This paper addresses the problem of link prediction in network-structured data, focusing on the limitations of traditional heuristic methods and proposing a novel approach to learn heuristics from local subgraphs. The authors develop a γ-decaying heuristic theory, which unifies various heuristics and proves that they can be approximated from local subgraphs. They then propose SEAL, a framework that uses a graph neural network (GNN) to learn general graph structure features from local enclosing subgraphs. SEAL outperforms existing heuristic methods, latent feature methods, and network embedding methods, demonstrating its effectiveness and robustness across various datasets. The key contributions include a new theory for learning heuristics from local subgraphs and the SEAL framework, which leverages GNNs to achieve superior performance in link prediction.This paper addresses the problem of link prediction in network-structured data, focusing on the limitations of traditional heuristic methods and proposing a novel approach to learn heuristics from local subgraphs. The authors develop a γ-decaying heuristic theory, which unifies various heuristics and proves that they can be approximated from local subgraphs. They then propose SEAL, a framework that uses a graph neural network (GNN) to learn general graph structure features from local enclosing subgraphs. SEAL outperforms existing heuristic methods, latent feature methods, and network embedding methods, demonstrating its effectiveness and robustness across various datasets. The key contributions include a new theory for learning heuristics from local subgraphs and the SEAL framework, which leverages GNNs to achieve superior performance in link prediction.