Time series similarity search based on Middle points and Clipping

Abstract

In this paper, we introduce a new time series dimensionality reduction method, MP_C (Middle points and Clipping). This method is performed by dividing time series into segments, some points in each segment being extracted and then these points are transformed into a sequence of bits. In our method, we choose the points in each segment by dividing a segment into sub-segments and the middle points of these sub-segments are selected. We can prove that MP_C satisfies the lower bounding condition and make MP_C indexable by showing that a time series compressed by MP_C can be indexed with the support of Skyline index. Our experiments show that our MP_C method is better than PAA in terms of tightness of lower bound and pruning power, and in similarity search, MP_C with the support of Skyline index performs faster than PAA based on traditional R*-tree.