Method of shapelet discovery for time series ordinal classification

Current methods for time series ordinal classification (TSOC) methods suffer from low efficiency because the measures used to evaluate the quality of the shapelet need to calculate Information Gain from the Euclidian distances between the shapelet and time series, which incurs tremendous computation for large datasets. This paper introduces a novel method of shapelet discovery for TSOC in which a new measure is adopted, which takes into account the coverage concentration and dominance of shapelet on SAX-represented time series datasets. Moreover, a trie-tree is constructed based on all candidate shapelets and aims to discover a diverse set of high-quality shapelets. The experimental results demonstrated the effectiveness and efficiency when compared to eight SOTA algorithms for time series classification/ordinal classification.