Sentiment Classification of Fuel Price Increase With Gated Recurrent Unit (GRU) and FastText
Aditya Andar Rahim*, Yuliant Sibaroni, Sri Suryani Prasetiyowati School of Computing, Telkom University, Bandung, Indonesia
Email: 1,*[email protected], 2[email protected], 3[email protected] Correspondence Author Email: [email protected]
Abstract-The government usually implements a policy of increasing fuel prices and reducing subsidized fuel every year. Rising fuel prices have had a mixed impact on society. The rapid development of information technology has led to easy access and an increase in the number of internet users. Social media platforms, such as Twitter, are widely used by people to express themselves in everyday life. Through this social media, the public can submit reviews regarding public policies implemented by the government regarding fuel prices. The reviews submitted varied, ranging from positive, neutral to negative. Sentiment analysis can analyze the types of reviews submitted by people, including positive, negative, or neutral. This research uses Gated Recurrent Unit and FastText feature expansion to classify sentiments related to rising fuel prices on Twitter. This system was developed through several stages, namely data crawling, data labeling, data initial processing, feature expansion, classification, and evaluation. This study aims to determine the classification performance using Gated Recurrent Unit and FastText. The data used was 8,635, and the highest accuracy reached 90.15% with an F1 score of 90.06%. The research results may help the government in determining how individuals feel about fuel price increases. By understanding public sentiment, the government can reevaluate its policies or even establish new ones that serve the public interest.
Keywords: Sentiment Classification; Fuel Price; Twitter; GRU; FastText; SMOTE
1. INTRODUCTION
Fuel utilization has become an important factor in facilitating human transportation. Apart from being used for fuel in personal transportation equipment, business entities also use BBM for the transportation of raw materials and goods delivery services. Every year the government usually implements a policy of increasing fuel prices and reducing fuel subsidies [1]. The increase in fuel prices has a broad impact, not only on transportation but also on the pricing of basic necessities. This phenomenon causes an increase in production costs, a decrease in individual purchasing power, and a spike in the unemployment rate [2].
Regarding the government's policy on increasing fuel prices, the public expressed their views through reviews on various social media platforms. Rapid technological developments have made the internet easy to access and affordable, thereby increasing the number of users. According to the survey conducted by APJII in 2021, 77.02% of Indonesian people have utilized and engaged with the Internet. Social media is the most accessed platform when using the internet [3]. Social media can be used to obtain responses or feedback from the public on public policies taken by the government [4]. Social medicine is a digital space created by and for society and provides a conducive environment to interact and network at various levels [5].
Twitter is a microblogging and social media service that is allowing users to upload posts with a limit of 240 characters which are called tweets [6]. Today, people tend to actively share their opinions about public policies using Twitter. Social media, such as Twitter, have become very significant in daily life [7].
Public sentiment on specific topics can be known and measured by sentiment analysis to determine the right message, intervention, or even policy. Sentiment analysis is defined as an automated process using a computer to determine reviews about a certain topic [8]. This makes sentiment analysis widely utilized in multiple fields, such as scientific, social, and commercial applications. The rapid growth of information on social media increases the urgency of sentiment analysis [9]. Public sentiment absorbed from social media is usually used to study public review patterns that can be used as a benchmark for future government policies and improve current service performance [10].
Sentiment analysis can be performed with deep learning. Deep learning is an application of Artificial Neural Networks (ANNs) that learn specific tasks using the network of layers. Deep learning can do much more learning or representation power of neural networks, better than its previous which only had one or two layers [11].
Research [12] uses LSTM and GloVe word embedding to perform sentiment analysis on Twitter. Another study [13] compared and combined the CNN and LSTM deep learning methods. Meanwhile, A. Hidayatullah et al. [14]
conducted a sentiment analysis in the 2019 presidential election using LSTM, CNN, and a combination of CNN and LSTM, a combination of GRU and LSTM, and Bidirectional LSTM.
Word embedding is the process of grouping words that have the same meaning into the closest position in one volume of vector space [15]. The word embedding process is required in the NLP deep learning model. While FastText is one of the methods capable to perform word embedding. E. Dharma et al. [16] conducted a study that compared several word embedding methods and found that FastText outperformed Word2Vec and GloVe.
One deep learning architecture that can be used is a Recurrent Neural Network. RNN is one of the most widely used architectures for handling variable-length text [17]. Studies conducted by Y. Wazery et al. [18] using K-Nearest Neighbor, Decision Tree, Naïve Bayes, Support Vector Machine, and RNN-LSTM, aims to compare
machine learning and deep learning approach to performing sentiment analysis. The conclusion obtained is that RNN-LSTM has better accuracy than other machine learning approaches. Using the Naïve Bayes method, A.
Isnain et al. [19] analyzed public sentiment regarding government policies throughout the Covid-19 pandemic using the Naïve Bayes method. The dataset used is 1.823 tweets about government policies during the Covid-19 pandemic. Meanwhile [10] assessed sentiment analysis of Omnibus Law policies using Naïve Bayes, SVM, and decision tree. In this research, it was concluded that SVM outperformed other machine learning methods with 91.80% accuracy.
Research conducted by A. Yadav et al. in 2020 [20] discusses sentiment analysis in deep learning using LSTM, GRU, and DBN. It states that the GRU is recognized as LSTM not having output gate, therefore GRU also called as variant of LSTM. Another study by S. Yang et al. in 2020 [21] compared LSTM and GRU, concluding that GRU is superior to LSTM in terms of cost-performance ratio. GRU has one less gate than LSTM, and this can reduce matrix multiplication and conserve significant time without sacrificing performance. Meanwhile, E.
Dharma et al. [16] compared several word embedding methods, namely Word2Vec, GloVe, and FastText, using a dataset of 20 newsgroups. FastText outperformed all other methods. FastText has the ability to generate text representations not found in the training data. In other words, FastText is able to deal with out-of-vocabulary problems. This capability allows FastText to outperform all other methods. While J. Lee et al. [22] used Gated Recurrent Unit as a classification method and obtained an accuracy of 87.2% accuracy and an F1 score of 91.74%.
This research will use the Gated Recurrent Unit model and FastText feature expansion for the classification of sentiment on fuel price increases from Twitter. The research results may help the government in determining how individuals feel about fuel price increases. By understanding public sentiment, the government can reevaluate its policies or even establish new ones that serve the public interest.
2. RESEARCH METHODOLOGY
2.1 System Design
The system was designed as a classification system using Gated Recurrent Unit. In this system, FastText will be used as a feature expansion method. Meanwhile, TF-IDF will be used for feature extraction. After the system has been built, classification sentiment results will be evaluated using a confusion matrix. Figure 1 illustrates an extensive representation of the system that is intended to be constructed. The first step is to crawl the data from Twitter. The process continued to data labeling, where data will be labeled with positive, neutral, or negative.
Labeled data will be processed through preprocessing to reduce the amount of irrelevant information that may interfere with the classification process. Feature extraction will be performed using TF-IDF then FastText feature expansion will be performed. Cross-validation using K-Fold was performed on the data splitting step, and then data modeling and data classification. The average of K-Fold result will be used to evaluate the result.
Figure 1. System of Classification Sentiment 2.2 Data Labeling
The data labelled manually by three person using majority vote. Features on tweets will be used to determine sentiment labels. The data labeling process's result will contain its sentiment, whether it is negative, neutral, or positive.
Table 1. Labeling Result Example
Tweet Person 1 Person 2 Person 3 Result
Bbm naik semua
barang naik -1 -1 -1 -1
Mantap subsidi BBM
dialihkan untuk BLT 1 0 1 1
Pertalite pertamina 0 0 0 0
Semoga BBM bisa
kembali normal 0 1 0 0
2.3 Pre-processing Data
Upon completion of data crawling, the raw data will undergo preprocessing. Preprocessing is a crucial step in (NLP) that aims to reduce the amount of irrelevant information that may interfere classification process. This study's preprocessing stages include a series of essential techniques, including data cleaning, case folding, tokenization, data normalization, stopword removal, and stemming.
The first step in preprocessing data is data cleaning, which removes punctuation, numbers, emoticon, and website link. The second step is case folding, a process of replacing all the letters to be lowercase. The third step is data tokenization, a method to split sentences into smaller units called tokens. The fourth step is data normalization, a process to normalize words based on KBBI. The fifth step is stopword removal, an approach to remove unnecessary words without implication on classification. The sixth step, or the last step, is stemming, a process of transforming words into their base form. An example is shown in Table 2.
Table 2. Pre-processing Data
Pre-processing Tweet Result
Data Cleaning Aku makin pusing belum siap sm harga bensin naik sementara gaji engga. Harga naik tp antrian makin panjang. Jd lama nunggu cma buat bensin.
Aku makin pusing belum siap sm harga bensin naik sementara gaji engga Harga naik tp antrian makin panjang Jd lama nunggu cma buat bensin
Case Folding Aku makin pusing belum siap sm harga bensin naik sementara gaji engga Harga naik tp antrian makin panjang Jd lama nunggu cma buat bensin
aku makin pusing belum siap sm harga bensin naik sementara gaji engga harga naik tp antrian makin panjang jd lama nunggu cma buat bensin
Data Tokenization aku makin pusing belum siap sm harga bensin naik sementara gaji engga harga naik tp antrian makin panjang jd lama nunggu cma buat bensin
[’aku’], [‘makin’], [‘pusing’], [‘belum’], [‘siap’], [‘sm’], [‘harga’], [‘bensin’], [‘naik’], [‘sementara’], [‘gaji’], [‘engga’], [‘harga’], [‘naik’], [‘tp’], [‘antrian’], [‘makin’], [‘panjang’], [‘jd’], [‘lama’], [‘nunggu’], [‘cma’], [‘buat’], [‘bensin’]
Data Normalization [’aku’], [‘makin’], [‘pusing’], [‘belum’], [‘siap’], [‘sm’], [‘harga’], [‘bensin’], [‘naik’], [‘sementara’], [‘gaji’], [‘engga’], [‘harga’], [‘naik’], [‘tp’], [‘antrian’], [‘makin’], [‘panjang’], [‘jd’], [‘lama’], [‘nunggu’], [‘cma’], [‘buat’], [‘bensin’]
[’aku’], [‘makin’], [‘pusing’], [‘belum’], [‘siap’], [‘sama’], [‘harga’], [‘bensin’], [‘naik’], [‘sementara’], [‘gaji’], [‘tidak’], [‘harga’], [‘naik’], [‘tapi’], [‘antrian’], [‘makin’], [‘panjang’], [‘jadi’], [‘lama’], [‘nunggu’], [‘cuma’], [‘untuk’], [‘bensin’]
Stopword Removal [’aku’], [‘makin’], [‘pusing’], [‘belum’], [‘siap’], [‘sama’], [‘harga’], [‘bensin’], [‘naik’], [‘sementara’], [‘gaji’], [‘tidak’], [‘harga’], [‘naik’], [‘tapi’], [‘antrian’], [‘makin’], [‘panjang’], [‘jadi’], [‘lama’], [‘nunggu’], [‘cuma’], [‘untuk’], [‘bensin’]
[’aku’], [‘makin’], [‘pusing’], [‘siap’], [‘harga’], [‘bensin’], [‘naik’], [‘gaji’], [‘harga’], [‘naik’], [‘antrian’], [‘makin’], [‘panjang’], [‘jadi’], [‘lama’], [‘nunggu’], [‘cuma’], [‘untuk’], [‘bensin’]
Stemming [’aku’], [‘makin’], [‘pusing’],
[‘siap’], [‘harga’], [‘bensin’], [‘naik’], [‘gaji’], [‘harga’], [‘naik’], [‘antrian’], [‘makin’], [‘panjang’], [‘jadi’], [‘lama’], [‘nunggu’], [‘cuma’], [‘untuk’], [‘bensin’]
[’aku’], [‘makin’], [‘pusing’], [‘siap’], [‘harga’], [‘bensin’], [‘naik’], [‘gaji’], [‘harga’], [‘naik’], [‘antri’], [‘makin’], [‘panjang’], [‘jadi’], [‘lama’], [‘tunggu’], [‘cuma’], [‘untuk’], [‘bensin’]
2.4 Feature Extraction TF-IDF
TF-IDF is a method used in this study to perform feature extraction by calculating the weight of every word. TF- IDF multiplies the Term Frequency value with the Inverse Document Frequency value to calculate the weight of every word [23]. Term frequency measures the frequency of word occurrences, whereas Inverse Document Frequency measures the significance of a word [24].
2.5 Feature Expansion FastText
Feature expansion is a technique for overcoming out-of-vocabulary problems by substituting the word with another word with a similar meaning [25]. One method that can be used to perform feature expansion is FastText. FastText is a method to represent words to vectors. FastText was initiated by Facebook as an open-source project [26]. In FastText, words with the same root are grouped together. This implies that FastText can organize lexically related words [27]. FastText becomes a fast and effective method to represent words as vectors [26]. When performing word embedding, FastText considers the internal structure of a word instead of learning its’ representation [26].
Figure 2 provides an illustration of FastText.
Figure 2. FastText Model Illustration
Sentence representation on FastText is formed from the average of N-Gram features that adhere to the hidden variable. The hidden variable can be used as a feature (input) in a classification model [28].
2.6 SMOTE
SMOTE is an oversampling method that is used to overcome imbalance class distribution on a dataset. This technique works by producing an original sample from the underrepresented class to achieve class balance in the dataset. New data is created based on two minority data samples through linear interpolation [29]. This technique reduces overfitting problems caused by the random oversampling method, balancing the class distribution and enhancing the classifier's generalization [29]. The result of SMOTE process can be shown in Table 3.
Table 3. SMOTE Process Result
Label Amount
Negatif 3.980
Netral 3.980
Positif 3.980
2.7 GRU Modelling
Gated recurrent unit (GRU) is one of the deep learning models. Cho et al. proposed GRU to address the vanishing gradient issue when using a standard RNN [30]. The illustration of GRU can be shown in Figure 3.
Figure 3. Gated Recurrent Unit
GRU has input and output similar to RNN vanilla but has similar internal structure to LSTM [21]. GRU has two gates, different from LSTM, which has three gates. Gates on GRU are update (z) gate and reset (r) gate.
On GRU, forget gate and input gate of LSTM are combined into update gate. Both gates coordinate the unit’s information flow, with the update gate determining how much data must be saved and passed to the next output.
The reset gate, in contrary, identifies and eliminates unnecessary information. In contrast to LSTM, which uses a memory unit and output gate to control the flow of data to the next layer, GRU bypasses that and directly exposes hidden contents without control to the next layer [22]. The equation of GRU shows below:
rt= σ(Wrhht−1+ wrxxt+ br) (1)
zt= σ(wzhht−1+ wzxxt+ bz) (2)
h = tanh (whh(rt∗ ht−1) + wxhxt+ bh]) (3)
h̃t= (1 − zt) ∗ ht−1+ zt∗ h̃t (4)
ot= σ(wo. ht) (5) rt is reset gate unit, zt is update gate. ht is the hidden state, and xt is input on t time.
2.8 Evaluation Matrix
In this step, the evaluation will be performed using a confusion matrix. The confusion matrix result is directly proportional to the quality of the classification process. The values that will be measured are accuracy and F1 score. Confusion matrix is a table that measures classification algorithm performance [31]. Confusion matrix can be shown in Table 4:
Table 4. Confusion Matrix Actual Class Prediction Class
Positive Negative Positive TP (True Positive) FN (False Negative) Negative FP (False Positive) TN (True Negative)
The confusion matrix can be utilized to determine values such as precision and F1 score. The following formula can be used to evaluate performance:
a. Accuracy
Accuracy is defined as comparison of total correct prediction, which is true positive and true negative, and total data amount.
accuracy = TP+TN
(TP+FP+FN+TN) (6)
b. Precision
Precision is defined as comparison of total correct positive prediction (True Positive) and all data predicted as positive.
Precision = TP
TP+FP (7)
c. Recall
Recall is defined as comparison of total accurate positive prediction (True Positive) and total of actual positive.
Recall = TP
TP+FN (8)
d. F1 score
F1 score is a metric that quantifies the balance between precision and recall.
F1 score = 2×(recall×precision)
(recall+precision) (9)
TP is True Positive, FP is False Positive. FN is False Negative, and TN is True Negative.
3. RESULT AND DISCUSSION
3.1 Data Distribution
Dataset in this study totaled 8.365 tweets in Indonesian, focusing on fuel price increases. Figure 4 shows the labeling result from the data:
Figure 4. Data Distribution 3.2 Test Result
Before executing the classification, the dataset will go through a few processes, including preprocessing, TF-IDF of feature extraction, and FastText of feature expansion. The results will be evaluated based on acuracy and F1
0 1000 2000 3000 4000 5000
Positive Neutral Negative
score. First scenario is conducted to determine the best baseline by comparing epoch values. Epoch values that will be compared are 5, 10, and 12. Table 5 shows that epoch 10 performs best with 57.39% accuracy and 57.11%
F1 score.
Table 5. Result of Baseline Experiment using GRU
Epoch Accuracy (%) F1 Score (%)
5 57.12 57.10
10 57.39 57.11
12 56.97 56.89
In the second scenario, classification was performed using the Gated Recurrent Unit model with ten epochs.
This scenario also adds TF-IDF feature extraction with 5.000 max features. As shown in Table 6, there are significant accuracy and F1 score improvement when feature extraction is added. Accuracy improves to 80.86%
and the F1 score improves to 81.58%.
Table 6. TF-IDF result
Method Accuracy (%) F1 Score (%)
GRU 57.39 57.11
GRU+TF-IDF 80.86 81.58
In the third scenario, the classification process was enhanced using FastText feature expansion. This enhancement improves accuracy to 84.54% and F1 score to 84.88%. The comparison is shown in Table 7.
Table 7. FastText result
Method Accuracy (%) F1 Score (%)
GRU + TF-IDF 80.86 81.58
GRU + TF-IDF + FASTTEXT 84.54 84.88
In the last scenario, SMOTE was added to the classification process to handle imbalance class on the dataset. The result of adding SMOTE into the classification process is shown in Table 8.
Table 8. SMOTE result
Method Accuracy (%) F1 Score (%)
GRU + TF-IDF+FASTTEXT 84.54 84.88
GRU + TF-IDF + FASTTEXT+SMOTE
90.15 90.06
3.3 Analysis of Test Results
Figure 5. Accuracy Graph
Figure 6. F1 Score Graph 0
50 100
1st scenario 2nd scenario 3rd scenario 4th scenario
Accuracy
Accuracy
0 50 100
1st scenario 2nd scenario 3rd scenario 4th scenario
F1 Score
F1 Score
Figure 5 and 6 shows the accuracy and F1 score result of all scenarios. Based on all four scenarios in this study, it was found that every test scenario positively impacts classification performance. The observed relationship between accuracy and F1 score is positively correlated, as evidenced by the upward trend in both Table 9 and Table 10. The accuracy improves from 57.39 percent in the first scenario to 90.15 percent in the last scenario. The accuracy has increased by 57.08 percent overall. The F1 score also increases from 57.11 percent in the first scenario to 90.06 percent in the last scenario. The F1 score has improved by 57.69% overall.
In the first scenario, some epoch values are compared to find the best epoch value, and it was found that epoch 10 has the best accuracy and F1 score. Epoch 10 then used as the baseline. In the second scenario, TF-IDF enabled the model to have knowledge about the weight of words based on their frequency and significance [24];
consequently, the accuracy and F1 score improved. In the third scenario, FastText was also added, allowing the model to overcome out-of-vocabulary problems by substituting a word with a similar meaning [25]; consequently, the accuracy and F1 also improved. SMOTE was implemented in the final scenario to increase the number of data for the underrepresented class. The dataset was 3.331 negative, 3.980 neutral, and 1.324 positive data. By implementing SMOTE in the classification process, the dataset was modified to contain 3.980 negative, 3.980 neutral, and 3.980 positive data; as a result of more balanced data, the accuracy and F1 score improved.
From four scenarios, the scenario with the best performance is the classification process using epoch 10, TF-IDF, FastText, and SMOTE oversampling with 90.15% accuracy and 90.06% F1 score. Compared with a previous study where J. Lee et al. [22] used Gated Recurrent Unit as a classification method and obtained accuracy of 87.2% accuracy and F1 score of 91.74%, this study shows accuracy and F1 score improvement.
4 CONCLUSION
This study conducted sentiment classification on Twitter regarding fuel price increases. The datasets used in this study were obtained from Twitter and consisted of 8.635 tweets labeled positive, neutral, or negative and have undergone a preprocessing process. The test consisted of four scenarios with GRU. Four classification scenarios consist of a baseline experiment to find the best epoch value, applying TF-IDF feature extraction, applying FastText feature expansion, and applying SMOTE oversampling. This study successfully performs sentiment classification using GRU and FastText. The fourth scenario was the best result, combining TF-IDF, FastText, and SMOTE. This scenario achieves 90.15% accuracy and 90.06% F1 score. Based on this result, it can be concluded that TF-IDF feature extraction, FastText feature expansion, and SMOTE oversampling implementation on classification significantly improve accuracy and F1 score. For future research, increasing the dataset amount and balancing the class distribution on dataset is recommended to achieve better performance.
REFERENCES
[1] F. Sandi, “Salah Satu Keputusan Berat Jokowi di 2022, Naikkan Harga BBM,” Jan. 01, 2023.
https://www.cnbcindonesia.com/news/20230101095550-4-401827/salah-satu-keputusan-berat-jokowi-di-2022-naikkan- harga-bbm (accessed Jun. 15, 2023).
[2] A. Agus Hidayat, L. Rahmawati, and A. M. Arlina, “Analisis Permintaan Transportasi Umum Modern Saat Kenaikan Harga Bbm Perspektif Permintaan Islam,” Al Iqtishod: Jurnal Pemikiran dan Penelitian Ekonomi Islam, vol. 5, no. 1, pp.
51–65, Jun. 2023, doi: 10.51339/iqtis.v5i1.930.
[3] Asosiasi Penyelenggara Jasa Internet Indonesia/APJII., “Indonesian Internet Profile 2022,” Jun. 2022.
[4] K. H. Manguri, R. N. Ramadhan, and P. R. Mohammed Amin, “Twitter Sentiment Analysis on Worldwide COVID-19 Outbreaks,” Kurdistan Journal of Applied Research, vol. 5, no. 3, pp. 54–65, May 2020, doi: 10.24017/covid.8.
[5] K. K. Kapoor, K. Tamilmani, and N. P. Rana, “Advances in Social Media Research: Past, Present and Future,” Information Systems Frontiers, vol. 20, pp. 531–558, Jun. 2018, doi: 10.1007/s10796-017-9810-y.
[6] H. Rosenberg, S. Syed, and S. Rezaie, “The Twitter pandemic: The critical role of Twitter in the dissemination of medical information and misinformation during the COVID-19 pandemic,” Canadian Journal of Emergency Medicine, vol. 22, no.
4. Cambridge University Press, pp. 418–421, Jul. 01, 2020. doi: 10.1017/cem.2020.361.
[7] L. Nemes and A. Kiss, “Social media sentiment analysis based on COVID-19,” Journal of Information and Telecommunication, vol. 5, pp. 1–15, 2020, doi: 10.1080/24751839.2020.1790793.
[8] C. A. Melton, O. A. Olusanya, N. Ammar, and A. Shaban-Nejad, “Public sentiment analysis and topic modeling regarding COVID-19 vaccines on the Reddit social media platform: A call to action for strengthening vaccine confidence,” J Infect Public Health, vol. 14, no. 10, pp. 1505–1512, Oct. 2021, doi: 10.1016/j.jiph.2021.08.010.
[9] L. Yue, W. Chen, X. Li, W. Zuo, and M. Yin, “A survey of sentiment analysis in social media,” Knowl Inf Syst, vol. 60, no. 2, pp. 617–663, Aug. 2019, doi: 10.1007/s10115-018-1236-4.
[10] E. A. Sukma, A. N. Hidayanto, A. I. Pandesenda, A. N. Yahya, P. Widharto, and U. Rahardja, “Sentiment Analysis of the New Indonesian Government Policy (Omnibus Law) on Social Media Twitter,” in International Conference on Informatics, Multimedia, Cyber and Information System (ICIMCIS), Institute of Electrical and Electronics Engineers Inc., Nov. 2020, pp. 153–158. doi: 10.1109/ICIMCIS51567.2020.9354287.
[11] L. Zhang, S. Wang, and B. Liu, “Deep learning for sentiment analysis: A survey,” Wiley Interdiscip Rev Data Min Knowl Discov, vol. 8, no. 4, Jul. 2018, doi: 10.1002/widm.1253.
[12] R. Monika, S. Deivalakshmi, and B. Janet, “Sentiment Analysis of US Airlines Tweets using LSTM/RNN,” 9th International Conference on Advanced Computing (IACC), pp. 92–95, 2019, doi: 10.1109/IACC48062.2019.8971592.
[13] S. Kamiş and D. Goularas, “Evaluation of Deep Learning Techniques in Sentiment Analysis from Twitter Data,” in Proceedings - 2019 International Conference on Deep Learning and Machine Learning in Emerging Applications, Deep- ML 2019, Institute of Electrical and Electronics Engineers Inc., Aug. 2019, pp. 12–17. doi: 10.1109/Deep- ML.2019.00011.
[14] A. F. Hidayatullah, S. Cahyaningtyas, and A. M. Hakim, “Sentiment Analysis on Twitter using Neural Network:
Indonesian Presidential Election 2019 Dataset,” IOP Conf Ser Mater Sci Eng, vol. 1077, no. 1, p. 012001, Feb. 2021, doi:
10.1088/1757-899x/1077/1/012001.
[15] K. Pasupa and T. Seneewong Na Ayutthaya, “Thai sentiment analysis with deep learning techniques: A comparative study based on word embedding, POS-tag, and sentic features,” Sustain Cities Soc, vol. 50, Oct. 2019, doi:
10.1016/j.scs.2019.101615.
[16] E. M. Dharma, F. Lumban Gaol, H. Leslie, H. S. Warnars, and B. Soewito, “The Accuracy Comparison Among Word2Vec, GloVe, and FastText towards Convolution Neural Network (CNN) Text Classification,” J Theor Appl Inf Technol, vol. 100, no. 2, pp. 349–359, 2022.
[17] G. S. N Murthy, S. Rao Allu, B. Andhavarapu, M. Bagadi, and M. Belusonti, “Text based Sentiment Analysis using LSTM; Text based Sentiment Analysis using LSTM,” International Journal of Engineering Research & Technology (IJERT), vol. 9, no. 5, pp. 299–303, 2020.
[18] Y. Wazery, H. Mohammed, and E. Houssein, “Twitter Sentiment Analysis using Deep Neural Network,” International Computer Engineering Conference (ICENCO), pp. 177–182, 2018, doi: 10.1109/ICENCO.2018.8636119.
[19] A. R. Isnain, N. S. Marga, and D. Alita, “Sentiment Analysis Of Government Policy On Corona Case Using Naive Bayes Algorithm,” IJCCS (Indonesian Journal of Computing and Cybernetics Systems), vol. 15, no. 1, p. 55, Jan. 2021, doi:
10.22146/ijccs.60718.
[20] A. Yadav and D. K. Vishwakarma, “Sentiment analysis using deep learning architectures: a review,” Artif Intell Rev, vol.
53, no. 6, pp. 4335–4385, Aug. 2020, doi: 10.1007/s10462-019-09794-5.
[21] S. Yang, X. Yu, and Y. Zhou, “LSTM and GRU Neural Network Performance Comparison Study: Taking Yelp Review Dataset as an Example,” in International Workshop on Electronic Communication and Artificial Intelligence (IWECAI), Institute of Electrical and Electronics Engineers Inc., Jun. 2020, pp. 98–101. doi: 10.1109/IWECAI50956.2020.00027.
[22] J. S. Lee, D. Zuba, and Y. Pang, “Sentiment analysis of Chinese product reviews using gated recurrent unit,” in Fifth International Conference on Big Data Computing Service and Applications, Institute of Electrical and Electronics Engineers Inc., Apr. 2019, pp. 173–181. doi: 10.1109/BigDataService.2019.00030.
[23] A. R. Isnain, J. Supriyanto, and M. P. Kharisma, “Implementation of K-Nearest Neighbor (K-NN) Algorithm For Public Sentiment Analysis of Online Learning,” IJCCS (Indonesian Journal of Computing and Cybernetics Systems), vol. 15, no.
2, pp. 121–130, Apr. 2021, doi: 10.22146/ijccs.65176.
[24] P. H. Prastyo, I. Ardiyanto, and R. Hidayat, “Indonesian Sentiment Analysis: An Experimental Study of Four Kernel Functions on SVM Algorithm with TF-IDF,” in International Conference on Data Analytics for Business and Industry:
Way Towards a Sustainable Economy (ICDABI), Institute of Electrical and Electronics Engineers Inc., Oct. 2020, pp. 1–
6. doi: 10.1109/ICDABI51230.2020.9325685.
[25] M. P. K. Dewi and E. B. Setiawan, “Feature Expansion Using Word2vec for Hate Speech Detection on Indonesian Twitter with Classification Using SVM and Random Forest,” JURNAL MEDIA INFORMATIKA BUDIDARMA, vol. 6, no. 2, pp. 979–988, Apr. 2022, doi: 10.30865/mib.v6i2.3855.
[26] N. Nedjah, I. Santos, and L. de Macedo Mourelle, “Sentiment analysis using convolutional neural network via word embeddings,” Evol Intell, vol. 15, no. 4, pp. 2295–2319, Dec. 2019, doi: 10.1007/s12065-019-00227-4.
[27] S. Thavareesan and S. Mahesan, “Sentiment Lexicon Expansion using Word2vec and fastText for Sentiment Prediction in Tamil texts,” in 2020 Moratuwa Engineering Research Conference (MERCon), 2020, pp. 272–276. doi:
10.1109/MERCon50084.2020.9185369.
[28] S. Tiun, U. A. Mokhtar, S. H. Bakar, and S. Saad, “Classification of functional and non-functional requirement in software requirement using Word2vec and fast Text,” J Phys Conf Ser, vol. 1529, no. 4, Jun. 2020, doi: 10.1088/1742- 6596/1529/4/042077.
[29] J. He, Y. Hao, and X. Wang, “An interpretable aid decision-making model for flag state control ship detention based on SMOTE and XGboost,” J Mar Sci Eng, vol. 9, no. 2, pp. 1–19, Feb. 2021, doi: 10.3390/jmse9020156.
[30] A. Dutta, S. Kumar, and M. Basu, “A Gated Recurrent Unit Approach to Bitcoin Price Prediction,” Journal of Risk and Financial Management, vol. 13, no. 2, p. 23, Feb. 2020, doi: 10.3390/jrfm13020023.
[31] P. Singh, N. Singh, K. K. Singh, and A. Singh, “Diagnosing of disease using machine learning,” in Machine Learning and the Internet of Medical Things in Healthcare, Elsevier, 2021, pp. 89–111. doi: 10.1016/B978-0-12-821229-5.00003-3.
