Visual perception of discriminative landmarks in classified time series

Abstract

Distance measures play a central role for time series data. Such measures condense two complex structures into a convenient, single number – at the cost of loosing many details. This might become a problem when the series are in general quite similar to each other and series from different classes differ only in details. This work aims at supporting an analyst in the explorative data understanding phase, where she wants to get an impression of how time series from different classes compare. Based on the interval tree of scales, we develop a visualisation that draws the attention of the analyst immediately to those details of a time series that are representative or discriminative for the class. The visualisation adopts to the human perception of a time series by adressing the persistence and distinctiveness of landmarks in the series.