S$^2$DN: Learning to Denoise Unconvincing Knowledge for Inductive Knowledge Graph Completion

TL;DR

S$^2$DN addresses inductive KGC on unseen entities by jointly smoothing semantic relation semantics and refining subgraph structure to reduce noise. The semantic smoothing module blends similar relations using a differentiable smoothing mechanism grounded in the Information Bottleneck, while the structure refining module learns task-dependent edge reliabilities to prune noisy interactions. Together, these components yield robust subgraph representations that, when fused with RGNN and GNN encoders, improve link prediction performance across standard inductive benchmarks and under contamination. The results demonstrate improved semantic consistency and reliable local structures, with practical implications for more robust knowledge graph reasoning in real-world noisy data settings. The approach offers a principled path to mitigate unconvincing knowledge and enhance inductive inference in KGs, with potential applications in biology and other noise-sensitive domains.

Abstract

Inductive Knowledge Graph Completion (KGC) aims to infer missing facts between newly emerged entities within knowledge graphs (KGs), posing a significant challenge. While recent studies have shown promising results in inferring such entities through knowledge subgraph reasoning, they suffer from (i) the semantic inconsistencies of similar relations, and (ii) noisy interactions inherent in KGs due to the presence of unconvincing knowledge for emerging entities. To address these challenges, we propose a Semantic Structure-aware Denoising Network (S$^2$DN) for inductive KGC. Our goal is to learn adaptable general semantics and reliable structures to distill consistent semantic knowledge while preserving reliable interactions within KGs. Specifically, we introduce a semantic smoothing module over the enclosing subgraphs to retain the universal semantic knowledge of relations. We incorporate a structure refining module to filter out unreliable interactions and offer additional knowledge, retaining robust structure surrounding target links. Extensive experiments conducted on three benchmark KGs demonstrate that S$^2$DN surpasses the performance of state-of-the-art models. These results demonstrate the effectiveness of S$^2$DN in preserving semantic consistency and enhancing the robustness of filtering out unreliable interactions in contaminated KGs.

Figures (8)

• Figure 1: (a) S$^2$DN outperforms GraIL in terms of Hits@10 on noisy KGs with different noise ratios (i.e., high robustness). (b) The relation edited_by shows a high percentage of being converted to other relations (enumerated on the x-axis) with similar semantics (i.e., high semantic consistency).
• Figure 2: The S$^2$DN framework comprises two modules for inductively predicting links in a given KG : (1) Smoothing relational semantics by blurring similar relations adaptively; (2) Refining the structure of subgraphs by learning reliable interactions dynamically.
• Figure 3: The architecture of Semantic Smoothing and Structure Refining modules of S$^2$DN.
• Figure 4: The transition ratio between the original and blurred relations on three datasets (V1 version). The element $m_{ij}$ in the matrix represents the proportion of the relation $i$ is smoothed to relation $j$.
• Figure 5: The big red and green nodes represent the source and target entities. The small nodes in red, orange, and blue are shared entities involved in $1-3$ hops between source and target nodes. The purple nodes indicate unshared entities.