Time Series Classification Experiments Repository

This repository contains code and experiments related to time series classification using shapelets. The experiments focus on evaluating various machine learning models and quality measures on different time series datasets. Below, you will find an overview of the contents and how to navigate this repository.

Table of Contents

1. Introduction
2. Contents
3. Usage
4. Dataset Descriptions
5. Experiment Configurations
6. Results

Introduction

This repository focuses on shapelet-based time series classification. Running demo.ipynb is an easy and fast way to explore our project's capabilities. Additionally, you can experiment with our Ipol demo.

Contents

• Utils: The utils directory contains Python scripts with the main functions used in the experiments. It includes code for shapelet selection, shapelet transformation of datasets and computing quality measures. There is also a Notebook json_to_table.ipynb usefull to translate the .json results into latex tables.

• Datasets: The datasets/raw_datasets directory contains the unprocessed series datasets used in the experiments. The datasets/preprocessed_datasets directory contains the processed datasets stored in NumPy format files for easy loading.

• Jupyter Notebook: The repository contains a range of Jupyter Notebooks, each tailored for specific aspects of our time series analysis project. The demo.ipynb offers a practical demonstration on a synthetic dataset, giving users an insight into the code's functionality. For more in-depth analysis, main.ipynb focuses on testing the efficacy of conventional classification algorithms on shapelet-transformed datasets. To ensure the robustness of our methods, sanity_checks.ipynb is dedicated to experimenting with key functions and verifying their performance. The preprocessing of our datasets is documented in datasets/datasets_processing.ipynb. Lastly, shapelet_visualization.ipynb presents a selection of the most informative shapelets discovered across various datasets.

• Experiments Results : results directory stores the results of the experiments, including classification accuracy on each dataset and for each quality measure shapelet selection, and selected shapelets. The results are organized by dataset name and quality measure. These files are generated by main.ipynb Notebook.

Usage

To reproduce or extend the experiments, follow these steps:

1. Clone this repository to your local machine:

   git clone https://github.com/VictorBaillet/shapelets-time-series-classification-experiments.git
   cd shapelets-time-series-classification-experiments

2. Install the required dependencies by creating a virtual environment and using pip:

   python -m venv venv
   source venv/bin/activate  # On Windows, use `venv\Scripts\activate`
   pip install -r requirements.txt

3. Run the Jupyter notebooks to conduct experiments.

Dataset Descriptions

Here is a list of datasets used in the experiments:

• GunPointAgeSpan
• Synthetic
• ECG200
• ECG Five Days
• Two Lead ECG
• Mote Strain
• Sony Robot
• Beef

Refer to the dataset-specific documentation for more details. These dataset (except for the synthetic one) can be download here.

Experiment Configurations

The dataset_to_parameters dictionary in the code defines the experiment configurations, including shapelet length ranges and the number of shapelets or clusters to consider. You can customize these configurations to tailor the experiments to your specific needs.

Results

The experiment results, including classification accuracy and quality measure evaluations, are stored in JSON files within the experiments_results directory. You can analyze these results to gain insights into the performance of different models and quality measures on various datasets.