A Relation Is an Operation

2026-01-11 · Source: Agus’s Substack · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Intermediate, medium

Summary

The article introduces knowledge graphs as structured memory for AI, contrasting them with ungrounded language models. It proposes a geometric interpretation where entities are points in space and relations are operations (e.g., translations, rotations) that move from a head entity to a tail. Link prediction, the task of inferring missing facts, is reframed as a geometric problem where a fact's plausibility is measured by the distance between the transformed head and the tail. The analysis highlights that different relation types, such as symmetric "spouse" or many-to-many "parentOf", necessitate distinct geometric operations. It also details critical measurement protocols for link prediction, including filtered ranking and average tie handling, to ensure reliable model evaluation. A `DistMult` model, trained with `dim=16` for `300` epochs at `lr=0.1` on a small family graph, demonstrates this concept by accurately predicting children for a parent.

Key takeaway

For AI Scientists and Machine Learning Engineers developing grounded AI systems, understanding knowledge graph embeddings as geometric operations is crucial. This perspective reframes link prediction as a spatial problem, guiding your choice of embedding models based on relation characteristics like symmetry or cardinality. Implement filtered ranking and average tie handling in your evaluation metrics to ensure robust and reproducible results, avoiding common pitfalls in model assessment.

Key insights

Knowledge graphs represent facts as geometric operations, enabling link prediction through spatial transformations.

Principles

• Relations are operations, not just labels.
• Different relation types demand distinct geometric operations.
• Reliable link prediction requires filtered ranking and average tie handling.

Method

Place entities as vectors in space; define relations as operations Fᵣ such that `s(h,r,t)=−‖Fr(h)−t‖`. Score candidate facts by distance, ranking them to predict missing links.

In practice

• Model "spouse" relations with half-turn rotations.
• Use `DistMult` for simple knowledge graph embeddings.
• Implement filtered protocol for link prediction evaluation.

Topics

• Knowledge Graphs
• Link Prediction
• Geometric Embeddings
• Relation Operations
• DistMult
• Evaluation Metrics

Code references

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

Best for: AI Scientist, Machine Learning Engineer, AI Student

Related on AIssential

• Leveraging Graph Structure in Seq2Seq Models for Knowledge Graph Link Prediction — Integrating graph structure with Seq2Seq models significantly improves knowledge graph link prediction.
• Compact Geometric Representations of Hierarchies — Compact geometric embeddings for hierarchical data achieve exact reachability with dimensions dependent on graph structure, outperforming prior methods.
• 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.
• Graph Neural Networks: Just Averaging Neighbors — Graph Neural Networks iteratively average neighbor features to create node vectors summarizing local graph structure.
• Model Graph Inductive Learning for Knowledge Graph Completion — MGIL improves knowledge graph link prediction by leveraging global structural patterns through a novel model graph and GNN approach.
• LLM Friendly Graphs for AI Memory — From Triples to Facts — Larger, context-rich labels and subgraph representations enhance LLM comprehension of complex facts in knowledge graphs.

Open in AIssential →

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