Optimization for Partitional Time-Series Clustering

Karolína Bakuncová, Luboš Marek
Statistika, 106(2): 152–163
https://doi.org/10.54694/stat.2025.47

Abstract
Time-series clustering is a convenient tool for analysing hidden structures in data. However, as is the case with clustering, it is possible to encounter a number of complications, especially with regards to the sensitivity to the initial algorithm conditions and the subjective choice of the number of groups. The aim of this article is to conduct an experiment using real life data of housing prices in the EU to suppress subjectivity, whether in terms of finding subgroups in the data or the validation of the result for partitional clustering. The proposed procedure is based on a modified bootstrapping principle, where the principle of stability via repetition is applied to the algorithm and its results. As such, this method is applied both to the group selection by monitoring the Calinski-Harabasz index and the final assignment of the resulting classification of clustered objects. The result of this process is a structure that has a better informative value about the relationships in the data.

Keywords
Partitional clustering DTW distance, k-means algorithm, time-series