The COVID-19 pandemic is having an unprecedented impact on society and the economy, affecting virtually every aspect of people's daily life and all sectors of the economy. In this situation, society and the health care system need help from modern technologies such as Artificial Intelligence, Big Data, and Machine Learning, which intended to help governments to choose and implement an adequate strategy to combat the spread of the disease by balancing between human safety and the constraints of social and economic life. This paper considers the recommendation rules extracted from the novel ensemble of machine learning methods such as regression tree and clustering. The merged Oxford COVID-19 Government Response Tracker and European Centre for Disease Prevention and Control Covid-19 Cases datasets have been used with the data ranged from January 01 to October 04, 2020. The conclusions and findings of the study could be helpful for decision making on appropriated state policy for reducing the spread of new Covid-19 cases.
2020 Copyright for this paper by its authors.
(iii) how many people in general will be affected by the disease [
In this situation, society and the health care system need help from modern technologies such as Artificial Intelligence (AI), Big Data, and Machine Learning to fight new diseases and make rapid progress. Governments and health organizations need effective decision support systems during a pandemic to effectively control the virus and make optimal real-time decisions to curb the spread of the disease. Artificial intelligence systems and methods can also provide significant assistance in the development of the COVID-19 vaccine. AI methods and algorithms are used to analyze the spread of the disease, predict and track patients and potential carriers of infection. Intelligent analysis of data on confirmed cases of illness, recovery and death is carried out.
At present, the vaccine against COVID-19 is still undergoing clinical trials, and there is no proven effective cure for this disease. In such circumstances, the only effective way to slow the spread of infection is "social distancing" and the use of antiseptics and personal protective equipment to prevent the transmission of the virus from person to person. At the same time, the role of modeling and predicting the spread of the disease, the time and duration of the maximum number of infections and the scale of the epidemic for each country is growing. The findings of such modeling and forecasting are the basis for sound public health decisions regarding appropriate measures and the allocation of resources, such as pulmonary ventilation systems or additional field hospitals.
Behavior analysis of COVID-19 requires the development of a powerful mathematical apparatus for tracking, including automated, its dissemination to develop timely, dynamic, and sound solutions. As shown above, the methods and tools of artificial intelligence and machine learning are an adequate and effective basis for such tasks. The current COVID-19 pandemic poses serious challenges for data scientists and AI professionals due to limited and incomplete data and the large amount of heterogeneous data, as well as the significant impact of various factors on disease behavior and spread.
The purpose of the paper is to build the recommendation rules for appropriated state policy for reducing the spread of new Covid-19 cases. The system of these rules is based on novel ensemble of machine learning methods such as regression tree and clustering.
The significant impact of the COVID-19 pandemic on all spheres of public life has led researchers
to a significant interest in modeling the spread, behavior and origin of the virus, understanding its
nature and properties, and finding ways to control the disease. Since the beginning of 2020, tens of
thousands of articles have been published in various fields of science on the COVID-19 pandemic.
Research institutions, foundations and governments are investing significant human, financial and
technical resources to gain new knowledge about the pandemic and close gaps in understanding the
nature of the disease and its consequences, including medical, social, and economic. Thus, a
largescale study and analysis of publications on COVID-19 using machine learning methods by A. Doanvo
et al. [
The development of new effective models for predicting the number and dynamics of new
COVID-19 infections, as well as epidemiological forecasting in general, is extremely important for
the health care system, as it enables effective planning to eliminate or reduce possible epidemics. The
main requirement for such models is the maximum possible accuracy and reliability of the forecast of
epidemiological time series. To meet these requirements, a number of AI-based models have been
used for a number of years [
Thus, Ribiero et al. [
Zhang et al. [
The dynamics of the COVID-2019 outbreak in China, Italy and France in January–March 2020 is
analyzed in [
Roosa et al. [
Vaishya et al. [
Basu and Campbell [13] propose a model based on Long Short-Term Memory, which was trained on cumulative number of COVID-19 cases and deaths separately. They train the developed model using the dataset containing number of confirmed COVID-19 cases during the period of more than four months. The parameters of the proposed model can be adjusted in such a way as to provide the required forecasting accuracy. In the article [13] using the developed LSTM-model, a quantitative assessment and analysis of the impact of measures taken by various US counties on the spread of the disease has been carried out.
Neher et al. [14] consider a seasonal model based on a susceptible-infectious-recovered approach to analyze and predict the spread of COVID-19 virus reinfection in the world in the coming years. In their model, they consider such factors as the rate of infection, the volume and speed of emigration and population movement, and so on. An updated version of their model [15] aims to analyze and forecast the necessary hospital resources to combat pandemics of such a scale as COVID-19.
Hu et al. [16] built and investigated a machine learning model based on composite autoencoders to describe the spread of coronavirus in China in April 2020 based on data from previous periods. They clustered cities and provinces based on the characteristics obtained from the developed auto-coding model. Similar LSTM-based models have been used to study patient-related statistics [17], as well as to identify trends in the spread of the epidemic in China [18].
The study by S. Tuli et al. [19] is based on an improved machine learning model for analyzing and predicting the spread of the COVID-19 epidemic in different countries. The authors of [19] have developed and studied the Robust Weibull model with iterative weighting. They conclude that the developed Weibull model is statistically better than the Gaussian baseline model for prediction of the COVID-19 outbreak. It was shown [19] that simple Gaussian model results in an overly optimistic COVID-19 spreading scenario. The developed model was deployed on a cloud computing platform using the FogBus framework for more accurate and realistic forecasting of the epidemic's growth dynamics.
In this study we have used data from the Oxford COVID-19 Government Response Tracker as well as European Centre for Disease Prevention and Control (ECDC) Covid-19 Cases [20].
The Oxford COVID-19 Government Response Tracker (OxCGRT) dataset provides information on which governments took action, what action they did, and when they took it. The OxCGRT systematically gathers information on a range of common policy responses by a given government and determines the extent to which the government is implementing these measures. The respective scores are combined into a set of policy indicators.
The second dataset contains new public data on the geography of COVID-19 cases worldwide from European Centre for Disease Prevention and Control. Each line or record contains the number of new cases per day, by country or region.
For the performed analysis we have merged the mentioned datasets and considered data from January 01, to October 04. 3.2.
Stochastic gradient enhancement is a solution to a regression problem by building an ensemble of "weak" predictive decision trees. At the first iteration, a decision tree with a limited number of nodes is built. The difference between that predicted tree multiplied by the learning rate and the desired variable at that stage is then calculated. And the next iteration builds on this difference. This continues until the result stops improving. At every step we try to fix the errors of the previous tree.
Dataset consists of attributes with different nature. An approach to processing large, distilled data was described in [21]. The noise data and outliers are presented in the dataset studied as well. That is why at the first stage the following steps are required: • feature selection, • empty data analysis, • data normalization and scaling.
Feature selection is made based on theory of information. The joint mutual information between each feature and target attribute ConfirmedCases is calculated as: where c is target class, I(f,c) is mutual information, is already selected feature, is processed feature.
The already selected featured were selected manually. The list of the already selected features looks like the following: countryName, loaddate. As result, 19 from 38 features were selected.
The next step is empty data analysis. The dataset consists of 66,998 rows, 26,003 of them have empty values in selected attributes. Due to small quantity of rows with empty data these rows were eliminated. In addition, binning was made as well.
It is desirable to bring all input variables to a single range and normalize (the maximum absolute
value of input variables should not exceed one). Otherwise, errors due to variables varying over a
wide range will be more influential than errors due to variables varying over a narrow range. By
ensuring that each feature changes within the same range, we ensure that each has an equal effect.
Therefore, the input variables, as a rule, are scaled so that the variables change in the range of the
function, as a rule, [
First, the regression decision tree is developed. Decision trees divide an object space according to a set of splitting rules. These rules are logical statements about a variable and can be true or false. Three circumstances are key here: 1. the rules make it possible to implement sequential dichotomous data segmentation, 2. two objects are considered similar if they appear in the same segment of the partition, 3. at each step of the partition, the amount of information about the variable under investigation (response) increases.
The main feature of the propose algorithm is its -arc structure. Branching on a chosen trait x splits training objects on subsamples, where is the number of different characteristic values.
Without loss of generality, we will assume that the feature has values from {0, 1, …, -1}, ≥ 2. In this case, when constructing of the decision tree, from the vertex there are arcs labeled with numbers from {0, 1, …, -1}. Let be the label of one of the arcs leaving the vertex , ∈ {0, 1, …, -1}. To form a new current subset of objects and a new current set of features, those objects from ̌ are deleted for which the value of feature is not equal to , and also feature itself is removed from the set of features.
Let be a hanging vertex generated by a branch of a tree with the inner vertices ,…, and let the arc outgoing from the vertex , ∈ {1,…, }, be labeled . Further let ( ) -the current set of objects that hit the vertex . The vertex is associated with a pair ( , ( )), where ( ) is equal to the mean value of the target variable over all objects from ( ), and B is an elementary conjunction of the form … . If the vertex is not pendant, then we assign to it the conjunction = … ... Truth interval of elementary conjunction B denote by . Let be a recognizable object. For each hanging vertex ( , ( )), a check is performed that the description of the test object belongs to the truth interval . If the description belongs to , then the object is associated with the value of the target variable ( ). Object is assigned the value of the target variable where (a) (b) Figure 1: Data distribution before (a) and after (b) preprocessing
The next step is clustering. The k-means method is used. The gap-statistics allows to find the appropriative number of clusters.
Cluster centroids allow to find “average” object in each group and to create the regularization rules. The cluster # 3 shows countries under restriction, cluster #1 consists of countries without almost any restriction (see Table 1, Table 2, Table 3). 2.
15000 t10000 n u o c 5000 0 12000 9000 tun6000 o c 3000 0
As it can be seen from the Figure 2 the clusters differ by the most distinct recommendations, which can be summarized as follows: • Cluster #1: recommended to close schools, and control international travels, • Cluster #2: recommended to stay at home, • Cluster #3: recommended to stay at home and cancel public events.
The influence of such clustering on COVID-19 spreading, peak position and duration as well as mortality rate will be the subject of our further study.
An example of countries along with the most frequent cluster number is given in Table 4. The same country can be joined to different clusters in different time slots. So, the clustering by country will not be so unambiguous. That is why time series for separated country can be interesting and will be the subject of the further study. At this stage we used the classification routine on the dataset. The comparison of rule-based, statistical and recurrent neural network methods can also be found at [22].
Three classifiers are analyzed: random forest (500 trees), logistic regression and XGBOOST (tree learning algorithm). The scores of the models are listed in Table 5.
The rules given below presents the strategy based on decision tree and can be used for strategic planning in the public health system to avoid deaths and severe consequences of the epidemic. if ( confirmedcases <= 11577.5 ) { if ( confirmedcases <= 3806.5 ) { [[confirmeddeaths=20.21931171]] }
else { [[ confirmeddeaths =206.84760845]] } } else { if ( confirmedcases <= 17153.5 ) { if ( e1_income_support <= 1.5 ) { else { if ( c8_international_travel_controls <= 2.5 ) { if ( c6_stay_at_home_requirements <= 0.5 ) { [[confirmeddeaths=452.20754717]] } else { [[confirmeddeaths=1410.8028169]] } } else { [[confirmeddeaths=501.40616622]] }
The current COVID-19 pandemic poses serious challenges for data scientists and AI professionals due to limited and incomplete data and the large amount of heterogeneous data, as well as the significant impact of various factors on disease behavior and spread.
This paper is devoted to building the recommendation rules for appropriated state policy for reducing the spread of new Covid-19 cases. The system of these rules is based on novel ensemble of machine learning methods such as regression tree and clustering. The merged data from the Oxford COVID-19 Government Response Tracker and ECDC Covid-19 Cases datasets were used in this study. [[confirmeddeaths=1346.375]]
The clustering was carried out using the k-means method. The gap-statistics allows to find the appropriative number of clusters, and in the case of the study three clusters were selected. The clusters differ by the recommendations and actions made by the correspondent governments. Thus the first cluster countries chose to close schools and to control international travels as main recommendations; the second cluster countries recommended to stay at home, while the major recommendations of governments of countries belonging to the third cluster were to stay at home and to cancel public events.
The same country can be joined to different clusters in different time slots. So, the clustering by country will not be so unambiguous. That is why time series for separated country can be interesting and will be the subject of the further study. Besides, the influence of such clustering on COVID-19 spreading, peak position and duration as well as mortality rate will also be the subject of our further study.
The regression decision tree was built, and the set of rules was extracted from the decision tree and can be used for strategic planning in the public health system.
We believe that the results obtained in this article will contribute to the public good for solving the current global problem. It is also seen that it is necessary to reduce the infection rate sharply and quickly to see a noticeable decrease in the epidemic peak and the death rate. 6. References [13] Basu, S., Campbell, R.H. Going by the numbers : Learning and modeling COVID-19 disease dynamics. Chaos, Solitons & Fractals 138 (2020), 110140. https://doi.org/10.1016/j.chaos.2020.110140 [14] R.A. Neher, R. Dyrdak, V. Druelle, E.B. Hodcroft, J. Albert. Potential impact of seasonal forcing on a SARS-CoV-2 pandemic. Swiss Med Weekly, 150 (1112) (2020). https://doi.org/10.4414/smw.2020.20224 [15] COVID-19 Scenarios. URL: https://neherlab.org/covid19/ [16] Hu Z, Ge Q, Jin L, Xiong M. Artificial intelligence forecasting of COVID-19 in China. arXiv preprint arXiv:200207112 2020. http://arxiv.org/abs/200207112 [17] Bandyopadhyay SK, Dutta S. Machine learning approach for confirmation of COVID-19 cases: positive, negative, death and release. medRxiv 2020. https://doi.org/10.1101/2020.03.25.20043505 [18] Z. Yang, Z. Zeng, K. Wang, S.-S. Wong, W. Liang, M. Zanin, et al. Modified SEIR and AI prediction of the epidemics trend of COVID-19 in China under public health interventions. J Thorac Dis, 12 (3) (2020), p. 165. https://doi.org/10.21037/jtd.2020.02.64 [19] S. Tuli, S. Tuli, R. Tuli, S.S. Gill. Predicting the growth and trend of COVID-19 pandemic using machine learning and cloud computing. Internet of Things 11 (2020), 100222. https://doi.org/10.1016/j.iot.2020.100222 [20] COVID-19 Data Lake. URL: https://azure.microsoft.com/en-US/services/opendatasets/catalog/covid-19-data-lake/ [21] Boyko, N., Mochurad, L., Stetsiv, I., Kryvenchuk, Yu.: Modeling of the Information System for Processing of a Large Distilled Data for the Investigation of Competitiveness of Enterprises. In: Proc. of the 4th Intl Conf. on Computational Linguistics and Intelligent Systems (COLINS 2020). Volume I: Main Conference Lviv, Ukraine, April 23–24, 2020, pp. 964–978. CEURWS.org, online CEUR-WS.org/Vol-2604/paper64.pdf [22] Boyko, N., Mochurad, L., Parpan, U., Basystiuk, O.: Usage of Machine-based Translation Methods for Analyzing Open Data in Legal Cases. In: Proc. of the Intl Workshop on Cyber Hygiene (CybHyg-2019) co-located with 1st International Conference on Cyber Hygiene and Conflict Management in Global Information Networks (CyberConf 2019), Kyiv, Ukraine, November 30, 2019, pp. 328–338. CEUR-WS.org, online CEUR-WS.org/Vol-2654/paper26.pdf