Time series clustering and characterisation

Abstract

In today’s data-driven society, the volume of time series data generated every second is enormous and is applied in various domains - from financial markets and sales forecasting and public health surveillance systems. Time series is a collection of data points recorded or observed at successive points in time. Time series clustering refers to grouping a collection of time series into clusters based on their temporal patterns. Clustering time series helps uncover shared patterns, trends and anomalies. In this report, I will focus on two main time series clustering methods: (1) shape-based clustering and (2) feature-based clustering along with various clustering algorithms such as K-Means, hierarchical clustering and DBSCAN. Evaluation metrics such as Silhouette score and Dunn index will be used to evaluate the clustering results. These methods and algorithms will be tested on the COVID-19 dataset and the Mpox dataset, which are major global health events in recent years. The clustering results obtained may provide insights in the development pattern of the diseases and aid in decision making when implementing new public health policies.