TransGCN: Coupling Transformation Assumptions with Graph Convolutional Networks for Link Prediction

Ling Cai, Bo Yan, Gengchen Mai, Krzysztof Janowicz, Rui Zhu

Abstract

Link prediction is an important and frequently studied task that contributes to an understanding of the structure of knowledge graphs (KGs) in statistical relational learning. Inspired by the success of graph convolutional networks (GCN) in modeling graph data, we propose a unified GCN framework, named TransGCN, to address this task, in which relation and entity embeddings are learned simultaneously. To handle heterogeneous relations in KGs, we introduce a novel way of representing heterogeneous neighborhood by introducing transformation assumptions on the relationship between the subject, the relation, and the object of a triple. Specifically, a relation is treated as a transformation operator transforming a head entity to a tail entity. Both translation assumption in TransE and rotation assumption in RotatE are explored in our framework. Additionally, instead of only learning entity embeddings in the convolution-based encoder while learning relation embeddings in the decoder as done by the state-of-art models, e.g., R-GCN, the TransGCN framework trains relation embeddings and entity embeddings simultaneously during the graph convolution operation, thus having fewer parameters compared with R-GCN. Experiments show that our models outperform the-state-of-arts methods on both FB15K-237 and WN18RR.

Introduction

Knowledge Graphs (KGs) serve as massive repositories of real-world facts, typically represented as triples (subject, relation, object). Despite their scale, KGs are often incomplete, making link prediction—the task of inferring missing facts—a critical research area.

Current methodologies generally follow two paths:

  1. Transformation-based models (e.g., TransE, RotatE) treat relations as geometric operations. While efficient, they often overlook the broader structural context of a node's neighborhood.
  2. Graph Convolutional Networks (e.g., R-GCN) excel at capturing local graph structure but are often parameter-heavy and fail to integrate relation embeddings directly into the encoding process.

TransGCN bridges these approaches by integrating transformation logic directly into a GCN framework, resulting in a model that is both structurally aware and parameter-efficient.

Methodology: Neighborhood Homogenization

The primary innovation of TransGCN is Neighborhood Homogenization. In a typical KG, a node's neighbors are connected via diverse relation types, creating a heterogeneous neighborhood that is difficult to aggregate. TransGCN addresses this by applying transformation assumptions—such as translation ($h + r \approx t$) or rotation in complex space—to "adjust" neighbor embeddings into a shared semantic space relative to the target node.

The TransGCN architecture illustrating simultaneous entity and relation updates.

Unlike previous models like R-GCN, TransGCN updates both entity and relation embeddings at every layer. Relation embeddings evolve through linear transformations, ensuring they remain synchronized with the entity vector space. The framework supports multiple variants, most notably TransE-GCN (real space) and RotatE-GCN (complex space).

Experimental Results

The models were evaluated on the FB15K-237 and WN18RR benchmarks. These datasets are specifically designed to be challenging by removing simple inverse relation patterns.

Link prediction performance comparison on FB15K-237 and WN18RR datasets.

As shown in the results, RotatE-GCN consistently achieves state-of-the-art performance. On FB15K-237, it reached a Mean Reciprocal Rank (MRR) of 0.356, a significant improvement over the base RotatE (0.338) and the R-GCN baseline (0.248). Similarly, on WN18RR, it achieved an MRR of 0.485.

Discussion and Insights

  • Structural Synergy: The performance gains over base transformation models confirm that explicitly modeling local graph structure provides a vital signal for link prediction.
  • Parameter Efficiency: By sharing weight matrices and utilizing transformation logic, TransGCN avoids the parameter explosion typical of relation-specific GCNs, reducing the risk of overfitting.
  • Neighborhood Density: Analysis revealed a "Goldilocks" effect regarding node degree; performance improves with more neighbors up to a certain threshold, after which noise from dense neighborhoods can hinder optimization.
  • Complex Space Advantages: The superior performance of RotatE-GCN suggests that rotation in complex space is more effective at capturing intricate relational patterns like symmetry and composition.

Conclusion

TransGCN demonstrates that coupling transformation assumptions with graph convolutions creates a powerful, efficient framework for knowledge graph completion. By learning entity and relation embeddings simultaneously within a homogenized neighborhood, it sets a new standard for link prediction tasks. Future work may explore more sophisticated relation evolution mechanisms and attention-based neighbor weighting to further refine performance in dense graph regions.

Summary updated

Abstract

Link prediction is an important and frequently studied task that contributes to an understanding of the structure of knowledge graphs (KGs) in statistical relational learning. Inspired by the success of graph convolutional networks (GCN) in modeling graph data, we propose a unified GCN framework, named TransGCN, to address this task, in which relation and entity embeddings are learned simultaneously. To handle heterogeneous relations in KGs, we introduce a novel way of representing heterogeneous neighborhood by introducing transformation assumptions on the relationship between the subject, the relation, and the object of a triple. Specifically, a relation is treated as a transformation operator transforming a head entity to a tail entity. Both translation assumption in TransE and rotation assumption in RotatE are explored in our framework. Additionally, instead of only learning entity embeddings in the convolution-based encoder while learning relation embeddings in the decoder as done by the state-of-art models, e.g., R-GCN, the TransGCN framework trains relation embeddings and entity embeddings simultaneously during the graph convolution operation, thus having fewer parameters compared with R-GCN. Experiments show that our models outperform the-state-of-arts methods on both FB15K-237 and WN18RR.
BibTeX
@inproceedings{cai2019transgcn,
  author = {Ling Cai and Bo Yan and Gengchen Mai and Krzysztof Janowicz and Rui Zhu},
  title = {{TransGCN: Coupling Transformation Assumptions with Graph Convolutional Networks for Link Prediction}},
  booktitle = {K-CAP},
  year = {2019},
  doi = {10.1145/3360901.3364441},
}
Figure