Time Series Similarity Search Based On Adaptive Cost Dynamic Time Warping Distance

With the development of a new round of information science and technology, we have entered the era of big data. Time-series data is one aspect of big data, and in recent years, time series data mining has received more and more attention.Time series similarity search is the basis of time series data mining. Dynamic time warping distance as time-series similarity measure has a better robustness and broad applicability. We use dynamic programming method minimizing the cumulative distance to solve dynamic time warping distance. This approach has two drawbacks, one is that, the original dynamic time warping distance in order to obtain a minimum cumulative distance may map a point in a time series to a plurality of points in another time series. That is, the original time series are overstretched. Another is that actually a time-series observations constitutes an image, while the original dynamic time warping distance only considers the results of global corresponding and local shape is not considered. In both cases, it may bring inaccuracies to characterize similarity between time series.In order to overcome the first drawback, a variety of dynamic time warping distance variants have been proposed, such as search window restriction method. However, these variants bring many new parameters to be determined, while the adaptive cost dynamic time warping distance in this paper can not only find the path through the process of change the cost of the current step to control the degree of distortion of time series, but just bring one new parameter. The new parameter is set as a constant value when testing the adaptive cost dynamic time warping distance on 17 UCR data sets, and you can get better 1NN classification accuracy.Considering the second drawbacks, some scholars have proposed derivative dynamic time warping distance. They use the original time series difference to describe the local shape of the time series. The shape context descriptors are good at describing the characters of images. Z. Zhang and P. Tang proposed dynamic time warping distance based on two-dimensional shape context, and has better 1NN classification accuracy compared to the derivative dynamic time warping distance. But dynamic time warping distance based on two-dimensional shape context can only handle one-dimensional time series. Two-dimensional time series can be find in many aspects of our production and social life. Therefore, this paper presents a dynamic time warping distance based on three-dimensional shape context. This new distance provides a new idea in searching the similarity between two-dimensional time series. Through tests on datasets from UCI, we proved that dynamic time warping distance based on three-dimensional shape context is a great measure of similarity between two-dimensional time series.