A tree-construction search approach for multivariate time series motifs discovery

This paper examines an unsupervised search method to discover motifs from multivariate time series data. Our method first scans the entire series to construct a list of candidate motifs in linear time, the list is then used to populate a sparse self-similarity matrix for further processing to generate the final selections. The proposed algorithm is efficient in both running time and memory storage. To demonstrate its effectiveness, we applied it to search for repeating segments in both music and sensory data sets. The experimental results showed that the proposed method can efficiently detect repeating segments as compared to well-known methods such as self-similarity matrix search and symbolic aggregation approximation approaches.