HGCNN: Deep Graph Convolutional Network for Sensor-Based Human Activity Recognition

Sensor-based human activity recognition(HAR) tasks often face issues of low-quality data, such as insufficient or incorrect data labels. But natural human activities are often con-textual, for example, an athlete needs to stretch before exercising. Therefore, we hypothesize unlabeled activity can be identified by contextual activity associations. Early statistical machine learning algorithms that rely on manual feature extraction frequently have knowledge constraints and can only capture superficial features, which prevents them from producing acceptable results in unsupervised or semisupervised circumstances. Deep learning on graph has made enormous strides in several disciplines in recent years because of its quick development. To address the issues highlighted above, we design a deep learning model called HGCNN based on graph neural network (GNN). Besides, we also make the sensor's time series data model as a fully connected subgraph with a sliding window size. Afterwards, a spectral graph convolution method is used to extract activity contextual relationship on the subgraph. The classification accuracy of this method on the Extra-Sensory dataset reaches 99.54% and HGCNN shows superior performance relative to previously CNN baseline, improving on macro F1-score reached up to 0.87.