Dynamic Early Time Series Classification Network

Early time series classification (ETSC) is of great significance for time-sensitive applications such as disaster prediction and gas leak detection. This task aims to classify time series with the least timestamps at desired accuracy. Recently, deep learning methods in ETSC usually used Convolutional Neural Networks to extract local features from fixed-length sequences for classification, and then set a threshold according to extensive expert experience for early exiting. However, the vanilla convolution operations cannot adapt to the data characteristics effectively. Moreover, the length variability of samples is also underestimated. To handle these problems, a dynamic convolution strategy is proposed to generate data adaptive convolution kernels for different samples. Moreover, we use random truncation based data augmentation techniques to enhance the convolution kernels in adapting to the data length variability. Experimental results on eight univariate datasets demonstrate the promising superiority of the proposed method.