Personalized task allocation based on graph neural networks in mobile crowd sensing

Personalized Task Allocation (PTA) has emerged as a pivotal challenge in mobile crowd sensing, aiming to balance platform costs and worker satisfaction through optimized task allocation, yet facing dual critical challenges: the exponential growth of problem scale driven by technological advancements and severe data sparsity caused by limited historical allocation and privacy constraints. To address these challenges, this paper proposes PTA-GNNTR, a novel two-stage graph neural network framework where Stage 1 develops a Graph Neural Network-based Task Recommendation (GNNTR) that integrates multi-dimensional factors (historical records, social relationships, and task attributes) using graph convolutional networks, attention mechanisms, and multi-layer perceptrons to enhance feature representations and generate satisfaction-oriented pre-allocation results, while Stage 2 constructs a worker-task allocation bipartite graph (WTABG) for globally optimal matching to minimize platform costs under multiple constraints. Experimental results demonstrate PTA-GNNTR’s superiority in large-scale sparse-data scenarios, revealing the importance hierarchy of allocation factors (historical records $>$ virtual $>$ physical social relationships) and outperforming baselines across worker satisfaction, completion rate, and execution cost metrics, ultimately presenting a paradigm-shifting methodology that fuses GNNs with bipartite graph optimization to deliver a scalable solution for next-generation crowd sensing systems.