Detecting trend motifs: an efficient framework for time series motif discovery

Abstract

The task of finding similar patterns in a long time series, commonly called motifs, has received continuous and increasing attention from diverse scientific fields. Although numerous approaches have been proposed for motif discovery, they cannot discover the motifs in an exact and efficient manner. Furthermore, domain knowledge is required from the experts for those methods to predefine the pattern length, which is also quite objective. In addiction, it is very time-consuming to extract the exact motifs and sometimes the extracted motif has no specific meanings. Especially in the field of financial and hydrology, many studies are focused on whether there is a fixed pattern including trend information hidden in the data. To address the above problems, we proposed a framework to automatically discovery the trend motifs without predefining the length of patterns. It has four main steps, (1) singular spectrum analysis is first applied to removed noise; (2) segmentation by extracting extreme points is then employed to automatically obtain the unequal length of time series pattern; (3) symbolic aggregate approximation is introduced to discretize the data and transform them into string sequences; (4) finally, the trend motifs are selected by measuring their similarity. Experimental results on the real-world time-series datasets reveal that our framework fit well in different circumstances, indicating our proposed framework is effective for trend motif discovery.