Real-world metrics.

The paper presents a decentralized, real-time clustering method designed for large-scale, distributed environments such as the Internet of Things (IoT). The approach combines compressed sensing for dimensionality reduction with a consensus protocol for distributed aggregation, enabling each node to generate compact, consistent summaries of the system’s clustering structure with minimal communication overhead. These representations are processed by a pre-trained neural network to reconstruct the global clustering state entirely without centralized coordination. Unlike traditional methods that depend on static topologies and centralized computation, this system adapts to dynamic network changes and supports on-the-fly processing. The system suits IoT applications where data must be processed locally, and immediate results are essential. Experiments on both synthetic and real-world datasets show that the method significantly outperforms baseline approaches in clustering accuracy, making it highly suitable for resource-limited, decentralized IoT scenarios.