Mitigating the influence of the curse of dimensionality on time series similarity measures

Time series are ubiquitous application domains that generate data including GPS, stock market, and ECG. Researchers concentrate on mining time series data to extract important knowledge and insights. Time series similarity search is a data mining technique that is widely used to compare time series data using similarity measurements, such as dynamic timewarping and Euclidean distance. The large number of sequences dimensions makes the mining process costly. Therefore, we need to extract fewer representative points, hence making the mining process manageable. In this paper, we investigate the application of three dimensionality reduction techniques (random projection, downsampling and averaging) on time series similarity search. Our study has been conducted based on very exhaustive experiments. Results show the performance of the reduction techniques on two similarity measures. Simulation shows that a high similarity matching accuracy can still be achieved after the reduction onto lower dimensions.