MMDL-Based Data Augmentation with Domain Knowledge for Time Series Classification

Xiaosheng Li, Yifan Wu, Wei Jiang, Ying Li, Jianguo Li

Abstract: Plenty of time series classification methods have been proposed in the past. Most methods utilize the labeled time series instances to build classifiers, ignoring the explicit domain knowledge. However, in real-world applications, practitioners may identify domain characteristics of the time series, and build the heuristic relationship between the class labels of the time series and these domain characteristics. In this paper, we investigate the possibility of incorporating the domain knowledge into time series classification for possible performance improvement. To this end, we propose a Modified Minimum Description Length (MMDL)-based data augmentation method to inject domain knowledge into time series classification. Based on the type of domain knowledge, the proposed method applies the MMDL shapes or residuals to augment the training data. Experimental results demonstrate that the proposed method consistently improves the classification accuracy across all tested datasets and achieves better results than other time series data augmentation methods.