ROCKET with Dynamic Convolution for Time Series Classification

Time series classification is an important research topic due to its prominent applications in industry, medicine, and finance. While in the early 2000s, techniques based on dynamic time warping (DTW) dominated this field, many recent works are based on Random Convolutional Kernel Transform (ROCKET). In this paper, we aim at combining the advantages of DTW and ROCKET. In particular, we incorporate dynamic convolution into ROCKET, thus we call the resulting approach DynamicROCKET. We perform experiments on 10 publicly available real-world time-series datasets and demonstrate that our approach, DynamicROCKET, may lead to statistically significant improvement in terms of classification accuracy. In order to promote the use of DynamicROCKET, we made our implementation publicly available in our github repository at https://github.com/kr7/DynamicROCKET .