Adaptive Law-Based Transformation (ALT): A Lightweight Feature Representation for Time Series Classification

Time series classification (TSC) is fundamental in numerous domains, including finance, healthcare, and environmental monitoring. However, traditional TSC methods often struggle with the inherent complexity and variability of time series data. Building on our previous work with the linear law-based transformation (LLT) - which improved classification accuracy by transforming the feature space based on key data patterns - we introduce adaptive law-based transformation (ALT). ALT enhances LLT by incorporating variable-length shifted time windows, enabling it to capture distinguishing patterns of various lengths and thereby handle complex time series more effectively. By mapping features into a linearly separable space, ALT provides a fast, robust, and transparent solution that achieves state-of-the-art performance with only a few hyperparameters.

💡 Research Summary

The paper introduces Adaptive Law‑Based Transformation (ALT), an extension of the previously proposed Linear Law‑Based Transformation (LLT) for time‑series classification (TSC). While LLT achieved competitive accuracy by extracting linear “laws” from training data via time‑delay embedding and spectral decomposition, it was limited to fixed‑length windows and therefore struggled to capture patterns that occur at multiple temporal scales. ALT overcomes this limitation by employing variable‑length, shifted time windows parameterized by a triplet (r, l, k): r denotes the window length, l the embedding dimension, and k the shift step. For each (r, l, k) combination, subsequences are extracted, evenly sampled to obtain 2l‑1 points, and embedded into an l‑dimensional delay‑matrix S. Spectral decomposition of S yields eigenvectors; the one associated with the smallest absolute eigenvalue (the “v‑shapelet”) represents the direction of minimal variability and is stored as a column in a shapelet matrix P. P is constructed separately for each class, thereby forming class‑specific partitions.

During transformation, a test instance is similarly embedded into a matrix A (size o × l, where o depends on r, l, k and the series length) and right‑multiplied by the appropriate P to obtain O = AP. Elements of O are driven toward zero by the competing class‑specific shapelets, effectively mapping the original series into a linearly separable feature space. From O, a rich set of statistical descriptors—percentiles, means, variances, higher‑order moments—is computed for each class partition, yielding a final feature tensor of size m × c × n × g (m channels, c classes, n statistical measures, g different (r, l, k) settings).

The authors evaluate ALT on eleven benchmark datasets from the UCR archive, covering both univariate and multivariate series of varying lengths and class counts. After generating ALT features, they train conventional classifiers (K‑Nearest Neighbours and Support Vector Machines) using Bayesian hyper‑parameter optimisation (30 iterations, 5‑fold CV). Results show that ALT consistently attains high validation and test accuracies, achieving perfect scores on several datasets (e.g., BasicMotions, Coffee, GunPoint4). Transformation times remain practical (seconds to a few minutes), with the most time‑consuming steps being shapelet vector generation and spectral decomposition. Classification after transformation is extremely fast. Compared against a strong neural‑network baseline (optimisable MLP with 500‑step Bayesian optimisation) and a range of state‑of‑the‑art TSC methods reported in the literature, ALT either matches or surpasses them while using far fewer hyper‑parameters and considerably less optimisation effort.

Key strengths of ALT include: (1) the ability to capture multi‑scale subsequence patterns via variable‑length windows; (2) a linear mapping that preserves interpretability—shapelet vectors can be inspected to understand discriminative patterns; (3) low computational overhead relative to deep learning approaches, especially during the classification phase; and (4) minimal hyper‑parameter tuning (essentially the (r, l, k) triplet). Limitations are noted: the current study selects (r, l, k) manually, which may be sub‑optimal; the transformation step can become costly for very long series; and the method has only been tested on standard benchmark data, leaving its performance on highly irregular or domain‑specific series (e.g., streaming IoT data, clinical EEG) to be explored.

Future work outlined by the authors includes automatic data‑driven optimisation of (r, l, k), pruning of redundant shapelets to reduce computational load, visualisation tools for the extracted shapelet vectors, and application to specialised domains such as multi‑channel EEG monitoring and IoT anomaly detection. In summary, ALT offers a lightweight, transparent, and effective feature representation for time‑series classification, bridging the gap between traditional feature‑engineered methods and heavyweight deep‑learning models.