Time series clustering and anomaly detection of COVID-19 global cases and deaths

• Main Author: Liew, Zhi Li
• Other Authors: Ke Yiping, Kelly
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2022
• Subjects: Engineering::Computer science and engineering
• Online Access: https://hdl.handle.net/10356/156387
• Institution: Nanyang Technological University

The spread and fatality patterns of COVID-19 behaviour varies amongst all the countries around the globe due to a multitude of reasons such as governments imposing differing strictness levels of health safety and quarantine measures at different times, geographical factors such as the population density, people’s attitude towards the virus, and the emergence of different variants of the virus with varying transmissibility and mortality. Time series clustering and anomaly detection are important analyses that identifies patterns in data, provides insightful knowledge on the similarity and dissimilarity of COVID-19 behaviours of different countries and identifies countries with anomalous spread patterns. In this contribution, two conventional time series clustering methods, K-Means++ and Agglomerative Hierarchical Clustering were performed on the global COVID-19 confirmed cases and deaths for all 194 countries from 22 January 2020 to 22 February 2022, for the longest possible period of 2 years and 1 months. This contribution is arguably the first to perform clustering for the longest possible time period for COVID-19 data using these clustering algorithms, and for all 194 countries available in the dataset. For the K-Means++ clustering algorithm, two different distance metrics were utilized, namely Euclidean Distance and Dynamic Time Warping, and the performance of the two algorithms and clustering results were compared and analysed. Finally, after the identification of anomalous clusters and evaluation of the clustering results, point anomaly detection were used to identify anomalous data points in the time series for the selected countries, Singapore, Faroe Islands, Peru, and Grenada. Three machine learning algorithms namely Isolation Forest, Clustering Based Local Outlier Factor, and One Class Support Vector Machine were performed to detect anomalous points in each individual time series. The performance and results from the three algorithms were analysed and the common point anomalies detected by all three algorithms are extracted for a more holistic evaluation.