[1J4-GS-2] Machine learning: Fundamental theory (1)

[1J4-GS-2-05] Time Series Classification with a Few Shapelets

〇Kento Kotera¹, Masayuki Numao¹, Ken-ichi Fukui¹ (1. Osaka university)

Keywords:Time Series Classification, Time Series, Shapelet

Research on subsequences called Shapelets has attracted attention. In recent years, it has shown significant improvements in classification accuracy while having the interpretability unique to Shapelets by using it as a feature transformation. However, applying these methods to large datasets requires hundreds of Shapelets. In this case, transparency of the classification basis, one of the advantages of Shapelets, is greatly impaired. What we want to know when interpreting a classification result is, at most, a few Shapelets. In this research, we propose a new classification method that can limit the number of Shapelets to a small number by combining the rules based on Shapelets and the deep learning method. Through experiments, it has been confirmed that the accuracy is comparable with that of the conventional Shapelets-based classification methods.