aeon

Aeon - Python Toolbox for Time Series Machine Learning

Skills Overview

Aeon is a scikit-learn-compatible Python toolbox for time series machine learning. It provides a comprehensive library of algorithms for classification, regression, clustering, forecasting, anomaly detection, segmentation, similarity search, and more, using a unified estimator API to simplify time series modeling.

Use Cases

1. Time Series Classification and Forecasting

When you need to identify class patterns in time series data or predict future values, Aeon offers a range of algorithm options from rapid prototyping to high-accuracy modeling. Whether working with sensor data classification, stock trend forecasting, or user behavior pattern recognition, you can use algorithms such as MiniRocket and HIVECOTEV2 to build models quickly.

2. Anomaly Detection and Pattern Discovery

In scenarios such as industrial equipment monitoring, financial fraud detection, and IT system monitoring, Aeon’s anomaly detection module can identify anomalous patterns and outliers in time series. Its similarity search functionality can also discover recurring motifs and rare discords (anomalous subsequences).

3. Time Series Feature Engineering and Clustering

When time series need to be converted into feature vectors for use with traditional machine learning, ROCKET or Catch22 can be used for feature extraction. For unlabeled time series data, clustering algorithms such as TimeSeriesKMeans can group sequences according to their similarity.

Core Features

1. Time Series Classification

Aeon provides a comprehensive range of classification algorithms, from lightweight methods to high-accuracy solutions. For rapid prototyping, MiniRocketClassifier offers a good balance between speed and performance. For maximum accuracy, HIVECOTEV2 or InceptionTimeClassifier can be used. When model interpretability is required, ShapeletTransformClassifier is a suitable choice. All classifiers support both univariate and multivariate time series.

2. Time Series Forecasting

Aeon supports multi-step time series forecasting, including baseline methods (NaiveForecaster), statistical models (ARIMA and AutoETS), and deep learning architectures (TCNForecaster). Note that the forecasting module is still experimental in aeon 1.x, and its API may change between minor versions.

3. Distance Measures and Similarity Analysis

Aeon provides distance measures designed specifically for time series, including elastic measures (DTW, WDTW, MSM, etc.) and lock-step measures (Euclidean and Manhattan distance). These distance measures can be used with KNN classifiers, as well as for similarity search and clustering analysis. DTW and its variants are particularly suitable for sequences with time-alignment issues.

Frequently Asked Questions

What is the difference between aeon and sktime?

Aeon is an independent toolbox that evolved from the sktime project. The main differences are that aeon uses a 1.x versioning scheme and has restructured the APIs for forecasting and transformations. If your code is based on sktime or aeon 0.x, import paths may need to be updated—for example, the forecasting module now uses from aeon.forecasting instead of the old path. It is recommended to pin the aeon version in your project, such as aeon>=1.4,<2, to ensure stability.

Which algorithm should I choose for time series classification?

The choice depends on the data size, accuracy requirements, and available computing resources:

Rapid prototyping: MiniRocketClassifier — fast training and strong performance

Highest accuracy: HIVECOTEV2 (traditional methods), InceptionTimeClassifier (deep learning)

Small datasets: KNeighborsTimeSeriesClassifier with DTW distance

Need for interpretability: ShapeletTransformClassifier, Catch22Classifier

Feature engineering: First use RocketTransformer to extract features, then apply a traditional ML classifier

How should missing values in time series be handled?

Aeon provides SimpleImputer for handling missing values. It is recommended to preprocess the data before analysis:

from aeon.transformations.collection import SimpleImputer
imputer = SimpleImputer(strategy='mean')
X_train = imputer.fit_transform(X_train)

Most algorithms also benefit from Z-normalization (standardization), which can be performed using Normalizer. In terms of data format, collection data uses the shape (n_cases, n_channels, n_timepoints), while a single series uses the shape (n_channels, n_timepoints).

Does time series forecasting support multi-step forecasting?

Yes, multi-step forecasting is supported. Set the prediction_horizon parameter in the constructor, then call predict or iterative_forecast. Note that the forecasting module is experimental, so for production environments it is recommended to prioritize the stable classification, regression, clustering, distance, and transformation modules. Deep learning forecasters require the additional aeon[all_extras] dependencies to be installed.

Which algorithms are suitable for small time series datasets?

For small datasets, distance-based methods are recommended: use KNeighborsTimeSeriesClassifier with DTW or WDTW distance measures. Avoid deep learning methods such as InceptionTimeClassifier, as they require large amounts of data to train effectively. Another option is to use feature extraction methods such as Catch22 to convert time series into features, then classify them with traditional ML algorithms, which tend to perform more reliably on small datasets.