Contrasting Global and Local Representations for Human Activity Recognition using Graph Neural Networks

Human Activity Recognition has achieved notable improvements with the emergence of Deep Learning models for automated feature extraction. Those models allow to extract complex translational-invariant features and to exploit the temporal dependencies from sensors' time series data. This work posits additional dependencies between sensors beyond the time dimension, such as physical proximity, which are equally important for the characterization of human activities. We leverage such spatial dependencies by modeling them as a graph. Using Graph Neural Networks (GNNs), we learn global and local representations of the intra- and intersensor dependencies. We empirically show that by maximizing the mutual information between the local and global representations, the performance of the recognition models can be significantly improved. Our results show a clear improvement over previous works based on CNNs, LSTMs, Attention-based and other more complex GNNs-based architectures. Our source code is available at: https://github.com/atello/GNNs4HAR.