Improving time series similarity measures by integrating preprocessing steps

摘要

While many application papers involving time series data report about the beneficial application of filters, filtering (and preprocessing in general) plays at best a minor role in the proposals of similarity measures for time series or the studies that compare them. We investigate the performance of basic Euclidean distance with an integrated preprocessing (filtering with automatically derived filters (supervised or unsupervised) and rescaling) and demonstrate that such measures can better respond to typical problems in time series similarity. By accounting for differences in both domains (time and value) we overcome some limitations of elastic measures that focus on time only. Using the proposed measure on real datasets we can achieve performance gains comparable to those of switching from a lock-step measure (Euclidean) to an elastic measure (DTW).