Graph Neural Networks-based Clustering for Social Internet of Things

TL;DR

This paper introduces a GNN-based clustering framework for large-scale social IoT devices, leveraging graph embeddings and unsupervised algorithms to improve device grouping based on social relations.

Contribution

It presents a novel GNN-based approach for clustering social IoT devices, combining graph embeddings with traditional clustering algorithms for enhanced grouping.

Findings

GNN embeddings reflect device characteristics and relations.

The framework achieves promising clustering performance.

Comparison with Louvain shows competitive results.

Abstract

In this paper, we propose a machine learning process for clustering large-scale social Internet-of-things (SIoT) devices into several groups of related devices sharing strong relations. To this end, we generate undirected weighted graphs based on the historical dataset of IoT devices and their social relations. Using the adjacency matrices of these graphs and the IoT devices' features, we embed the graphs' nodes using a Graph Neural Network (GNN) to obtain numerical vector representations of the IoT devices. The vector representation does not only reflect the characteristics of the device but also its relations with its peers. The obtained node embeddings are then fed to a conventional unsupervised learning algorithm to determine the clusters accordingly. We showcase the obtained IoT groups using two well-known clustering algorithms, specifically the K-means and the density-based algorithm for discovering clusters (DBSCAN). Finally, we compare the performances of the proposed GNN-based clustering approach in terms of coverage and modularity to those of the deterministic Louvain community detection algorithm applied solely on the graphs created from the different relations. It is shown that the framework achieves promising preliminary results in clustering large-scale IoT systems.