Dimensionality Reduction for Visualization of Time Series and Trajectories

Abstract

Visualization is essential for data analysis and it is particularly challenging for data in high-dimensional space, especially for temporal information. Many techniques have been employed in an attempt to transform multivariate time series data to one-dimensional data by reducing the number of features in order to visualize their time-dependent behaviors. However, the applicability of these approaches is restricted to a limited number of data instances that can be visualized simultaneously. We present a technique to visualize time series and trajectories that overcomes these limitations by transforming these data into subspaces which allows data analysts to easily select the instance of interest from a bunch of data. The benefits of our proposed method are threefold: it provides (i) a visual representation of time-dependent data in a massive amount simultaneously, (ii) a very concise feature representation and (iii) an easy identification of anomalies. The results are demonstrated by employing this technique to various data traits from public archives, they are (i) univariate time series data from the UCR archive, (ii) multivariate time series data from several sources, and (iii) human motion trajectories from two motion capture (MoCap) datasets.