Scalable classifier-agnostic channel selection for multivariate time series classification

Accuracy is a key focus of current work in time series classification. However, speed and data reduction are equally important many applications, especially when the scale storage requirements rapidly increase. Current multivariate classification (MTSC) algorithms need hundreds compute hours to complete training prediction. This due nature which grows with number series, their length channels. In not all channels useful for task, hence we require methods that can efficiently select thus save computational resources. We propose evaluate two channel selection. Our techniques by representing each class prototype performing selection based on distance between classes. The main hypothesis enable better separation classes; hence, larger prototypes more useful. On UEA MTSC benchmark, show these achieve significant classifier speedup similar levels accuracy. Channel applied as pre-processing step before state-of-the-art saves about 70% computation preserved Furthermore, our efficient classifiers, such ROCKET, accuracy than using no or greedy forward To further study impact techniques, present experiments classifying synthetic datasets 100 channels, well real-world case dataset 50 both cases, result improved