<p>Link prediction in complex networks underpins many computer and information science applications, such as recommendation, social network analysis, transportation and infrastructure planning, and biological interaction discovery. Subgraph-based methods such as SEAL encode an enclosing subgraph around each candidate edge and then score the edge based on the GNN-encoded representations of its endpoints. However, they typically rely on a single subgraph view per edge, which makes the learned edge representations brittle to small local perturbations (e.g., adding or removing a few non-endpoint nodes or edges), even though the underlying structural pattern of the edge should remain stable. We propose PREM-SEAL, an edge-level multi-view subgraph aggregation framework that learns locally perturbation-robust edge representations on top of SEAL style encoders. For each candidate edge, starting from the standard h-hop enclosing subgraph with the target edge removed, PREM-SEAL generates multiple local perturbation views by randomly dropping a subset of non-endpoint nodes and their incident edges. A pre-trained SEAL encoder is then used as a frozen feature extractor to obtain edge embeddings from all views, which are fused by a lightweight edge-level aggregator to produce the final link score. Under a strict train-graph protocol on multiple benchmark networks, PREM-SEAL consistently matches or outperforms single-view SEAL in terms of AUC and average precision, while naïve node-level EMA-based multi-context aggregation often degrades performance. Further analyses show that a small number of moderately perturbed views, especially with attention-based aggregation, is particularly beneficial for structurally weak edges, indicating that edge-level local multi-view subgraphs provide an effective way to enhance the robustness of subgraph-based link prediction to such local node-drop perturbations in complex networks.</p>