Improving U-shapelets Clustering Performance: An Shapelets Quality Optimizing Method

Abstract

Unsupervised shapelets (u-shapelets) are time series subsequences that can best separates between time series coming from different clusters of data set without label. Because of the high computational cost, the u-shapelets are prohibited for many large dataset. Nevertheless, almost all of the current methods try to improving the u-shapelets based clustering method through reducing the computation time of u-shapelets candidate set. In this paper, we proposed a novel method improving efficiency of u-shapelets in terms of improving the u-shapelets quality. There are three contributions in our work: firstly, we show that by using internal evaluation measure instead gap score can improve quality of u-shapelets. Secondly, a novel method was proposed that applying diversified top-k query technology to filter similar u-shapelets, especially selecting the k most representative u-shapelets on the entirely shapelets candidates. Lastly, extensive experimental results show that combining internal evaluation measure and diversified top-k u-shapelets technology, our proposed method outperforms not only u-shapelet based methods, but also typical time series clustering approaches.