Section Min-Hash Approximating Time Series Search based on Dynamic Time Warping

抄録

Dynamic Time Warping (DTW) is a well-known similarity measure between time series data. Although DTW can calculate the similarity between time series with different lengths, it is computationally expensive. Therefore, fast algorithms that approximate the DTW have been desired. SSH (Sketch, Shingle & Hash) is a representative hash-based approximation algorithm. It extracts a set of quantized subsequences from a given time series and finds similar time series by means of Min-Hash, a hash-based set similarity search. However, Min-Hash does not care about the location of set elements (i.e., quantized subsequences) in the time series, so that hash collisions have rather weak correlation with DTW. In this paper, to strengthen the correlation between hash collisions and DTW, we propose a new method termed Section Min-Hash that can couple the hash values with the positions of quantized subsequences. After quantizing subsequences in a time series based on Euclidean distance, Section Min-Hash explicitly specifies multiple sections within the time series and generates the hash values from all the sections.