SCAGOT: Semi-Supervised Disentangling Context and Activity Features Without Target Data for Sensor-Based HAR

Classic deep learning methods for human activity recognition (HAR) from wearable sensors struggle with cross-person and cross-position challenges due to nonidentical data distributions caused by context variations (e.g., user, sensor placement). Existing solutions show promise but usually require extensive labeled data from source and target contexts, which is often unavailable in real-world scenarios. To address these limitations, we introduce semi-supervised context agnostic representation learning without target (SCAGOT), a novel semi-supervised approach that learns context-agnostic activity representations without relying on target context data. SCAGOT uses a dual-stream architecture with adversarial disentanglement and a contrastive clustering mechanism. This effectively separates context features from context-agnostic activity features, maximizing intraclass compactness and interclass separability in the activity representation space. In addition, a new inverse cross-entropy loss further refines the representations by removing residual context information. Extensive evaluations on four benchmark datasets demonstrate that SCAGOT outperforms state-of-the-art methods in cross-person and cross-position HAR, offering a promising solution for robust real-world activity recognition.