Hindawi
Advances in Human-Computer Interaction
Volume 2021, Article ID 5561204, 7 pages
https://doi.org/10.1155/2021/5561204
Research Article
Statistical Analysis of Public Sentiment on the Ghanaian
Government: A Machine Learning Approach
John Andoh, Louis Asiedu , Anani Lotsi, and Charlotte Chapman-Wardy
Department of Statistics & Actuarial Science, School of Physical and Mathematical Sciences, University of Ghana, Legon,
Accra, Ghana
Correspondence should be addressed to Louis Asiedu; lasiedu@ug.edu.gh
Received 5 January 2021; Revised 27 January 2021; Accepted 18 February 2021; Published 3 March 2021
Academic Editor: Francesco Bellotti
Copyright © 2021 John Andoh et al. 'is is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Gathering public opinions on the Internet and Internet-based applications like Twitter has become popular in recent times, as it
provides decision-makers with uncensored public views on products, government policies, and programs. 'rough natural
language processing and machine learning techniques, unstructured data forms from these sources can be analyzed using
traditional statistical learning.'e challenge encountered in machine learning method-based sentiment classification still remains
the abundant amount of data available, which makes it difficult to train the learning algorithms in feasible time. 'is eventually
degrades the classification accuracy of the algorithms. From this assertion, the effect of training data sizes in classification tasks
cannot be overemphasized. 'is study statistically assessed the performance of Naive Bayes, support vector machine (SVM), and
random forest algorithms on sentiment text classification task. 'e research also investigated the optimal conditions such as
varying data sizes, trees, and kernel types under which each of the respective algorithms performed best. 'e study collected
Twitter data from Ghanaian users which contained sentiments about the Ghanaian Government. 'e data was preprocessed,
manually labeled by the researcher, and then trained using the aforementioned algorithms.'ese algorithms are three of the most
popular learning algorithms which have had lots of success in diverse fields. 'e Naive Bayes classifier was adjudged the best
algorithm for the task as it outperformed the other two machine learning algorithms with an accuracy of 99%, F1 score of 86.51%,
and Matthews correlation coefficient of 0.9906. 'e algorithm also performed well with increasing data sizes. 'e Naive Bayes
classifier is recommended as viable for sentiment text classification, especially for text classification systems which work with
Big Data.
1. Introduction considering the reviews of users on such online platforms.
'rough the use of machine learning techniques, data an-
'e explosion of blogging, microblogging, social media, and alysts can extract and classify this wealth of information to
review sites has armed data analysts with valuable infor- make informed inferences. 'is process of making the
mation on users’ preferences. Information is now shared all computer understand human language in texts is largely
over the world at ever-increasing speeds, volume, and di- called natural language processing (NLP). NLP techniques
versity. 'is connectivity leaves “data prints” which we can can also be used to perform sentiment analysis to summarize
use to describe almost everything in our world today. opinions from online platforms.
Consequently, one type of data that has become increasingly 'e conjoining of news with social networking and
important in recent times is the opinions and preferences of blogging has made Twitter a hotbed for the discussion of
Internet users regarding products, subjects, and views. 'is events in real time. Twitter currently serves as a medium for
type of data aggregates in e-commerce sites, blogs, social the discourse of a wide variety of societal issues such as
media, and other online platforms. Traditional methods of sports, governance, advocacy, religion, and especially poli-
collecting data on product feedback from customers such as tics. Public views expressed in the form of text on these
interviews and polling are gradually being phased out by societal issues are called sentiment texts. For instance,
2 Advances in Human-Computer Interaction
before, during, and after the 2016 US presidential elections, One of the earliest uses of sentiment analysis using
Twitter proved itself as the major election news destination. Twitter as corpus was in [5]. 'eir work highlights the
A record 40 million tweets were posted regarding the importance of preprocessing techniques as they are neces-
elections and its “immediacy and speed” was unmatched by sary to achieve higher accuracies. 'ey employed “emoti-
any other traditional news network [1]. cons” as noisy labels to achieve maximum accuracies of
'ere is an active Ghanaian presence on Twitter and 82.7% for Naive Bayes when using unigram and bigram
other social media platforms who update their statuses re- features, 82.7% for max entropy using unigram and bigram
garding happenings in their social circles and conversations features, and 82.9% for SVM when using only unigram
on politics. Assuming opinions shared on Twitter mirror features. 'ey concluded by highlighting the shortcomings
public perception as it is unbiased and unrestricted, a of sentiment analysis at the time, which included the han-
sentiment analysis task trained on data from Twitter would dling of neutral tweets, internalization (so as to be able to use
yield interesting results for policy analysts and political them for multiple languages), and the utilization of emo-
parties. ticon data.
As one of the pioneering works in this field, the paper [2] 'e authors in [6] also asserted that the challenge en-
classified reviews by sentiments using Naive Bayes, max countered in machine learning method-based sentiment
entropy, and SVM and analyzed the difficulty under each classification is the abundant amount of data available. 'ey
classification task for sentiment analysis. 'ey sought to explained that this amount makes it difficult to train the
recognize whether sentiment classification was a special learning algorithms in a feasible time and degrades the
topic-based categorization, which was a technique for text classification accuracy of the built model. 'ey recom-
classification or special sentiment categorization methods mended feature selection as essential in developing robust
needed to be developed to address the novel challenges and efficient classification models whilst reducing the
sentiment analysis tasks presented. Even though all three training time.
methods outperformed human classification, they could not 'e effect of training data sizes in classification tasks has
reach accuracies achieved by the same methods for topic been of interest to researchers obviously because of its
categorization. 'e classification task in sentiment analysis purported influence on accuracy. 'e authors in [7], for
becomes challenging if the texts are rhetoric and sarcastic. instance, measured the effect of training data sizes on
Features exclusive to sentiment analysis will be needed to classification using SVM and Naive Bayes. 'ey concluded
accommodate such words. that the effect was not significant. 'e authors in [8] also
'e authors in [3] evaluated the effectiveness of statis- found that the complexity of the features can affect accuracy
tical keyword extraction methods in conjunction with en- and that some classifiers could even work better with less
semble learning algorithms. 'ey compared base learning data. 'is study among others will investigate training sizes
algorithms (Naive Bayes, support vector machines, logistic effect.
regression, and random forest) with five widely used en- Generally, the study seeks to identify the most suitable
semble methods (AdaBoost, bagging, dagging, random machine learning processes for collecting, analyzing, and
subspace, and majority voting). 'eir study revealed that the predicting public sentiments from Twitter. 'e study does
bagging and random subspace ensembles of random forest this specifically by analyzing tweets on sentiments about
yield promising results. 'ey also found that the use of political discourse in the country, analyze the various
keyword-based representation of text documents in con- conditions under which the algorithms work well with the
junction with ensemble learning enhances the predictive tweet data, and statistically evaluate the performance of the
performance and scalability of text classification schemes. study algorithms.
In another comparative study of various classification
techniques for sentiment analysis, the authors in [4] pointed 2. Materials and Methods
out that, in selecting a particular algorithm, one will need to
consider the type of specific input required. 'is implies 2.1. Data Acquisition and Authorization. Twitter returns a
that, in order to achieve higher accuracies, it is important to collection of tweets that match a specified query. 'e
know which algorithm will be appropriate given the avail- standard search Application Program Interface (API)
able input data. 'eir study identified Naive Bayes, max accessed from the Twitter developer page is free but de-
entropy, boosted trees, and random forest classifiers as the velopers do not have access to the entire database of tweets.
most widely used algorithms in sentiment analysis. 'ey Only tweets from the last 30 days can be accessed with this
concluded by noting that each classifier had its advantages standard search API.
and disadvantages and could all be assessed on the basis of We secured assess to the standard API for some period to
accuracy, resources (computing power), data input, time for extract tweets related to the subject for this study. 'is was
training, etc. 'e random forest classifier achieved the done through creation of a developer account which was
highest accuracy and exhibited improvements over time. It is eventually approved.
however costly in terms of resources as it uses longer
training times and requires high computing power. It will be
interesting to investigate how well the random forest clas- 2.1.1. $e Tweets. 'e ease in obtaining data from tweets
sifier will perform when dealing with short text classification through the Twitter API for developers was one key mo-
in Twitter data which will be considered in this study. tivation for performing this sentiment analysis. R packages
Advances in Human-Computer Interaction 3
capable of accessing the Twitter API which were used for this
study include “twitteR” and “rtweet.” After obtaining au-
thorization, the tweets were collected using the keywords,
“NPP,” (the ruling party) and “nanakuffoaddo” (President of
Ghana) and making it specific to Ghana by tagging the
geolocation for Ghana. 3,000 tweets were collected over a
two-month period (January 2020 to March 2020).
Figure 1 shows a word cloud diagram of tweets used in
the study.
'e word cloud diagram like the bar plot explores fre-
quent words; however, the word cloud is often desirable
because it represents the words with their relative fre-
quencies aesthetically.
'e word cloud from our data suggests “Ghana” is the most
popular word found in tweets regarding the governance,
probably the central theme ofmost Ghanaian public sentiments.
Some other relevant keywords are also shown in Figure 1.
'e tweets extracted had various useful attributes like the
screen name of user, tweet text, time stamp, geolocation, and
number of “retweets” and “likes.” For this study, only the
texts were extracted. 'e following are samples of the tweets
in their raw forms: Figure 1: Word cloud of tweets.
(i) “Npp has a vision for Ghana.”
(ii) “No reasonable Ghanaian will vote for NDC or NPP
again!” 100
(iii) “When its NPP its a different Narative. When its the 80
NDC, then yeah the NDC Is corrupt.”
53
(iv) “@CheEsquire All hail the NPP government.” 60
(v) “While we’re busy with NDC vs NPP - Ghana is 40 33
losing.”
20 14
'ese tweets presented were processed in stages to
remove unwanted characters like numbers, punctuation, 0
special characters, and stopwords in order to reduce noise Negative Neutral Positive
and prepare them for the classifiers. Each of the tweets above Sentiments 
conveys some form of sentiment, which shall be classified as Figure 2: 'e prior distribution of the sentiment texts or tweets.
positive, negative, or neutral.
2.2. Random Forest Model. 'e random forest model clas-
2.1.2. Annotation of Tweets. 'e tweets were manually sifier is actually a bagging method of various classifiers or
annotated by one researcher and cross-referenced by an- decision trees. 'e idea is to average the results of various
other, before being considered for sentiment classification. decision trees in order to reduce overall variance. Each tree is
All the tweets which were classified differently were re- independent and identically distributed (i.i.d.) and the ex-
moved. Tweets that were regarded as positive towards the pectation of a number of trees is the same as that of the
government were classified as positive sentiments (e.g., “Npp individual trees. 'e random forest is the collection of these
has a vision for Ghana”) while those regarded as negative individual trees and the results of a classification represent
towards the government were regarded as negative senti- the majority votes of the trees.
ments (e.g., “No reasonable Ghanaian will vote for NDC or Given an ensemble of classifiers h1((x)), h2
NPP again!”). Tweets that had no sentiment, or could neither ((x)), . . . , hk((x)) with training set drawn randomly from
be classified as positive nor negative, were regarded as the random vector X,Y, we define the margin function as
neutral sentiments.
Out of the 3,000 tweets, 990 tweets were prepared for mg((X,Y)) � avkI( hk(X) � Y􏼁 − max avkI( hk(X) � j􏼁,j≠Y
sentiment classification after the cleaning and annotation
(1)
phases. Figure 2 shows the prior distribution of the senti-
ment texts. where I(·) is the indicator function.
From Figure 2, 14% of the tweets had positive conno- 'ismarginmeasures the extent to which the average votes
tations, about 33% of the tweets were negative, and 53%were at X,Y for the actual class exceeds the average vote for any
neutral. other class.'e larger themargin, themore confidence we have
Words (%)
4 Advances in Human-Computer Interaction
in the classification [9]. 'e performance of the algorithm will Matthews Correlation Coefficient (MCC). According to [11],
be assessed by varyingm, the number of variables considered at the MCC is a more reliable statistical rate which produces a
split, within √� /2 √�􏽮 p , p 􏽯 and the number of trees. high score only if the prediction obtained good results in all
categories of the confusionmatrix categories. Formore details
about the adopted performance metrics, please refer to [11].
2.3. Naive Bayes Model. 'e Naive Bayes algorithm is based
on Bayes’ 'eorem, a probabilistic method used for calcu-
lating likelihoods of events based on conditional probabil- 3. Results and Discussion
ities. 'e probability of a document d being assigned to a 3.1. Model Training. 'e random forest, Naive Bayes, and
category or class c is given by SVM classifiers are trained using manually annotated tweets.
P(c | d)αP(c) 􏽙 P( t |c􏼁, 'e data was divided into three (150, 300, and 540 tweets) tok (2)
1≤ ≤ investigate the algorithms’ performances with varying datak nd
sizes. 70% of each dataset was used for training the algo-
where P(tk|c) is the conditional probability of a term tk in d rithms and the remaining 30% for testing.
of a certain class c [10]. In line with the objectives of the 'e dataframe of the document term matrix created
study, the Naive Bayes model was tested on various data with from the corpus which has been split into 3 has these
different feature space sizes and number of observations. It is dimensions:
our expectation to find the optimal conditions so as to get From Table 1, the observations are the individual tweets
the best accuracy out of the model. whereas the variables are obtained from 99.5% of the most
common words/items in the tweets.
We shall adopt a baseline of 33% to compare the ac-
2.4. Support Vector Machine Model. If we have N training curacy of the algorithms. 'is baseline is obtained by di-
data of pairs, (x1, y1), (x2, y2), . . . , (xn, yn), with xi ∈ Rk and viding the tweets in three. For any model, a performance
yi ∈ { ± 1}, i � 1, 2, . . . , n, we can define the hyperplane as above 33% implies it is better than a random selection of the
T
x: f(x) � x β + β � 0, (3) label by the respective classifier.0
where β is a unit vector with ‖β‖ � 1. 'e classification is
then determined by 3.2. Results of the Random Forest Algorithm. 'e random
forest algorithm was run on the three randomly shuffled sets
G(x) � sign xT􏽨 β + β0􏽩. (4) of data. Dataset 1 contained 150 tweets; Dataset 2, 300
tweets; and Dataset 3, 540 tweets. As stated earlier, 70: 30
If the classes are easily separable, the function f(x) � ratio was adopted to split each dataset for training and
xTβ + β0 with yif(xi)> 0,∀i. 'e hyperplane with the testing, respectively. 'e datasets were also trained con-
biggest margin between points of different classes is now sidering the set √� √�m � 􏽮 p /2, p 􏽯 (p is the number variables
reduced to the optimization problem considered) and the set of trees (500, 1000).
max M 'e plot in Figure 3 also shows the Out-of-Bag (OOB)
β,β0 ,‖β‖�1 (5) estimate of error for all the classes in the model. 'e black
subject to xTβ β ≥ 1 2 line shows the OOB estimate for the model as a whole whileyi􏼐 i + 0􏼑 M, i � , , . . . , n. the green, red, and blue represent negative, neutral, and
Kernels are used to modify dimensionality of the data to positive classes, respectively.
find the flat affine dimensional subspace hyperplane to 'e best performances of the random forest algorithm in
correctly determine the accurate support vector classifier. terms of Out-of-Bag estimates of error are given in Table 2.
'e structure of the data determines the kind of kernel to Using the best performance of the random forest model√�
use, whether to use a linear, polynomial, radial basis (m � p and 1000 trees), we get the following confusion
function, or sigmoid. In this study, the results present the matrix and some performance statistics shown in Tables 3
optimal conditions for this algorithm on sentiment text and 4, respectively.
classification based on kernel type. From Table 4, using
√�
m � p and 1000 trees for the
random forest result in an overall accuracy of 52.22% with a
runtime of 10.22 seconds.
2.5. Performance Metrics of Machine Learning Models. 'e random forest model was now tested on the training
Comparison of performances of various machine learning data to further investigate the model. Table 5 compares the
models is very important and does not need to be just kappa and accuracy.
superficial. For instance, just comparing “Accuracy” To optimize random forest models, we vary the number
amongst different models may be inadequate and statisti- of trees and number of variables at split (m). 'e R package
cally insufficient especially when the accuracies are close. “randomForest” uses a default of 500 trees and
√�
m � p,
In this study, we compare performances of the study where p is the number of variables considered. From Table 5,
algorithms using Cohen’s kappa statistic (measure of reli- the model has a kappa statistic of 0.3 (fair reliability) and an
ability), F1 score (which strikes a balance between precision accuracy of 53.33% when used to classify the test data.'is is
and recall), sensitivity, specificity, classification accuracy, and slightly above the baseline of 33%.
Advances in Human-Computer Interaction 5
Table 1: Dimensions of datasets.
Dataframe (df) Observations Variables (p)
df1 150 845
df2 300 466
df3 540 399
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0 200 400 600 800 1000
Trees
Model Positive
Negative Neutral
Figure 3: Random forest plot of OOB error estimates of classes.
Table 2: Random forest training.
Dataset m Trees OOB estimate (%) Kappa Accuracy
150 √�p 500 64.76 0.067 0.3778
300 √�p 1000 53.81 0.2924 0.5222
540 √�p 1000 46.32 0.2414 0.4875
Table 3: Confusion matrix of the random forest model.
Actual (%)
Negative Neutral Positive
Negative 54.55 22.73 22.73
Predicted (%) Neutral 15.38 84.62 0.00
Positive 23.64 32.73 43.64
Table 4: Performance statistics of the random forest model.
Sensitivity Specificity
Runtime (sec.) 95% accu. CI Accuracy
Neg Neu Pos Neg Neu Pos
10.22 (0.143, 0.6287) 0.5222 0.4444 0.3235 0.8276 0.8413 0.9643 0.4918
Table 5: Prediction comparison with testing on training data.
Kappa 95% accu. CI Accuracy
Model on testing data 0.3085 (0.4251, 0.6393) 0.5333
Model on training data 0.8642 (0.8623, 0.9446) 0.9095
Error
6 Advances in Human-Computer Interaction
Table 6: Confusion matrix of the Naive Bayes model.
Actual (%)
Negative Neutral Positive
Negative 98.33 1.67 0.00
Predicted (%) Neutral 0.00 100.00 0.00
Positive 0.003 0.00 100.00
Table 7: Performance statistics of the random forest model.
Sensitivity Specificity
Runtime (sec.) 95% accu. CI Accuracy
Neg Neu Pos Neg Neu Pos
0.09 (0.9657, 0.9998) 0.9938 1.0000 0.9783 1.0000 0.9901 1.0000 1.0000
'e high variation between the accuracy of the al-
gorithm when used to classify the train data and the test Table 8: Support vector machine training.
data shows evidence of overfitting. In general, the per-
formance of the random forest model is not too appre- Dataset Kernel type No. of support vectors Kappa Accuracy
ciable.'e huge variance from the 95% confidence bounds 150 RBF 105 0.1 0.4
and the high OOB estimates of error rates are also not 300 Linear 91 0.3506 0.5667
ideal. From the tests, we can conclude that the best 540 Linear 317 0.3115 0.5375
random forest models were achieved with 1000 trees and
√�
m � p.
Table 9: Confusion matrix of the SVM model.
3.3. Results of the Naive Bayes Algorithm. 'e Naive Bayes
model was run similarly on the three randomly shuffled Actual (%)
datasets (150, 300, and 540 tweets). 'e confusion matrix of Negative Neutral Positive
the best Naive Bayes algorithm and some performance Negative 45.00 35.00 20.00
metrics are shown in Tables 6 and 7, respectively. Predicted (%) Neutral 24.00 72.00 4.00
It is evident from Tables 7 that the overall accuracy of Positive 26.67 20.00 53.33
the Naive Bayes algorithm is 99.38% which is highly
appreciable. 'e results from Tables 6 and 7 show that the kernel outperformed all the other kernels in Datasets 2
Naive Bayes model performs remarkably well. It is and 3 and consequently had the best accuracy.
worthy of note that the performance of the algorithm It is evident from Table 10 that the SVM model had the
increases with increasing dataset size. 'e runtime for best performing accuracy of 56.67% with 95% CI of 45.8%–
the Naive Bayes algorithm (shown in Table 7 as 0.09 67.08% when used to classify the study data. 'e runtime
seconds) is relatively better than the runtime of the averaged around 0.11 seconds. 'e SVM model performs
random forest algorithm (shown in Table 4 as 10.22 slightly above the 33% baseline and the huge variability in
seconds). the confidence interval of the accuracy is an evidence of low
precision.
3.4. Results of the Support Vector Machine (SVM) Algorithm.
An SVM classifier was also used on the same datasets 3.5. Comparison ofModels. Table 11 shows the results of the
(150, 300, and 540 tweets). SVM can be extended to solve best performing models under the three different
multiclass categories problems, not just binary, as has algorithms.
been discussed in the methodology. 'e appropriateness From Table 11, the Naive Bayes model outperformed
of the various kernel methods for this task is also the random forest and SVM, recording the highest kappa
explored with the most suitable ones reported in Table 8. statistic value of 0.9906 (near perfect reliability), F1 score
'e SVM model had its best performance when ran on value of 86.51%, MCC of 0.9906, accuracy of 99.38%, and
Dataset 2, with an accuracy of 56.67% and kappa statistic the lowest runtime (0.09 seconds). 'e SVM model had a
value of 0.3506 (fair reliability). 'e confusion matrix and slight edge in performance over the random forest
other performance statistics of the SVM algorithm are algorithm.
shown in Tables 9 and 10, respectively. 'e random forest classifier had a relatively low per-
In training the SVM model, the various kernel types formance with an F1 score of 52.18%, MCC of 0.3404, and
were run on all three datasets and the best performing an accuracy of 53.33%. 'e relatively high computational
kernels are reported in Table 8. 'e RBF kernel worked time of 10.22 seconds is as a result of the numerous av-
best on Dataset 1 with 150 observations and 845 variables eraging of trees which further makes the classifier
but was not suitable for Datasets 2 and 3. 'e linear unattractive.
Advances in Human-Computer Interaction 7
Table 10: Performance statistics of the SVM model.
Sensitivity Specificity
Runtime (sec.) 95% accu. CI Accuracy
Neg Neu Pos Neg Neu Pos
0.11 (0.458, 0.6708) 0.5667 0.3333 0.5294 0.8276 0.8254 0.8750 0.6557
Table 11: Comparing performance results of the three classifiers.
Kappa F1 score MCC Accuracy Runtime (sec.)
Random forest 0.3085 0.5218 0.3404 0.5333 10.22
Naive Bayes 0.9906 0.8651 0.9906 0.9938 0.09
SVM 0.3506 0.5473 0.3638 0.5667 0.11
4. Conclusion and Recommendations Data Availability
As stated earlier, the study considered about 990 tweets 'e data used to support the findings of this study are
collected from Ghanaian Twitter users from January to available from the corresponding author upon request.
February of 2020. 'e tweets were collected using keywords
that identify with the government in order to gather public Conflicts of Interest
sentiment. Prior analysis of the data showed that 14% of the
tweets had positive connotations, about 33% of the tweets 'e authors declare that there are no conflicts of interest.
were negative, and 53% were neutral. 'is indicates some
sort of public disapproval of the government. Most of the References
tweets also were centered around keywords like “free” from
the free SHS policy the government implemented and [1] M. Isaac and S. Ember, “For election day influence, twitter
“cathedral” from the plan of the government to build a ruled social media,” $e New York Times, vol. 8, p. 11, 2016.
national cathedral. [2] B. Pang and L. Lee, “Opinion mining and sentiment analysis,”
From Table 11, the SVM classifier performed slightly Computational Linguistics, vol. 35, pp. 311-312, 2009.
better (F1 score of 0.5473, MCC of 0.3638, and accuracy of [3] A. Onan, S. Korukoğlu, and H. Bulut, “Ensemble of keyword
0.5667) than the random forest classifier. Literature on in- extraction methods and classifiers in text classification,” Ex-pert Systems with Applications, vol. 57, pp. 232–247, 2016.
vestigating and optimizing the various conditions like kernel [4] A. Gupte, S. Joshi, P. Gadgul, A. Kadam, and A. Gupte,
types of the SVM, tree sizes, and variable split sizes for “Comparative study of classification algorithms used in
random forest and other ensemble methods could be ex- sentiment analysis,” International Journal of Computer Science
plored in quest to improve their performance for sentiment and Information Technologies, vol. 5, pp. 6261–6264, 2014.
text classification tasks. [5] A. Go, R. Bhayani, and L. Huang, “Twitter sentiment clas-
'e results of this study as shown in Table 11 also sification using distant supervision,” CS224N Project Report,
revealed that the Naive Bayes classifier has the highest Stanford University, Stanford, CA, USA, 2009.
Cohen’s kappa statistic value of 0.99 (near perfect reliability [6] A. Onan and S. Korukoğlu, “A feature selection model based
of the algorithm), F1 score of 86.51%, Matthews Correlation on genetic rank aggregation for text sentiment classification,”
Coefficient of 0.9906, and classification accuracy of 99 38%. Journal of Information Science, vol. 43, no. 1, pp. 25–38, 2017..
'e algorithm also recorded the lowest run time/compu- [7] O. Abdelwahab, M. Bahgat, C. J. Lowrance, andA. Elmaghraby, “Effect of training set size on svm and naive
tational time of 0.09 seconds. 'is makes the Naive Bayes bayes for twitter sentiment analysis,” in Proceedings of the
algorithm relatively the best classifier for the sentiment text 2015 IEEE International Symposium on Signal Processing and
classification task. Information Technology (ISSPIT), pp. 46–51, IEEE, Abu
'e findings of the study are consistent with existing Dhabi, UAE, December 2015.
literature which suggests Naive Bayes models perform well [8] A. Hadjarian, J. Zhang, and S. Cheng, “An empirical analysis
with high-dimensional feature spaces and with little data. of the training and feature set size in text categorization for
'e study therefore recommends the Naive Bayes model as a e-discovery,” in Proceedings of the International Conference on
viable algorithm for text classification. Artificial Intelligence and Law, Rome Italy, June, 2013.
'is study brings to light the potential benefits of [9] L. Breiman, “Random forests,” Machine Learning, vol. 45,
harvesting social media data from Twitter, for instance, no. 1, pp. 5–32, 2001.[10] C. D. Manning, P. Raghavan, and H. Schütze, “Probabilistic
and making analysis on them. 'e study provides an information retrieval,” in Introduction to Information Re-
avenue for monitoring product performance on the trieval, pp. 220–235, Cambridge University Press, Cambridge,
markets, public sentiment, and track progress of policies, UK, 2009.
having many countless applications. Future studies can [11] D. Chicco and G. Jurman, “'e advantages of the matthews
consider the creation of a web tool for performing sen- correlation coefficient (mcc) over f1 score and accuracy in
timent analysis using the Naive Bayes classifier on tweets binary classification evaluation,” BMC Genomics, vol. 21,
in real time. pp. 1–13, 2020.