Semi-supervised Human Activity Recognition with individual difference alignment

Human Activity Recognition (HAR) is a crucial application for wearable devices, providing essential guidance for various intelligent scenarios. Currently, HAR predominantly relies on supervised models powered by labeled data. However, due to the constraints imposed by annotation costs, the available labeled data often do not represent individuals with diverse activity habits across numerous scenarios, thereby frequently resulting in overfitting issues. Consequently, this paper focuses on semi-supervised method that extracts additional information from unlabeled data. Traditional semi-supervised training paradigms primarily focus on identifying sample-level discriminative features, yet they often neglect the individual differences inherent inhuman activity data, which results in limited improvements in generalization. To address this issue, we propose a semi-supervised training task named individual difference alignment, aimed at making features across different individuals more robust. Specifically, our designed Difference Alignment Contrastive Loss (DAC Loss) aligns features of similar individuals and reduces intra-class variances, thereby enhancing the model's generalization capabilities across diverse individuals. Moreover, we introduce a sampling strategy tailored to the individual difference alignment task to prevent the model from learning incorrect features. Extensive experiments demonstrate that our method surpasses other weakly supervised methods, achieving an average performance improvement in F1-Score of 6.86, 6.44, and 15.56, respectively, on the UCI-HAR, RealWorld, and MotionSense datasets under the condition of scarce labeled data compared to the supervised baseline.