DendroTime: Progressive Hierarchical Clustering for Variable-Length Time Series

Kurzfassung

Many effective dissimilarity measures for variable-length time series, such as DTW, MSM, or TWED, are expensive to compute because their runtimes increase quadratically with the time series' lengths. When used in hierarchical agglomerative clustering algorithms that need to compute all pairwise time series dissimilarities, they cause slow runtimes and do not scale to large time series collections. However, there are use cases, where fast, interactive hierarchical clustering is necessary. For these use cases, progressive hierarchical clustering algorithms can improve runtimes and interactivity. Progressive algorithms are incremental algorithms that produce and continuously improve an approximate solution, which eventually converges to the exact solution. In this paper, we present DendroTime, the first (parallel) progressive clustering system for variable-length time series colections. The system incrementally computes the pairwise dissimilarities between the input time series and supports different ordering strategies to achieve progressivity. Our evaluation demonstrates that DendroTime's progressive strategies are very effective for clustering scenarios with expensive time series dissimilarity computations.