Model approach for custering count time series
Conference
65th ISI World Statistics Congress 2025
Format: CPS Abstract - WSC 2025
Keywords: clustering, count, series, time
Session: CPS 5 - Time Series Analysis
Tuesday 7 October 4 p.m. - 5 p.m. (Europe/Amsterdam)
Abstract
The clustering of time series has proven to be of interest in various fields, ranging from economics and finance to environment and medicine, among others. The objective is to group similar items according to a criterion suitable for the problem under analysis. The complexity of clustering time series is substantially higher than this task in the context of cross-sectional objects.
Clustering can be performed according to three approaches: observation-based, feature-based, and model-based. This last approach considers dissimilarities between the series by evaluating the proximities between the fitted statistical models, such as parameter estimates.
Much of the work developed over the past 3-4 decades has been conducted within the framework of continuous-valued time series, with few studies on clustering for count time series. In this work, the aim is to establish and apply model-based clustering to appropriately defined discrete-valued time series, particularly those that allow for overdispersion and/or zero inflation. The idea is to use a finite mixture model, that accommodates the mentioned characteristics, and several existing techniques such as the selection of the number of clusters, estimation using expectation-maximization and model selection are applicable.