Journal of Artificial Intelligence and Engineering Applications

Website: https://ioinformatic.org/
th
15 June 2024. Vol. 3. No. 3; e-ISSN: 2808-4519

Optimization of Classification Algorithm with GridSearchCV

and Hyperparameter Tuning for Sentiment Analysis of the
Nusantara Capital City
Achmad Baroqah Pohan1*, Irmawati2, Aliyah Kurniasih3
1,2Universitas
Bina Sarana Informatika, Indonesia
3
Universitas Ary Ginanjar, Indonesia
[email protected]*, [email protected], [email protected]

Abstract

The relocation of the country's capital has implications for social, economic, geographical and political aspects. This raises the public's
views which are expressed in Twitter tweets that contain public sentiment. Sentiment analysis has a significant influence on understanding
public views. Therefore, further research is needed related to the sentiment of analysis of the IKN. In this study, a model that can predict
public sentiment is created by developing a model at the training stage using the GridSearchCV algorithm and hyperparameter tuning with
several classification algorithms. The best model was produced with the Support Vector Machine classification algorithm that was able to
outperform, compared to probability-based and tree-based classification algorithms with an accuracy of 69.95%.

Keywords: Hyperparameter Tunning, Sentiment, Classification Algorithm

1. Introduction
The Nusantara Capital City (IKN) of the Archipelago is a city that is projected to be the center of government for a country. The construction
of the IKN for the relocation of the state capital from Jakarta to Kalimantan Island will have implications not only on the socio-economic
and geographical aspects of the region, but with the realization of the political relocation of the state capital will create a new constitutional
dynamic [1]. This constantly makes all layers of people comment through Twitter, the views of the community can have elements of
sentiment that are positive, negative or neutral.

Sentiment analysis is the process of extracting and looking for patterns from textual data to obtain information contained in the text [2].
Machine learning modeling related to public sentiment analysis is still an interesting research topic. Sentiment analysis is carried out
because it has significant relevance in understanding public views and from understanding the public can provide valuable insights for
improvement and development [3].

Machine learning algorithms are still the main players in determining the results of model performance. Algorithms that are still trending
to use such as the Support Vector Machine, where the SVM model's calculation of the weight of each word in sentiment greatly determines
the accuracy produced [4]. Then Naïve Bayes' algorithm works by assuming the independence of variables so that only the variances
associated with variables in a label are the focus in determining the classification and none take into account the entire covariance matrix
[5]. Similarly, the Logistic Regression algorithm is a statistical analysis method used to predict the possible outcome of binary dependent
variables (two categories) based on one or more independent variables. This method models the relationship between the predictor variable
and the response variable through a logistic function, which results in a probability between 0 and 1 [6].

The Decision Tree algorithm is a predictive analysis method that uses a tree-like structure to make data-driven decisions. This structure
consists of nodes, branches, and leaves, where each node represents an attribute, each branch represents the value of that attribute, and each
leaf represents a class or result. Decision tree are easy to implement and their visualization makes it easy to understand the decision flow
[7], [8]. There is also the Random Forest algorithm, which is an ensemble learning method used for classification, regression, and various
other tasks by building a large number of decision trees during the training process. This method works by combining the results of various
decision trees to make more accurate predictions and reduce the risk of overfitting that often occurs in single decision tree models [9].

In this study, classification algorithms such as Logistic Regression, Naïve Bayes Classifier, Support Vector Machine, Decision Tree, and
Random Forest are applied by making tuning on hyperparameters for model training with the GridSearchCV algorithm process in
improving the overall model evaluation performance results.
Journal of Artificial Intelligence and Engineering Applications 809

2. Research Method
2.1. Dataset

The data used in this study is Twitter tweet data in Indonesian which is only related to the Capital City of the Archipelago (IKN). Data was
obtained through crawling with the keyword ikn language id (meaning data searched in Indonesian) in the crawling time period, which is
March 1, 2024 to March 18, 2024. The crawling results obtained data of 8,616 instances and 12 features.

2.2. Exploratory Data Analysis

Exploratory data analysis is a stage where what is done is with the aim of understanding data, recognizing a problem from data, and gaining
insights from data. Based on the results of the check, there are 7 features with data types of objects (strings) including features created_at,
id_str, full_text, lang, user_id_str, username, and tweet_url, then there are 5 features with data types of intergers including quote_count,
reply_count, retweet_count, favorite_count, conversation_id_str. The results of the subsequent check did not find any missing value data,
but there were 85 duplicates data as a whole and 178 duplicates data in feature full_text.

Table 1 is a description of each feature along with examples of the data produced. In the text data itself from the full_text attribute, it can
be seen from one example of the data row presented that the narrative of the tweet in the data contains URLs, capital letters, punctuation,
and even numbers and slang words in other text data rows that cannot be presented in its entirety.

Table 1: Features Description

No Features Description Data
1 Created_at UTC standard format date and time Sun Mar 17 18:55:41 +0000 2024
2 id_str Unique ID of the tweet in the form of a string 1769437468242440400
full_text Bandara VVIP IKN nantinya tidak bisa dipakai untuk rakyat,
melainkan khusus untuk penerbangan kepresidenan dan tamu-
3 Full text of a tweet tamu penting negara. RAKYAT TERUTAMA WARGA
KALIMANTAN ITU BUKAN TAMU/ORANG
PENTING !!! Waras lo ? https://t.co/HLyU6mWKLn
quote_count The number of times the tweet was quoted by
4 5
other users
reply_count The number of replies or how many people
5 13
responded directly to the tweet
6 retweet_count The number of retweets from the tweet 16
7 favorite_count The number of 'likes' received by the tweet 52
8 lang The language indicated by the tweet in
user_id_str The unique ID of the user who created the
9 1038483432391634944
tweet in the form of a string
conversation_id_str The unique ID of the conversation associated
10 1769437468242440400
with the tweet in the form of a string
11 username Username of the tweet creator RomitsuT
12 tweet_url Tweet URL or direct link https://twitter.com/RomitsuT/status/1769437468242440400

2.3. Data Preprocessing

2.3.1. Convert Datetime

In this process, the crated_at feature will be converted from the UTC standard format to the ISO 8601 standard format (created_at_new),
i.e. the data will be in the form of a date in the year-month-date format, followed by the time portion in the hour-minute-second format,
and the time zone offset from UTC to +07:00, which means that the time is 7 hours ahead of UTC. Furthermore, the feature created_at_new
is further divided into feature date (date_created), time (time_created), year, month, and day. The purpose of this process is to gain insight
into what day the tweet occurred the most in that time span. Table 2 is the data from the first 5 rows of the conversion results.

Table 2: Convert Datetime Results

created_at … created_at_new data_created time_created year month day
Mon Mar 18 00:11:17 +0000 2024 … 2024-03-18 07:11:17+07:00 2024-03-18 07:11:17 2024 March Monday
Mon Mar 18 00:10:57 +0000 2024 … 2024-03-18 07:10:57+07:00 2024-03-18 07:10:57 2024 March Monday
Mon Mar 18 00:10:18 +0000 2024 … 2024-03-18 07:10:18+07:00 2024-03-18 07:10:18 2024 March Monday
Mon Mar 18 00:10:13 +0000 2024 … 2024-03-18 07:10:13+07:00 2024-03-18 07:10:13 2024 March Monday
Mon Mar 18 00:06:20 +0000 2024 … 2024-03-18 07:06:20+07:00 2024-03-18 07:06:20 2024 March Monday

Based on the results of the time conversion, it can be known the distribution of the amount of data by day, the most tweets occurred on
Friday with 2,102 instances, followed by Tuesday 1,832 instances, Monday 1,738 instances, Sunday 1,449 instances, Saturday 781
instances, Thursday 377 instances, and Wednesday 337 instances. The tweet is data generated from crawling for a period of 19 days
starting from February 29, 2024 to March 18, 2024.

2.3.2. Removing Duplicate Data and Missing Values

In this step, it is to delete duplicate data which is as many as 85 instances, after which there is an indication of duplicate data in feature
full_text from the previous 178 instances to as many as 93 instances. Furthermore, duplicate data was also deleted based on the 90 full_text
feature, so that the data dimensions became 8,438 instances and 18 features.
810 Journal of Artificial Intelligence and Engineering Applications

2.4. Text Preprocessing

1. Text Normalization
At this stage, the tweet text data on feature full_text that will be used as input to the model is improved in the form of removing
mentions, removing hashtags, removing RTs, removing URLs that are replaced with single spaces, converting non-ASCII
characters to ASCII, removing certain patterns from text strings and then recombining the words generated using single spaces,
lowercase, remove punctuation, remove single character, remove double spaces in the middle of sentences, remove double spaces
at the beginning and end of sentences, and change English words to Indonesian.
2. Changing the Slang
Furthermore, correct the word commonly called slang by using a corpus in the form of a txt document containing slang and
standard words and then the tweet text data will be corrected based on the data on the corpus.
3. Stopword Removal
Stopword removal is the process of removing common words in the tweet data based on the stopword data in the NLTK (Natural
Language ToolKit) corpus.
4. Converting Number to Strings
This process converts the numbers in the tweet data into words in Indonesian.
5. Lemmatization
This process converts the words in the tweet data into root words based on the WordNetLemmatizer corpus data in NLTK
(Natural Language ToolKit).

2.5. Labeling

In machine learning modeling, especially in the sentiment analysis task in this study, where a data from crawling Twitter tweets does not
have a label, while training in machine learning modeling for sentiment analysis tasks requires labeled data. At this stage, the labeling
process is carried out using the polarity score sentiment technique based on positive and negative corpus lexicon [10]. Table 3 is the first
10 rows of the results of the text data before and after text preprocessing, then feature polarity_score along with the labeling of sentiment
(feature polarity) from the text data that has been cleaned based on the generation of feature polarity_score.

Table 3: Text Preprocessing and Labeling Results

No full_text text_preprocessed polarity_score polarity
1 @rasa2086 Dukung IKN dukung ikn 4 positive
2 Bukan Kampus, Kepala Otorita Tegaskan Stanford Hanya kampus kepala otorita tegaskan stanford 0 neutral
Bangun Pusat Riset di IKN. Mio Mirza JATENG IS RED bangun pusat riset ikn mio mirza jateng be
Nahan Pagiii Mawar Fajri https://t.co/eVkNXbFK92 redaksi nahan pagiii mawar fajri
3 @tempodotco alhamdulillah untung SUMATERA tercinta alhamdulillah untung sumatera tercinta 13 positive
ngk ikut2an IKN. selamat menikmati ya kalimantan. ngk ikut2an ikn selamat menikmati iya
kalimantan
4 Luar Biasa! Investasi Mengalir di IKN Sudah Tembus Rp 49,6 investasi mengalir ikn tembus rp empat 3 positive
Triliun. Barca Joao Felix Mio Mirza JATENG IS RED hyeri ratus sembilan puluh enam triliun barca
Mawar Fajar https://t.co/UepfC1Uttc joao felix mio mirza jateng be redaksi
hyeri mawar fajar
5 @tempodotco Warga kalimantan timur akan menjadi yg warga kalimantan timur dirugikan -11 negative
paling dirugikan dengan belum disahkannya RUU masyarakat disahkannya ruu masyarakat adat terkait
adat terkait IKN. Masih pingsan semua orang2 Kaltim itu ikn pingsan orang2 kaltim paham malang
rupanya !!! MASIH BELUM PAHAM JUGA KALIAN !!! kali nasib pingsanlah
Malang kali nasib kalian, pingsanlah terus !!!
6 @aiiptu Dukung pembangunan IKN dukung pembangunan ikn 4 positive
7 @MurtadhaOne1 Bagimana sy bs percaya anda ini soal IKN? bagimana percaya ikn suaranya keributan 0 neutral
Dari suaranya sangat jelas itu keributan antara warga dngan warga dngan perusahaan sawit
perusahaan sawit.
8 Kenapa beliau berbicara seperti teman sekelas beliau berbicara teman sekelas -3 negative
https://t.co/8jI33BXtOR
9 Sebenernya kasihan. Kalian terusir dari tanah kalian sendiri. sebenernya kasihan terusir tanah 1 positive
Tapi kan kalian dulu mendukung IKN. mendukung ikn
https://t.co/WEu5TcubAg
10 @DokterTifa Knp sih kamu d jawa sana sirik kl di luar jawa d sih jawa sirik jawa bangun dr tirta sok -12 negative
bangun? Dr tirta yg sok tau krn pindah agama lgsg bisa koar pindah agama langsung teriak teriak hello
koar? Hello.. Nanti juga cari makan d IKN., pret cari makan ikn pret

After the text preprocessing and labeling process, duplicate data was generated as many as 771 instances based on clean text data (feature
text_preprocessed). Furthermore, the duplicate data is deleted so that the data dimensions become 7,667 instances and 18 features. Then a
re-check was carried out regarding the missing value data and indicated that there was data that was missing value of 1 instance in feature
text_preprocessed and deleted, then the data dimensions became 7,666 instances and 18 features.

2.6. Data Transformation

Data transformation is the process of changing data from string categorical to numerical categorical, this process is carried out on feature
polarity as a sentiment feature for the target prediction model from machine learning classification. Table 4 below is the before and after
process of data transformation on features polarity.
Journal of Artificial Intelligence and Engineering Applications 811

Table 4: Data Transformation Results

Before After
negative 0
neutral 1
positive 2

2.7. Data Visualization

Figure 1 is the amount of data for each label based on days, this number is the data that has gone through the data and text preprocessing
stage that is ready to be used for the machine learning model creation process. Based on this image, the distribution of the most tweet data
on Friday was 1,974 instances, followed by Tuesday 1,665 instances, Monday 1,378 instances, Sunday 1,255 instances, Saturday 726
instances, Thursday 344 instances, and Wednesday 324 instances.

Fig. 1: Number of Label Data in Days

In Table 5, the amount of data based on the day is presented, namely in the data of the crawling results compared to the amount of data
after going through the data process and text preprocessing, in the difference column is the column about the number of data deleted. It can
be seen that there was a decrease in the amount of data from before and after due to indications of 949 instances of duplicate data and 1
instance of missing value data carried out in the previous process. So there is a difference in the overall amount of data from 8,616 instances
to 7,666 instances, and the amount of data deleted is 950 instances.

Table 5: Comparison of the Amount of Data Before and After Preprocessing

Day Before After Difference
Friday 2.102 1.974 128
Tuesday 1.832 1.665 167
Monday 1.738 1.378 360
Sunday 1.449 1.255 194
Saturday 781 726 55
Thursday 377 344 33
Wednesday 337 324 13
Sum 8.616 7.666 950

Figure 2 is the percentage of the amount of data based on labels, namely negative labels of 46.9% for 3,597 instances, positive labels of
34.5% for 2,646 instances, and neutral labels of 18.6% for 1,423 instances. Based on the distribution of the amount of data, it is indicated
that there is an imbalances class data, but in this study, the process of handling imbalance classes is not carried out because the process
does not always have a significant influence on machine learning modeling on text data [11].

Fig. 2: Data Label Percentage

2.8. Training and Testing Data Split Ratio

Data division is the process of dividing data by determining which features will be the input of the model (X) and which features will be
the target of the model (Y), then the data on X and Y will be re-divided with a percentage of 80% for train data and 20% for test data. In
812 Journal of Artificial Intelligence and Engineering Applications

this study, the input model is feature text_preprocessed, namely text data that has been preprocessed, and the target model is feature polarity
as the sentiment of the text data. Then the results of data split on the data train were produced as many as 6,132 instances, and test data as
many as 1,534 instances. The data train is used for the machine learning model training process and the test data is used for the machine
learning model evaluation process. The amount of data from the train data and test data based on sentiment labels is presented in Table 6.
Table 6: Results of Data Split
Data Label Sum
Negative [0] 2.868
Train Neutral [1] 2.127
Positive [2] 1.137
Negative [0] 729
Test Neutral [1] 519
Positive [2] 286

2.9. TF-IDF

TF-IDF (Term Frequency–Inverse Document Frequency) is one of the extraction features to convert text data into numeric which works
by calculating the weight of words in a document or data line based on the frequency of words appearing in a document or data line and
inverse document frequency, where documents or data lines that often contain certain words will have a higher weight while words that
appear in the entire document (data) or appear too rarely in the document will have more weight low. Formula (1) to calculate the TF-IDF
value is the result of tfv,d and tfv,d , where tfv,d is the sum of the frequencies of the word t (for example) in a document or data row d (for
example), and tfv,d is the frequency inverse of the entire document. TF increases with the number of word occurrences in a document or
data row and IDF increases with word step across a document collection or entire data row [11].

M
xv,d → tfv,d − idfv,d = (1 + log10 tcv,d ) × (log10 df ) (1)
v

In this study, the model input data (X) on the data train and data test was carried out by TfidfVectorizer using max_features=1000, but the
data train was carried out with the fit_transform process and in the data test only the transform process was carried out.

2.10. Training Model with Hyperparameter Tuning

Hyperparameter tuning is the process of adjusting hyperparameters and hyperparameter values that are set before the training process [12].
Hyperparameter tuning is used with the GridSearchCV algorithm where this algorithm works by looking for a certain subset of the
hyperparameter space of a machine learning model algorithm so that it can directly affect the performance results of the training model
[12].

In this study, machine learning algorithms such as Naïve Bayes Classifier, Support Vector Machine, Decision Tree, and Random Forest
were used to train the model by tuning on the hyperparameters of each algorithm. This training uses GridSearchCV with model algorithms,
variables that contain tuning of each algorithm's hyperparameters, and 10-fold cross validation [13].

Table 7 is the hyperparameters and the variation of values used in each machine learning algorithm used. The best hyperparameter value
was produced along with the score of the training model, namely the best model training score on the Logictic Regression algorithm of
71.38% with the best hyperparameter penalty L2 and the value of C=10, then followed by the next best model training score on the SVM
algorithm 70.79%, Complement NB 67.06%, Random Forest 66.54%, Multinomial NB 64.94%, Bernoulli NB 63.37%, Gaussian NB is
59.72%, and the lowest result on the Decision Tree algorithm is 56.10%.

Table 7: Model Training Results with Hyperparameter Tunning

Algorithm Hyperparameter and Values Best Hyperparameters Best Score
penalty= L1 dan L2 penalty=L2 71,38%
Logistic Regression
C=0.1, 1, 10, 100, 1000 C=10
Gaussian NB var_smoothing=(0, -9, num=10) var_smoothing=0.01 59,72%
Multinomial NB alpha=0.01, 0.1, 1, 10 alpha=0.01 64,94%
Complement NB alpha=0.01, 0.1, 1, 10 alpha=0.1 67,06%
Bernoulli NB alpha=0.01, 0.1, 1, 10 alpha=0.1 63,37%
Kernel= rbf, linear, poly Kernel=rbf 70,79%
SVM C=0.1, 1, 10 C=10
gamma (RBF)=1, 0.1, 0.01 gamma=0.1
criterion=gini, entropy criterion=gini 56,10%
Decision Tree max_depth= none, 10, 20 max_depth=none
min_sample_split=2, 5, 10 min_samples_split=10
n_estimators=100, 200, 300 n_estimators=200 66,54%
Random Forest max_depth=none, 10, 20 max_depth=none
min_samples_split=2, 5, 10 min_samples_split=10

3. Result and Discussion

Table 8 is the result of the overall model evaluation on each algorithm used on data that has three labels that are imbalance class. The
metrix used in precision, recall and f1-score uses macro weighting. Of the overall performance metrics used, the best model was produced
in the Support Vector Machine algorithm with an accuracy value of 69.95%, precision of 67.67%, recall of 54.96% and f1-score of 66.86%.
These results prove that SVM on Indonesian text data can create the best hyperplane in separating data on the RBF kernel and a C value
of 10 with a gamma of 0.1 compared to other probability-based algorithms such as the Naïve Bayers Classifier, and other tree-based
Journal of Artificial Intelligence and Engineering Applications 813

algorithms such as Decision Tree and Random Forest, although Random Forest works by maximizing each tree compared to Decision
Tree, but this study proves that the SVM algorithm is still superior.

Table 8: Model Evaluation Results

Algoritma Accuracy Precision Recall F1-Score
Logistic Regression 68,90% 65,91% 65,75% 65,78%
Gaussian NB 61,54% 62,43% 63,85% 60,45%
Multinomial NB 63,43% 62,89% 54,67% 55,41%
Complement NB 66,88% 63,71% 63,06% 63,35%
Bernoulli NB 63,17% 61,55% 63,14% 61,13%
SVM 69,95% 67,67% 66,27% 66,86%
Decision Tree 56,78% 54,17% 54,96% 54,38%
Random Forest 66,30% 64,56% 61,47% 62,24%

Figure 3 is the result of the confusion matrix with the SVM algorithm, where through the test data of 20% (1,534 instances) the data
predicted as negative sentiment (0) resulted in true as negative as 570 instances, but produced false as neutral (1) as many as 97 instances
and false as positive (2) as many as 124 instances. Likewise, the data predicted as neutral sentiment (1) produces true as neutral for 151
instances, but produces false as negative (0) for 66 instances and false as positive (2) for 43 instances. Then in the data predicted as positive
sentiment (2) resulted in true as positive for 352 instances, but produced false as negative (0) for 93 instances and false as neutral (1) for
38 instances.

So the actual data generated for negative sentiment (0) was 729 instances, neutral sentiment (1) was 286 instances, and positive sentiment
(2) was 519 instances. However, the prediction data generated for negative sentiment (0) was 791 instances, neutral sentiment (1) was 260
instances, and positive sentiment (2) was 483 instances. Based on these results, there was a prediction error for negative sentiment of 62
instances, neutral 26 instances, and positive 36 instances.

Fig. 3: Confusion Matrix SVM Result

4. Conclusion
Based on the performance metric value generated from the model evaluation, it is proven in this study that tuning the hyperparameter to
the classification algorithm can maximize the overall model performance. However, from the results obtained, the best model, namely the
SVM classification algorithm with an accuracy of 69.95%, can still be improved. These results can occur due to the existence of imbalance
class data, or the use of feature extraction that is not suitable for this data. But overall, when viewed from the loading process and the order
of use of methods, it can be said that it is quite good and maximum, which means that where it can be concluded, namely hyperparameter
tuning with GridSearchCV and then 10-fold cross validation is able to collaborate well on classification algorithms such as Naïve Bayes
Classifier, Support Vector Machine, Decision Tree, and Random Forest.

References
[1] E. Nugrohosudin, “Kedudukan Kepala Otorita Ibu Kota Nusantara,” J. Legis., vol. 5, no. 2, pp. 79–90, 2022.
[2] A. K. Santoso, A. Noviriandini, A. Kurniasih, B. D. Wicaksono, and A. Nuryanto, “Klasifikasi Persepsi Pengguna Twitter Terhadap Kasus Covid-
19 Menggunakan Metode Logistic Regression,” JIK (Jurnal Inform. dan Komputer), vol. 5, no. 2, pp. 234–241, 2021.
[3] A. Supian, B. Tri Revaldo, N. Marhadi, L. Efrizoni, and R. Rahmaddeni, “Perbandingan Kinerja Naïve Bayes Dan Svm Pada Analisis Sentimen
Twitter Ibukota Nusantara,” J. Ilm. Inform., vol. 12, no. 01, pp. 15–21, 2024.
[4] P. Arsi and R. Waluyo, “Analisis Sentimen Wacana Pemindahan Ibu Kota Indonesia Menggunakan Algoritma Support Vector Machine (SVM),” J.
Teknol. Inf. dan Ilmu Komput., vol. 8, no. 1, pp. 147–156, 2021.
[5] C. Mulia and A. Kurniasih, “Teknik SMOTE untuk Mengatasi Imbalance Class dalam Klasifikasi Bank Customer Churn Menggunakan Algoritma
Naïve bayes dan Logistic Regression,” Pros. Semin. Ilm. Nas. Online Mhs. Ilmu Komput. dan Apl., pp. 552–559, 2023.
[6] W. F. Hidayat, T. Asra, and A. Setiadi, “Klasifikasi Penyakit Daun Kentang Menggunakan Model Logistic Regression,” Indones. J. Softw. Eng., vol.
8, no. 2, pp. 173–179, 2022.
[7] I. Irmawati, H. Hermanto, E. H. Juningsih, S. Rahmatullah, and F. Aziz, “Prediksi Lama Tinggal Pasien Rawat Inap Di Rumah Sakit Pada Masa
Pandemi Covid-19 Menggunakan Metode Ensemble Learning Dan Decission Tree,” J. Inform. Kaputama, vol. 5, no. 2, pp. 391–397, 2021.
[8] E. Priyanti, “Penerapan Decision Tree Untuk Klasifikasi Tingkat Pendapatan,” IJCIT (Indonesian J. Comput. Inf. Technol., vol. 7, no. 1, pp. 7–12,
814 Journal of Artificial Intelligence and Engineering Applications

2022.
[9] N. Wuryani and S. Agustiani, “Random Forest Classifier untuk Deteksi Penderita COVID-19 berbasis Citra CT Scan,” J. Tek. Komput. AMIK BSI,
vol. 7, no. 2, pp. 187–193, 2021.
[10] Fajri Koto and Gemala Y. Rahmaningtyas, “InSet Lexicon: Evaluation of a Word List for Indonesian Sentiment Analysis in Microblogs,” 2017 Int.
Conf. Asian Lang. Process., pp. 391–394, 2017.
[11] M. P. Pulungan, A. Purnomo, and A. Kurniasih, “Penerapan SMOTE untuk Mengatasi Imbalance Class dalam Klasifikasi Kepribadian MBTI
Menggunakan Naive Bayes Classifier,” J. Teknol. Inf. dan Ilmu Komput., vol. 10, no. 7, pp. 1493–1502, 2023.
[12] Fatihah Rahmadayana and Yuliant Sibaroni, “Sentiment Analysis of Work from Home Activity using SVM with Randomized Search Optimization,”
J. RESTI (Rekayasa Sist. dan Teknol. Informasi), vol. 5, no. 5, pp. 936–942, 2021.
[13] A. B. Pohan, Irmawati, and A. Kurniasih, “Penerapan Teknik SMOTE dalam Memprediksi Kegagalan Mesin Menggunakan Support Vector Machine
dan Logistic Regression,” J. Inform. Kaputama, vol. 7, no. 2, pp. 181–187, 2023.