Heterogeneous graph-enhanced multi-objective dependency-aware task offloading in vehicular edge computing

Vehicular edge computing (VEC) enables computation-intensive vehicular applications through collaborative vehicle-edge-cloud resources. However, dependency-aware task offloading remains challenging due to coupled subtask precedence constraints, heterogeneous computing resources, dynamic system scales, and diverse latency-energy preferences. Existing methods usually rely on fixed scalarization, heuristic subtask ordering, or metaheuristic solvers, limiting their adaptability to changing user preferences and resource configurations. To address these challenges, this paper proposes HGATE-EMOPPO, a graph-based multi-objective reinforcement learning framework for dependency-aware task offloading in VEC systems. The problem is formulated as a multi-objective Markov decision process to jointly optimize task completion delay and energy consumption. A heterogeneous graph is constructed to jointly model directed acyclic task dependencies and undirected competition relationships among server cores. Based on this representation, a heterogeneous graph attention auto-encoder (HGATE) with decoupled unsupervised pretraining is developed to extract reusable structural and relational embeddings. An envelope-based multi-objective proximal policy optimization algorithm (EMOPPO) is further designed to learn a unified preference-aware policy, where a feature-wise linear modulation mechanism enables adaptive decision-making under different latency-energy trade-offs. Unlike conventional approaches that depend on predefined subtask priorities, HGATE-EMOPPO dynamically selects both executable subtasks and offloading destinations, enabling size-agnostic and heuristic-free scheduling. Extensive simulations show that HGATE-EMOPPO achieves competitive multi-objective offloading performance in terms of hypervolume and adaptation error. Although some baselines obtain better results for specific metrics under particular settings, HGATE-EMOPPO provides a flexible single-policy preference-aware solution and generalizes effectively to unseen task topologies, varying communication-computation characteristics, and heterogeneous server-core configurations without retraining, indicating its robustness and scalability in realistic VEC environments.