ALOHA: Leveraging Additional Information to Learn Robust Representations for Human Activity Recognition

Human Activity Recognition using wearable sensors has applications in health monitoring, entertainment, and industrial settings. However, the performance of Human Activity Recognition models in real-life settings is usually lower than in laboratory settings due to the reduced quantity and quality of the sensors available in the former. Here, we propose using a suitable shared representation space to incorporate the information of additional sensors available during training time to address these limitations. We evaluate two representation spaces: one created using Feature Agglomeration and the other using Uniform Manifold Approximation and Projection (UMAP) under three conditions to evaluate their performance and robustness to noise: clean data, Gaussian noise, and Magnitude Warping noise using three datasets: Opportunity, Cooking, and PAMAP2. Our results consistently show that the representation spaces enhances performance relative to the conventional single-sensor method. The UMAP approach outperforms Feature Agglomeration, achieving up to a $14 \%$ improvement in the F1-Score metric when using clean data. In the presence of Gaussian noise, the UMAP representation space not only improves classification performance but also exhibits resilience to noise in the Opportunity and PAMAP2 datasets. While the UMAP method exhibits lower robustness to noise in the Cooking dataset, it still achieves the highest performance. When experimenting with Magnitude Warping noise, the UMAP representation space shows varying levels of robustness across datasets but still enhances performance to some extent. Using shared representations, we leverage the higher number and quality of sensors available in laboratory settings for training HAR models, while releasing the usual requirement of using the same number of sensors at the final deployment.