Graph neural network enhanced Internet of Things node classification with different node connections

Abstract

Internet of Things (IoT) has profoundly impacted human life by providing ubiquitous connectivity and unique advantages. As the demand for IoT applications continues to grow, the number of connected devices is increasing at a rapid pace. This growth poses challenges in identifying data sources and managing data in large networks. The graph data structure offers a meaningful way to represent IoT networks, where nodes represent devices and edges represent their connections. In this study, we convert IoT networks into graph representations, considering two approaches: fully connected node graphs and randomly connected node graphs. Graph neural networks (GNNs) are highly effective for processing graph data, as they can capture relationships within graph structures based on their topological properties. We utilize GNNs to perform node classification tasks for IoT networks. Seven different GNN models were investigated to perform node classification tasks on both complete and random graphs. The experimental results indicate that the SAGEConv model achieves high classification accuracy under dense network conditions. Additionally, the CHEBYSHEVConv model performs well with fully connected graphs, while the TAGConv model demonstrates strong performance with randomly connected graphs.