Research map / Label-efficient and robust learning / Semi-supervised learning
Uncertain Knowledge Graph Completion via Semi-Supervised Confidence Distribution Learning
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 ssCDL, a semi-supervised UKG completion method that turns each triple confidence into a Gaussian confidence distribution and uses meta self-training to generate pseudo-labeled negative triples, improving confidence and link prediction on NL27k and CN15k.
Problem
Uncertain knowledge graph completion suffers from extremely imbalanced triple confidence distributions (mostly high-confidence triples), so embedding models fit low-confidence samples poorly and lack supervision for unseen confidences.
Approach
Transform each triple confidence into a Gaussian-shaped discrete confidence distribution so neighboring confidence levels supply extra supervision, and augment training with negative-sampled unseen triples given pseudo confidence distributions from a meta-learned generator (PCDG), which is optimized by the performance of the relational learner (CDL-RL) in a meta self-training loop.
Key results
- On NL27k, ssCDL reaches MSE 0.009 and MAE 0.042 for confidence prediction, reducing MSE by 52.6% versus the best baseline; WMRR 0.727 and Hits@1 0.636 for link prediction.
- On CN15k, ssCDL reduces MSE by 63.8% and MAE by 53.2% relative to the best baseline, with WMRR 0.207 and Hits@1 0.133.
- Ablations show removing confidence distribution learning (w/o cdl) gives the largest drop, e.g., NL27k MSE rising from 0.009 to 0.015.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add confidence distribution learning (ablation) | ssCDL · NL27k | MSE: 0.015 -> 0.009 | Helped |
| add confidence distribution learning (ablation) | ssCDL · NL27k | WMRR: 0.586 -> 0.727 | Helped |
| add confidence distribution learning (ablation) | ssCDL · CN15k | MSE: 0.044 -> 0.034 | Helped |
| add confidence distribution learning (ablation) | ssCDL · CN15k | WMRR: 0.149 -> 0.207 | Helped |
| add meta self-training (ablation) | ssCDL · NL27k | MSE: 0.010 -> 0.009 | Helped |
| add meta self-training (ablation) | ssCDL · NL27k | WMRR: 0.718 -> 0.727 | Helped |
| add meta self-training (ablation) | ssCDL · CN15k | MSE: 0.035 -> 0.034 | Helped |
| add meta self-training (ablation) | ssCDL · CN15k | WMRR: 0.200 -> 0.207 | Helped |
Surprising findings
- Link prediction gains are much smaller than confidence prediction gains, which the authors attribute to implicit optimization of link prediction through shared embeddings.
- The authors note that on CN15k the triple confidences are not clearly distinguishable in reliability, so the imbalance-based motivation may not hold there.
Limitations
- Evaluated on only two UKG datasets; CN15k's confidences are argued to lack meaningful reliability differences.
- Link prediction improvement over baselines is modest relative to confidence prediction.
- Hyperparameters (sigma, w_p, selection threshold) were tuned per dataset using validation sets.
Experimental setup
Datasets: NL27k CN15k
Benchmarks: NL27k CN15k
Metrics: MSE MAE WMRR Hits@1
Baselines: UKGE PASSLEAF UKGsE BEUrRE UPGAT
Builds on: UKGE PASSLEAF label distribution learning uncertainty weights multi-task loss DistMult
Closest papers by meaning
- KGCRR: An Effective Metric-Driven Knowledge Graph Completion Framework by Designing a Novel Upper Bound Function with Adaptive Approximation to Reciprocal Rank (AAAI 2025 · similarity 0.86)
- Towards Global-Topology Relation Graph for Inductive Knowledge Graph Completion (AAAI 2025 · similarity 0.85)
- Structure Is All You Need: Structural Representation Learning on Hyper-Relational Knowledge Graphs (ICML 2025 · similarity 0.85)
- Towards Continual Knowledge Graph Embedding via Incremental Distillation (AAAI 2024 · similarity 0.85)
- Towards Trustworthy Knowledge Graph Reasoning: An Uncertainty Aware Perspective (AAAI 2025 · similarity 0.84)
- Stability and Generalization Capability of Subgraph Reasoning Models for Inductive Knowledge Graph Completion (ICML 2025 · similarity 0.84)
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.