Biclustering Multivariate Time Series

Sensor networks are able to generate large amounts of unsupervised multivariate time series data. Understanding this data is a non-trivial task: not only patterns in the time series for individual variables can be of interest, it can also be important to understand the relations between patterns in different variables. In this paper, we present a novel data mining task that aims for a better understanding of the prominent patterns present in multivariate time series: multivariate time series biclustering. This task involves the discovery of subsets of variables that show consistent behavior in a number of shared time segments. We present a biclustering method, BiclusTS, to solve this task. Extensive experimental results show that, in contrast to several traditional biclustering methods, with our method the discovered biclusters respect the temporal nature of the data. In the spirit of reproducible research, code, datasets and an experimentation tool are made publicly available to help the dissemination of the method.