This paper proposes a hybrid optimization framework for link prediction algorithms based on graph machine learning, integrating subgraph representation learning and dynamic coarsening strategies to overcome the shortcomings of traditional methods. Firstly, a two-layer neighborhood aggregation model is constructed by extracting local subgraphs (such as 2-hop neighborhoods) of target node pairs, and fusing node attributes and topological distance labels to enhance the expression ability of local context information. Secondly, to address the over-smoothing problem caused by the high repetition of nodes in subgraph learning, an attention-driven and structure-enhanced coarse-graining strategy is proposed, and combined with a triadic closure structure enhancement strategy, high similarity nodes are iteratively merged to construct virtual nodes to compress the subgraph scale, thereby reducing computational complexity while enhancing the model’s sensitivity to key topological patterns. Experimental results show that the modified model has achieved improved prediction accuracy on both the Cora and Cora indices, and it has achieved certain optimization results.