A new minimalist time series reduction technique

We propose in this paper a minimalist time series reduction technique that comes with a new distance. The proposed technique relies on the extraction of a minimal number of salient features of a time series, which are leveraged by the new distance to perform time series classification. We prove in this paper that the new proposed distance lower bounds the Euclidean distance and has a tighter lower bound than Piecewise Aggregation Approximation. Furthermore, we conduct experiments using standard UCR univariate time series datasets and include a comparative study of the new proposed technique with related SAX and deep learning based time series reduction and classification techniques. The experiments show that the new distance based time series classification enjoys better accuracy results than classifications based on well-known distances and is competitive to deep learning based time series classification techniques.