Enhancing E-commerce Insights: Sentiment

Analysis Using Machine Learning and

Ensemble Techniques
Shubham Shedekar Sahil K. Shah* Vidya Kumbhar
Symbiosis Institute of Symbiosis Institute of Symbiosis Institute of
Geoinformatics Geoinformatics Geoinformatics
Symbiosis International Symbiosis International Symbiosis International
(Deemed University) (Deemed University) (Deemed University)
Pune, India Pune, India Pune, India
[email protected] [email protected] [email protected]

Abstract- This research aims to explore the impact of reviews posted by other customers to make their
customer reviews on consumer decisions in the e- purchase decisions.[3][1] Therefore, e-commerce
commerce era. Machine learning techniques, businesses must monitor and analyze customer
including Naive Bayes, Random Forest, Decision
feedback to improve their products or services and
Tree, Extra Trees, and Logistic Regression, are
utilized for sentiment analysis of customer reviews. A
build brand loyalty. One of the best ways to analyze
comprehensive dataset is collected through web customer feedback is through sentiment analysis.
scraping and subjected to exploratory data analysis [3] Sentiment analysis employs a combination of
(EDA) to gain a deeper understanding of its natural language processing (NLP) and machine
characteristics. Feature extraction techniques are learning techniques to evaluate the sentiment
applied to convert raw text data into meaningful expressed in customer reviews shared on e-
numerical representations. The ensemble learning commerce platforms.[4] This analysis aims to
approach, specifically the voting ensemble method, is extract insights from customer feedback, identify
employed to combine individual model predictions,
customer sentiment toward products or services,
enhancing overall performance and robustness. The
findings contribute valuable insights into customer and identify areas for improvement [5]. This article
sentiment, empowering businesses to understand will discuss the benefits of sentiment analysis of
preferences and enhance their offerings. customer reviews of e-commerce and the steps
involved in performing this analysis. Sentiment
Keywords - Machine Learning; TF-IDF; SVM; Naïve analysis allows e-commerce businesses to
Bayes; Decision Tree; Logistic Regression; Ensemble understand customer sentiment toward their
Models; Sentiment Analysis; Natural Language products or services. By analyzing customer
Processing reviews, businesses can identify positive and
I. INTRODUCTION negative feedback and take action to address areas
of concern. By analyzing customer feedback, e-
E-commerce, also recognized as electronic commerce businesses can identify areas for
commerce, pertains to the online transactional improvement in their products or services.[5] This
exchange of goods and services.[1] It has brought feedback can be used to enhance product features,
about a transformative shift in the operational improve customer service, and provide a better
landscape of businesses and consumer shops by overall customer experience. By addressing areas
providing a convenient and accessible platform for of concern highlighted in customer reviews, e-
transactions. Online shopping, a major component commerce businesses can enhance customer
of e-commerce, allows customers to browse and satisfaction. Satisfied customers are inclined to
purchase products or services through websites or share positive reviews, which in turn have the
mobile applications. One of the significant potential to draw in new customers and increase
advantages of e-commerce and online shopping is sales.[6] By monitoring and responding to
the convenience it offers.[2] Customers can shop customer feedback, e-commerce businesses can
anytime, anywhere, without being limited by build a positive brand reputation. Positive reviews
physical store hours or location. Consumers can have the potential to attract fresh customers while
conveniently compare prices, access product addressing negative reviews showcases a
reviews, and make well-informed purchasing dedication to customer service and fosters customer
choices without leaving the comfort of their own loyalty.[5] The rapid growth of e-commerce has
homes or while on the go[2]. In the era of e- presented businesses with the challenge of
commerce, online reviews have become an effectively utilizing customer feedback to improve
essential part of the decision-making process for their products or services and build customer
customers. Nowadays, customers rely heavily on

1
loyalty.[6] In the context of online reviews, pool of 425 publications from reputable journals.
businesses struggle to analyze and extract The findings indicated that sentiment analysis in
meaningful insights from the vast amount of the context of COVID-19 is predominantly tailored
unstructured data available. The problem at hand is to specific domains and applications. Ensemble
how to leverage sentiment analysis techniques to models and deep learning (DL) classifiers
understand customer sentiment toward products or demonstrated superior performance compared to
services in the e-commerce industry. By addressing single classifier models. Notably, pre-training
this problem, businesses can identify areas for BERT or RoBERTa models on Twitter data yielded
improvement, enhance the overall customer promising outcomes[4]. This paper discusses the
experience, and ultimately drive sales.[5] use of Chinese social e-commerce for hotel
Additionally, it is crucial to explore the challenges bookings and the importance of customer reviews.
specific to sentiment analysis in the e-commerce The study uses hotel review data from Ctrip and
domain, such as dealing with the unique language applies emotion analysis using a BERT model.[5]
and nuances found in customer reviews. Finding The paper introduces a novel approach that
effective strategies to overcome these challenges is combines feature extraction techniques to achieve
paramount to gaining accurate and actionable precise sentiment analysis of unprocessed online
insights from customer feedback.[7] By delving reviews. By leveraging temporal sentiment
into these problem areas, businesses can leverage analysis, researchers can gain insights into the
sentiment analysis to its fullest potential, leading to fluctuating emotional states and their
improved customer satisfaction and a competitive interconnectedness with various contextual factors,
advantage in the e-commerce market[8]. The main thereby facilitating a more comprehensive
contribution of this paper is to do a sentiment understanding of the subject matter.[6] The study
analysis of real customer reviews available on e- suggests that developers should focus on improving
commerce sites. Classify the polarity of the usability and affection levels in non-gamified
sentiment in reviews, use machine learning models applications and consider users' emotional values to
to classify the reviews, and compare the results of enhance their satisfaction. However, the study has
machine learning algorithms. some limitations, such as the lack of user
demographic information and the generalizability
II. LITERATURE REVIEW of results to other app stores, and future research
To accurately perform the sentiment analysis of the can address these limitations using more advanced
raw reviews this paper uses the feature extraction natural language processing techniques. [7] The
methods. Results show that a good approach study also finds that consumers are more strict
greatly improves sentiment analysis performance. when assigning positive sentiment toward star
finding the performance of the models using ratings when viewing them as consumers.[8] This
datasets and learning more NLP techniques [1]. The paper proposes a method for sentiment analysis of
experimental findings demonstrate that the Twitter tweets using a stochastic gate neural
proposed algorithm outperforms existing methods network (SGNN) and a pre-trained word
in terms of performance. Moving forward, the embedding feature based on the polarity of the
researchers aim to develop innovative approaches lexicon and n-grams. The approach applies to
utilizing tags and knowledge graphs as diverse social media platforms and serves as a
supplementary information to further enhance valuable tool for researchers operating within the
recommendation accuracy. Additionally, they plan specified domain.[9] The research explores the
to introduce a tag recommendation model that connections between physician characteristics and
leverages dynamic user preferences, offering sentiment analysis scores obtained from reviews
personalized recommendations based on evolving and ratings. Additionally, it explores the influence
user interests and preferences. These future of frequently used words on the overall positivity
advancements hold promise for improving the of reviews. The results indicate to receive higher
overall effectiveness of the RS. [2] Understanding ratings, and the inclusion of pleasant personality
the limitation of the study is most important, which descriptors in reviews has a positive effect on their
primarily focused on attractions and relied solely sentiment.[10]
on data from TripAdvisor, potentially introducing
Data Collection: -
platform bias. To validate the findings, future
research should explore information and services. The data collection process involved gathering data
[3] The study encompassed a comprehensive on 13 different smart TVs. The dataset consists of
literature review aimed at identifying and 24,000 rows and was obtained through web
classifying existing studies that employed scraping techniques.[10] To begin the data
sentiment analysis on COVID-19 Twitter data, collection process, a set of 13 specific smart TVs
particularly focusing on machine learning (ML) was selected for analysis. These TVs were chosen
techniques. Through a multistage screening based on various criteria such as their popularity,
process, 40 papers were selected out of an initial customer ratings, or specific brands and models of

2
interest. Web scraping, a method of automatically iterated through multiple pages of reviews for each
extracting data from websites, was employed to smart TV. This involved navigating to subsequent
collect the necessary information.[10] pages and systematically extracting the reviews
until all available data was collected. The resulting
III. METHODOLOGY dataset comprises 24,000 rows, indicating a
substantial volume of customer reviews for the 13
selected smart TVs. This large dataset can provide
a robust foundation for sentiment analysis,
allowing for in-depth insights into customer
sentiments toward the different smart TV models.
Throughout the web scraping process, it is essential
to consider ethical considerations and adhere to
legal requirements. Compliance with the terms and
conditions of the websites being scraped is crucial.
[10] Additionally, it is important to be mindful of
any rate limits or restrictions imposed by the
websites to avoid overloading their servers or
violating their policies. In summary, the data
collection phase of the paper involved
systematically scraping customer reviews for 13
smart TVs from e-commerce websites. The process
automated the extraction of relevant information
from multiple pages, resulting in a substantial
dataset of 24,000 rows suitable for sentiment
analysis.
Data Preprocessing: -
After collecting the data preprocessing is an
important part of data analysis in preparing data for
further analysis.
• Removing duplicates: It's important to check for
and eliminate any duplicate rows or reviews within
the dataset. Duplicates can skew the analysis and
introduce bias. By comparing the content of each
review, you can identify and remove duplicate
entries.
• Handling missing values: Missing data occurs
when fields are not filled out or captured during
web scraping. Proper handling involves removing
rows, imputing values using statistical measures, or
Fig. 1 Methodology using advanced imputation techniques to avoid bias
or impact on analysis.
The web scraping script or tool accessed relevant e-
commerce websites that sell smart TVs. It Review text cleaning: -
navigated to the product pages of the selected 13 • Removing HTML tags: When scraping data
models and extracted specific details required for from websites, the collected text data often contains
sentiment analysis, with a particular focus on HTML tags used for formatting and styling.
customer reviews.[10] The extracted information Removing these tags is essential to focus solely on
encompassed various elements such as product the textual content of the reviews.[9] This can be
names, descriptions, prices, customer ratings, done using regular expressions or HTML parsing
reviews, and any other pertinent data that could libraries.
contribute to sentiment analysis. During the web
scraping process, the script or tool systematically • Removing special characters: Special characters
accessed each smart TV's web page. It located the such as punctuation marks, symbols, or non-
customer review section on the page and retrieved alphanumeric characters can add noise to the text
relevant content, including the review text, data and may not contribute significantly to
reviewer's name, rating, and any other associated sentiment analysis.[1] Removing these special
information. To ensure comprehensive data characters helps in reducing noise and improving
coverage, the web scraping script or tool likely the accuracy of subsequent analyses[1].

3
• Lowercasing the words: Text data often contains reduced to "good." By performing lemmatization,
words in various cases (uppercase, lowercase, or we can improve the accuracy and effectiveness of
mixed). Converting all the words to lowercase various natural language processing tasks such as
standardizes the text, ensuring that the same word text classification, sentiment analysis, and
in different cases is treated as the same word. This information retrieval
process helps in avoiding duplication and improves
• Polarity: The emotional tone or mood
the efficiency of subsequent analysis.
communicated in a text is referred to as its polarity,
• Remove Stop words: Stop words are common indicating whether it is positive, negative, or
words with no significant meaning or contribution neutral [2]. Polarity is often represented on a
to sentiment analysis.[6] To reduce the numerical scale, such as -1 to +1, to indicate the
dimensionality of the data we can remove the stop emotional sentiment associated with a text or data.
words that are in our dataset so that the model can In this scale, a value of -1 signifies a highly
move precisely to make prediction. Some stop negative emotion, +1 represents a strongly positive
words are ‘is’, ‘and’, ‘the’ like this stop words we emotion, and 0 indicates a neutral emotion. [10] By
can remove. [6][10]. examining the words and phrases used and taking
into consideration the sentiment they are linked
• Stemming words: In the steaming words process
within a predetermined sentiment lexicon, or by
convert the words to their base forms. For example,
using machine learning algorithms trained on
stemming the word "running," "runs," and "ran"
labeled data, it is possible to compute the polarity
would result in the base form "run." Stemming
score of a text. [4] [10]
helps in consolidating words with similar
meanings, reducing vocabulary size, and improving • After calculating the Polarity, we classify the
the accuracy of sentiment analysis [6]. reviews as positive negative, and neutral. Based on
that we can understand the actual feeling in the
Sentiment Analysis: -
review text.
Sentiment analysis, a natural language processing
technique, detects emotions expressed in text, also
known as opinion mining. It helps organizations
analyze and categorize sentiments related to
products, services, or concepts. Sentiment analysis
classifies opinions to determine a writer's attitude
towards a subject, product, or entity, determining
whether it is positive, negative, or neutral. Tools for
sentiment analysis are crucial for identifying and
comprehending emotions. These tools can help
businesses and organizations better understand how
people feel and boost productivity. Utilizing
sentiment analysis techniques, businesses can learn
how to increase customer satisfaction and
experience. Fig.2 Word cloud of positive reviews
• Tokenisation: Tokenization in sentiment analysis The above diagram Is of word cloud of the positive
is the process of breaking down a piece of text, reviews posted on the e-commerce platform. From
such as a customer review, into individual units this word cloud, we can observe that the peoples’
called tokens. These tokens can be words, phrases, are satisfied with the TV's picture quality and the
or even individual characters.[6] The first and sound quality of the TVs. By this word cloud, we
important step in Natural language processing is understand that people are happy about the prices
Tokenization, as Tokenisation makes the data into of TVs also people are happy about the speaker,
small small tokens so that the machine learning smooth performance, display, and so on.
model can understand it easily.
• Lemmatization: The lemmatization technique is
used. It is a text processing technique that reduces
words to their base or dictionary form, known as
the lemma. It aims to normalize words by
considering their morphological variants. Unlike
stemming, which simply chops off word endings,
lemmatization takes into account the word's context
and grammar to produce meaningful lemmas.[6]
For example, lemmatizing the word "running"
would result in "run," and "better" would be

4
 2. Logistic regression(LR): - LR is a statistical
method that is widely used to solve the
classification problem by predicting the probability
of an outcome falling into one of two categories [2]
[3]. The output is derived by applying a logistic
function to a linear combination of input variables,
enabling the transformation of the inputs into a
probability-based outcome.[3] This model is
trained using maximum likelihood and based on
that tries to find the best-fit line.
3. K-nearest neighbors (KNN) Classifier is a
versatile machine learning algorithm capable of
handling both classification and regression tasks.
Fig. 3 Word cloud of negative reviews [11] It operates as a non-parametric approach,
relying on the similarity between input data points
The above word cloud diagram is of the negative and their neighboring data points to make accurate
reviews posted on the e-commerce platform. By predictions. The KNN is identified based on the
this word cloud, we can understand that the users minimum distances and assigned the class label
are also not satisfied with the products. Some of the that is most prevalent among its K nearest
consumers have problems related to customer neighbors.[12] The choice of K is an important
service some of the users have problems related to parameter in the algorithm. It is suitable for online
the price. Does not satisfied with the customer care learning scenarios and can capture complex
service, delivery, screen display, picture quality, decision boundaries.
and so on.
4. Decision Tree (DT): - The machine learning
algorithm called Decision Tree is generally used for
classification tasks basically Decision Tree from a
tree-like structure it makes nodes and sub nodes.
[11] Decision trees are known for their
interpretability Also decision Tree can be able to
hand both numerical data and also categorical data.
[12] However, a potential drawback of decision
trees is their tendency to overfit the training data.
To mitigate this issue, methods like pruning,
ensemble learning, and regularization can be
employed to enhance the generalization and
performance of decision tree models [12].
5. Extra tree (ET): - The Extra Trees Classifier is
Fig. 4 Word cloud of neutral reviews an ensemble machine learning algorithm that works
by constructing multiple decision trees and
The above word cloud is of the neutral review selecting random subsets of features for each split.
posted on the e-commerce platform. By this word This randomization reduces the variance of the
cloud, we can understand that the reviews have model, making it less prone to overfitting.[12]
words like product, quality, money, TV, amazon, During the prediction phase, the class label with the
service, install replace, and many more. By this majority vote from the decision trees is selected as
word cloud, we can understand that people are the predicted class. The key advantages of the Extra
happy about the product, product quality, and Trees Classifier include its ability to handle high-
customer care. dimensional datasets, its resistance to overfitting,
Models Used: - and its efficiency in terms of training time.[12]
However, the randomization may lead to a slight
1. Multinomial Naïve Bayes (MNB):- decrease in predictive accuracy and reduced
Multinomial Naive Bayes is one of the most well- interpretability.
liked supervised learning classifiers used to analyze
categorical text data. The method predicts the tag or 6. Random Forest (RF): - Random Forest is a
category of a text by applying the principles of combination of different decision trees.[11] It can
Bayes theorem [2]. It calculates the probabilities of be used for classification as well as for regression.
each class for a given sample and assigns the class Random Forest is most suitable for high-dimension
with the highest probability as the output noisy data. Random Forest Algorithm works with
prediction. the majority votes in classifying classes using
different decision trees.[12]

5
Ensemble model (voting ensemble): - predictions by accurately identifying positive
sentiments and avoiding misclassifications as
To get the accurate result of the models we have
negative or neutral.
used ensemble learning methods. Ensemble
learning methods are used to get more accurate Recall = TP/(TP+FN) [5]
results than the traditional models.[11] The voting
ensemble is a popular ensemble learning method it 4. F1 Score: The harmonic mean of recall and
gives the output by combining the results of all precision, and also provides the balance of model
models used it aggregates the results to make the performance. F1 score consists of precision as well
predictions. The aim behind the voting ensemble is as recall to make the predictions. The F1 score
to combine models’ predictions and aggregate them considers the trade-off between precision and recall
to get more accurate results and predictions. This and provides a single value that represents the
method has three types majority voting, weighted model's overall performance. [2] The F1 score
voting, and soft voting. In this study, we have used proves valuable in scenarios where there is an
the soft voting technique. Soft voting sums up the imbalanced class distribution or when the
probabilities of each class predicted by models and significance of false positives and false negatives is
selects the highest summed probability. equal[9].
F1score=(2×Precision×Recall)/(Precision + Recall)
Evaluation parameter: - [5]
1. Accuracy: The accuracy is used to understand IV. RESULTS AND DISCUSSION
the overall correctness of the machine learning
model. The calculation entails evaluating the Data Balancing: - As we can see our data is
precision of sentiment predictions (positive, imbalanced. Imbalanced data can cause lower
negative, neutral) by contrasting the count of accuracy of the machine learning models or can be
accurately predicted sentiments with the total performed overfit to the data. so firstly we will
number of reviews present in the dataset.[9] balance our data. for the balancing, we have used
Accuracy provides a general overview of the resample techniques. There are different types of
model's performance, indicating how well it resampling techniques present we have used
predicts sentiment across all classes.[2] However, random over-sampling technique. in this technique
accuracy may not be the best metric when dealing we have the minority class distribution and
with imbalanced datasets, where one sentiment majority class distribution so on that minority class
class dominates. Accuracy generally defines the we have randomly increased the minority class by
ratio of the correct perfection of models to the total replacing them randomly and getting enough
number of predictions. classes the same as the majority class.

Accuracy = (TP+TN)/(TP+TN+FP+FN) [5]

2. Precision: Precision is a metric that focuses on
the positive outcomes made by the model. The
precision is calculated by taking the ratio of the
instances that are positively predicted to the
instances of all positives means True positive and
false positive.[9] It is generally used to measure the
working of the model to ignore the false positive.
[2]i.e., correctly identifying positive sentiments
without misclassifying negative or neutral
sentiments as positive. A low rate of false positives
means it will give a high value.
Fig 5. Classification of Balanced Classes
Precision = TP/(TP+FP)
Model result: -
3. Recall: Recall refers to the true positive and
sensitivity, and also evaluates the model capacity to The ML models were built to classify the sentiment
correctly capture all positive instances. This recorded. The models used were (MNB), (RF),
method identifies the relation between the actual (ETC), (LR), (DTC) and (KNC). The
positive data present in the dataset with the performance matrices like precision, recall, F1
correctly predicted positive dataset.[9] By score, and accuracy were used. Out of the all
measuring recall, the model's performance in algorithms used for sentiment analysis, we can
accurately identifying all positive sentiments, understand that the random forest algorithm
without overlooking any, is assessed.[2] A high performs best. We can see that the multinomial
recall value signifies that the model effectively Naïve Bayes and k neighbor classifier before was
minimizes the occurrence of false negative not good because the shows less accuracy than

6
 other algorithms. Other algorithms like (ETC), (MNB), Decision Tree Classifier (DTC), Random
(LR), (and DTC) classifiers perform well and have Forest (RF), Logistic Regression (LR), and K-
the accuracy same as the random forest but as we Nearest Neighbor Classifier (KNC), were
can compare the confusion matrix of each model employed, along with ensemble learning using the
we can say that these models can be overfitted or voting ensemble method. Feature extraction was
can have less ability to predict true positive performed using TFIDF Vectorizer. Among the
elements.
models evaluated using various metrics such as
TABLE I: EVALUATION PARAMETERS FOR MODEL COMPARISON accuracy, precision, recall, and F1 score, the
Random Forest classifier demonstrated the highest
Model MNB RF ET LR DT KNN
performance with an accuracy of 99% and an F1
Accuracy 0.86 0.98 0.99 0.97 0.99 0.78
score of 0.99. Other models, such as Logistic
Precision 0.87 0.98 0.99 0.97 0.99 0.85
Regression (LR), Multinomial Naïve Bayes
Recall 0.86 0.98 0.99 0.97 0.99 0.78 (MNB), Decision Tree Classifier (DTC), Extra Tree
F1 Score 0.86 0.98 0.99 0.97 0.99 0.78 Classifier (ETC), and K-Nearest Neighbor
The below figure shows the evaluation parameter Classifier (KNC), achieved accuracies of 87%,
of each model which is used for the sentiment 99%, 98%, 99%, and 79%, respectively.
analysis. Here we can see that as compared to other Furthermore, the research implemented an
models K neighbors have less value in the ensemble method using the voting ensemble
evaluation matrix. The random forest has good technique, which performed well with an accuracy
values as compared to other machine learning of 99%. The ensemble method also achieved the
algorithms. Therefore, we can say that random highest precision and recall and F1-scores of 0.98,
forest performs better as compared to other
algorithms. Table III shows the classification report 0.98, and 0.98 respectively. The future of sentiment
of the voting ensemble technique. Here we can see analysis in e-commerce holds exciting possibilities,
that the accuracy of this model is 99% and the good with advancements in technology leading to more
score of precision-recall and f1 score. Here the sophisticated algorithms that better understand and
precision value for class 0,1,2 is 1, 0.98, 0.98 classify customer reviews. These advancements
respectively. will enable businesses to gain clearer insights into
customer sentiment across different languages and
1.2
platforms, fostering increased customer loyalty and
1 success in the online marketplace.
0.8
REFERENCES
0.6
[1] Kaur, G. and Sharma, A. (2023) A deep
0.4 learning-based model using hybrid feature
0.2 extraction approach for consumer sentiment
0
analysis - Journal of Big Data, SpringerOpen.
MNB RF ET LR DT KNN Available at: https://doi.org/10.1186/s40537-022-
Accuracy Precision Recall F1 Score 00680-6 (Accessed: 27 September 2023).

Fig. 6 Model Evaluation Parameters [2] Venkataraman, T.K. and Jadhav, A. (2023)
Classifying the sentiment polarity of Amazon
TABLE II: ENSEMBLE MODEL EVALUATION PARAMETER
mobile phone reviews and their ratings, AIP
Precision Recall F1-Score Publishing. Available at:
Model Accuracy
0 1 2 0 1 2 0 1 2 https://doi.org/10.1063/5.0110605 (Accessed: 27
Voting
September 2023).
Ensemble
1 0.98 0.98 0.96 1 1 0.98 0.99 0.99 0.99

[3] Aravindan, T.R., Vigneshwar, C.N. and

CONCLUSION Suganeshwari G (2023) ‘Sentiment classification
The research concludes that customer reviews play for Amazon Fine Foods reviews using Pyspark’,
a crucial role in determining customer satisfaction Advances in Transdisciplinary Engineering
and sentiment toward products. The study [Preprint]. doi:10.3233/atde221293.
specifically focused on analyzing the sentiment
polarity of Smart TV purchases using reviews from [4] Braig, N. et al. (2023) ‘Machine learning
e-commerce platforms. Machine learning techniques for sentiment analysis of COVID-19-
classification algorithms, including Extra Tree related Twitter data’, IEEE Access, 11, pp. 14778–
Classifier (ETC), Multinomial Naïve Bayes 14803. doi:10.1109/access.2023.3242234.

7
[5] Wen, Y., Liang, Y. and Zhu, X. (2023)
‘Sentiment analysis of hotel online reviews using
the Bert Model and Ernie Model—data from
China’, PLOS ONE, 18(3).
doi:10.1371/journal.pone.0275382.

[6] Rodríguez-Ibánez, M. et al. (2023) ‘A review

on sentiment analysis from social media platforms’,
Expert Systems with Applications, 223, p. 119862.
doi:10.1016/j.eswa.2023.119862.

[7] Ahn, H. and Park, E. (2023) ‘Motivations for

user satisfaction of Mobile Fitness Applications: An
analysis of user experience based on online review
comments’, Humanities and Social Sciences
Communications, 10(1). doi:10.1057/s41599-022-
01452-6.

[8] Noh, Y.-G., Jeon, J. and Hong, J.-H. (2023)

‘Understanding of customer decision-making
behaviors depending on online reviews’, Applied
Sciences, 13(6), p. 3949.
doi:10.3390/app13063949.

[9] Vidyashree, K.P. and Rajendra, A.B. (2023) ‘An

improvised sentiment analysis model on Twitter
data using stochastic gradient descent (SGD)
optimization algorithm in Stochastic Gate Neural
Network (SGNN)’, SN Computer Science, 4(2).
doi:10.1007/s42979-022-01607-x.

[10] Jo, J.J. et al. (2023) Physician review websites:

Understanding patient satisfaction with
ophthalmologists using natural language
processing, Journal of Ophthalmology. Available at:
https://doi.org/10.1155/2023/4762460 (Accessed:
27 September 2023).

[11] Jain, V. and Kashyap, K.L. (2022) ‘Ensemble

hybrid model for Hindi COVID-19 text
classification with metaheuristic optimization
algorithm’, Multimedia Tools and Applications,
82(11), pp. 16839–16859. doi:10.1007/s11042-022-
13937-2.

[12] Sharma, A. and Ghose, U. (2023) ‘Toward

machine learning based binary sentiment
classification of movie reviews for Resource
Restraint Language (rrl)—Hindi, IEEE Access, 11,
pp. 58546–58564.
doi:10.1109/access.2023.3283461.