Research map / Graph learning / Knowledge and temporal graphs
S²DN: Learning to Denoise Unconvincing Knowledge for Inductive Knowledge Graph Completion
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes S²DN, an inductive KGC model that combines a Gumbel-softmax relation smoothing module (semantic smoothing) with a learnable Bernoulli edge-reliability module (structure refining) over enclosing subgraphs, evaluated on WN18RR, FB15k-237 and NELL-995 inductive splits.
Problem
Inductive knowledge graph completion for emerging entities suffers from inconsistent representations of semantically similar relations and noisy, unreliable interactions in enclosing subgraphs.
Approach
Adaptively merge similar relations into shared representations via Gumbel-softmax smoothing (regularized by a KL loss), and learn edge reliability with a reparameterized Bernoulli gate to prune noisy edges; the smoothed-relation RGNN and refined-structure GCN representations are concatenated for classification.
Key results
- S²DN achieves average Hits@10 of 81.23 on WN18RR inductive splits versus 79.19 for the best baseline SNRI.
- On FB15k-237 inductive splits S²DN reaches average Hits@10 of 81.25 versus 79.87 for SNRI.
- Under 50% semantic noise on WN18RR V1, S²DN's Hits@10 drops 5.2% versus 8.7% for RMPI.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Semantic Smoothing (ablation) | S²DN · WN18RR (Avg.) | Hits@10: 78.49 -> 81.23 | Helped |
| add Structure Refining (ablation) | S²DN · WN18RR (Avg.) | Hits@10: 76.31 -> 81.23 | Helped |
| add Semantic Smoothing (ablation) | S²DN · FB15k-237 (Avg.) | Hits@10: 79.68 -> 81.25 | Helped |
| add Structure Refining (ablation) | S²DN · FB15k-237 (Avg.) | Hits@10: 78.97 -> 81.25 | Helped |
| add Semantic Smoothing (ablation) | S²DN · NELL-995_V1 | Hits@10: 57.00 -> 63.50 | Helped |
| add Structure Refining (ablation) | S²DN · NELL-995_V1 | Hits@10: 58.50 -> 63.50 | Helped |
| add 50% semantic noise to the training KG | S²DN · WN18RR_V1 | Hits@10: 87.64 -> 83.12 | Hurt |
| add 50% structural noise to the training KG | S²DN · WN18RR_V1 | Hits@10: 87.64 -> 84.98 | Hurt |
| add 50% semantic noise to the training KG | S²DN · FB15k-237_V1 | Hits@10: 67.34 -> 64.97 | Hurt |
| add 50% structural noise to the training KG | S²DN · FB15k-237_V1 | Hits@10: 67.34 -> 64.89 | Hurt |
| replace Attention reliability estimator with MLP (ablation) | S²DN · WN18RR_V1 | Hits@10: 87.64 -> 86.70 | Hurt |
| replace Attention reliability estimator with Weighted Cosine (ablation) | S²DN · WN18RR_V1 | Hits@10: 87.64 -> 87.50 | Hurt |
Surprising findings
- On NELL-995, S²DN is only comparable to baselines and RMPI outperforms it on several splits, and on the V1 split of NELL-995 S²DN's Hits@10 is not the best.
Limitations
- Gains on NELL-995 are modest and not consistently best across splits.
- Relies on enclosing subgraph extraction, which scales with subgraph size and degree.
- Ranking uses 50 sampled negatives rather than all entities, limiting comparability to other protocols.
Experimental setup
Datasets: WN18RR FB15k-237 NELL-995
Benchmarks: WN18RR inductive V1-V4 FB15k-237 inductive V1-V4 NELL-995 inductive V1-V4
Metrics: Hits@1 Hits@10 MRR
Baselines: NeuralLP DRUM A*Net CoMPILE TAGT SNRI RMPI GraIL KAPING (LLaMA2-7B)
Builds on: GraIL RMPI SNRI CoMPILE RGCN
Closest papers by meaning
- Towards Global-Topology Relation Graph for Inductive Knowledge Graph Completion (AAAI 2025 · similarity 0.92)
- Context-aware Inductive Knowledge Graph Completion with Latent Type Constraints and Subgraph Reasoning (AAAI 2025 · similarity 0.90)
- Stability and Generalization Capability of Subgraph Reasoning Models for Inductive Knowledge Graph Completion (ICML 2025 · similarity 0.89)
- Replacing Paths with Connection-Biased Attention for Knowledge Graph Completion (AAAI 2025 · similarity 0.89)
- Towards Graph Foundation Models: Training on Knowledge Graphs Enables Transferability to General Graphs (NeurIPS 2025 · similarity 0.88)
- Knowledge Graph Error Detection with Contrastive Confidence Adaption (AAAI 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.