Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
MoEMeta: Mixture-of-Experts Meta Learning for Few-Shot Relational 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 MoEMeta, a meta-learning framework for few-shot KG relational learning that combines a globally shared sparse mixture-of-experts relation-meta learner with a per-task lightweight projection adaptation, reporting state-of-the-art results on Nell-One, Wiki-One and FB15K-One.
Problem
Few-shot relational learning on knowledge graphs, where meta-learning methods learn relation meta-knowledge in isolation (ignoring cross-task shared patterns) and use a single global initialization that cannot adapt well to task-specific local contexts.
Approach
A sparsely-gated mixture of MLP experts is globally trained to produce relation-meta from support triplets (capturing shared relational prototypes), and a task-specific set of three projection vectors (TransD-style) is fine-tuned in the inner loop to adapt head, relation and tail embeddings locally.
Key results
- On Nell-One 1-shot, MoEMeta reaches MRR 0.322 and Hits@1 0.228, versus 0.288 and 0.184 for the second-best HiRe.
- On Wiki-One 5-shot, MoEMeta achieves MRR 0.377 versus 0.371 for HiRe.
- On FB15K-One 3-shot, MoEMeta achieves MRR 0.423 versus 0.405 for RelAdapter, and Hits@10 0.651 versus 0.575.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use MoEMeta instead of HiRe | few-shot relational learning · Nell-One 1-shot | MRR: 0.288 -> 0.322 | Helped |
| use MoEMeta instead of HiRe | few-shot relational learning · Nell-One 5-shot | MRR: 0.306 -> 0.339 | Helped |
| use MoEMeta instead of MetaR-P | few-shot relational learning · Wiki-One 1-shot | MRR: 0.314 -> 0.325 | Helped |
| use MoEMeta instead of HiRe | few-shot relational learning · Wiki-One 5-shot | MRR: 0.371 -> 0.377 | Helped |
| use MoEMeta instead of RelAdapter | few-shot relational learning · FB15K-One 3-shot | MRR: 0.405 -> 0.423 | Helped |
| add attentive neighbor aggregation (ablation) | MoEMeta · Nell-One 1-shot | MRR: 0.311 -> 0.322 | Helped |
| add attentive neighbor aggregation (ablation) | MoEMeta · Nell-One 5-shot | MRR: 0.328 -> 0.339 | Helped |
| add attentive neighbor aggregation (ablation) | MoEMeta · FB15K-One 3-shot | MRR: 0.413 -> 0.422 | Helped |
| use MoE instead of an MLP-based relation-meta learner (ablation) | MoEMeta · Nell-One 1-shot | MRR: 0.291 -> 0.322 | Helped |
| use MoE instead of an MLP-based relation-meta learner (ablation) | MoEMeta · Nell-One 5-shot | MRR: 0.293 -> 0.339 | Helped |
| use MoE instead of an MLP-based relation-meta learner (ablation) | MoEMeta · FB15K-One 3-shot | MRR: 0.401 -> 0.422 | Helped |
| add task-tailored local adaptation (ablation) | MoEMeta · Nell-One 1-shot | MRR: 0.301 -> 0.322 | Helped |
Surprising findings
- Hits@1 on FB15K-One 3-shot (0.302) is only marginally above GANA (0.301), and MoEMeta's Hits@1 gain is small there despite the large MRR gain.
- Removing the MoE and replacing it with an MLP relation-meta learner causes the largest drop on Nell-One 5-shot (MRR 0.339 to 0.293).
Limitations
- Improvements on Hits@1 for FB15K-One are marginal relative to GANA.
- Experts count M is tuned on Nell-One 5-shot only and the paper notes M should stay below the number of tasks to avoid overfitting.
Experimental setup
Datasets: Nell-One Wiki-One FB15K-One
Benchmarks: Nell-One Wiki-One FB15K-One
Metrics: MRR Hits@1 Hits@5 Hits@10
Baselines: TransE TransD DistMult ComplEx GMatching FSKGC FAAN MetaR GANA HiRe RelAdapter
Builds on: MAML TransD MetaR GMatching RelAdapter
Closest papers by meaning
- Automated Relational Meta-learning (ICLR 2020 · similarity 0.90)
- Hierarchical Relational Learning for Few-Shot Knowledge Graph Completion (ICLR 2023 · similarity 0.89)
- Few-shot Relational Reasoning via Connection Subgraph Pretraining (NeurIPS 2022 · similarity 0.87)
- UNITE: Universal kNowledge Integration from Task-specific Experts (ICLR 2026 · similarity 0.87)
- Adaptive Compositional Continual Meta-Learning (ICML 2023 · similarity 0.87)
- Few-Shot, No Problem: Descriptive Continual Relation Extraction (AAAI 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.