Traditional GNNs struggle with heterogeneous graphs due to their inability to handle multi-type features and semantic noise. While meta-path-based methods exist, they often miss fine-grained heterogeneity and nuanced node dependencies, leading to poor link prediction. To address this, we propose a reinforcement learning (RL)-guided framework that dynamically selects optimal local heterogeneous subgraphs for target node pairs. Additionally, a heterogeneous GNN module aggregates neighborhood features using type-aware relevance metrics. Experiments on three real-world datasets show our method outperforms state-of-the-art baselines.