You are on page 1of 4

2015 International Conference on Intelligent Computing and Internet of Things (IC1T)

One Methodology for Spam Review Detection Based


on Review Coherence Metrics

Xinkai Yang
College of Information, Mechanical and Electronical Engineering, Shanghai Normal University
Shanghai, China, 200234
E-mail: xkyang@shnu.edu.cn

Abstract-In this paper, we propose an iterative computation breakfast with espresso machine, outstanding service, and
framework to detect spam reviews based on coherent great price. Can't fmd a thing to complain about it. "
examination. We first define some reviews' coherent metrics to
analyze review coherence in the granularity of sentence. Then The rest of this paper is organized as follows. In Section II,
the framework and its evaluation process are discussed in details. some related research works on spam reviews detection are
discussed. Based on coherence examination, we propose a
Keywords- spam review detection; coherence metric; word general framework in Section Ill, followed by the coherent
transition probability; word concurrence probability metrics defined in Section IV. We discuss our evaluation
process in Section V and conclude this paper in section VI.
I. INTRODUCTION
IT. RELATED WORKS
Nowadays product (including service, e.g. hotel, restaurant.)
reviews play an important role for consumers' online shopping Mostly, the features of review contents and reviewers'
activities. Many people read these reviews before making behaviors are used in early research works to detect spam
purchase decision. Normally, the reviews for a product may be reviews [9]. By representing a review using a set of review­
positive or negative. Customers will pay more attention to the level, reviewer-level and product-level features, a supervised
products with positive reviews and avoid the negative ones. classification method is applied to detect three types of fake
This effect will bring larger amount of business or lead to reviews in ref. [4]. Duplicated reviews are first identified using
potential financial losses [9]. Although lots of product reviews the Jaccard ratio, and these reviews are then used as the
are posted by real consumers to express their views and share training examples for the logistic regression classifier [4, 7]. In
their shopping experience with other people, more untruthful particular, untruthful review detection is performed by using
reviews appear in the e-business web sites because of financial duplicate reviews as labeled data.
reasons. For example, supposing a lodger writes a very Opinion and sentiment analysis of reviews extract and
negative review for a certain hotel on a review website due to aggregate positive and negative opinions from product reviews
its bad service. This review will present an unfavorable [5]. Researchers study the problem of generating feature-based
impression of this hotel to its potential customers and damage summ aries of customer reviews of products sold online. Given
its business. In order to avoid business losses caused by this a data set of customer reviews of any particular product, the
kind of truthful reviews, the owner of this hotel could employ task involves three subtasks: (1) identifying features of product
or entice some people to write untruthful reviews to promote its that customers have discussed or expressed their opinion on. (2)
reputation. Because of these pervasive untruthful reviews, For each feature, identifYing review sentences which gives an
customers are easy to be misled to buy low-quality opinion (positive or negative) and (3) producing a summary
services/products, while decent services/products could be using the discovered information [5].
defamed by malicious reviews. Usually, the writers who post
these deceptive reviews to intentionally mislead consumers or The other works study spam reviews by detecting
opinion analysis systems are called spammers and the fraudulent or unfair ratings. The review ratings are clustered
untruthful reviews are called spam reviews [9]. In order to into unfairly high ratings and unfairly low ratings by using
show how difficult spam detection is, we first depict one third party ratings products [4]. Those reviews with unfair
example reviews in the following part, which was posted to ratings can be removed to restore a fair item evaluation system.
ctrip.com (English version). This review is posted by one A new classification approach is developed to solve helpful
person who tries to promote that hotel and it is hard to identify prediction using review content and rating features. The
it as one deceptive review: reviews' rating features include the difference between the
review rating and the average rating of all reviews of the
"If you can put up with the small size of the room (which product. These features are then used to train a helpful
is normal in Japan) then you will love this place. Five minute classification method [4]. These works do not address review
walk from the main station, modem clean room, and great

978-1-4799-7534-1114/$31.00 ©2015 1EEE 99 Harbin, China


spammer detection directly. Some of them do not conduct assume that the review set D = {dj, d2, d3 ... dn} needs to be
evaluation of their techniques on real data. detected.
Xie and others tries to identify singleton reviews which are Defmition 1. Store-sentiment word pair Pi! consists of two
the only one review written by each user. A correlated temporal elements, one is from the online store set Sf, and the other one
pattern is proposed to detect such attacks via multidimensional is from the sentiment word set WOo
time series based on aggregate statistics, in order to depict and
mine such correlations [9]. In this way, they map the problem Pi! {<tj, OJ> I <tjESt, OJ E �J
=
(1)
of singleton spam detection to an abnormally correlated pattern This word pair is used to maintain the associative
detection problem. Then a hierarchical algorithm is proposed to information between a specific store and the sentiment word.
detect the time interval when such singleton spam attacks are We can infer that spammers will use positive sentiment words
likely to happen. to praise less reliable stores, or use negative sentiment words to
unreasonably complain about more reliable stores. This kind of
Recently, a more sophisticated graph-based method is
proposed in ref. [6] to detect review spammers by considering store-sentiment word pair may be helpful to find the clues of
such suspicious phenomena.
the relationships among reviews, reviewers and reviewed stores.
This kind of review spammer detection is preferred because We adopt a model containing hubs and authorities, which
gathering behavioral evidence of spammers is much easier and have some similarity with the popular graph ranking algorithm
scalable than that of spam reviews. A heterogeneous graph to [12]. There are two layers in this proposed model. The word
capture such relationships is constructed based on thorough pair layer is considered as hubs and the review layer as
analysis [6]. authorities. Fig.2 gives the bipartite graph representation of this
model. If a word pair occurs in one sentence of a review, there
Ill. OUR MODEL will be an edge between them. In Fig.l, the word pair that has
links to many reviews can be assigned a high weight to denote
Here we propose a general framework to detect spam
a strong associative degree between the store and a sentiment
reviews based on coherence examination. The assumptions we
word.
used in the proposed model is discussed firstly. Then several
reviews' coherent metrics are defmed. The iterative
computation framework based on these defmitions is provided . . . . . . . (0
fmally.
Usually spammers will benefit from their activity by
promoting some specific stores' prominence or defaming other
stores. The spam reviews have inner connections with certain
stores, which are either positive or negative depending on the
spammer's intention. Untruthful spam reviews always deviate
from the truth since that lousy stores suffer if customers know
the truth about them [2]. So the spam reviews will bear subtle
semantic incoherence among sentences in the reviews of the
Fig. 1. Link graph
same store. The main opinions shared by most surrounding
reviews, rather than outliers' opinion are helpful for customers
We denote the bipartite graph as G <Hp, Rd. Edp>, where
=

when they read reviews online.


Hp {Pij}
= is the set of all store-sentiment word pairs, Rr{dk}
In order to detect such spam reviews which are inconsistent is the set of reviews, Edp {eij I pi!EHp, dkERd} corresponds to
=

with the surrounding reviews, we propose a model to analyze the connection between reviews and store-sentiment word
review coherence in the granularity of sentence. Intuitively, pairs. Each edge ei! is associated with a weight w/E [0, 1]
spam reviews always use some untruthful sentiment words to denoting the contribution of Pi! to the review dk • The weight
express their deceptive feelings, to promote or defame a store w / is computed by the contribution of word pair Pi! in all

through such positive or negative feedback. These sentiment sentences of dk• Further metrics will be defined to investigate
words are linked with the targeted stores in the spam reviews. the coherence between these sentiment words and other related
So it is crucial to find the coherent relationships between the words to find more information.
specific store and the sentiment words among them.
IV. COHERENCE METRICS AND ALGORITHM
Based on the above assumptions, we consider each review
as a bag of sentences, and define a sentence set as S {Sj. S2.
=
In this part we provide several metrics to measure the
S3, .... sn } . For each sentence, we capture the intra-sentence coherence of a review based on the flow smoothness
information through the store-sentiment word pair and the information between sentences: 1) Word transition probability
coherent features which are pair-wise sentence coherence. In - conditional probability; 2) Word concurrence probability -
addition, we construct a sentiment word lexicon Vo. We also joint probability. Then the framework is provided.

100
A. Word transition probability OJ J
Con(spsj+l) max log(--'- ) (6)
Usually, human writings will demonstrate certain word =

lV;E,Sl,lV E,S2
j GiD;
transition patterns naturally between two consecutive
sentences. When a word is given in one sentence, certain The abnormal word concurrence phenomena in spam
words could be observed in its following sentence with some reviews could be observed normally. For example, the fixed
probability [3]. However, such transition patterns in spam pattern of two associative words in a spam review can hardly
reviews can be impaired due to their deceptive nature. The be found in truthful reviews. Therefore, the value of coherent
consecutive sentences sand s' are denoted as a pair (s, s'). The metric for spam reviews is often lower than that of truthful
symbol S, represents the sentence set formed by s'. Let f(w, SJ reviews. So this coherent metric is also very helpful to detect
denote the frequency of W in Sj. Given a set of words W, we spam reviews and it can be jointly used with the former one.
can define one step transition probability p(Wj I wJ as the 1 "n-l
Can ( r) = -- L- 1-1 Can (S"S'+I)
-
(7)
probability of observing Wj in one sentence when the word Wj n -I
appears in the previous sentence.
f(wj,SJ (2)
p (Wj I wJ =

Lw W f(w ,SJ
k
ALGORITHM: Coherence Metrics Computation Process
kE
Input: The reviews set V, store S, and the times to of
For any two sentences s I and s2, the transition probability iterations
to observe S2 after s], i.e. s] S2, can be estimated by
-- Output: reviews' coherence metric
choosing the edge with the highest one step transition counter 0 repeat
=

probability since s] and S2 have multiple words. for r E V do


P(SI � S2 ) = max p (Wj I wj)· f(Wj,S2) (3) compute H(r) using (5);
If;E5,'l,lf E5,'z
j compute Con(r) using (7);
Based on the above defmition, we provide a coherent end
metric of a review r in the following. This metric can be used update r's coherence scores using H(r) and Con(r);
to detect spam reviews. We assume the review r includes end
sentences {sJ, S2, S3, "',Sn }. So its coherent metric H(r) can be counter++;
defined as: until counter < to.
f loge pes1 --+ s 2 --+ ... --+ S n )) } (4)
H (r -)
-
exp, -
n-I Fig. 2 The proposed algorithm based on iterative computation
I i�1 1 S'+1 I
In the spam reviews, the word transition patterns can be
The above algorithm provides a framework to calculate
impaired; their coherent values will become larger than those
reviews' coherent score iteratively, which is self-explanatory.
on truthful reviews. In order to make the calculation process
In practice, this algorithm can converge fast. In every round, it
simple, we have one assumption that the transition probability
is only needed to access each review once and calculate its
between two consecutive sentences is independent of each
coherent value based on word transition probability and word
other, so we can get the following formula.
L :'�-lllog( P(Si ---+ Si+I))
concurrence probability. So the time complexity of this
H(r)= exp{-
(5) algorithm is O(Nvlo), where to is the number of iterations. It is
",,-I }
L... i�11 Si+1 I linear to the product of reviews' number N" and iteration times
to which is usually small.
B. Word concurrence probability V. OUR EVALUATION
We first define three probabilities between two
To evaluate our method, we built a labeled dataset using
consecutive sentences which have some similarity with ref. [3]: expert human judges. In our evaluation process, human
the probability of observing word w, (denoted as 0, ), the experts help to label dataset with reviews from ctrip.com
probability of observing word Wj (denoted as OJ ), and the (website of English version) . This website is famous for its
probability of observing word Wj in one sentence and word Wj large amount of traveling, hotel services. It is a typical
in the other (denoted as Ojj ) [3]. We defme the word ecommerce site. Our review data set was crawled from
concurrence metric for word Wj and Wj as log(Ojj / OJ OJ). This ctrip.com in June 2009 only for hotel services. We invite 3
metric does not only measure the concurrence probability in students who major in computer science to take part in our
one consecutive sentence pair, but also include the situation evaluation. The above algorithm is implemented to identifY
that the same word might naturally con-occur with many other 100 most suspicious reviews. Then the students begin to judge
words. The concurrence possibility of the w, and wi in the these suspicious reviews. They work independently on spam
same sentence is also considered. reviews identification. They are asked to give their
independent judgments based on several rules: (i) reviews are
Then the sentence concurrence measure for two full of empty adjectives, (ii) reviews with merely praises
consecutive sentences can be defined by using the highest without faults, or purely negative comments (iii) reviews are
word concurrence measure: left within a short period of time of each other, etc [II]. These

101
infonnation is used to identifY spam revIews from truthful REFERENCES
reviews by human evaluators. [I] Castillo c., Donato D., "A reference collection for web spam", SIGIR
Forum 2006,40(2),ppll-24.
Tn our evaluation, we label a review as a suspicious spam
[2] Wang, Guan, Sihong Xie, Bing Liu, and Philip S. Yu., "Identify Online
review if more than one student has an agreement. Since the
Store Review Spammers via Social Review Graph", ACM Transactions
students can identifY 46 spam reviews from the 100 suspicious on Intelligent Systems and Technology,2012.
reviews which are identified by our algorithm, the accuracy of [3] Sun, Huan, Alex Morales, and Xifeng Yan., "Synthetic review
this our methodology is 46%. Although this accuracy is not spamming and defense", Proceedings of the 19th ACM SIGKDD
very high compared to other similar works, it is still international conference on Knowledge discovery and data mining -
KDD 13,2013.
meaningful to us. Because we are trying to detect spam
[4] http://www.cs.uic.edu.
reviews on the semantic level which needs to handle more [5] Sharma, Kuldeep, and King-Ip Lin., "Review spam detector with rating
complex situations. Few existing research is found in this consistency check", Proceedings of the 51st ACM Southeast Conference
direction. In addition, the students evaluators agreed on each on - ACMSE 13,2013.
other's judgment. [6] Xu, Chang, "Detecting collusive spammers in online review
communities", Proceedings of the sixth workshop on Ph D students in
TABLE I. RESULT OF OUR EvALUAnON information and knowledge management - PIKM 13,2013.
[7] Lau, Raymond Y. K., S. Y. Liao, Ron Chi-Wai Kwok,Kaiquan Xu,
Student 1 Student 2 Student 3 Yunqing Xia, and Yuefeng Li. "Text mining and probabilistic language
Student t 46 29 36 modeling for online review spam detection", ACM Transactions on
Management Information Systems,2011.
Student 2 38 30 [8] 1. Liu, Y. Cao, C.-Y. Lin, Y. Huang, and M. Zhou,"Low-quality product
review detection in opinion summarization",In EMNLP-CoNLL,2007.
Student 3 40 [9] Xie, Sihong,Guan Wang,Shuyang Lin, and Philip S. Yu. "Review spam
The evaluatIOn result IS hsted m table I WhiCh shows the detection via temporal pattern discovery",Proceedings of the 18th ACM
SIGKDD international conference on Knowledge discovery and data
agreement level among evaluators. For example, student 1
mining - KDD 12,2012.
identified 46 suspicious reviews; out of which 29 were [10] G. Wang,S. Xie, B. Liu, and P. S. Yu,"Review graph based online store
recognized by student 2 and 36 were caught by student 3. The review spammer detection",IEEE 11th International Conference on Data
different results also shows that it is hard to get an agreement Mining,pp 1242-1247,2011.
on some revIews among all the elevators without enough [11] A. Mukhe�jee, B. Liu, and N. Glance, "Spotting fake reviewer groups in
consumer reviews", In Proceedings of the 21st International Conference
strong evidences.
on World Wide Web (WWW),2012.
[12] N. Jindal and B. Liu., "Opinion spam and analysis", In Proceedings of
VI. CONCLUSION the International Conference on Web Search and Web Data Mining
(WSDM),pp2I9-230,2008.
Until today, there are a few attempts have been made on [13] N. Jindal and B. Liu, "Analyzing and detecting review spam", In
review spam detection which is a challenging and under­ Proceedings of the 7th IEEE International Conference on Data Mining,
exploration area [2]. In this paper, we propose a general pp547-552,2007.
[14] NJindal and B.Liu, "Review spam detection", In Proceedings of the
framework to detect spam reviews based on coherent 16th International Conference on World Wide Web, pp1l89-1190, 2007.
examination. We first discuss some assumptions of spam [15] E. Lim, V. Nguyen, N. Jindal, B. Liu, and H. Lauw, "Detecting product
reviews, and then we define some review coherent metrics. review spammers using rating behaviors", In Proceedings of the 19th
The iterative computation framework is also provided. Our ACM International Conference on Information and Knowledge
Management (CIKM), pp 939-948,2010.
proposed model tries to analyze review coherence in the [16] c. Dellarocas, "Immunizing online reputation reporting systems against
granularity of sentence. We define some metrics to investigate unfair ratings and discriminatory behavior",In ACM EC,2000.
the coherence between sentiment words and other related [17] S.-M. Kim, P. Pantel, T. Chklovski, and M. Pennacchiotti,
words based on the flow smoothness between sentences: word "Automatically assessing review helpfulness",In EMNLP,2006.
[18] Kleinberg, J. ,"Authoritative sources in a hyperlinked environent". In J.
transition probability and word concurrence probability. ACM 46,5,pp604-632,1999.
Because we try to identifY spam reviews on the semantic level,
our proposed model can reveal more important clues of spam
reviews based on their word pattern. This work provides a
novel viewpoint for spam review detection and more potential
approaches could be explored in the future.

102

You might also like