[Paper Review] Line Graph Neural Networks for Link Prediction

Lei Cai, Jundong Li+2 more|arXiv (Cornell University)|Oct 20, 2020

Advanced Graph Neural Networks39 references4 citations

TL;DR

This paper proposes Line Graph Neural Networks (LGLP) for link prediction by transforming the original graph into a line graph, where each edge becomes a node, enabling direct link feature learning via node classification without graph pooling. The method outperforms state-of-the-art models like SEAL on 14 diverse datasets, achieving higher AUC and AP with fewer parameters and faster convergence.

ABSTRACT

We consider the graph link prediction task, which is a classic graph analytical problem with many real-world applications. With the advances of deep learning, current link prediction methods commonly compute features from subgraphs centered at two neighboring nodes and use the features to predict the label of the link between these two nodes. In this formalism, a link prediction problem is converted to a graph classification task. In order to extract fixed-size features for classification, graph pooling layers are necessary in the deep learning model, thereby incurring information loss. To overcome this key limitation, we propose to seek a radically different and novel path by making use of the line graphs in graph theory. In particular, each node in a line graph corresponds to a unique edge in the original graph. Therefore, link prediction problems in the original graph can be equivalently solved as a node classification problem in its corresponding line graph, instead of a graph classification task. Experimental results on fourteen datasets from different applications demonstrate that our proposed method consistently outperforms the state-of-the-art methods, while it has fewer parameters and high training efficiency.

Motivation & Objective

• Address the information loss caused by graph pooling layers in existing deep learning-based link prediction models.
• Overcome the limitations of heuristic methods that are manually designed and not generalizable across different graph types.
• Develop a novel framework that reformulates link prediction as a node classification problem in line graphs to enable direct, efficient feature learning for links.
• Improve training efficiency and model generalization by eliminating the need for graph pooling while maintaining high predictive performance.

Proposed method

• Transform the original graph into its corresponding line graph, where each node in the line graph represents an edge in the original graph.
• Use a node labeling function to assign structural importance scores to nodes in the enclosing subgraph, enhancing feature representation.
• Apply graph convolutional networks (GCNs) directly on the line graph to learn node embeddings that represent the original edges.
• Use the learned embeddings from the line graph as features for predicting link existence, treating link prediction as a node classification task.
• Integrate node attributes by concatenating them with node labels in the line graph, enabling performance on attributed graphs.
• Train the model end-to-end using cross-entropy loss to optimize link prediction performance.

Experimental results

Research questions

• RQ1 Can reformulating link prediction as a node classification problem in line graphs eliminate information loss from graph pooling layers?
• RQ2 Does the proposed line graph neural network framework achieve better performance than state-of-the-art models like SEAL on diverse real-world datasets?
• RQ3 How does the proposed method perform under low-data regimes, such as using only 50% of training links?
• RQ4 Can the model maintain high performance on both plain and attributed graphs without significant degradation?
• RQ5 How does the training convergence speed of the proposed method compare to models that rely on graph pooling layers?

Key findings

• The proposed LGLP model outperforms all baseline methods, including SEAL, on 14 diverse datasets in terms of both AUC and AP metrics.
• On the Cora dataset with 50% training links, LGLP achieved an AUC of 81.45 and AP of 81.99, significantly outperforming SEAL’s AUC of 75.33 and AP of 77.69.
• With only 50% of training links, LGLP maintained AUC and AP values close to those achieved with 80% training data, indicating strong generalization under low-data settings.
• The model converged in 10–15 epochs, while SEAL required 50 epochs, demonstrating significantly faster training speed and lower parameter count.
• On attributed graphs like Cora, LGLP maintained stable performance, whereas SEAL’s performance degraded when node attributes were included.
• The line graph transformation preserved topological information effectively, enabling GCNs to learn high-quality edge representations without pooling operations.