aeon Toolkit

aeon hosts various short to medium duration projects involving development, research, and community interaction. These projects are suitable for internships, academic projects (undergraduate/postgraduate), Google Summer of Code (GSoC), or personal side projects.

Requirements:

• Proficiency in Python and Git/GitHub.
• Most projects require knowledge of machine learning and time series analysis.

How to participate:

• Join the Discord to discuss ideas with the community.
• Propose your own project ideas, but discuss them with the maintainers first.
• Funding opportunities are advertised as they become available.

aeon is a scikit-learn compatible Python library designed for time series machine learning. It provides a comprehensive suite of tools for various time series tasks, including:

• Classification and Regression
• Clustering and Similarity Search
• Forecasting
• Anomaly Detection
• Segmentation
• Distances and Transformations
• Benchmarking

The library is maintained by the researchers who developed many of the underlying algorithms, ensuring access to state-of-the-art models, including deep learning approaches.

aeon is a scikit-learn compatible toolkit designed for time series machine learning tasks. It provides a broad library of algorithms, including state-of-the-art methods, implemented efficiently (e.g., using numba). Because it follows the scikit-learn API, it integrates easily with the broader Python machine learning ecosystem.

Supported tasks include:

• Classification
• Regression (extrinsic)
• Clustering
• Anomaly detection
• Segmentation
• Similarity search
• Forecasting
• Transformations
• Distances

Time series segmentation partitions a series into regions that are dissimilar to their neighboring regions. You can use the aeon.segmentation module to access various segmentation algorithms.

Available segmenter implementations include:

• BinSegmenter
• ClaSPSegmenter
• FLUSSSegmenter
• InformationGainSegmenter
• GreedyGaussianSegmenter
• EAggloSegmenter
• HMMSegmenter
• HidalgoSegmenter
• RandomSegmenter

The forecasting module received a major update in v1.3.0, introducing several new forecasters including:

• ARIMA/AutoARIMA
• SETAR-Tree/Forest
• Theta
• TAR/SETAR
• Deep Learners (via BaseDeepForecaster and specific implementations like TCN)

Additionally, regression forecasters now support exogenous variables, and a Time-Varying Parameter (TVP) regression forecaster using a Kalman filter has been implemented.

As of version 1.4.0, several new native implementations of estimators are available for time series analysis:

• Anomaly Detection: ROCKAD (whole-series anomaly detector).
• Clustering: K-Shape clusterer.
• Classification: ElasticEnsemble now supports TS-QUAD configuration.
• Forecasting: ETS, AutoETS, and DeepARF forecasters.
• Transformations: STL, MSTL, and LOWESS transformers.

The aeon toolkit provides a consistent API across ten major time series task areas:

• Classification: Predict labels for time series.
• Regression: Predict continuous values from time series.
• Clustering: Group similar series without labels.
• Forecasting: Predict future values.
• Anomaly detection: Find unusual points or subsequences.
• Segmentation: Split a series into homogeneous regions.
• Similarity search: Find similar subsequences or whole series in collections.
• Transformations: Feature extraction and preprocessing.
• Distances & kernels: Time series similarity measures.
• Benchmarking: Reproducible experimental evaluation.

aeon is a toolkit for time series machine learning that supports several distinct learning tasks:

• Classification: Predicting a discrete label for a collection of time series.
• Regression: Predicting a continuous value for a collection of time series.
• Clustering: Grouping unlabeled time series into clusters.
• Similarity search: Finding nearest neighbors among subsequences or whole series.
• Anomaly detection: Identifying values or regions in a single time series that deviate from the norm.
• Forecasting: Predicting future values of a single time series.
• Segmentation: Splitting a single time series into dissimilar regions.

Core modules used across these tasks include Transformations (changing representations), Distances (measuring dissimilarity), and Networks (deep learning models).

As of version 0.6.0, aeon introduced several new capabilities and structural changes:

• Classification: New algorithms available in RED CoMETS and LITETime (Deep Learning classifiers).
• Clustering: A new clustering subpackage for deep learning clustering has been added.
• Similarity Search: An experimental similarity search subpackage is now available.
• Segmentation & Anomaly Detection: The previous annotation module has been split into two distinct modules: segmentation and anomaly detection.
• Distances: Minkowski distance has been added to the distances module.