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 1 1 2 Saleh Nagi Alsubari , Sachin N. Deshmukh , Mosleh Hmoud Al-Adhaileh , 3 3 Fawaz Waselalla Alsaade, and Theyazn H. H. Aldhyani Department of Computer Science & Information Technology, Dr. Babasaheb Ambedkar Marathwada University, Aurangabad, India Deanship of E-Learning and Distance Education King Faisal University Saudi Arabia, Al-Ahsa, Saudi Arabia Community College of Abqaiq, King Faisal University, P.O. Box 400, Al-Ahsa, Saudi Arabia Correspondence should be addressed to Saleh Nagi Alsubari; firstname.lastname@example.org 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 for detecting and discriminating between fake and truthful customer reviews to detect product issues and to discover business intelligence knowledge about their opponents . 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 necessary for developing a model to detect and eliminate to aﬀect business by manipulating potential reputation of 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 . 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 . 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.  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 . 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.  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.  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.  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.  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  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 results were 82% and 81% in terms of accuracy for DFFN and tree by using syntactic analysis and implemented them to the 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 the inﬂuences of social relations of reviewers for deception et al.  have examined the involvement of reviewers who 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 algorithm for classifying reviews into genuine and fake with the experiments. Duan and Yang  explored fake review identiﬁcation based on reviews of the hotels. Through their a detection rate of 95% accuracy. Ren and Ji  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.  have concentrated on dissem- (GRNN-CNN) for detecting deceptive opinion spam on in- domain and in-cross domain datasets. They used doctors, ination footprints of reviewers and giving an association 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.  have proposed framework of signiﬁcant features reviews into spam and nonspam reviews. The best classiﬁca- tion result obtained was 83% in terms of accuracy. Using the for deceptive review detection. Based on online Yelp product reviews, they carried out experiments using diﬀerent super- same datasets used in , Zeng et al.  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 sentence. The best-achieved results of their method were mental result showed that the AdaBoost algorithm provided the best performance with obtained 82% accuracy. Noekhah 85% in terms of accuracy. et al.  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- Figure 1 shows the proposed methodology for fake review tures individually. Then, they asked three experts for assign- identiﬁcation system that is consisted of six modules, namely, ing weight for every feature. Finally, they are multiplying the datasets, preprocessing, CNN-LSTM method, data splitting, weight of the feature with its average value for calculating the evaluation metrics, and results. The details of the framework spamicity for the review text and reviewer. Their approach are discussed below. achieved 93% accuracy. Alsubari et al.  have proposed 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.  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. . 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 . The dataset includes 9461 reviews and reviewers with 4 Applied Bionics and Biomechanics 3.2.4. Removing Contractions. This process is aimed at removing a word that has been written with the short form Datasets and replaces it with full form. Example “when’ve” will become “when have.” Data preprocessing 3.2.5. Tokenization. This process can be deﬁned as dividing each textual review sentence into small pieces of words Training Training Data splitting or tokens. data data 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 experiments. Each used dataset has divided into 70% as a Results analysis training set, 10% as a validation set, and 20% as testing set. Then, we have adopted a hyperneural network model that Figure 1: A Framework for the proposed methodology. 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 (A) Word Embedding. The embedding layer is an initial advisor. The authors of this dataset have reﬁned all 5- and 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 dataset are constructed and transformed into 3.2. Data Preprocessing. The aim of preprocessing is applied 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- nents that are the vocabulary size (maximum fea- 3.2.1. Lowercase. It is the process of converting whole words 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 the dimensions of each word that is transformed signiﬁcant information for the model; they have been 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 review text contain a sequence of words that can be ing all punctuation marks in the review text. 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 Word embeddings Embedding layer …… x x x x 1 2 3 t Figure 2: The structure of the CNN-LSTM model. annotated as X , X , X ….X as shown in Figure 2 These matrices can be expressed in equations (1), 1 2 3, t cited above section, and each word is assigned a spe- (2), and (3) that are given below. ciﬁc index integer number. The embedding layer lxw converts the indices of each word into D dimensional ð1Þ P = R , word vector representation. In our proposed model, l×m we have used dissimilar domain datasets and for ð2Þ F = R , each dataset, we have created diﬀerent embedding l×d matrix sizes [V × D] where V represents the vocabu- ð3Þ O = R , 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 V×D symbolized as E ∈ R . 30000×100 the input matrix that is presented as R for Amazon 10000×100 (B) Convolution Layer. In the CNN-LSTM model, the and Yelp datasets and R 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 the sequence knowledge and decrease the dimensions of the objective functions, which produces a third function. 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 Output gate operation is given as follows. n m Forget gate t = 〠 〠 f ⨂P + l − 1, j + w − 1, ð4Þ i,j i l,w Memory cell w=1 l=1 where ⨂ represents element-wise cross multiplication, l×d i t ∈ R is indicating tth element of output matrix, f ∈ i,j l,w Input modulator n×m R denotes the elements of the weight matrix, P + l − 1, j Input gate lxw + w − 1 ∈ R is represented pth elements of the input matrix. (C) LSTM Layer. Long short-term memory network (LSTM) is one type of recurrent neural network x h t t-1 (RNN) that has the capability for learning long-term Figure 3: The structure of LSTM Unit. dependence and contextual information of the input 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 σ = : ð7Þ LSTM cell. 2x 1 − e In every cell, four discrete computations are conducted based on four gates: input (i ), forget (f ), candidate (c ), 3.5. Evaluation Matrices. This subsection presents an evalua- t t t and output (o ). The equations for these gates are introduced tion of how proﬁciently the proposed model can classify and as follows . distinguish between fake and truthful review text in terms of false-positive and false-negative rates. For measurement of the performance of the classiﬁcation capability of the CNN- f = sig Wf + Uf − 1+ b , t xt h LSTM model, we employed dissimilar performance metrics as follows. i = sig Wi + Ui − 1+ b , t xt h i O = sig Wo + Uo − 1+ b , TP + TN t xt h o Accuracy = × 100, FP + FN + TP + TN c ~ t = tanh wc + Uc − 1+ bc , ð5Þ xt h TP Precision = × 100, C = f ct − 1+ i c ~ t, t to to TP + FP h = O ∗ tanh C , TP ðÞ t to t Sensitivity = × 100, ð8Þ 2x TP + FN 1 − e tanh ðÞ x = , TN 2x 1 − e Specificity = × 100, TN + FP where sig and tanh are sigmoid and tangent activation precision × sensitivity F1 − score = 2 ∗ × 100: functions, respectively. X is the input data. W and b repre- precision + sensitivity sent the weight and bias factors, respectively. C is cell state, c ~ t is candidate gate, and h 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 . 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% fxðÞ = maxðÞ 0, x : ð6Þ as testing. Based on the learning of n-grams of the review Applied Bionics and Biomechanics 7 True neg True neg 1750 False pos False pos 8 2098 36.36% 49.95% 5.74% 13.64% 6 1250 True pos False neg True pos False neg 286 1575 6.81% 37.50% 40.91% 9.09% Fake Truthful Fake Truthful Figure 7: Confusion matrix for Amazon dataset. Figure 4: Confusion matrix for restaurant dataset. Table 2: Classiﬁcation results for in-domain experiment. In-domain Sensitivity Speciﬁcity Precision F1-score Accuracy True neg False pos datasets (%) (%) (%) (%) (%) Restaurant 82 72 75 78 77 45.94% 4.06% Hotel 77.5 92 90 83 85 Yelp 87 86 88 87 86 Amazon 85 90 87 86 87 True pos False neg 38.75% 11.25% 20 90 Fake Truthful Figure 5: Confusion matrix for hotel dataset. True neg False pos Restaurants Hotels Yelp Amazon 39.32% In-domain datasets 6.50% Sensitivity F1-score Specificity Accuracy 400 Precision True pos False neg Figure 8: Visualization of the classiﬁcation results for in-domain 47.15% 7.03% experiment. Fake Truthful Figure 6: Confusion matrix for Yelp dataset. True neg False pos 2500 2987 332 46.43% 5.16% text, we create a speciﬁc word-embedding matrix for every dataset using a hidden neural network-embedding layer, which is one component of the proposed CNN-LSTM model. True pos 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 42.94% 5.47% topwords selected as features from the dataset) and D refers to an embedding dimension. For example, the restaurant Fake Truthful and hotel datasets have an input embedding matrix of size 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 Truthful Fake Truthful Fake Truthful Fake Accuracy Truthful Fake Truthful Fake 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 89.8 89.6 89.4 89.2 88.8 88.6 88.4 Cross-domain datasets Sensitivity F1-score Accuracy Specificity 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 0.30 0.850 0.25 0.825 0.20 0.800 0.15 0.775 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 reviews, 110 restaurant reviews, and 1600 hotel reviews. We incorrectly predicted as truthful reviews. True positive 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 Accuracy Axis title Loss 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.  Restaurant Linguistic features from review content MLP 73% Ren Y et al.  Hotel Review content and pretrained word embedding (bag of word) CNN 84% Barushka et al.  Hotel Review content with pretrained word embedding (skip-gram) DFNN 83% Garcia L.  Amazon Review content with TF-IDF SVM 63% DFFNN 82% Hajek et al.  Amazon Review content with pretrained word embedding (skip-gram) CNN 81% Barbado et al.  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 Amazon). Preprocessing methods such as lowercase, remov- dimensions of each word in V) which is equal to 50000 × ing of stopword and punctuation, and tokenization have 100. Further, the convolutional and max-pooling layers of been conducted for the dataset cleaning as well as padding CNN are utilized for sliding over an input matrix and extract sequence method was used to make a ﬁxed length for all the feature maps from input sequences. Then, LSTM layer input sequences. Further, an embedding layer as one compo- receives the output from the max-pooling layer and performs nent of the proposed model was applied to create diﬀerent the processing task for handling of contextual information of types of word-embedding matrices of size V ∗ D (V is the the sequences based on gate mechanism. Finally, last layer is vocabulary size of the dataset, and D is an embedding dimen- the sigmoid function that is applied for classiﬁcation of the sion of each word in V) for in-domain and cross-domain input sequence into truthful or fake. The experimental results experiments. Convolutional and max-pooling layers of the show that CNN-LSTM model provides better performance in CNN technique perform the feature extraction and selection. cross-domain than an in-domain datasets. Figure 9 below 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- the-art fake review identiﬁcation techniques in terms of accu- 4. Comparative Analysis racy and F1-score measures for in-domain experiment. The experimental results also revealed that the proposed model In this section, we compare the results of in-domain experi- provides better performance in a cross-domain experiment ments performed by the proposed model (CNN-LSTM) with than an in-domain experiment because the ﬁrst one is imple- the existing works based on accuracy metric. Table 4 reports mented to a large-size dataset with more features. According the comparative analysis using the accuracy metric. to the literature review of fake review detection methods, According to the literature review of fake review detec- there is no research work has used the same datasets in a tion, there is no research work has used the same datasets cross-domain experiment. Thus, we are unable to make com- in a cross-domain experiment. Thus, we are unable to make parative analyses with cross-domain datasets. comparative analyses for cross-domain datasets. Data Availability 5. Conclusion The data are available in the following links: https://www This paper presents a hyperneural network model compris- .kaggle.com/rtatman/deceptive-opinion-spam-corpus; https ing of convolutional neural network along with long short- ://github.com/asiamina/FakeReviews-RestaurantDataset; htt term memory (CNN-LSTM) techniques for detecting and ps://github.com/aayush210789/Deception-Detection-on-A classifying the review text into fake or truthful. In the 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.  S. Feng, “Distributional footprints of deceptive product reviews,” in Sixth International AAAI Conference on Weblogs and Social Media, Dublin, Ireland, 2012. References  R. Barbado, O. Araque, and C. A. Iglesias, “A framework for fake review detection in online consumer electronics retailers,”  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.  S. Noekhah, E. Fouladfar, N. Salim, S. H. Ghorashi, and A. A.  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,  S. N. Alsubari, M. B. Shelke, and S. N. Deshmukh, “Fake  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.  S. Rayana and L. Akoglu, “Collective opinion spam detection:  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,  M. Young, The Technical Writer’s Handbook, University Science, Mill Valley, CA, 1989.  C. Miller, Company settles case of reviews it faked, New York  Y. Ren and D. Ji, “Neural networks for deceptive opinion spam Times, 2009. detection: an empirical study,” Information Sciences, vol. 385,  Y. Ren, D. Ji, and H. Zhang, “Positive unlabeled learning for pp. 213–224, 2017. deceptive reviews detection,” in Proceedings of the 2014 confer-  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,  D. Streitfeld, For $2 a star, an online retailer gets 5-star product p. 243, 2019. reviews, vol. 26, New York Times, 2012.  L. Garcia, Deception on Amazon on an NL exploration, 2018,  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.  S. Kim, H. Chang, S. Lee, M. Yu, and J. Kang, “Deep semantic  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.  M. Ott, Y. Choi, C. Cardie, and J. T. Hancock, “Finding  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.  R. Y. Lau, S. Y. Liao, R. C. Kwok, K. Xu, Y. Xia, and Y. Li, “Text  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.  M. Ott, C. Cardie, and J. Hancock, “Estimating the prevalence  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,  M. Ott, C. Cardie, and J. T. Hancock, “Negative deceptive Salangor, Malaysia, 2013. opinion spam,” in Proceedings of the 2013 conference of the  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.  M. Allahbakhsh, A. Ignjatovic, B. Benatallah, S. M. R. Beheshti,  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.  H. Duan and P. Yang, “Building robust reputation systems for  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  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.  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.
Applied Bionics and Biomechanics – Hindawi Publishing Corporation
Published: Apr 15, 2021