Improving human activity recognition via graph attention network with linear discriminant analysis and residual learning

Abstract

This paper proposes a model based on the graph attention network (GAT) with the linear discriminant analysis (LDA) and the residual learning (AResGAT) for improving the human activity recognition (HAR). This approach can address the challenges due to the data scarcity and improve the robustness of the classification with respect to the adversarial attacks due to the operations in the adversarial environments by applying the adversarial loss principle. As a result, the reliability of the classification is improved. Moreover, by using the minimum spanning trees (MST) for constructing the graph, our proposed method can efficiently capture the information related to the complex interactions among the nodes which are used for representing different parts of the human body. Furthermore, since the gradient is computed in a more efficient way, the convergence of the training algorithm is fastened. Hence, the AResGAT tackles the vanishing gradient problem found in the training of the deep learning model. Finally, the computer numerical simulation results show that the AResGAT significantly outperforms the existing models for the various datasets.