Joint Structural Balance Feature Representation on Graph Convolution Networks for Link Prediction

Meixia He, Jianrui Chen, Zhihui Wang, Zhongshi Shao

Abstract: Link prediction identifies missing or future connections in networks. Existing methods often neglect varying neighbor contributions and lack mechanisms to reclassify networks based on these differences. To address these issues, a link prediction algorithm by jointing structural balance feature representation on graph convolution network (GCN) is proposed. GCN intersects with cognitive computing, which simulates human cognitive functions such as learning and decision-making. Specifically, the algorithm begins by calculating node similarities based on their attribute features. It then applies interactive importance filtering within the neighbor set to partition a typical network into positive and negative subnetworks, effectively creating a specialized signed network. Building upon this, the algorithm maintains a balanced structure in the divided signed network while employing signed GCN to extract both positive and negative neighborhood features of nodes. Concurrently, it filters out weak interaction node features. Subsequently, node feature representations are transformed into corresponding edge feature representations. Finally, these edge features are processed through a multi-layer perceptron to yield sign prediction results. The algorithm achieves strong performance on real datasets (e.g., citation/social networks), particularly in preserving structural balance and modeling neighbor contributions. This work bridges structural balance theory and GCNs, enhancing link prediction accuracy with interpretable neighbor interaction analysis. Future directions include scalability improvements for larger networks.