Heterogeneous Graph Reinforcement Learning for Dependency-aware Multi-task Allocation in Spatial Crowdsourcing

Spatial Crowdsourcing (SC) has emerged as a significant paradigm for executing complex real-world projects and tasks, which are often decomposed into interdependent subtasks requiring diverse worker skills. A key challenge lies in the efficient allocation of multiple tasks to workers while respecting intricate subtask dependencies and skill requirements within a limited timeframe. This paper formally defines and addresses this Dependency-aware Multi-task Allocation (DMA) problem. We propose a novel framework, termed Heterogeneous Graph Reinforcement Learning for Task Allocation (HGRL-TA), to derive optimal allocation policies. Central to our framework is a multi-relation graph model that uniformly represents the complex problem state, and a Compound-path-based Heterogeneous Graph Attention Network (CHANet) designed to generate comprehensive state embeddings by capturing the intricate relationships among tasks and workers. The allocation decisions are made sequentially by a policy network, which is trained jointly with CHANet using the Proximal Policy Optimization (PPO) algorithm. Extensive experiments on both real-world and synthetic datasets demonstrate that the proposed HGRL-TA framework significantly outperforms seven representative baselines, showcasing its effectiveness and generalizability for the DMA problem.