The global impact of the COVID-19 pandemic has been significant, which requires data analysis to understand trends and patterns. However, this endeavor is challenging due to the complex transmission dynamics and diverse factors that influence the virus's spread. The data associated with COVID-19 is extensive and constantly evolving, and extracting meaningful insights from it is dificult. Therefore, the objective of this study is to analyze the impact of COVID-19 in various European countries, to identify common patterns, and to make predictions within the relevant context. To accomplish this, we used clustering techniques to reveal patterns in COVID-19 cases among European countries. The implementation involved cluster analysis to estimate labels based on cluster size and density while considering relevant background information. Subsequently, a classification model was applied to the labeled dataset. Using the K-Prototypes algorithm and leveraging the Silhouette score for identification, we determined the optimal number of clusters. These clusters were then combined based on density, and the degree of sparsity was assessed. As a result, two clusters emerged: one labeled as "low chance of infection" and the other as "high chance of infection." Using these results, we implemented a classification algorithm, achieving an accuracy rate of 90%. For this study, we gathered data from five diferent sources, consolidating them into a single dataset. Our findings demonstrate that combining COVID-19 datasets with diverse features enables trend analysis, while the use of clustering algorithms facilitates successful label identification in unsupervised learning scenarios involving unlabeled data. The density and size of clusters prove valuable in estimating labels, enhancing our overall understanding of the data. Our code is publicly available here.
implemented various measures such as decontamination, curfews, travel bans, and vaccination campaigns.
The COVID-19 pandemic has had a profound impact on As governments implement various containment and the global population, causing significant disruptions in social distancing measures, the demand for healthcare healthcare systems, industries, and societies around the systems has increased significantly. This poses a chalworld. To control the spread of the virus and alleviate lenging problem in efectively managing infected papressure on healthcare systems, numerous countries have tients in hospitals. Having an efective modeling method implemented strict measures. In Europe, like in other re- that can identify patterns and predict the spread of the gions, the COVID-19 outbreak emerged in January 2020 virus within the population would be highly valuable for and quickly escalated, leading to a surge in cases and preparing and formulating health and economic policies fatalities in hospitals [1]. While some European coun- for governments, administrators, and decision-makers. tries are currently experiencing new waves of infections, This would aid in slowing down or halting the spread of others are still dealing with a relatively low number of the virus. With the increase in cases of COVID-19 and COVID-19 cases. Throughout the epidemic, France, Italy, the availability of more data, several studies have utilized Spain and the United Kingdom have documented a sig- mathematical models [3], [4], [5] to analyze the spread nificant number of cases and fatalities[ 2]. To combat the of the virus. In addition, [6], [7] have also used LSTM virus, the European Union and many member states have models to forecast. However, these models often rely on outdated data from the same country, which limits ITAT’23: Information Technologies – Applications and Theory, Septem- their efectiveness. In a cluster analysis study [ 8], simiber 22–26, 2023, Tatranské Matliare, Slovakia larities were observed in the dynamics of the spread of * Corresponding author. the disease between countries such as Italy, France, and †$Thne0snedanui@thoinrfs.ecloten.thruib(uAte.dToesqauyaellvya.); ermiyasbirihanu@inf.elte.hu Germany, which implemented similar intervention strate(E. Birihanu); t.m.tashu@rug.nl (T. M. Tashu) gies. In another study [9], supervised machine-learning 0009-0004-0113-9298 (A. Tosayeva); 0000-0001-7081-0365 approaches were used to predict the future of COVID-19. (E. Birihanu); 0000-0002-4498-2486 (T. M. Tashu) Furthermore, [10] demonstrated the potential of Machine CPWrEooUrckReshdoinpgs IhStpN:/c1e6u1r3-w-0s.o7r3g ©ACt2tEr0i2bU2utCRioonpWy4r.0igoIhnrttekfornsrahtthiooisnppaalp(PCerCrboByYcite4s.0ea)u.dthionrsg.Usse( CpeErmUittRed-uWndeSr.Correagti)ve Commons License Learning and cloud computing to improve the prediction of epidemic growth. consider qualitative data, which can influence the spread
Most existing works are confined to forecasting within of the disease. specific countries or regions, overlooking the inclusion of In the study conducted by Lai et al. [14], the authors newly reported cases, recovery rates, and mortality num- analyzed the incidence and mortality rates of 57 countries bers that vary across diferent countries and over time. in 2020. They used Spearman’s rank-order correlation to In this study, we utilized COVID-19 data from diverse examine the relationship between cases and deaths. Howsources. We employed clustering techniques for pattern ever, this research overlooked trends and patterns, which recognition and applied the concept of estimating class can be crucial for further analysis. Several other research labels derived from the work of [11] for estimating class works have focused on clustering COVID-19 cases for labels. Subsequently, we applied classification methods diferent countries, employing the same K-Means algoto develop, train, and assess a supervised model aimed at rithm. predicting COVID-19 cases. Labeling unlabeled data has been a significant area of
The remainder of this study is organized as follows. research, with various ideas and approaches proposed. Section 2 discusses the existing research conducted in According to Fredriksson et al., [15], approximately 80% the relevant area, along with its limitations. Section 3 of engineering tasks in a machine learning (ML) project outlines in detail the proposed research methodology. involve data preparation and labeling. Data preparation Section 4 presents information on the datasets used in the and labeling often require extensive efort due to incomresearch, along with the preprocessing and analysis steps plete datasets or the lack of labels for some or all intaken. Section 5 provides a comprehensive overview of stances. Moreover, even when labels are available, they the conducted experiments. Finally, Section 6 presents may not be of good quality, leading to incorrect or parthe conclusions drawn from the research and outlines tially correct labels for data points. High-quality labels potential future work in the given research area. are crucial for successful supervised machine learning, as the model’s performance during operations is directly influenced by the quality of the training data. 2. Related Works Diferent techniques have been suggested by previous researchers to address the labeling challenge. Cui et al.
Clustering techniques have been utilized by researchers [16] proposed an approach where samples are divided since the initial spread of COVID-19 cases. These tech- into clusters, and classification models are applied to niques have been instrumental in grouping and identi- each cluster based on the dataset’s behaviors. The results fying distinct patterns to discern diferences or similar- from each cluster are then combined. To improve clasities among country-specific cases. Cheema et al. [ 12] sification performance, the swarm algorithm was used proposed the implementation of the K-Means algorithm for clustering, classification, and ensemble learning. The to establish diverse patterns among countries based on cluster-based ensemble learning method proved efective various features such as disease prevalence, health sys- with cross-validation practices. However, this research tems, and environmental indicators. The elbow method only utilized labeled samples and did not consider unlawas employed to determine the optimal number of clus- beled ones. ters, considering the sum of squared distances between In another study by Kusumaningrum et al., [17], Chisamples and their closest cluster centers. Their study Square was used for labeling with the assistance of Ksuccessfully demonstrated the reliability of Centroid- Means clustering. The homogeneity test was conducted based Partition clustering in identifying patterns among using the Silhouette coeficient, followed by employing country-specific cases. However, some limitations of this the Chi-Square Test for automatic cluster labeling in genresearch include the need to consider both categorical eral. and numerical features, as K-Means may not be the most In the study conducted by Yogesh [9], the aim was to suitable choice in such cases. Additionally, as the data develop an LSTM (long-short-term memory) model for only covers the year 2021, it may not accurately reflect forecasting COVID-19 deaths and cases, specifically for the current situation for further analysis. Gohari et al. Italy and the United States. The model was subsequently [13] also employed the K-Means algorithm to analyze lon- evaluated using data from Germany, France, Brazil, India, gitudinal patterns of change in quantitative COVID-19 and Nepal. On the other hand, Zeroual et al. [18] conincidence and mortality rates. They utilized dimension- ducted a comparative study of five deep learning methods reduction techniques to identify correlations between (RNN, LSTM, BILSTM, GRU, and VAR) to forecast the features, which enhanced the model’s performance. The number of new cases and recovered cases. However, it research successfully identified three distinct patterns should be noted that both researchers focused on a limthrough experiments and compared diferent trajectories. ited set of features during their investigation, potentially Although this research presents an innovative approach limiting the scope and comprehensiveness of their findin the field, it is limited to quantitative data and does not ings.The work by [19] introduced the concept of creating a transmission dynamics predictor that exploits tempo- In our research, we aimed to assess the prevalence of ral variations among diferent countries in relation to specific clusters and determine their commonality. To the disease’s spread. This is significant because certain achieve this, we utilized the k-prototypes clustering techcountries encountered outbreaks before others. How- nique, which is suitable for both categorical and numerever, it’s important to note that the data collected by the ical data. We manipulated the size and density of the researchers spanned only a duration of three days. clusters by varying the values of parameters and . By applying the k-prototypes clustering algorithm, we were able to cluster the data efectively. Following the clus3. Methodology tering process, we assigned a class label to each cluster based on the features it contained. In COVID-19 cases, 3.1. Pattern Recognition infections are categorized as high or low chance of infecAssociating a classification with a label is known as recog- tion. We assign labels to clustered dataset points using nition. Pattern recognition, as the science of drawing two rules: (1) Dense or sparse clusters are designated conclusions based on data, aims to categorize items or as "high chance," and (2) while others are categorized as events into groups based on shared characteristics. In "low chance." This involves classifying clusters into "low" our work, we are utilizing clustering, which falls under or "high" likelihood of infection groups based on their unsupervised learning, to create patterns and uncover characteristics. As a result, both the cluster and the data commonalities among the data. Clustering is an efective points contained within it are assigned identical labels, technique for identifying inherent structures or group- reflecting their infection likelihood. ings within a dataset without the need for pre-existing In general, if a cluster exhibits high density or sparsity, labels or categories. it is labeled as having a high chance of infection. On the other hand, if a cluster is not extremely dense or sparse, 3.2. Label Estimation it can be labeled as having a low chance of infection. This approach allows us to categorize clusters based on their Labeling data for COVID-19 manually is a time- characteristics and assign relevant infection likelihood consuming task that demands significant human re- labels accordingly. sources. However, our proposed approach tackles this challenge by estimating relevant labels for data points, 3.3. Context Prediction enabling supervised learning without the need for explicit user-provided labels. To separate the instances in the training data into appro
When it comes to defining the exact formula for small priate classes, SVM is used. SVM utilizes a hyperplane clusters, it is important to note that there is no univer- defined as + = 0, where is the weight vector sally accepted formula. In our approach, we consider the and is the bias term. The marginal hyperplanes, 1 number of data objects within a cluster to determine its and 2, are given as [20]: size. Let us assume that we have clusters and a cluster set containing data points. 1 : ( + ) = 1
Let the number of data points in each cluster be and be the parameter used to determine whether a cluster 2 : ( + ) = − 1 is considered small or not. a cluster is considered small Thus, correctly classified points satisfy the inequality: if (1)
( + ) ≥ 1 || < · where || is the number of data points of the cluster. For example, =0.2 indicates that if a cluster contains less than 20% then the cluster is considered small. As a next step, we are exploring whether the cluster is sparse or dense. For partitioning-based clustering, we used the sum of squares within the cluster .
In order to find the degree of sparsity, we have , and we assume that the cluster is sparse when: < · median()
(2) where is the set of for all . As a result, if =2.0 means that of a cluster is greater than 2.0 · () and it is sparse.
In SVM, the margin refers to the distance between the marginal hyperplanes, also known as the decision boundary. Specifically, for a linear SVM, the margin is equal to |2| , where represents the weight vector of the hyperplanes. Support vectors are the data points that lie on either the 1 or 2 hyperplanes, which denfie the margin. These data points play a crucial role in determining the position and orientation of the decision boundary.
1and [2] Our World in Data 2. In total, five datasets were used for our investigation.
number of newly reported COVID-19 cases and deaths by EU/EEA country." This dataset covers the period from February 2020 to October 2022 from 30 European countries. The second dataset contains vaccination information, providing details on the COVID-19 vaccination progress across diferent countries. The third dataset includes information on the Gross Domestic Product (GDP) of countries impacted by COVID-19. The fourth dataset focuses on travel restrictions implemented by each country. The fifth dataset concentrates on school restrictions in European countries. It considers diferent levels and time periods of school closures, reopening, and other related restrictions. Finally, we created a comprehensive compilation of all five datasets, resulting in a final dataset with a total of 2,246,240 entries. This comprehensive dataset serves as the basis for our analysis and research ifndings. It encompasses a diverse range of numerical and categorical units with varying scales.
2022. However, there was a significant decrease in cases towards the end of 2022, indicating a decline in COVID19 cases. Figure 2 shows the overall death figures caused by the spread of COVID-19 from 2021 to 2022. Initially, during the early stages of this period, the number of deaths was notably high. This could be attributed to the limited knowledge and understanding of the pandemic, and the lack of efective solutions and methods to handle it. However, as vaccinations and other measures were implemented, the number of COVID-19 deaths began to decline. Nevertheless, in 2022, there were still considerable rates of deaths reported. Towards the end of 2022, there was a significant decrease in these rates, as several decisive steps were taken in response to the increased severity of the pandemic.
During the data preprocessing phase, we encountered missing values in some features of the dataset. To address these missing values, we applied imputation techniques, which involve estimating the missing values based on the available data. In our case, we found that median and mode imputation were suitable approaches. Therefore, we performed median or mode imputation wherever necessary, substituting the missing values with the median or mode of the corresponding feature. Additionally, we opted to drop insignificant missing values to ensure the integrity of the data.
for duplicate columns in the dataset. Any duplicate columns identified were removed from the dataset to avoid redundancy. Additionally, we applied normalization and standardization techniques to scale the data within specific ranges using Z-score standardization.
and comparability of the data for further analysis.
= −
value, is the dataset we want to normalize represents mean of dataset or column and demonstrates standard deviation of column or dataset.
Figure 3 shows the direct impact of vaccination on COVID-19 cases during the entire period of analysis. The results were obtained by calculating the total number of vaccinations, which involved summing the administered doses. We further analyzed the percentage of the population that received at least one dose of the vaccination.
In Figure 4, we observed a correlation between the daily new COVID-19 cases and vaccination rates. When the daily new cases were high, the percentage of vaccinated individuals remained low. Conversely, as the vaccination rates increased, the number of new cases decreased. As time progressed, with a gradual decline in the number of COVID-19 cases, the demand for vaccinations also decreased. These findings demonstrate the significant role vaccination plays in curbing the spread of COVID-19 and reducing the number of cases over time. vice versa. By utilizing the Pearson correlation coeficient and examining the correlation graph, we can gain insights into the relationships between diferent features and identify patterns within the data. This assists in understanding the underlying structure and dependencies among the variables, thereby aiding the cluster analysis process [21].
In Figure 4, we observe and interpret the correlation between various features. A positive correlation exists between Covid-19 cases and deaths (textbf0.62), as well as between Covid-19 cases and the Population feature (textbf0.22). On the other hand, we find a negative correlation between international school closures and Covid-19 cases (textbf-0.24). This indicates that when international school closures are implemented, there is a reduction in the number of Covid-19 cases. Similarly, a negative correlation exists between travel controls and Covid-19 cases (textbf-0.037), suggesting that as travel controls become more stringent, the number of Covid-19 cases decreases.
These correlations provide valuable insights into the relationships between diferent factors and Covid-19 cases, Figure 3: Vaccination Vs COVID-19 cases deaths, and control measures. Understanding these connections can aid in making informed decisions and formulating efective strategies to manage and mitigate the
In cluster analysis, various methods can be utilized, impact of the pandemic. such as Chi-Square, ExtraTreeClassifier, Forward Feature Selection, and Correlation-based Feature Selection [21].
For our analysis of the clustering method, we employed a 5.3. Experimentation correlation graph, specifically utilizing the Pearson Cor- In our experiment, we employed two approaches: Krelation Coeficient. This coeficient is a measure of the Prototypes [22] and SVM [23]. For both approaches, we strength of the relationship between diferent features. utilized specific hyperparameters, which are listed in The Pearson correlation coeficient ranges between -1 Table 1. These hyperparameters were chosen to optimize and +1. A value close to +1 indicates a strong positive the performance and accuracy of the models. correlation between features. In such cases, if one feature Rather than using the default parameters for Kincreases, the other feature is also likely to increase, and Prototypes and SVM, we chose to determine the best vice versa. Conversely, a value close to -1 indicates a hyperparameters through grid search. The specific pastrong negative correlation. In this scenario, when one rameter values we chose are presented in Table 1. feature increases, the other feature tends to decrease, and
= − max(, )
nearest cluster And, is the mean intra-cluster distance to all the points.
To assess the SVM model’s performance, we employed metrics such as accuracy, precision, and recall. =
+ + + + =
+ =
+
served the distribution of countries across diferent clusters as follows. Cluster0, Cluster1, Cluster4, Cluster5, Cluster7, and Cluster9 each consist of one country. Cluster10 and Cluster11 contain three countries each. Cluster6 is assigned to two countries and Cluster2 is assigned to four countries. Cluster8 includes nine countries and has the highest mean of confirmed cases. During the Label Estimation phase, all twelve clusters were analyzed based on their sparsity and density. As a result, it was determined that Cluster2 and Cluster8 exhibited factors (4) (5) (6) (7) associated with a low risk of infection. The specific characteristics of these clusters contributed to a lower risk of COVID-19 infection.
On the other hand, the remaining clusters were combined to represent the high-risk category. These clusters likely had higher case densities and lower overall population vaccinations, indicating a higher risk of COVID-19 spread in those regions. By categorizing the clusters into high and low-risk categories, we gain valuable insights into the diferent patterns of COVID-19 prevalence and can further explore factors contributing to these variations. As indicated in Table 2, certain countries are present in multiple clusters. This phenomenon can be attributed to the fact that certain regions within a country pose higher risks, while other areas exhibit lower risks. This pattern emerges due to variations in risk levels across diferent parts of the same country.
For context prediction, we employed the Support Vector Machine (SVM) algorithm. We split the dataset into a testing set comprising 25 percent of the data and a training set comprising 75 percent. The table below dis- accuracy and efectiveness of our clustering analysis. By plays the performance metrics of the model. The SVM refining data scaling, enhancing labeling techniques, and model achieved an accuracy of 0.84 , indicating that it exploring advanced label estimation approaches, we can correctly predicted the context in 85 percent of the cases. advance our understanding of COVID-19 patterns and The precision, which measures the proportion of cor- provide valuable insights for future studies in this field. rectly predicted positive instances, was 0.90. The recall, representing the proportion of actual positive instances correctly identified, was 0.71. These metrics provide in- 8. Conclusion and future work sights into the performance of the SVM model for context prediction.
COVID-19 data has garnered significant attention in various research domains. However, to the best of our knowledge, the integration of clustering techniques for pattern estimation, label exploration, and context prediction has not been previously investigated. We have successfully achieved our objective of grouping countries based on various COVID-19 features, labeling the data using data characteristics, and implementing context prediction. However, there is still much more to be done in this research. One crucial aspect that deserves consideration is the integration of heterogeneity and coherence of the clusters to enhance accuracy. Combining clustering results from multiple algorithms or incorporating ensemble methods can lead to more robust and reliable clustering outcomes.
In our work, we have utilized the K-Prototypes algorithm for handling both numeric and categorical features.
However, there are other algorithms, such as the Cluster Ensemble algorithm (CEBMC), that can handle both types of features and may provide additional insights. Exploring and implementing multiple algorithms can lead to a comprehensive understanding of the data and improve the analysis. Furthermore, to enhance the performance of the model, we aim to incorporate multiple clustering algorithms in a combined manner. Ensemble learning techniques, where several clustering algorithms work together, can produce more accurate and stable results.
Additionally, we seek to enhance the robustness of label estimation to ensure more accurate results. Exploring and implementing various label estimation techniques, considering the characteristics of the dataset, can lead to more meaningful clustering analysis.
Overall, our research has made significant progress in the domain of COVID-19 data analysis using clustering techniques. However, there are still exciting avenues to explore and improvements to be made. By addressing the limitations and considering the integration of multiple algorithms and ensemble methods, we can advance our understanding of COVID-19 patterns and contribute valuable insights to the research community.
Based on the results obtained from our experiments, it is important to acknowledge and discuss the potential challenges and areas for improvement in our work. One significant challenge we encountered was obtaining positive results for the Silhouette Score, which is an important measure to evaluate the quality of cluster grouping. Initially, we faced negative or low scores due to the lack of appropriate scaling in our dataset. Scaling the data correctly is crucial to ensure reliable and meaningful results. Going forward, it is essential to pay attention to data scaling techniques and implement them efectively to enhance the accuracy of our analysis.
Another challenge we faced was the labeling of the dataset. Since the initial dataset did not include labels, we had to undertake the task of labeling the data ourselves.
Accurate labeling is essential for cluster analysis, as it provides meaningful interpretation and understanding of the clusters. It required careful analysis of the features of each country to assign appropriate labels. In future research, it would be valuable to explore automated or semi-automated labeling techniques that can expedite the process and enhance the accuracy of labeling. Furthermore, while label estimation techniques were not the focus of our study, it is an aspect that could beneift from improvement. Considering the specific features of the dataset, exploring alternative label estimation approaches can provide more nuanced and detailed analysis objectives. Incorporating advanced label estimation techniques could improve the overall analysis and provide 9. Online Resources more comprehensive insights into the patterns and dynamics of COVID-19 spread. Our dataset and code are publicly available here. at https:
Addressing these challenges and limitations in our //github.com/Tsegaye-misikir/NCC. research will contribute to further improvements in the