Comparative Study of Graph Neural Networks for Device Classification in Social IoT

The Social Internet of Things (SIoT) integrates the principles of social networks with the Internet of Things, enabling devices to interact based on social relationships. The information network produced from such connectivity has shown efficient query processing and collaborative tasks. Motivated by these characteristics and prospects of such a compiled relationship, this work studies the construction of a graph-compatible dataset from a public IoT platform's access records for device classification. The proposed graph-formatted dataset-building process utilises a two-level time-based approach. The first level uses a time-based resampling mechanism for graph definition, while the second one exploits nodes' historical co-location records for functional connectivity. Following the dataset construction, this work studies the performance of Graph Neural Networks (GNN) to predict a device type by employing two operators for node classification problems, e.g., Graph Convolutional Networks (GCN) and GraphSAGE. Our results show that models utilised both graph operators achieve satisfactory results, with an accuracy ≥ 0.8 and an $F1-\mathbf{score }\geq 0.7$ on validation and test sets defined.