Conference Paper Time Series Classification Based on Random Shapelets

Suehiro, Daiki  ,  Kuwahara, Kengo  ,  Hatano, Kohei  ,  Takimoto, Eiji

Time-Series Shapelets are shown to be useful features to learn accurate classifiers for various time series. There are many algorithms for searching or optimizing shapelets, however, they use limited classes of features based on subsequences in the training data. In this paper, we consider an infinitely large set of sequences as a class of shapelets and discuss the generalization bound of the shapelets-based hypothesis class. In practice, instead of using infinitely many sequences, we propose a heuristic approximate algorithm using random sequences that allows us to obtain good classification rules and shapelets. In preliminary experiments over real data sets, we obtained accurate classifiers which are comparable to the state-of-the-art methods.

Number of accesses :  

Other information