledge
would also allow us to better evaluate negative influencers in order to understand who give plausible
reasons and who, instead, are prevented, regardless it happens. As a matter of fact, a business could
make an effective and efficient recovery work on the former category of influencers, while it could
decide not to intervene on the latter one, because the possibility of making them neutral or positive
is low.
Another limitation of our approach, which is, at the same time, a potential future development
of our research concerns stereotypes. In this paper, we have presented three of them, namely the
k-bridges, the power users and the double-life users. Their identification was driven by our research
needs. However, we believe that several other stereotypes could be defined and that it could be even
possible to go so far as to define a real taxonomy of stereotypes for both Yelp and other online (review)
platforms. These would become a real toolbox available to decision makers when they need to make
decisions regarding the products/services provided by their business (for instance, to determine those
ones to be removed from catalogues, new ones to be proposed, existing ones to be modified for making
them more in line with user needs and desires, etc.).
A third limitation of our approach, which is also linked to current technological limitations expected
to become less impacting in the future, concerns the possibility of studying all these phenomena over
time. In fact, our current approach is based on a temporal (albeit wide) photograph of the negative
reviews of Yelp. It is not incremental and, if we want to study the evolution of a phenomenon over time,
we should take more datasets referring to different times and study them separately. However, this
does not allow us to have a continuous monitoring of the phenomenon, in order to capture any changes
regarding it (for instance, any change of how some products/services are perceived by users) as soon as
possible. The weight of this limitation (and, consequently, the relevance of overcoming it) is smaller in
substantially stable socio-economic conditions, because user perceptions of products/services change
very slowly over time in this scenario. Instead, it becomes crucial in historical periods characterized by
sudden and disruptive phenomena (think, for instance, of the current COVID-19 pandemic), capable
of upsetting all previous mental patterns of people’s judgement. In this case, having the possibility
of immediately understanding the changed perceptions of users about products/services and/or the
appearance of new needs, with the consequent demand for new products/services, can allow a business
to gain a huge advantage over its competitors. More importantly, this feature would allow the whole
ecosystem of public and private product/service providers to be efficient and effective in responding
to people demands.
35

7 Conclusion
In this paper, we dealt with the phenomenon of negative reviews in Yelp and outlined the profile of
negative influencers. To this end, we used a new multi-dimensional social network based model of
Yelp, several stereotypes of Yelp users derived from it, and a Negative Reviewer Network. Then, we
formulated several hypotheses and we evaluated their correctness through an experimental campaign.
In particular, at the end of our activities, we obtained the following knowledge patterns: (i) the
star-based review system of Yelp is positively biased; (ii) bridges and double-life users play a key
role in negative reviews; (iii) a user has a high influence on her friends when doing negative reviews;
(iv) the main negative influencers in Yelp are score-dl-users who simultaneously are top central users
(according to degree and/or eigenvector and/or PageRank centrality measures).
In this paper, we have proposed a multi-dimensional investigation of negative reviews in Yelp.
In our analysis, the dimensions into consideration are co-reviews, friendships and business categories.
First, we have proposed a preliminary analysis of Yelp data to understand the distribution of categories
and reviews in the macro-categories of Yelp. Then, we have focused on three types of users, namely
k-bridges, power users and double-life users. After that, we have studied how users can influence each
other in making negative reviews on the same businesses and/or on the same macro-categories. Finally,
we have built a network that takes into account only negative reviewers and, after a series of analyses
and studies that we have made on this network, we have determined a possible identikit of a negative
influencer in Yelp. Thanks to our investigations, we have seen that there are two important categories
of users in Yelp, namely power users (who are very active in several macro-categories) and double-
life users (who show different behaviors in different macro-categories). We have found that a user
who reviews one or more businesses of a given macro-category negatively can influence their friends to
perform the same. This can be explained according to the principle of homophily characterizing several
social network phenomena. We have seen that the influence exerted by bridges is greater than the one
exerted by non-bridges. Finally, we have detected that negative influencers in Yelp are people who
are score-dl-users and simultaneously top X% users in degree centrality and/or eigenvector centrality
and/or PageRank. In the future, we plan to extend our research in various directions. First of all, we
think of analyzing the phenomenon of negative reviews in other social media, such as TripAdvisor, to
understand the similarities and differences with Yelp. Then, we plan to extend our model to make it
suitable for analyzing other aspects and other peculiarities of Yelp. Finally, we think of defining an
approach that exploits the anti-monotonic property characterizing the definition of k-bridge to allow
the detection of negative influencers related to a business, a macro-category or a group of target users.
References
[1] A.K. Agarwal, A.P. Pelullo, and R.M. Merchant. “ld”: the Word Most Correlated to Negative Online Hospital
Reviews. Journal of General Internal Medicine, pages 1–2, 2019.
[2] R. Aggarwal, R. Gopal, A. Gupta, and H. Singh. Putting Money Where the Mouths Are: The Relation Between
Venture Financing and Electronic Word-of-Mouth. Information Systems Research, 23(3):976–992, 2012. INFORMS.
[3] A. Alexandrov. Characteristics of single-item measures in Likert scale format. The Electronic Journal of Business
Research Methods, 8(1):1–12, 2010.
36

[4] S. Angelidis and M. Lapata. Multiple instance learning networks for fine-grained sentiment analysis. Transactions
of the Association for Computational Linguistics, 6:17–31, 2018.
[5] J.R. Arthur, D. Etzioni, and A.J. Schwartz. Characterizing extremely negative reviews of total joint arthroplasty
practices and surgeons on yelp.com. Arthroplasty Today, 2019.
[6] C. Aslay, L.V.S. Lakshmanan, W. Lu, and X. Xiao. Influence maximization in online social networks. In Proc. of
the ACM International Conference on Web Search and Data Mining (WSDM’18), pages 775–776, Marina del Rey,
CA, USA, 2018. ACM.
[7] S. Basuroy, S. Chatterjee, and S.A. Ravid. How critical are critical reviews? The box office effects of film critics,
star power, and budgets. Journal of Marketing, 67(4):103–117, 2003. SAGE Publications.
[8] K. Bauman and A. Tuzhilin. Discovering contextual information from user reviews for recommendation purposes. In
Proc. of the International Workshop on New Trends in Content-Based Recommender Systems (CBRecSys @ RecSys
2014), pages 2–9, Foster City, CA, USA, 2014.
[9] J. Berger, A.T. Sorensen, and S.J. Rasmussen. Positive effects of negative publicity: When negative reviews increase
sales. Marketing science, 29(5):815–827, 2010. INFORMS.
[10] M. Berlingerio, M. Coscia, F. Giannotti, A. Monreale, and D. Pedreschi. Foundations of Multidimensional Network
Analysis. In Proc. of the International Conference on Advances in Social Networks Analysis and Mining (ASONAM
2011), pages 485–489, Kaohsiung, Taiwan, 2011. IEEE.
[11] M. Berlingerio, F. Pinelli, and F. Calabrese. Abacus: frequent pattern mining-based community discovery in
multidimensional networks. Data Mining and Knowledge Discovery, 27(3):294–320, 2013.
[12] D. Bertram. Likert scales. Retrieved November, 2:2013, 2007.
[13] P.K. Bhanodia, A. Khamparia, B. Pandey, and S. Prajapat. Online social network analysis. In Hidden Link
Prediction in Stochastic Social Networks, pages 50–63. IGI Global, 2019.
[14] A.K. Bhowmick, S. Suman, and B. Mitra. Effect of information propagation on business popularity: A case study
on yelp. In Proc. of the International Conference on Mobile Data Management (MDM’17), pages 11–20, Daejeon,
South Korea, 2017. IEEE.
[15] F. Buccafurri, V.D. Foti, G. Lax, A. Nocera, and D. Ursino. Bridge Analysis in a Social Internetworking Scenario.
Information Sciences, 224:1–18, 2013. Elsevier.
[16] F. Buccafurri, G. Lax, S. Nicolazzo, and A. Nocera. Comparing Twitter and Facebook user behavior: Privacy and
other aspects. Computers in Human Behavior, 52:87–95, 2015. Elsevier.
[17] F. Buccafurri, G. Lax, A. Nocera, and D. Ursino. SISO: a conceptual framework for the construction of “stereotypical
maps” in a Social Internetworking Scenario. In Proc. of the International Workshop on New Frontiers in Mining
Complex Knowledge Patterns at ECML/PKDD 2012 (NFMCP 2012), Bristol, UK, 2012.
[18] F. Buccafurri, G. Lax, A. Nocera, and D. Ursino. Moving from social networks to social internetworking scenarios:
The crawling perspective. Information Sciences, 256:126–137, 2014. Elsevier.
[19] D. Cai, Z. Shao, X. He, X. Yan, and J. Han. Community mining from multi-relational networks. In Proc. of the
European Conference on Principles of Data Mining and Knowledge Discovery (PKDD’05), pages 445–452, Porto,
Portugal, 2005. Springer.
[20] Q. Cao, W. Duan, and Q. Gan. Exploring determinants of voting for the “helpfulness” of online user reviews: A
text mining approach. Decision Support Systems, 50(2):511–521, 2011. Elsevier.
37

[21] Y.C. Chang, C.H. Ku, and C.H. Chen. Social media analytics: Extracting and visualizing Hilton hotel ratings and
reviews from TripAdvisor. International Journal of Information Management, 48:263–279, 2019. Elsevier.
[22] C.M.K. Cheung and M.K.O Lee. What Drives Consumers to Spread Electronic Word of Mouth in Online Consumer-
Opinion Platforms. Decision Support Systems, 53(1):218–225, 2012. Elsevier.
[23] C.M.K. Cheung and D.R. Thadani. The impact of Electronic Word-of-Mouth Communication: A Literature Analysis
and Integrative Model. Decision Support Systems, 54(1):461––470, 2012. Elsevier.
[24] Y. Cui. An Evaluation of Yelp Dataset. arXiv preprint arXiv:1512.06915, 2015.
[25] D. Davis, R. Lichtenwalter, and N.V. Chawla. Multi-relational link prediction in heterogeneous information net-
works. In Proc. of the International Conference on Advances in Social Networks Analysis and Mining (ASONAM
2011), pages 281–288, Kaohsiung, Taiwan, 2011. IEEE.
[26] J. Fogel and S. Zachariah. Intentions to use the yelp review website and purchase behavior after reading reviews.
Journal of Theoretical and Applied Electronic Commerce Research, 12(1):53–67, 2017.
[27] C. Forman, A. Ghose, and B. Wiesenfeld. Examining the Relationship Between Reviews and Sales: The Role of Re-
viewer Identity Disclosure in Electronic Markets. Information Systems Research, 19(3):291—-313, 2008. INFORMS.
[28] D.W. Franks, J. Noble, P. Kaufmann, and S. Stagl. Extremism propagation in social networks with hubs. Adaptive
Behavior, 16(4):264–274, 2008.
[29] M.S. Granovetter. The strength of weak ties. American Journal of Sociology, 78(6):1360–1380, 1973. JSTOR.
[30] J. Guerreiro and P. Rita. How to predict explicit recommendations in online reviews using text mining and sentiment
analysis. Journal of Hospitality and Tourism Management, 2019. Forthcoming. Elsevier.
[31] A. Gulati and M. Eirinaki. With a Little Help from My Friends (and Their Friends): Influence Neighborhoods
for Social Recommendations. In Proc. of the World Wide Web Conference (WWW’19), pages 2778–2784, San
Francisco, CA, USA, 2019. ACM.
[32] A. Hicks, S. Comp, J. Horovitz, M. Hovarter, M. Miki, and J.L Bevan. Why people use Yelp. com: An exploration
of uses and gratifications. Computers in Human Behavior, 28(6):2274–2279, 2012. Elsevier.
[33] Y.C. Ho, J. Wu, and Y. Tan. Disconfirmation Effect on Online Rating Behavior: A Structural Model. Information
Systems Research, 28(3):626––642, 2008. INFORMS.
[34] L. Hu, A. Sun, and Y. Liu. Your neighbors affect your ratings: on geographical neighborhood influence to rating
prediction. In Proc. of the International ACM SIGIR Conference on Research & development in information retrieval
(SIGIR’14), pages 345–354, Gold Coast, Queensland, Australia, 2014. ACM.
[35] C.J. Hutto and E. Gilbert. Vader: A parsimonious rule-based model for sentiment analysis of social media text.
In Proc. of the International AAAI Conference on Weblogs and Social Media (ICWSM’14), pages 216–225, Ann
Arbor, MI, USA, 2014.
[36] Y.S. Kang, J. Min, J. Kim, and H. Lee. Roles of alternative and self-oriented perspectives in the context of the
continued use of social network sites. International Journal of Information Management, 33(3):496–511, 2013.
Elsevier.
[37] W. Kasper and M. Vela. Sentiment analysis for hotel reviews. In Proc. of the International Computational
Linguistics-Applications Conference, volume 231527, pages 45–52, Jachranka, Poland, 2011.
[38] A.L. Kavanaugh, D.D. Reese, J.M. Carroll, and M.B. Rosson. Weak ties in networked communities. The Information
Society, 21(2):119–131, 2005.
38

[39] K. Kaviya, C. Roshini, V. Vaidhehi, and J.D. Sweetlin. Sentiment analysis for restaurant rating. In Proc. of
the International Conference on Smart Technologies and Management for Computing, Communication, Controls,
Energy and Materials (ICSTM’17), pages 140–145, Chennai, India, 2017. IEEE.
[40] C. Ke-Jia, Z. Pei, Y. Zinong, and L. Yun. iBridge: Inferring bridge links that diffuse information across communities.
Knowledge-Based Systems, 192, 2020. Elsevier.
[41] J. Kim, J. Bae, and M. Hastak. Emergency information diffusion on online social media during storm Cindy in US.
International Journal of Information Management, 40:153–165, 2018. Elsevier.
[42] J. Knoll and J. Matthes. The Effectiveness of Celebrity Endorsements: A Meta-Analysis. Journal of the Academy
of Marketing Science, 45(1):55–75, 2017. Springer.
[43] N. Kumar and I. Benbasat. Research note: the influence of recommendations and consumer reviews on evaluations
of websites. Information Systems Research, 17(4):425–439, 2006. INFORMS.
[44] K. Lee, J. Ham, S. Yang, and C. Koo. Can You Identify Fake or Authentic Reviews? An fsQCA Approach. In
Information and Communication Technologies in Tourism 2018, pages 214–227, Jonkoping, Sweden, 2018. Springer.
[45] X. Lei and X. Qian. Rating prediction via exploring service reputation. In Proc. of the International Workshop on
Multimedia Signal Processing (MMSP’15), pages 1–6, Xiamen, China, 2015. IEEE.
[46] J. Leskovec, L.A. Adamic, and B.A. Huberman. The dynamics of viral marketing. ACM Transactions on the Web,
1(1):5, 2007. ACM.
[47] M.X. Li, C.H. Tan, K.K. Wei, and K.L. Wang. Sequentiality of Product Review Information Provision: An
Information Foraging Perspective. MIS Q., 41(3):867–892, 2017. Management Information Systems Research
Center.
[48] Y. Lim and B. Van Der Heide. Evaluating the wisdom of strangers: The perceived credibility of online consumer
reviews on Yelp. Journal of Computer-Mediated Communication, 20(1):67–82, 2014. Oxford University Press.
[49] X. Lin and X. Wang. Examining gender differences in people’s information-sharing decisions on social networking
sites. International Journal of Information Management, 50:45–56, 2020. Elsevier.
[50] M. Luca. Reviews, reputation, and revenue: The case of Yelp.com. Harvard Business School Working Paper,
(12-016), 2016.
[51] M. Luca and G. Zervas. Fake it till you make it: Reputation, competition, and yelp review fraud. Management
Science, 62(12):3412–3427, 2016.
[52] X. Luo. Quantifying the long-term impact of negative word of mouth on cash flows and stock prices. Marketing
Science, 28(1):148–165, 2009. INFORMS.
[53] W. Maharani, Adiwijaya, and A.A. Gozali. Degree centrality and eigenvector centrality in twitter. In Proc. of
the International Conference on Telecommunication Systems Services and Applications (TSSA’14), pages 1–5, Bali,
Indonesia, 2014. IEEE.
[54] J. Malbon. Taking fake online consumer reviews seriously. Journal of Consumer Policy, 36(2):139–157, 2013.
Springer.
[55] M. McPherson, L. Smith-Lovin, and J.M. Cook. Birds of a feather: Homophily in social networks. Annual Review
of Sociology, 27:415–444, 2001. JSTOR.
[56] A. Mukherjee, V. Venkataraman, B. Liu, and N. Glance. What yelp fake review filter might be doing? In Proc. of
the International AAAI Conference on Weblogs and Social Media (ICDSM’13), Boston, MA, USA, 2013.
39

[57] M. Nakayama and Y. Wan. The cultural impact on social commerce: A sentiment analysis on yelp ethnic restaurant
reviews. Information & Management, 56(2):271–279, 2019. Elsevier.
[58] H. Nam, Y.V. Joshi, and P.K. Kannan. Harvesting brand information from social tags. Journal of Marketing,
81(4):88–108, 2017.
[59] P. Nokhiz and F. Li. Understanding rating behavior based on moral foundations: The case of Yelp reviews. In Proc.
of the International Conference on Big Data (Big Data 2017), pages 3938–3945, Boston, MA, USA, 2017. IEEE.
[60] A. Parikh, C. Behnke, M. Vorvoreanu, B. Almanza, and D. Nelson. Motives for reading and articulating user-
generated restaurant reviews on yelp. com. Journal of Hospitality and Tourism Technology, 5(2):160–176, 2014.
[61] A.A. Parikh, C. Behnke, B. Almanza, D. Nelson, and M. Vorvoreanu. Comparative content analysis of professional,
semi-professional, and user-generated restaurant reviews. Journal of Foodservice Business Research, 20(5):497–511,
2017.
[62] G. Peeters and J. Czapinski. Positive-negative asymmetry in evaluations: The distinction between affective and
informational negativity effects. European review of social psychology, 1(1):33–60, 1990. Taylor & Francis.
[63] M. Potamias. The warm-start bias of Yelp ratings. arXiv preprint arXiv:1202.5713, 2012.
[64] J. Qiu, Y. Li, and Z. Lin. Does Social Commerce Work in Yelp? An Empirical Analysis of Impacts of Social
Relationship on the Purchase Decision-making. In Proc. of the Pacific Asia Conference on Information Systems
(PACIS’18), page 16, Yokohama, Japan, 2018.
[65] J. Qiu, Y. Li, and Z. Lin. Detecting Social Commerce: An Empirical Analysis on Yelp. Journal of Electronic
Commerce Research, 21(3):168–179, 2020. Journal of Electronic Commerce Research.
[66] A. Saxena, R. Gera, I. Bermudez, D. Cleven, E.T. Kiser, and T. Newlin. Twitter Response to Munich July 2016
Attack: Network Analysis of Influence. Frontiers in Big Data, 2:17, 2019. Forthcoming. Frontiers.
[67] D. Schuff and S. Mudambi. What makes a helpful online review? A study of customer reviews on Amazon.com.
Social Science Electronic Publishing, 34(1):185–200, 2012. Elsevier.
[68] V. Setyani, Y.Q. Zhu, A.N. Hidayanto, P.I. Sandhyaduhita, and B. Hsiao. Exploring the psychological mechanisms
from personalized advertisements to urge to buy impulsively on social media. International Journal of Information
Management, 48:96–107, 2019. Elsevier.
[69] W. Shen, Y.J. Hu, and J.R. Ulmer. Competing for Attention: An Empirical Study of Online Reviewers’ Strategic
Behavior. MIS Q., 39(3):683–696, 2015. Management Information Systems Research Center.
[70] X. Shi, B.L. Tseng, and L.A. Adamic. Looking at the blogosphere topology through different lenses. In Proc. of
the International Conference on Weblogs and Social Media (ICWSM’07), Boulder, CO, USA, 2007.
[71] R. Singh, J. Woo, N. Khan, J. Kim, H.J. Lee, H.A. Rahman, J. Park, J. Suh, M. Eom, and N. Gudigantala.
Applications of machine learning models on yelp data. Asia Pacific Journal of Information Systems, 29(1):117–143,
2019.
[72] Y. Sun and J.D.G. Paule. Spatial analysis of users-generated ratings of yelp venues. Open Geospatial Data, Software
and Standards, 2(1):5, 2017.
[73] P.L. Ting, S.L. Chen, H. Chen, and W.C. Fang. Using big data and text analytics to understand how customer
experiences posted on yelp. com impact the hospitality industry. Contemporary Management Research, 13(2), 2017.
Academy of Taiwan Information Systems Research.
[74] Q. Xuan, X. Shu, Z. Ruan, J. Wang, C. Fu, and G. Chen. A self-learning information diffusion model for smart
social networks. IEEE Transactions on Network Science and Engineering, 2019. Forthcoming.
40

[75] Y. Yang, N. Chawla, Y. Sun, and J. Hani. Predicting links in multi-relational and heterogeneous networks. In Proc.
of the International Conference on Data Mining (ICDM’12), pages 755–764, Bruxelles, Belgium, 2012. IEEE.
[76] D. Yin, S.D. Bond, and H. Zhang. Anxious or angry? Effects of discrete emotions on the perceived helpfulness of
online reviews. MIS quarterly, 38(2):539–560, 2014. JSTOR.
[77] D. Yin, S. Mitra, and H. Zhang. When do consumers value positive vs. negative reviews? An empirical investigation
of confirmation bias in online word of mouth. Information Systems Research, 27(1):131–144, 2016. INFORMS.
[78] K.Z. Zhang, S.J. Zhao, C.M. Cheung, and M.K. Lee. Examining the influence of online reviews on consumers’
decision-making: A heuristic–systematic model. Decision Support Systems, 67:78–89, 2014. Elsevier.
[79] Y. Zhang, S. Shi, S. Guo, X. Chen, and Z. Piao. Audience management, online turbulence and lurking in social net-
working services: A transactional process of stress perspective. International Journal of Information Management,
56:102233, 2021. Elsevier.
[80] Z. Zhang, Q. Li, D. Zeng, and H. Gao. User community discovery from multi-relational networks. Decision Support
Systems, 54(2):870–879, 2013. Elsevier.
[81] M. Zhou, X. Cai, Q. Liu, and W. Fan. Examining continuance use on social network and micro-blogging sites:
Different roles of self-image and peer influence. International Journal of Information Management, 47:215–232,
2019. Elsevier.
41