Location-Awareness in Time Series Compression

Abstract

We present our initial findings regarding the problem of the impact that time series compression may have on similarity-queries, in the settings in which the elements of the dataset are accompanied with additional contexts. Broadly, the main objective of any data compression approach is to provide a more compact (i.e., smaller size) representation of a given original dataset. However, as has been observed in the large body of works on compression of spatial data, applying a particular algorithm “blindly” may yield outcomes that defy the intuitive expectations – e.g., distorting certain topological relationships that exist in the “raw” data [7]. In this study, we quantify this distortion by defining a measure of similarity distortion based on Kendall’s T. We evaluate this measure, and the correspondingly achieved compression ratio for the five most commonly used time series compression algorithms and the three most common time series similarity measures. We report some of our observations here, along with the discussion of the possible broader impacts and the challenges that we plan to address in the future.