Time series learning network using fast dynamic time warping

Abstract

Mining time series data has been revived in the last decade due to the increasing availability of time series datasets. This paper presents an online incremental learning algorithm for time series data using fast dynamic time warping (FastDTW), called time series learning network (TSLN), which is a neural network model for online incremental learning of time series. It automatically learns appropriate prototypes from input data series which are of equal or unequal length, and it does not ask to predefine the amount of prototypes and similarity thresholds. Besides, TSLN allows the arising of the new prototype, which is much superior to other existing unsupervised neural network algorithm. We test our model with UCR time series datasets, and experimental results show that, from the respect of classification accuracy, the proposed TSLN gets better or approximate results relative to the state-of-the-art approaches. Meanwhile, TSLN greatly mitigates demanding CPU time, which is much faster than most of the state-of-the-art methods.