CrossHAR: Generalizing Cross-dataset Human Activity Recognition via Hierarchical Self-Supervised Pretraining The increasing availability of low-cost wearable devices and smartphones has significantly advanced the field of sensor-based human activity recognition (HAR), attracting considerable research interest. One of the major challenges in HAR is the domain shift problem in cross-dataset activity recognition, which occurs due to variations in users, device types, and sensor placements between the source dataset and the target dataset. Although domain adaptation methods have shown promise, they typically require access to the target dataset during the training process, which might not be practical in some scenarios. To address these issues, we introduce CrossHAR, a new HAR model designed to improve model performance on unseen target datasets. CrossHAR involves three main steps: (i) CrossHAR explores the sensor data generation principle to diversify the data distribution and augment the raw sensor data. (ii) CrossHAR then employs a hierarchical self-supervised pretraining approach with the augmented data to develop a generalizable representation. (iii) Finally, CrossHAR fine-tunes the pretrained model with a small set of labeled data in the source dataset, enhancing its performance in cross-dataset HAR. Our extensive experiments across multiple real-world HAR datasets demonstrate that CrossHAR outperforms current state-of-the-art methods by 10.83% in accuracy, demonstrating its effectiveness in generalizing to unseen target datasets. CrossHAR is a novel model that leverages unlabeled sensor data and the sensor data generation principles to improve cross-dataset HAR. It consists of three components: (i) Physically-informed Sensor Data Augmentation, which improves the generalization capability of models trained on the source dataset by simulating device movement and generating diverse data. (ii) Hierarchical Sensor Data Pretraining, which uses masked sensor modeling and contrastive regularization to capture local and global patterns in IMU data. (iii) Fine-Tune Activity Recognition, which fine-tunes the pretrained model with a small set of labeled data from the source dataset to recognize activities on the target dataset. The results show that CrossHAR outperforms state-of-the-art methods in cross-dataset HAR, achieving a 10.83% improvement in accuracy. Additionally, CrossHAR is deployed on a smartphone, demonstrating its high efficiency for real-world usage.CrossHAR: Generalizing Cross-dataset Human Activity Recognition via Hierarchical Self-Supervised Pretraining The increasing availability of low-cost wearable devices and smartphones has significantly advanced the field of sensor-based human activity recognition (HAR), attracting considerable research interest. One of the major challenges in HAR is the domain shift problem in cross-dataset activity recognition, which occurs due to variations in users, device types, and sensor placements between the source dataset and the target dataset. Although domain adaptation methods have shown promise, they typically require access to the target dataset during the training process, which might not be practical in some scenarios. To address these issues, we introduce CrossHAR, a new HAR model designed to improve model performance on unseen target datasets. CrossHAR involves three main steps: (i) CrossHAR explores the sensor data generation principle to diversify the data distribution and augment the raw sensor data. (ii) CrossHAR then employs a hierarchical self-supervised pretraining approach with the augmented data to develop a generalizable representation. (iii) Finally, CrossHAR fine-tunes the pretrained model with a small set of labeled data in the source dataset, enhancing its performance in cross-dataset HAR. Our extensive experiments across multiple real-world HAR datasets demonstrate that CrossHAR outperforms current state-of-the-art methods by 10.83% in accuracy, demonstrating its effectiveness in generalizing to unseen target datasets. CrossHAR is a novel model that leverages unlabeled sensor data and the sensor data generation principles to improve cross-dataset HAR. It consists of three components: (i) Physically-informed Sensor Data Augmentation, which improves the generalization capability of models trained on the source dataset by simulating device movement and generating diverse data. (ii) Hierarchical Sensor Data Pretraining, which uses masked sensor modeling and contrastive regularization to capture local and global patterns in IMU data. (iii) Fine-Tune Activity Recognition, which fine-tunes the pretrained model with a small set of labeled data from the source dataset to recognize activities on the target dataset. The results show that CrossHAR outperforms state-of-the-art methods in cross-dataset HAR, achieving a 10.83% improvement in accuracy. Additionally, CrossHAR is deployed on a smartphone, demonstrating its high efficiency for real-world usage.