Adaptive Structural Similarity Guided GCN–GAT Framework for Robust Link Prediction

Social Network Analysis (SNA) has experienced major advancements with the growth of technology, which has made link prediction an important aspect of SNA, particularly regarding social network applications, recommendation systems, and various biological networks. Traditional Graph Neural Networks (GNN), such as Graph Convolutional Networks (GCN) and Graph Attention Networks (GAT) which primarily focus on aggregating node features and often neglect structural similarity (SS) between node pairs. This paper proposes a Structural Similarity-Infused Multi-Head (MH) Graph Attention Network (SS-MH-GAT) for link prediction that integrates graph-topology-based features into a hybrid GCN-GAT model. SS-MH-GAT introduces a structural-similarity-weight-extraction module, which learns the relative importance of the adjacency, Jaccard Index, clustering coefficient, and path-distance similarity metrics at the graph level. These globally weighted structural features are fused to form a similarity-aware graph representation, which is further refined through a Graph Convolutional Network. This is followed by the use of a Multi-Head Graph Attention Network to impart a locality-focused adaptability to the network by assigning importance weights to neighboring nodes, culminating in a DenseNet Classifier to infer probabilities of links based on the enhanced feature representations. Experiments on five diverse real-world datasets demonstrate that SS-MH-GAT surpasses baseline models. There is a significant performance improvement in accuracy (+12%), precision (+14%), recall (+17%), and AUC (+13%), which validates the effectiveness of the proposed approach.