An Adaptive Filtering Technique for Time Series Search

Time series data has been rapidly aggregated in many domains, such as meteorology, astrophysics, geology, multimedia, and economics. Similarity search is a core module of the tasks of time series data mining, such as classification and clustering. Dynamic time warping (DTW) is a robust distance measure method for time series data, minimizing the effects of shifting and distortion in time. Unfortunately, DTW does not satisfy the triangle inequality, so that spatial indexing techniques cannot be applied. We propose an adaptive multilevel filter technique by using a novel lower bound technique based on DTW for time series, which measures the distance between original sequence reduced dimensionality by PAA approximation method and query sequence reduced dimensionality by GMBR representation approach. The thorough experimental results show that, comparing with state of the art method, the proposed technique yields bigger lower bounding distance, more tightness of bound, stronger power pruning ability and shorter run time.