Directed Link Prediction Using GNN With Local and Global Feature Fusion

Yuyang Zhang, Xu Shen, Yu Xie, Ka-Chun Wong, Weidun Xie, Chengbin Peng

Abstract

Directed link prediction is a fundamental task in network analysis, yet it remains challenging due to the inherent asymmetry of directed edges and the limitations of existing Graph Neural Networks (GNNs) in capturing multi-scale structural information. This paper presents the Feature Fused Directed-line-graph (FFD) model, a framework designed to bridge the gap between local connectivity and global topology. By fusing path-based labels, community affiliations, and contrastive node embeddings, and subsequently transforming the graph into a directed line graph, the FFD model enables more expressive edge-centric representations. Experimental evaluations on six benchmark datasets show that FFD consistently outperforms state-of-the-art methods, achieving an average gain of 15% in AUC under data-scarce conditions.

Introduction

Link prediction—the task of predicting the existence of a connection between two nodes—is essential for applications ranging from social network recommendations to the reconstruction of metabolic pathways. In directed graphs, this task is complicated by the fact that a link from node A to node B does not imply a reciprocal link from B to A.

Existing deep learning models for link prediction often suffer from two primary issues: local myopia and global over-smoothing. Models that rely solely on immediate neighborhoods are highly sensitive to noise and missing data, while those focusing on global structures often fail to capture the specific local motifs that signal a potential connection. The FFD framework addresses these challenges by integrating information across three distinct scales: local (paths), medium (communities), and global (embeddings).

Methodology

The FFD framework operates through a three-stage pipeline designed to maximize the structural information available to the GNN.

1. Multi-Scale Hybrid Node Fusion

Instead of relying on a single feature vector, FFD concatenates three types of node information to create a comprehensive representation:

  • Local (Path-based labels): These capture immediate connectivity patterns and structural roles within a small radius.
  • Medium (Community-based labels): Utilizing unsupervised community detection, the model identifies the broader "neighborhood" or functional group a node belongs to. This provides a structural prior that local paths might overlook.
  • Global (Contrastive Node Embeddings): Through self-supervised learning, the model captures the node's relative position within the entire graph topology.

2. Directed Line Graph Transformation

A core innovation of the FFD model is the transformation of the directed subgraph into a directed line graph. In this representation, every edge in the original graph is treated as a node in the transformed graph. This shift allows the GNN to perform convolutions directly on the relationships (edges) rather than the entities (nodes). By treating edges as first-class citizens, the model can aggregate features from "links between links," significantly increasing the information density available for the prediction task.

3. GNN Classification

The transformed line graph is processed by a GNN, followed by a fully connected network. This architecture outputs the final probability of a link's existence between any two given nodes. To support this approach, the authors provide a formal theoretical proof using a Stochastic Blockmodel (SBM), demonstrating that the inclusion of community information mathematically improves the probability of correct link prediction.

Experimental Evaluation

Datasets

The FFD model was benchmarked against six diverse datasets to ensure robustness across different network types:

  • Citation Networks: Cora_ml, Cora, Citeseer, and Pubmed.
  • Transaction Networks: Bitcoin (a dynamic network of trust and transactions).
  • Communication Networks: p2p-Gnutella04 (a large-scale peer-to-peer network).

Performance and Key Results

The FFD model demonstrated significant improvements over traditional baselines such as GAE, GVAE, and Gravity-based models.

  • Robustness to Data Scarcity: The model's performance advantage was most pronounced when training data was limited. With only 30% of the training data, FFD achieved an average improvement of 15% in AUC and 7.1% in Average Precision (AP) over the best-performing baselines.
  • Ablation Insights: Ablation studies confirmed the necessity of the fusion approach. Removing community labels or global embeddings resulted in a marked decline in accuracy. Notably, the addition of community information alone provided a 9% AUC boost on specific datasets, highlighting the importance of "medium-scale" structural data.
  • Efficiency: Despite the complexity of the transformation, the model demonstrated faster convergence rates compared to competing deep learning approaches.

Discussion and Conclusion

The FFD model proves that integrating multi-scale features—particularly community structures—is vital for accurate directed link prediction. By utilizing directed line graphs, the framework effectively shifts the analytical focus to the edges themselves, providing a more expressive medium for GNNs to learn from.

However, the approach does introduce computational challenges. The line graph transformation increases the node count relative to the original graph's edge count, which may lead to overhead in extremely dense, large-scale networks. Future research will focus on simplifying the line graph representation to maintain performance while reducing the computational footprint, as well as adapting the framework for dynamic, time-varying graphs.

Summary updated

BibTeX
@article{zhang2025directed,
  author = {Yuyang Zhang and Xu Shen and Yu Xie and Ka-Chun Wong and Weidun Xie and Chengbin Peng},
  title = {{Directed Link Prediction Using GNN With Local and Global Feature Fusion}},
  journal = {IEEE Trans. Network Science \& Engineering},
  year = {2025},
  doi = {10.1109/tnse.2024.3498434},
}
Figure