PETSC: pattern-based embedding for time series classification

Efficient and interpretable classification of time series is an essential data mining task with many real-world applications. Recently several dictionary- shapelet-based methods have been proposed that employ contiguous subsequences fixed length. We extend pattern to efficiently enumerate long variable-length sequential patterns gaps. Additionally, we discover at multiple resolutions thereby combining cohesive vary in length, duration resolution. For construct embedding based on occurrences learn a linear model. The discovered form the basis for insight into each class series. pattern-based (PETSC) supports both univariate multivariate datasets varying length subject noise or missing data. experimentally validate MR-PETSC performs significantly better than baseline such as DTW, BOP SAX-VSM On series, our method comparably recent methods, including BOSS, cBOSS, S-BOSS, ProximityForest ResNET, only narrowly outperformed by state-of-the-art HIVE-COTE, ROCKET, TS-CHIEF InceptionTime. Moreover, PETSC current CIF none which are interpretable. scales large total training making predictions all 85 ‘bake off’ UCR archive under 3 h it one fastest available. particularly useful learns model where feature represents domain, human oversight ensure trustworthy fair financial, medical bioinformatics