2023 World Conference on Communication & Computing (WCONF)
Raipur, India. July 14-16, 2023
Fake Product Review Detection and Elimination
using Opinion Mining
A.Thilagavathy, P.R.Therasa, J.Jeno Jasmine, Sneha M, Shree Lakshmi R, Yuvanthika S
2023 World Conference on Communication & Computing (WCONF) | 979-8-3503-1120-4/23/$31.00 ©2023 IEEE | DOI: 10.1109/WCONF58270.2023.10234996
Department of Computer Science and Engineering, R.M.K. Engineering College, Tamil Nadu, India
atv.cse@rmkec.ac.in, prt.cse@rmkec.ac.in, jje.cse@rmkec.ac.in, sneh19407.cs@rmkec.ac.in, shre19405.cs@rmkec.ac.in,
yuva19441.cs@rmkec.ac.in
Abstract - Identification and elimination of fake reviews and
their removal from the dataset provided using the supervised
machine learning algorithm and natural language processing
techniques based on a vast variety of aspects. In this proposed
paper, we trained the counterfeit review dataset by the process
of using two independently developed machine learning
algorithm models for assessing the extent to which the
information being provided is real. The counterfeit product
evaluations can be found on numerous online retailers are
mostly influencing the customers to buy those products and
profit for those products is probably dependent on the reviews
of those products. Hence these counterfeit reviews must be
noticed so that large E-commerce companies like Meesho,
Amazon, Flipkart, Nykaa, etc. can address this issue so that
fraudsters and fraudulent critics are taken out, sustaining
users' credibility in shopping sites. This approach may be
utilized for websites and apps with relatively few consumers,
estimating the authenticity of reviews so that online businesses
can respond to them suitably. This model is developed using
Naïve Bayes, Support Vector Machine,and TF-IDF (term
frequency-inverse document frequency )Vectorizer. To detect
spam reviews on a website or application instantly, one can
make use of these models. However, effectively countering
spammers requires a sophisticated model that has to undergo
training on a large dataset of millions of reviews. In this work ”
Reviews of 20 Hotels in Chicago hotel dataset” a limited
dataset is utilized to train the models on a small scale, but it
can be expanded to achieve greater accuracy and authenticity
in the reviews.
Index Terms – Opinion Mining, Data Preprocessing,
Supervised Machine Learning Algorithm.
I. INTRODUCTION
The trend of people giving reviews for the product they
are buying online has become a day-to-day activity
nowadays. Based on the feedback consumers are buying
products through various e-commerce websites. But when
the reviews given by the critics are counterfeit there is no
way that the consumers would not know the authenticity of
the reviews provided by the critics to the customers. So
consumers are being manipulated to buy a product that is not
trustworthy product. The task is straightforward but time-
consuming because each review must be read and marked as
a counterfeit or ambiguous category to identify the issue’s
true cause. By teaching a machine learning model that deals
with the review section to flag a specific review as genuine
or spam, this issue can be solved. The intriguing part is that
979-8-3503-1120-4/23/$31.00 ©2023 IEEE
Authorized licensed use limited to: China University of Petroleum. Downloaded on April 10,2024 at 07:35:52 UTC from IEEE Xplore. Restrictions apply.
this method can be used to catch spammers who didn't use the
product. In order to falsely filter reviews of the product and give
it a high rating, spam reviews or the use of different customer ids
may be used. This can be filtered by looking at how often words
like "awesome," "so good," "fantastic," etc. are used. This
encourages us to create a system that uses a review's text and
rating information to identify fake customer reviews of a product.
The credibility grading and evaluation for a fraudulent
review will be calculated utilizing machine learning models. By
deriving topics and convictions from online reviews, a
computerized system could be used to monitor consumer
analyses. It might additionally block out fraudulent critiques.
Hence this issue of fake review identification and removal
takes a lot of data to train and be effective, along with additional
subject knowledge such as the sarcastic sentences users employ
to convey their displeasure of the product. In some cases, what is
being reviewed may be good, but the method of distribution or
shipping may not be, which impacts the review classification. As
contrasted with classification errors as an adverse rating as in the
evaluation of sentiment, currently an NLP method has been
utilized for detecting such reviews. The preliminary processing
of data is utilized to eradicate irrelevant or old evaluations of
products. Because the number of users on these
websites/applications is growing daily,Evaluation of sentiment is
used by businesses like Twitter, WhatsApp, and Facebook for
recognizing fake news and harmful or disparaging messages and
to ban perpetrators. The primary objective of this research project
is to build a platform for internet shopping where users may
develop confidence in a system where the goods they purchase
are real and consumer testimonials are precise and often
validated by the company themselves. In addition, businesses in
the e-commerce (Walmart, Amazon), logistics, travel, job search
(LinkedIn, Glassdoor, Indeed), and food (Shopsy, Swiggy,
Zomato) sectors use algorithms to combat Fraudulent who trick
customers into purchasing subpar goods and services by posting
false reviews. Users shouldn't worry about such fake users
considering that they will be apprised of scammers like "not
verified listings." Instruction Manual labeling of the reviews
consumes a lot of effort and is less efficient. As a result, the
evaluations are assigned labels using an algorithm for supervised
learning, and the designation then appears to be untenable. The
Naïve Bayes,SVM,and TF-IDF vectorizer methods have been
utilized to identify and remove fake reviews. The fake review
identification problem is addressed fairly and helps consumers to
view authenticated reviews.
1
 II. LITERATURE REVIEW
In [1], a preliminary study is conductedendeavoring to
fabricate an ingenious LMS ( Learning Management System
) for distance learning which employs techniques from NLP.
This study is based on a Systematic Literature Review (SLR)
focused on Recommender Systems (RS). In [2], two
Machine Learning (ML) models are applied to train a dataset
of fake reviews in order to predict their authenticity. In [3]
proposes an algorithm to track customer reviews and extract
topic and sentiment information from online reviews. The
algorithm is also capable of identifying and blocking fake
reviews.
The proposed system in [4], called ICF++, is designed to
measure a review's honesty, the reviewer's trustworthiness,
and the product's reliability. In [5] examines review-centric
features proposed for detecting fake reviews, with a
particular focus on approaches that employ supervised
machine learning techniques. The authors of reference [6]
have extended a recently proposed method for detecting
opinion spam that relies on n-gram techniques. Their
approach involves introducing feature selection and utilizing
different methods for representing opinions.
Reference [7] presents a new and reliable system for
identifying spam reviews. The proposed approach effectively
utilizes three distinct features: (i) the sentiment of the review
and its accompanying comments, (ii) content-based factors,
and (iii) rating deviation.The purpose of [8] is to serve as a
literature review for beginners and a survey for identifying
opportunities in the field. In [9] proposes a Fake Product
Review Monitoring and Removal System (FPRMS) that uses
an Intelligent Interface and Uniform Resource Locators
(URLs) to remove fake reviews and provide users with
genuine reviews and ratings.
In [10], analyzes Yelp's filtered reviews to understand its
filtering algorithm and concludes that it is reasonable and
linked to unusual spamming behaviors. [11] Suggests a
comprehensive technique called SPEAGLE that makes use
of relational data and information including text, timestamps,
and ratings to find suspicious reviews, individuals, and goods
that have been the victim of spam. The purpose of [12] is to
develop a machine-learning model that can tell the difference
between real and false reviews in the Yelp dataset. In [13]
uses Naïve Bayes and Logistic Regression to classify Twitter
reviews and assess the algorithms' performance based on
accuracy, precision, and throughput.
III. METHODOLOGY
Utilizing the Naive Bayes, TF-IDF Vectorizer, and
extracted dataset feature extraction techniques. False
reviewer identification has been created. package contains
the kaggle.com dataset, which includes the following
characteristics::
2
Authorized licensed use limited to: China University of Petroleum. Downloaded on April 10,2024 at 07:35:52 UTC from IEEE Xplore. Restrictions apply.
Title: the title of a review
Author: the creator of the review.
Text: the review’s text; it may be incomplete.
Label: a label indicating that the review may not be reliable
1: Untrustworthy
0: Dependable
Data pre-processing: The process of processing and
improving data by removing superfluous and unnecessary
information as well as noisy and inaccurate data from the view
dataset.Data preparation, a fundamental stage in machine
learning methods, is the first step in the suggested strategy. Pre-
processing data involves cleaning up the view dataset of
extraneous and pointless information as well as noisy and
inaccurate data. One of the crucial phases in machine learning
approaches is data preparation, which is the first stage in the
recommended strategy.
Step 1: Sentence tokenization
The NLTK software is used to tokenize the full review into
sentences after receiving it as input.Tokenization is a popular
natural language processing technique that serves as a critical
first step before implementing any additional preprocessing
techniques. This process involves breaking down the text into
individual words, which are referred to as tokens. For example,
tokenization would separate the sentence "Going to college for
attendance" into the following tokens: "Going", "to", "college",
"for", and "attendance".
Step 2: Removal of punctuation marks
The reviews' initial and final punctuation, as well as
additional whitespace, have been removed. Stop words are the
most typically used terms, although they are useless. Typical
instances of stop words are (an, a, the, this). Before moving on to
the procedure of identifying false reviews in this study, all data
are cleared of stop words.
Step 3: Word Tokenization
Each individual review is tokenized into terms and
maintained in a list to aid retrieval.
Step 4: Removal of stop words
The stem is free of affixes. For instance, "cooking" has the
word "cook" as its stem, andthestemmingalgorithm is aware that
the "ing" suffix might be dropped.
After converting our text strings into numerical
representations with TF-IDF Vectorizer, we initialize a Naive
Bayes Classifier to fit the model. In conclusion, the accuracy
score and confusion matrix show how well our model performs.
TF-IDF. Vectorizer is a well-known approach for turning text
into intelligible appropriate units. We assume that a term will
become more important inside the given text as it appears more
frequently. We normalizethe frequency of a term based on the
size of the document, and we call this term frequency.
 Calculated definition:
݀‫ܿ݋‬Ǥ ܿ‫ݐ݊ݑ݋‬
ܶ‫ܨ‬ሺ‫ݓ‬ሻ ൌ 
‫ݐ݊݁݉ݑܿ݋̴݄݀݁ݐ̴̴݊݅ݏ݀ݎ݋ݓ̴݈ܽݐ݋ݐ‬

While assessing term frequency, each syllable is assigned a

comparable weight. Due to their high frequency, words that are
often used in papers may be less useful in interpreting the
meaning of the text. Words like "a," "the," and other similar
words may lower the weights of more significant words. To
mitigate this effect, TF is discounted by the inverse document
factor..
‫ݏݐ݊݁݉ݑܿ݋̴݂݀݋̴ݎܾ݁݉ݑ̴݈݊ܽݐ݋ݐ‬
‫ܨܦܫ‬ሺ‫ݓ‬ሻ ൌ ݈‫݃݋‬ሺ ሻ
݊‫݀ݎ݋ݓ̴݃݊݅݊݅ܽݐ݊݋̴ܿݏݐ݊݁݉ݑܿ݋̴݂݀݋̴ݎܾ݁݉ݑ‬

Thus, one can get TF-IDF by multiplying TF and IDF. With

more important phrases, the TF-IDF score would rise
ܶ‫ ܨ‬െ ‫ܨܦܫ‬ሺ‫ݓ‬ሻ ൌ ܶ‫ܨ‬ሺ‫ݓ‬ሻ ‫ܨܦܫ כ‬ሺ‫ݓ‬ሻ
Simple machine learning is a subset of naive Bayes classifiers
in artificial intelligence. The renowned Naive Bayes algorithm is
extended to multinomial NB and pipelining concepts. An
Fig. 1. Flowchart for fake review detection
algorithm decides whether the news is accurate and reliable.
There are other strategies that can be employed for training these
classifiers that emphasis on common concepts, so this is not the
only one.
To determine whether the news is true or false, adopt Naive
Bayes. It is a variety of algorithm used to classify texts into
various people. The Bayes theorem is utilised to ascertain
whether the news is true after the usage of tokens is applied to
information that may or may not be inaccurate. The components
for naiveté are as follows:The likelihood of the prior occurrence
is used in Bayes classification, which contrasts it with the current
event. After computing each and every probability of the event, a
final calculation is done to establish the overall likelihood of the
news when compared to the dataset. Hence, by calculating the
overall likelihood, we may estimate the value and decide if the
news is real or fake.
௉ሺ஺ሻ
ܲሺ‫ܤ‬ሻ ൌ ܲሺ‫ܣ‬ሻ െ  (1)
௉ሺ஻ሻ

Finding the probability of an event, A when event B is true.

ܲሺ‫ܣ‬ሻ ൌ ܴܲ‫ܻܶܫܮܫܤܣܤܱܴܴܱܲܫ‬
ܲሺ‫ܣ‬ȁ‫ܤ‬ሻ ൌ ܱܲܵܶ‫ܻܶܫܮܫܤܣܤܱܴܴܱܲܫܴܧ‬
Finding probability:
ܲሺ‫ͳܤ‬ሻ ൌ ܲሺ‫ͳܣ‬ȁȁ‫ͳܤ‬ሻǤ ܲሺ‫ʹܣ‬ȁȁ‫ͳܤ‬ሻǤ ܲሺ‫͵ܣ‬ȁȁ‫ͳܤ‬ሻ (2)
ܲሺ‫ʹܤ‬ሻ ൌ ܲሺ‫ͳܣ‬ȁȁ‫ʹܤ‬ሻǤ ܲሺ‫ʹܣ‬ȁȁ‫ʹܤ‬ሻǤ ܲሺ‫͵ܣ‬ȁȁ‫ʹܤ‬ሻ (3)
If the probability is 0
ܹ‫ ݐ݊ݑ݋ܿ݀ݎ݋‬൅ ͳ
ܲሺܹ‫݀ݎ݋‬ሻ ൌ
ܶ‫݋݈݊ܽݐ݋‬Ǥ ‫ ݏ݀ݎ݋ݓ݂݋‬൅ ܰ‫݋‬Ǥ ‫ݏ݀ݎ݋ݓ݁ݑݍ݅݊ݑ ݂݋‬

As a result, using this approach, one may assess the reviews'

accuracy.
Fig. 2. Workflow for identifying and analyzing reviews using different
algorithms

3
Authorized licensed use limited to: China University of Petroleum. Downloaded on April 10,2024 at 07:35:52 UTC from IEEE Xplore. Restrictions apply.
 Support Vector Machine (SVM) is a commonly used
supervised learning approach that is utilized for classification
and regression problems. Although it can be used for both, it
is primarily used in Machine Learning Classification
problems. SVM algorithm aims to identify the optimal line
or decision boundary, known as a hyperplane, that can divide
n-dimensional space into classes so that new data points can
be categorized quickly in the future. The SVM algorithm
chooses the extreme points and vectors of the hyperplane.
The name "support vector machine" comes from these
exceptional circumstances, which are described by support
vectors. Below is an example of how a decision boundary or
hyperplane can be used to separate two different categories:
Linear SVM classifier is employed for data that can be
separated by a single straight line into two classes, while
non-linear data and non-linear SVM classifier refer to data
that cannot be classified using a straight line.The terms "non-
linear data" and "non-linear SVM classifier" refer to data that
cannot be categorized using a straight line. For non-linearly
separated
p data,, non-linear SVM is utilized.
Fig. 3. Statistical analysis of review
IV. RESULTS & DISCUSSION
The identification of fake reviews has become
increasingly common on websites and social media
networks. To address this issue, our team utilized text
processing and Naive Bayes to develop a model that can
detect fake reviews. By leveraging machine learning tools,
we were able to classify news as fake or not fake in a shorter
amount of time by drawing upon prior data set values. This
provides users with a greater sense of trust in reviews that
appear on social media and other sources.
TABLE I. SUMMARY OF THE DATASET
Total number of reviews 5853 reviews
Number of fake reviews 1144 reviews
Number of real reviews 4709 reviews
Number of distinct reviews 102739 words
Total number of tokens 103052 tokens
The maximum review length 875 words
The minimum review length 4 words
The average review length 439.5 words
4
Authorized licensed use limited to: China University of Petroleum. Downloaded on April 10,2024 at 07:35:52 UTC from IEEE Xplore. Restrictions apply.
The spread of erroneous information seriously harms users
and the social environment. It is difficult to spot a false review in
the first place because it is meant to deceive the user. Many
different avenues are used to spread false information, which
disturbs society and the lives of its residents. Further
improvements would include identifying the source of the
erroneous information and stopping its spread on social media
and online platforms. It would also be able to find and pinpoint
the sources of misleading information in order to stop those who
are seeking to deceive the public. Also, they would track down
the social media profiles of those spreading rumors and false
information so they could halt them before they spread.
V. CONCLUSION
In this proposed work, independently working machine
learning algorithm models was developed for assessing the
reviews of the products. The propsed model was developed using
Naïve Bayes, Support Vector Machine,and TF-IDF (term
frequency-inverse document frequency )Vectorizer. The
proposed model efficiently detected the spam reviews on a
website or application instantly. The proposed work was tested
using” Reviews of 20 Hotels in Chicago hotel dataset” and
achieved greater accuracy and authenticity in the reviews.
REFERENCES
[1] D. F. Murad, Y. Heryadi, B. D. Wijanarko, S. M. Isa and W. Budiharto,
"Recommendation System for Smart LMS Using Machine Learning: A
Literature Review," 2018 International Conference on Computing,
Engineering, and Design (ICCED), Bangkok, Thailand, 2018, pp. 113-118,
doi: 10.1109/ICCED.2018.00031.
[2] S. M. Anas and S. Kumari, "Opinion Mining based Fake Product review
Monitoring and Removal System," 2021 6th International Conference on
Inventive Computation Technologies (ICICT), Coimbatore, India, 2021,
pp. 985-988, doi: 10.1109/ICICT50816.2021.9358716.
[3] Jain, Piyush &Chheda, Karan & Lade, Mihir. (2019). Fake Product Review
Monitoring System. International Journal of Trend in Scientific Research
and Development. Volume-3. 105-107. 10.31142/ijtsrd21644.
[4] Wahyuni, Eka &Djunaidy, Arif. (2016). Fake Review Detection From a
Product Review Using Modified Method of Iterative Computation
Framework. MATEC Web of Conferences. 58. 03003.
10.1051/matecconf/20165803003.
[5] Kashid, Aishwarya & Lalwani, Ankita &Gaikawad, Saniksha& Patil,
Rajal&Sonkamble, Rahul & More, Shivaprasad. (2021). Fake Review
Detection System Using Machine Learning.
[6] R. Patel and P. Thakkar, "Opinion Spam Detection Using Feature
Selection," 2014 International Conference on Computational Intelligence
and Communication Networks, Bhopal, India, 2014, pp. 560-564, doi:
10.1109/CICN.2014.127.
[7] Saumya, S., Singh, J.P. Detection of spam reviews: a sentiment analysis
approach. CSIT 6, 137–148 (2018). https://doi.org/10.1007/s40012-018-
0193-0
[8] N. Sodera and A. Kumar, "Open problems in recommender systems
diversity," 2017 International Conference on Computing, Communication
and Automation (ICCCA), Greater Noida, India, 2017, pp. 82-87, doi:
10.1109/CCAA.2017.8229776.
[9] Ata-Ur-Rehman et al., "Intelligent Interface for Fake Product Review
Monitoring and Removal," 2019 16th International Conference on
Electrical Engineering, Computing Science and Automatic Control (CCE),
Mexico City, Mexico, 2019, pp. 1-6, doi: 10.1109/ICEEE.2019.8884529.
[10] Mukherjee, A., Venkataraman, V., Liu, B., & Glance, N. (2021). What
Yelp Fake Review Filter Might Be Doing?. Proceedings of the
International AAAI Conference on Web and Social Media, 7(1), 409-418.
[11] 2015. Proceedings of the 21th ACM SIGKDD International Conference on
Knowledge Discovery and Data Mining. Association for Computing
Machinery, New York, NY, USA.
[12] A. Sihombing and A. C. M. Fong, "Fake Review Detection on Yelp
Dataset Using Classification Techniques in Machine Learning," 2019
International Conference on contemporary Computing and
Informatics (IC3I), Singapore, 2019, pp. 64-68, doi:
10.1109/IC3I46837.2019.9055644
[13] A. Prabhat and V. Khullar, "Sentiment classification on big data using
Naïve bayes and logistic regression," 2017 International Conference
on Computer Communication and Informatics (ICCCI), Coimbatore,
India, 2017, pp. 1-5, doi: 10.1109/ICCCI.2017.8117734.

5
Authorized licensed use limited to: China University of Petroleum. Downloaded on April 10,2024 at 07:35:52 UTC from IEEE Xplore. Restrictions apply.