Local feature based dynamic time warping

Time series is a ubiquitous form of data and the analytics of time series is attracting increasing interest recently. In the context of time series data mining, similarity measure has the fundamental importance because many data mining techniques depend on it. Dynamic time warping (DTW) is considered to be the most popular similarity measure for time series and it is alignment-based. However, it has its inherent weak point that it always tends to explain the variability in observed values by warping the time axis, hence leading to a pathological alignment. In this paper, we propose an algorithm framework to enable different kinds of local features instead of the raw value used by DTW to align time series and thus getting a more robust similarity measure. Almost any kind of local feature can be adopted in our framework and we start from the simplest ones, average, maximum, and minimum. One nearest neighbor (1NN) classification on UCR time series archive is carried out to demonstrate the superiority of our proposed method.