Binary Shapelet Transform for Multiclass Time Series Classi cation


Authors: Aaron Bostrom and Anthony Bagnall

Volume 32 (2017)

Abstract


Shapelets have recently been proposed as a new primitive for time series classi cation. Shapelets are subseries of series that best split the data into its classes. In the original research, shapelets were found recursively within a decision tree through enumeration of the search space. Subsequent research indicated that using shapelets as the basis for transforming datasets leads to more accurate classi ers. Both these approaches evaluate how well a shapelet splits all the classes. However, often a shapelet is most useful in distinguishing between members of the class of the series it was drawn from against all others. To assess this conjecture, we evaluate a one vs all encoding scheme. This technique simpli es the quality assessment calculations, speeds up the execution through facilitating more frequent early abandon and increases accu- racy for multi-class problems. We also propose an alternative shapelet evaluation scheme which we demonstrate signi cantly speeds up the full search.