A Context-Aware Distance Analysis Approach for Time Series

Dynamic Time Warping (DTW) is a widely used elastic distance measure for time series. It can warp the time axis to cope with local time shift, but it also causes singularities due to its warping ability. The singularities indicate pathological warping. One kind of approaches for limiting the singularities is to mechanically limit the warping ability without considering the values of time series, and these approaches are likely to miss a good warping path. Another kind of approaches is to use the derivative of the values of time series to limit the singularities, but the derivative is easily affected by noise, and the derivation is also de-informatized. This kind of approaches tends to miss a good warping path, too. The ideal situation is to achieve a balance between limiting the singularities and finding a good warping path. To this end, we propose Context-aware DTW (CDTW), which uses the context information of the current point in the time series, and can find the right warping path while limiting singularities. For illustrating that our idea can be easily applied to other elastic distances, we also introduce context information in MSM and propose Context-aware MSM (CMSM). The experiments on the UCR datasets show that our CDTW and CMSM can achieve better accuracy than the original DTW and MSM respectively, and demonstrate the effectiveness of our context-aware distance analysis approach for time series.