Research map / Graph learning / Knowledge and temporal graphs
Modeling Knowledge Graphs with Composite Reasoning
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.
Unifies several knowledge graph completion model families under a composite reasoning framework and proposes CompilE, which penalizes erroneous entity composition in tensor factorization models to improve link prediction.
Problem
Tensor factorization KGC models can learn embeddings that compose predictions from unrelated entities, violating the collaborative filtering assumption and hurting generalization.
Approach
Views translational, TF, instance-based (CIBLE) and DURA regularizer models as composite reasoning over known facts of other entities, defines a composition risk for TF models, approximates its lower bound via connected entities, and adds it as a penalty term to the training loss.
Key results
- CompilED reaches 0.372 MRR on FB15k-237 vs 0.371 for DURA and 0.367 for N3 (ComplEx base), and 0.563 Hits-based MRR on WN18RR vs 0.560 for DURA (MRR 0.563 reported in table).
- On Kinship, CompilEN improves MRR over other TF-based models by 3.2% per the authors; CompilED reaches 0.861 MRR on UMLS vs 0.841 for DURA.
- Approximated composition risk shows a stronger Spearman correlation with MRR than entity frequency across FB15k-237 and WN18RR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add composition-risk alleviation to ComplEx with DURA regularization | ComplEx with DURA regularization · FB15k-237 | MRR: 0.371 -> 0.372 | Helped |
| add composition-risk alleviation to ComplEx with N3 regularization | ComplEx with N3 regularization · FB15k-237 | MRR: 0.367 -> 0.368 | Helped |
| add composition-risk alleviation to ComplEx with DURA regularization | ComplEx with DURA regularization · WN18RR | MRR: 0.491 -> 0.495 | Helped |
| add composition-risk alleviation to ComplEx with N3 regularization | ComplEx with N3 regularization · WN18RR | MRR: 0.488 -> 0.492 | Helped |
| add composition-risk alleviation to ComplEx with DURA regularization | ComplEx with DURA regularization · UMLS | MRR: 0.841 -> 0.861 | Helped |
| add composition-risk alleviation to ComplEx with N3 regularization | ComplEx with N3 regularization · UMLS | MRR: 0.842 -> 0.868 | Helped |
| add composition-risk alleviation to ComplEx with DURA regularization | ComplEx with DURA regularization · Kinship | MRR: 0.670 -> 0.724 | Helped |
| add composition-risk alleviation to ComplEx with N3 regularization | ComplEx with N3 regularization · Kinship | MRR: 0.697 -> 0.713 | Helped |
Surprising findings
- CP and ComplEx learn embeddings whose composition relies on unrelated entities (e.g., Macau, TC Islands for Mexico), violating collaborative filtering despite fitting training data.
Limitations
- Relies on the assumption that connected entities are a weaker restriction of reliable entities; lower bound is only an approximation of composition risk.
- Gains are small on large datasets; NBFNet remains better on larger benchmarks.
- Theory assumes one-to-one relations cannot be represented under the composition assumption.
Experimental setup
Datasets: FB15k-237 WN18RR UMLS Kinship
Benchmarks: FB15k-237 WN18RR UMLS Kinship
Metrics: MRR Hits@1 Hits@3 Hits@10 MR Spearman correlation residual ratio
Baselines: TransE RotatE NeuralLP RNNLogic CIBLE NBFNet DistMult CP ComplEx DURA N3
Builds on: ComplEx N3 regularization DURA CIBLE CP
Closest papers by meaning
- SPAC: Sparse Partitioning and Adaptive Core Tensor Pruning Model for Knowledge Graph Completion (AAAI 2025 · similarity 0.89)
- Graph Reasoning Transformers for Knowledge-Aware Question Answering (AAAI 2024 · similarity 0.87)
- DuetGraph: Coarse-to-Fine Knowledge Graph Reasoning with Dual-Pathway Global-Local Fusion (NeurIPS 2025 · similarity 0.87)
- Context-aware Inductive Knowledge Graph Completion with Latent Type Constraints and Subgraph Reasoning (AAAI 2025 · similarity 0.87)
- Amplifying Diversity and Quality in Commonsense Knowledge Graph Completion (Student Abstract) (AAAI 2024 · similarity 0.87)
- Toward a Unified Statistical Theory of Unsupervised Pretraining and Supervised Neural Knowledge Graph Learning (NeurIPS 2026 · 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.