🤖 machine learning

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

Original authors: Balázs P. Halmos, Balázs Hajós, Vince Á. Molnár, Marcell T. Kurbucz, Antal Jakovác

Published 2026-02-06

📖 5 min read🧠 Deep dive

Original authors: Balázs P. Halmos, Balázs Hajós, Vince Á. Molnár, Marcell T. Kurbucz, Antal Jakovác

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a long, messy recording of a heartbeat, a stock market ticker, or a sensor on a robot. This recording is a "time series"—a string of numbers changing over time. Your goal is to figure out what kind of event this recording represents (e.g., "Is the robot walking or running?" or "Is this stock going up or down?").

This is the problem of Time Series Classification (TSC).

The paper introduces a new Python tool called ALT (Adaptive Law-Based Transformation). Think of ALT as a smart "translator" that turns these messy, confusing recordings into a neat, organized map where different events are clearly separated from one another.

Here is how it works, using simple analogies:

1. The Problem: The "One-Size-Fits-All" Trap

Previous methods tried to analyze these recordings using fixed rules. Imagine trying to find a specific pattern in a song by only listening to exactly 5 seconds at a time, no matter what.

If the pattern you are looking for is a short drum beat, a 5-second window works.
If the pattern is a slow, swelling melody, a 5-second window cuts it off and misses the point.

The old method (called LLT) was like that fixed window. It was good, but it couldn't adapt if the patterns in the data changed size or speed.

2. The Solution: The "Adaptive Camera"

The new ALT package is like a camera with an auto-zoom lens. Instead of forcing the data into a fixed size, it looks at the data through windows of different lengths.

It can zoom in to catch fast, tiny details.
It can zoom out to catch slow, big trends.

By looking at the data through these "variable-length windows," ALT finds the hidden "laws" or rules that govern the data, regardless of how fast or slow they happen.

3. The Magic Trick: Turning Chaos into Clusters

Once ALT has looked at the data through these smart windows, it performs a mathematical transformation.

Before ALT: Imagine a bag of mixed-up marbles (red, blue, and green) all jumbled together in a dark room. It's hard to tell them apart.
After ALT: ALT acts like a magnet that sorts the marbles. It rearranges them so that all the red marbles form a tight, neat pile on the left, the blue ones form a pile in the middle, and the green ones on the right.

In technical terms, it transforms the raw data into a "linearly separable feature space." In plain English: it rearranges the data so that a simple straight line can easily draw a boundary between the different groups.

4. Why This Matters (The "Why Should I Care?")

The paper highlights three main benefits of this tool:

It's Fast and Light: Unlike massive Deep Learning systems (which are like super-computers that need huge amounts of electricity and data to learn), ALT is lightweight. It's like using a bicycle instead of a rocket ship to get to the store. It gets the job done quickly without needing a massive computer.
It's Transparent (You Can See How It Works): Deep learning is often a "black box"—you put data in, and it gives an answer, but you don't know why. ALT is different. Because it sorts the data into clear piles based on specific rules, humans can actually look at the result and understand why the computer made that decision.
It Works on Real-World Mess: Real data is noisy (like a radio with static). ALT is built to ignore the static and focus on the signal, making it very reliable for physics experiments, sensors, and other scientific data.

5. What the Paper Actually Says It Does

The authors tested this tool on standard datasets (like the "GunPoint" dataset, which records the motion of a hand, and "BasicMotions," which records walking vs. running).

The Result: When they used ALT to sort the data, even very simple, old-school math (like drawing a straight line) could separate the groups with extremely high accuracy (sometimes 100% or 94%).
The Claim: They claim this tool is better than many complex, expensive methods because it is faster, easier to understand, and just as accurate.

What It Does Not Claim (Yet)

The paper is careful to stick to what it has proven:

It does not claim to be a medical diagnostic tool for doctors to use on patients right now (though it mentions it could be useful for ECG analysis in the future).
It does not claim to predict the stock market for making money.
It does not claim to be perfect for every possible dataset, but it works very well for the physics and sensor data they tested.

In summary: ALT is a new, smart tool that takes messy time-based data, zooms in and out to find the right patterns, and rearranges that data into neat, separate piles so that simple math can easily tell them apart. It's fast, clear, and effective.

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

Problem Statement
Time series classification (TSC) remains a fundamental challenge in computational physics and related domains, where data streams from simulations, experiments, and sensor networks often exhibit variable temporal patterns, noise, and complex dependencies. Existing solutions face distinct trade-offs: feature-based methods (e.g., statistical descriptors, spectral transforms) often struggle with multi-scale patterns; distance-based methods (e.g., Dynamic Time Warping) suffer from high computational complexity and lack interpretability; and deep learning approaches, while accurate, require large labeled datasets, significant computational resources, and offer limited transparency. There is a specific need for a robust, scalable, and interpretable method tailored to physics-related time series that avoids the computational overhead of deep learning while maintaining high accuracy.

Methodology: Adaptive Law-Based Transformation (ALT)
The paper introduces ALT, an open-source Python package implementing the Adaptive Law-Based Transformation algorithm. This method extends the authors' previous Linear Law-Based Transformation (LLT) by introducing variable-length shifted time windows to better capture patterns across multiple temporal scales.

The core algorithm proceeds through the following steps:

Data Partitioning & Sequence Extraction: Instances are split into learning, training, and test sets. For specific parameters (window size $r$ , embedding dimension $l$ , and shift step $k$ ), sequences are extracted using time-shifted windows and downsampled to $2l-1$ points.
Shapelet Vector Computation: The algorithm performs $l$ -dimensional time-delay embedding to construct a symmetric matrix. Unlike Principal Component Analysis (PCA), which seeks maximum variance, ALT computes the spectral decomposition to identify the eigenvector corresponding to the smallest absolute eigenvalue. This vector (the "shapelet") represents a low-variability dimension, effectively capturing the governing laws of the time series.
Transformation: The training data is transformed into a feature space where rows of the data matrix are multiplied by shapelet matrices derived from specific classes. The goal is to map instances of a specific class toward near-null vectors in the transformed space, thereby creating linear separability.
Feature Extraction & Classification: Statistical methods (e.g., mean, percentiles, moments) are applied to the transformed matrices to extract features. These features are then fed into traditional classifiers (e.g., K-Nearest Neighbors, SVM) for final classification.

Key Contributions

Software Implementation: The authors provide the ALT Python package, designed for scalability, interpretability, and ease of use. It supports both univariate and multivariate time series and can utilize GPU acceleration.
Adaptive Windowing: By replacing fixed-length windows with variable-length shifted windows, the method overcomes the limitations of its predecessor (LLT), allowing it to handle datasets with varying temporal structures more effectively.
Interpretability: The method generates interpretable features based on "governing laws" extracted from the data, offering a transparent alternative to "black-box" deep learning models.
Efficiency: The package achieves state-of-the-art performance with minimal computational overhead, making it suitable for environments where resources are constrained or explainability is critical.

Results and Evaluation
The paper evaluates ALT using benchmark datasets from the UCR Time Series Classification Archive:

Linear Separability: Visualizations of the transformed feature space (e.g., on the BasicMotions and GunPoint datasets) demonstrate that instances of the same class form distinct, linearly separable clusters.
Classification Accuracy: On the GunPoint dataset, a simple linear separator applied to the transformed features achieved 94.0% test accuracy. On the BasicMotions dataset, a Fine KNN classifier achieved 100% validation and test accuracy using only two extracted features.
Parameter Sensitivity: Experiments show that the window size ( $r$ ) and embedding dimension ( $l$ ) significantly impact accuracy, while the shift step ( $k$ ) offers a trade-off: increasing $k$ reduces runtime but can decrease accuracy, though a value of $k=2$ was found to reduce runtime by one-third with only a slight accuracy drop in some cases.

Significance and Claims
The paper positions ALT as a practical, lightweight solution for TSC in scientific domains. Its primary significance lies in its ability to map complex time series data into a linearly separable feature space using only a few hyperparameters, thereby enabling the use of simple, fast, and interpretable classifiers.

The authors claim that ALT:

Achieves state-of-the-art accuracy on benchmark TSC datasets.
Offers superior speed and transparency compared to existing deep learning and distance-based methods.
Efficiently captures complex patterns across diverse temporal scales.
Is particularly well-suited for physics-related applications, biomedical signal processing (e.g., ECG), and financial pattern identification where transparency is crucial.

The paper concludes by noting that while the current implementation requires manual hyperparameter tuning, future work will focus on automating this process and implementing shapelet pruning to further enhance scalability for large datasets.