A Shapelet Selection Algorithm for Time Series Classification: New Directions

Abstract

Time series classification (TSC) has attracted significant interest over the past decade. One of the promising recent approaches are shapelet based algorithms, which are interpretable, more accurate, and faster than most classifiers. However, the high time complexity of shapelet selection process hinders its application in real time data procession. To overcome this, in this paper we propose a fast shapelet selection algorithm (FSS), which sharply reduces the time consumption of shapelet selection. In our algorithm, we first sample some time series from training data set with the help of a subclass splitting method. Then FSS identifies the local farthest deviation points (LFDPs) for sampled time series and selects the subsequences between two nonadjacent LFDPs as shapelet candidates. Through these two steps, the number of shapelet candidates is greatly reduced, which leads to an obvious reduction in time complexity. Unlike other methods which accelerate shapelet selection process at the expense of reducing accuracy, the experimental results demonstrate that FSS is thousands of times faster than ST with no accuracy reduced. Our results also demonstrate that our methods is the fastest method among the shapelet-based methods which have the leading level of accuracy.