UOKGE: Bridging the Gap Between Ontological Certainty and Real-World Uncertainty in KGE

TL;DR

The paper introduces UOKGE, the first Knowledge Graph Embedding (KGE) framework designed to handle uncertain ontology-aware KGs. By mapping entities to points and classes/properties to spheres, it simultaneously captures structured RDFS semantics and their respective confidence scores. It outperforms traditional models like Complex and UKGE, demonstrating that geometry is a powerful tool for modeling "fuzzy" knowledge.

Background & Motivation: Why Point-Vectors Aren't Enough

In the traditional KGE landscape (e.g., TransE, TransR), entities are treated as isolated points in a vector space. While efficient, this approach faces two critical hurdles when dealing with modern, automated Knowledge Bases:

1. Ontological Blindness: Standard models ignore the "rules" of the graph, such as SubclassOf or Domain/Range constraints, treating all relations as flat links.
2. The Uncertainty Crisis: Real-world KGs (like ConceptNet or NELL) are probabilistic. A triple like (John, type, Loyal) might only have a confidence score of 0.5. Forcing a model to embed this as a "hard" fact introduces noise that degrades embedding quality.

The authors argue that we need a region-based approach where the "certainty" of a relation is reflected by the geometric proximity of points to spheres.

Methodology: The Geometry of Uncertainty

UOKGE represents different KG elements through specific geometric primitives:

• Entities ($e$): Points in $n$-dimensional space.
• Classes ($c$): $n$-spheres defined by a center and a radius $\rho$.
• Properties ($p$): $2n$-spheres, representing the relationship between a domain center ($p^d$) and a range center ($p^r$).

The Core Insight: The Gap Function

Instead of a simple distance metric, UOKGE defines a Gap function for various triple formats (type, subclass, domain, etc.). This function calculates how well the geometric positioning of an entity or class matches its confidence score.

Figure 1: Conceptual visualization of UOKGE where entities are points and classes are spheres. The distance between them reflects the confidence score.

For a type triple, if the confidence score $s_l$ is high, the entity point $e$ should be nestled deep within the class sphere $c$. If $s_l$ is low, the point should reside near the boundary or outside it. The training objective is a Mean Squared Error (MSE) loss that minimizes the difference between the predicted geometric gap and the ground-truth confidence score.

Experimental Results: Setting New Benchmarks

The model was evaluated on CN15K (a subset of ConceptNet).

1. Confidence Prediction

UOKGE achieved the lowest error rates among all competitors, proving it can accurately "guess" how reliable an unseen triple is.

2. Triple Classification

This task determines if a triple is "strong" (Confidence > 0.7) or "weak." UOKGE achieved a staggering 95.5% accuracy, significantly higher than the ontology-aware (but uncertainty-blind) EmbedS (91.1%).

Figure 2: Accuracy comparison across different embedding baselines.

Critical Insight & Conclusion

The success of UOKGE stems from its Inductive Bias. By using spheres, the model inherently understands containment (Subclass) and membership (Type). When uncertainty is added to the mix, the radius of the sphere acts as a soft boundary, allowing the model to be "less sure" about entities on the periphery.

Limitations: The current model is tested primarily on RDFS-level semantics. Future work will need to address more complex logical operators (AND/OR) and scale to graphs with millions of classes.

Final Takeaway: For practitioners building KGs via web-scraping or LLM-extraction, UOKGE provides a robust blueprint for keeping the "semantics" while embracing the "noise."