JURIKOM (Jurnal Riset Komputer), Vol. 9 No. 3, Juni 2022 e-ISSN 2715-7393 (Media Online), p-ISSN 2407-389X (Media Cetak) DOI 10.30865/jurikom.v9i3.4143 Hal 543−549 http://ejurnal.stmik-budidarma.ac.id/index.php/jurikom
Sentiment Analysis Of Development Jakarta-Bandung High-Speed Train Using Twitter Social Media With BNN Method
Arya Pratama Anugerah, Yuliant Sibaroni*
Faculty of Informatics, Informatics, Telkom University, Bandung, Indonesia Email: 1[email protected], 2,*[email protected]
Email Penulis Korespondensi: [email protected] Submitted 19-05-2022; Accepted 29-05-2022; Published 30-06-2022
Abstract
The Jakarta-Bandung high-speed train is one of the infrastructure development projects currently being carried out by the Indonesian government. The project is a large project that requires a long processing time and very large costs. Therefore, infrastructure development has reaped a lot of public opinions, both positive and negative. The purpose of writing this Final Project is to analyze sentiment on public opinion about the construction of the Jakarta-Bandung high-speed train. With data sourced from Twitter social media, the data will be analyzed in three classes, namely positive, negative, and neutral classes where the weighting will use the TF- IDF. The classification method used in this study is the Backpropagation Neural Network method. The best results were obtained in this study using a hyper tuning scenario with an accuracy of 74.56%.
Keywords: Jakarta-Bandung High-Speed Train; Sentiment Analysis; Twitter; TF-IDF; Backpropagation Neural Network
1. INTRODUCTION
High-speed trains consist of two components, namely high-speed trains and special high-speed lines. The combination of the two will result in fast trains that are twice as fast as driving by car and more comfortable than short-haul flights [1].
The Jakarta-Bandung high-speed train is a high-speed rail project in Indonesia that connects the capital city of Jakarta and the city of Bandung with 4 stopping stations Halim, Karawang, Padalarang, and Tegalluar [2][3].
The Jakarta-Bandung high-speed rail project has attracted a lot of attention and opinions from the public. For example, the opinion of this project will cause the subsidy budget to increase when the project is completed because there will be a transportation subsidy budget for KAI operations at the beginning of the high-speed train operation and high- speed train tickets later [4]. Or opinions about the many opportunities created by the presence of existing stations such as bringing in a lot of new investment and the development of supporting industries that will open up many new jobs [5].
The public also expressed their opinion about the Jakarta-Bandung high-speed rail project on social media such as Twitter.
Therefore, the author wants to analyze how public sentiment is on the Jakarta-Bandung high-speed rail project, especially on Twitter.
Sentiment analysis also called opinion mining, is a study that analyzes how a person's opinions, sentiments or feelings, behavior, and emotions towards an entity are expressed through writing [6]. In research [7], in the field of infrastructure, public opinion is indirectly involved in the feasibility study, construction, and post-evaluation of the project. The results of this study were that 47.7% of the messages were negative sentiments toward large hydro infrastructure projects using the lexicon-based classification method, an accuracy of 88.1% was obtained. Another research in the field of infrastructure is research [8], examining the sentiment analysis of the people of Surabaya towards Mrs. Tri Rismaharini's policy as the mayor of Surabaya regarding infrastructure issues, the results show that most of the people of Surabaya express negative sentiments with a focus on topics regarding electricity and toll roads as highways.
Research [9] raises the issue of President Jokowi's leadership which will be analyzed by sentiment on Twitter. As a result, 51% negative sentiment, 31% positive sentiment, and 18% neutral sentiment towards President Jokowi's leadership using the CNN classification method and also the expansion of the word2vec obtained 57% accuracy. Another research on sentiment analysis is on covid-19 using the Backpropagation Neural Network method, which obtained f1-score results of positive 0.77, negative 0.75, and neutral 0.5 using 100 hidden layers obtained 70% accuracy [10].
Therefore, the writer in this Final Project raises the issue regarding the construction of the Jakarta-Bandung high- speed train because there are still few studies that address this issue. By using the classification method, namely Backpropagation Neural Network. The advantage is that Backpropagation Neural Network can more easily formulate experiences and predictions and is very flexible in changing prediction rules and learning by adjusting the weights of each perceptron network [11]. Backpropagation Neural Network also has better performance than other classifications that do not include NN such as SVM [8]. The author also uses TF-IDF feature extraction because, in research [7], more detailed function features are needed to analyze data so that classification performance is better. To measure the performance of the method, the author uses a confusion matrix to find the accuracy and f1-score value of the model made.
2. RESEARCH METHODOLOGY
2.1 Research Stage
JURIKOM (Jurnal Riset Komputer), Vol. 9 No. 3, Juni 2022 e-ISSN 2715-7393 (Media Online), p-ISSN 2407-389X (Media Cetak) DOI 10.30865/jurikom.v9i3.4143 Hal 543−549 http://ejurnal.stmik-budidarma.ac.id/index.php/jurikom The stages carried out in sentiment analysis research on the construction of the Jakarta-Bandung high-speed train with social media Twitter using the BNN method can be described as shown in Figure 1.
Figure 1.Overview of Research Stages 2.2 TF-IDF
TF-IDF (Term Frequency Inverse Document Frequency) is a weighting matrix that measures the relationship between words where the weighting process combines two parts, namely Term Frequency (TF) to count the number of occurrences of each word in the document and Inverse Document Frequency (IDF) the number of related documents containing a certain word [12]. The TF-IDF feature is one of the extraction techniques, namely by assigning a value to each word in the training data, so that we can find out how important the word that represents the sentence is, it will be given a calculation value. The value of the calculation depends on the size of the frequency of occurrence of words in the document [13][14]. The calculation of TF-IDF can be done using the equation formula 1.
𝑊𝑖,𝑗= 𝑡𝑓𝑖,𝑗× 𝑖𝑑𝑓 = 𝑡𝑓𝑖𝑗× log (𝑁
𝑑𝑓𝑖) (1)
Keterangan:
𝑊𝑖,𝑗 = word weight t𝑗 untuk document d𝑖
𝑡𝑓𝑖,𝑗 = number of occurrences of words from t𝑗 in documents d𝑖 N = number of documents
𝑑𝑓𝑖 = number of documents containing searched word 2.3 Backpropagation Neural Network (BNN)
BNN is an algorithm to develop a model of ANN (Artificial Neural Network), in BNN there are 3 layers, namely the input layer where the data will be entered into the network, the hidden layer where the incoming data will be processed and the last is the output layer, where the results of the inputs have been processed [15]. algorithm Backpropagation is used for data training, this algorithm is also quite simple and very good to use even with complex data. The stages of the process of forming the Backpropagation Neural Network (BNN) algorithm can be summarized into several stages, the first is initializing the network, then forward propagating, Back Propagate Error, training the network, and the last is making predictions with the trained network [16].
Figure 2. Backpropagation Neural Network Architecture
JURIKOM (Jurnal Riset Komputer), Vol. 9 No. 3, Juni 2022 e-ISSN 2715-7393 (Media Online), p-ISSN 2407-389X (Media Cetak) DOI 10.30865/jurikom.v9i3.4143 Hal 543−549 http://ejurnal.stmik-budidarma.ac.id/index.php/jurikom 2.4 Grid Search
The stages of Hyperparameter tuning are the stages to determine the optimal parameters for the data and the prediction model used to produce predictions with the good ability [17]. In this study, hyperparameter tuning uses the grid search.
In the grid search the models will be trained using the cross-validation, namely by dividing the training data from the preprocessing stage into folds consisting of training data and folds, then calculating the average MSE score for each fold.
) obtained during the training process for each model from the hyperparameter combinations that have been formed [17].
the first step is to initialize hyperparameters to be optimized, then use the grid search to determine which hyperparameters are the most optimal The hyperparameters that will be optimized in the Backpropagation Neural Network in this study are batch size, epochs, and dropout rate each of which can be seen in table 1.
Table 1. Parameters Used in Grid Search
Hyperparameter Value
Batch size [ 8,16 ], [ 16,32 ], [ 32,64 ] Epochs [ 50,100 ], [ 100,150 ], [ 150,200 ] Dropout rate [ 0.3,0.4 ], [ 0.4,0.5 ], [ 0.5,0.6 ] 2.5 Performance Measurement
There are 2 objectives of performance measurement, the first is to measure the performance of the training model and the second is to compare two or more different performances and then determine which algorithm model is the best [8]. In this study, the author uses a confusion matrix as a tool for measuring performance. the confusion matrix is a measuring instrument in the form of a 2 x 2 matrix which serves to measure the amount of accuracy of the algorithm used [18].
Table 2. Confusion Matrix Actual Class
TRUE FALSE
Prediction Class
TRUE TP (True Positive) FP (False Positive) FALSE FN (False Negative) TN (True Negative)
3. RESULTS AND DISCUSSION
3.1 Crawling Data
The dataset is taken from social media Twitter using keywords such as #keretacepatJakarta-Bandung and #fast train development. For the collection method, the author uses the data crawling API Twitter by using the library programming language python such as tweepy. The results of data collection or crawling are in the form of a CSV file which contains tweets on Twitter regarding their opinions on the Jakarta-Bandung high-speed rail infrastructure.
3.2 Labeling Data
After the data regarding the construction of the Jakarta-Bandung high-speed train has been collected, then the data is labeled manually. The labeling consists of 2 labels, namely positive and negative labels.
Table 3. Twitter Dataset Labeling Example
Label Sentence
Positif Pembangunan ini dapat memberikan kesempatan baik untuk pelaku UMKM disekitar wilayah stasiun pembangunan kereta cepat
Negatif Kok begini??? KECEWA gw, pemerintah hanya membuang-buang anggaran saja dengan kereta cepat Jakarta-Bandung ini! @jokowi
3.3 Preprocessing
At the preprocessing, the data that has been taken before doing the classification first goes through the preprocessing, which is the stage to clean, delete, or change the data to simplify the classification process and make the classification results better. stages preprocessing are as follows:
a. Cleaning, the process of cleaning, deleting or filling in data if there are numbers, emoticons, URLs in text or missing values , or duplicates so as not to interfere with or reduce the results of the classification.
b. Case Folding, the process of uniforming the letters in the text into lowercase.
c. Normalization, the process of changing the format of the text in the data to return to the standard language
d. Stopword Removal, the process of removing words that have no meaning such as "this", "and", "yang" so as not to interfere with the classification process
e. Stemming, changing each word into its original form without affixes
JURIKOM (Jurnal Riset Komputer), Vol. 9 No. 3, Juni 2022 e-ISSN 2715-7393 (Media Online), p-ISSN 2407-389X (Media Cetak) DOI 10.30865/jurikom.v9i3.4143 Hal 543−549 http://ejurnal.stmik-budidarma.ac.id/index.php/jurikom Table 4. Stages and Results of Preprocessing
Tahap Results
After Labeling Kok begini??? KECEWA gw,pemerintah hanya membuang-buang anggaran saja dengan keretacepat Jakarta-Bandung ini! @jokowi
Cleaning Kok begini KECEWA gw,pemerintah hanya membuang-buang anggaran saja dengan keretacepat Jakarta-Bandung ini
Case Folding kok begini kecewa gw pemerintah hanya membuang-buang anggaran saja dengan kereta cepat jakarta-bandung ini
Normalization kok begini kecewa saya pemerintah hanya membuang-buang anggaran saja dengan kereta cepat jakarta-bandung ini
Stopword Removal kecewa saya pemerintah membuang-buang anggaran dengan kereta cepat jakarta-bandung ini temming kecewa saya pemerintah buang anggaran dengan kereta cepat jakarta-bandung
3.4 TF-IDF
After doing the preprocessing which is the stage to clean, delete, or change the data to simplify the classification process and make the classification results better, then enter the next stage feature extraction using the TF-IDF method. Examples of TF-IDF calculations for tweet can be seen in tables 5 and 6 with the keyword used is "fast train".
Table 5. TF-IDF Document Example
Dokumen Sentence
Dokumen 1 (D1) kereta keren dan canggih pertama indonesia Dokumen 2 (D2) indonesia punya kereta cepat Dokumen 3 (D3) Transportasi canggih salut bapak presiden
Table 6. TF-IDF Calculation Example
TF IDF weight
token keywords D1 D2 D3 df N/df Log(N/df) D1 D2 D3
kereta 1 1 1 0 2 1,5 0,176 0,176 0,176 0
keren 0 1 0 0 1 3 0,477 0,477 0 0
canggih 0 1 0 1 2 1,5 0,176 0,176 0 0,176
pertama 0 1 0 0 1 3 0,477 0,477 0 0
indonesia 0 1 1 0 2 1,5 0,176 0,176 0,176 0
punya 0 0 1 0 1 3 0,477 0 0,477 0
cepat 1 0 1 0 1 3 0,477 0 0,477 0
transportasi 0 0 0 1 1 3 0,477 0 0 0,477
salut 0 0 0 1 1 3 0,477 0 0 0,477
bapak 0 0 0 1 1 3 0,477 0 0 0,477
predisen 0 0 0 1 1 3 0,477 0 0 0,477
3.5 Test Results
There are several stages in this research, the first is crawling the data and labeling the data obtained from the social media Twitter. Then the data goes through the preprocessing and data split stages, then performs the feature extraction stage using TF-IDF. After that, the data classification stage uses the Backpropagation Neural Network method with a total of 5875 data. The following are several scenarios carried out in this study:
a. Scenario 1 : Testing the number of Max Features values on TF-IDF unigram b. Scenario 2 : Testing the number of epochs on the BNN model
c. Scenario 3 : Testing the comparison of hyperparameter tuning grid search on the Backpropagation Neural Network model
3.5.1 Scenario 1 Testing the number of Max Features on the TF-IDF unigram
In scenario 1, a test was conducted to find out how the influence of the number of max features parameters on the Backpropagation Neural Network model was carried out. The N-grams used are unigrams with max feature values of 500, 1000, 1500, 2000, 2500, 3000, 3500, 4000, 4500, 5000. The results of the test scenario 1 data are shown in table 7.
Table 7. TF-IDF Max Feature Test Results
Number of Max Features Accuracy(%) F1-Score
500 68,57 66,97
1000 69,28(+0,71) 66,74(-0,23)
1500 71,93(+2,65) 69,87(+3,13)
2000 70,96(-0,97) 69,21(-0,66)
JURIKOM (Jurnal Riset Komputer), Vol. 9 No. 3, Juni 2022 e-ISSN 2715-7393 (Media Online), p-ISSN 2407-389X (Media Cetak) DOI 10.30865/jurikom.v9i3.4143 Hal 543−549 http://ejurnal.stmik-budidarma.ac.id/index.php/jurikom
2500 71,49(+0,53) 70,05(+0,84)
3000 71,93(+0,44) 69,84(-0,21)
3500 72,02(+0,09) 70,14(+0,30)
4000 72,19(+0,17) 69,70(-0,44)
4500 71,05(-1,14) 68,90(-0,80)
5000 71,84(+0,79) 70,37(+1,47)
Figure 1. Test Results Of Accuracy Max Feature TF-IDF
From table 7 and Figure 3, it can be concluded that there is an increase in the accuracy of the model from every increase in the number of max features in TF-IDF, where there is the most significant increase in the max feature 1500 to 5000 with an average accuracy of 71-72% compared to the smaller number of max features. With the most optimal accuracy at max feature 4000 which is 72.19% with an f1-score of 69.70%.
3.5.2 Scenario 2 Testing the number of epochs values in the BNN model
In scenario 2, a test was conducted to determine how the influence of the many epochs in the Backpropagation Neural Network. value epochs used in the test are 50, 100, 150, 200, and 250. The results of the scenario testing 2 data are shown in table 8.
Table 8. Epochs Value Test Results
Number of Epochs Accuracy(%) F1-Score
50 70,07 67,16
100 71,66(+1,59) 70,19(+3,03)
150 72,55(+0,89) 70,43(+0,24)
200 71,57(-0,98) 69,69(-0,74)
250 70,52(-1,05) 69,10(-0,59)
Figure 2. Graph of Epochs Value Test Results Accuracy
From table 8 and figure 4, it can be concluded that a large number of epochs in the model can affect the accuracy of the model, whereas in the number of epochs 50,100, and 150 there is an increase in model accuracy and a decreased model accuracy when the number of epochs is 200 and 250. The most optimal accuracy in this test is when there are 150 epochs with an accuracy of 72.55% with an f1-score of 70.43%.
66 67 68 69 70 71 72 73
500 1000 1500 2000 2500 3000 3500 4000 4500 5000 Accuracy
68,5 69 69,5 70 70,5 71 71,5 72 72,5 73
50 100 150 200 250
Accuracy
JURIKOM (Jurnal Riset Komputer), Vol. 9 No. 3, Juni 2022 e-ISSN 2715-7393 (Media Online), p-ISSN 2407-389X (Media Cetak) DOI 10.30865/jurikom.v9i3.4143 Hal 543−549 http://ejurnal.stmik-budidarma.ac.id/index.php/jurikom 3.5.3 Scenario 3 Testing on the comparison of the hyperparameter tuning gridsearch on the Backpropagation Neural
Network
In scenario 3, a test was conducted to find out how the influence of the number of epochs on the Backpropagation Neural Network. value epochs used in the test are 50, 100, 150, 200, and 250. The parameters used during the test can be seen in table 9 and the results from the scenario 3 data test are shown in table 10.
Table 9. Parameter Value Used
Parameter Batch_size Epochs Dropout_rate
Parameter 1 [ 8,16 ] [ 50,100 ] [ 0.3,0.4 ] Parameter 2 [ 16,32 ] [ 100,150 ] [ 0.4,0.5 ] Parameter 3 [ 32,64 ] [ 150,200 ] [ 0.5,0.6 ]
Table 10. Hyperparameter With Grid Search Test Results
Method Accuracy(%) F1-Score
Backpropagation Neural Network 70,07 67,16
Backpropagation Neural Network + hyperparameter tuning Parameter 1 74,21(+4,14) 71,70(+4,54) Backpropagation Neural Network + hyperparameter tuning Parameter 2 74,56(+0,35) 71,21(-0,49) Backpropagation Neural Network + hyperparameter tuning Parameter 3 74,47(-0,09) 71,05(-0,16) From table 10, it can be concluded that the model generated in the Backpropagation Neural Network using hyperparameter tuning has a greater value than the Backpropagationion Neural Network without hyperparameter tuning.
Hyperparameter tuning parameter 2 produces the highest accuracy with an accuracy value of 74.56% with an increase of 4.49.
3.6 Analysis of Test Results
After conducting the test scenarios in the study, it can be concluded that the stage being tested can affect the performance value of the model created. test max feature on TF-IDF where it can be concluded that the higher the max feature of the TF-IDF model, the higher the accuracy of the model will be. In testing scenario 1, it can be seen that the number of max features 1500 to 5000 can increase the accuracy of the model compared to the number of max features 500 and 1000. The second is a test on the effect of the epochs on the model, where the results are that the more epochs on the model, the better the accuracy of the model. However, too many epochs can also cause the model's performance to decrease because too many epochs will cause the model not to learn the data but instead only remember the data and there will be an over- fit on the data. It can be seen in the test scenario 2 where the epoch 50, 100, and 150 always get an increase in the accuracy value, but when the epoch 200 and 250 there is a decrease in the accuracy value. The third is the hyperparameter tuning with grid search on the model which obtained a significant increase in accuracy compared to the model without hyperparameter tuning with an accuracy increase of 4.49%.
Figure 3. Comparison Graph of Accuracy of Each Scenario
From the scenario testing that has been carried out, it can be seen in the figure that in scenario 1 the effect of the number of the max feature of TF-IDF gets 72.19% accuracy, scenario 2 the effect of the number of epochs gets 72.55%
accuracy and the last scenario 3, namely the comparison of the model using hyperparameter tuning using grid search to get the highest accuracy value compared to the two scenarios that have been carried out, which is 74.56%. The significant increase in model accuracy in models using hyperparameter tuning is because in the NN model we never know how many layers, neurons, or what is the most optimal optimizer in the dataset, and with hyperparameter tuning, we can find the best possibilities of the parameters to be used in datasets.
71 71,5 72 72,5 73 73,5 74 74,5 75
Scenario 1 Scenario 2 Scenario 3
JURIKOM (Jurnal Riset Komputer), Vol. 9 No. 3, Juni 2022 e-ISSN 2715-7393 (Media Online), p-ISSN 2407-389X (Media Cetak) DOI 10.30865/jurikom.v9i3.4143 Hal 543−549 http://ejurnal.stmik-budidarma.ac.id/index.php/jurikom
4. CONCLUSION
After the research on sentiment analysis on the construction of the Jakarta-Bandung high-speed train with social media Twitter using the Backpropagation Neural Network was completed, it can be concluded that the number of max features in TF-IDF can affect the level of accuracy in the model whereby increasing the number of max features in TF -IDF can improve the accuracy of the results on the model. Then the number of epochs is also proven to affect the level of accuracy in the model, where ephocs too manyThe use of hyperparameter tuning with the grid search is also proven to affect the accuracy of the model, with hyperparameter tuning the accuracy of the NN model can be increased compared to the Backpropagation Neural Network without using hyperparameter tuning. Backpropagation Neural Network using hyperparameter tuning gets the highest accuracy results compared to the results of other scenarios, which is 74.56%.
REFERENCES
[1] “High-Speed Tracks | High Speed Rail Alliance.” [Online]. Available: https://www.hsrail.org/high-speed-tracks
[2] “Jakarta-Bandung High-Speed Rail - Railway Technology.” [Online]. Available: https://www.railway- technology.com/projects/jakarta-to-bandung-high-speed-rail/
[3] “Kereta Cepat Jakarta Bandung - Kereta Cepat Indonesia China.” [Online]. Available: https://kcic.co.id/proyek/kereta-progres- pengembangan/
[4] T. Ahmad, “Dampak Negara Danai Kereta Cepat : Subsidi Naik , Tiket Mahal,” pp. 1–7, 2021.
[5] O. R. Kasali, “Menyoal Ribut-ribut Kereta Cepat Jakarta- Bandung,” pp. 1–10, 2021.
[6] B. Liu, “Introduction,” in Sentiment Analysis, Cambridge University Press, 2020, pp. 1–17. doi: 10.1017/9781108639286.002.
[7] H. Jiang, P. Lin, and M. Qiang, “Public-Opinion Sentiment Analysis for Large Hydro Projects,” J. Constr. Eng. Manag., vol.
142, no. 2, p. 05015013, 2016, doi: 10.1061/(asce)co.1943-7862.0001039.
[8] Z. Reyhana, K. Fithriasari, M. Atok, and N. Iriawan, “Linking Twitter Sentiment Knowledge with Infrastructure Development,”
Matematika, vol. 34, no. 3, pp. 91–102, 2018, doi: 10.11113/matematika.v34.n3.1142.
[9] H. Tsaniya, R. Rosadi, and A. S. Abdullah, “Sentiment analysis towards Jokowi`s government using twitter data with convolutional neural network method,” J. Phys. Conf. Ser., vol. 1722, no. 1, 2021, doi: 10.1088/1742-6596/1722/1/012017.
[10] T. Hendrawati and C. P. Yanti, “Analysis of Twitter Users Sentiment against the Covid-19 Outbreak Using the Backpropagation Method with Adam Optimization,” J. Electr. Electron. Informatics, vol. 5, no. 1, p. 1, 2021, doi: 10.24843/jeei.2021.v05.i01.p01.
[11] M. D. Sapanta, J. T. Elektro, F. T. Industri, and U. I. Indonesia, “Backpropagation Pada Studi Peramalan Beban Menggunakan Metode Artificial Neural Network,” 2018.
[12] G. Z. Nabiilah, S. Al Faraby, and M. Dwifebri Purbolaksono, “Classification of Hadith Topic of Indonesian Translation Using K-Nearest Neighbor and Chi-Square,” Int. J. Inf. Commun. Technol., vol. 7, no. 2, pp. 11–22, 2021, doi: 10.21108/ijoict.v7i2.573.
[13] VINDI DWI ANTONIO, “Analisis Kinerja Ekstrasi Fitur Tf-Idf (Term Frequency – Inverse Document Frequency) Untuk Algoritma Klasifikasi Stochastic Gradient Descent Pada Analisis Sentimen Teks Indonesia,” pp. 7–37, 2021.
[14] V. Amrizal, “Penerapan Metode Term Frequency Inverse Document Frequency (Tf-Idf) Dan Cosine Similarity Pada Sistem Temu Kembali Informasi Untuk Mengetahui Syarah Hadits Berbasis Web (Studi Kasus: Hadits Shahih Bukhari-Muslim),” J.
Tek. Inform., vol. 11, no. 2, pp. 149–164, 2018, doi: 10.15408/jti.v11i2.8623.
[15] S. P. Siregar and A. Wanto, “Analysis of Artificial Neural Network Accuracy Using Backpropagation Algorithm In Predicting Process (Forecasting),” IJISTECH (International J. Inf. Syst. Technol., vol. 1, no. 1, p. 34, 2017, doi: 10.30645/ijistech.v1i1.4.
[16] J. Brownlee, “How to Code a Neural Network with Backpropagation In Python (from scratch),” Machine Learning Mastery. p.
1, 2016. [Online]. Available: https://machinelearningmastery.com/implement-backpropagation-algorithm-scratch-python/
[17] E. Eka Patriya, “Implementasi Support Vector Machine Pada Prediksi Harga Saham Gabungan (Ihsg),” J. Ilm. Teknol. dan Rekayasa, vol. 25, no. 1, pp. 24–38, 2020, doi: 10.35760/tr.2020.v25i1.2571.
[18] F. A. Hizham, Y. Nurdiansyah, and D. M. Firmansyah, “Implementasi Metode Backpropagation Neural Network (BNN) dalam Sistem Klasifikasi Ketepatan Waktu Kelulusan Mahasiswa (Studi Kasus: Program Studi Sistem Informasi Universitas Jember),”
Berk. Sainstek, vol. 6, no. 2, p. 97, 2018, doi: 10.19184/bst.v6i2.9254.