This paper presents a simple yet effective approach to visualizing ensemble multivariate time series as 3D traces. Ensemble multivariate time series data are common in many areas. This type of data contains large amount of information which is often crucial to both knowledge discovery and decision making. Visualization can be employed to help the researchers quickly gain insight from the data. First, we project all multivariate data points to a 2D projection plane with a dimension reduction algorithm. Then we expand the data points of any ensemble member back into a trace in the 3D space spanned by the 2D projection plane and time. The resulting 3D ensemble traces provide a holistic and consistent view of the original ensemble multivariate time series. These traces are useful for revealing differences between ensembles, identifying groups and outliers, and catching temporal trends. In addition, we interactively link the ensemble traces to a panel of single variable plots. The combined visualization of raw data plots and multivariate ensemble traces provide a unique perspective to patterns and trends. We studied 3 different dimension reduction algorithms, i.e., t-Distributed Stochastic Neighbor Embedding (t-SNE), classic Multidimensional Scaling (MDS), and Locally Linear Embedding (LLE). We demonstrated our approach with two different datasets and evaluated our methods with domain experts.
Swastik Singh, Song Zhang, William Andrew Pruett, Robert Hester, "Ensemble Traces: Interactive Visualization of Ensemble Multivariate Time Series Data" in Proc. IS&T Int’l. Symp. on Electronic Imaging: Visualization and Data Analysis, 2016, https://doi.org/10.2352/ISSN.2470-1173.2016.1.VDA-505