| | --- |
| | task_categories: |
| | - other |
| | tags: |
| | - human-activity-recognition |
| | - sensor-data |
| | - time-series |
| | - out-of-distribution |
| | --- |
| | |
| | # HAROOD: A Benchmark for Out-of-distribution Generalization in Sensor-based Human Activity Recognition |
| |
|
| | [**Paper**](https://huggingface.co/papers/2512.10807) | [**GitHub Repository**](https://github.com/AIFrontierLab/HAROOD) |
| |
|
| | HAROOD is a modular and reproducible benchmark framework for studying generalization in sensor-based human activity recognition (HAR). It unifies preprocessing pipelines, standardizes four realistic OOD scenarios (cross-person, cross-position, cross-dataset, and cross-time), and implements 16 representative algorithms across CNN and Transformer architectures. |
| |
|
| | ## Key Features |
| |
|
| | - **6 public HAR datasets** unified under a single framework. |
| | - **5 realistic OOD scenarios**: cross-person, cross-position, cross-dataset, cross-time, and cross-device. |
| | - **16 generalization algorithms** spanning Data Manipulation, Representation Learning, and Learning Strategies. |
| | - **Backbone support**: Includes both CNN and Transformer-based architectures. |
| | - **Standardized splits**: Provides train/val/test model selection protocols. |
| |
|
| | ## Usage |
| |
|
| | The benchmark is designed to be modular. Below are examples of how to run experiments using the official implementation: |
| |
|
| | ### Run with a YAML config |
| |
|
| | ```python |
| | from core import train |
| | results = train(config='./config/experiment.yaml') |
| | ``` |
| |
|
| | ### Run with a Python dict |
| |
|
| | ```python |
| | from core import train |
| | config_dict = { |
| | 'algorithm': 'CORAL', |
| | 'batch_size': 32, |
| | } |
| | results = train(config=config_dict) |
| | ``` |
| |
|
| | ### Override parameters |
| |
|
| | ```python |
| | from core import train |
| | results = train( |
| | config='./config/experiment.yaml', |
| | lr=2e-3, |
| | max_epoch=200, |
| | ) |
| | ``` |
| |
|
| | ## Supported Algorithms |
| |
|
| | The benchmark implements 16 algorithms across three main categories: |
| |
|
| | - **Data Manipulation**: Mixup, DDLearn. |
| | - **Representation Learning**: ERM, DANN, CORAL, MMD, VREx, LAG. |
| | - **Learning Strategy**: MLDG, RSC, GroupDRO, ANDMask, Fish, Fishr, URM, ERM++. |
| |
|
| | ## Citation |
| |
|
| | If you use HAROOD in your research, please cite the following paper: |
| |
|
| | ```bibtex |
| | @inproceedings{lu2026harood, |
| | title={HAROOD: A Benchmark for Out-of-distribution Generalization in Sensor-based Human Activity Recognition}, |
| | author={Lu, Wang and Zhu, Yao and Wang, Jindong}, |
| | booktitle={The 32rd ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD)}, |
| | year={2026} |
| | } |
| | ``` |