multi_user_har_dataset

This dataset supports research in multi-user human activity recognition (HAR) using wearable sensors, with a focus on complex, composite, and interaction-driven activities. The dataset was developed for the study: "Hybrid Classification for Complex and Composite Activity Recognition in Multi-User Environments" It consists of multi-user sessions (2–4 users per session) with synchronized and asynchronous activities captured through simulated inertial measurement unit (IMU) signals from wrist, waist, and chest sensors. The dataset includes hierarchical annotations for atomic activities, composite activities, and inter-user interactions. To enable realistic evaluation, the dataset incorporates controlled variations such as sensor noise, synchronization drift (up to 500 ms), and sensor dropout (up to 10%). This allows benchmarking of robustness, temporal modeling, interaction recognition, and user attribution methods. The dataset is intended for:- Multi-user HAR benchmarking- Sensor fusion and multimodal learning- Graph neural network-based interaction modeling- Temporal sequence modeling and symbolic reasoning- Robustness analysis under real-world sensing conditions This is a dataset designed for controlled experimentation and reproducibility. For comprehensive evaluation, it is recommended to complement this dataset with real-world benchmarks such as PAMAP2. The dataset is provided in CSV format with session-level, segment-level, and sample-level annotations, along with predefined train/validation/test splits. Keywords: human activity recognition, wearable sensors, multi-user HAR, sensor fusion, graph neural networks, temporal modeling, composite activities, interaction recognition.