The paper aims to understand the logical expressivity of state-of-the-art Graph Neural Networks (GNNs) for knowledge graph reasoning. It first unifies popular GNN methods like NBFNet and RED-GNN into a common framework called QL-GNN, which scores triplets based on the tail entity representation.
The key insights are:
QL-GNN can learn rule structures described by formulas in the graded modal logic (CML) with the query entity as a constant. This class of rule structures includes chain-like rules and some more complex structures.
To learn rule structures beyond the capacity of QL-GNN, the paper proposes EL-GNN, which applies a labeling trick to additional entities in the graph besides the query entity. This allows EL-GNN to learn a broader class of rule structures.
Experiments on synthetic datasets validate the theoretical findings, showing that EL-GNN can learn rule structures that QL-GNN fails to capture. On real-world datasets, EL-GNN also demonstrates improved performance over QL-GNN methods.
The paper provides a formal analysis of the logical expressivity of state-of-the-art GNNs for knowledge graph reasoning, explaining their empirical success and inspiring novel GNN designs to learn more complex rule structures.
To Another Language
from source content
arxiv.org
Key Insights Distilled From
by Haiquan Qiu,... at arxiv.org 04-11-2024
https://arxiv.org/pdf/2303.12306.pdfDeeper Inquiries