Drug Recommendation System Based on Sentiment Analysis of Drug

Reviews Using Machine Learning

Rishitha Bontha, Sabreen Taj Bandar, Siddu Veera Venkatesh Barigela,

Susmitha Gangappagari and Nalini Priyanka Kummara

Department of Computer Science and Engineering (Data Science), Srinivasa Ramanujan Institute of Technology,

Anatapur‑515001, Andhra Pradesh, India

Keywords: Sentiment Analysis, Drug Recommendation System, Machine Learning, Patient Reviews, Vectorization,

Classification Algorithms, Precision, Recall.

Abstract: People Self-Medicate Without Physicians Advice, Making Their Conditions Worse in Some Cases While The

COVID-19 Pandemic Has Further Exposed Inadequacies in The System. To address this, this study develops

a machine learning-based drug recommendation system that uses sentiment analysis of patient reviews to

recommend drugs. The system uses vectorization methods Bag of Words (Bo W), TFIDF, Word2Vec to

convert textual drugs reviews into organized sentiment information. Classification models such as MLP then

evaluate sentiments and create drug recommendations. Results evaluated by precision, recall, F1-score,

accuracy, and AUC scores confirm that MLP classifier model out performs the rest of models in accuracy.

This model offers an inherent advantage over current systems that rely on patient demographics and risk

groups, greatly alleviating cold start issues, computational resource consumption, and information sparsity. It

comprises a set of classifiers and uses a useful count, that is, a number that measures the number of times a

particular drug has been reviewed to ensure that only the most reliable drugs are recommended to each patient.

The hybrid approach featured in our model improves predictive robustness, outperforming traditional methods

by yielding superior performance and reliability. Notably, it also fosters computational efficiency by choosing

the fastest algorithms, resulting in greatly minimized training times and enhanced prediction accuracies. This

novel framework provides a scalable, data driven method for generating automated pharmaceutical

suggestions.

1 INTRODUCTION

The COVID-19 pandemic has put a tremendous stress

on systems of global health care, leading to shortages

of healthcare workers, equipment and

pharmaceuticals. In this crisis, many people cannot

seek a timely professional consultation, and this

contributes to self-medication, which causes most

people to worsen their health status because they

select the wrong medicines or take incorrect doses

Aggarwal, C. C. (2016). Most of the existing

frameworks for drug recommendation make use of

some pre-defined set of medical details including

immunity levels, underlying risk factors as well as

comorbidities to recommend appropriate

medications. These models, however, often do not

consider the wide variability in individual responses

to drugs, resulting in suboptimal recommendations.

Furthermore, traditional systems are computational

ly inefficient, requiring massive analysis of risk

factors which drives up processing costs and slow the

decision-making process (Bermingham, A., &

Smeaton, A. F. (2010)). The cold start problem in

these models is a crucial limitation, with the absence

of adequate historical data for emerging drugs or

conditions causing unreliable recommendations

(Bermingham, A., & Smeaton, A. F. (2010)).

Moreover, standard systems face information

sparsity, where the partial embodiment of patients

makes predictions difficult (Borth, D, et.al., 2013)

Overcoming these drawbacks demands a flexible and

data-informed methodology linking data from actual

patient experiences to the recommender framework

admitting this patient-based information into the

recommendation routine, yielding enhanced

precision, robustness, and scalability. One way to

solve this problem is to use machine learning on

sentiment analysis to extract meaningful insights

516

Bontha, R., Bandar, S. T., Barigela, S. V. V., Gangappagari, S. and Kummara, N. P.

Drug Recommendation System Based on Sentiment Analysis of Drug Reviews Using Machine Learning.

DOI: 10.5220/0013901000004919

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Research and Development in Information, Communication, and Computing Technologies (ICRDICCT‘25 2025) - Volume 3, pages

516-523

ISBN: 978-989-758-777-1

from patient-generated drug reviews to enable more

informed and personalized pharmaceutical

recommendations (Brusilovsky, et.al., 2007).

The proposed framework makes use of sentiment

analysis techniques that can analyse unlabelled drug

reviews by transforming unstructured feedback into

numerical data that can be used directly to train

machine learning classifiers (Cambria, et.al., 2009).

Bag of Words (Bo W), term frequency-inverse

document frequency (TF-IDF), and Word2Vec are

advanced vectorization methods used in patient

reviews to convert text into a structured numerical

format for automated analysis (Cambria, E., et.al.,

2009). The work then uses various classification

models (SVM, Logistic Regression, Naïve Bayes,

etc) to classify if a drug is perceived positively or

negatively from patient sentiment (Dey, L., &

Haque, S. K. M. (2009)). They are trained on a set of

labeled data for drug reviews with sentiment labels

and can classify new reviews at high accuracy (Feng,

et.al., 2019). To ensure robustness and reliability, the

system evaluates model performance using

precision, recall, F1-score, accuracy, and AUC score

(Garg, S. (2021)). Out of all tested classifiers, Linear

SVC with TF-IDF vectorization performed best with

the highest accuracy, and beating other models in

predicting sentiment (Hu, M., & Liu, B. (2004)). In

contrast to traditional methods that depend on static

and predeterminate clinical criteria, this strategy can

adapt recommend drug treatment through the ongoing

updates of real world data, the system is able to

advance quickly and be applicable for widespread

diseases and new drugs (Jakob, N., & Gurevych, I.

(2010)).

Based on this ensemble method, the system can

achieve a better accuracy of recommendation, since it

integrates the prediction results from a series of

classifiers to avoid biases of each single model and

enhances reliability as a whole (Lamba, M., &

Madhusudhan, M. (2019).). To determine the final

sentiment predictions, they are summed up and

weighted using "useful count" metric which estimates

the trustfulness of each drug review by counting the

number of times it was cited by users in (Lei, et.al.,

2008). This in turn allows drugs that would

potentially require more engagement from a user to

have an increased impact on recommendation,

making the model more in-line with what individuals

experience in the real-world (Liu, et.al., 2008).

Especially, this hybridization addresses misusage of

traditional systems, for instance, computational

inefficiencies and sparsity of data by capitalizing on

the shared intelligence brought by extensive patient

reviews rather than solely relying on defined risk

factors (Morency, et.al., 2011). Moreover, by

drastically cutting the number of classifiers that need

to be trained, this approach leads to lower

computational expenses, since the fewer cadidates

have very similar predictions maintaining high

accuracy, which allows the system to respond in real-

time (Nguyen, et.al., 2024). The system uses patient-

derived information to supplement any

recommendation where it be drug effectiveness

derived from both the clinical community and patient-

driven conversational knowledge websites,

presenting a broader scope of understanding about

drug effectiveness in the context of both the clinical

community and patient derived solution driven

processes (Pang, B., & Lee, L. (2008)).

Their results show that the drug recommendation,

based on sentiment analysis, showing considerably

useful as its more accuracy and scalability by

claiming that it overcomes limitations of traditional

models such as cold start and incomplete records

(Poria, S., et.al., 2017). Best practices were

developed that involved the integration of multiple

machine learning classifiers as a richer decision-

making framework ensuring that drug

recommendations were grounded in data, rather than

limited by accompanied medical protocols. In

addition, the optimized computational efficiency of

the proposed system paves the way towards rapid

decision-making, making it a robust lead for real-time

pharmaceutical counsels as well. This allows to

reduce the burden on healthcare workers but at the

same time offers patients trustworthy, evidence-based

medication proposals, established via a plurality of

user events. Forthcoming enhancements may also

consider deep learning techniques to potentially

improve sentiment classification accuracy and

implement context-aware recommendations for

complex medical conditions (Poria, S., et.al.,).

Integrating such advancements can provide pathways

towards improved predictive performance and a

patient-centric method for automated drug

recommendation.

2 RELATED WORKS

In recent years, there has been extensive research in

the field of drug-recommendation systems utilizing

machine learning alongside many studies indicating

improvement in their accuracy and efficiency.

Conventional drug recommendation methods have

suggested drugs for patients based on various medical

data, risk factors, and manual rules. Nevertheless,

such systems are often limited with accuracy, high

Drug Recommendation System Based on Sentiment Analysis of Drug Reviews Using Machine Learning

517

computational cost and low personalization. So, how

to fix them such that we widen the chances of drug

recommendation reliability WEAN So, the

researchers have compared these reviews of patient

and they give their trust (sentiment analysis) which

helps to resolve these issues. A notable study by

(Zhang et al.) developed a drug recommendation

system based on sentiment analysis of patient

reviews to predict the effectiveness of medicines.

Using multiple methods, such as text mining and

natural language processing (NLP), for sentiment

analysis, the authors examined reviews and

identified their sentiment as either positive or

negative. Yet the low performance of multi-data

sources in some fields indicates that not every aspect

of the sentiment analysis is effective; therefore, when

a researcher wants to do sentiment analysis in

combination with drug recommendation or

recommendation systems, he/she should use

combination data sources and careful merger of the

new merged dataset. In the same vein, a study

conducted by Xie et al. (2019) discussed the

application of machine learning algorithms in drug

recommendation systems. A variety of classification

algorithms were tested, including Naive Bayes, SVM,

and Decision Trees, to predict the sentiment of

patient reviews. They found that the SVM algorithm

had the best accuracy and precision and is thus a

promising algorithm for drug recommendation

system. In a separate study, Lee et al. (2020)

conducted a study which created a drug

recommendation system that used a combination of

patient reviews in conjunction with medical data to

recommend drugs. The authors implemented a

hybrid approach that utilized both sentiment analysis

and clinical data to increase overall accuracy of the

system. These results indicated that hybrid systems

that integrate multiple data sources may boost the

performance of drug recommendation systems.

Vectorization techniques have also been

extensively researched regarding in drug

recommendation systems. Methods like BoW, TF-

IDF, and Word2Vec turn textual data into numerical

features to be utilized by machine learning models.

A study by Tang et al. (2021) cycled through several

vectorization methods within sentiment analysis

tasks. 6. Bag of Words (BoW)BoW relies on the

frequency of words in a document without

considering the order of those words or their

semantic meaning. Additionally, the cold start

problem is a well-studied research line in drug

recommendation systems. Other solutions have also

been proposed in many studies to handle this issue,

including hybrid models combining collaborative

filtering and content-based approaches. When new

drugs are available, there is often little feedback from

patients, but these models can still help to make

recommendations. In summary, the findings suggest

that drug recommendation systems can greatly

benefit from developments in sentiment analysis,

machine learning algorithms and vectorization

techniques. The proposed system in this paper

combines these methods to make drug

recommendations that are not only precise but also

quick and personalized.

3 METHODOLOGY

3.1 Theoretical Structure

Building upon STON-recommendation, this study

proposes a comprehensive approach that considers

more sophisticated data collection and processing

techniques to create a drug recommendation model

based on patient reviews employing sentiment

analysis. The method combines cutting-edge machine

learning algorithms to improve forecast accuracy

and reduce computational burden. The first step

consists of data collection and pre-process. The data

is extracted into a wholesome dataset from trusted

drug review sources. Data cleaning is a necessity for

removing inconsistencies, redundancies and

anomalies that can affect model quality. These

include how to handle missing values, remove

duplicate entries, and standardise textual components

so that they all match. In addition, second, one

approach to preparing textual data to be vectorised is

to process textual data using natural language

processing techniques for example tokenisation,

lemmatisation, and stop-word. Various vectorisation

methods such as Bag-of-Words, Term Frequency-

Inverse Document Frequency and Word2Vec are

used to perform feature extraction These approaches

allow for a mapping of textual information into

numerical formats, allowing for the training of

machine learning algorithms. We need to empirically

assess which vectorisation technique will provide the

most feasible result.

Various machine learning classification

algorithms are applied to build a quality classification

model. Some of them are Support Vector Classifier,

Random Forest, Logistic Regression, Naïve Bayes,

to name a few. The vectorised feature set is used to

train each classifier, enabling the model to understand

complex relationships in the data. It is then tuned for

hyperparameters to improve performance and

predictive power. Each model's prediction ability is

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

518

carefully evaluated by standard measures. Metrics

like precision, recall, F1-score, accuracy and Area

Under Curve score are used to quantify model

performance. These metrics provide holistic

information about the classification performance that

allows us to determine which of the models is the

most capable. Empirical results suggest that the

Linear Support Vector Classifier with Term

Frequency-Inverse Document Frequency

vectorisation outperforms other methodologies

yielding both higher accuracy and reliability. The

top-performing sentiment predictions are aggregated

to strengthen the recommendation mechanism.

Amalgamation: In this stage, we compute

amalgamation process that discards the least accurate

classifications and integrates the remaining ones into

a single prediction. Such a way, ensures that the

system for Million Developers does not only depend

on help from a single model but rather on the aid of

multiple points of view to improve the reliability of

that information. We refine the final recommendation

process using a normalisation technique that

includes a useful count metric. This number

represents how helpful users found a review, and is a

proxy for its credibility. An overall score is generated

for each drug related to a specific medical condition

by applying the action sentiment prediction to a

normalised useful count. By relying on this scoring

mechanism, the most effective drugs appear at the

top for the users making it a trusted recommendation.

Several considerations underlie the

methodological framework of the proposed system.

The first, computational efficiency through using

classification algorithms with optimised training

times. This avoids the challenge of computationally

heavy deep learning models while maintaining high

accuracy. This helps mitigate the limitations of any

individual vectorisation method, providing a holistic

approach to evaluating textual characteristics.

Thirdly, incorporating a credibility metric in the

ranking process improves the reliability of

recommendations and avoids fuzzy results due to

attractive-looking false reviews. Our system

improves upon previous approaches by incorporating

the most recent advances, as well as overcoming key

drawbacks of classical drug recommender systems.

Most of existing systems depend on patient

physiological parameters only םממ and failing to

utilize sentiment-based results generated from

reviews given by the end users. On the other hand, the

proposed framework combines sentiment analysis

with quantitative scoring, thereby providing a more

rounded drug recommendation model. ` Additionally,

leveraging machine learning classifiers offers greater

stability in predictions, addressing the cold start

problem and data sparsity challenges commonly

associated with traditional models. Implementing this

methodology requires a well- organised computatio-

nal workflow. A typical first step is data preparation

and pre-processing, where remaining raw textual

data must be structured for analysis preferably in a

usable format (until October 2023).

Figure 1: Schematic flow of theoretical structure.

The feature extraction, model training,

performance evaluation, and the eventual

recommendation follow at the further stages of the

recommendation process. Every step has been

carefully crafted to maximise the precision, speed,

and trustworthiness of our results. The practical

implications of this are significant. This system

alleviates the burden on medical practitioners by

automating drug recommendations, leading to more

effective allocation of resources. Patient have

informed medication selection, as they can rely on

accumulated user experiences for drug treatment and

make the effective dose. Additionally, the machine's

convenience makes it applicable for various diseases

across the medical field. While the system is very

beneficial, its performance depends on the entire

dataset and its quality. Sentiment prediction could be

hindered by biases in the user reviews, which would

require more work on bias mitigation approaches.

Drug Recommendation System Based on Sentiment Analysis of Drug Reviews Using Machine Learning

519

Moreover, further work can address the incorporation

of deep learning architectures to enhance

classification performance. The current

methodology, however, is a major step forward in

drug recommendation systems and caters to modern

healthcare needs.

4 RESULTS AND EVALUATION

Figure 1: Drug recommendation system GUI.

In the Figure 1 screen, the app UI shows an “Upload

Drug Review Dataset” button for users to select and

upload an existing drug review dataset. Users are

provided with a button that, when clicked, brings up

a window to select a file from their local storage to

load the DRUG dataset. After selecting the dataset, it

is loaded by clicking the ‘Open’ button. It is an

important step as it makes sure that the application is

reading the dataset correctly before plotting or other

analysis. This system processes the uploaded file and

builds the interface for data visualization and pre-

processing of drug reviews data. This graph

represents the distribution of ratings within the

dataset uploaded. Drug ratings are plotted on the x-

axis and the total number of reviews for each

respective rating is represented on the y-axis. It gives

us an idea of what the distribution of sentiments

looks like in the data set. You can see a very clear

pattern of which ratings are-by far-the most popular.

After analysing, Now the user can close the graph

and click ‘Read & Pre-process Dataset’ to proceed. It

starts the clean, getting rid of stop words and special

symbols, and converting the data appropriate for

processing. Figure 2 shows the Top 20 drug names

bar graph.

Figure 2: Top 20 drug names bar graph.

After preprocessing, all unnecessary stop words and

special characters are removed from the reviews,

ensuring a clean dataset. The above graph visually

represents the Top 20 medicines appearing most

frequently within the dataset. On the x-axis, drug

names are displayed, and on the y-axis, the

corresponding count of each drug in the dataset is

plotted. This visualization provides insights into the

most commonly mentioned medications. After

reviewing this information, the graph can be closed,

and the user should click on the ‘TF-IDF Features

Extraction’ button to convert the cleaned reviews

into numerical representations for machine learning

processing.

Figure 3: TF-IDF feature matrix display.

In the Figure 3 graph, drug reviews have been

transformed into TF-IDF (Term Frequency-Inverse

Document Frequency) vectors. The first row of the

dataset represents review words, while the

remaining columns display the computed frequency

values for each word in different reviews. If a word

does not appear in a review, the corresponding

column value remains zero. This process converts

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

520

textual data into numerical vectors, making it

suitable for machine learning algorithms. The next

step involves scrolling down the interface to

examine non-zero frequency values, ensuring that

significant words retain meaningful numerical

representations for predictive analysis. All

reviews converted to TF- IDF vector where first row

represents review WORDS and remaining columns

will contains that word average frequency and if

word not appear in review then 0 will put. Now scroll

down above screen to view some non-zero

frequency values you can see some columns

contains non- zero average frequency values and now

TF-IDF vector is ready and now click on ‘Train

Machine Learning Algorithm’ button to train all

algorithm and get below output.

Figure 4: ML model performance metrics.

In Figure 4 shows screen for each algorithm we

calculate accuracy, precision, recall and F1-SCORE

and in all algorithm MLP - Multilayer Perceptron

Classifier has got high performance and now click

on ‘Comparison Graph’ button to get below graph.

Figure: 5 ML algorithms performance comparison.

Figure 5 shows the ML algorithms performance

comparison. In the above graph, the x-axis

represents different machine learning algorithms,

while the y-axis displays their corresponding

accuracy, precision, recall, and F1-score. Each

metric is depicted in a distinct color, allowing for

clear differentiation between the performance

measures of various models. As shown, MLP

achieves the highest performance across all

evaluation metrics. This confirms its superiority in

predicting drug recommendations based on

sentiment analysis. After analyzing the graph, the

user can close the visualization and proceed by

clicking the ‘Recommend Drug from Test Data’

button, which enables drug prediction based on new

test input data.

Figure 5: Predicted drug for disease output.

Drug Recommendation and Ratings Prediction In the

above screen, the system predicts and displays

recommended drugs along with their corresponding

ratings for each disease. Figure 6 shows the predicted drug

for disease output. Based on sentiment analysis of past

reviews, the application suggests the most suitable

medications for specific medical conditions. The

output showcases disease names alongside their top-

ranked drug recommendations, ensuring that patients

receive the most relevant pharmaceutical options.

These recommendations are derived from the trained

machine learning model, which assesses sentiment

polarity and assigns ratings accordingly. This

functionality significantly enhances decision-making

in pharmaceutical selection, supporting both medical

practitioners and patients in choosing the most

effective treatment options.

5 DISCUSSION

These results validate the hypothesis that integrating

positive and negative sentiments of patient reviews

improves drug recommendation quality. As can be

observed, MLP (Multilayer Perceptron Classifier)

gave the best results as compared to other classifiers

in terms of accuracy, precision, recall and F1-score.

Drug Recommendation System Based on Sentiment Analysis of Drug Reviews Using Machine Learning

521

While other classifiers (like Support Vector

Classifier, Random Forest and Naïve Bayes)

performed fairly well, their limitations in capturing

complex relationships in sentiment-based data were

mitigated with MLP’s performance.

The research notes that conventional drug

recommendation systems mainly focus on biological

metrics and specialists, thus insufficient in addressing

patient-reported outcomes. And user-generated

sentiments complement the medical prescriptions

well, as it is more holistic and patient-centric

approach for selecting the right drugs. The results also

point out that certain machine learning models have

difficulty in capturing nuanced sentiment

expressions, particularly in ambiguous, sarcastic, or

context-dependent reviews.

While the model has shown some very promising

results, there remain some challenges to be dealt

with. User review biases like fake reviews or

misleading feedback can affect sentiment

classification. Moreover, the reliance on available

data and quality will determine its effectiveness in

scenarios where reviews are meagre. The model

selection process was guided by computational

efficiency as well, making sure that the

recommendation system is scalable and practical for

real-world applications.

Traditional drug recommendation systems are still

in use today, yet this research highlights the

importance of sentiment-driven AI models could

change how healthcare decisions are made.

Furthermore, as digital health platforms are

developing continuously, there is a high demand for

supplementary utilization of sophisticated NLP

methods and deep learning settings to enhance the

potential of sentiment classification and cross-

domain adaptation. Potential extensions may involve

the integration of multilingual review analyses, online

learning systems, and expert-level validation

processes that add layers of credibility to the

generated recommendations.

Results indicate that traditional recommendation

frameworks are not yet entirely superseded, but will

need significant adaptation if they are to compete

amidst AI-facilitated decision support systems in the

future. Healthcare professionals and pharmaceutical

companies must count use of sentiment analysis to

improve satisfaction and efficiency of the treatment.

If this is not the case, recommendation methods may

become stale and suppress the possibility of

personalising to new patient needs. As you can see

from this context, sentiment-based drug

recommendations further supports the overall trend

of AI-driven personalization in healthcare, by

transforming patient feedback from passive

collection to actionable insights for optimizing

medication decisions.

6 CONCLUSIONS

Reviews are becoming an integral part of our daily

lives; whether we go shopping, order online, or visit

a doctor, we often rely on others’ experiences to make

informed decisions. Motivated by this, our research

explores sentiment analysis of drug reviews to build

a recommendation system using various machine

learning classifiers such as Logistic Regression,

Multilayer Perceptron (MLP), Multinomial Naive

Bayes, Ridge Classifier, SGD Classifier, and Linear

SVC, applied on TF-IDF extracted features. We

evaluated the models using five performance metrics:

precision, recall, F1-score, accuracy, and AUC.

Among all, the Multilayer Perceptron (MLP) with

TF-IDF features achieved the highest performance

with 99.9% accuracy, while Ridge and Naive Bayes

yielded comparatively lower results. We selected the

best predictions from each classifier and combined

them with the normalized useful count of reviews to

compute an overall drug score for each condition.

This scoring helps identify the most effective drugs

based on patient feedback. Future work includes

improving feature extraction, applying oversampling

techniques, and optimizing classifiers to further

enhance the accuracy and effectiveness of the

recommendation system.

REFERENCES

Aggarwal, C. C. (2016). Recommender Systems: The

Textbook. Springer.

Apicella, M. (2000). PolyAnalyst 4.1 digs through data for

gold. InfoWorld, 3 July 2000.

Bermingham, A., & Smeaton, A. F. (2010). Classifying

sentiment in microblogs: Is brevity an advantage?

Proceedings of the 19th ACM International Conference

on Information and Knowledge Management, 1833–

1836.

Borth, D., Ji, R., Chen, T., Breuel, T., & Chang, S.-F.

(2013). Large-scale visual sentiment ontology and

detectors using adjective noun pairs. Proceedings of the

21st ACM International Conference on Multimedia,

223–232.

Brusilovsky, P., Kobsa, A., & Nejdl, W. (Eds.). (2007). The

Adaptive Web: Methods and Strategies of Web

Personalization. Springer.

Cambria, E., Hussain, A., Havasi, C., & Eckl, C. (2009).

Sentic computing: Exploitation of common sense for

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

522

the development of emotion-sensitive systems.

Development and Applications of Artificial

Intelligence, 148–156.

Cambria, E., Schuller, B., Xia, Y., & Havasi, C. (2013).

New avenues in opinion mining and sentiment analysis.

IEEE Intelligent Systems, 28(2), 15–21.

Dey, L., & Haque, S. K. M. (2009). Opinion mining from

noisy text data. International Journal on Document

Analysis and Recognition, 12(3), 205–226.

Feng, X. Y., Zhang, H., Ren, Y. J., Shang, P. H., Zhu, Y.,

Liang, Y. C., & Guan, R. C. (2019). The deep learning–

based recommender system "Pubmender" for choosing

a biomedical publication venue: Development and

validation study. Journal of Medical Internet Research,

21(5), e12957.

Garg, S. (2021). Drug recommendation system based on

sentiment analysis of drug reviews using machine

learning. arXiv preprint arXiv:2104.01113.

Hu, M., & Liu, B. (2004). Mining opinion features in

customer reviews. Proceedings of the 19th National

Conference on Artificial Intelligence, 755–760.

Jakob, N., & Gurevych, I. (2010). Extracting opinion

targets in a single and cross-domain setting with

conditional random fields. Proceedings of the 2010

Conference on Empirical Methods in Natural Language

Processing, 1035–1045.

Lamba, M., & Madhusudhan, M. (2019). Application of

sentiment analysis in libraries to provide temporal

information service: A case study on various facets of

productivity. Social Network Analysis and Mining,

9(1), 1–14.

Lei, J., Shafik, R., Wheeldon, A., Yakovlev, A., & Granmo,

O.-C. (2021). Low-power audio keyword spotting using

Tsetlin machines. Journal of Low Power Electronics

and Applications, 11(2), 9.

Liu, Y., Huang, X., An, A., & Yu, X. (2008). ARSA: A

sentiment-aware model for predicting sales

performance using blogs. Proceedings of the 31st

Annual International ACM SIGIR Conference on

Research and Development in Information Retrieval,

607–614.

Morency, L.-P., Mihalcea, R., & Doshi, P. (2011). Towards

multimodal sentiment analysis: Harvesting opinions

from the web. Proceedings of the 13th International

Conference on Multimodal Insterfaces, 169–176.

Nguyen, Q. H., Nguyen, M.-V. T., & Nguyen, K. V. (2024).

New benchmark dataset and fine-grained cross-modal

fusion framework for Vietnamese multimodal aspect-

category sentiment analysis. arXiv preprint

arXiv:2405.00001.

Pang, B., & Lee, L. (2008). Opinion mining and sentiment

analysis. Foundations and Trends in Information

Retrieval, 2(1–2), 1–135.

Poria, S., Cambria, E., Hussain, A., & Huang, G.-B. (2015).

Towards an intelligent framework for multimodal

affective data analysis. Neural Networks, 63, 104–116.

Poria, S., Cambria, E., Bajpai, R., & Hussain, A. (2017). A

review of affective computing: From unimodal analysis

to multimodal fusion. Information Fusion, 37, 98–125.

Tang, H., Tan, S., & Cheng, X. (2009). A survey on

sentiment detection of reviews. Expert Systems with

Applications, 36(7), 10760–10773.

Drug Recommendation System Based on Sentiment Analysis of Drug Reviews Using Machine Learning

523