Performance of ANN and RNN in Predicting the Classification of Covid-19 Diseases based on Time Series Data
Ridho Isral Essa, Sri Suryani Prasetyowati, Yuliant Sibaroni* School of Compting, Informatic, Telkom University, Bandung, Indonesia
Email: 1ridhokdd@student.telkomuniversity.ac.id, 2srisuryani@telkomuniversity.ac.id, 3,*yuliant@telkomuniversity.ac.id Coresspondence Author Email: yuliant@telkomuniversity.ac.id
Submitted 16-01-2023; Accepted 08-02-2023; Published 17-02-2023 Abstract
Indonesia is one of the countries with the highest confirmed cases of COVID-19. The city of Bandung is an area in Indonesia where the number of confirmed cases have continued to increase from 2021 to 2023. Currently there are around 103,574 cases with a total of deaths of around 1485 people. This is bad news for the city of Bandung because of the increasing number of confirmed cases. Various precautions against factors that might affect the rapid spread of COVID-19 in the city of Bandung have been carried out. But the confirmation cases still can't be stopped. Therefore, in this study we made a classification of the spread of COVID-19 in the city of Bandung with 25 features which later be expanded using feature expansion techniques. This aims to analyze what factors have a major influence on the spread of COVID-19 in the city of Bandung. The method used are ANN and RNN methods. Where in this study the two methods were compared to determine which model had the best performance. Modeling is done by building models 2, 3, 4, and 5 months then the best model accuracy results from the ANN method are 79% and 81% for the RNN method. The purpose of this research is to analyze the performance between both ANN and RNN models so the most efficient model can be determined. The contribution of this research is to help the government of Bandung to better manage the spread of the COVID-19 disease.
Keywords: Artificial Neural Network; Recurrent Neural Network; Accuracy, COVID-19, Feature expansion.
1. INTRODUCTION
COVID-19 is a disease that has a high infection rate. This disease has occurred since 2019 and is caused by SARS-CoV- 2. The disease COVID-19 was declared a Public Health Emergency of International Concern (PHEIC) [1]–[4] in 2020 by the World Health Organization (WHO), the virus originated in 2019 in a lab located in Wuhan, China [2], [5], [6]. Then the virus began to spread throughout the world, so it became a pandemic. Since 2019 the number of confirmed cases has continued to rise, one country that also has a high number of confirmed cases is Indonesia. The 3 province in Indonesia with the most confirmed cases are DKI Jakarta, West Java, and Central Java. This is due to the fact that these provinces have many occupants. This research mainly focuses on the cases of COVID-19 in West Java specifically in Bandung.
Based on the data from the government of Bandung that can be accessed online, it is stated that the amount of confirmed cases in Bandung continues to rise since 2021. In early 2021 the number of confirmed cases in Bandung reached around 10 thousand confirmed cases, then it is stated that in early 2023 the number of cases has reached around 100 thousand cases. The number of deaths tend to stabilize since early 2021 where the number of deaths are around 200 and in early 2023 the total number of deaths has reached around 1000 people. This data can be accessed from Indonesia’s COVID-19 information center based around Bandung “covid19.bandung.go.id”.
This disease spreads through the secretions of the mouth and nose, so it can spread infection through sneezing, coughing, and also talking. However, just touching does not necessarily mean that you will be infected. If the traces of the touch enter through the mouth or nose, the transmission will occur. In addition, the disease can spread if you touch a contaminated location and then touch your face [1]–[3]. This pandemic also has made several governments around the world able to control the spread of the virus by reducing the number of interactions and looking for alternative ways to be able to continue daily life [1], [7]. There are also several other businesses such as Non-profit Organizations (NPOs), health agencies, and providing datasets related to health materials needed by various healthcare systems [4]. The government has done various ways to be able to better regulate the situation so that the number of victims can decrease.
There are many possible reasons as to why certain areas have a high infection rate. It could be due to the size of the population in Bandung, the occupants not behaving safety protocols or not wearing masks, or the lack of education in terms of preventing the spread of the virus itself. In order to be able to identify the cause or main factors of the high infection rate, there needs to be a classifcation to the problem. Which is why we propose to use a method that uses machine-learning models to help identify the classification of the disease in order to be able to make predictions on the spread of COVID-19. We propose to use machine-learning models such as Artificial Neural Network (ANN) and Recurrent Neural Network (RNN). RNN models tend to perform better for time series-based datasets, however we believe there is a way to gain a better performance from ANN models in this aspect by using feature expansion. Artificial Neural Network (ANN) is a concept method based on a biological nervous system to manage information like a brain [8]–[11].
ANN takes information from neural networks in the brain and builds different networks based on its methods. In addition, there is the Recurrent Neural Network (RNN) method which is the concept of storing system output results in nodes and feeding that information back to the model [1], [2], [9]. RNN is a Deep Learning (DL) model like ANN in that it works
by storing data information training results into nodes and returning the training results to the model for further training [2], [9], [12]–[15].
In several papers, the ANN model has been compared with other deep learning methods to determine which model is most often used in forecasting the COVID-19 classification, which means making prediction results using past and present data which was done by Carmela Comito in the year 2021 [1]. In another paper, the performance of ANN and RNN models is compared using methods such as RMSE with new deep learning models such as Transformer by Yiming Fei during 2021 [2]. In another method, there are papers that propose to use ANN or RNN algorithms to make a new deep learning model namely ARIMA as stated by Othman Istaiteh and Pratima Kumari in the year 2020 [16], [17]. There are also papers that discuss the use of ANN, RNN, and various other models to diagnose the COVID-19 virus using image processing. Many models have compared their performance with the two models above, such as the LSTM, MLP, LR models, and various other deep learning models. The existing paper produces Confusion Matrix values with ANN and RNN models having values of around 90% for sensitivity (93.78% and 92.04%), specificity (91.76% and 90.87%), accuracy (86% and 84.16%), and f- scores (91.34% and 90.61%) analyzed by K. Shankar in 2021 [3]. However, comparisons have not been made using data based on time [1]–[3].
This study aims to be able to analyze the performance of the ANN and RNN models in predicting the classification of COVID-19 disease based on time data to make it easier for the government to control the spread of the disease. As previously explained, there are no papers that discuss the performance of the ANN and RNN models regarding the prediction of the classification of COVID-19 disease based on time. Our contribution to this research is to help develop new models based on ANN and RNN models, especially for models that can process data based on time. Creating models with feature expansion using ANN and RNN as well as building these models based on previous months to make future predictions. This contribution is also for the purpose of helping the government of Bandung to better manage the way this disease spread amongst the people of Bandung.
2. RESEARCH METHODOLOGY
In Figure 1, the research method is shown in a flowchart manner to further describe the process of this research. In this research method, the data is first collected and then preprocessed in order to be able to be processed within the model.
Then the model itself needs to be built, where both the ANN and RNN model each have 4 models for each of their respective months. After the model and the dataset are set up, The model can then begin training and testing the dataset that is already prepared beforehand. Once the model has finished training, evaluation for the model as well as the prediction result for each model can then be compared between the ANN and RNN model.
Figure 1. Research Methodology
2.1 Dataset
The dataset used is a dataset on the spread of the COVID-19 disease throughout Indonesia from 2020 to April 2022. The dataset was taken from 153 villages for 18 months, with each village having 25 attributes there are genders (male, female), amount of rainfall, solar radiation, average temperature, maximum temperature, minimum temperature, school statuses like not yet at school, not yet finished elementary school, graduated from elementary school, junior high school, high school, 1-year and 2-year diploma, 3-year diploma, Undergraduate (Bachelor), Graduate (Magister), Postgraduate (Doktor), vaccine statuses like vaccinated dose 1, vaccinated dose 2, vaccinated dose 3, wearing a mask, compliance with physical distancing, confirmed cases, recovered cases, and death cases. There are 3 classes namely low, medium, and high.
2.2 Preprocessed Data
Data preprocessing is a way to convert data that can be read by humans into data that can be read by computers easily.
This is because there is data that is ambiguous according to the programming language, such as data that has a choice of
"Yes" or "No". There is also data that is numeric but does not fall within the range of data values to be processed by the intelligent systems modeling language. Therefore, data preprocessing is needed to make it easier to input into the intelligent system model. The data preprocessing steps are described as follows:
a. Distribution of Train Data and Test Data
First, the dataset is divided into training data and test data. Train data is part of the dataset that is used as a parameter.
The test data is a dataset after the training data which aims to be included in the test. Then the two datasets are compared to find out if the model works correctly. The distribution of the dataset can be done by Random Sampling, where the distribution is chosen randomly.
b. Eliminate Missing Values and Unused Attributes
In this step, empty or null values and unused attributes in the dataset are removed. This is done to produce more accurate and robust model predictions so that when the dataset is plotted, the results are not random or random.
Columns that are not used are columns from the dataset that are not needed in the testing phase.
c. Data Normalization
Normalization is the process of changing data to be in the same range of values. Normalization is needed to reduce the number of errors resulting from the modeling system and make the dataset tidier.
2.3 Build Model
Before the model is trained using ANN or RNN, we first build each model first. We divide the model into 4 types based on the months required to reach a target prediction. We then generate 4 more models from each month type namely a, b, c, and d through data selection. These models are made by dividing the data based on x model and using the best feature using the SelectKBest method and feature expansion method. The following is a detailed description of the build model.
a. Classification Prediction Model
Table 1. Prediction Model based on 2 to 5 Months of Prediction
Data Model Target Month Train Data
2 Months Before 2a Aug Jun, Jul
Oct 2b Oct Jul, Aug
Nov 2c Nov Aug, Oct
Dec 2d Dec Oct, Nov
3 Months Before 3a Aug May, Jun, Jul
Oct 3b Oct Jun, Jul, Aug
Nov 3c Nov Jul, Aug, Oct
Dec 3d Dec Aug, Oct, Nov
4 Months Before 4a Aug Apr, May, Jun, Jul
Oct 4b Oct May, Jun, Jul, Aug
Nov 4c Nov Jun, Jul, Aug, Oct
Dec 4d Dec Jul, Aug, Oct, Nov
5 Months Before 5a Aug Mar, Apr, May, Jun, Jul
Oct 5b Oct Apr, May, Jun, Jul, Aug
Nov 5c Nov May, Jun, Jul, Aug, Oct
Dec 5d Dec Jun, Jul, Aug, Oct, Nov
b. Artificial Neural Network
Artificial Neural Network (ANN) is an intelligent system model consisting of several nodes that have mutually interconnected networks connected between nodes. The explanation of the ANN model process can be seen in the following figure.
Figure 2. Process diagram for the Artificial Neural Network model
As seen in Figure 2, the input data entered into the model is reduced in number based on the required data and then each input data that has been processed is entered into the hidden layer along with the weighted value of each data.
Because the dataset is based on data for 2020 and 2021, the data for training and testing is based on monthly data. The prediction of month t requires data from months 𝑡 − 1, 𝑡 − 2, 𝑡 − 3, . . . , 𝑡 − 𝑚 and the attributes of each data (𝑥1, 𝑥2, 𝑥3, 𝑥4, 𝑥5, 𝑥6, 𝑥7) are processed into the hidden layer with each weighted value along with the weighted value bias to reduce the error and do the sigmoid (σ) calculation [18], [19].
ℎ𝑖= 𝜎(∑𝑚𝑗=1 ∑𝑛𝑖=1 𝑤𝑡𝑗𝑖𝑥𝑡𝑗𝑖 + 𝑤0) (1)
The weighted value that has been added up is then entered into the formula and produces the output value. The formula of the ANN model is as follows. [15]
𝑦𝑠𝑢𝑚= ∑𝑛𝑖=1 𝑤𝑖ℎ𝑖+ 𝑤0 (2)
Then, the 𝑦𝑠𝑢𝑚 is entered into the activation function calculation as in the following formula.
𝑦𝑡= 𝑓(𝑦𝑠𝑢𝑚) = {1, 𝑥 ≥ 1 − 1, 𝑥 < 1 (3) From the formula above, it can be concluded that the output value is marked with a value of 1 or -1.
Description:
𝑤0 : Weighted value bias
𝑤𝑡𝑗𝑖 : Weighted value in the j-th month of the i-th node 𝑥𝑡𝑗𝑖 : Data value in the j-th month of the i-th node 𝑤𝑖 : The i-th weighted value
ℎ𝑖 : The value of the data in the i-th hidden layer
𝑦𝑠𝑢𝑚 : The result of the total value of the i-th data multiplied by its weighted value 𝑦𝑡 : The result of the t-th output value
c. Recurrent Neural Network
Recurrent Neural Network (RNN) is an intelligent system model that generates a value from a recalculation based on the iterations it has passed. The iterations are then used in a way that will store information from previous iterations to obtain better results that is then passed down toward the next iteration. This information passing is done with the use of weighted values. An explanation of the RNN model process can be seen in the following in Figure 3.
Figure 3. Recurrent Neural Network Process Diagram
As seen in Figure 3, the process of the RNN model is carried out using iterative iterations to determine the results, namely the i-th hidden state and the output value. With the use of feature expansion, this model is then further extended in to different models based on the different amount of features that are being used in the model. Because there are many results and outputs that are returned from the many iterations, only the iteration with the best accuracy and performance that is returned to then be further analyzed. This model of RNN will also use the “relu” function to help further calculate the accuracy between each iteration. The equation is shown as follows [2], [9], [12], [20].
𝑦𝑡𝑗𝑖= 𝜎 (𝑈𝑥𝑡𝑗𝑖+ 𝑊𝑦𝑡𝑗(𝑖−1)) (5)
𝑧𝑜𝑢𝑡= 𝑆𝑜𝑓𝑡𝑚𝑎𝑥 (𝑉𝑦𝑡𝑗𝑖) (6)
Description :
𝑦𝑡𝑗𝑖 : The i-th hidden value of the j-th feature and the t-th model 𝑧𝑜𝑢𝑡 : The output value
U : The weighted value to the hidden layer V : The weighted value to the output layer W : The weighted value between each hidden node
𝑥𝑡𝑗𝑖 : The i-th data value of the j-th feature and the t-th model d. Feature Expansion
Feature expansion is the usage of an extended amount of features made using feature selection and iterations that are based by months and features to help further analyze which feature has the most impact to the spread of the COVID- 19 disease [21]–[23]. For example, the feature “dose3” which is used to describe the amount of people that has been
vaccinated with the third dosage, is then extended to “dose3_1” and “dose3_2” to differentiate between the 2 months in a 2 month model. To shorten the time complexity of the model, the feature selection method is used to select the relevant attributes to be included in the test. In addition, this method can also improve the accuracy of the model results. The selection method that can be used is SelectKBest where the selected attribute is based on the largest k value. This method can be used in the form of a library in the Python language. The following is an example of a combination of samples according to the SelectKBest selection.
Table 2. Example of a sample of the results of attribute selection
No Features Combination
1 1 x3
2 2 x2,x7
3 3 x1,x2,x4
4 4 x2,x3,x4,x5
5 5 x1,x3,x4,x5,x6
… … …
n n x1,x2,x3,…,xn
The result of the feature selection will be a combination of features that will then be used in the modeling process as seen in Table 2. This feature selection process is crucial in the use of feature expansion through random selection. The feature selection method alongside with the model itself is then run through a series of iterations in order to expand the features for feature expansion. This will help determine which features are the main factors that are causing the disease to have a high infection rate.
3. RESULT AND DISCUSSION
3.1 Result
After implementing the models and generating test results for the time series dataset, the performances are then evaluated using the confusion matrix of each model (model 2,3,4,5 bulan sebelum target pada metode ANN dan RNN) and return the best model to be used for the prediction classification on the spread of COVID-19. Following are the graph results of the accuracy pattern of each best model:
Figure 4. Best accuracy plotting of each month from ANN models. (a) 2A Model, (b) 3D Model, (c) 4D Model, (d) 5A Model
Figure 5. Best accuracy plotting of each month from RNN models. (a) 2B Model, (b) 3B Model, (c) 4A Model, (d) 5C Model
From Figure 4 and Figure 5, we can retrieve the model with the best performance from both ANN and RNN models. The following are the test results from both of these models as shown in Table 3.
Table 3. The Performance Result of the Two Best Models from Either ANN and RNN Model
Precision Recall F1-Score Accuracy
0 1 2 0 1 2 0 1 2
ANN Model (4A) 0,96 0,66 0,85 0,7 0,91 0,73 0,81 0,77 0,79 0,79 RNN Model (3B) 0,87 0,45 0,91 0,93 0,59 0,64 0,9 0,51 0,75 0,81
In Table 3, it is shown that the RNN model of 3B which is a model used in the 3 months model, that the model was able to reach an accuracy of 81% while the ANN model of 4A which is a model used in the 4 month model, was able to reach an accuracy of 79%. This shows that although the RNN model still has the higher accuracy result for a time series based data, the ANN model was able to have a close enough accuracy to the RNN model especially in the classification prediction with the usage of 4 months.
3.2 Discussion
From the test results that are shown in Figure 4 and Figure 5 as well as Table 3, we can conclude that in terms of which model has the best performance, the 3B RNN Model has the best accuracy and confusion matrix results out of the other performance models. This model gained an accuracy of 81% at an index of 32, which means the best performance for this model is within the use of 35 features through the help of feature expansion. However, some things need to be further analyzed and perfected, such as the RNN model using the “MSE” loss parameter instead of the “binary cross-entropy”
loss parameter. Besides that, the ANN model has an accuracy of 79% for the 4A model, which is less than 3B from the RNN model. Other research has made predictions for COVID-19 using the SVM method and RNN using LSTM method and both of these methods obtained an accuracy of 80% [24] and 78% [6], while in our research, our best model is the 3B RNN model with an accuracy of 81%. Then there is a paper that compares the ANN method with the RNN in the Intrusion Detection System and the result is that the RNN model is better than the ANN model, as well as in this study the RNN model has better performance than ANN [9].
This analysis has shown that with the usage of feature expansion and SelectKBest for the feature selection phase, we can further analyze the importance of each feature in generating the results to determine whether or not the features play an important role in becoming a main factor of the problem. The most common features that are used in each of the models are 'dose3_2', 'Compliance in wearing a masks_1', 'Compliance in wearing a masks_2', 'Confirmed Case_1', and 'Confirmed Case_2' with each of these features being used in at least 7 models out of each model with the best accuracy.
After further analysis, it would make sense why these features are of importance to the factor in the spread of COVID- 19.
4. CONCLUSION
After creating a model based on 2, 3, 4, and 5 months before the target month and successfully classifying predictions of the spread of COVID-19, we get the best model accuracy results from the ANN model, which is 91%, but when we analyze the confusion matrix, the model with an accuracy of 91% does not have good values of the confusion matrix. So we decided to use the ANN model with an accuracy of 79% because it has better confusion matrix values. While the best model of the RNN model has an accuracy value of 81%. So that the best model between RNN and ANN in this study is won by RNN. The best model of this RNN model has 35 features from the results of the feature expansion. Furthermore, further research is needed to find a better method for increasing the performance of ANN models for the usage of time series-based datasets classifications. Further work would also need to be done in the model implementation. Such as the usage of “binary cross-entropy” for the loss parameter of the RNN model and building a more refined model either by using a different activation function or adding more nodes for each model.
REFERENCES
[1] C. Comito and C. Pizzuti, “Predicting COVID-19 with AI techniques: Current research and future directions,” in Proceedings of the 2021 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining, ASONAM 2021, Nov.
2021, pp. 518–524. doi: 10.1145/3487351.3490958.
[2] Y. Fei, “Analysis the use of machine learning algorithm-based methods in predicting COVID-19 infection,” in ACM International Conference Proceeding Series, Oct. 2021, pp. 88–94. doi: 10.1145/3500931.3500948.
[3] K. Shankar, E. Perumal, M. Elhoseny, F. Taher, B. B. Gupta, and A. A. A. El-Latif, “Synergic Deep Learning for Smart Health Diagnosis of COVID-19 for Connected Living and Smart Cities,” ACM Trans Internet Technol, vol. 22, no. 3, Nov. 2021, doi:
10.1145/3453168.
[4] G. Thakur et al., “COVID-19 Joint Pandemic Modeling and Analysis Platform,” in Proceedings of the 1st ACM SIGSPATIAL International Workshop on Modeling and Understanding the Spread of COVID-19, COVID-19 2020, Nov. 2020, pp. 43–52. doi:
10.1145/3423459.3430760.
[5] L. Fell, “Trust and COVID-19,” Digital Government: Research and Practice, vol. 2, no. 1, pp. 1–5, Jan. 2021, doi:
10.1145/3428472.
[6] S. Bodapati, H. Bandarupally, and M. Trupthi, “COVID-19 Time Series Forecasting of Daily Cases, Deaths Caused and Recovered Cases using Long Short Term Memory Networks,” in 2020 IEEE 5th International Conference on Computing Communication and Automation, ICCCA 2020, Oct. 2020, pp. 525–530. doi: 10.1109/ICCCA49541.2020.9250863.
[7] M. Kiamari, G. Ramachandran, Q. Nguyen, E. Pereira, J. Holm, and B. Krishnamachari, “COVID-19 Risk Estimation using a Time-varying SIR-model,” in Proceedings of the 1st ACM SIGSPATIAL International Workshop on Modeling and Understanding the Spread of COVID-19, COVID-19 2020, Nov. 2020, pp. 36–42. doi: 10.1145/3423459.3430759.
[8] R. A. al Rashid Agha, M. N. Sefer, and P. Fattah, “A comprehensive study on sign languages recognition systems using (SVM, KNN, CNN and ANN),” in ACM International Conference Proceeding Series, Oct. 2018. doi: 10.1145/3279996.3280024.
[9] N. Chaibi, B. Atmani, and M. Mokaddem, “Deep Learning Approaches to Intrusion Detection: A new Performance of ANN and RNN on NSL-KDD,” in ACM International Conference Proceeding Series, Oct. 2020. doi: 10.1145/3432867.3432889.
[10] A. T. M. T. S. Hendri, I. T. M. T. S. Suprayogi, M. Zulfakar, and A. Ongko, “Comparisons of Tidal Prediction Analysis by Using Adaptive Neuro Fuzzy Inference System (ANFIS) and Artificial Neural Network (ANN),” in ACM International Conference Proceeding Series, Dec. 2017, pp. 164–168. doi: 10.1145/3168390.3168393.
[11] L. Yuan and S. Zeng, “Optimization of machining parameters based on principal component analysis and artificial neural network,” in ACM International Conference Proceeding Series, 2019, vol. Part F147614, pp. 46–49. doi:
10.1145/3314493.3318454.
[12] W. W. T. Fok, L. C. W. Chan, and C. Chen, “Artificial intelligence for sport actions and performance analysis using Recurrent Neural Network (RNN) with Long Short-Term Memory (LSTM),” in ACM International Conference Proceeding Series, Nov.
2018, pp. 40–44. doi: 10.1145/3297097.3297115.
[13] I. Aliyu, Y. B. Lim, and C. G. Lim, “Epilepsy detection in EEG signal U sing recurrent neural network,” in ACM International Conference Proceeding Series, Mar. 2019, pp. 50–53. doi: 10.1145/3325773.3325785.
[14] M. A. H. Ashour and R. A. Abbas, “Improving time series’ forecast errors by using recurrent neural networks,” in ACM International Conference Proceeding Series, Feb. 2018, pp. 229–232. doi: 10.1145/3185089.3185151.
[15] J. M. Nazzal, I. M. El-Emary, S. A. Najim, and A. Ahliyya, “Multilayer Perceptron Neural Network (MLPs) For Analyzing the Properties of Jordan Oil Shale,” World Appl Sci J, vol. 5, no. 5, pp. 546–552, 2008.
[16] O. Istaiteh, T. Owais, N. Al-Madi, and S. Abu-Soud, “Machine Learning Approaches for COVID-19 Forecasting,” in 2020 International Conference on Intelligent Data Science Technologies and Applications, IDSTA 2020, Oct. 2020, pp. 50–57. doi:
10.1109/IDSTA50958.2020.9264101.
[17] P. Kumari and D. Toshniwal, “Real-time estimation of COVID-19 cases using machine learning and mathematical models-The case of India,” in 2020 IEEE 15th International Conference on Industrial and Information Systems, ICIIS 2020 - Proceedings, Nov. 2020, pp. 369–374. doi: 10.1109/ICIIS51140.2020.9342735.
[18] X. Huang, Y. Fu, Z. Dong, W. Chen, and J. Chen, “Cooperative Analysis on Artificial Neural Network with Optimization Algorithm Based on Social Network,” in ACM International Conference Proceeding Series, Jul. 2020, pp. 120–124. doi:
10.1145/3414752.3414799.
[19] X. Chen, L. Huang, W. Liu, P. C. Shih, and J. Bao, “Automatic Surgery Duration Prediction Using Artificial Neural Networks,”
in ACM International Conference Proceeding Series, Oct. 2021. doi: 10.1145/3487075.3487128.
[20] L. Wang, X. He, W. Zhang, and H. Zha, “Supervised reinforcement learning with recurrent neural network for dynamic treatment recommendation,” in Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Jul. 2018, pp. 2447–2456. doi: 10.1145/3219819.3219961.
[21] E. Saha, H. Schaeffer, and G. Tran, “HARFE: Hard-Ridge Random Feature Expansion,” Feb. 2022, [Online]. Available:
http://arxiv.org/abs/2202.02877
[22] J. Guo and S. Guo, “A Novel Perspective to Zero-Shot Learning: Towards an Alignment of Manifold Structures via Semantic Feature Expansion,” IEEE Trans Multimedia, vol. 23, pp. 524–537, 2021, doi: 10.1109/TMM.2020.2984091.
[23] Naufal Adi Nugroho and Erwin Budi Setiawan, “Implementation Word2Vec for Feature Expansion in Twitter Sentiment Analysis,” Jurnal RESTI (Rekayasa Sistem dan Teknologi Informasi), vol. 5, no. 5, pp. 837–842, Oct. 2021, doi:
10.29207/resti.v5i5.3325.
[24] P. Subramani, S. K, K. R. B, S. R, and P. B.D, “Prediction of muscular paralysis disease based on hybrid feature extraction with machine learning technique for COVID-19 and post-COVID-19 patients,” Pers Ubiquitous Comput, 2021, doi: 10.1007/s00779- 021-01531-6.
