Conflict‐Aware Composition of IoT Services: An Approach Based on MetaPath‐Guided Trust Learning

ABSTRACT In modern intelligent cyber‐physical systems (ICPS), the integration of internet of things (IoT), service‐oriented computing (SOC), machine learning (ML), and foundational computing paradigms (edge/fog, sensor clouds) has enhanced the encapsulation and orchestration of diverse IoT resources into cooperative IoT services, referred to as composite IoT applications. However, components of these applications can encounter conflicting conditions, such as sensor incompatibility, data aggregation issues, privacy risks, regulatory compliance violations, vendor lock‐in, and business model disagreements, that critically impair IoT workflow effectiveness. Existing studies fail to jointly address trust propagation and conflict resolution across heterogeneous ICPS entities, which limits the reliability of composed IoT applications. Aiming to ensure seamless and dependable execution of composite IoT workflows, this paper introduces Trust‐MPGNN, a unified conflict‐ and trust‐aware framework for handling IoT service composition requests. The framework relies on a trust knowledge graph (TKG) that explicitly models conflict and trust relations among ICPS entities (e.g., service providers, IoT services, and underlying IoT resources). Building on this representation, we resort to a metapath‐guided variant of graph neural networks (GNNs) to infer latent trust dependencies and propagate trust across ICPS entities through semantically guided neighborhood sampling. Finally, we define algorithms that filter and compose non‐conflicting and trustworthy IoT services and resources with respect to user‐defined workflow requirements and constraints. The experimental studies, conducted on a hybrid ICPS dataset, show that Trust‐MPGNN consistently surpasses three baseline methods (Trust‐GNN, FFCA‐IoTSC, and TQoSC) across all evaluation metrics. Specifically, Trust‐MPGNN achieves a composition success rate of up to 98.5%, a trust score of 0.95, and a conflict severity as low as 0.04 under different conflict density conditions, representing improvements of up to 7.6%, 8.0%, and 56% over the best‐performing competing baseline, respectively.