This work focuses on the comparison and clustering of irregular, multivariate (i.e., with events over many dimensions)  sequences.
The classical Dynamic Time Warping (DTW) distance, and its extension Drop-DTW (DTW where dropping elements is possible when trying to match two sequences), can be used in this context.
Recently Drop-DTW was extended over timed sequences and applied to the clustering of care pathways. 
Motivated by a similar problem, in this work, we consider the extension of Drop-DTW over multivariate sequences, where match costs are set by dimensions. 
The clustering pipeline combines this new distance measure with the DTW Barycenter Averaging (DBA) method to estimate centroids under temporal alignment, followed by K-means.
We normalize both the value and time components in order to stabilize the hyperparameters and prevent any single scale from dominating the distance. 
We will discuss limitations and avenues for methodological improvement (Sakoe–Chiba windows,  parallelization), as well as 
evaluation metrics.