aeon_skill

Overview

This skill provides a scikit-learn compatible toolkit for time series machine learning tasks including classification, regression, clustering, forecasting, anomaly detection, segmentation, and similarity search. It targets both univariate and multivariate time series and exposes feature extraction, distance metrics, and neural architectures with familiar estimator APIs. Use it to apply specialized temporal algorithms beyond standard tabular ML for patterns that evolve over time. The skill is optimized for quick prototyping with ROCKET variants and for higher-accuracy workflows using deep learning or ensemble methods.

How this skill works

The skill exposes estimator classes and transformers that accept time-indexed arrays (shape: n_samples, n_channels, n_timepoints) and implement fit/predict/transform interfaces. It includes convolutional feature transformers (ROCKET), distance-based classifiers (DTW-based KNN), clustering (TimeSeriesKMeans), forecasters (ARIMA, TCN), anomaly detectors, segmentation tools, motif search, and statistical feature extractors (Catch22). You can combine transformers and estimators in sklearn pipelines, extract features for traditional ML, or train neural networks for end-to-end learning.

When to use it

• Classifying sequences or predicting labels from temporal observations
• Forecasting future values from time-indexed series
• Detecting anomalies, change points, or unusual subsequences
• Clustering or segmenting similar time series patterns
• Comparing series with temporal distance measures (DTW, ERP, SBD)
• Extracting temporal features for downstream ML models

Best practices

• Normalize (z-normalize) time series before modeling to stabilize distances and features
• Impute or handle missing values before training (SimpleImputer or domain-specific methods)
• Start with ROCKET/MiniRocket for fast, strong baselines before using deep nets
• Validate hyperparameters with held-out splits or cross-validation designed for temporal data
• Match algorithm choice to dataset size: distance-based for small sets, deep models for large labeled sets

Example use cases

• Classify sensor recordings into activity labels using RocketClassifier + Normalizer
• Forecast monthly sales with ARIMA or TCN forecasters and evaluate multi-step horizons
• Detect production anomalies by computing anomaly scores (STOMP) and visualizing thresholds
• Cluster multivariate patient time series with TimeSeriesKMeans and DTW for subgroup discovery
• Extract ROCKET or Catch22 features, then train a RandomForest for fast, interpretable pipelines