Hindawi

Applied Bionics and Biomechanics

Volume 2021, Article ID 5522574, 11 pages
https://doi.org/10.1155/2021/5522574

Research Article
Development of Integrated Neural Network Model for
Identification of Fake Reviews in E-Commerce Using
Multidomain Datasets

Saleh Nagi Alsubari ,1 Sachin N. Deshmukh ,1 Mosleh Hmoud Al-Adhaileh ,2

Fawaz Waselalla Alsaade,3 and Theyazn H. H. Aldhyani 3
1
Department of Computer Science & Information Technology, Dr. Babasaheb Ambedkar Marathwada University,
Aurangabad, India
2
Deanship of E-Learning and Distance Education King Faisal University Saudi Arabia, Al-Ahsa, Saudi Arabia
3
Community College of Abqaiq, King Faisal University, P.O. Box 400, Al-Ahsa, Saudi Arabia

Correspondence should be addressed to Saleh Nagi Alsubari; salehalsubri2018@gmail.com

Received 28 February 2021; Revised 20 March 2021; Accepted 5 April 2021; Published 15 April 2021

Academic Editor: Fahd Abd Algalil

Copyright © 2021 Saleh Nagi Alsubari et al. This is an open access article distributed under the Creative Commons Attribution
License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is
properly cited.

Online product reviews play a major role in the success or failure of an E-commerce business. Before procuring products or services,
the shoppers usually go through the online reviews posted by previous customers to get recommendations of the details of products
and make purchasing decisions. Nevertheless, it is possible to enhance or hamper specific E-business products by posting fake
reviews, which can be written by persons called fraudsters. These reviews can cause financial loss to E-commerce businesses and
misguide consumers to take the wrong decision to search for alternative products. Thus, developing a fake review detection
system is ultimately required for E-commerce business. The proposed methodology has used four standard fake review datasets
of multidomains include hotels, restaurants, Yelp, and Amazon. Further, preprocessing methods such as stopword removal,
punctuation removal, and tokenization have performed as well as padding sequence method for making the input sequence has
fixed length during training, validation, and testing the model. As this methodology uses different sizes of datasets, various input
word-embedding matrices of n-gram features of the review’s text are developed and created with help of word-embedding layer
that is one component of the proposed model. Convolutional and max-pooling layers of the CNN technique are implemented
for dimensionality reduction and feature extraction, respectively. Based on gate mechanisms, the LSTM layer is combined with
the CNN technique for learning and handling the contextual information of n-gram features of the review’s text. Finally, a
sigmoid activation function as the last layer of the proposed model receives the input sequences from the previous layer and
performs binary classification task of review text into fake or truthful. In this paper, the proposed CNN-LSTM model was
evaluated in two types of experiments, in-domain and cross-domain experiments. For an in-domain experiment, the model is
applied on each dataset individually, while in the case of a cross-domain experiment, all datasets are gathered and put into a
single data frame and evaluated entirely. The testing results of the model in-domain experiment datasets were 77%, 85%, 86%,
and 87% in the terms of accuracy for restaurant, hotel, Yelp, and Amazon datasets, respectively. Concerning the cross-domain
experiment, the proposed model has attained 89% accuracy. Furthermore, comparative analysis of the results of in-domain
experiments with existing approaches has been done based on accuracy metric and, it is observed that the proposed model
outperformed the compared methods.

1. Introduction reviews posted on E-commerce sites represent the opinions

of customers, and now these opinions play a significant role
The development of Web 4.0 has increased the activity of in E-businesses because they could potentially influence
internet shopping through E-commerce platforms. Online customer-buying decisions. Business owners use online
2 Applied Bionics and Biomechanics

customer reviews to detect product issues and to discover for detecting and discriminating between fake and truthful
business intelligence knowledge about their opponents [1]. reviews in online E-commerce websites. In order to miti-
Fraudsters post fake comments termed misleading reviews gate the problems of online review mining systems, it is
to affect business by manipulating potential reputation of necessary for developing a model to detect and eliminate
product brands. Fake reviews are divided into 3 types: (1) online fake reviews due to their effect on customers and
untrusted (fake) reviews, (2) review on product name only, E-commerce companies.
and (3) nonreviews. The fake reviews are posted deliberately
to mislead and deceive buyers and consumers. These reviews 2. Literature Review
contain unjust positive reviews for particular desired prod-
ucts to promote them and provide unfavorable reviews to This section sheds light on methods and datasets used in pre-
worthy products for deprecating. Hyperactive fake reviews vious studies for fake/spam review detection. Online product
are linked to this type of review. Reviews on products brand reviews are defined as guidelines that are widely used by a
only are the second version of fake reviews that can be potential customer to make online purchasing that involves
created to manipulate the brands of products. Nonreviews choosing or not to purchase a particular product, identifying
are composed of two subsets, namely, (a) advertisement the problems of manufacturing companies’ products, and
and (b) unrelated reviews [2]. Larger amounts of positive gaining intelligent information of their competitors in mar-
reviews lead to making the shoppers and customers buy keting research. Recently media news from the New York
products and enhance companies’ financial benefits, whereas Times and the BBC have reported that counterfeit reviews
negative reviews can make customers to search for substitute are very widespread on E-commerce, for example, a photog-
products that way resulting in revenue loss. However, a sig- raphy company has recently been targeted by fake reviews of
nificant number of review comments are generated across thousands of fraudulent [11]. Over the last two decades,
social media applications, adding complications for extract- fake/spam review detection has become a popular topic of
ing views and difficulty in obtaining accurate findings. In study. Since fake reviews have such a significant effect on E-
addition, there is no monitoring on the reliability of digital commerce and customers, several researchers have conducted
content generated on the E-commerce websites, and this several types of research on spam/fake review analysis.
encourages the creation of several low-quality reviews possi-
ble. Various companies hire persons to write fake reviews for 2.1. Fake Review Detection Based on Machine Learning
rising the purchasing of their online products and services. Methods. Jindal et al. [2] have presented first research
Such persons are known as fake reviewers or spammers, towards spam review detection. The authors dealt with dupli-
and the activities they perform are called review faking [3]. cate or near-duplicate in Amazon product reviews as fake
Therefore, the existence of fake and spam reviews makes reviews that were comprised attributes regarding the review
the issue more considerable to be handled because they affect text and reviewer. It has been applied the logistic regression
the possible changing of buying decision to customers and technique for classifying reviews into truthful or fake with
shoppers. A huge amount of positive reviews enable a con- reaching 78% in the terms of accuracy.
sumer to purchase a product and improve the manufacture’s Ott et al. [10] have utilized the crowdsourcing website
financial profits, whereas negative reviews encourage (Amazon Mechanical Turk) to create a dataset, and the nat-
consumers to search for substitutes and therefore causing ural language processing tool was also used to obtain linguis-
financial losses [3, 4]. Consumer-generated reviews can get tic features from the review contents. They trained and
a huge influence on the reputation of products and brands, compared several types of supervised machine learning tech-
and hence, E-business companies would be motivated to pro- niques. However, the obtained results on real market datasets
duce positive fake reviews over their products and negative have not been very good. Lau et al. [11] have presented model
deceptive reviews over their competitors’ products [5–7]. for fake opinion identification using an LDA algorithm,
Electronic commerce sites have numerous ways of spamming namely, Topic Spam that can categorize the text of the review
with spam reviews, for instance, hiring expert persons who by calculating the likelihood of spam index to the little
are specialized in generating fraud reviews, utilizing crowd- dissimilarity between the distribution of the keywords of
sourcing websites to utilize review fraudsters, and using auto- the spam and the nonspam reviews.
mation tool bots for feedback [8, 9]. The capability of vendors Shojaee et al. [12] have proposed syntactic grammar and
to produce misleading opinions as a way of either promoting lexical-based attributes named stylometric attributes. These
their products or defame the reputation of their competitors attributes are utilized to distinguish fake reviews from online
is indeed worth remembering. Fake reviews have a tremen- hotel reviews. Using lexical features, the authors imple-
dous influence on consumer satisfaction. For example, when mented SMO (sequential minimal optimization) and Naive
a consumer is tricked or mislead via a fake review, a Bayes methods for classifying the reviews into fake or truthful
consumer will not utilize that E-commerce website again and the obtained results were 81% and 70% in the terms of
for purchasing. Ott et al. [10] reported that about 57% is F1-score, respectively. However, then, they have enhanced
the total average of testing accuracy of human judges for the performance of the model by merging lexical and syntac-
distinguishing fake reviews from truthful ones; therefore, tic features, and the SMO technique attained 84% F1-score.
further research is required in identifying misleading (fake) Xu and Zhao [13] suggested a parsing tree-based model for
reviews. The limitations of existing studies of fake/decepti- detecting and classifying fake reviews. They used textual fea-
ve/spam review detection are proposing automated methods tures of the review text that were taken out from the parsing
Applied Bionics and Biomechanics 3

tree by using syntactic analysis and implemented them to the results were 82% and 81% in terms of accuracy for DFFN and
model for identifying fake reviews. They just concentrated on CNN methods, respectively. Goswami et al. [20, 21] have
textual features and ignored behavioral features. Allahbakhsh proposed Artificial Neural Network model to investigate
et al. [14] have examined the involvement of reviewers who the influences of social relations of reviewers for deception
place prejudiced score reviews on online rating classification recognition at online customer reviews, and in their experi-
systems collected through some attributes that can assist to ment, Yelp’s review dataset was gathered and preprocessed.
point out a set of spammers. In their model, they utilized Then, they mined behavioral and social relation features of
the Amazon log (AMZLog) with its dataset for carrying out customers and applied the backpropagation neural network
the experiments. Duan and Yang [15] explored fake review algorithm for classifying reviews into genuine and fake with
identification based on reviews of the hotels. Through their a detection rate of 95% accuracy. Ren and Ji [22] have pro-
method, they measured and used three features of the review posed a hyper deep learning model that is consisted of a gated
text for detecting spam actions, general score, subscore, and recurrent neural network and convolutional neural network
review content. Feng et al. [16] have concentrated on dissem- (GRNN-CNN) for detecting deceptive opinion spam on in-
ination footprints of reviewers and giving an association domain and in-cross domain datasets. They used doctors,
between distribution abnormalities and spammer’s actions. restaurants, and hotels with a size of 432, 720, and 1280
Using the Markov model, they assessed the product review reviews, respectively. By combining all these datasets, they
dataset collected from the Trip Advisor website. Barbado applied their proposed method for classification of the
et al. [17] have proposed framework of significant features reviews into spam and nonspam reviews. The best classifica-
for deceptive review detection. Based on online Yelp product tion result obtained was 83% in terms of accuracy. Using the
reviews, they carried out experiments using different super- same datasets used in [22], Zeng et al. [23] have proposed a
vised machine learning techniques. In terms of features, recurrent neural network-bidirectional long-short technique
reviewer (personal, social, review activity, and trust) and for deceptive review detection. They divided the review text
review features (sentiment score) were used. Their experi- into three parts: a first sentence, middle context, and last
mental result showed that the AdaBoost algorithm provided sentence. The best-achieved results of their method were
the best performance with obtained 82% accuracy. Noekhah 85% in terms of accuracy.
et al. [18] have presented a novel approach-based graph for
detecting opinion spam in Amazon product reviews. First, 3. Methodology
they calculated an average value for review and reviewer fea-
tures individually. Then, they asked three experts for assign- Figure 1 shows the proposed methodology for fake review
ing weight for every feature. Finally, they are multiplying the identification system that is consisted of six modules, namely,
weight of the feature with its average value for calculating the datasets, preprocessing, CNN-LSTM method, data splitting,
spamicity for the review text and reviewer. Their approach evaluation metrics, and results. The details of the framework
achieved 93% accuracy. Alsubari et al. [3] have proposed are discussed below.
different models based on supervised machine learning algo-
3.1. Datasets. This module presents the datasets used in these
rithms such as Random Forest, AdaBoost, and Decision tree.
experiments that are performed for the identification of
They used the standard Yelp product review dataset. The
deceptive/fake reviews. We have employed four standard
information gain method was applied as feature selection.
fake review datasets: hotel, restaurant, Amazon, and Yelp.
From their experimental results, it is observed that the
AdaBoost algorithm has provided the best performance by 3.1.1. Amazon-Based Dataset. This dataset is standard fake
recording 97% accuracy. Amazon product reviews consists of 21,000 reviews (10500
truthful and 10500 fake), and each review has metafeature
2.2. Fake Review Detection Based on Deep Learning Methods. such as product Id, product name, reviewer name, verified
The use of deep learning neural network models for fake purchase (no or yes), and rating value as well as a class label,
review identification has three key points. The first point is while in the statistical analysis of the dataset, we found that
that deep learning models utilize real-valued hidden layers the average rating value of the reviews was 4.13, and 55.7%
for automated feature compositions that can catch compli- of the data was recognized as verified purchases. The reviews
cated global semantic data, which is difficult by utilizing of this dataset are equally distributed through 30 discrete
typical specific handcrafted features. This provides an effec- product classifications (e.g., wireless, PC, health, etc.). Each
tive way in solving the shortcomings of different traditional product has 700 reviews (350 fake and 350 truthful reviews).
models aforementioned above. The second point is that neu- Furthermore, the reference for labeling this dataset is the
ral networks consider clustered word embedding as inputs Amazon filtering algorithm that is employed by the Amazon
that can be learned from raw text, hence mitigating the short- website [20, 21, 24].
age of labeled data. The third point is that neural models can
learn consistent text structure instantaneously. Based on 3.1.2. Yelp-Based Dataset. This dataset is standard fake elec-
Amazon electronic product review dataset, Hajek et al. [19] tronic products reviews combined from four USA cities
have implemented two neural network methods that were (Los Angeles, Miami, NY, and San Francisco) by Barbado
Deep Feed-Forward Neural Network and convolution neural et al. [17]. A reference for labeling this dataset is the Yelp
network. Then, they extracted features from the review text filtering algorithm utilized by the http://Yelp.com/ website
set such as word emotions and n-grams. Their methodology [25]. The dataset includes 9461 reviews and reviewers with
4 Applied Bionics and Biomechanics

3.2.4. Removing Contractions. This process is aimed at

Datasets removing a word that has been written with the short form
and replaces it with full form. Example “when’ve” will
Data become “when have.”
preprocessing
3.2.5. Tokenization. This process can be defined as dividing
each textual review sentence into small pieces of words
Training Training
data
Data splitting
data or tokens.

3.2.6. Padding Sequences. The deep learning algorithms

require input sequences in text classification to have the same
CNN-LSTM model length; therefore, for this purpose, we have used the padding
sequence method and set the maximum length of the review
text to 500 words.
Evaluation metrics
3.3. Data Splitting. This subsection introduces the details of
dividing the multidomain datasets that are evaluated in our
Results analysis experiments. Each used dataset has divided into 70% as a
training set, 10% as a validation set, and 20% as testing set.
Figure 1: A Framework for the proposed methodology.
Then, we have adopted a hyperneural network model that
is consisting of a convolutional neural network integrated
with long short-term memory (CNN-LSTM) for detecting
features such as rating value, reviewer name, verified and classifying the review text into a fake or truthful review.
purchase (yes or no), reviewer Id, product Id, review title, Table 1 summarizes the splitting of each dataset individually.
and review text as well as the class label.
3.4. CNN-LSTM-Based Fake Review Identification. The sug-
3.1.3. Restaurant-Based Dataset. This dataset is fake restau- gested method applies and assists the performance of inte-
rant reviews developed by Abri et al. [26, 27]. It includes grated convolution neural network with long short-term
110 reviews belong to three local Indian restaurants and has memory (CNN-LSTM) to detect and identify the review
organized a way to have an equivalent distribution of fake text comprising content with fake linguistic clues. For this
and real reviews (55 fake and 55 truthful). The metafeatures purpose, we train the deep learning-based neural network
of the dataset are sentiment polarity that means positive or model for classifying the input review text of different
negative review, review text, reviewer Id, restaurant name, domain datasets. Figure 1 illustrates the structure of the
and a class label. CNN-LSTM model.
Figure 2 presents the structure of the proposed model
3.1.4. Hotel-Based Dataset. This is a publicly available stan- used in this research work for identifying the fake reviews
dard dataset developed by Ott et al. [10, 28, 29]. It contains in different domain datasets. The components of the CNN-
1600 hotel reviews (800 truthful and 800 fake) collected from LSTM model are discussed in detail as follows.
one of the popular hotel booking websites, that is, a Trip
advisor. The authors of this dataset have refined all 5- and (A) Word Embedding. The embedding layer is an initial
3-star rated reviews from 20 hotels in Chicago city. The layer of the proposed CNN-LSTM model that is used
features of the dataset consist of review text, reviewer name, for the transformation of each word presented in
hotel name, sentiment polarity, and class label. training data into an actual-valued vector represen-
tation that means a set of words as features of the
3.2. Data Preprocessing. The aim of preprocessing is applied dataset are constructed and transformed into
to make the data clean and easy to process. For this purpose, numerical form. This process is named word embed-
the following preprocessing techniques are implemented on ding. The word embedding is inputted as a matrix of
whole datasets. sequences to the following layer. An embedding
layer used in this model has made of three compo-
3.2.1. Lowercase. It is the process of converting whole words nents that are the vocabulary size (maximum fea-
of the review text into lowercase words. tures), embedding dimension, and input sequence
length. Maximum features which can keep the most
3.2.2. Stopword Removal. Stopwords are a collection of widely frequent and topwords represent the size of the
utilized words in a language, as these words do not carry any vocabulary. Embedding dimension demonstrates
significant information for the model; they have been the dimensions of each word that is transformed
removed from the contents of the review. Instances of stop- and by using the embedding layer into real-valued
words are “the,” “a,” “an,” “is,” “are,” etc. vector representations. Further, the input sequence
length defines the maximum length of the input
3.2.3. Punctuation Removal. This process is aimed at remov- sequence of the review text. The sentences of the
ing all punctuation marks in the review text. review text contain a sequence of words that can be
Applied Bionics and Biomechanics 5

Table 1: Splitting of datasets used in the experiments.

Dataset Total of Training set Validation set Testing set Total of deceptive Total of truthful
name samples (70%) (10%) (20%) reviews reviews
Amazon 21,000 15120 1680 4200 10500 10500
Yelp 9460 6622 946 1892 4730 4730
Restaurants 110 80 8 22 55 55
Hotels 1600 1152 128 320 800 800

Sigmoid layer Classification

LSTM layer LSTM unit Contextual information

Max pooling Reducing the dimensionality

layer

Feature sequence
Information extraction

Convolutional layer
Convolutional neural network

Embedding layer Word embeddings

x1 x2 x3 …… xt

Figure 2: The structure of the CNN-LSTM model.

annotated as X 1 , X 2 , X 3 ,….X t as shown in Figure 2 These matrices can be expressed in equations (1),
cited above section, and each word is assigned a spe- (2), and (3) that are given below.
cific index integer number. The embedding layer
converts the indices of each word into D dimensional P = Rlxw , ð1Þ
word vector representation. In our proposed model,
we have used dissimilar domain datasets and for F = Rl×m , ð2Þ
each dataset, we have created different embedding
matrix sizes [V × D] where V represents the vocabu- O = Rl×d , ð3Þ
lary size and D is the dimension vector representa-
tions of each word in V. For input sequence length,
where P, F, and O indicate the input, filter, and output
we assigned a fixed sequence length for all datasets
matrices, respectively, R is representing entirely real num-
with 500 words. The embedding matrix can be
bers, l is the sequence length, and w denotes the width of
symbolized as E ∈ RV×D .
the input matrix that is presented as R30000×100 for Amazon
(B) Convolution Layer. In the CNN-LSTM model, the and Yelp datasets and R10000×100 for restaurant and hotel
convolution layer is a second layer and performing datasets. M is the width of the filter matrix, and d is the width
a mathematical operation that is applied on two of the output matrix. A convolutional layer is utilized to mine
objective functions, which produces a third function. the sequence knowledge and decrease the dimensions of the
The convolutional operation is calculated on the input sequences [30–32]. It has parameters such as filters
dimension vectors of various matrices such as input with window size. Here, we set the window size to 2 × 2 and
matrix (I), filter matrix (F), and output matrix (O). the number of filters to 100, which passes over the input
6 Applied Bionics and Biomechanics

matrix to extract the features. The formula for convolutional ht

Output gate
operation is given as follows.
O
n m Forget gate
t i, j = 〠 〠 f l,w ⨂Pi + l − 1, j + w − 1, ð4Þ
f c Memory cell
l=1 w=1

where ⨂ represents element-wise cross multiplication,

t i, j ∈ Rl×d is indicating tth element of output matrix, f l,w ∈ i
Input modulator
Rn×m denotes the elements of the weight matrix, Pi + l − 1, j Input gate
+ w − 1 ∈ Rlxw is represented pth elements of the input matrix.

(C) LSTM Layer. Long short-term memory network

(LSTM) is one type of recurrent neural network xt ht-1
(RNN) that has the capability for learning long-term
dependence and contextual information of the input Figure 3: The structure of LSTM Unit.
sequence. We have utilized LSTM as one layer of the
CNN-LSTM model and assigned it with different (E) A sigmoid activation function is the last layer of the
values which include 50 cells in the case in-domain model that is applied to detect and classify output
experiment and 100 cells in the cross-domain experi- classes (fake or truthful review). The equation for a
ment. LSTM cell executes precalculations for input sigmoid function is given as follows
sequence before giving an output to the last layer of
the network. Figure 3 depicts the structure for the 1
LSTM cell. σ= : ð7Þ
1 − e2x
In every cell, four discrete computations are conducted
based on four gates: input (it ), forget (f t ), candidate (ct ), 3.5. Evaluation Matrices. This subsection presents an evalua-
and output (ot ). The equations for these gates are introduced tion of how proficiently the proposed model can classify and
as follows [31]. distinguish between fake and truthful review text in terms of
  false-positive and false-negative rates. For measurement of
f t = sig W f xt + U f ht − 1 + b f , the performance of the classification capability of the CNN-
LSTM model, we employed dissimilar performance metrics


it = sig Wixt + Uiht − 1 + bi , as follows.


Ot = sig Woxt + Uoht − 1 + bo , TP + TN

Accuracy = × 100,
c ~ t = tanh wcxt + Ucht − 1 + bc , FP + FN + TP + TN
ð5Þ
TP
Ct = ð f to ct − 1 + ito c ~ t , Precision = × 100,
TP + FP
ht = Oto ∗ tanh ðC t Þ, TP
Sensitivity = × 100, ð8Þ
1 − e2x TP + FN
tanh ðxÞ = , TN
1 − e2x Specificity = × 100,
TN + FP
where sig and tanh are sigmoid and tangent activation precision × sensitivity
functions, respectively. X is the input data. W and b repre- F1 − score = 2 ∗ × 100:
precision + sensitivity
sent the weight and bias factors, respectively. Ct is cell state,
c ~ t is candidate gate, and ht refers to the output of the 3.6. Experimental Results and Analysis. We assessed the
LSTM cell. proposed CNN-LSTM model in two different types of exper-
iments (in-domain and cross-domain) based on four stan-
(D) Dense Layer. The dense layer (fully connected layer)
dard fake review datasets (Amazon, Yelp, restaurant, and
is one of the hidden layers in the CNN-LSTM model.
hotel). We also analyze the performance of the model on each
It consists of N artificial connected neurons and is
dataset and across datasets.
used to connect all neurons of the network [33].
The function applied to this layer is Rectified 3.6.1. In-Domain Experiment. In this section, we introduce
Linear Unit (RLU). This function is used to speed the results of the experiments executed to assess the efficiency
up the training process of the model. It has the of the proposed integrated CNN-LSTM model on the four
following equation. publicly available fake review datasets individually. We have
split each dataset as 70% training, 10 as validation, and 20%
f ðxÞ = max ð0, xÞ: ð6Þ as testing. Based on the learning of n-grams of the review
Applied Bionics and Biomechanics 7

9 2000

True neg 8 True neg False pos 1750

Fake
False pos
Fake

8 2098 241
3 7 1500
36.36% 49.95% 5.74%
13.64%
6 1250

5 1000

Truthful
Truthful

True pos False neg True pos

False neg 4 750
9 286 1575
2
40.91% 6.81% 37.50% 500
9.09% 3

2 250
Fake Truthful Fake Truthful

Figure 4: Confusion matrix for restaurant dataset. Figure 7: Confusion matrix for Amazon dataset.

Table 2: Classification results for in-domain experiment.

140
In-domain Sensitivity Specificity Precision F1-score Accuracy
True neg
False pos 120 datasets (%) (%) (%) (%) (%)
Fake

147
13 Restaurant 82 72 75 78 77
45.94% 100
4.06%
Hotel 77.5 92 90 83 85
80 Yelp 87 86 88 87 86
Amazon 85 90 87 86 87
Truthful

True pos 60
False neg
36 124
40
11.25% 38.75% 100
20 90
80
Fake Truthful 70
Accuracy

60
Figure 5: Confusion matrix for hotel dataset. 50
40
30
20
800 10
True neg False pos 0
Restaurants Hotels Yelp Amazon
Fake

744 123 700

39.32% 6.50% In-domain datasets
600
Sensitivity F1-score
500
Specificity Accuracy
True pos 400 Precision
Truthful

False neg
133 892 300
47.15% Figure 8: Visualization of the classification results for in-domain
7.03% experiment.
200

Fake Truthful

Figure 6: Confusion matrix for Yelp dataset.

True neg False pos 2500
Fake

2987 332
46.43% 5.16% 2000
text, we create a specific word-embedding matrix for every
dataset using a hidden neural network-embedding layer,
1500
which is one component of the proposed CNN-LSTM model.
True pos
Truthful

In this experiment, we create different embedding matrices of False neg

2763 1000
size V × D, where V is the vocabulary size (number of the 352
42.94%
topwords selected as features from the dataset) and D refers 5.47%
500
to an embedding dimension. For example, the restaurant
and hotel datasets have an input embedding matrix of size Fake Truthful
10000 × 100, the Yelp dataset has 20000 × 100, and the
Figure 9: Confusion matrix for cross-domain datasets.
Amazon dataset has 30000 × 100. Further, convolutional
and max-pooling layers of CNN technique are applied to
extract and select the features of input sequences. The
8 Applied Bionics and Biomechanics

Table 3: Classification results for cross-domain experiment.

In-cross domain datasets Sensitivity (%) Specificity (%) Precision (%) F1-score (%) Accuracy (%)
Restaurant+hotel+Yelp+Amazon 89 90 90 89 89

90.2
90
89.8
Axis title 89.6
89.4
89.2
89
88.8
88.6
88.4
Cross-domain datasets

Sensitivity F1-score
Specificity Accuracy
Precision

Figure 10: Visualization of the classification results for cross-domain experiment.

Model accuracy Model loss

0.950 0.45
0.925
0.40
0.900
0.35
0.875
Accuracy

Loss

0.30
0.850
0.25
0.825

0.800 0.20

0.775 0.15

0.750 0.10
1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0
Epoch Epoch
Train
Val

Figure 11: The performance and loss of the CNN-LSTM model on cross-domain datasets.

LSTM layer with sigmoid function is used for learning and fake reviews. Table 2 and Figure 8 summarize and visualize
classifying an input sequences into fake or truthful reviews. the results for the in-domain experiments.
Figures 4–7 show the confusion matrices for restaurant,
hotels, Yelp, and Amazon datasets. 3.6.2. Cross-Domain Experiment. In this experiment, we have
In confusion matrices depicted in above Figures 4–7, true gathered all domain datasets into a single data frame for
negative (TN) represents the total numbers of samples that discovering features that are more robust. The size of this
the model successfully predicted as fake reviews. False nega- dataset is 32170 review text distributed as 21,000 different
tive denotes the total number of samples that the model Amazon product reviews, 9460 Yelp electronic product
incorrectly predicted as truthful reviews. True positive reviews, 110 restaurant reviews, and 1600 hotel reviews. We
denotes the total number of samples that the model success- have split the datasets into 70% as a training set, 10% as a
fully predicted as truthful reviews. FP represents the total validation set, and 20% as a testing set. Based on word
number of samples that the model incorrectly predicted as embedding of n-gram features of the review text, we have
Applied Bionics and Biomechanics 9

Table 4: Comparing the results of an in-domain datasets with existing work.

Paper id Domain dataset Features used Method Accuracy

Faranak Abri et al. [27] Restaurant Linguistic features from review content MLP 73%
Ren Y et al. [22] Hotel Review content and pretrained word embedding (bag of word) CNN 84%
Barushka et al. [33] Hotel Review content with pretrained word embedding (skip-gram) DFNN 83%
Garcia L. [24] Amazon Review content with TF-IDF SVM 63%
DFFNN 82%
Hajek et al. [19] Amazon Review content with pretrained word embedding (skip-gram)
CNN 81%
Barbado et al. [17] Yelp Review content with TF-IDF AdaBoost 82%
Restaurant 77%
Hotel n-grams of the review content with word-embedding matrix 85%
This study CNN-LSTM
Yelp using embedding layer 86%
Amazon 87%

created an input embedding matrix that has the size of V × D in-domain and cross-domain have been carried out on four
(V is vocabulary size of the dataset, and D is embedding standard fake review datasets (hotel, restaurant, Yelp, and
dimensions of each word in V) which is equal to 50000 × Amazon). Preprocessing methods such as lowercase, remov-
100. Further, the convolutional and max-pooling layers of ing of stopword and punctuation, and tokenization have
CNN are utilized for sliding over an input matrix and extract been conducted for the dataset cleaning as well as padding
the feature maps from input sequences. Then, LSTM layer sequence method was used to make a fixed length for all
receives the output from the max-pooling layer and performs input sequences. Further, an embedding layer as one compo-
the processing task for handling of contextual information of nent of the proposed model was applied to create different
the sequences based on gate mechanism. Finally, last layer is types of word-embedding matrices of size V ∗ D (V is the
the sigmoid function that is applied for classification of the vocabulary size of the dataset, and D is an embedding dimen-
input sequence into truthful or fake. The experimental results sion of each word in V) for in-domain and cross-domain
show that CNN-LSTM model provides better performance in experiments. Convolutional and max-pooling layers of the
cross-domain than an in-domain datasets. Figure 9 below CNN technique perform the feature extraction and selection.
Further, the LSTM technique is combined with the CNN for
presents the confusion matrix for cross-domain datasets.
contextual information processing of input sequences that
From the experimental results carried out in this research
are based on gate mechanisms and forward the output to
work, we conclude that a large number of n-gram features
the last layer. A sigmoid function as last layer of the proposed
lead to better accuracy with deep learning neural network model is used to classify the review text sequences into fake
techniques. Table 3 and Figure 10 show the classification or truthful. For in-domain experiments, the proposed model
and visualization of results in cross-domain experiment. is applied to each dataset individually for fake review detec-
In the above Figure 11 and on the left plot, the X-axis rep- tion. Further, a cross-domain experiment was performing
resents the training and validation accuracy and Y is the on mixed data of restaurants, hotels, Yelp, and Amazon
number of epochs, which indicate the number of iterations reviews. From experimental results, we conclude that a large
that the CNN-LSTM model has trained and tested on the number of features lead to better accuracy while using deep
dataset. The right plot shows the model loss. learning neural network (DLNN) algorithms. Outstandingly,
the proposed model surpassed existing baseline and state-of-
4. Comparative Analysis the-art fake review identification techniques in terms of accu-
racy and F1-score measures for in-domain experiment. The
In this section, we compare the results of in-domain experi- experimental results also revealed that the proposed model
ments performed by the proposed model (CNN-LSTM) with provides better performance in a cross-domain experiment
the existing works based on accuracy metric. Table 4 reports than an in-domain experiment because the first one is imple-
the comparative analysis using the accuracy metric. mented to a large-size dataset with more features. According
According to the literature review of fake review detec- to the literature review of fake review detection methods,
tion, there is no research work has used the same datasets there is no research work has used the same datasets in a
in a cross-domain experiment. Thus, we are unable to make cross-domain experiment. Thus, we are unable to make com-
comparative analyses for cross-domain datasets. parative analyses with cross-domain datasets.

5. Conclusion Data Availability

This paper presents a hyperneural network model compris- The data are available in the following links: https://www
ing of convolutional neural network along with long short- .kaggle.com/rtatman/deceptive-opinion-spam-corpus; https
term memory (CNN-LSTM) techniques for detecting and ://github.com/asiamina/FakeReviews-RestaurantDataset; htt
classifying the review text into fake or truthful. In the ps://github.com/aayush210789/Deception-Detection-on-A
proposed methodology, two different experiments that are mazon-reviews-dataset.
10 Applied Bionics and Biomechanics

Conflicts of Interest ference on Privacy, Security and Trust (PST 2012), Paris,
France, 2012.
The authors declare that they have no conflicts of interest. [16] S. Feng, “Distributional footprints of deceptive product
reviews,” in Sixth International AAAI Conference on Weblogs
and Social Media, Dublin, Ireland, 2012.
References [17] R. Barbado, O. Araque, and C. A. Iglesias, “A framework for
fake review detection in online consumer electronics retailers,”
[1] D. U. Vidanagama, T. P. Silva, and A. S. Karunananda, Information Processing & Management, vol. 56, no. 4,
“Deceptive consumer review detection: a survey,” Artificial pp. 1234–1244, 2019.
Intelligence Review, vol. 53, no. 2, pp. 1323–1352, 2020.
[18] S. Noekhah, E. Fouladfar, N. Salim, S. H. Ghorashi, and A. A.
[2] N. Jindal and B. Liu, “Opinion spam and analysis,” in Pro- Hozhabri, “A novel approach for opinion spam detection in e-
ceedings of the 2008 international conference on web search commerce,” in Proceedings of the 8th IEEE international con-
and data mining, pp. 219–230, Palo Alto, California, USA, ference on E-commerce with focus on E-trust, Mashhad, Iran,
2008. 2014.
[3] S. N. Alsubari, M. B. Shelke, and S. N. Deshmukh, “Fake [19] P. Hajek, A. Barushka, and M. Munk, “Fake consumer review
reviews identification based on deep computational linguistic,” detection using deep neural networks integrating word embed-
International Journal of Advanced Science and Technology, dings and emotion mining,” Neural Computing and Applica-
vol. 29, pp. 3846–3856, 2020. tions, vol. 32, no. 23, pp. 17259–17274, 2020.
[4] S. Rayana and L. Akoglu, “Collective opinion spam detection: [20] K. Goswami, Y. Park, and C. Song, “Impact of reviewer social
bridging review networks and metadata,” in Proceedings of interaction on online consumer review fraud detection,” Jour-
the 21th acm sigkdd international conference on knowledge dis- nal of Big Data, vol. 4, no. 1, pp. 1–9, 2017.
covery and data mining, pp. 985–994, Sydney, NSW, Australia, [21] M. Young, The Technical Writer’s Handbook, University
2015. Science, Mill Valley, CA, 1989.
[5] C. Miller, Company settles case of reviews it faked, New York [22] Y. Ren and D. Ji, “Neural networks for deceptive opinion spam
Times, 2009. detection: an empirical study,” Information Sciences, vol. 385,
[6] Y. Ren, D. Ji, and H. Zhang, “Positive unlabeled learning for pp. 213–224, 2017.
deceptive reviews detection,” in Proceedings of the 2014 confer- [23] Z. Y. Zeng, J. J. Lin, M. S. Chen, M. H. Chen, Y. Q. Lan,
ence on empirical methods in natural language processing and J. L. Liu, “A review structure based ensemble model
(EMNLP), pp. 488–498, Doha, Qatar, 2014. for deceptive review spam,” Information, vol. 10, no. 7,
[7] D. Streitfeld, For $2 a star, an online retailer gets 5-star product p. 243, 2019.
reviews, vol. 26, New York Times, 2012. [24] L. Garcia, Deception on Amazon on an NL exploration, 2018,
[8] A. Heydari, M. Ali Tavakoli, N. Salim, and Z. Heydari, “Detec- https://medium.com//@lievgarcial/deception-on-amazon-
tion of review spam: a survey,” Expert Systems with Applica- cle30d977cfd.
tions, vol. 42, no. 7, pp. 3634–3642, 2015. [25] S. Kim, H. Chang, S. Lee, M. Yu, and J. Kang, “Deep semantic
[9] M. Arjun, V. Vivek, L. Bing, and G. Natalie, “What yelp fake frame-based deceptive opinion spam analysis,” in Proceedings
review filter might be doing,” in Proceedings of The Interna- of the 24th ACM International on Conference on Information
tional AAAI Conference on Weblogs and Social Media and Knowledge Management, pp. 1131–1140, Melbourne,
(ICWSM-2013), Massachusetts USA, 2013. Australia, 2015.
[10] M. Ott, Y. Choi, C. Cardie, and J. T. Hancock, “Finding [26] L. Gutierrez-Espinoza, F. Abri, A. S. Namin, K. S. Jones, and
deceptive opinion spam by any stretch of the imagination,” D. R. Sears, “Ensemble learning for detecting fake reviews,”
in Proceedings of the 49th Annual Meeting of the Association in 2020 IEEE 44th Annual Computers, Software, and Applica-
for Computational Linguistics: Human Language tions Conference (COMPSAC), pp. 1320–1325, Madrid, Spain,
Technologies-Volume 1, Oregon, USA, 2011. 2020.
[11] R. Y. Lau, S. Y. Liao, R. C. Kwok, K. Xu, Y. Xia, and Y. Li, “Text [27] F. Abri, L. F. Gutierrez, A. S. Namin, K. S. Jones, and D. R.
mining and probabilistic language modeling for online review Sears, “Fake reviews detection through analysis of linguistic
spam detection,” ACM Transactions on Management Informa- features,” 2020, https://arxiv.org/abs/2010.04260.
tion Systems (TMIS), vol. 2, no. 4, pp. 1–30, 2011. [28] M. Ott, C. Cardie, and J. Hancock, “Estimating the prevalence
[12] S. Shojaee, M. A. A. Murad, A. B. Azman, N. M. Sharef, and of deception in online review communities,” in Proceedings of
S. Nadali, “Detecting deceptive reviews using lexical and syn- the 21st international conference on World Wide Web,
tactic features,” in 2013 13th international conference on intel- pp. 201–210, Lyon, France, 2012.
ligent systems design and applications (ISDA), pp. 53–58, [29] M. Ott, C. Cardie, and J. T. Hancock, “Negative deceptive
Salangor, Malaysia, 2013. opinion spam,” in Proceedings of the 2013 conference of the
[13] Q. Xu and H. Zhao, “Using deep linguistic features for finding north american chapter of the association for computational
deceptive opinion spam,” Proceedings of COLING 2012: Post- linguistics: human language technologies, pp. 497–501, Atlanta,
ers, pp. 1341–1350, 2012. Georgia, 2013.
[14] M. Allahbakhsh, A. Ignjatovic, B. Benatallah, S. M. R. Beheshti, [30] S. Ahmad, M. Z. Asghar, F. M. Alotaibi, and I. Awan, “Detec-
N. Foo, and E. Bertino, “Detecting, representing and querying tion and classification of social media-based extremist affilia-
collusion in online rating systems,” 2012, https://arxiv.org/abs/ tions using sentiment analysis techniques,” Human-centric
1211.0963. Computing and Information Sciences, vol. 9, no. 1, p. 24, 2019.
[15] H. Duan and P. Yang, “Building robust reputation systems for [31] Understanding LSTM cells using C#https://msdn.microsoft
travel-related services,” in Proceedings of the 10th Annual Con- .com/en-us/magazine/mt846470.aspx.
Applied Bionics and Biomechanics 11

[32] H. Alkahtani, T. H. Aldhyani, and M. Al-Yaari, “Adaptive

anomaly detection framework model objects in cyberspace,”
Applied Bionics and Biomechanics, vol. 2020, 14 pages, 2020.
[33] A. Barushka and P. Hajek, “Review spam detection using word
embeddings and deep neural networks,” in Artificial intelli-
gence applications and innovations. AIAI 2019, vol 559. IFIP
advances in information and communication technology, J.
Mac Intyre, I. Maglo-giannis, L. Iliadis, and E. Pimenidis,
Eds., pp. 340–350, Springer, Cham, 2019.