Drug Review Sentiment Analysis using Boosting Algorithms

International Journal of Trend in Scientific Research and Development (IJTSRD)
Volume 5 Issue 4, May-June 2021 Available Online: www.ijtsrd.com e-ISSN: 2456 – 6470
@ IJTSRD | Unique Paper ID – IJTSRD42429 | Volume – 5 | Issue – 4 | May-June 2021 Page 937
Drug Review Sentiment Analysis using Boosting Algorithms
Sumit Mishra
Electrical and Electronics Department, Bharati Vidyapeeth’s College of Engineering, New Delhi, India
ABSTRACT
Sentiment Analysis of the Reviews is important to understand the positive or
negative effect of some process using their reviewsaftertheexperience.Inthe
study the sentiment analysis of the reviewsofdrugsgiven bythepatientsafter
the usage using the boosting algorithms in machine learning. The Dataset
used, provides patient reviews on some specific drugs along with the
conditions the patient is suffering from and a 10-star patient rating reflecting
the patient satisfaction. Exploratory Data Analysis is carried out to get more
insight and engineer features. Preprocessing is done togetthedata ready.The
sentiment of the review is given according to the rating of the drugs. To
classify the reviews as positive or negative three Classification models are
trained LightGBM, XGBoost, and CatBoost and the feature importance is
plotted. The result showsthatLGBMisthebestperformingBoostingalgorithm
with an accuracy of 88.89%.
KEYWORDS: Sentiment Analysis, NLP, Classification, textblob, Features
Engineering
How to cite this paper: Sumit Mishra
"Drug Review Sentiment Analysis using
BoostingAlgorithms"
Published in
International Journal
of Trend in Scientific
Research and
Development(ijtsrd),
ISSN: 2456-6470,
Volume-5 | Issue-4,
June 2021, pp.937-941, URL:
www.ijtsrd.com/papers/ijtsrd42429.pdf
Copyright © 2021 by author (s) and
International Journal ofTrendinScientific
Research and Development Journal. This
is an Open Access article distributed
under the terms of
the Creative
Commons Attribution
License (CC BY 4.0)
(http://creativecommons.org/licenses/by/4.0)
I. INTRODUCTION:
Sentiment Analysis of the reviews is a very important aspect
as it helps companies perceive how positively the customers
are perceiving the productand to make changes accordingto
user satisfaction and help them grow. Doing the sentiment
analysis on the drug review is also an application where the
drugs which don’t have any effect on the condition can be
changed with something effective and the side effect of the
drug can also be analyzed.
The dataset used, contains the reviews of the drugs given by
the patients according to their experience after the usage of
the drug and a rating from 1 to 10 is given for the specific
drug [8]. The purpose of this study is to do the sentiment
analysis on the drug reviews given by the patients using the
gradient boosting algorithms in machine learning.
Initially, the dataset was segregated into two parts whichare
train and test set but as the main task is to do the sentiment
analysis so both the dataset is merged to get more data to
train and test overall (In this study we'll talk about the
merged dataset only). Exploratory Data Analysis is done on
the dataset with different features to gain insight about the
data and the correlation between them which will help in
feature engineering. The features like 'uniqueID' are not of
much use as they are just the identity given to each of the
data point or patient review. some preprocessing is done
before the EDA so the data is ready for the Exploratory Data
Analysis. Different bar graphs, Histogram and Word Clouds
are plotted. The Reviews then are cleaned so that
unnecessary words and elements are removed and the
features are generated. Feature Engineering is done on both
the uncleaned and cleaned reviews. Textblob module is also
used to give the polarity to both the cleaned and uncleaned
reviews and use it as a feature as well. The classification
models are trained and their performances are comparedon
the basis of the evaluation metrics given in the study.
Accuracy score is sometimes misleading when the dataset is
skewed so metrics like recall and precision are also given for
deeper analysis. The Classification algorithms that are used
are LightGBM, XGBoost, and CatBoost.
Previous research which are related to the drug review
sentiment analysis were carried out byJin-CheonNaandWai
Yan Min Kyaing where they did the sentiment analysis with
three valuespositive,negativeandneutralandimplementthe
SVM [10]. In the early studies like Pang et al [11] which was
focused on document level analysis for the sentiment
assigning. Some recent researcher like Jo & Oh [12] and Ding
et al [13] carried out the sentence level sentiment analysisto
examine the opinion or review. Some researcherlikeLiu[14]
have used linguistic features in addition to the word features
to overcome the limitation of the Bag-of-Word approach. In
this study, I used thegradientboostingalgorithmsinmachine
learning for the sentiment analysis. The results of the three
models are also compared based on the evaluation metrics.
II. Dataset
The Dataset is taken from the UCI Machine Learning
Repository [7]. The dataset containsthereviewsofthedrugs
given by the patients according to their experience after the
usage of the drug and a rating from 1 to 10 is given for the
specific drug [8]. The Drug Review Data Set is of shape
(215063, 7) i.e. It has 7 features including the review
and 215,063 Data Points or entries. The features are
'drugName' which is the name of the drug, 'condition' which
is the condition the patient is suffering from, 'review' is the
patients review, 'rating' is the 10-star patient rating for the
drug, 'date' is the date of the entry and the 'usefulcount' is
IJTSRD42429

International Journal of Trend in Scientific Research and Development (IJTSRD) @ www.ijtsrd.com eISSN: 2456-6470
the number of users who found the review useful. The
sentiment of the Drug review is the target variable that the
models will train to predict. Here we can notice that the
sentiment of any review is not given, so we have to give the
sentiment to the rating first and then use it as the target
variable. The drugName and condition are categorical
features, the date is date object, rating and usefulcount are
numerical features, and the review is text.Thesereviews are
from the year 2008 to 2017.
III. Exploratory Data Analysis
Exploratory Data Analysis (EDA) is donetogetaninsightinto
the data and summarize the main characteristics. To
understand dependency or correlation of the features. The
plots are generated using the matplotlib [5] and seaborn [6]
library.
Fig 1 Top 20 drugs with 10/10 ratings
Figure 1 is a bar graph which shows the top 20 drugs given
in the data set with a rating of 10/10. 'Levonorgestrel' is the
drug with the highest number of 10/10 ratings, about 1883
Ratings in the data set for 'Levonorgestrel'. It’s followed by
‘Phentermine’ with 1079 ratings.
Levonorgestrel (LNG) is a synthetic progestogen similar to
Progesterone used in contraception and hormone therapy.
Also known as Plan B, it’s used as a single agent for
emergency contraception, and as a contraceptive hormone
released from the intrauterine device, known as the IUD.
Fig 2 Top 20 drugs with 1/10 rating
Figure 2 is a bar graph that shows the top 20 drugs given in
the data set with a rating of 1/10. 'Miconazole' is the drug
with the highest number of 1/10 ratings, about 767. It's
followed by 'Ethinyl estradiol/norethindrone' and
'Etonogestrel'.
Fig 3 Count and Distribution of ratings
Figure 3 shows a distribution plot on the right-hand side and
a bar graph of the same on the left-hand side. This shows the
distribution of the ratings from 1 to 10 in the data set. It can
be inferred that mostly it's 10/10 rating and after that 9 and
1. The data points with rating of the drugs from 2 to 7 is
pretty low.
Fig 4 Top 10 Conditions in the Dataset
Figure 4 is a bar graph which exhibits the top 10 conditions
the people are suffering from.Inthisdataset'BirthControl'is
the most prominent condition by a very big margin followed
by Depression and pain. The 'Birth Control' condition has
occurred about 38,436 and the depressions have occurred
about 12,164. It can easily be noticed that the ‘Birth Control’
is more than 3 time the depression in the whole data set. In
the top 10 conditions, ADHD is in the 10th Position, ADHD
stands for Attention deficit hyperactivity disorder.
Fig 5 Number of reviews per year
Figure 5 is a Bar graph that shows the number of reviews in
the data set per year. It can be inferred that most ratings are
given in 2016 and 2008 has the least number of reviews.
2016 have 46507 reviewswhereas2008have 6700reviews.
A. Word Clouds
A Word Cloud is a visual representation of the frequency of
the words occurring in the text or speech i.e., text data. It’s
also known as a tag cloud. Higher the frequency of the word
in the text bigger its size will be and vice versa.

Fig 6 Word Cloud for the reviews with a rating of 10
Figure 6 shows the word cloud for the Drug reviews which
has a rating of 10/10. We can see that words like‘sideeffect’,
‘now’ and ‘year’ are occurring very frequently.
Fig 7 Word Cloud for the reviews with rating 1
Figure 7 shows the word cloud for the Drug reviews which
has a rating of 1/10. We can see that words like ‘side effect’,
‘pill’ and ‘period’ are occurring very frequently.
IV. Preprocessing
Dataset contains 6 Features about the drugs which are
'drugName', 'condition', 'rating', 'date', 'usefulcount' andthe
'review' itself. The sentiments ofthereviewsare notgiven so
we have to generate them based on the ratingsanduseitasa
target value which is to be predicted.
The train and test sets are merged so that we have a large
combined dataset as the sentiments were not given in either
of the sets. The size of the dataset is 215,063 Rows and 7
columns. The data set is then sorted based on the unique ID
of the drugs (data points). The Data points with the null
values in any of the given features are dropped as
the dropped rows were only 0.55% of the total data. The
shape of the dataset after dropping is (213,869,7).Thedates
given in the dataset are not in the Date Time format, so I
have changed it to the datetime64 format for further
processing. The sentiment for the reviews is given based on
the rating. If the rating is greater than 5 then it's positive
sentiment and if the rating is less than or equal to 5 then it's
a negative sentiment.
The reviews are cleaned before the feature engineering.
Regular expressions are used to clean the reviews. The
reviews are changed into the lower case first so that there's
uniformity. After analyzing the reviews, it’s found there's a
repeating pattern "'" which is occurringinmostof the
reviews hence they are removed. All the Special characters
are removed. Some special characterswerestill lefthenceall
the non-ASCII characters are removed. Trailing and leading
whitespaces are removed from the reviews. Multiple
whitespaces are replaced with a single space for more
clarity.
The stopwords are also removed from the reviews as it'll be
not very useful in the modelling. Only English stopwordsare
removed. The words in the review are also stemmed using
the snowball stemmer. For example, the word running will
be replaced with run.
V. Feature Engineering
There are initially 7 features given in the dataset which are
'drugName', 'condition', 'rating', 'date', 'usefulcount' andthe
'review'. The heatmap of the correlation matrix of the
numerical features is plotted before the feature engineering
which is given in figure 8.
Fig 8 Correlation Heatmap before Feature Engineering
It's plotted with seaborn [6]. It can be inferred that the
correlation between the 'usefulcount' and 'rating' is
significant that is 0.24. 'uniqueID' is just the Unique ID given
to each data point that is the consumer of the drug.
The textblob library is used to generate the sentiment
polarity of the drug review [9]. This polarity is given to both
the cleaned and uncleaned review. The interestingfactisthat
the correlation coefficient of the rating and the uncleaned
review is 0.348 and with cleaned reviews is 0.233 hence it's
greater for uncleaned review so, I have dropped the cleaned
review columns and Cleaned it again but this time without
removing the stopwords and stemming the words. Now the
correlation coefficient of the cleaned review with the rating
is 0.346 which is very good when compared to thelastresult.
The new features engineered are 'count_word' which is the
number of words in each review, 'count_unique_word'which
is the number of the unique words in the reviews.
'count_letters' is the letter count, 'punctuation_count' is the
punctuation count, 'count_words_upper' is the upper-case
word count,'count_words_title' is the title case word counts,
'count_stopwords' is the number of stop wordsinthereview,
and the 'mean_word_len' is the averagelengthofthewordsin
the review. The date isalso divided intothreecolumnswhich
are day, month and year for separate features for training.
A new correlation heatmap is plotted using seaborn which
contains all the new features engineered and the old
features. It’s given in figure 9.

Fig 9 Correlation Heat map after Feature Engineering
The Label Encoder is used to changethecategorical valuesof
Drug Names and the conditions into numerical valuesforthe
machine learning modelling. Thereare3,667uniquedrugsin
the dataset that's why One hot encoder is not used as it
would generate 3,667 new features and it would be very
computationally expensive.
VI. Modelling
The shape of the dataset after the deletion of the null values
is (213,869, 7). 70% of the dataset is used for the training
and the rest of the data i.e., 30% is used for the testing
purpose. The shape of the training set is (149708, 15) and
the shape of the test set is (64161, 15). It can be seen that
before feature engineering that there were only 7 features
but now there are 15. Three Machine learning models are
trained which are LightGBM, XGBoost,andtheCatBoost.The
feature importance is also plotted for LightGBM and the
description of these algorithms and their hyperparameters
are given below. The feature importancegivenbythemodels
XGBoost and LightGBM are also plotted.
A. LGBM
LightGBM stands for Light Gradient Boosting Machine. It's a
gradient boosting framework which is based on the tree-
based learning algorithms. It's a very efficient boosting
algorithm. There are certain advantages for LGBM like fast
training speed and high efficiency, lower memory usageand
support of parallel and GPU learning as it is based on
decision tree algorithms, it splits the tree leaf wise in
accordance to the best fit. The LightGBMusesthe XGBoostas
a baseline. The LGBM algorithm outperformsmany boosting
algorithms in terms of efficiency and the size ofthedatasetit
can comprehend easily. The hyperparameters of the
LightGBM used are,
LGBMClassifier (n_estimators = 10000, learning_rate=0.10,
num_leaves = 30, subsample = .9, max_depth =7,reg_alpha =
.1, reg_lambda = .1, min_split_gain = .01, min_child_weight =
2, silent = -1, verbose = -1)
Fig 10 Feature Importance Plot by LGBM
Figure 10 depicts the feature importance plot using the
LightGBM. It can be inferred that the most importantfeature
in the dataset is the mean word length and after that the
condition of the patient. The condition of the patient and the
useful-count are very comparableinfeatureimportance. The
least important feature of them all is the upper-case word
count.
B. XGBoost
XGBoost stands for extreme Gradient Boosting. XGBoost is a
boosting algorithm used in many tasks in machine learning
[2]. It is an optimized gradient boosting library which is
basically designed to be highly efficient and flexible. It's also
a Gradient Boosting framework which is under the machine
learning algorithms. XGBoost bring forth the parallel tree
boosting. It's open-sourced gradient boosting framework
available for C++, Java, Python, R, Julia, Perl, and Scala. Most
of the operating systems can be used for working on
XGBoost. From the project description, it intends to produce
a "Scalable, Portable and Distributed Gradient Boosting
Library" [3].
Recently, XGBoost has earned a lot of popularityandbecame
the choice of algorithm for many winning teams of machine
learning competitions. It is an optimized Gradient Boosting
Machine Learning library. The hyperparameters of the
XGBoost are,
XGBClassifier (n_estimators = 10000, learning_rate = 0.10,
num_leaves = 30)
Fig 11. Feature Importance Plot by XGBoost
Figure 11 depicts the feature importance plot using the
XGBoost. It can be inferred thatthemostimportantfeatureis
the condition of the patient and it's far more important than
the features following it. The features like sentiment,
usefulcount and the year are equally important for the
training.

C. CatBoost
CatBoost is an algorithm for gradient boosting on decision
trees. It is developed by Yandex researchers and is used for
many applications like search, recommendation systems,
weather prediction and many other tasks at Yandex and in
other companies well [4]. It's open-source as well. The
hyperparameters of the CatBoost are,
CatBoostClassifier (iterations = 10000, learning_rate = 0.5)
VII. Evaluation Metrics
A. Accuracy
It is the ratio of the correct predictions i.e., the correct
predicted values over the total prediction or total values.
Accuracy = (TP + TN) / (TP +TN + FP + FN)
B. Precision
Precision is defined as the ratio of true positive to the sum of
true positive and false positive. It defines how often the
classifier is correct when it predicts positive.
Precision = TP / (TP + FP)
C. Recall
Recall is defined as the ratio of true positive to the sum of
true positive and false negative [1]. It defines how the
classifier is correct for all positive instances.
Recall = TP / (TP + FP)
D. F1 Score
The F1 score can be interpreted or defined as a weighted
average of the precision and recall as given in the equation,
where an F1 score has its best value at 1 and worst score at
0.
F1 Score = 2*(Precision*Recall) / (Precision + Recall)
VIII. Results
Three machine learning models aretrained whichareLGBM,
XGBoost and CatBoost. Given are some boosting algorithms
in machine learning. The aim is to classify the sentiment of
the drug reviews given by the patient as negativeorpositive.
The results of the experiment are shown in table 1. It can be
inferred that the best performing model is the LGBM
followed by CatBoost. The accuracy of the LGBM is 88.89%
with a good F1 Score of 0.922. The CatBoost algorithm also
has a very good result and very close to LGBM. The XGBoost
is not able to perform better in the task as compared to the
other two models as the accuracy of the Model was 76.85%.
Hence, LGBM is the best boosting algorithm in machine
learning for the Drug review sentiment analysis.
Table I. Results
Algorithm Accuracy F1-Score Precision Recall
LGBM 0.888 0.922 0.902 0.942
XGBoost 0.768 0.846 0.786 0.917
CatBoost 0.882 0.916 0.904 0.929
IX. Conclusion
The main aim of the study is to predict the sentiment of the
drug reviews given by the patients using the Boosting
algorithms in Machine learning and compare them. Hence
Exploratory Data Analysis was done to get more insight into
the dataset and preprocessing was done to get the data
ready for both the modelling and EDA. Initially, 7 features
were given, hence feature engineering was done based on
the EDA and reviews by the patients. The reviews were
cleaned, and features are generated. The features were
generated by both the cleaned and uncleanedreviews.Inthe
Machine Learning modelling, three classification models
were trained which were LightGBM, XGBoost, and the
CatBoost. The performance metrics used here are Accuracy,
F1-Score, Precision and Recall. Thebestperformingmodel is
the LGBM Classifier, but its accuracy and the classification
report are comparable to the CatBoost Classifier. The
accuracies were 0.888 and 0.882 respectively. The features
importance is also plotted for LGBM and CatBoost. The
XGBoost was not able to perform better than the other two.
References
[1] Duman, Ekrem & Ekinci, Yeliz & Tanriverdi, Aydin.
(2012). Comparingalternativeclassifiersfordatabase
marketing: The case of imbalanced datasets. Expert
Syst. Appl. 39. 48-53. 10.1016/j.eswa.2011.06.048.
[2] “XGBoost Hyperparameters Overview” [Online].
Available:https://www.numpyninja.com/post/xgboo
st-hyperparameters-overview.
[3] “XG Boost Documentation” [Online]. Available:
https://xgboost.readthedocs.io/en/latest/index.html.
[4] “CatBoost Documentation” [Online]. Available:
https://catboost.ai.
[5] “Matplotlib Documentation” [Online]. Available:
https://matplotlib.org/stable/index.html.
[6] “Seaborn Documentation” [Online]. Available:
https://seaborn.pydata.org/.
[7] Felix Gräßer, Surya Kallumadi, Hagen Malberg, and
Sebastian Zaunseder. 2018. Aspect-Based Sentiment
Analysis of Drug ReviewsApplyingCross-Domainand
Cross-Data Learning. In Proceedings of the 2018
International Conference on Digital Health (DH '18).
ACM, New York, NY, USA, 121-125.
[8] Agrawal, A., Yang, X., Agrawal, R., Shen, X., & Menzies,
T. (2020). Simpler Hyperparameter Optimization for
Software Analytics: Why, How, When? ArXiv,
abs/2008.07334.
[9] Madhu, Shrija. (2018). An approach to analyze
suicidal tendency in blogsandtweetsusing Sentiment
Analysis. International Journal of Scientific Research
& Management Studies. 6. 34-36.
10.26438/ijsrcse/v6i4.3436.
[10] Na, Jin-Cheon & Kyaing, Wai. (2015). Sentiment
Analysis of User-Generated Content on Drug Review
Websites. Journal of Information Science Theory and
Practice. 3. 6-23. 10.1633/JISTaP.2015.3.1.1.
[11] Pang, B., Lee, L., & Vaithyanathan, S. (2002). umbs
up? Sentiment classification using machine-learn-ing
techniques. Proceedings of the 2002 Conference on
Empirical Methods in Natural Language Process-ing
(pp. 79-86).
[12] Jo, Y., & Oh, A. H. (2011). Aspect and sentiment uni-
cation model for online review analysis. Proceed-ings
of the Fourth International Conferenceon WebSearch
and Data Mining (WSDM) (pp. 815-824). Hong Kong.
[13] Ding, X., Liu, B., & Yu, P. S. (2008). A holistic lexi-
22JISTaP Vol.3 No.1, 06-23con-based approach to
opinion mining. Proceed-ings of the International
Conference on Web Search and Web Data Mining (pp.
231-240). New York: ACM.
[14] Liu, B. (2012). Sentimentanalysisandopinionmining.
San Rafael, CA: Morgan & Claypool.

Drug Review Sentiment Analysis using Boosting Algorithms

More Related Content

What's hot (20)

Similar to Drug Review Sentiment Analysis using Boosting Algorithms (20)

More from ijtsrd (20)

Recently uploaded (20)

Drug Review Sentiment Analysis using Boosting Algorithms