Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Inductive Link Prediction in Knowledge Graphs using Path-based Neural Networks

Canlin Zhang, Xiuwen Liu

TL;DR

This work tackles inductive link prediction in knowledge graphs by proposing SiaILP, a pair of path-based, embedding-free models that operate with a Siamese architecture to compare relation/path patterns rather than entity embeddings. By using an inverse-added KG and two complementary encoders—the connection-based model for connecting paths and the subgraph-based model for out-reaching paths—SiaILP achieves strong, scalable performance on inductive versions of WN18RR, FB15K-237, and Nell-995, without fine-tuning for new entities. Empirical results under both AUC-PR and ranking metrics demonstrate state-of-the-art performance in several settings, with ablations showing that three paths per entity provide a robust representation of local topology. The approach offers a practically efficient alternative to embedding-based or GNN-based methods for inductive link prediction, enabling direct transfer to new graphs with the same topological structure.

Abstract

Link prediction is a crucial research area in knowledge graphs, with many downstream applications. In many real-world scenarios, inductive link prediction is required, where predictions have to be made among unseen entities. Embedding-based models usually need fine-tuning on new entity embeddings, and hence are difficult to be directly applied to inductive link prediction tasks. Logical rules captured by rule-based models can be directly applied to new entities with the same graph typologies, but the captured rules are discrete and usually lack generosity. Graph neural networks (GNNs) can generalize topological information to new graphs taking advantage of deep neural networks, which however may still need fine-tuning on new entity embeddings. In this paper, we propose SiaILP, a path-based model for inductive link prediction using siamese neural networks. Our model only depends on relation and path embeddings, which can be generalized to new entities without fine-tuning. Experiments show that our model achieves several new state-of-the-art performances in link prediction tasks using inductive versions of WN18RR, FB15k-237, and Nell995. Our code is available at \url{https://github.com/canlinzhang/SiaILP}.

Inductive Link Prediction in Knowledge Graphs using Path-based Neural Networks

TL;DR

This work tackles inductive link prediction in knowledge graphs by proposing SiaILP, a pair of path-based, embedding-free models that operate with a Siamese architecture to compare relation/path patterns rather than entity embeddings. By using an inverse-added KG and two complementary encoders—the connection-based model for connecting paths and the subgraph-based model for out-reaching paths—SiaILP achieves strong, scalable performance on inductive versions of WN18RR, FB15K-237, and Nell-995, without fine-tuning for new entities. Empirical results under both AUC-PR and ranking metrics demonstrate state-of-the-art performance in several settings, with ablations showing that three paths per entity provide a robust representation of local topology. The approach offers a practically efficient alternative to embedding-based or GNN-based methods for inductive link prediction, enabling direct transfer to new graphs with the same topological structure.

Abstract

Link prediction is a crucial research area in knowledge graphs, with many downstream applications. In many real-world scenarios, inductive link prediction is required, where predictions have to be made among unseen entities. Embedding-based models usually need fine-tuning on new entity embeddings, and hence are difficult to be directly applied to inductive link prediction tasks. Logical rules captured by rule-based models can be directly applied to new entities with the same graph typologies, but the captured rules are discrete and usually lack generosity. Graph neural networks (GNNs) can generalize topological information to new graphs taking advantage of deep neural networks, which however may still need fine-tuning on new entity embeddings. In this paper, we propose SiaILP, a path-based model for inductive link prediction using siamese neural networks. Our model only depends on relation and path embeddings, which can be generalized to new entities without fine-tuning. Experiments show that our model achieves several new state-of-the-art performances in link prediction tasks using inductive versions of WN18RR, FB15k-237, and Nell995. Our code is available at \url{https://github.com/canlinzhang/SiaILP}.
Paper Structure (14 sections, 1 figure, 6 tables, 1 algorithm)

This paper contains 14 sections, 1 figure, 6 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Description on the Problem and Concepts
  5. Structure of the Connection-based Model
  6. Structure of the Subgraph-based Model
  7. Recursive Path-finding Algorithm
  8. Experimental Results
  9. Datasets and Baseline Models
  10. Training Protocols
  11. Evaluation Metrics
  12. Performances
  13. Ablation Studies
  14. Conclusion

Figures (1)

  • Figure 1: The architectures of our proposed models. (a): Given the connection paths from $s$ to $t$, our connection-based model will predict the target relation $r$. (b): Given out-reaching paths from $s$ and $t$ respectively, our subgraph-based model will predict the target relation $r$. (c): The architecture of one stack in our siamese neural network. (d): The architecture of our connection-based model. (e): The architecture of our subgraph-based model.