A Relation Is a Rotation

2026-01-11 · Source: Agus’s Substack · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Advanced, long

Summary

RotatE is a knowledge graph embedding model that represents relations as rotations in the complex plane, addressing the "symmetry problem" where translation-based models like TransE collapse symmetric relations. By restricting complex relation channels to the unit circle, RotatE makes relations pure phase rotations. This allows symmetric relations, such as "spouse" and "siblingOf" in a family graph, to be represented as 180° (π) half-turns without collapsing, exhibiting an involution defect of 0.000, while asymmetric relations like "parentOf" show a defect of 3.203. RotatE also naturally handles inverse relations via complex conjugates and composition through angle addition. Its use of ℓ₂ distance scoring enables highly efficient evaluation, reducing computation by a factor of |ℰ| (e.g., four orders of magnitude on WN18RR with 40,000 entities). Benchmarks on UMLS and WN18RR demonstrate RotatE's significant improvement in Hits@1 scores (0.686 vs. TransE's 0.535 on UMLS; 0.416 vs. TransE's 0.01 on WN18RR) by resolving entities that TransE collapses.

Key takeaway

For Machine Learning Engineers evaluating knowledge graph embedding models, RotatE provides a robust solution for graphs rich in symmetric and inverse relations. You should adopt RotatE to prevent entity collapse, which significantly boosts Hits@1 performance over translation-based models like TransE. If your graph also requires order-dependent relation compositions, explore QuatE for its non-commutative properties.

Key insights

RotatE models knowledge graph relations as rotations in the complex plane, resolving symmetry issues and enabling efficient inference.

Principles

• Symmetric relations are half-turns (π rotation).
• Inverse relations are complex conjugates.
• Relation composition adds rotation angles.

Method

RotatE restricts complex relation channels to the unit circle (|rₖ|=1), making relations pure phase rotations rₖ = eⁱᶿᵏ. It scores facts using s(h,r,t)=γ−‖h⊙r−t‖.

In practice

• Use RotatE for graphs with many symmetric relations.
• Consider QuatE for order-dependent relation compositions.
• Leverage ℓ₂ distance for faster RotatE evaluation.

Topics

• Knowledge Graph Embeddings
• RotatE Model
• Symmetric Relations
• Relation Composition
• Quaternions
• Model Evaluation

Code references

• asudjianto-xml/Knowledge-Graph-Geometry
• asudjianto-xml/Knowledge-Graph-Geometry

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, AI Architect

Related on AIssential

• A Relation Is a Rotation On a Knowledge Graph — Knowledge graph relations can be modeled as rotations in high-dimensional space, scoring fact plausibility by geodesic distance.
• TransE: Translation as a Decoder — TransE models relations as simple vector translations, but this geometric rigidity causes entity collapse for many-to-many and symmetric relations.
• Multi-hop Relational Contrastive Learning: Extending Spatial Contrastive Pre-training Beyond Pairwise Relations — Multi-hop relational contrastive learning captures complex spatial dependencies beyond pairwise relations, improving scene understanding.
• When Predicates Lie: SHACL, Reification, and the Event-First Ontology — Complex domains are records of events and their conditions, not just entities with properties.
• From Relational Database to Knowledge Graph: A Three-Pass Migration That Actually Works — Existing relational databases and data catalogs are foundational drafts for knowledge graph development.
• Generating Graph-Like Logical Rules for Knowledge Graph Reasoning via Diffusion Models — Graph-like logical rules for Knowledge Graph reasoning can be efficiently generated by combining discrete diffusion models with reinforcement learning.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Agus’s Substack.