Understanding Expressivity of GNN in Rule Learning
Haiquan Qiu, Yongqi Zhang, Yong Li, Quanming Yao
TL;DR
This paper analyzes the expressivity of Graph Neural Networks in rule learning for knowledge graph reasoning, providing theoretical insights and proposing a new labeling strategy to enhance rule learning capabilities.
Contribution
It unifies GNNs with tail entity scoring into a common framework and characterizes their rule learning expressivity, introducing a novel labeling strategy to improve rule discovery.
Findings
Theoretical demonstration of GNNs' superiority in rule learning
Proposed labeling strategy enhances rule learning in KG reasoning
Experimental results confirm theoretical insights and method effectiveness
Abstract
Rule learning is critical to improving knowledge graph (KG) reasoning due to their ability to provide logical and interpretable explanations. Recently, Graph Neural Networks (GNNs) with tail entity scoring achieve the state-of-the-art performance on KG reasoning. However, the theoretical understandings for these GNNs are either lacking or focusing on single-relational graphs, leaving what the kind of rules these GNNs can learn an open problem. We propose to fill the above gap in this paper. Specifically, GNNs with tail entity scoring are unified into a common framework. Then, we analyze their expressivity by formally describing the rule structures they can learn and theoretically demonstrating their superiority. These results further inspire us to propose a novel labeling strategy to learn more rules in KG reasoning. Experimental results are consistent with our theoretical findings and…
Peer Reviews
Decision·ICLR 2024 poster
This paper leads a technical and thoughtful analysis to what kinds of relationships GNN-based models can effectively represent and effectively predict. Strengths of the paper include: * **Formalism** for link prediction in knowledge graphs using CML. This allows the authors to describe which kinds of rule structures each class of model is able to represent. It allows for the generalization of existing methods to represent broader classes of rules. * **Empirical Successes** are demonstrated acros
My main concern with this paper is the presentation / structure and the way in which that presentation and structure limits the reader from connecting both the clear theoretical advantages of the proposed class of GNN to both the limitations of other classes and the empirical successes. Please correct me if you think I have misinterpreted or misunderstood things or put emphasis points inappropriately. I am mentioning these presentation points because I think the paper has a number of very nice p
- The paper presents a novel approach to understanding RED-GNN and NBFNet from a rule-learning perspective. - The theoretical analysis reveals the advantages and limitations of existing popular GNNs in KG reasoning. The paper also provides experimental results to support the theoretical conclusion. - Furthermore, the paper proposes two GNNs for KG reasoning, which outperform state-of-the-art models on several datasets.
- In my opinion, the datasets used in Section 6.2 appear to be well-suited for rule-based methods. However, the most popular link prediction dataset, FB15K-237, was not included in this experiment. Therefore, I believe that the experimental results of Section 6.2 are insufficient to evaluate the effectiveness of the proposed method.
S1 The paper provides a thorough analysis of the expressivity of state-of-the-art GNNs used for KG reasoning. By unifying these GNNs into a common framework (QL-GNN), the authors offer a structured approach to understanding their capabilities and limitations in terms of rule structure learning. S2 The introduction of the QL-GNN framework and the subsequent EL-GNN model showcases the authors' innovative approach to addressing the gaps in the current understanding of GNNs for KG reasoning. The EL-
W1 While the EL-GNN model is introduced as an improvement over QL-GNN, there's limited discussion on its scalability. How does EL-GNN perform when applied to very large-scale KGs? Are there any computational constraints or challenges that users should be aware of? Moreover, the experiments in the paper employ relatively small datasets. It would greatly benefit the research to include larger datasets to demonstrate the effectiveness of the proposed methods on a more substantial scale. W2 The pa
Code & Models
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsAdvanced Graph Neural Networks · Bayesian Modeling and Causal Inference · Explainable Artificial Intelligence (XAI)