Research map / Graph learning / Knowledge and temporal graphs
Lambda: Learning Matchable Prior For Entity Alignment with Unlabeled Dangling Cases
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 Lambda, a framework pairing a selective-aggregation GNN encoder (KEESA) with spectral contrastive learning and an iterative positive-unlabeled learner (iPULE) to align knowledge graphs with unlabeled dangling entities.
Problem
Entity alignment across knowledge graphs when dangling (unmatched) entities exist but are unlabeled, so only matchable pairs are labeled and the class prior of matchable entities is unknown.
Approach
Combines a GNN encoder with learnable dangling-indicator weights and relation projection attention, a spectral contrastive loss shared by alignment and detection, and a PU learning loss whose class priors are estimated iteratively with an EM-style procedure, without any labeled dangling entities.
Key results
- On DBP2.0 consolidated setting, Lambda achieves F1 of 0.836 on ZH-EN and 0.875 on EN-ZH for dangling detection, outperforming baselines that use 30% labeled dangling entities (best baseline MTransE+BR 0.767 and 0.785).
- For entity alignment on DBP2.0, Lambda reaches F1 0.344 on ZH-EN and 0.303 on EN-ZH, best or near best, though its precision is slightly lower than baselines.
- Lambda's dangling-unaware comparison on GA16K is superior or state-of-the-art on Hits@K, with Dual-AMN as the strongest competitor.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Lambda instead of the strongest listed dangling-detection baseline | DBP2.0 dangling detection · DBP2.0, six language directions | F1: .767 -> .836; .785 -> .875; .772 -> .821; .801 -> .843; .604 -> .685; .713 -> .740 | Helped |
| Use Lambda instead of the strongest listed entity-alignment baseline | DBP2.0 entity alignment · DBP2.0, six language directions | F1: .335 -> .344; .294 -> .303; .338 -> .362; .295 -> .310; .238 -> .271; .222 -> .251 | Helped |
| Remove the adaptive dangling indicator r_ei (ablation) | Lambda · DBP2.0 entity alignment | alignment performance | Hurt |
| Add relation projection attention for relation embeddings (ablation) | Lambda · DBP2.0 entity alignment | alignment performance | Helped |
| Replace relation projection attention with the original relation embedding (ablation) | Lambda · DBP2.0 entity alignment | alignment performance | Hurt |
| Increase the proportion of pre-aligned anchor nodes (ablation) | Lambda · DBP2.0 entity alignment | alignment performance | Helped |
| Use a 5% anchor ratio (ablation) | Lambda · Entity alignment datasets other than FR-EN and EN-FR | accuracy: >30% | Helped |
Surprising findings
- Lambda's precision is slightly inferior to baselines despite higher recall, attributed to dangling entities lying at the decision boundary.
- Removing the adaptive dangling indicator has minor impact on DBP2.0 because dangling entities there have lower degrees and are more isolated.
Limitations
- Precision of entity alignment and dangling detection is slightly lower than some baselines, due to dangling entities near the decision boundary.
- Theoretical convergence relies on strong EM-style assumptions, including a globally optimal classifier and a preference condition approximation.
- Embedding dimension limited to 96 on larger datasets due to out-of-memory issues.
Experimental setup
Datasets: DBP2.0 DBP15K GA16K GA-DBP15K DBP2.0-plus DBP2.0-minus
Benchmarks: DBP2.0 GA16K DBP15K GA-DBP15K
Metrics: Hits@K Precision Recall F1
Baselines: GCN-Align RSNs MuGNN KECG BootEA TransEdge MRAEA AliNet Dual-AMN MTransE
Builds on: MTransE AliNet GCN CSLS
Closest papers by meaning
- HLMEA: Unsupervised Entity Alignment Based on Hybrid Language Models (AAAI 2025 · similarity 0.87)
- Entity Alignment with Noisy Annotations from Large Language Models (NeurIPS 2024 · similarity 0.87)
- NeuSymEA: Neuro-symbolic Entity Alignment via Variational Inference (NeurIPS 2025 · similarity 0.85)
- Knowledge Graph Completion with Relation-Aware Anchor Enhancement (AAAI 2025 · similarity 0.85)
- MCL-NER: Cross-Lingual Named Entity Recognition via Multi-View Contrastive Learning (AAAI 2024 · similarity 0.84)
- UniversalNER: Targeted Distillation from Large Language Models for Open Named Entity Recognition (ICLR 2024 · 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.