<p>Influence maximization (IM) seeks a small seed set that maximizes influence spread. Deploying IM across independent platforms calls for platform-agnostic solutions. Traditional heuristics are limited by scalability, theoretical guarantees, and cross-network generalization. Learning-based approaches also face challenges: Reinforcement learning methods are computationally costly and rely on predefined diffusion models, while supervised models struggle to generalize due to graph-specific architectures. We propose <b>TDDL-PIM</b>, a dual-stage framework for platform-agnostic IM. TDDL-PIM shifts most computation to an offline stage that can be parallelized across networks, where it learns transferable diffusion regularities from heterogeneous seed–activation data. The resulting models support retraining-free deployment: The online stage performs sparsity-constrained inference on new platforms, completing in seconds on million-node graphs with a single GPU. Experiments on eight datasets under different diffusion models demonstrate improved influence spread and deployment efficiency. They also show that this high-performance-computing (HPC)-friendly parameterization-level topology decoupling effectively addresses the scalability and efficiency challenges of large-scale IM, enabling cross-platform reuse at deployment time.</p>