train

HELIX: Hybrid Encoding with Learnable Identity and Cross-dimensional Synthesis for Time Series Imputation

The official experiment code for the paper HELIX: Hybrid Encoding with Learnable Identity and Cross-dimensional Synthesis for Time Series Imputation, which has been accepted as a Spotlight paper at the International Conference on Machine Learning (ICML) 2026.

Fengming Zhang, Wenjie Du, Huan Zhang, Ke Yu, Shen Qu. HELIX: Hybrid Encoding with Learnable Identity and Cross-dimensional Synthesis for Time Series Imputation. In Proceedings of the International Conference on Machine Learning (ICML), 2026.

🤗 If you find HELIX helpful to your work, please cite it and ⭐️star this repository to help others discover it. Thank you!

❖ Motivation and Highlights

⦿ Motivation: Existing time series imputation methods often treat temporal and feature dimensions in isolation, limiting their ability to capture the complex inter-dependencies in multivariate incomplete data. HELIX is proposed to bridge this gap by jointly encoding temporal position, feature identity, and cross-dimensional interactions within a unified framework.

⦿ Highlights:

• Learnable Feature Identity (FeatID): A learnable embedding module that assigns each feature a unique identity representation, enabling the model to distinguish and specialize across heterogeneous variables.
• Hybrid Encoding: Combines sinusoidal positional encoding with learnable identity embeddings, capturing both absolute temporal positions and feature-specific characteristics simultaneously.
• Cross-dimensional Synthesis: A fusion mechanism that synthesizes information across the temporal and feature dimensions, allowing bidirectional information flow for more accurate imputation.
• State-of-the-art Performance: HELIX achieves competitive or superior results on multiple benchmark datasets for time series imputation.

❖ Model Availability

pip install pypots  # v1.5+ when released; for now install from main branch

from pypots.imputation import HELIX

# initialize the model
helix = HELIX(
    n_steps=48,           # number of time steps
    n_features=37,        # number of features
    # ... (refer to PyPOTS documentation for full parameter list)
    epochs=100,
)

# train
helix.fit(train_set, val_set)

# impute
imputation = helix.impute(test_set)

Additional variants available at milaogou/PyPOTS@add-helix-model:

• Ablation variants (w/o FeatID, w/o Fusion, w/o Hybrid, w/o Sinusoidal PE)
• Gated fusion variant with learnable gate mechanism

❖ Reproducing Experiments

All experiments were conducted on a Linux aarch64 HPC cluster with NVIDIA A100 GPUs. Tasks were submitted in parallel via SLURM. The shell scripts (.sh) in this repo reflect this cluster setup — please adapt them to your own environment as needed.

Training logs are omitted due to their large size. Each script is self-explanatory; the overall workflow is described below.

Step-by-step Workflow

1. Prepare datasets: Follow the data preprocessing instructions in benchmark_code/.
2. Configure experiments: Adjust hyperparameters and paths in the config/script files as needed.
3. Run training: Submit the .sh scripts via SLURM or run them directly.
4. Evaluate: Results will be saved automatically. Compare with baselines using the provided evaluation scripts.

See benchmark_code/ for full details.

❖ Repository Structure

HELIX/
├── benchmark_code/                    # Full experiment code and evaluation
│   ├── train_model.py                 # Main training script (5 seeds)
│   ├── train_model_tuning.py          # HPO training script
│   ├── in_sample_exp.py               # Batch SLURM job submission
│   ├── global_config.py               # Global configuration
│   ├── utils.py                       # Utilities
│   ├── data/                          # Dataset generation scripts
│   ├── hpo_results/                   # Tuned hyperparameters per dataset
│   ├── PyPOTS_tuning_configs/         # HPO search spaces and scripts
│   ├── reproduce_imputation/          # Result collection and analysis
│   ├── generate_latex_tables.py       # LaTeX table generation
│   ├── feature_embedding_analysis.py  # Embedding analysis (Fig. 2)
│   ├── extract_attention.py           # Attention patterns (Fig. 3, 4)
│   ├── imputation_visualization.py    # Imputation visualization (Fig. 5)
│   ├── physionet_embedding_analysis.py# Clinical grouping analysis (Fig. 6)
│   └── README.md                      # Detailed experiment instructions
├── LICENSE                            # MIT License
└── README.md                          # This file

❖ Citing HELIX

If HELIX is helpful to your research, please cite our paper as below and ⭐️star this repository. 🤗 Thank you!

@inproceedings{zhang2026helix,
    title     = {{HELIX}: Hybrid Encoding with Learnable Identity and Cross-dimensional Synthesis for Time Series Imputation},
    author    = {Fengming Zhang and Wenjie Du and Huan Zhang and Ke Yu and Shen Qu},
    booktitle = {Forty-third International Conference on Machine Learning},
    year      = {2026},
    url       = {https://arxiv.org/abs/2605.02278}
}

or

Fengming Zhang, Wenjie Du, Huan Zhang, Ke Yu, and Shen Qu. HELIX: Hybrid Encoding with Learnable Identity and Cross-dimensional Synthesis for Time Series Imputation. In Proceedings of the Forty-third International Conference on Machine Learning (ICML), 2026.

🔥 If you use PyPOTS in your research, please also cite:

@article{du2023pypots,
    title   = {{PyPOTS: a Python toolbox for data mining on Partially-Observed Time Series}},
    author  = {Wenjie Du},
    journal = {arXiv preprint arXiv:2305.18811},
    year    = {2023}
}

❖ Acknowledgments

Thanks to all co-authors for their valuable contributions. Thanks to the PyPOTS community for providing the ecosystem and infrastructure. Thanks to all reviewers for helping improve the quality of this paper.

✨ Stars/forks/issues/PRs are all welcome!

❖ License

This project is released under the MIT License.

❖ Contact

If you have any questions or are interested in collaboration, please feel free to open an issue or contact us. 😊