Grouping Multivariate Time Series: A Case Study

We present a case study to demonstrate a process for grouping massive multivariate time series based on nonparametric statistical summaries aided by information visualization. We want a method that allows us to quickly find approximate groups in time series, both to identify typical aggregate behaviors and to find aberrant outliers. We use simple statistical summaries to capture the temporal nature and variability of the time series, as well as the interaction between the various multivariate components. Each individual time series is mapped to a fixed-length vector of summaries. The summary vectors are then clustered using any fast clustering algorithm like -means. Appropriate information visualization techniques are used at every stage to guide the analyst. Because the method is nonparametric, it is customizable and flexible, and it generalizes easily. When choosing the statistical summaries, we can incorporate domain knowledge that may enhance the clustering. We demonstrate with a massive real life telecommunications application.