[2G1-GS-2d] 機械学習：シーケンシャルデータの処理

[2G1-GS-2d-04] Time Series Classification using the Order of Shapelet Appearance

〇Kohei Fujioka¹, Masayuki Okabe¹ (1. Prefectural University of Hiroshima)

Keywords:time series classification, shapelet, multiple kernel learning

In this study, we propose a classification learning method that uses the appearance order information of Shapelet, which is a characteristic sub-series in time-series data. In the proposed method, the order relationship between Shapelets is quantified using the n-gram kernel that quantifies the similarity of the appearance order of words in the character string data. The generated Shapelet version of the n-gram kernel is integrated with the kernel generated by the conventional minimum distance feature by multi-kernel learning. Multi-kernel learning is performed by a method based on kernel alignment. As a result of comparing the proposed method with the conventional method using benchmark data sets for multiple time series data classification, the accuracy was improved and the standard deviation was reduced in all the used data sets, confirming the effectiveness of the proposed method.