Temporal MDS Plots for Analysis of Multivariate Data

Abstract

Multivariate time series data can be found in many application domains. Examples include data from computer networks, healthcare, social networks, or financial markets. Often, patterns in such data evolve over time among multiple dimensions and are hard to detect. Dimensionality reduction methods such as PCA and MDS allow analysis and visualization of multivariate data, but per se do not provide means to explore multivariate patterns over time. We propose Temporal Multidimensional Scaling (TMDS), a novel visualization technique that computes temporal one-dimensional MDS plots for multivariate data which evolve over time. Using a sliding window approach, MDS is computed for each data window separately, and the results are plotted sequentially along the time axis, taking care of plot alignment. Our TMDS plots enable visual identification of patterns based on multidimensional similarity of the data evolving over time. We demonstrate the usefulness of our approach in the field of network security and show in two case studies how users can iteratively explore the data to identify previously unknown, temporally evolving patterns.