Leveraging Multi-Hop Neighborhoods in VGAE for Enhanced Link Prediction

Graph Neural Networks (GNNs) face numerous challenges when dealing with complex graph-structured data, such as insufficient capture of graph structure information, over-smoothing, and gradient vanishing. Although the traditional Variational Graph Autoencoder (VGAE) performs well in link prediction tasks, these issues still limit its application to complex graph data. To address this, this study proposes an improved VGAE model named MN-VGAE by introducing additional neighborhood aggregation in the encoder and incorporating residual connections, which enhances the ability to capture complex graph structure information. These improvements effectively alleviate over-smoothing and gradient vanishing issues, thereby enhancing the model's expressiveness and stability. Experimental results on multiple benchmark datasets show that MN-VGAE achieves significant performance gains in link prediction tasks, especially in terms of AP and AUC scores. This study demonstrates that the combination of deep architectures and stabilization techniques can significantly improve VGAE's performance on complex graph data.