GANCDE: Neural networks based on graphs and attention neural control differential equations for human activity recognition

The success of human activity recognition relies on the extraction of effective features, which can provide richer information for subsequent downstream tasks after embedding spatial fusion. However, due to the disruption of the original temporal relationships when integrating features from different variables, existing methods exert pressure on capturing temporal characteristics. In response to this issue, we propose a deep learning-based multi-channel architecture that combines two stages graph neural network and stacked attentional neural control differential equation method, named GANCDE. First, an adaptive graph architecture is employed to calculate similarity weights between different variables. Then, to obtain global information, multi-channel features are squeezed, and attention weights are calculated to overcome the drawbacks of graph convolution operations. This enables a more precise graph structure while eliminating redundant information. Next, the expression of the derivative of the temporal attention weight curve is obtained using tricubic interpolation and multi-scale convolution. Finally, the downstream task is completed using an N-layer convolutional architecture. Extensive experiments prove that GANCDE not only alleviates the issue of motion and non-motion features becoming similar caused by graph convolution, but also produces attention weight curves that better adhere to the continuity of actions. As a result, GANCDE achieves state-of-the-art performance in human activity recognition tasks. Specifically, on datasets such as WISDM, UCIHAR, and PAMAP2, the accuracy can reach 94.75, 93.50, and 96.37%, respectively.