Task-Specific Trust Prediction with GNN for Minimized Risk of Task Completion in Dynamic Collaborative Systems

Complex task completion in resource-constrained Internet of Things (IoT) systems relies exclusively on trusted collaboration among distributed devices. However, dynamic task requirements and collaborator features inevitably introduce collaboration uncertainties, thus increasing task completion risks. To minimize such risks, we propose TST-GNN, a Task-Specific Trust prediction method based on Graph Neural Networks (GNN) to construct reliable collaborator groups for task completion with reduced collaboration uncertainties. Specifically, the collaboration uncertainties are quantified in this paper by our proposed risk assessment metric, i.e., Risk of Task Completion. Then an optimization problem for reliable collaborator selection is formulated to minimize the long-term risk of task completion while completing various tasks with diverse goals. To solve this problem, we first model the dynamic collaborative system as a set of trust-guided bipartite graphs. Then a task-specific trust evaluation model is trained based on GNN for accurate trust prediction and reliable collaborator selection, where the excessive training overhead is reduced by only training the graph knowledge from the most relevant tasks. Moreover, a risk-aware trust update policy is proposed to continuously validate and refine the collaborator selection strategy for effective risk control. Simulation results demonstrate the effectiveness and robustness of the proposed strategy in concurrently reducing the long-term risk of task completion and completing complex tasks.