DeepSense++: Robust HAR with Missing Data

Human Activity Recognition (HAR) using wearable sensors is increasingly applied in healthcare, sports, and intelligent environments. Performance is however hindered in the majority of the cases by absent sensor values, class imbalance, and inter-subject variability. We present a robust HAR pipeline that utilizes Principal Component Analysis (PCA) for reducing dimensions and Generative Adversarial Networks (GANs) for realistic imputation of absent values and minority-class oversampling. This is integrated into an improved DeepSense architecture with convolutional and recurrent layers for spatial–temporal feature learning. Comparisons on the OPPORTUNITY dataset, in terms of K-Fold, Leave-One-Session-Out (LOSEO), and Leave-One-Subject-Out (LOSO) schemes, demonstrate improved accuracy (+3.7%) and F1 score (+2.9%) over baseline DeepSense. The results highlight the applicability of hybrid imputation-augmentation pipelines in bringing HAR to practical, noisy sensing scenarios.