This repository contains the implementation for the paper "Benchmarking Time Series Feature Extraction for Algorithm Selection" by Moisés Santos, Vitor Cerqueira, and Carlos Soares. The paper evaluates the effectiveness of different meta-feature extraction methods for time series forecasting algorithm selection.
The study investigates how meta-features derived from time series influence the selection of forecasting algorithms. Four meta-feature extraction frameworks were compared: TSFRESH, TSFEATURES, TSFEL, and Catch22. The experiments utilized data from the M4 Competition, applying a range of forecasting models, including neural network-based methods, and analyzed the meta-level performance of these feature extractors.
-
base_models.py
Implements time series forecasting using the base learners, such as Multi-Layer Perceptron (MLP), DeepAR, NHITS, and TCN. -
base_performance.py
Extracts the forecasting performance metrics (e.g., MAE) for each algorithm and dataset. -
mfe.py
Handles the extraction of meta-features from the time series data using the selected frameworks. -
metadata.py
Combines the extracted meta-features with the target values (based on forecasting performance) to generate the metadata required for meta-learning. -
analysis.ipynb
A Jupyter notebook containing all meta-analyses, visualizations, and statistical evaluations performed during the study.