Content uploaded by Jian Wu
Author content
All content in this area was uploaded by Jian Wu on Apr 05, 2021
Content may be subject to copyright.
Noname manuscript No.
(will be inserted by the editor)
A group consensus-based travel destination
evaluation method with online reviews
Jian Wu ·Qing Hong ·Mingshuo Cao ·
Yujia Liu ·Hamido Fujita
Received: date / Accepted: date
Abstract With the help of the massive online information, non-expert decision-
making problems (such as travel destination selection) can be solved. Online
reviews provide decision-making opinions and weight information for tourists
who have never been to the alternative travel destinations. To deal with the
increasing tourism products and group tourists, this study proposes a novel
group consensus-based travel destination evaluation method with online re-
views, which considers the missing preference estimating and group consensus
reaching process. Firstly, decision opinions are represented through the senti-
ment matrix with the percentage distribution. Secondly, to obtain the weight
vector of attributes, the incomplete complementary matrices with the prefer-
ence of attributes are given by group users. Subsequently, the missing pref-
erence values in the matrix are estimated. Thirdly, all users are required to
Jian Wu
School of Economics and Management, Shanghai Maritime University, Shanghai 201306,
China
E-mail: jyajian@163.com
Qing Hong
School of Economics and Management, Shanghai Maritime University, Shanghai 201306,
China
E-mail: hqing9977@163.com
Mingshuo Cao
School of Economics and Management, Shanghai Maritime University, Shanghai 201306,
China
E-mail: caomingshuo0606@163.com
Yujia Liu
School of Economics and Management, Shanghai Maritime University, Shanghai 201306,
China
E-mail: lyj71@126.com
Hamido Fujita
Iwate Prefectural University, Iwate, Japan.
E-mail: HFujita-799@acm.org
2 Jian Wu et al.
reach group consensus based on the minimum adjustment cost feedback mech-
anism. Finally, the sentiment matrix with the percentage distribution can be
aggregated by the weight vectors. So that the ranking of alternatives can be
obtained. In this study, an example of travel destination evaluation based on
the online reviews of Dazhong.com and Ctrip.com is given to illustrate the use
of the proposed method.
Keywords Online reviews ·Group decision making ·Incomplete preference
values ·Group consensus ·Travel destination evaluation
1 Introduction
Benefit from tourism websites and APPs, the rapid growth of tourism infor-
mation makes tourism more and more popular, at the same time, generates
massive online review data for most aspects of travel destination evaluation[1].
Besides, the rapid development of social networks makes the relationship be-
tween people closer. So that more and more people tend to choose to travel
together, such as family travel, group travel of schools and companies, which
means that group tourism has become the main way of tourism. The group
travel destination evaluation is a multi-attribute group decision-making prob-
lem, which involves a group of users and several attributes. Online reviews
provide information for users to make decisions. To make a reasonable and
acceptable decision by group travelers, the two issues need to be solved.
The first issue is to obtain the preference values for the unknown travel
destination. Different from the traditional experts’ decision-making problem,
which has relatively complete knowledge and information about the decision
problem, group consensus-based travel destination evaluation always lacks
knowledge and information about decision problem. Users know little about
the details of the travel destination, so that they cannot provide complete
decision information and weight vector of decision attributes. To obtain a rea-
sonable decision-making result, users need to access online reviews on tourism
websites to make decision. The forms of online reviews are diversified, in-
cluding numerical rating, text reviews, and image evaluations[2], likewise, the
evaluation content is comprehensive, covering multiple attributes of tourism
products: eating, living, traffic, traveling, shopping, and entertainment[3]. Nev-
ertheless, the large amount of online data and the sparse useful information
can make it extremely difficult for tourists to find the most appropriate al-
ternative. Therefore, how to process online reviews to help users understand
the details of each travel destination intuitively is one of the problems to be
solved. On the other hand, to obtain the weight vector of multi-attributes of
travel destination, the evaluation method of incomplete complementary judg-
ment matrix is another problem to be solved[4,5]. However, extant literature
rarely considers unknown decision opinions and the incomplete weight vectors
of group tourists at the same time. This study aims to propose an incomplete
preference of weight vectors for selecting unknown travel destination based on
online reviews.
Title Suppressed Due to Excessive Length 3
The second issue is to reach a group consensus. With the increasing group
tourists, it is necessary for tourism product provider to consider a decision
support system to help group users to make a reasonable decision. Users may
have different backgrounds and from different organizations, which may lead
to group inconsistent in obtaining the weight vectors of multi-attributes[6,
7]. Group inconsistent may lead to inaccurate results for travel destination
evaluation. So that a minimum group consensus is required by the independent
moderator to ensure the rationality of travel destination evaluation[8,9]. Thus,
how to generate recommended opinions to help inconsistent users to reach a
minimum group consensus degree is a significant topic to be studied[10–14].
However, conventional travel destination evaluation method does not include
the group consensus reaching processes.
The objective of this paper is to propose a novel group consensus-based
travel destination evaluation method with online reviews. In the method, the
online reviews are represented as the sentiment matrix with percentage dis-
tribution, then the group of users provide their complementary matrix on
pairwise comparisons of attributes. Therefore, the main purpose of this study
can be concluded as follows:
1) To propose a novel travel destination evaluation method with online
reviews, a percentage distribution on evaluations concerning each alternative
based on online reviews is constructed. A missing preference value complete
method is utilized for constructing the complementary judgment matrix of the
weight vector of the multiple attributes[15].
2) To reach a minimum group consensus for travel destination evaluation,
an interactive group consensus reaching process based on the minimum cost
feedback mechanism is utilized. Once, the group consensus is reaching, the
weight vectors can be obtained.
Thus, the rest of this paper is arranged as follows. Section 2 provides an
overview of the related work. Section 3 proposes a framework for solving the
problem and presents a new method according to the framework. In section
4, an example is given to illustrate the use of the proposed method. Finally,
section 5 summarizes and highlights the major contributions of this paper.
2 Related work
2.1 Multi-attribute decision-making method based on online reviews
To have a reasonable decision-making result in the big data era, more and
more researchers pay attention to the multi-attribute decision-making method
with online reviews. Kang and Park[16] studied an APPs selection method,
in which customer satisfaction on the APPs is measured based on the online
reviews, and then VIKOR with sentiment analysis is utilized to rank alterna-
tive APPs. Liang et al.[17] proposed an improved AHP method for evaluating
the satisfaction of the O2O takeaway platforms, in which, the PLTSs and
improved AHP calculate weight vector by utilizing online reviews. Chen et
4 Jian Wu et al.
al.[18] conducted a positive and negative sentiment analysis based on the on-
line reviews of mobile phones, and then rank the alternative mobile phones
using the TOPSIS method. Hotel selection is a hot topic in the study of multi-
attribute decision-making methods with online reviews. Liang et al.[19] devel-
oped a quantitative method for hotel selection based on sentiment analysis
and DL-VIKOR method, in which, the tourists’ preferences are obtained by
the sentiment levels of online reviews, and the alternative hotels are rank by a
proposed DL-VIKOR method. Kwok and Lau[20] develop a novel VS-TOPSIS
method for hotel selection based on online reviews, which let travelers express
their preference of the selection criteria. Peng et al.[21] developed an applica-
ble hotel decision support model for tourists based on online reviews, in which
probabilistic linguistic term sets (PLTSs) are introduced to summarize the re-
view information statistically, and then a cloud model is employed to deal with
probabilistic linguistic information. Zhang et al.[1] designed a new hotel deci-
sion support model using online social information for a travel website in the
United States that provides travel catalog information and travel-related re-
view content to help the independent traveler to find satisfactory restaurants.
The model makes full use of social information including online reviews and
social relationships and considers the connection between attributes. Online
reviews provide decision-makers with information on unknown tourist destina-
tions. Hence, the multi-attribute decision-making method with online reviews
is an effective method for the tourist destination decision-making problem.
Existing studies denoted that the multi-attribute decision-making method
based on online reviews is a complicated process, which is mainly reflected
in two aspects: First, the preference of users are uncertain and incomplete.
Second, there is a situation of inconsistency among experts in groups. Even
though researchers realized that obtaining attribute weights is a very im-
portant problem for travel destination evaluation[22], existing multi-attribute
decision-making methods based on online reviews didn’t consider the incom-
plete decision opinions of attributes’ weights, nor the group consensus in the
decision processes.
2.2 Researching consensus in group decision making
Zhang et al.[23] proposed a novel personalized restaurant recommendation ap-
proach that combines group correlations between customer group and restau-
rant group, which shows the interaction between group users in tourism de-
cision problem has attracted the attention of researchers. However, the group
consensus reaching process, which is necessary for reasonable group travel des-
tination evaluation, has been ignored by researchers. In the processes of group
tourism destination decision-making, due to the different backgrounds and ob-
jectives, the preference for attributes weights of all the users may not reach
a consensus. Travelers will not accept the decision results without reaching a
minimum group consensus level[24]. Therefore, a key problem that needs to be
considered in group tourism destination decision-making based on online re-
Title Suppressed Due to Excessive Length 5
views is how to reach the group consensus during the evaluating of attributes’
weights. Li et al.[25] proposed the unique axiomatic distance in the linguistic
context for measuring group consensus. The inconsistent decision makers will
be asked to revise the decision opinions to reach the group consensus. Due to
the different psychology and behavior of decision-makers in different situations,
the opinion recommendation and acceptance mechanism are the most impor-
tant topic of group consensus research processes[26,27]. A three-level group
consensus measurement and feedback mechanism were proposed and utilized
to reach group consensus[28]. Dong et al.[10] proposed a minimum cost feed-
back mechanism for reaching group consensus. The advantage of this feedback
mechanism is that it balanced the consistency and adjustment costs. Liu et
al.[29] believe that trust makes the group consensus reaching processes makes
the recommendation mechanism more personalized. So that a trust-induced
feedback mechanism is constructed, in which the inconsistent experts only ac-
cept recommendations from the experts they trust. Furthermore, Liu et al.[30]
developed a recommendation mechanism based on the subjective and objec-
tive trust relationship. Urena[31] developed a similarity-confidence-consistency
based social network that enables the agents to provide their opinions with
intuitionistic fuzzy preference values and achieve an agreement among users.
Gong et al.[32] presented a consensus model based on interval preference. It
makes the group consensus reaching processes based on the minimum cost
and maximum return with duality theory. Wu et al.[33] proposed a group con-
sensus model of a minimum cost feedback mechanism based on a distributed
language trust social network. The advantage of this method is that it only
requires inconsistent users to adjust their opinions to optimize the balance
between consistency and adjustment costs. Since the minimum cost of group
consensus reaching is an essential goal of travel destination evaluation decision-
making problems, this study uses the group consensus model of minimum cost
feedback mechanism to reach a consensus on the evaluation of group travel
destination.
In summary, the existing multi-attribute decision-making method based on
online reviews for group travel destination evaluation still has the following
deficiencies: (1) There are few studies on group decision-making method based
on online reviews for travel destination evaluation, in which, the text analysis
for online reviews of travel destinations needs to be studied. What’s more, the
incomplete information for obtaining weight vectors of attributes wasn’t con-
sidered in the travel destination evaluation problems. To estimate the unknown
decision opinions and the incomplete weight vector of attributes, this study
will propose an incomplete preference of weight vectors for selecting unknow
travel destination based on online reviews. (2) Conventional multi-attribute
decision-making methods with online reviews ignored the group consensus pro-
cesses, which is the fundamental condition for obtaining a reasonable decision
result. An effective and personalized group consensus reaching method needs
to be applied before the decision results have been made. This study will con-
sider a minimum cost-based group consensus reaching process for group travel
destination evaluation.
6 Jian Wu et al.
3 Methodology
3.1 The structure of the proposed group travel destination evaluation model
To solve the group travel destination evaluation problem, a framework of group
consensus-based travel destination evaluation is constructed, which contains
the following four parts: (1) construct the percentage distribution of each al-
ternative based on online reviews; (2) estimate incomplete preference value for
group travel destination evaluation; (3) conduct the group consensus reaching
processes based on minimum adjustment cost feedback mechanism; and (4)
select the most appropriate alternatives. The proposed framework of group
consensus-based travel destination evaluation can be shown in Fig.1.
Fig. 1: The structure of group travel destination evaluation model
The decision framework of group travel destination evaluation with on-
line reviews aims to select the most appropriate travel destination for group
users. The following notations are used to denote the sets and variables in the
problem, which will be used throughout this paper:
–X={X1, X2, ..., Xm}: the set of mtravel destinations, where Xtdenotes
the t-th alternative, t= 1,2, ..., m.
–C={C1, C2, ..., Cn}: the set of ntourism attributes, where Cjdenotes
the j-th attribute, j= 1,2, ..., n.
–E={e1, e2, ..., ek}: the set of kusers in the group, where ehdenotes the
h-th user, h= 1,2, ..., k.
–Eh= (eh
ij )n×n: a complementary preference matrix given by user eh, each
element eh
ij expresses his/her fuzzy preference value for Cion Cj,h=
1,2, ..., k,i= 1,2, ..., n,j= 1,2, ..., n.
–R={R1, R2, ..., Rm}: online reviews set, where Rtdenotes the set of
online reviews concerning alternative travel destination Xt,t= 1,2, ..., m.
–w= (w1, w2, ..., wn): the vector of attribute weights, where wjdenotes the
weights of attribute Cj,wj≥0 and
n
P
j=1
wj= 1, j = 1,2, ..., n.
Title Suppressed Due to Excessive Length 7
3.2 The sentiment analysis based on online reviews
The online reviews were crawled from Ctrip.com and Dazhong.com by python,
which are the most popular travel website and consumer reviews website in
China. They have the advantages of massive user base and a large number of
reviews. Before crawling the online reviews, the travel destination should be
determined. This study selects the travel destinations with the same market
positioning and similar pricing levels as the alternatives.
3.2.1 Text-preprocess
To identify the sentiment orientations of the online reviews, it is necessary to
preprocess the obtained online reviews. The preprocessed process includes two
parts: (1) word segmentation and part-of-speech (POS) tagging, and (2) stop
words elimination.
(1) Word segmentation and part-of-speech tagging. The jieba word-segmentation
tool is implemented on the online reviews set R={R1, R2, ..., Rm}con-
cerning the alternative travel destinations for word segmentation and POS
tagging, each sentence is decomposed into some words and POS of each
word is marked after the word. For example, if jieba word-segmentation
tool is implemented on “the service is very user-friendly”, then the result
is “service/vn, very/d, user-friendly/a”, where n(noun); vn(verb noun);
d(adverb); a(adjective).
(2) Stop words elimination. Stop words refer to the frequent words with little
practical meaning. To improve the efficiency and effectiveness of sentiment
analysis, stop words need to be removed. The obtained words after word
segmentation and POS tagging are compared with stop word list from
HowNet(A Chinese Vocalbulary for Sentiment Analysis), then the words
that belong to the stop word list are deleted.
3.2.2 Feature word extraction
In this paper, feature words are refined based on word frequency statistics[34],
which sort product features according to the frequency of nouns, and then,
extract the top several features as feature words. According to the result of
word frequency sorting, the top Ntop features are selected as the feature words
set. The details are given below.
First, each review in online reviews set (denoted as R={R1, R2, ..., Rm})
is decomposed into several words through text preprocessing steps. Then, the
nouns are sorted by word frequency and the top Ntop features are extracted as
feature words set F={F1, F2, ..., Fm}, where Ftdenotes the feature words set
concerning alternative travel destination Xt,t= 1,2, ..., m. Next, the feature
words set is classified manually by referring to the tourism ontology[34,35],
which is divided into six major attributes of tourism[34]: food, accommodation,
transportation, sightseeing, shopping, and entertainment. Finally, the feature
8 Jian Wu et al.
word list (shown in Table.3) is obtained. The pseudo code of the algorithm for
building feature word list is shown in Algorithm.1.
Algorithm 1 Building feature word list by word frequency statistics
Input:
The online reviews set R={R1, R2, ..., Rm}
Output:
The feature word list
1: for review in Rtdo
2: text preprocess to review
3: end for
4: sort the frequency of nouns
5: return top Ntop features Ft
6: classify Ftmanually into C={C1, C2, ...Cn}
7: return feature word list
3.2.3 Opinion sentences extraction
A review usually contains several short sentences, but not every sentence con-
tains tourism attributes and opinion words. For example, the review “Very
beautiful. It takes 3 minutes to take the cable car. The service is very user-
friendly in Emei” contains three sentences, but only the last sentence can be
extracted as the effective content for sentiment analysis. In this paper, the sen-
tences containing tourism attributes and opinion words in online reviews are
classified into the opinion sentences set (denoted as R0={R10, R20, ..., Rm0},
where Rt0denotes the set of opinion sentences concerning alternative travel
destination Xt) based on the feature word list (shown in Table.3) and senti-
ment word list (denoted as O={O+, O−}, where O+and O−respectively
denotes the set of positive and negative sentiment words in HowNet). The
details are given below.
First, a review usually contains several sentences, the arbitrariness and
irregularity of Chinese increase the difficulty of subsequent analysis, so each
review in online reviews set R={R1, R2, ..., Rm}is separated from several
sentences according to the punctuation“,.!?;...”. Then, each sentence is de-
composed into several words through text preprocessing steps. Next, if one
sentence contains features in feature words set F={F1, F2, ..., Fm}and sen-
timent words in sentiment word list O={O+, O−}, the sentence needs to be
classified into opinion sentences set R0={R10, R20, ..., Rm0}. The pseudo code
of the algorithm for obtaining opinion sentences set is shown in Algorithm.2.
Title Suppressed Due to Excessive Length 9
Algorithm 2 Obtain opinion sentences set
Input:
The online reviews set R={R1, R2, ..., Rm};
The sentiment word list O={O+,O −};
The feature words set F={F1, F2, ..., Fm}
Output:
opinion sentences set R0={R10, R20, ..., Rm0}
1: for review in Rtdo
2: separate sentences in review
3: return sentences
4: for sentence in sentences do
5: text preprocess to sentence
6: return words
7: for word in words do
8: if word ∈Ftthen
9: for word in words do
10: if word ∈Othen
11: classify sentence into opinion sentences set Rt0
12: end if
13: end for
14: end if
15: end for
16: end for
17: end for
18: return opinion sentence set R0={R10, R20, ..., Rm0}
3.2.4 Sentiment analysis
To identify the sentiment orientation on the tourism attribute Cj, sentiment
analysis is performed on each sentence in opinion sentences set. First, the
feature-opinion pairs in each sentence are extracted based on the distance be-
tween features and opinion words, then the dependency parsing is implemented
on the sentence to extract the modifiers modifying the opinion words. Next,
the sentiment orientation of the opinion word is identified based on HowNet (A
Chinese Vocabulary for Sentiment Analysis). Furthermore, sentiment intensity
analysis and orientation analysis of the modifier of opinion word are carried
out, which may have an impact on the intensity or invert the orientation of
opinion words. Then, the sentiment orientation of the feature in each opinion
sentence is obtained. The details are given below.
(1)Feature-opinion pairs extraction and dependency analysis
First, each sentence in opinion sentences set R0={R10, R20, ..., Rm0}is
decomposed into several words through text preprocessing steps. Then, ac-
cording to the distance between features(nouns) in feature words set F=
{F1, F2, ..., Fm}and opinion words, feature-opinion pairs can be obtained. For
example, the sentence “service is very user-friendly”, where the opinion word
closest to feature “service” is “user-friendly”, so feature-opinion pair is “ser-
vice, user-friendly”. For the situations that the feature-opinion pairs of some
sentences cannot be correctly identified based on the distance between features
and opinion words, the manual identification can be employed to determine
10 Jian Wu et al.
the feature-opinion pairs of some sentences. Next, Stanford CoreNLP is im-
plemented on the modifiers and opinion words in each sentence to conduct de-
pendency parsing, in which four typical dependency relations are considered:
advmod adverb modifier, amod adjective modifier, nsubj nominal subject, and
neg negative modifier. Finally, by matching the modifiers in sentence with the
modifiers in list of adverbs of degree D={D1, D2, D3}and negation words set
Nw, the opinion words in sentence with the sentiment words in sentiment word
list O={O+, O−}, the corresponding sentiment orientation is calculated. For
example, the sentence ”service is very user-friendly”, whose dependency pars-
ing is advmod, indicating that ”very” as the modifier modifies the opinion
word ”user-friendly”, then by matching ”very” and ”user-friendly” with the
modifiers in list of adverbs of degree D={D1, D2, D3}and sentiment words
in sentiment word list O={O+, O−}, the sentiment orientation of feature
“service” is obtained.
(2) Identify the sentiment orientation of opinion words
Four sets of HowNet (A Chinese Vocalbulary for Sentiment Analysis) are
employed to determine the positive or negative sentiment orientation of the
opinion words describing reviewers’ opinions on the tourism attributes in
each opinion sentence. In details, positive evaluation words set (e.g. beau-
tiful, friendly), negative evaluation words set (e.g. shallow, cruel), positive
sentiment set (e.g. convenient, satisfactory) and negative sentiment set(e.g.
regretful, disappointed). Then the sentiment orientation of opinion words can
be expressed as follows:
Definition 1 Let Sq
tj be the opinion word on the tourism attribute Cjin the
q-th online review of alternative travel destination Xt, then the sentiment ori-
entation of opinion word Sq
tj can be expressed as follows[33]:
P(Sq
tj ) = 1, Sq
tj ∈O+
−1, Sq
tj ∈O−(1)
where O+denotes positive evaluation words set and positive sentiment set, O−
denotes negative evaluation words set and negative sentiment set.
(3)Sentiment intensity and sentiment orientation analysis
We define two variables (adverbs of degree and negation words) to express
the modifiers of opinion words. The adverbs of degree express the strength
of orientation. Negation words can invert the orientation of the opinion word.
Let adv(Sq
tj ) and neg(Sq
tj ) respectively denote the adverbs and negation words
modifying the opinion word Sq
tj .
Adverbs of degree are words that have impact on the sentiment inten-
sity, such as “extremely”, “especially”, “greatly”. HowNet provides a list of
adverbs of degree D={D1, D2, D3}, which can be divided into three levels
(shown in Table.1) according to their different intensity, where D1,D2and
D3respectively denotes the intensity set of adverbs of degree in HowNet.
Title Suppressed Due to Excessive Length 11
Table 1: A list of adverbs of degree from the HowNet
Score of the adverbs Adverbs of degree
2 very, extremely, especially, greatly, strongly, exceedingly, spe-
cially.. .
1 more, so, too, quite, pretty, rather, almost, fairly, consider-
ably.. .
0.5 slightly, briefly, a little, a bit. . .
Negation words can invert the sentiment orientation of the opinion word,
such as “no” and “not”. If only a negation word appears before the opinion
word, the sentiment orientation of the opinion word is reversed accordingly.
But if a negation word appears before the negation word, it indicates affirma-
tion and the sentiment orientation of the opinion word remains unchanged.
For example, “Tickets are not so expensive”, where “expensive” belongs to
the set O−, however, “not” as the negation word modifies the opinion word
“expensive”, so the sentiment orientations of the opinion word “expensive”
is reversed. The pseudo code of the algorithm for sentiment analysis of Sq
tj
is shown in Algorithm3. Then the sentiment orientation can be expressed as
follows:
Definition 2 The sentiment orientation on the tourism attribute Cjin the
q-th online review of alternative travel destination Xtis defined as[33]:
Sensibilityq
tj =P(Sq
tj )×V(Sq
tj )×(−1)q
tj N(2)
where P(Sq
tj )represents the sentiment orientation of the opinion word Sq
tj ,
V(Sq
tj )represents the adverb of degree modifying Sq
tj ,(−1)q
tj Ncontrols the
orientation of the opinion word Sq
tj .
Thus, the value range of Sensibilityq
tj is {−2,−1,−0.5,0.5,1,2}.
Algorithm 3 Sentiment analysis
Input:
The online reviews set R={R1, R2, ..., Rm};
The sentiment word list O={O+, O−};
The adverbs of degree list D={D1, D2, D3};
The negation words set Nw;
Output:
Sensibilityq
tj
1: for sentence in Rt0do
2: text preprocess to sentence
3: return words
4: for word in words do
5: if word ∈Ftthen
6: find opinion word Sq
tj closest to word
7: dependency parsing
12 Jian Wu et al.
8: return adv(Sq
tj ) OR neg(Sq
tj )
9: if Sq
tj ∈O+then
10: P(Sq
tj ) = 1
11: else if Sq
tj ∈O−then
12: P(Sq
tj ) = −1
13: end if
14: if adv(Sq
tj )∈D1then
15: V(Sq
tj )=2
16: else if adv(Sq
tj )∈D2then
17: V(Sq
tj )=1
18: else if adv(Sq
tj )∈D3then
19: V(Sq
tj )=0.5
20: else
21: V(Sq
tj )=1
22: end if
23: N= 0
24: if neg(Sq
tj )∈Nwthen
25: N=N+ 1
26: end if
27: end if
28: end for
29: end for
30: Sensibilityq
tj =P(Sq
tj )×V(Sq
tj )×(−1)q
tj N
31: return Sensibilityq
tj
3.2.5 Emotional score probability distribution representation
Let H1=−2, H2=−1, H3=−0.5, H4= 0.5, H 5= 1, H6= 2be the set of
the evaluation scales. To represent the evaluation scales of different travel des-
tinations on attributes more accurately, this study uses probability distribution
to express sentiment orientation.
Definition 3 Let nqH1
tj , qH2
tj , ..., qH6
tj obe the set of frequencies of reviews with
different evaluation scales of travel destination Xt(t= 1, ..., m)on the at-
tribute Cj, where qHr
tj denotes the frequency of reviews with r-th evaluation
scale. Qt(t= 1, ..., m)indicates the total number of reviews for the t-th travel
destination. Where uHr(Sensibilityq
tj ) = 1, Sensibilityq
tj =Hr
0, Sensibilityq
tj 6=Hr. The values
of qH1
tj , qH2
tj , ..., qH6
tj can be respectively calculated[37]:
qH1
tj =
Qt
X
q=1
uH1(Sensibilityq
tj ), t = 1,2, ..., m, j = 1,2, ..., n. (3)
qH2
tj =
Qt
X
q=1
uH2(Sensibilityq
tj ), t = 1,2, ..., m, j = 1,2, ..., n. (4)
Title Suppressed Due to Excessive Length 13
qH3
tj =
Qt
X
q=1
uH3(Sensibilityq
tj ), t = 1,2, ..., m, j = 1,2, ..., n. (5)
qH4
tj =
Qt
X
q=1
uH4(Sensibilityq
tj ), t = 1,2, ..., m, j = 1,2, ..., n. (6)
qH5
tj =
Qt
X
q=1
uH5(Sensibilityq
tj ), t = 1,2, ..., m, j = 1,2, ..., n. (7)
qH6
tj =
Qt
X
q=1
uH6(Sensibilityq
tj ), t = 1,2, ..., m, j = 1,2, ..., n. (8)
Definition 4 Let npH1
tj , pH2
tj , ..., pHr
tj odenote the percentages of reviews with
different evaluation scale on travel destination Xtconcerning the attribute
Cj. Then the probability distribution npH1
tj , pH2
tj , ..., pHr
tj ocan be respectively
calculated:
pH1
tj =qH1
tj
qH1
tj +qH2
tj +qH3
tj +qH4
tj +qH5
tj +qH6
tj
, t = 1,2, ..., m, j = 1,2, ..., n,
(9)
pH2
tj =qH2
tj
qH1
tj +qH2
tj +qH3
tj +qH4
tj +qH5
tj +qH6
tj
, t = 1,2, ..., m, j = 1,2, ..., n,
(10)
pH3
tj =qH3
tj
qH1
tj +qH2
tj +qH3
tj +qH4
tj +qH5
tj +qH6
tj
, t = 1,2, ..., m, j = 1,2, ..., n,
(11)
pH4
tj =qH4
tj
qH1
tj +qH2
tj +qH3
tj +qH4
tj +qH5
tj +qH6
tj
, t = 1,2, ..., m, j = 1,2, ..., n,
(12)
pH5
tj =qH5
tj
qH1
tj +qH2
tj +qH3
tj +qH4
tj +qH5
tj +qH6
tj
, t = 1,2, ..., m, j = 1,2, ..., n,
(13)
pH6
tj =qH6
tj
qH1
tj +qH2
tj +qH3
tj +qH4
tj +qH5
tj +qH6
tj
, t = 1,2, ..., m, j = 1,2, ..., n.
(14)
Thus, we know that pH1
tj +pH2
tj +pH3
tj +pH4
tj +pH5
tj +pH6
tj = 1 and pH1
tj , pH2
tj , ..., pH6
tj ≥
0, t = 1,2, ..., m, j = 1,2, ..., n.
14 Jian Wu et al.
3.3 Estimate incomplete preference value for group travel destination
evaluation
Different from the general group decision, group travel destination evaluation
usually involves incomplete preference of some attributes[38,39]. Due to lack
of information, users may have no preference of attributes, such as service,
food. Meanwhile, users may have more information on attributes, such as traf-
fic and entertainment. To have a reasonable decision result, effective methods
are proposed to evaluate the missing preference[15,33, 40–46]. Herrera et al.[47]
proposed a method to complete missing preference value using the information
he/she provides based on the principle of consistency. For ensuring the great-
est consistency of preference, Chiclana’s incomplete information evaluation
method is applied in this study. Thus, the estimation processes of incomplete
preference value include the following two parts: (1) Construct the comple-
mentary preference matrix; (2) Estimate incomplete preference values.
3.3.1 Construct the complementary preference matrix
This study allows decision-makers in the group to use a complementary judg-
ment matrix to give the incomplete preference of each attribute.
Definition 5 (Complementary preference matrix) Complementary pref-
erence matrix for attribute Cito Cjcan be denoted as E= (eij )n×n, where
eij represents preference value and the specific meaning is as follows[12]:
eij = 1 indicates that Ciis completely superior to Cj,
eij ∈(0.5,1) shows that Ciis better than Cjto a certain extent,
eij = 0.5means that there is no difference between Ciand Cj.
eij +eji = 1∀i, j ∈ {1,2, ..., n}
3.3.2 Estimate incomplete preference values
In reality, users may not be able to give accurate values due to time pressure or
not accurately distinguish the importance of two attributes, which may result
in missing values. So that it is necessary to estimate the missing preference
values. In this study, the method proposed by Chiclana[15] is used to estimate
the missing value based on the principle of preference consistency. The advan-
tage of this method is that the estimation of the missing value of a user is only
based on the value of the user’s preference without considering other users’ in
the group.
Definition 6 Let Ube a uninorm operator with strong negation N(x)=1−x,
and the fuzzy preference values of all attributes are consistent with U[15], if
∀i, j, l : (eil, elj )/∈ {(0,1),(1,0)} ⇒ eij =U(eil, elj ) (15)
U(x, y) = 0,(x, y)∈ {(0,1),(1,0)}
xy
xy+(1−x)(1−y), otherwise (16)
Title Suppressed Due to Excessive Length 15
Then, the value eij can be estimated by the intermediate attribute Cl
url
ij =U(eil, elj ),∀i, j, l : (eil, elj )/∈ {(0,1),(1,0)}(17)
For estimating the missing value, the following symbols are introduced as
follows:
A={(i, j)|i, j ∈ {1,2, ..., n} ∧ i6=j}
MV ={(i, j)∈A|eij ∧eji unknown}
EV =A\M V
H01
ij =l∈E01
ij |(i, j)∈M V ∧(i, l),(l, j )∈EV
where E01
ij ={l6=i, j|(eil , elj )/∈ {(0,1),(1,0)}},M V represents the miss-
ing value in the complementary judgment matrix, EV means the known value
in the matrix, H01
ij indicates the missing value eij, which is estimated by the
uninorm operator with the help of the intermediate attribute Clunder the
premise that eil and elj are known.
Definition 7 The estimated value urij of the missing value is defined as the
average of the estimated values predicted by using all possible intermediate
attributes Cl[12]
urij =P
l∈H01
ij
url
ij
#H01
ij
, if H01
ij 6= 0 (18)
3.4 Group consensus reaching processes
In the process of group tourism decision-making, the preference for attributes
of all the users may not reach a consensus. To have an acceptable decision-
making results with a minimum adjustment cost, this study utilizes an in-
teractive consensus reaching method based on the minimum adjustment cost
feedback mechanism, which is consist of the following processes: (1) Consensus
measurement; (2) Group optimization model based on minimum adjustment
cost feedback mechanism[33].
3.4.1 The three-level’ consensus degree measurement
To measure the consensus degree, the collective matrix can be calculated as:
Definition 8 Let Eh= (eh
ij )n×n(h= 1,2, ..., k)be a complementary prefer-
ence matrix given by user eh. When each member expresses his/her preference
values for all attributes, the collective matrix E= (eij )n×ncan be calculated
as:
eij =
k
X
h=1
1
k·eh
ij , i, j = 1,2, ..., n (19)
16 Jian Wu et al.
Therefore, the consensus degree of each user can be obtained by three
levels:
Level 1. Consensus degree on preference values level. The consensus degree
of user eh’s preference value eh
ij relative to the group is:
CE h
ij = 1 −
eh
ij −eij
(20)
Level 2. Consensus degree on attribute level. The consensus degree of user
eh’s preference for attribute Cirelative to the group is:
CAh
i=1
n−1
n
X
j=1,j6=i
CE h
ij (21)
Level 3. Consensus degree on user level. The consensus degree of user eh
relative to the group is:
CI h=1
n
n
X
i=1
CAh
i(22)
Based on the consensus degree, the user who is inconsistent with the group
can be identified. If all users reach the consensus, the collective matrix can
be obtained, and then the weight of each attribute can be calculated to rank
the alternatives. Otherwise, the inconsistent user will be identified and recom-
mendations will be generated for the inconsistent user based on the minimum
cost feedback adjustment mechanism[33] until the group reaches consensus.
3.4.2 Group optimization model based on minimum adjustment cost feedback
mechanism
(1)Identification of the inconsistent preference value
Level 1. The set of users (denoted as E XP C H) whose C I his below the
threshold γare obtained as:
EX P CH =h|CI h< γ
Level 2. After the inconsistent users are identified, the set of attributes (de-
noted as ALT ) which are lower than the threshold γare identified as:
ALT =(h, i)|h∈EX P CH ∧CAh
i< γ
Level 3. Finally, the set of preference values (denoted as AP S) which are
lower than the threshold γcan be identified at the preference level:
AP S =(h, i, j)|(h, i)∈ALT ∧C Eh
ij < γ
Title Suppressed Due to Excessive Length 17
(2)The generation of recommendation based on boundary feedback param-
eters.
After the identification process is done, the personalized recommendation
can be conducted for the inconsistent user and the new preference values for
the inconsistent preference values in AP S can be generated. So that after the
user accepts the recommendation, the consensus level will be improved.
For all (h, i, j)∈AP S , the inconsistent user ehneeds to change the pref-
erence value eh
ij to a value is closer to eh0
ij .
eh0
ij = (1 −λh)eh
ij +λheij (23)
where λhis the feedback parameter that controls the degree of recommen-
dation, λh∈[0,1], and eij is calculated by Eq.(19).
When λ= 0, the inconsistent individual maintains the original preference
and does not accept the recommendation. Otherwise, when λ= 1, the indi-
vidual abandons the original preference completely and replaces the original
preference with the collective value. It indicates that the larger the feedback
parameter λ, the more recommendations will be adopted, and the higher the
adjustment cost of inconsistent users will pay. Therefore, this study employs
the boundary feedback parameter λbased on the minimum cost feedback
mechanism[33] in the group interaction, which improves the group consensus
level and reduces the individual adjustment cost at the same time.
The new preference values matrices for the weight of attributes are consist
of two parts: inconsistent user’s preference matrix Ehand consistent users’
preference matrix Es. After the inconsistent user ehaccepts the recommen-
dation, the new preference value {Eh0= (eh0
ij )|eh0
ij = (1 −λh)eh
ij +λheij , i, j ∈
AP S}and the set of unchanged values {Es0=Es= (es0
ij )|es0
ij =es
ij , s =
1,2, ..., k, s 6=h}can be obtained. Then the new collective matrix can be
expressed as E0= (eij 0)n×n, in which the eij0can be calculated as:
eij 0=1
k·eh0
ij +
k
X
s=1
1
k·es
ij , i, j = 1,2, ..., n, s = 1,2, ..., k, s 6=h(24)
Definition 9 After the inconsistent user accepts the recommendations, the
adjustment cost can be calculated as follows[33]:
F=
Eh0−Eh
=X
h,i,j∈AP S
λh
eij −eh
ij
(25)
Therefore, the minimum cost feedback model is constructed as:
min P
h,i,j∈AP S
λh
eij −eh
ij
CI hEh0, E 0≥γ
CI sEs0, E 0≥γ, s = 1, ..., k, s 6=h
0≤λh≤1
(26)
18 Jian Wu et al.
By resolving model(26), the boundary feedback parameter λmin can be
obtained.
3.5 The selection process
Once all users reach the consensus, the new collective matrix E0can be ob-
tained. Then, the vector of attribute weights w= (w1, w2, ..., wn) of each at-
tribute can be obtained by transit method (MTM)[48]. By assembling the
sentiment matrix, the collective evaluation matrix Xt= (Hr, xr)1×6(t=
1,2, ..., m;r= 1,2, ..., 6) of each alternative travel destination can be cal-
culated by a weighted arithmetic averaging (WAA) operator.
wj=
n
P
i=1
eij 0+n
2−1
n(n−1) (27)
xr
t=
n
X
j=1
pHr
tj ∗wj(28)
Then, the expected score[49] of each alternative can be calculated as:
EXt=
6
X
r=1
Hr
t·xr
t(29)
Finally, the ranking of alternative travel destination can be obtained.
3.6 The processes of the proposed group travel destination evaluation model
To solve the problem of selection of travel destination for group travel, a group
travel destination evaluation model based on online reviews is proposed, which
is shown in Fig.2. The model can be summarized as:
(1) Construct the percentage distribution on evaluations concerning each al-
ternative based on online reviews. This study crawls and processes online
text reviews and obtains the sentiment orientations. After calculating the
sentiment orientation, the online text reviews are converted into a sen-
timent matrix with percentage distributions. Thus, the tourists’ opinions
and sentiments on the tourist attributes of all the alternatives are obtained
to construct a sentiment matrix.
(2) Complete the missing preference value in the complementary judgment
matrix. In this study, the incomplete complementary matrix is used to
represent users’ preference on attributes. The method proposed by Chi-
clana et al.[15] is used to predict the missing value, in which the uninorm
operater is used to estimate missing values.
Title Suppressed Due to Excessive Length 19
(3) Conduct the group consensus reaching processes. This study employs three-
level consensus degrees to measure the consensus level: the preference val-
ues’ level, the attributes’ level, and users’ level. According to the proposed
threshold value of consensus, users who are inconsistent with the group
are identified. Based on the minimum adjustment cost feedback mech-
anism, recommendations are generated for inconsistent users. After all
users’ consensus degree reaches the given threshold value, the group con-
sensus reaching processes end.
(4) Select the most appropriate alternatives. After all the members reached
the consensus, the new collective preference matrix can be calculated.
Then, the weight vector of attributes can be obtained. By aggregating the
percentage distribution with the weight vector of attributes, the ranking
of all alternatives can be obtained.
Fig. 2: The structure of group travel destination evaluation model
20 Jian Wu et al.
In summary, the pseudo code of the algorithm for the group consensus-
based travel destination evaluation method with online reviews is shown in
Algorithm4.
Algorithm 4 The group consensus-based travel destination evaluation
method with online reviews
Input:
The online reviews set R={R1, R2, ..., Rm};
The sentiment word list O={O+,O −};
The adverbs of degree list D={D1, D2, D3};
The negation words set Nw;
The members’ complementary matrices on pairwise comparisons of attributes Eh=
(eh
ij )n×n
Output:
The most satisfied destination for group travel
Step 1: Each review in R={R1, R2, ..., Rm}is preprocessed using jieba word-
segmentation tool;
Step 2: Establish feature word list (shown in Table.3) using Algorithm1;
Step 3: Obtain the opinion sentences set R0={R10, R20, ..., Rm0}using Algorithm2;
Step 4: Identify the sentiment orientation of Sq
tj using Algorithm3;
Step 5: Calculate the probability distribution matrix Sof each travel destination Xt
using Eqs.(9)-(14);
Step 6: Estimate the missing preference in complementary matrix Eh= (eh
ij )n×nusing
Eqs.(16)-(18);
Step 7: Genarate recommendations for the identified inconsistent user based on the mini-
mum adjustment cost feedback mechanism using Eqs.(25)(26) until all the members reach
the consensus;
Step 8: Obtain the new collective matrix E0using Eq.(24);
Step 9: Aggregate the sentiment matrix Swith the weight vector of attributes calculated
using Eq.(27);
return The ranking of all alternatives.
4 Example
In this section, an example of a group travel destination evaluation with online
reviews is given to illustrate the use of the proposed method.
Company A is a medium-sized enterprise in China. In order to improve the
work enthusiasm of employees, the company selects five outstanding employees
every year and rewards these employees for group tours. To select the most
satisfactory travel destination for all outstanding employees, the group deci-
sion will be made by the five employees (denoted as E={e1, e2, e3, e4, e5}), in
which six attributes are considered: food, accommodation, traffic, sightseeing,
shopping, entertainment (denoted as C={C1, C2, C3, C4, C5, C6}). The com-
pany gives four alternative travel destinations (denoted as X={X1, X2, X3, X4}).
4.1 Sentiment analysis based on online reviews
In this case, online text reviews of the four travel destinations of Hua Mountain
(X1), Emei Mountain (X2), Taishan Mountain (X3)and Huang Mountain
Title Suppressed Due to Excessive Length 21
(X4) are extracted from Ctrip.com and Dazhong.com, including 7900 reviews
of travel destination X1, 7200 reviews of travel destination X2, 7500 reviews of
travel destination X3, 9500 reviews of travel destination X4. Fig.3 illustrates
an example of travel destination reviews on Dazhong.com, which displays the
representing image and some details, including its total score, surroundings,
service, number of reviews, and average price of Mount Emei. After data fil-
tering and cleaning, there are 7798, 7069, 7346, and 9336 reviews left for each
travel destination respectively. Fig.4 shows the details of reviews from Mount
Emei on Dazhong.com. The userId and the text reviews are extracted in Fig.4.
Fig. 3: A screenshot from Dazhong.com
Fig. 4: The details of reviews from Mount Emei on Dazhong.com
22 Jian Wu et al.
4.1.1 Feature words and opinion sentences extraction
By preprocessing the online reviews, each review is decomposed into several
words. Take the alternative Mount Emei X2as an example, the obtained re-
processed sets of online reviews concerning X2are shown in Table.2. Next,
using Algorithm 1, 173 features are extracted as feature words set, and then
they are classified manually to obtain the feature word list, which is shown
in Table.3. For example, the “restaurant” is a feature to evaluate the “dining
environment” in the secondary level of the feature word list, so that it should
be sorted to the “food” attribute in the six attributes of tourism. Similarly,
the “sunset” is a feature to evaluate a “scenery” in the secondary level of the
feature word list, so that it should be sorted to the “sightseeing” attribute.
According to Algorithm.2, the opinion sentences set is obtained after the sen-
tence of reviews is preprocessed and classified.
4.1.2 Identifying the sentiment orientation of opinion sentences
The feature-opinion pairs are extracted and the dependency parsing is imple-
mented on each opinion sentence to extract the modifiers. Then HowNet is
introduced to identify the sentiment orientation of each opinion sentence. Ac-
cording to Algorithm.3, the sentiment orientation of each feature is obtained.
Take the alternative Mount Emei X2as an example, the obtained sentiment
orientations concerning each feature are shown in Table.4.
4.1.3 Sentiment score probability distribution representation
The probability distribution under each evaluation scale with all travel at-
tributes of each travel destination is obtained by Eqs.(9)-(14), the results are
shown in Table.5.
Table 2: The obtained reprocessed sets of online reviews on X2
The online reviews on travel desti-
nation X2
The reprocessed results on travel
destination X2
q= 1 The scenery is very beautiful, jind-
ing is extremely spectacular
scenery/n, very/d, beautiful/a,
jinding/n, extremely/d, spectacu-
lar/a
q= 2 The environment is very good, the
tickets are a bit expensive
environment/n, very/d, good/a,
tickets/n a bit/d, expensive/a
q= 3 The service is very user-friendly service/vn, very/d, user-friendly/a
... ... ...
q= 7069 ... ...
Note. 1 : n(none); vn(verb none); d(adverb); a(adjective).
Title Suppressed Due to Excessive Length 23
Table 3: The feature word list of the four travel destinations
Tourism at-
tribute
Second level Features
Food
Food Food, gourmet, snacks, xiaochi, breakfast,
lunch, dry food, instant noodles, paomian, ra-
men, sausage, kaochang, bread, biscuits, pan-
cakes, steamed meat, tofu, fruit, cucumber,
mineral water
Taste Taste
Environment Hotel, restaurant, diner, dining hall
Accommodation Accommodation Accommodation, hotel, guesthouse, lvguan, lv-
dian, inn, homestay, room, standard room, bed
Traffic
Travel mode High-speed rail, gaotie, dongche, train, bus,
minibus, xiaoba, daba, taxi, car, shuttle, black
car, walk
Transportation
fee
Ticket, bus ticket, train ticket
Convenience
degree
Station, train station, bus station, parking lot,
traffic, line, route, driver, drive, distance, ride
Sightseeing
Ticket price Price, buy tickets, sell tickets, goupiao, book
tickets, fares, shoupiao, tickets, packages, half
price, cost-effective
Scenery Snow, jixue, heavy snow, heavy fog, clouds,
fog, rime, heavy rain, sunset, sunset, sunrise,
beautiful scenery, scenery, wonders, jingse,
jingxiang, fengguang, air, climate, environ-
ment, wonderland, weather, temperature, qi-
wen
Attractions Jinding, foding, shanding, peak, main peak,
famous mountain, bodhisattva, foxiang, jinx-
iang, holy land, temple, baoguo temple,
miaoyu, simiao, daochang, daoguan, monkey,
monkey group, monkey area, squirrel, xixi-
ang Pool, jinsuoguan, furrows, canyons, lion
peaks, jiyubei, guangmingding, tianwai village,
attractions, landscapes, scenic spots, observa-
tion decks
Sightseeing
value
Culture, cultural heritage, humanities, cul-
tural landscape, history
Scenic service ID card, tour guide, staff, internet, queuing,
toilet, restroom, garbage, street lamp, service
center, service attitude
Shopping Goods Crampons, raincoats, bamboo poles, shoe cov-
ers, gloves, tea, walking sticks, tents, souvenirs,
shopping
Price Price
Entertainment Entertainment
facilities
Cable car, sightseeing car, touring car, roller
coaster, sliding rod, huagan, mountain road,
ground rail
Entertainment
project
Mountain climbing, climbing to the top,
changkong zhandao, yaozi fanshen, ladder,
yunti, tianti, project, Taohuayuan ropeway
24 Jian Wu et al.
Table 4: The obtained sentiment orientation concerning each feature on travel
destination X2
Opinion sentences Features Tourism at-
tributes
sentiment
orientation
q= 1 The scenery is particu-
larly beautiful
scenery sightseeing 2
q= 2 The air is very good air sightseeing 2
q= 3 Jinding is so wonderful jinding sightseeing 2
q= 4 The sunrise is very
beautiful
sunrise sightseeing 2
q= 5 Accommodation is a
bit expensive
accommodation accommodation -0.5
q= 6 Taste is good taste food 1
q= 7 The cable car is too
crowded
cable car entertainment -1
q= 8 Tickets are a bit expen-
sive
tickets sightseeing -0.5
... ... ... ... ...
Table 5: The percentage distribution pH1
tj , pH2
tj , ..., pH6
tj on evaluations with each
alternative travel destination based on online reviews
C1C2
H1H2H3H4H5H6H1H2H3H4H5H6
X10.102 0.136 0.033 0.068 0.339 0.322 0.208 0.179 0.052 0.029 0.150 0.382
X20.222 0.148 0.037 0.000 0.130 0.463 0.226 0.185 0.058 0.045 0.131 0.355
X30.167 0.097 0.028 0.028 0.333 0.347 0.204 0.178 0.045 0.028 0.137 0.408
X40.189 0.203 0.041 0.027 0.189 0.351 0.167 0.221 0.036 0.060 0.123 0.393
C3C4
H1H2H3H4H5H6H1H2H3H4H5H6
X10.161 0.194 0.096 0.065 0.129 0.355 0.122 0.157 0.039 0.032 0.094 0.556
X20.150 0.300 0.075 0.075 0.150 0.250 0.126 0.131 0.033 0.039 0.122 0.549
X30.269 0.205 0.026 0.051 0.090 0.359 0.079 0.149 0.027 0.035 0.113 0.597
X40.149 0.170 0.043 0.021 0.213 0.404 0.132 0.143 0.031 0.031 0.116 0.547
C5C6
H1H2H3H4H5H6H1H2H3H4H5H6
X10.112 0.224 0.016 0.096 0.104 0.448 0.217 0.072 0.030 0.058 0.058 0.565
X20.098 0.170 0.012 0.053 0.125 0.542 0.246 0.246 0.000 0.052 0.123 0.333
X30.126 0.167 0.058 0.086 0.126 0.437 0.295 0.169 0.037 0.054 0.120 0.325
X40.106 0.165 0.026 0.043 0.133 0.527 0.303 0.184 0.026 0.053 0.118 0.316
Note.2 : The feature words of tourism destination evaluation can be classified into a two-level
evaluation attributes system (show as Table 3). Due to the limited space of this article,
there are too many second-level indexes to display. In this example, the feature words
abstracted have been categorized into the first-level attributes of the attributes system to
construct the probability distribution matrices.
Title Suppressed Due to Excessive Length 25
4.2 Estimation of the missing value of preference matrix
E1=
0.50 −0.80 0.70 0.70 0.60
−0.50 0.80 0.60 − −
0.20 0.20 0.50 0.40 0.30 0.30
0.30 0.40 0.60 0.50 0.40 0.40
0.30 −0.70 0.60 0.50 0.50
0.40 −0.70 0.60 0.50 0.50
;E2=
0.50 0.80 0.80 −0.70 0.10
0.20 0.50 0.60 −0.40 0.20
0.20 0.40 0.50 0.40 0.60 0.00
− − 0.60 0.50 0.40 0.10
0.30 0.60 0.40 0.60 0.50 0.20
0.90 0.80 1.00 0.90 0.80 0.50
;
E3=
0.50 0.30 0.40 0.40 0.20 0.50
0.70 0.50 0.80 0.40 0.60 0.50
0.60 0.20 0.50 0.50 0.40 0.40
0.60 0.60 0.50 0.50 0.30 0.40
0.80 0.40 0.60 0.70 0.50 0.60
0.50 0.50 0.60 0.60 0.40 0.50
;E4=
0.50 1.00 −0.70 0.80 −
0.00 0.50 0.60 0.70 0.80 0.70
−0.40 0.50 0.80 0.70 0.60
0.30 0.30 0.20 0.50 0.50 0.60
0.20 0.20 0.30 0.50 0.50 0.60
−0.30 0.40 0.40 0.40 0.50
;
E5=
0.50 0.60 0.80 0.30 0.30 0.50
0.40 0.50 0.80 0.30 0.30 −
0.20 0.20 0.50 0.20 0.20 0.30
0.70 0.70 0.80 0.50 0.50 0.60
0.70 0.70 0.80 0.50 0.50 0.50
0.50 −0.70 0.40 0.50 0.50
The missing preference value is calculated by Eqs.(16)-(18) based on the
known value in the matrix, then the complete preference matrix is shown as
follows:
E1=
0.50 0.56 0.80 0.70 0.70 0.60
0.44 0.50 0.80 0.60 0.59 0.57
0.20 0.20 0.50 0.40 0.30 0.30
0.30 0.40 0.60 0.50 0.40 0.40
0.30 0.41 0.70 0.60 0.50 0.50
0.40 0.43 0.70 0.60 0.50 0.50
;E2=
0.50 0.80 0.80 0.71 0.70 0.10
0.20 0.50 0.60 0.56 0.40 0.20
0.20 0.40 0.50 0.40 0.60 0.00
0.29 0.44 0.60 0.50 0.40 0.10
0.30 0.60 0.40 0.60 0.50 0.20
0.90 0.80 1.00 0.90 0.80 0.50
;
E3=
0.50 0.30 0.40 0.40 0.20 0.50
0.70 0.50 0.80 0.40 0.60 0.50
0.60 0.20 0.50 0.50 0.40 0.40
0.60 0.60 0.50 0.50 0.30 0.40
0.80 0.40 0.60 0.70 0.50 0.60
0.50 0.50 0.60 0.60 0.40 0.50
;E4=
0.50 1.00 0.67 0.70 0.80 0.85
0.00 0.50 0.60 0.70 0.80 0.70
0.33 0.40 0.50 0.80 0.70 0.60
0.30 0.30 0.20 0.50 0.50 0.60
0.20 0.20 0.30 0.50 0.50 0.60
0.15 0.30 0.40 0.40 0.40 0.50
;
E5=
0.50 0.60 0.80 0.30 0.30 0.50
0.40 0.50 0.80 0.30 0.30 0.43
0.20 0.20 0.50 0.20 0.20 0.30
0.70 0.70 0.80 0.50 0.50 0.60
0.70 0.70 0.80 0.50 0.50 0.50
0.50 0.57 0.70 0.40 0.50 0.50
26 Jian Wu et al.
4.3 Consensus interaction among group members
Based on the preference matrix, the consensus degrees of five users are calcu-
lated using Eqs.(19)-(22).
(1) The consensus degrees at preference value level are:
CE 1=
1.00 0.91 0.89 0.86 0.84 0.91
0.91 1.00 0.92 0.91 0.95 0.91
0.89 0.92 1.00 0.94 0.86 0.98
0.86 0.91 0.94 1.00 0.98 0.98
0.84 0.95 0.86 0.98 1.00 0.98
0.91 0.91 0.98 0.98 0.98 1.00
;CE 2=
1.00 0.85 0.89 0.85 0.84 0.59
0.85 1.00 0.88 0.95 0.86 0.72
0.89 0.88 1.00 0.94 0.84 0.68
0.85 0.95 0.94 1.00 0.98 0.68
0.84 0.86 0.84 0.98 1.00 0.72
0.59 0.72 0.68 0.68 0.72 1.00
;
CE 3=
1.00 0.65 0.71 0.84 0.66 0.99
0.65 1.00 0.92 0.89 0.94 0.98
0.71 0.92 1.00 0.96 0.96 0.92
0.84 0.89 0.96 1.00 0.88 0.98
0.66 0.94 0.96 0.88 1.00 0.88
0.99 0.98 0.92 0.98 0.88 1.00
;CE 4=
1.00 0.65 0.98 0.86 0.74 0.66
0.65 1.00 0.88 0.81 0.74 0.78
0.98 0.88 1.00 0.66 0.74 0.72
0.86 0.81 0.66 1.00 0.92 0.82
0.74 0.74 0.74 0.92 1.00 0.88
0.66 0.78 0.72 0.82 0.88 1.00
;
CE 5=
1.00 0.95 0.89 0.74 0.76 0.99
0.95 1.00 0.92 0.79 0.76 0.95
0.89 0.92 1.00 0.74 0.76 0.98
0.74 0.79 0.74 1.00 0.92 0.82
0.76 0.76 0.76 0.92 1.00 0.98
0.99 0.95 0.98 0.82 0.98 1.00
(2) The consensus degrees at attribute level are:
CA1= (0.88,0.92,0.92,0.93,0.92,0.95);
CA2= (0.80,0.85,0.85,0.88,0.85,0.68);
CA3= (0.77,0.88,0.89,0.91,0.86,0.95);
CA4= (0.78,0.77,0.80,0.81,0.80,0.77);
CA5= (0.87,0.87,0.86,0.80,0.84,0.94)
(3) The consensus degrees at user level are:
CI 1= 0.92; C I 2= 0.82; C I 3= 0.88; C I 4= 0.79; C I 5= 0.86
The threshold value of consensus γ= 0.8 has been given in this example.
Then, the inconsistent user e4is identified. Subsequently, the feedback mech-
anism is activated to help modify his/her preference values to improve the
consensus level.
First, the inconsistent preferences of 3-tuples AP S can be identified by a
three-level identification method as:
AP S ={(4,1,2),(4,1,5),(4,1,6),(4,2,1),(4,2,5),(4,2,6),(4,3,6),(4,5,1),(4,5,2),(4,6,1),(4,6,2),(4,6,3)}
Title Suppressed Due to Excessive Length 27
Based on the assumed threshold value of γ= 0.8, the above model (26)
can be further optimized as:
min P
h,i,j∈AP S
λh
¯eij −eh
ij
CI hEh0, E 0≥0.8
CI sEs0, E 0≥0.8, s = 1, ..., k, s 6=h
0≤λh≤1
By solving the model, the boundary feedback parameter λmin = 0.12. Thus,
the recommendations for user e4are:
You should change preference value of attribute C1on attribute C2close to
the value 0.96;
You should change preference value of attribute C1on attribute C5close to
the value 0.77;
You should change preference value of attribute C1on attribute C6close to
the value 0.81;
You should change preference value of attribute C2on attribute C1close to
the value 0.04;
You should change preference value of attribute C2on attribute C5close to
the value 0.77;
You should change preference value of attribute C2on attribute C6close to
the value 0.67;
You should change preference value of attribute C3on attribute C6close to
the value 0.57;
You should change preference value of attribute C5on attribute C1close to
the value 0.23;
You should change preference value of attribute C5on attribute C2close to
the value 0.23;
You should change preference value of attribute C6on attribute C1close to
the value 0.19;
You should change preference value of attribute C6on attribute C2close to
the value 0.33;
You should change preference value of attribute C6on attribute C3close to
the value 0.43.
After user e4adopts the recommendations, the new collective preference
matrix E0would be
E0=
0.50 0.64 0.69 0.56 0.53 0.50
0.36 0.50 0.72 0.51 0.53 0.47
0.31 0.28 0.50 0.46 0.44 0.31
0.44 0.49 0.54 0.50 0.42 0.42
0.47 0.47 0.56 0.58 0.50 0.48
0.50 0.53 0.69 0.58 0.52 0.50
28 Jian Wu et al.
The new consensus degrees of all the members are calculated using Eqs.(19)-
(22):
CI 1= 0.92; C I 2= 0.82; C I 3= 0.88; C I 4= 0.80; C I 5= 0.87
4.4 The selection process of tourist alternatives
According to the above results, all the members have reached the consensus
level. Hence, the weight vector of attributes would be computed based on the
new collective matrix E0by Eq.(27): w= (0.18,0.17,0.14,0.16,0.17,0.18). By
assembling the percentage distribution (shown as Table.5) of the four alterna-
tives under six attributes of six evaluation scales with above weights, we can
get the collective evaluation values:
X1=(H1
1,0.15),(H2
1,0.16),(H3
1,0.04),(H4
1,0.06),(H5
1,0.15),(H6
1,0.44)
X2=(H1
2,0.18),(H2
2,0.19),(H3
2,0.04),(H4
2,0.04),(H5
2,0.13),(H6
2,0.42)
X3=(H1
3,0.19),(H2
3,0.16),(H3
3,0.03),(H4
3,0.05),(H5
3,0.16),(H6
3,0.41)
X4=(H1
4,0.18),(H2
4,0.18),(H3
4,0.03),(H4
4,0.04),(H5
4,0.15),(H6
4,0.42)
Then, the expected scores of each travel destination can be calculated by
Eq.(29) as:
E(X1)=0.58, E(X2)=0.42, E (X3)=0.45, E(X4)=0.46
Thus, the ranking of the alternatives is: X1X4X3X2
Therefore, Hua Mountain is the most satisfied destination for group travel of
outstanding employees.
5 Conclusions
This study proposed a novel group consensus-based travel destination evalua-
tion method with online reviews. In this method, the percentage distribution
on evaluations of each alternative based on the sentiment analysis of online
reviews is calculated to construct a sentiment matrix. Due to the lack of infor-
mation and time, the missing preference values in the complementary matrix
of attributes are estimated by the consistency-based method. Then, the incon-
sistent users are asked to revise their preference values based on the minimum
adjustment cost feedback mechanism, so that the group consensus could be
reached. Finally, the decision matrix of the percentage distribution can be ag-
gregated with the weight vectors. Therefore, the ranking of travel destinations
can be obtained. The proposed method has the following advantages:
(1) Different from the expert decision-making problem, travel destination eval-
uation is a representative non-expert group decision-making problem, which
denotes that decision-makers barely have decision opinion. Instead of the
questionnaire method in the traditional GDM, the decision opinions and
the attributes are obtained by sentiment analysis of online reviews, which
resolves the non-expert decision-making problems and makes the decision
results more objective.
Title Suppressed Due to Excessive Length 29
(2) As group travel is an emerging way of traveling, the group destination
evaluation is the problem to be resolved. Group consensus and the in-
complete preference for obtaining the weight vector of attributes are two
key issues, which were ignored by the existing travel destination evalu-
ation method. In this study, to obtain the weight vector of attributes
for evaluating group travel destinations, two steps are conducted. First,
a consistency-based method is adopted to estimate missing values in the
complementary matrix of attributes. Second, the group users are asked to
reach group consensus based on the minimum adjustment cost feedback
mechanism. These two steps ensure the weight vector obtaining process
pays the minimum adjustment cost, at the same time, reaches consensus.
This study makes the above contributes but also has some limitations,
which may serve as avenues for future research. First, the data are extracted
from two travel websites, which may raise external validity issue, therefore, it is
necessary to extend the analysis to data from multiple websites for enhancing
the application of the proposed method. Second, fake reviews are universal
on websites, however, the issue of identifying fake reviews involves complex
research, therefore, the identification and treatment of fake reviews should be
paid more attention.
Acknowledgements The authors are very grateful to the anonymous referees for their
valuable comments and suggestions. This work was supported by National Natural Science
Foundation of China (NSFC) under the Grant No. 71971135, 72001134, 71571166. And
industrial and Informationalization Ministry of China for Cruise Program (No. 2018-473),
and Key Project of National Social and Scientific Fund Program (18ZDA052).
References
1. Zhang HY, Ji P, Wang JQ, Chen XH, A novel decision support model for satisfactory
restaurants utilizing social information: A case study of TripAdvisor.com, Tourism Man-
agement, 59, 281-297 (2017)
2. Sun XY, Han MX, Feng J, Helpfulness of online reviews: Examining review informative-
ness and classification thresholds by search products and experience products, Decision
Support Systems, 124 (2019)
3. Yu SM, Wang J, Wang JQ, Li L, A multi-criteria decision-making model for hotel selec-
tion with linguistic distribution assessments, Applied Soft Computing, 67, 741-755 (2018)
4. Li CC, Gao Y, Dong YC, Managing ignorance elements and personalized individual se-
mantics under incomplete linguistic distribution context in group decision making, Group
Decision and Negotiation (2020)
5. Zhang HJ, Xiao J, Dong YC, Integrating a consensus-reaching mechanism with bounded
confidences into failure mode and effect analysis under incomplete context, Knowledge-
Based Systems, 183, 1-20 (2019)
6. Dong YC, Zha QB, Zhang HJ, Herrera F, Consensus reaching and strategic manipulation
in group decision making with trust relationships, IEEE Transactions on Systems, Man,
and Cybernetics: Systems, 1-15 (2020)
7. Zha QB, Dong YC, Zhang HJ, Chiclana F, Herrera-Viedma E, A personalized feedback
mechanism based on bounded confidence learning to support consensus reaching in group
decision making, IEEE Transactions on Systems, Man, and Cybernetics: Systems, 1-11
(2019)
30 Jian Wu et al.
8. Cao MS, Wu J, Chiclana F, Ure˜na R, Herrera-Viedma E, A personalized feedback mech-
anism based on maximum harmony degree, IEEE Transactions on Systems, Man, and
Cybernetics: Systems, 1-13 (2020)
9. Gai TT, Cao MS, Cao QW, Wu J, Yu GF, Zhou M, A joint feedback strategy for con-
sensus in large-scale group decision making under social network, Computers & Industrial
Engineering, 147 (2020)
10. Dong YC, Chen X, Herrera F, Minimizing adjusted simple terms in the consensus reach-
ing process with hesitant linguistic assessments in group decision making, Information
Sciences, 297, 95-117 (2015)
11. Lu YL, Xu YJ, Herrera-Viedma E, Han YF, Consensus of large-scale group decision
making in social network: The minimum cost model based on robust optimization, Infor-
mation Sciences, 547, 910-930 (2021)
12. Wu J, Zhao ZW, Sun Q and Fujita H, A maximum self-esteem degree based feedback
mechanism for group consensus reaching with the distributed linguistic trust propagation
in social network,Information Fusion,67,80-93(2021)
13. Wu J, Sun Q, Fujita H and Chiclana F, An attitudinal consensus degree to control the
feedback mechanism in group decision making with different adjustment cost, Knowledge-
Based Systems,164,265-273(2019)
14. Dong YC, Zha QB, Zhang HJ, Kou G, Fujita H, Chiclana F and Herrera-Viedma E,
Consensus reaching in social network group decision making: Research paradigms and
challenges, Knowledge-Based Systems,162,3-13(2018)
15. Chiclana F, Herrera-Viedma E, Alonso S, Herrera F, A note on the estimation of missing
pairwise preference values: A uninorm consistency based method, International Journal of
Uncertainty, Fuzziness and Knowledge-Based Systems, 16(2), 19-32 (2018)
16. Kang D, Park Y, Review-based measurement of customer satisfaction in mobile service:
Sentiment analysis and VIKOR approach, Expert Systems with Applications, 41(4), 1041-
1050 (2014)
17. Liang D, Dai ZY, Wang MW, Assessing customer satisfaction of O2O takeaway based on
online reviews by integrating fuzzy comprehensive evaluation with AHP and probabilistic
linguistic term sets, Applied Soft Computing, 98 (2021)
18. Chen K, Kou G, Shang J, Chen Y, Visualizing market structure through online product
reviews: Integrate topic modeling, TOPSIS, and multi-dimensional scaling approaches,
Electronic Commerce Research and Applications, 14(1), 58-74 (2015)
19. Liang X, Liu PD, Wang ZH, Hotel selection utilizing online reviews: A novel decision
support model based on sentiment analysis and DL-VIKOR method, Technological and
Economic Development of Economy, 25(6), 1139-1161 (2019)
20. Kwok P, Laub H, Hotel selection using a modified TOPSIS-based decision support
algorithm, Decision Support Systems, 120, 95-105 (2019)
21. Peng HG, Zhang HY, Wang JQ, Cloud decision support model for selecting hotels on
TripAdvisor.com with probabilistic linguistic information, International Journal of Hospi-
tality Management,68, 124-138 (2018)
22. Zhou M, Liu XB, Chen YW, Qian XF, Yang JB, Wu J, Assignment of attribute weights
with belief distributions for MADM under uncertainties, Knowledge-Based Systems, 189,
1-18 (2020)
23. Zhang CB, Zhang HY, Wang JQ, Personalized restaurant recommendation method
combining group correlations and customer preferences, Information Sciences, 454-455,
128-143 (2018)
24. Wu J, Li X, Chiclana F, Yager R, An attitudinal trust recommendation mechanism to
balance consensus and harmony in group decision making, IEEE Transactions on Fuzzy
Systems, 27(11), 2163-2175 (2019)
25. Li Y, Chen X, Dong YC, Herrera F, Linguistic group decision making: Axiomatic dis-
tance and minimum cost consensus, Information Sciences, 541, 242-258 (2020)
26. Wu J, Cao MS, Chiclana F, Dong YC, Herrera-Viedma E, An optimal feedback model
to prevent manipulation behaviours in consensus under social network group decision
making, IEEE Transactions on Fuzzy Systems, 1-14 (2020)
27. Zhang BW, Dong YC, Feng X, Pedrycz W, Maximum fuzzy consensus feedback mecha-
nism with minimum cost and private interest in group decision making, IEEE Transactions
on Fuzzy Systems, 1-12 (2020)
Title Suppressed Due to Excessive Length 31
28. Wu J, Liu YJ, Liang CY, A consensus-and harmony-based feedback mechanism for
multiple attribute group decision making with correlated intuitionistic fuzzy sets, Inter-
national Transactions in Operational Research, 22(6): 1033-1054 (2014)
29. Liu YJ, Liang CY, Chiclana F, Wu J, A trust induced recommendation mechanism
for reaching consensus in group decision making, Knowledge-Based Systems, 119, 221-231
(2017)
30. Liu YJ, Liang CY, Chiclana F, Wu J, A knowledge coverage-based trust propagation
for recommendation mechanism in social network group decision making, Applied Soft
Computing Journal, 101 (2021)
31. Ure˜na R, Chiclana F, Melan¸con G, Herrera-Viedma E, A social network based approach
for consensus achievement in multiperson decision making, Information Fusion, 47, 72-87
(2019)
32. Gong ZW, Xu XX, Zhang HH, Ozturk U, Herrera-Viedma E, Xu C, The consensus
models with interval preference opinions and their economic interpretation, Omega, 55,
81-90 (2015)
33. Wu J, Dai LF, Chiclana F, Fujita H, Herrera-Viedma E, A minimum adjustment cost
feedback mechanism based consensus model for group decision making under social net-
work with distributed linguistic trust, Information Fusion, 41, 232-242 (2018)
34. Tu HL, Tang XB, Tourist sentiment analysis model building based on online reviews,
Journal of Modern Information, 36(4), 70-77 (2016)
35. Morente-Molinera J, Kou G, Pang C, Cabrerizo F, Herrera-Viedma E, An automatic
procedure to create fuzzy ontologies from users’ opinions using sentiment analysis proce-
dures and multi-granular fuzzy linguistic modelling methods, Information Sciences, 476,
222-238 (2019)
36. Liang X, Jiang YP, Gao M, Product selection methods based on online reviews, Journal
of Northeastern University(Natural Science), 38(1), 143-147 (2017)
37. Liu Y, Bi JW, Fan ZP, Ranking products through online reviews: A method based on
sentiment analysis technique and intuitionistic fuzzy set theory, Information Fusion, 36,
149-161 (2017)
38. Qin JD, Liu XW, Pedrycz W, An extended TODIM multi-criteria group decision making
method for green supplier selection in interval type-2 fuzzy environment, European Journal
of Operational Research, 258(2), 626-638 (2017)
39. Lin MW, Xu ZS, Zhai YL, Yao ZQ, Multi-attribute group decision-making under prob-
abilistic uncertain linguistic environment, Journal of the Operational Research Society,
69(2), 157-170 (2018)
40. Wang JQ, Zhang HY, Zhang Z, Fuzzy multi-criteria decision-making approach with
incomplete information based on evidential reasoning, Journal of Systems Engineering
and Electronics, 21(4), 604-608 (2010)
41. Liu D, Li TR, Zhang JB, A rough set-based incremental approach for learning knowl-
edge in dynamic incomplete information systems, International Journal of Approximate
Reasoning, 55(8), 1764-1786 (2014)
42. Liu JP, Liao XW, Yang JB, A group decision-making approach based on evidential
reasoning for multiple criteria sorting problem with uncertainty, European Journal of
Operational Research, 246(3), 858-873 (2015)
43. Bao TT, Xie XL, Long PY, Wei ZK, MADM method based on prospect theory and
evidential reasoning approach with unknown attribute weights under intuitionistic fuzzy
environment, Expert Systems with Applications, 88, 305-317 (2017)
44. Liang Q, Liao XW, Liu JP, A social ties-based approach for group decision-making
problems with incomplete additive preference values, Knowledge-Based Systems, 119, 68-
86 (2017)
45. Wu J, Chiclana F, Herrera-Viedma E, Trust based consensus model for social network in
an incomplete linguistic information context, Applied Soft Computing, 35, 827-839 (2015)
46. Liu HB, Ma Y, Jiang L, Managing incomplete preferences and consistency improvement
in hesitant fuzzy linguistic preference values with applications in group decision making,
Information Fusion, 51, 19-29 (2019)
47. Herrera-Viedma E, Chiclana F, Herrera F, Alonso S, Group decision-making model with
incomplete fuzzy preference values based on additive consistency, IEEE Transactions on
Systems, Man, and Cybernetics, Part B (Cybernetics), 37(1), 176-189 (2017)
32 Jian Wu et al.
48. Xu Y, Da L, Liu L, Normalizing rank aggregation method for priority of a fuzzy prefer-
ence relation and its effectiveness, International Journal of Approximate Reasoning, 50(8)
(2009)
49. Zhang GQ, Dong YC, Xu YF, Consistency and consensus measures for linguistic pref-
erence relations based on distribution assessments, Information Fusion, 17, 46-55 (2014)
