Extracting Features from Random Subseries: A Hybrid Pipeline for Time Series Classification and Extrinsic Regression

In time series classification (TSC) literature, approaches which incorporate multiple feature extraction domains such as HIVE-COTE and TS-CHIEF have generally shown to perform better than single domain approaches in situations where no expert knowledge is available for the data. Time series extrinsic regression (TSER) has seen very little activity compared to TSC, but the provision of benchmark datasets for regression by researchers at Monash University and the University of East Anglia provide an opportunity to see if this insight gleaned from TSC literature applies to regression data. We show that extracting random shapelets and intervals from different series representations and concatenating the output as part of a feature extraction pipeline significantly outperforms the single domain approaches for both classification and regression. In addition to our main contribution, we provide results for shapelet based algorithms on the regression archive datasets using the RDST transform, and show that current interval based approaches such as DrCIF can find noticeable scalability improvements by adopting the pipeline format.