Novel representation approaches for time series classification

This dissertation presents three novel Time Series Classification (TSC) frameworks—NSIBOF, ALTS, and ARTSC—that address key challenges in computational efficiency, interpretability, and robustness by improving the representation of time series data. Effective representation is crucial in TSC because it directly impacts the ability to identify meaningful patterns, reduce noise, handle temporal distortions, and support efficient and interpretable classification. The NSIBOF framework introduces a novel approach that combines scale-space theory with the bag-of-features technique. It identifies key points in the time series using the Significant Zero Crossings (SiZer) method to detect scale-space extrema and then builds a bag-of-features representation from local features around these points, which improves robustness to noise and distortions. The ALTS framework improves classification by using highly discriminative subsequences of the data, sometimes called shapelets, which are learned directly through an optimization process combined with fused lasso regularization. This approach avoids exhaustive searches, preserves class-discriminative patterns, and reduces computational complexity. The ARTSC framework integrates autoregressive (AR) modeling with a Random Forest (RF) classifier to extract features and dynamically adjusts parameters to handle both stationary and non-stationary time series, which improves robustness across domains. Experiments on benchmark datasets show that NSIBOF improves robustness to noise and distortions, ALTS offers better efficiency and accuracy than traditional subsequence-based classifiers, and ARTSC delivers a simple, effective TSC framework.