MS-SRALAT: Multi-granularity SubStructure-aware Representation Learning Algorithm for Time-series

Abstract

Time-series representation is essential in many data mining algorithms, such as clustering, classification, motif discovery, which have been used to discover knowledge from the time-series data. In this paper, we propose an algorithm to learn the semantic representation of a symbol sequence which is generated corresponding to a time-series by an approximation algorithm that can capture the structure of original data. However, the granularity of structure (coarse-to fine-grained) approximated by such an algorithm is defined by a parameter which affects the quality of resulting representation, and therefore impacts the performance of its subsequent tasks. We then propose a multi-granularity substructure-aware representation learning algorithm for time-series (MS-SRALAT) which is an ensemble model that incorporates the trained models with different granularity to produce more robust representations. The resulted experiments on the benchmark datasets showed the superiority of MS-SRALAT over single-granularity learning models, and comparable performances compared to the exact baseline methods while suggesting good scalability for the similar search task.