DUAL DENOISING LOGICAL REASONING FOR INDUCTIVE KNOWLEDGE GRAPH COMPLETION

Sun Kai, Huajie Jiang, Yongli Hu, Baocai Yin

Primary Area: general machine learning (i.e., none of the above)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Inductive Knowledge Graph Completion, Dual Denoising

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract: In recent years, message-passing Graph Neural Networks (GNNs) have been widely used to tackle the problem of inductive knowledge graph completion. Though great progress has been made in GNN-based knowledge graph reasoning, it still suffers from the noise existing in irrelevant entities. These noises accumulated exponentially as the reasoning process continues, significantly impacting the overall performance of the model. To tackle this problem, several node-based sampling methods have been proposed for denoising. However, they do have inherent limitations. Firstly, they rely on node scores to evaluate node importance, which cannot effectively assess the quality of paths in GNN-based reasoning. Secondly, they often overlook noise interference caused by irrelevant edges. To address these problems, we propose a dual denoising logical reasoning (DDLR) framework, which integrates path-based and edge-based sampling to achieve comprehensive denoising. Specifically, DDLR employs a path-scoring mechanism to evaluate the importance of paths, aiming to remove irrelevant paths. Moreover, DDLR leverages rules within the knowledge graph to remove irrelevant edges. Through the dual denoising process, we can achieve more effective logical reasoning. To demonstrate the effectiveness of the DDLR framework, conduct experiments on three benchmark datasets, and our approach achieves state-of-the-art performance.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 4984