StreamStory: Exploring Multivariate Time Series on Multiple Scales

In visualizing multivariate time series, it is difficult to simultaneously present both the dynamics and the structure of the data in an informative way. This paper presents an approach for the interactive visualization, exploration, and interpretation of multivariate time series. Our approach builds an abstract representation of the data based on a hierarchical, multiscale structure, where each scale is modeled as a continuous time Markov chain. All the aspects of the visualization are designed with the multiple scales in mind. Encoding visual cues consistently across scales enables intuitive exploration of the different scales, allowing a user to quickly find appropriate scales for their data. The construction uses a combination of machine learning methods so that it requires minimal user input. We also present a number of coordinated views and tools which help the user understand how structure in the abstract representation maps to the data. Some of these include attribute distribution and time histograms, a time series matrix providing a cross-scale historical view of the data, and automatic labeling of the states. The visualization of the representation and the associated tools are integrated into an interactive, web-based tool called StreamStory. Using this tool, we show how this approach can be used to understand and find interesting long term and recurrent behavior in data using four different datasets, coming from weather data (i.e. rainfall and temperature), traffic monitoring sensors, GPS traces and wind velocity data. These represent a wide range of data and complexity but all show various interesting recurring patterns which we interpret with the help of StreamStory.