ALT: A Python Package for Lightweight Feature Representation in Time Series Classification

TL;DR

The paper tackles efficient, interpretable time series classification in contexts with multi-scale temporal structure. It introduces the adaptive law-based transformation (ALT), an extension of LLT that uses variable-length shifted windows to map time series into a linearly separable feature space via time-delay embedding and spectral shapelet extraction. The authors provide a Python package implementing ALT, demonstrate state-of-the-art performance with low overhead on benchmark datasets, and show that simple classifiers can achieve high accuracy on the transformed features. The work offers scalable, interpretable tooling with GPU support and outlines future directions including automated hyperparameter tuning, shapelet pruning, and broader applications such as EEG and IoT data.

Abstract

We introduce ALT, an open-source Python package created for efficient and accurate time series classification (TSC). The package implements the adaptive law-based transformation (ALT) algorithm, which transforms raw time series data into a linearly separable feature space using variable-length shifted time windows. This adaptive approach enhances its predecessor, the linear law-based transformation (LLT), by effectively capturing patterns of varying temporal scales. The software is implemented for scalability, interpretability, and ease of use, achieving state-of-the-art performance with minimal computational overhead. Extensive benchmarking on real-world datasets demonstrates the utility of ALT for diverse TSC tasks in physics and related domains.

Figures (4)

• Figure 1: Software architecture. The blocks correspond to the points in Section \ref{['sec:algorithm']}. Original source: kurbucz2025adaptive.
• Figure 2: The transformation makes the data form clusters. The BasicMotions dataset, $r=53,\, l=27,\, k=1$, 2nd sensor, meanAll as the feature extraction method. The instances corresponding to different classes are marked with different colors, and can be separated from each other easily, the train instances are marked with dots, and the test instances with triangles.
• Figure 3: The transformation tries to make the data linearly separable. The GunPoint dataset, $r=25,\, l=4,\, k=1$, 1st sensor, mean with 5$^\text{th}$ percentile as the feature extraction method. The instances corresponding to different classes are marked with different colors, the train instances are marked with dots, and the test instances with triangles. A linear separator (shown as a black line with its equation) is able to classify the data with 94.0% test accuracy. Note: some of the correctly classified points are not plotted for better visibility.
• Figure 4: The effect of changing the parameters $r, l$ and $k$ on the GunPoint database. The accuracies are calculated using linear discriminant on 2 features, similar to Figure \ref{['fig:GunPoint_plot']}. Parameter $r$ describes the physical scale in which the shapelets are found, subfigure (a). The embedding dimension $l$ describes the depth of the linear recursion, subfigure (b). Increasing $k$ will lower the accuracy and the transformation time, subfigure (c). It depends on the database which $r$ and $l$ values separate the instances well and which feature extraction method should be used.