Effects of dimensionality reduction techniques on time series similarity measurements

Time Series are ubiquitous, hence, similarity search is one of the biggest challenges in the area of mining time series data. This is due to the vast data size, number of sequences and number of dimensions that lead to a very costly querying process. In this paper, we demonstrate, for the first time, the use of three dimensionality reduction techniques (Random Projection (RP), Down sampling (DS) and Averaging (Avg)) in time series similarity searches. Two different similarity measurements are used for this investigation; Dynamic Time Warping (DTW) and Euclidean distance. A thorough study has been conducted in this paper based on very exhaustive experiments. Results show the individual performance of Avg, RP, and DS in the two similarity measurements in different dimensions. Simulation shows that a high similarity matching accuracy can still be achieved after a significant dimension reduction onto lower dimensions.