10th IEEE International Conference on Communication Systems and Network Technologies

Sentiment analysis of Twitter data by making

2021 10th IEEE International Conference on Communication Systems and Network Technologies (CSNT) | 978-1-6654-2306-9/21/$31.00 ©2021 IEEE | DOI: 10.1109/CSNT51715.2021.9509679

use of SVM, Random Forest and Decision Tree

algorithm
Jyotsna Singh Pradeep Tripathi
Computer Science & Engineering Computer Science & Engineering
Vindhya Institute of Technology and Science, Satna, Vindhya Institute of Technology and Science, Satna,
India India
[email protected] [email protected]
A vast amount of data is found on the internet, which is in
Abstract: Data production over the internet has been increased an unstructured format. In this present data, almost 80% of
tremendously because of social media sites' growth. A the data is unstructured, which is needed to be formatted in
significant number of users share their thoughts, images, the required format [4]. The structuring of the unstructured
videos etc., on their particular accounts. Among these, all data is required as the formless. Data becomes difficult for
varieties available for the thoughts expression twitter provide a
concise way to share users' thoughts. This thought shared by
analysis, and accuracy with such data becomes significantly
users is called tweets. These tweets may lead to a big revolution less. Sentiment analysis is a type of analysis that works of
for a good change or, at the same time, lead to a big issue over text placed on the network [5].
any particular topic. So sentiments involved in such tweets Twitter is a blogging site which gives the user to share their
needed to be classified, and then only the tweet may be shared thoughts, and this site is used by a considerable number of
with other users. Here in our paper, we have taken a dataset users [6]. The thoughts or tweets done on any particular
from the Kaggle website, which was collected based on KFC's topic can be useful for making an opinion. This opinion can
challenge and McDonald's of AI. We have used more than be of any type that can be used for any purpose. Some of the
14000 tweets for analysis. The cleaning of the dataset is done major areas are politics, brands, campaign etc. [7]. The
using Term Frequency- Inverse Document Frequency (TF-
IDF). After cleaning the dataset, we have applied three
tweets made on such topics can be useful for analysis.
classification algorithms for the testing purpose that are Opinions can be useful when it comes to taking any decision
Support Vector Machine (SVM), Random Forest (RF) and regarding any particular topic.
Decision Tree (DT). Among these three algorithms, the highest So this paper presents the Work done for the sentiment
accuracy is obtained from the Decision tree algorithm is analysis of the data taken from Twitter. The paper does the
88.51%. The experiment calculates accuracy, recall, precision sentiment analysis after cleaning the dataset. For cleaning,
and F1 measures. we have used TF-IDF, and then this data has been used to
implement the classifier for the classification of the
Keywords: Twitter, Decision Tree, Sentiment Analysis, Supervised sentiments. The dataset used here is taken from the Kaggle
Learning, Random Forest, Support Vector Machines.
website.
I. INTRODUCTION

Social media platforms are a craze for everyone in today's

scenario. Facebook, Twitter, Instagram, Telegram etc., have
been used in a huge amount, and data created due to these
social media platforms is vast [1]. Sentiment analysis is also
termed, in other words, as Opinion mining. Opinion mining
is also becoming an essential task for companies to manage
social media platforms [2]. One wrong message Spread over
the internet can harm humanity in any of the ways. So we
need to apply opinion mining here to get the user's actual
mood. It may be happy, angry, etc. Also, the behaviour and
attitude of the user needed to be understood. Sentiment
analysis plays an essential role in this Work. Reactions of
users can be Analyzed for various fields such as politics, Figure1. Some of the social media sites
finance, sociology and economy [3].

978-1-6654-2306-9/21/$31.00 ©2021 IEEE 193

DOI: 10.1109/CSNT.2021.35
Authorized licensed use limited to: Somaiya University. Downloaded on December 28,2023 at 13:38:52 UTC from IEEE Xplore. Restrictions apply.
 II. LITERATURE REVIEW 3.1 Data Preprocessing:
Here authors have done sentiment analysis for the tweets Text mining needs preprocessing of the data on a necessary
made on donations, fundraising or some charities. They basis in the overall procedure of the analysis. The data
have implemented a natural language processing toolkit to provided to the algorithm as an input makes the basis for the
check tweet sentiment, which is neutral, positive or negative accuracy obtained after a comprehensive evaluation. This
[8]. The sentiment analysis for any area helps in getting the preprocessing of data helps remove the stopwords, hashtags,
information of that area. Here the authors have used and various languages can be removed through this step.
machine learning algorithms for the sentiment analysis. Making whole text in lower case, removing Unicode and
They implemented an unsupervised learning approach for obtaining the required data format has been done in this
this Work. The algorithm here used by them is the rule- step. The dataset should be free of tags, URLs, along with
based or lexical approach. They have taken tweets having this space, and tabs also need to remove.
hashtags as #haryanaassemblypolls, #theskyispink, etc. [9]. 3.2 Computation of Features:
Social media have become an open platform to share • The next step after the data preprocessing is to
personal thoughts, and the shortest form to share is provided obtain features from the Twitter data, and the
by Twitter. 90% of peoples share their thought on social representation of such data in vector form is to be
media platforms. Here authors have implemented sentiment done.
analysis on statistical tools by use of R programming. Some • This textual information is converted into sectors to
of the feelings analyzed in this Work are disgust, trust, represent this data in numerical format. This Work
surprise, anticipation etc. [10]. In this paper, the authors is useful in machine learning algorithms. If we see
have done sentiment analysis for the Twitter data. For this, the ML can't understand text very well and
the authors have used TFIDF along with natural language numerical representations are more feasible.
processing. They have implemented the proposed approach, • We have used TF-IDF for the vectorization of the
and the accuracy achieved with the NLP technique is dataset.
85.25% [11]. Sentiment analysis has been performed in TF-IDF: Term frequency and inverse document frequency
various ways earlier. But here, the authors have are the terms that are used widely for retrieving the
implemented a sentiment analysis of the Twitter dataset in information from the textual dataset by making use if
real-time. The authors have used 2 algorithms for this weighs. Weight calculation of any particular term in the text
purpose that us simple voter and naïve Bayes. Among both is done by multiplicating the TF with IDF.
algorithms, the Naïve Bayes algorithm has proved to be In this procedure, any keyword is founded by getting the
more efficient [12]. importance of that particular word, and the TF-IDF
algorithm defines the number of occurrences of that word.
III. PROPOSED ALGORITHM One more thing it does is corpus which means that the
Here we have presented the approach given by us, which is similar word to the keywords found. Let us have any
used for sentiment analysis. The overall Work is carried out document d, and weight is WT of any term t in that
in four stages: data collection, preprocessing of the data, document.
computation of the features and the classification of the WT, d = TF (t), d log (N/DF(t))
sentiments have been done in the very last stage. A Here, TF(t) is the number of times any term occurs in a
flowchart for the Work that is carried out in this paper has document, in whole corpus number of documents present in
been given below. N, number of total documents that have that term present in
it DF(t).

Random Forest: It is a supervised algorithm for the

classification task. Random forest gives higher accuracy if
several trees formed in random forest algorithm is large.
The pseudo-code for the algorithm has been provided
below.
1. Select k out of m features available.
2. From these k features, calculate d node using the
split points.
3. Split further node into daughter nodes
4. Repeat step 1 to step 3 until we reach i several
nodes.
5. Repeat step 1 to step 4 for n times to create n
number of trees.
Figure 2. Flowchart for the proposed system

194

Authorized licensed use limited to: Somaiya University. Downloaded on December 28,2023 at 13:38:52 UTC from IEEE Xplore. Restrictions apply.
Support Vector Machine Algorithm: SVM is an algorithm Data Collection: Here, we have explained the evaluation
that is majorly used for the task of classification. Various carried out by our proposed methodology to analyze Twitter
data points have been extracted through SVM, and mapping data. Our experimental analysis is done on the Kaggle
was done in high dimensional space with a non-linear kernel dataset. Dataset has been obtained from the challenge given
function. The algorithm for SVM has been stated with the by the school of AI (Artificial Intelligence). We have used a
help of the pseudo-code given below. portion of 4000 tweets for training purpose and 10000
tweets for testing purpose [13].
Inputs: provide different datasets for training and testing of
the classifier Data Set Processing: This section provides the details of
Outputs: determine the accuracy calculated the experiments that we have performed to analyze the
1. Choose gamma and cost optimally for SVM proposed methodology in the context of Twitter analysis.
2. While (condition = true) do We have done tests on the Kaggle Twitter data set. Data set
3. SVM training for every data point based on the challenge launched by the KFC and
4. SVM testing for every data points McDonald's of AI - Algiers, which consists of building a
5. End while system that can classify tweets as Sad or Happy. Currently,
6. Return accuracy we have check tweets that are correct or incorrect. In the
KFC and McDonald's dataset, we have taken 14000 tweets
Decision Tree Algorithm: Another classifier we have is a for our research. 10000 tweet for training and 4000 tweets
decision tree classification algorithm. This algorithm can be for testing purpose. The data set link mention below.
majorly used for textual classification. https://www.kaggle.com/mcdonalds/nutrition-facts, details
figure 3 and figure 4.
Input: training and testing dataset (tweets).
Output: accuracy, precision, F1 measure, recall
Start
Pre-processing and data normalization;
For training dataset to do
Features calculation;
Decision tree algorithm;
Classifier building;
End
Value is used for particular tweets;
For testing dataset
Analyze accuracy
End
(training, testing)
End
Figure 3.Without Processed Data
IV. IMPLEMENTATION
There are URL's, own usernames, special characters and
Requirements: The required software and hardware has repeated words and symbols. We have to remove all the
been stated below. Hashtags identified by the # symbols, all the special
For the design Python Programming Language, 15.6 in HD characters, URL's, own usernames and repeated words.
WLED touchscreen (1366 x 768), 10-finger multi-touch
support. 10th Generation Intel Core i7-1065G7 1.3 GHz up
to 3.9 GHz. 8GB DDR4 SDRAM 2666MHz, 512GB SSD,
No Optical Drive. Intel Iris Plus Graphics, HD Audio with
stereo speakers. HP TrueVision HD camera. Realtek
RTL8821CE 802.11b/g/n/ac, Bluetooth 4.2, 1 HDMI 1.4, 1
USB 3.1 Gen 1 Type-C, 2 USB 3.1 Gen 1 Type-A. The
Python Programming was run on Windows 10 64-bit
Operating System platform. The python library was used
during implementation like NumPy, Pandas, Matplotlib,
SciPy, Scikit-Learn, PyTorch, Seaborn, XGBoost, Plotly,
TensorFlow, Keras, Seaborn, TextBlob, Stanford CoreNLP,
Gensim, and Afterword.
Figure 4. After Processed Data

195

Authorized licensed use limited to: Somaiya University. Downloaded on December 28,2023 at 13:38:52 UTC from IEEE Xplore. Restrictions apply.
We have removed all private usernames, special characters,
all hashtags identified by # symbols, repeated words and Recall
symbols. 100
85.7384.8885.67
V. RESULT ANALYSIS 80

Performance of each procedure is done no. of times, and 60

results have been calculated using different classification 44

%
procedures on a Twitter dataset. Figures have been drawn 40 33
using other products as calculated.
Our proposed method evaluated in the below parameters: 20 11

0
▪ Recall
Existing Work Proposed Work
▪ Precision
▪ Accuracy SVM Decision Tree Random Forest
▪ F1-Score
Figure 5 Recall Figure between Existing Work and
Table 1Evaluate Metric with Contingency Table Proposed Work

Figure 5 calculates the recall value for all algorithms

Existing Work [14] and Proposed Work. And results shown
with the help of a diagram. We find that the proposed
approach shows true positive rate compare to collaborative
and content-based approaches.

Precision: Precision measures the number of positive class

expectations that have a place with the positive class. The
precision is calculated using equation 2.

Precision
90 80 80.6679.6281.67
80
70
tp 60 50
recall = ……………….. (1) 50
tp+fn
%

tp 33
precision = …………………(2) 40
tp+fp
30
tp+tn 20
accuracy = …………….. (3) 10
tp+tn+fp+fn
Recall: Recall evaluates the quantity of positive class 0
expectations made from every single positive model in the Existing Work Proposed Work
dataset. The recall is calculated using equation 1.
SVM Decision Tree Random Forest

Figure 6 Precision Figures between Existing Work and

Proposed Work

Figure 6 calculates a precision value for all algorithms

Existing Work [16] and Proposed Work and results shown
with a diagram's help. We find that the proposed approach
shows a true positive rate compare to collaborative and
content-based approaches.

196

Authorized licensed use limited to: Somaiya University. Downloaded on December 28,2023 at 13:38:52 UTC from IEEE Xplore. Restrictions apply.
Accuracy: Accuracy is essentially a proportion of the Figure 8 here, we can say from the above results that the
accurately anticipated groupings (True Positives + True proposed approach is efficient. And running time is reduced
Negatives) to the absolute Test Dataset. The accuracy is to an extent by keeping the quality of recommendation as to
calculated using equation 3. its best. This concludes that the proposed method is scalable
and can be applied to a large dataset.
Accuracy
VI. CONCLUSION
100

80 Sentiment analysis is necessary to do on social media

87.7188.5186.55
platforms to stop spreading hatred messages or issues. We
60 have proposed our Work to do the sentiment analysis of
%

56 54 58 tweets made by users over any particular topic. We have

40
done our data cleaning with the help of the available TF-
20 IDF algorithm. The dataset used here was of McDonald's
and KFC customers. More than 14000 tweets were
0 analyzed, and from these, 4000 tweets were used for
Existing Work Proposed Work training and the remaining 10000 for testing. The results
were carried out for three different SVM algorithms,
SVM Decision Tree Random Forest
Decision Tree and Random Forest, among which best
accuracy was obtained for the decision tree algorithm,
Figure7 Accuracy Figures between Existing Work and which valued 88.51%.
Proposed Work
Reference:
Figure 7 calculates a precision value for all algorithms
Existing Work [14] and Proposed Work and results shown [1] L. Mandloi and R. Patel, "Twitter sentiments analysis using
machine learning methods," 2020 Int. Conf. Emerg. Technol.
with a diagram's help. We find that the proposed approach
INCET 2020, pp. 1–5, 2020, doi:
shows true positive rate compare to collaborative and 10.1109/INCET49848.2020.9154183.
content-based strategies. [2] V. Prakruthi, D. Sindhu, and S. Anupama Kumar, "Real-Time
Sentiment Analysis of Twitter Posts," Proc. 2018 3rd Int. Conf.
Comput. Syst. Inf. Technol. Sustain. Solut. CSITSS 2018, pp. 29–
F1_Measure 34, 2018, doi: 10.1109/CSITSS.2018.8768774.
The F1 measure states the accuracy of the test; this is [3] C. W. Park and D. R. Seo, "Sentiment analysis of Twitter corpus
specially used in the binary classification. The precision and related to artificial intelligence assistants," 2018 5th Int. Conf. Ind.
the recall are used for the calculation of the F1 measure. Eng. Appl. ICIEA 2018, pp. 495–498, 2018, doi:
10.1109/IEA.2018.8387151.
All samples that should have been identified as positive in [4] S. Dhawan, K. Singh, and P. Chauhan, "Sentiment Analysis of
Figure 4 Twitter Data in Online Social Network," Proc. IEEE Int. Conf.
F1_Score = 2 * ((precision * Recall) / (precision * Recall)) Signal Process. Control, vol. 2019-October, pp. 255–259, 2019,
doi: 10.1109/ISPCC48220.2019.8988450.
[5] Kusrini and M. Mashuri, "Sentiment analysis in twitter using
lexicon-based and polarity multiplication," Proceeding - 2019 Int.
F1_Measure Conf. Artif. Intell. Inf. Technol. ICAIIT 2019, pp. 365–368, 2019,
100 doi: 10.1109/ICAIIT.2019.8834477.
[6] R. Wagh and P. Punde, "Survey on Sentiment Analysis using
78.1 Twitter Dataset," Proc. 2nd Int. Conf. Electron. Commun. Aerosp.
80 73.02
Technol. ICECA 2018, no. Iceca, pp. 208–211, 2018, doi:
57 55.57 10.1109/ICECA.2018.8474783.
60 [7] S. A. El Rahman, F. A. Alotaibi, and W. A. Alshehri, "Sentiment
40
%

Analysis of Twitter Data," 2019 Int. Conf. Comput. Inf. Sci. ICCIS
40 2019, 2019, doi: 10.1109/ICCISci.2019.8716464.
16 [8] A. Shelar and C. Y. Huang, "Sentiment analysis of twitter data,"
20 Proc. - 2018 Int. Conf. Comput. Sci. Comput. Intell. CSCI 2018,
pp. 1301–1302, 2018, doi: 10.1109/CSCI46756.2018.00252.
0 [9] S. Zahoor and R. Rohilla, "Twitter Sentiment Analysis Using
Existing Work Proposed Work Lexical or Rule-Based Approach: A Case Study," ICRITO 2020 -
IEEE 8th Int. Conf. Reliab. Infocom Technol. Optim. (Trends
SVM Decision Tree Random Forest Futur. Dir., pp. 537–542, 2020, doi:
10.1109/ICRITO48877.2020.9197910.
[10] S. Saini, R. Punhani, R. Bathla, and V. K. Shukla, "Sentiment
Analysis on Twitter Data using R," 2019 Int. Conf. Autom.
Figure 8 F1_Score Figures between Existing Work and Comput. Technol. Manag. ICACTM 2019, pp. 68–72, 2019, doi:
Proposed Work 10.1109/ICACTM.2019.8776685.
[11] M. R. Hasan, M. Maliha, and M. Arifuzzaman, "Sentiment

197

Authorized licensed use limited to: Somaiya University. Downloaded on December 28,2023 at 13:38:52 UTC from IEEE Xplore. Restrictions apply.
 Analysis with NLP on Twitter Data," 5th Int. Conf. Comput.
Commun. Chem. Mater. Electron. Eng. IC4ME2 2019, pp. 1–4,
2019, doi: 10.1109/IC4ME247184.2019.9036670.
[12] A. S. Al Shammari, "Real-time Twitter Sentiment Analysis using
a 3-way classifier," 21st Saudi Comput. Soc. Natl. Comput. Conf.
NCC 2018, pp. 1–3, 2018, doi: 10.1109/NCG.2018.8593205.
[13] https://www.iflexion.com/blog/sentiment-analysis-python.
[14] S. A. El Rahman, F. A. AlOtaibi and W. A. AlShehri, "Sentiment
Analysis of Twitter Data," 2019 International Conference on
Computer and Information Sciences (ICCIS), Sakaka, Saudi
Arabia, 2019, pp. 1-4, doi: 10.1109/ICCISci.2019.8716464.

198

Authorized licensed use limited to: Somaiya University. Downloaded on December 28,2023 at 13:38:52 UTC from IEEE Xplore. Restrictions apply.