The formation and analysis of clusters in multivariate time series can reveal interesting patterns and complex correlations in temporal data. However, traditional clustering methods based on distance metrics fall short to discover interpretable characteristics and structures reflected by these clusters. This paper provides a new method for semi- supervised time point clustering based on the temporal proximity of time points and the correlation of their corresponding values. For this purpose, we utilize CoExDBSCAN, a recently developed density-based clustering algorithm with constrained expansion. CoExDBSCAN allows to identify clusters of temporal neighbourhoods that are only expanded with regards to a priori constraints in defined subspaces. Adopting this algorithm to time series data and grouping segments with similar correlations allows us to find accurate and interpretable structures. We provided a comparison to state-of-the-art methods and verification of our approach on a synthetic dataset and an experimental evaluation on a real-world dataset. The experimental assessment shows that our clustering results can further serve as an effective basis for time series classification.
Article ID: 2021L05
Venue: Canadian Conference on Artificial Intelligence
Publisher: Canadian Artificial Intelligence Association