Heinrich-Heine-Universität Düsseldorf

Clustering is a fundamental unsupervised learning technique for grouping similar instances in unlabeled datasets. When data evolves over time, clustering can be applied at each time point, resulting in an over-time clustering (ot-clustering). Several approaches incorporate temporal context into ot-clustering in different ways, all aiming to uncover unknown patterns in multivariate time series data. However, many of these methods require tuning of a weighting parameter for temporal context, which must be set manually or determined through computationally intensive algorithms, depending on the method. More recently, a study introduced a concept of over-time stability for such ot-clusterings and proposed a method that measures changes in clusterings over time, eliminating the need for a weighting parameter for temporal context. In this work, we propose an alternative to the existing definition of over-time stability and introduce a framework for evaluating ot-clusterings based on this definition. Beyond producing stability scores, our approach enables the visual analysis of ot-clusterings in a two-dimensional space, regardless of the number of dimensions in the original dataset. Our experiments demonstrate that applying the proposed framework yields alternative yet meaningful, ot-clusterings compared to not taking temporal context into account. Compared to the only other ot-clustering stability evaluation method called CLOSE, our approach offers a greater flexibility in defining temporal context, produces ot-clusterings with less granularity but also reduced noise, and achieves better runtimes.