Analyzing Sentiment in Terms of Online Feedback on Top of Users' Experiences

Full text

(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 14, No. 11, 2023
144 | P a g e
www.ijacsa.thesai.org
Analyzing Sentiment in Terms of Online Feedback on
Top of Users' Experiences
Mohammed Alonazi
Department of Information Systems, College of Computer Engineering and Sciences,
Prince Sattam bin Abdulaziz University, Al-Kharj, 16273, Saudi Arabia
Abstract—Since most businesses today are conducted online,
it is crucial that each customer provide feedback on the various
items offered. Evaluating online product sentiment and making
suggestions using state-of-the-art machine learning and deep
learning algorithms requires a comprehensive pipeline. Thus,
this paper addresses the need for a comprehensive pipeline to
analyze online product sentiment and recommend products using
advanced machine learning and deep learning algorithms. The
methodology of the research is divided into two parts: the
Sentiment Analysis Approach and the Product Recommendation
Approach. The study applies several state-of-the-art algorithms,
including Naïve Bayes, Logistic Regression, Support Vector
Machine (SVM), Decision Tree, Random Forest, Bidirectional
Long-Short-Term-Memory (BI-LSTM), Convolutional Neural
Network (CNN), Long-Short-Term-Memory (LSTM), and
Stacked LSTM, with proper hyperparameter optimization
techniques. The study also uses the collaborative filtering
approach with the k-Nearest Neighbours (KNN) model to
recommend products. Among these models, Random Forest
achieved the highest accuracy of 95%, while the LSTM model
scored 79%. The proposed model is evaluated using Receiver
Operating Characteristic (ROC) - Area under the ROC Curve
(AUC). Additionally, the study conducted exploratory data
analysis, including Bundle or Bought-Together analysis, point of
interest-based analysis, and sentiment analysis on reviews (1996-
2018). Overall, the study achieves its objectives and proposes an
adaptable solution for real-life scenarios.
Keywords—Sentiment analysis; product review; machine
learning; recommendation system; collaborative filtering;
exploratory data analysis
I. INTRODUCTION
In this era of modern computational technology,
technological advancement can be seen everywhere; even the
business sector is taking the initiative to enhance its revenue in
computing and technology [1]. The term sentiment analysis has
extensively been utilized to track out social media, allowing
businesses to extract hidden information from the recorded
data or identify critical information before coming into the
limelight [2]. Thinking about giant tech companies such as
Facebook, Google, Apple, and Microsoft, they have a huge
amount of datasets. Every day, lots of data is being recorded to
their central database. Besides, manually analyzing these data
is time-consuming [3]. With a massive amount of dataset, it is
quite difficult to manually extract meaningful insights that can
help them make a business oriental decision. Turning into a
product-based company, it can be stated that the product-based
company tends to develop its product and launch it into the
market [4]. If the thing is grocery or jewelry items so in this
case, users will be giving their opinion based on the items
whether the product is caught their attention or not. Suppose
the clients explore a large e-commerce platform like
Amazon.com. In this case, it is noticeable that the end-users
threw their comments or reviews related to a specific item.
Other individuals take themselves towards the advertisement
phenomenon [5]. In turn, this brings us to a big question,
whether it or not possible to handle such massive amounts of
ratings manually and extract the business insights. So, the
automated system can be the possible solutions to overcome
these issues [6]. Thus, the impact of information on user
sentiment and physical environments is not limited to modern
technology.
The computational approach can be taken into
consideration. Nowadays, machine learning algorithms have
widely been utilized in biomedical imaging, forecasting things
or even critical disease prognosis [7]. For the case of product
analysis, it has shown their promising performance beforehand.
Researchers are now using computing power to take their
analysis to a satisfactory level from where meaningful insights
can be extracted easily by analyzing a large number of datasets
[8]. The priority of this research is to analyze the product
sentiment using machine learning algorithms and propose a
recommendation system for the stakeholder to make a better
decision while doing online business. In this research,
conventional Machine Learning (ML) algorithms were adopted
to analyze the online product, and our study will significantly
contribute to the research community. This is the motivation of
this proposed study.
On the other hand, there are three contributions have been
addressed in this study, and these are following:
 Three types of data analysis have been completed and
through which business owners can make a variety of
decisions towards their product.
 Various machine learning algorithms were applied to
check their credibility to analyze the Sentiment, and
satisfactory accuracy was turned out to be successful.
This research also ensures the robust machine learning
pipeline that achieved a good accuracy, and the
concentrated model can be deployed to a webserver to
achieve sustainable goals.
 The different assessment pointers bend assessed the
proposed show, and at long last, a proposal framework
has been submitted by coordination overall sifting
strategy. By taking after the proposal framework, the
This study is supported via funding from Prince Satam bin Abdulaziz
University project number (PSAU/2023/R/1444).

(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 14, No. 11, 2023
145 | P a g e
www.ijacsa.thesai.org
partner will be able to supply important data to their
enlisted client, which can improve the request for
specific things.
The manuscript is classified into six interconnected
sections. Section II presents the exciting works with research
gap analysis. Section III depicts the overall methodology of the
proposed system with proper discussion. Section IV shows the
results associated with our proposed solutions, including data-
driven analysis and approaches. Section V represents the
observations and discussion on the results. Finally, the
conclusion of the research with future work will be discussed
in Section VI.
II. LITERATURE REVIEW
This section illustrates the background study of the
previous works related to the proposed model. In this section,
the research gaps have been extracted with proper discussion.
Many great contributors have traced fruitful online product
contributions or sentiment analysis contributions.
The authors of the paper in [9] worked on an efficient way
to optimize the accuracy of the sentiment analysis in Egyptian
Arabic. The proposed work was identically based on the
conventional semantic orientation and machine learning
techniques, and the authors had achieved the highest accuracy
of 92.98% while working with Support Vector Machine
(SVM). The purpose of the article in [10] was to examine the
attitudes of buyers regarding electrical devices by analyzing
various sale tweets. The experimental results of the proposed
research will be valuable to a variety of business organizations
in making business decisions that will ultimately increase the
sales of the products they offer. The author of the paper also
claimed that they had achieved the highest accuracy of 86%,
91%, and 91% with the Logistic Regression (LR) in the phone,
laptop, and television, respectively.
The paper in [11] aimed to extract the text features into the
semantics of words. The authors adopted a Word Sense
Disambiguation (WSD) technique to extract the features from
the reviewing sentences. A supervised learning approach has
been adopted to analyze the product reviews and utilized 10-
fold cross-validation to validate the results. The authors had
significantly optimized the performance by 10.6%, while
precision was 10.9% higher and recall was 9.2% higher than
baseline approaches. The author of the paper in [12] had
presented a significant comparison among several conventional
deep learning-based models for word embedding in product
sentiment analysis. Thus, they adopted data augmentation
techniques to enrich the dataset and classify it into identical
classes. The research also claimed they found the highest
accuracy of 96% while working with CNN-RNN based BI-
LSTM algorithms.
In paper [13] proposed an Adaptive Neuro-Fuzzy
Inferences System (IANFIS) model to produce a way of
analyzing the sentiment of online products. The method is
identically based on natural language processing to track the
user's opinion. The authors classified the dataset into three
interconnected parts: contents, grades, and collaborations.
Then, they applied deep learning algorithms to make a
prediction on the negative and positive comments from the
users. The research also performs a comparison among the
existing solutions. The paper in [14] aimed to implement a
model for analyzing the sentiment of the users in movie
reviews. The authors extracted the feeling and feedback from
existing text patterns. The models had the ability to detect
several types of feeling like negative, positive, and even
neutral. To accomplish this goal, they utilized different
machine learning algorithms and classifiers, and mechanisms
of natural language processing.
The authors in [15] have presented a machine learning-
based online product sentiment analysis. In this work, they
showed the labeled product reviews in several websites with
the help of supervised and unsupervised (lexicon-based) based
algorithms. The models were then applied to the iPhone 5s
reviews collected from the existing popular online shops. The
authors further extracted the combination of unigram and
bigram features, which placed the best results while dealing
with machine learning-based classifiers.
Paper in [16] identified three subtasks that must be
addressed: the definition of the target; the separation of good
and bad news content from good and bad sentiment expressed
on the target; and the analysis of clearly marked opinion that is
defined explicitly, without the need for interpretation or the use
of world knowledge. The authors in [17] created a new strategy
that combines previous approaches to provide the best
coverage results and competitive agreement. They had also
proposed iFeel, a free Web service that provides an open API
for retrieving and comparing findings from several sentiment
methods for a given text. In paper [18], researchers categorized
movie reviews using features based on these taxonomies paired
with traditional "bag-of-words" features, and reported 90.2
percent accuracy. Furthermore, they discover that some types
of assessment appear to be more critical for sentiment
classification than others.
The contributions of the papers in [19] had only focused on
the development of a notable features selection on online
product sentiment analysis. But the researchers didn‟t focus on
the correct terms of algorithms and algorithm tuning to
optimize sentiment analysis accuracy level. In sharp contrast,
the manuscript presented three forms of data analysis. These
analyses will allow business owners to make several judgments
regarding their specific products. Various machine learning
methods were used to check their reliability and determine
adequate accuracy. This research also assures a robust machine
learning pipeline that the condensed model can be deployed to
a webserver to fulfill long-term objectives for product
sentiment analysis.
Based on the literature review, several studies have been
conducted to improve sentiment analysis and product
recommendation using machine learning techniques. The
authors of one study achieved the highest accuracy of 92.98%
in sentiment analysis of Egyptian Arabic using Support Vector
Machine. Another study analyzed buyer attitudes towards
electronic devices through sale tweets and achieved high
accuracy of 86%, 91%, and 91% for phone, laptop, and
television respectively using Logistic Regression. One study
aimed to extract text features through Word Sense
Disambiguation and achieved a significant performance

(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 14, No. 11, 2023
146 | P a g e
www.ijacsa.thesai.org
improvement compared to baseline approaches. Lastly, a study
compared several deep learning-based models for word
embedding in product sentiment analysis and achieved an
accuracy of 96% using a CNN-RNN based BI-LSTM
algorithm.
Overall, the studies show the effectiveness of machine
learning techniques in sentiment analysis and product
recommendation. These approaches can help businesses make
data-driven decisions to improve sales and customer
satisfaction. However, there is still room for further research to
optimize the accuracy and efficiency of these techniques.
III. METHODOLOGY
This section presents the overall design of the proposed
model, including the illustration. Fig. 1 illustrates a block
diagram of the proposed research through which this study was
conducted. By looking at Fig. 1, it can be observed that this
research was carried out through three interconnected stages.
The experimental dataset was collected and preprocessed to
model the data in the first stage. In the second phase, product
sentiment analysis was accomplished. Finally, in the third
stage, a recommendation system is proposed as the priority of
this research is to provide a model that can analyze the product
sentiment and transform the traditional business into a data-
driven approach.
Fig. 1. Block diagram of the proposed research, including the experimental
analysis to product recommendation approach.
A. Sentiment Analysis Approach (SAP)
This section explains the steps of approach, including
experimental data, data preparation and concentrated research
algorithms in more details.
1) Experimental dataset: From May 1996 through July
2018, Amazon's "Clothing, Shoes, and Jewelry" category
received 2.5 million product ratings and information from 2.5
million customers. This collection includes reviews (scores,
description, and sentiment comments), product metadata
(descriptions, controls and monitors, pricing, branding, and
image attributes), and links [20].
2) Data preparation: Perusing different JSON records
from a single JSON record, 'ProductSample.json,' and
including them in the list in such a way that each list of the list
has the substance of a single JSON record [21]. Following
that, iterate over the list, loading each index as JSON,
extracting the data from each index, and creating a list of
Tuples containing all the data from the JSON files. Again,
each cycle begins with a clean JSON file converted to the
right JSON format using some substitutions. Finally, a data
frame is created using the list of Tuples obtained in the
previous step.
3) Concentrated research algorithms (CRA): The
Concentrated Research Algorithms (CRA) indicates the
suggested model that has been adopted in this study to analyze
the research data [22]. It is to be specified that various
conventional techniques were applied in this investigation;
among them, the Naïve Bayes and Decision tree algorithms
were found to be satisfactory [23]. So, the mathematical
interpretation and model optimization procedure were
highlighted in this section. These models have a benchmark
performance that appeared in the previous research. In
addition, model performance is depending on the data
distribution, furthermore, the Naïve based and Decision tree
are capable enough to handle the product sentiment analysis
data.
 Naïve Bayes Algorithm: Nave Base Classifier is a
classification-type machine learning algorithm [24].
This algorithm is based on the Base Theorem. Simply
put, the base theorem is a method of determining the
probability of one event (X) occurring and another
event (Y) occurring. If clouds are seen in the sky, there
is a possibility of rain. The base theorem can be
mathematically written as,
󰇛󰇜󰇛󰇜
󰇛󰇜 (1)
The above equation is a simple base equation used to
determine probability in the case of a conditional event only. In
practice, most datasets are multivariate, in which case the
equation becomes a bit more complicated. Then we can write
the equation like this:
󰇛     󰇜󰇛󰇜󰇛󰇜󰇛󰇜󰇛󰇜
󰇛󰇜󰇛󰇜󰇛󰇜 (2)
A few of the highlights of Naïve Base Classifier: It is
exceptionally simple to execute and works moderately quick,
works well indeed on small datasets, gives a small less
exactness than other calculations, and all traits in Naïve Base
are considered commonly autonomous but within the genuine
world Isn't.
 Decision Tree: Both classification and regression
problems can be solved with the classification and
regression tree or CART algorithm [25]. In short, many
people call it the Decision Tree. The decision tree looks

(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 14, No. 11, 2023
147 | P a g e
www.ijacsa.thesai.org
a lot like the branches of a tree, which is why the word
'tree' is associated with its name. The decision trio starts
from the 'root node' just as the tree starts from the root.
From the root node, the branches of this tree spread
through different decision conditions; such nodes are
called decision nodes, these nodes are called leaf nodes
after making a final decision. Other Parameters of the
Decision Tree: Splitting - The process of moving a
dataset across a series of variables, starting from the
root node, is called splitting [26]. Entropy - Entropy is
the amount of chaos. When the tree is split, the amount
of data of the same type/class in each node is purity. All
the data in a pure node are of the same class. The lower
the purity, the higher the entropy. Again, the lower the
entropy, the higher the purity. Information Gain - The
measure of righteousness is information gain [27]. The
higher the information gain, the purer nodes the tree can
create. Guinea Index - Guinea is the probability of all
node members being in the same class. This value
ranges from 0 to 1. Guinea value 0 means all the
members of that node belong to the same category, and
Guinea value 1 means the members of that node are
randomly distributed or of different classes, i.e., entropy
is much higher. If the value of Guinea is 0.5, then the
members of the two classes are equal (if the number of
classes is 2) [28].
  󰇛󰇜󰇛󰇛󰇜  󰇛󰇜󰇛󰇛 (3)
Information Gain= Entropy Before Split - Entropy After
Split
    󰇛󰇛  󰇜󰇜 (4)
B. Product Recommendation Approach (PRA)
This section highlighted the product recommendation
procedure. It can be said that the PRA is vital towards business
transformation because user behavior and pattern cannot easily
be identified if the stakeholder did not design any
recommendation system. By looking at Fig. 2, it is noticeable
that a flow chart has been proposed in terms of the
recommendation system. The collaborative filtering approach
is selected that will filter out the user ID based on age, gender,
location and rating score etc. After having all of that
information, the system will make a comparison set for the
specific users. However, the detailed sequence and
consequences are shown in Fig. 2. The diagram in Fig. 2
suggests that the proposed pipeline aims to improve e-
commerce and enhance customer experience by using a
collaborative filtering method for item recommendation. It
implies that the pipeline could help businesses identify the
most relevant products for their customers, thus improving
their overall shopping experience. The use of collaborative
filtering suggests that the pipeline may leverage the behavior of
similar users to provide personalized recommendations,
ultimately leading to increased customer satisfaction and
potentially higher sales. Overall, the diagram title hints at a
promising approach to improving the online shopping
experience and driving business growth.
Fig. 2. Efficient pipeline for the item proposal approach towards commerce
change and client design acknowledgment through the collaborative sifting
method.
The research system utilizes several methods like K-
Nearest Algorithm (KNN) [29], the Jaccard's coefficient, the
Dijkstra algorithm, and the cosine similarity. The aim is to
suggest based on users' behavior patterns. In recommendation
systems, the most common types are the Collaborative
Filtering Method (CFM), the Content-Based Filtering
Approach (CFA), and Hybrid Recommendation System (HRS)
[30]. This filtering approach generally focuses on collecting
and analyzing user experience information, behaviors, or
interests and predicting what they would like based on
similarity with other users. The collaborative sifting approach's
imperative advantage is that it does not depend on machine
analyzable substances and can accurately prescribe complex
things without requiring an "understanding" of the thing itself.
A typical recommendation engine processes data over the
following four steps: selection, storage, analysis, and filtering.
We have applied the K-Nearest Algorithm (KNN), the
Jaccard's coefficient, the Dijkstra algorithm, and the cosine
similarity to forecast the shortest path from the user‟s current
location to the user‟s desired destination and as well as to
suggest places to the users based on the rating. In Figure 3 a)
shows the corresponding KNN's cluster filtering working
procedure and b) shows the Dijkstra algorithm „s workflow.

(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 14, No. 11, 2023
148 | P a g e
www.ijacsa.thesai.org
(a)
(b)
Fig. 3. a) Perform K –nearest algorithm's cluster filtering, b) Dijkstra
algorithm.
Fig. 4. Architecture diagram of the product recommendation system using
various filtering methods.
In Fig. 4, it can be stated that the suggested
recommendation engine is the result of a number of
interconnected methodologies being used. In order to utilize
the system, the user must first complete their registration
before being allowed to proceed. After logging into a system, a
log file will be created automatically to keep track of the user's
patterns of behavior. At this point, we have implemented a
cooperating filtering method as well as a content-based
filtering approach. Users' comments and product ratings will be
taken into consideration by the recommendation system, which
will take action based on both forms of data. After obtaining all
of the necessary parameters, the system will proceed to the
next stage of proposing a specific product to a user who has
expressed interest in it. The architecture also assures that
collaborative and content-based filtering are the middle layer of
this architecture that are the responsible for the product
sentiment analysis. However, the hybrid parameters will then
send to the model for identifying the sentiment.
IV. RESULT ANALYSIS
The result analysis section is categorized into several parts:
Classification Metrics Interpretation (CME), Measuring the
Efficiency, Interpretation of Sentiment Analysis, Observation
& Discussion. The precision of expectations from the
classification calculations is evaluated by applying a
classification report. The report illustrates the exactness,
review, and f1-score of the key classification measurements per
lesson. These measurements are computed by utilizing genuine
and untrue positives and genuine and wrong negatives. The
measurements comprise of four components: genuine positive,
untrue positive, genuine negative, wrong negative, and wrong
negative. The taking after Condition (1), (2), (3), and (4) was
considered for finding the exactness, review, and f1-score. In
Table I, the classification report of the machine learning
calculation, is depicted. In this table, we can clearly observe
that the machine learning algorithms like Random Forest (RF)
and Logistic Regression provide better results, such as 95%
and 94%, compared to the other algorithms like LSTM and
CNN_LSTM. This is because for this dataset, we have found
that low-cost classifiers work far better at deep computation
because of the small size of the dataset. Thus, we have
achieved the highest accuracy from conventional classifiers.
Precision: It is the relationship between the true positive
estimate of the model and the overall positive estimate (both
accurate and wrong) .It is articulated as:
󰇛󰇜  
 (5)
Recall / Sensitivity: The probability of being capable of
predicting is a positive ratio. It is given in mathematical form
as:
󰇛󰇜  
 (6)
F1-score: As a general rule, the harmonic mean for
Accuracy and Review provides a much better; a significantly
better; higher; a stronger and more enhanced gauge than the
Precision Metric of the incorrectly categorized occurrences. It
is given, mathematically, as:
      
 (7)

(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 14, No. 11, 2023
149 | P a g e
www.ijacsa.thesai.org
Accuracy: It is the sum of all the cases in which the
predictions were right. It is given as:
󰇛󰇜  
 (8)
TABLE I. CLASSIFICATION REPORT OF THE CONCENTRATED
ALGORITHMS
Algorithm
Accuracy
(%)
Precision
(%)
Recall
(%)
F1-score
(%)
Naïve Bayes (NB)
93
92
93
93
Logistic Regression (LR)
94
93
94
94
SVM
93
93
93
93
Decision Tree (DT)
91
89
89
90
Random Forest (RF)
95
94
95
95
BI-LSTM
76
70
71
70
CNN_LSTM
77
77
75
76
Stacked LSTM
76
76
76
76
LSTM
79
78
78
77
Fig. 5. Performance analysis from the different models.
On the other hand, Fig. 5 highlights the confusion matrix
on top of the Random Forest and LSTM model. The confusion
matrix consists of four values: TP, TN, FP, and FN. We can
identify the total sensitivity and specificity ratio by following
the Confusion matrix. This is another model evaluation
indicator, and in the field of data science, this matrix has been
utilized extensively to measure a specific model. The confusion
matrix consists of four values, True positive, True negative,
False positive, and False negative are the four values in the
confusion matrix. Fig. 6 (a) and (5) illustrates the significant
proportion of true positive and false negative values.
ROC-AUC curve is used to determine how good a model
is. This evaluation indicator distinguishes the positive and
negative data points from the dataset. If the ROC curve goes to
1.0, the model can accurately differentiate the positive and
negative data points. Fig. 7 (a) and (b) are almost close to 1.0
or the area under the curve. It can be stated that this model is
applicable to use in real life.
(a)
(b)
Fig. 6. (a) Confusion matrix on CNN+LSTM (b) Visualizing the confusion
matrix on top of the BI-LSTM model towards product sentiment analysis.
(a)
(b)
Fig. 7. (a) Evaluating curve on top of the Random Forest model (b)
Measuring the model and visualizing the ROC-AUC curve on top of the
LSTM model.

(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 14, No. 11, 2023
150 | P a g e
www.ijacsa.thesai.org
A. Interpretation of Sentiment Analysis
This analysis is divided into three stages: Sentiment
analysis on reviews (1996-2018), exploratory analysis on
product reviews (1996-2018). This is of particular note since it
is the one with the greatest number of prior sales on Amazon,
as determined by an examination of "Bundles" or "Bought-
Together."
B. Analysis 1: Sentiment Analysis on Reviews (1996-2018)
Fig. 8 indicates the exploratory product data analysis in
terms of user feedback. In Fig. 8 (a) and Fig. 8 (b), the
sentiment analysis on user reviews have been demonstrated for
the year of between 1996 to 2018. It is noticeable that, the year
2000 was the least number and 2001 reach out at the highest
number of negative product reviews based on the sentiment. At
the same time, the year 2000 was the top year where significant
number of positive reviews based on the users sentiment was
recorded. After that, the percentage was decreased and
remained same for the next consecutive years. Besides, Fig. 7
(c) and (d) shows the word cloud visualization based on the
positive and negative reviews where it the keywords are tagged
and, in this way, it would be convenient to find out the negative
and positive keywords from the dataset.
For Positive Word Cloud, some well-known terms such as
adored, idealized, decent, wonderful, best, outstanding, and so
on were used to describe the goods. A large number of people
who were polled were pleased with the prices of things sold on
Amazon. Bra, coat, bag, and outfit are some of the most
commonly discussed goods. Disappointment, terrible fit,
horrible deformity, return and etc. are some of the well-known
adjectives used to describe the things. Some of the most talked-
about goods were shoes, binoculars, bras, batteries, and so on.
Predominant item in terms of how others feel about it. There
are 953 positive reviews for the Speak Unisex Chuck Taylor
Classic Colors Sneaker. There are 672 positive reviews for the
Talk Unisex Chuck Taylor All-Star Howdy Best Dark
Monochrome Sneaker. There are 65 negative reviews for the
Yaktrax Walker Footing Cleats for Snow and Ice. There are 44
negative reviews for the Speak Unisex Chuck Taylor Classic
Colors Sneaker. Welcome to the Best Dark Monochrome
Sneaker, with a total of 247 honest reviews. [31], which state
achieved that 72.7 % was positive, 5 % was negative, and 22.3
% was neutral. Overall, Sentiment for reviews on Amazon is
on the positive side as it has very few negative sentiments.
The drift for Rate of Audit over a long time, positive
surveys rate has been lovely reliable between 70-80 all through
a long time. Negative surveys have been diminishing recently
since final three a long time; possibly they worked on the
administrations and issues.
C. Analysis 2: Point of Interest-based Analysis
The lexical density of a language is a concept in
computational linguistics that measures the structure and
complexity of human conversation in [31]. Functional and
content words are used to calculate a written or spoken
composition's lexical density.
a) Negative reviews over the years based on sentiments.
b) Positive reviews over the years based on Sentiments.
c) Visualizing the negative observation.
d) Visualizing the postive observation.
Fig. 8. a) Negative reviews over the years based on the Sentiment (b)
Positive reviews based on the Sentiment (c) Visualizing the negative
observation through the word cloud visualization approach (d) Illustrating the
positive observation with the help of word cloud visualization.

(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 14, No. 11, 2023
151 | P a g e
www.ijacsa.thesai.org
Fig. 9. Lexical density over years.
By looking at the Fig. 9, it can be observed that lexical
density over years have been displayed. In 2018, the significant
amount just over the 40 but in 1996, there is a downward trend
at nearly 36 and remained steady for the next subsequent years.
D. Analysis 3: Bundle’ or ‘Bought- Together’ based Analysis
In the Table II, Bundle or Bought together based analysis
has been interpreted in terms of up vote, helpful rating, total
votes, and percentage. Based on the reviewer ID, it is clearly
demonstrated the helpful rating, and these has been illustrated
owing to the fact that for the case of analysis, product
sentiment analysis and review records are considerably
required so that essential information can be extracted.
Taking a close look at the Fig. 10, helpfulness and average
length have been displayed. The findings show that the
effectiveness of review length is influenced by product
category; longer evaluations are more helpful for think
products. Furthermore, review helpfulness is linked to the
degree of consistency between individual review ratings and
total product ratings. In contrast the Fig. 11 shows the data
analysis with correlation between the ASIN.
E. Analysis 3: Exploratory Data Analysis
1a) The common survey rating for the foremost commonly
checked items is between 4.5 and 4.8, with little variation. 1b)
Whereas there's a little converse affiliation between the
recurrence level of ASINs and regular audit appraisals for the
primary four ASINs, this relationship isn't critical since the
normal survey for the prior four ASINs is evaluated between
4.5 and 4.8, which is respected greatly by and large audits. 2a)
As illustrated within the bar chart (beat), ASINs with lower
frequencies have much more change in their routine survey
evaluations on the point-plot chart (foot), as demonstrated by
the length of the vertical lines. As a result of the tall
fluctuation, we accept that our research's normal audit
evaluations for ASINs with lower frequencies are not
imperative. 2b) On the other hand, we assume that the lower
frequencies for ASINs are inferable to lesser quality items. 2c)
Moreover, the final four ASINs have no change due to their
significantly lower frequencies. Whereas the survey appraisals
are a culminate 5.0, we ought not to consider these audit
appraisals' significance due to the lower recurrence as shown in
2a). Based on our information examination between ASINs
and audits. Rating, we have taken note that numerous ASINs
with the common event had huge fluctuations; in this way, we
decided that these moo event ASINs are not imperative in our
think about due to the little test measure. Additionally, we
found nearly no interface between ASINs and surveys. Rating
in our relationship considers which is reliable with our
findings.
Fig. 10. Helpfulness and average length.
TABLE II. EXPLORARYTORY DATA ANALYSIS WITH THE DATASET AND THEIR FEATURES
Reviewer ID
Rating helpful
Upvote
Total Votes
Percentage
Rating
A2XVJBSRI3SWDI
5.0
0.0
N/A
0.0
N/A
A2G0LNLN79Q6HR
4.0
1.0
0.0
N/A
N/A
A2R3K1KX09QBYP
2.0
1.0
100.0
N/A
N/A
A19PBP93OF896
1.0
1.0
100.0
N/A
N/A
A19PBP93OF896
0.0
0.0
0.0
N/A
N/A
A0000188NWOSI5X2PMS
0.0
N/A
0.0
1.0
N/A
A000063614T1OE0BUSKUT
N/A
N/A
N/A
0.0
5.0
A00031045Q68JAQ1UYT
N/A
N/A
N/A
4.0
N/A
A00028781NF0U7YEN9U19
N/A
N/A
N/A
0.0
5.0
A00031045Q68JAQ1UYT
N/A
N/A
N/A
100.0
1.0

(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 14, No. 11, 2023
152 | P a g e
www.ijacsa.thesai.org
(a)
(b)
Fig. 11. (a) Exploratory data analysis on top of the ASIN (b) Correlation
analysis between ASIN.
V. DISCUSSIONS
The foremost regularly looked into products have routine
audits within the 4.5 to 4.8 extend, with small fluctuation. Even
though there's a slight converse relationship between the
ASINs recurrence level and normal audit evaluations for the
primary 4 ASINs, this relationship is immaterial. The normal
survey for the prior 4 ASINs is evaluated between 4.5 to 4.8,
which large surveys consider great. For ASINs with lower
frequencies, we see that they're comparing normal audit
evaluations on the point-plot chart (foot) have an essentially
bigger change, as appeared by the length of the vertical lines.
As a result, we recommend that the normal audit appraisals for
ASINs with lower frequencies are not critical for our
examination due to high variance.
On the other hand, due to their lower frequencies for
ASINs with lower frequencies, we recommend that this result
from more second-rate quality items 2c). Moreover, the final 4
ASINs have no fluctuation due to their altogether lower
frequencies. Even though the audit evaluations are a culminate
5.0, we ought not to consider the centrality of these review
evaluations due to lower recurrence. I am able to see that
certain items have much more reviews than others based on the
ASIN investigation, proposing a greater deal for those items.
Ready to see that the ASINs have "right-tailed" dissemination,
demonstrating that specific things have bigger deals, which can
be connected to the higher recurrence of ASINs within the
audits. Moreover, we took the log of the ASINs to normalize
the information so that we seem to get a more nitty-gritty to see
each ASIN and see that the dissemination is still "right-tailed."
In our study, product sentiment analysis has been carried out
towards in the year between 1996 to 2018. Three types of data
analysis have been completed and through which business
owners can make a variety of decisions towards their particular
product. The different assessment pointers bend assessed the
proposed show, and at long last, a proposal framework has
been submitted by coordination overall sifting strategy. By
taking after the proposal framework, the partner will be able to
supply important data to their enlisted client, which can
improve the request for specific things.
VI. CONCLUSION AND FUTURE WORK
Nowadays, product sentiment is very important because,
when a business is run online, it is important for every user to
recommend their various products through pattern recognition.
In order to use cutting-edge machine learning and deep
learning algorithms to evaluate online product sentiment and
make recommendations, a thorough pipeline is needed. This
research proposes a pipeline for analyzing the online product.
Also, a recommendation system has been presented through
which a similar product can be filtered out for users. The study
approach comprises two distinct components, namely the
sentiment analysis approach and the product recommendation
approach. The study uses appropriate hyperparameter
optimization techniques to apply a number of cutting-edge
algorithms, such as Naïve Bayes, Logistic Regression, Support
Vector Machine (SVM), Decision Tree, Random Forest,
Bidirectional Long-Short-Term Memory (BI-LSTM),
Convolutional Neural Network (CNN), Long-Short-Term
Memory (LSTM), and Stacked LSTM. The k-Nearest
Neighbors (KNN) model is combined with the collaborative
filtering approach in the study to make product
recommendations. Of these models, the Random Forest model
had the highest accuracy (95%), followed by the LSTM model
(79%). The Area under the ROC Curve (AUC), also known as
the Receiver Operating Characteristic (ROC) curve, is used to
evaluate the proposed model. In addition, the study carried out
exploratory data analysis on reviews (1996–2018) using point-
of-interest-based analysis, sentiment analysis, and bundle or
bought-together analysis. Overall, the study meets its goals and
suggests a flexible fix for practical situations.
Different machine learning and deep learning algorithms
were applied to analyze the sentiment in this research. The
Random Forest was found to be satisfactory through the
investigation and can recommend any product effectively. In
addition, three types of analysis have been carried out in this
study. This research has several limitations, such as
experimenting with just one dataset, but experimenting on
multiple datasets was required, which we will complete in the
future. In addition, a software system will be developed where
a recommendation system will be integrated. Also, various loss
optimization formulas will be applied to ensure model
efficiency. Evaluation indicator approaches will later justify the
type of model followed for recommendation in this phase.
REFERENCES
[1] G. S. Blair, K. Beven, R. Lamb, R. Bassett, K. Cauwenberghs et al.,
“Models of everywhere revisited: a technological perspective,”
Environmental Modelling & Software, vol. 122, pp. 104521, 2019.
[2] N. Mantel and W. Haenszel, “Statistical aspects of the analysis of data
from retrospective studies of disease,” National Cancer Institute, vol. 22,
no. 4, pp. 719–748, 1959.
[3] B. Jiang, “Investigating the market strategy of smart food‟s latest
product based on business analysis models," in 2021 3rd International
Conference on Economic Management and Cultural Industry (ICEMCI
2021), Guangzhou, China, vol. 43, pp. 2218-2223. 2021.

(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 14, No. 11, 2023
153 | P a g e
www.ijacsa.thesai.org
[4] A. Schwarz, S. Isaksson, U. Källman and M. Rusner, “Enabling patient
safety awareness using the Green Cross method: A qualitative
description of users‟ experience,” Journal of clinical nursing, vol. 30, no.
5, pp.830-839, 2021.
[5] D. Gavilan and M. Gema, “Exploring user‟s experience of push
notifications: a grounded theory approach,” Qualitative Market
Research: An International Journal, vol. 3, pp. 35-57, 2022.
[6] S. Abuelenin, S. Elmougy and E. Naguib, “Twitter sentiment analysis
for arabic tweets,” in International Conference on Advanced Intelligent
Systems and Informatics, London UK, pp. 467-476 .2017.
[7] S. Kumar, V. Koolwal and K. K. Mohbey, “Sentiment analysis of
electronic product tweets using big data framework,” Jordanian Journal
of Computers and Information Technology, vol. 5, pp. 43-59, 2019.
[8] B. Setya Rintyarna, R. Sarno and C. Fatichah, “Semantic features for
optimizing supervised approach of sentiment analysis on product
reviews,” Computers, vol. 8, p. 55, 2019.
[9] V. Balakrishnan, Z. Shi, C. L. Law, R. Lim and Y. Fan, “A deep
learning approach in predicting products‟ sentiment ratings: a
comparative analysis,” The Journal of Supercomputing, vol. 4, pp. 1-21,
2021.
[10] P. Sasikala and L. Mary Immaculate Sheela, “Sentiment analysis of
online product reviews using dlmnn and future prediction of online
product using IANFIS," Journal of Big Data, vol. 7, pp. 1-20, 2020.
[11] Y. Zhu, J. Jiang, W. Han, Y. Ding and Q. Tian, “Interpretation of users‟
feedback via swarmed particles for content-based image retrieval,”
Information Sciences, vol. 375, pp. 246-257, 2017.
[12] W. Muhammad, M. Mushtaq, K. N. Junejo and M. Y. Khan, “Sentiment
analysis of product reviews in the absence of labelled data using
supervised learning approaches,” Malaysian Journal of Computer
Science, vol. 33, pp. 118-132, 2020.
[13] L. Monsurrò and L. Dezi, “Elderly experience of smart objects: how
technology and family support can make senior users overcome their
limits,” In 2021 IEEE International Conference on Technology
Management, Operations and Decisions (ICTMOD), Marrakech,
Morocco, pp. 1-5, 2021.
[14] K. Vembandasamy, R. Sasipriya and E. Deepa, “Heart diseases
detection using naive bayes algorithm,” International Journal of
Innovative Science, Engineering & Technology, vol. 2, pp. 441-444,
2015.
[15] B. Charbuty and A. Abdulazeez, “Classification based on decision tree
algorithm for machine learning,” Journal of Applied Science and
Technology Trends, vol. 2, pp. 20-28, 2021.
[16] Z. Zhang, “Introduction to machine learning: k-nearest neighbours,”
Annals of translational medicine, vol. 4, pp. 345- 356, 2016.
[17] M. Almaliki, C. Ncube and R. Ali, “The Design of adaptive acquisition
of users feedback: an empirical study," 2014 IEEE Eighth International
Conference on Research Challenges in Information Science (RCIS),
Marrakech, Morocco, pp. 1-12, 2014.
[18] N. Sherief, W. Abdelmoez, K. Phalp and R. Ali, “Modelling users‟
feedback in crowd-based requirements engineering: an empirical study,”
In IFIP Working Conference on The Practice of Enterprise Modelling,
pp. 174-190. Springer, Cham, 2015.
[19] S. Francesco, Lorenzo, B. Cristian, L. Danza, M. Ghellere et al.,
“Integrated method for personal thermal comfort assessment and
optimization through users‟ feedback, IoT and machine learning: A case
study,” Sensors, vol. 18, no. 5, pp. 1602, 2018.
[20] S. Manikandan, A. Delphincarolinarani, C. Rajeswari, T. Suma and D.
Sivabalaselvamani, “Recognition of font and tamil letter in images using
deep learning,” Applied Computer Science, vol. 17, no. 2, pp. 90–99,
2021.
[21] P. N. Tan, M. Steinbach and V. Kumar, “Introduction to data mining,” in
The Pearson. Upper Saddle River, NJ, USA: Pearson,vol. 54, pp. 37–69,
2006.
[22] A. Saurav, “Amazon Review Data (2018),” Amazon review, 06 March,
2022, Available: https://nijianmo.github.io/amazon/index.html.
[23] K. McCartan, A. Danielle, F. Harris and S. David, “Seen and not heard:
The service user‟s experience through the justice system of individuals
convicted of sexual offenses,” International journal of offender therapy
and comparative criminology, vol. 65, no. 12, pp. 1299-1315, 2021.
[24] D. Gavilan and G Martinez-Navarro, “Exploring user‟s experience of
push notifications: a grounded theory approach. Qualitative Market
Research,” An International Journal, vol. 4, 2022.
[25] P. Sajjadi, L. Hoffmann, P Cimiano, and S. Kopp, “A personality-based
emotional model for embodied conversational agents: Effects on
perceived social presence and game experience of users,” Entertainment
Computing, vol. 32, pp.100313, 2019.
[26] B. Sovacool, J. Osborn, M. Martiskainen and M. Lipson, “Testing
smarter control and feedback with users: Time, temperature and space in
household heating preferences and practices in a Living Laboratory,”
Global Environmental Change, vol. 65, pp.102185, 2021.
[27] W. Wanyuan, Y. Jiang and W. Wu, “Multiagent-based resource
allocation for energy minimization in cloud computing systems,” IEEE
Transactions on Systems Man and Cybernetics: Systems, vol. 47, no. 2,
pp. 205–220, 2016.
[28] S. Kumar and M. Singh, “Big data analytics for healthcare industry:
Impact, applications, and tools,” Big Data Mining and Analytics, vol. 2,
no. 1, pp. 1–20, 2019.
[29] K. Sahu, F. A. Alzahrani, R. K. Srivastava and R. Kumar, “Evaluating
the impact of prediction techniques: Software reliability
perspective,” computers,” Materials & Continua, vol. 67, no. 2, pp.
1471–1488, 2021.
[30] J. Athinarayanan, V. S. Periasamy, M. Alhazmi, K. A. Alatiah and A. A.
Alshatwi, “Synthesis of biogenic silica nanoparticles from rice husks for
biomedical applications,” Ceramics International, vol. 41, no. 1, pp.
275–281, 2015.
[31] S. O. Proksch, W. Lowe, J. Wäckerle and S. N. Soroka, “Multilingual
sentiment analysis: A new approach to measuring conflict in legislative
speeches,” Legislative Studies Quarterly, vol. 44, no. 1, pp. 97–131,
2019.