TPGC Enhances Multi-Task Graph Pre-Training with Dual-Prior Prompts

This paper introduces TPGC (Task-Specific Prompt with Global Context), a novel approach designed to enhance multi-task graph pre-training, particularly in low-resource scenarios. Existing methods often rely on randomly initialized prompts, which can lead to poor alignment with specific tasks, graph structures, and overall transferability. TPGC addresses this by proposing a dual-prior prompt initialization solution that explicitly models the synergy between task-specific knowledge and global structural context. The framework incorporates a Task-Prior Injection Module, which conducts a short homologous multi-task pre-training on an auxiliary graph. This process allows prompt initialization to inherit optimization preferences relevant to various pretext tasks. Building on these task-aware representations, the Structure-Prior Injection Module then extracts transferable global structural context from the auxiliary graph, converting it into layer-wise prompt vectors by aggregating structurally informative node embeddings. Extensive experiments across six benchmarks for node and graph classification demonstrate that TPGC consistently outperforms state-of-the-art baselines in few-shot settings, while also requiring fewer downstream tunable parameters and lower runtime.