DiversifyAnalysis

Replication of the experiments from the ICLR 2023 paper: "Out-of-distribution Representation Learning for Time Series Classification" (Lu et al., 2023)

This project was completed as part of the AI and Cybersecurity course at the University of Luxembourg.

What This Project Is About

Time series classification models often fail when test data comes from a different distribution than training data — a common problem in real-world deployments. This paper proposes a representation learning approach (Diversify) that improves model robustness against out-of-distribution (OOD) data.

This repo reproduces the core experiments from the paper using the EMG (Electromyography) dataset to validate the authors' findings on OOD generalisation for time series classification.

What I Did

• Reproduced the preprocessing, training, and evaluation pipeline from the original paper
• Ran experiments on the EMG dataset using the Diversify algorithm
• Validated the OOD representation learning approach and analysed model accuracy across test groups

Key Concepts

• Out-of-Distribution (OOD) Learning — training models to generalise beyond the training distribution
• Time Series Classification — classifying sequential data (e.g. sensor readings, network traffic, physiological signals)
• Representation Learning — learning meaningful feature representations that improve generalisation
• Diversify Algorithm — latent domain-based approach to improve OOD robustness

How to Run

1. Prerequisites

pip install -r requirements.txt

2. Preprocess Dataset

Download and prepare the EMG dataset:

python prepare_dataset.py --data_dir ./data/ --dataset EMG

If executed successfully, the dataset will be saved to ./data/emg

3. Train

python train.py --data_dir ./data/ --task cross_people --test_envs 0 --dataset emg --algorithm diversify --latent_domain_num 10 --alpha1 1.0 --alpha 1.0 --lam 0.0 --local_epoch 3 --max_epoch 5 --lr 0.01 --output ./data/train_output/act/cross_people-emg-Diversify-0-10-1-1-0-3-50-0.01 --output_model ./data/model_output

Key parameters:

• --data_dir — path to your dataset directory
• --dataset — dataset name (e.g. emg)
• --test_envs — specify a different test group if applicable
• --output — training log will be saved here
• --output_model — generated models will be saved here
• Run --help to see the full parameter list

4. Evaluate

For each training round, the output value target acc is the accuracy value for the model.

Technologies Used

• Python
• PyTorch
• NumPy / Pandas
• Machine Learning / OOD Representation Learning
• Time Series Classification

Course Context

This replication was completed as part of the AI and Cybersecurity course at the University of Luxembourg, exploring the intersection of machine learning robustness and security-relevant applications such as anomaly detection and out-of-distribution generalisation in sensor and network data.

Reference

@inproceedings{lu2022out,
  title={Out-of-distribution Representation Learning for Time Series Classification},
  author={Lu, Wang and Wang, Jindong and Sun, Xinwei and Chen, Yiqiang and Xie, Xing},
  booktitle={International Conference on Learning Representations},
  year={2023}
}