A Method for Predicting Links in Complex Networks by Integrating Enclosure Subgraphs With High-Frequency Graph Information

• Published in: IEEE Access
• Main Authors: Zhiwei Zhang, Guangliang Zhu, Wenbo Qin
• Format: Article
• Language: English
• Published: IEEE 2024-01-01
• Subjects: Complex network; link prediction; graph neural network; enclosure subgraph; high-frequency graph information
• Online Access: https://ieeexplore.ieee.org/document/10517583/

Link prediction in complex networks, crucial for uncovering hidden or upcoming links between nodes and widely applicable in fields such knowledge graphs, faces challenges with current techniques. Predominantly, graph neural networks (GNN) based methods focus on learning node representations and use predictive components to assess the similarity of these representations for achieving link prediction. However, these approaches tend to accumulate errors in the predictive model and complicates the training process. Additionally, existing GNNs often display a low-pass filtering effect during network data processing, prioritizing low-frequency information while overlooking high-frequency details in node representations. These bias make GNNs mainly used for link prediction in strongly assortative networks and limit their performance on highly disassortative networks. Addressing these issues, this article introduces a novel framework that redefines the link prediction problem. By extracting enclosure subgraphs of both &#x2018;observed&#x2019; and &#x2018;unobserved&#x2019; links, we represent these links by corresponding enclosure subgraphs and transform link prediction into a problem of subgraphs classification. We innovate by combining high- and low-frequency graph information from the subgraphs, using an attention mechanism for integration, and constructing a graph neural network tailored to learn these subgraph representations, thus accomplishing the task of link prediction indirectly and enhancing link subgraphs classification accuracy. Our extensive experiments on recognized benchmark datasets, evaluated using the <inline-formula> <tex-math notation="LaTeX">$Hits\text{@}n$ </tex-math></inline-formula> metric, demonstrate that our method not only shows remarkable performance but also possesses strong generalization capabilities, positioning it as a potent baseline for link prediction tasks.