HyperMIL: Hypergraph-based channel reasoning for Multiple Instance Learning on Multivariate Time Series

Livia Del Gaudio¹  Vittorio Cuculo¹  Rita Cucchiara¹ 

¹University of Modena and Reggio Emilia  

ICPR 2026

HyperMIL is a Multiple Instance Learning framework for multivariate time series classification that explicitly models channel interactions through learnable hypergraphs. Higher-order relationships among variables are captured through prototype-driven hyperedges and further integrated within a temporal MIL architecture which ensures inherent interpretability.

---

📌 Overview

HyperMIL is a modular MIL pipeline for multivariate time-series classification that combines:

1. Intra-sample channel reasoning with a channel-level hypergraph encoder.
2. Temporal aggregation with a transformer-like MIL head (Nyström attention).

This design enables multiple levels of interpretability by design: prototype-channel affinities reveal latent variable groupings, hypergraph structures expose higher-order dependencies, and temporal attention highlights discriminative intervals driving the final prediction. Overall, HyperMIL introduces a principled and extensible approach to structured reasoning in multivariate time series, unifying hypergraph learning and MIL under a single framework.

✨ Highlights

• HyperMIL is the first MIL framework that leverages hypergraph reasoning to capture high-order channel dependencies in multivariate time series.
• Data-driven hypergraph construction using latent prototypes enables the discovery of complex channel groupings directly from raw data (no domain-specific prior knowledge needed).
• Evaluation on multiple benchmark MTS datasets shows that HyperMIL outperforms state-of-the-art MIL and temporal deep learning baselines in both classification accuracy and model interpretability.

🧐 Project structure

HyperMIL/
├── main.py                  # CLI entrypoint for training on aeon datasets
├── train_model.py           # training/evaluation loops and metrics
├── lookhead.py              # Lookahead optimizer wrapper
├── utils.py                 # IO/logging/reproducibility helpers
├── requirements.txt
└── models/
    ├── hypermil.py
    ├── channels_hypergraph.py
    ├── timemil.py
    ├── nystrom_attention.py
    └── common.py

⚙️ Installation

pip install -r requirements.txt

PyTorch should be installed separately to match your CUDA version — see the official instructions. The code requires torch>=1.13.

💡 Quick start

Data is loaded directly from the aeon time-series library — no manual download required.

python main.py --dataset PenDigits --num_epochs 50 --batchsize 64 --encoding wavelet --pooling cls

Main CLI arguments

• --dataset: aeon dataset name (train/test split fetched automatically)
• --num_epochs, --batchsize, --lr, --optimizer, --scheduler
• --encoding: wavelet | sinusoidal | none
• --pooling: cls | mean | max | attention | conjunct
• --k_prototypes, --num_convs, --tau, --intra_embed, --embed
• --seed: random seed for reproducibility (default: 0)
• --save_dir: checkpoint output directory (default: ./savemodel/)

Current scope

This release version targets aeon-based equal-length classification workflows while preserving support paths for variable-length collation in the training utilities.

🙏 Acknowledgements

This work builds on TimeMIL by Chen et al. — we thank the authors for releasing their code, which we used as the foundation for the temporal MIL head in this repository.

📚 Citation

If you use this code, please cite:

@inproceedings{del2026hypermil,
  title={HyperMIL: Hypergraph-based channel reasoning for Multiple Instance Learning on Multivariate Time Series},
  author={Del Gaudio, Livia and Cuculo, Vittorio and Cucchiara, Rita},
  booktitle={Proceedings of the 28th International Conference on Pattern Recognition},
  year={2026}
}

---

License

This dataset is released under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license.

You are free to share and adapt the material for non-commercial research purposes only, provided appropriate credit is given.