Cross-KG Link Prediction by Learning Substructural Semantics

Wen Wen,Shiyuan Wu,Ruichu Cai,Zhifeng Hao

DOI: https://doi.org/10.1007/s11063-024-11537-9

IF: 2.565

2024-02-16

Neural Processing Letters

Abstract:Link prediction across different knowledge graphs (i.e. Cross-KG link prediction) plays an important role in discovering new triples and fusing multi-source knowledge. Existing cross-KG link prediction methods mainly rely on entity and relation alignment, and are challenged by the problems of KG incompleteness, semantic implicitness and ambiguosness. To deal with these challenges, we propose a learning framework that incorporates both node-level and substructure-level context for cross-KG link prediction. The proposed method mainly consists of a neural-based tensor-completion module and a graph-convolutional-network module, which respectively captures the node-level and substructure-level semantics to enhance the performance of cross-KG link prediction. Extensive experiments are conducted on three benchmark datasets. The results show that our method significantly outperforms the state-of-the-art baselines and some interesting analysis on real cases are also provided in this paper.

computer science, artificial intelligence

What problem does this paper attempt to address?

The paper primarily focuses on addressing the issue of cross-knowledge graph (Cross-KG) link prediction. Specifically, the research targets the following challenges:

1. **Knowledge Graph Incompleteness**: Existing knowledge graphs often contain incomplete information, i.e., missing fact triples, which hinders the integration of information and link prediction across different knowledge graphs.

2. **Semantic Ambiguity**: The consistency of entity and relation naming and granularity differences make it difficult to find relationships between basic elements in different knowledge graphs.

3. **Semantic Implicitness**: The semantics in knowledge graphs are usually implicit, making it hard to directly discover potential connections between entities based on entity names.

To address the above issues, the paper proposes a learning framework that combines node-level and substructure-level contexts. This framework mainly includes two modules:

- **Node-level Representation (NR) Module**: Utilizes tensor decomposition methods to obtain node-level representations to alleviate the problem of knowledge graph incompleteness.

- **Substructure-level Representation (SR) Module**: Designs a graph convolutional network module to extract higher-order neighbor contexts, enhancing node representations at the substructure level to reduce the negative impacts of semantic ambiguity and implicitness.

The experimental section evaluates the performance of the proposed method on three benchmark datasets and compares it with several advanced baseline methods. The results show that the proposed model achieves significant advantages in the cross-knowledge graph link prediction task. Additionally, experiments on datasets with different entity overlap ratios verify the robustness of the proposed method to changes in entity overlap degree, especially maintaining good performance even with low entity overlap. Finally, ablation experiments further validate the importance of each component of the model.