Link prediction, as a core downstream task of network analysis, mainly evaluates the connection probability between two nodes. In heterogeneous networks, the diverse types of nodes and relations increase the complexity and challenges of link prediction. Existing methods based on Graph Neural Network (GNN), largely focus on transductive link prediction-forecasting connections between a graph’s existing nodes. However, these methods underperform in real-world heterogeneous networks for inductive link prediction-key to predicting links of new isolated nodes whose link information is unavailable during training. Recent inductive link prediction approaches largely rely on node features, often overlooking graph structural information. Therefore, we propose an adversarial distillation-based heterogeneous graph inductive link prediction model (HGADI). The teacher model, a more expressive GNN, integrates local structures and high-order dependencies via contrastive learning to capture comprehensive structural context. The student model, a lightweight Multi-Layer Perceptron (MLP) that predicts missing links using node features, employs adversarial distillation to align its feature distributions with those of the teacher. By combining node feature learning with distilled structural knowledge, the student model effectively performs inductive link prediction, leveraging the GNN’s expressiveness and MLP’s fast inference. Experiments on three real-world heterogeneous datasets show that HGADI noticeably outperforms state-of-the-art (SOTA) methods in both the Area Under Curve (AUC) and Average Precision (AP) metrics. Additionally, HGADI achieves remarkable rapid inference and less memory usage compared to GNN-based approaches.