5* Knowledge Graph Embeddings with Projective Transformations

Mojtaba Nayyeri, Sahar Vahdati, Can Aykul, Jens Lehmann

Introduction

Knowledge Graph Embeddings (KGE) are fundamental for link prediction, mapping entities and relations into a continuous vector space. While traditional models rely on Euclidean transformations like translation (TransE) or rotation (RotatE), they often struggle with the diverse topological motifs found in real-world data. This paper introduces $5^*E$, a novel KGE model that leverages projective geometry and Möbius transformations to provide a more flexible and expressive embedding framework.

The Challenge of Geometric Rigidity

Most existing KGE models are "stiff," constrained by a single type of geometric transformation. For instance, a "hypernym" relation typically forms a hierarchical path (a line), whereas a "similar_to" relation might form a cluster or loop (a circle). Standard Euclidean models cannot mathematically map a line-like structure onto a circle-like structure without losing information or collapsing the data.

Figure 1: Comparison of geometric mappings across different KGE models.

As shown in Figure 1, $5^*E$ overcomes this by utilizing projective transformations that can seamlessly transition between paths and loops, allowing it to capture complex graph patterns that traditional models fail to represent.

Methodology: The Möbius Transformation

The core of the $5^*E$ model is the Möbius transformation, a fractional linear transformation defined on the complex projective line: $$\vartheta(x) = \frac{ax + b}{cx + d}$$ By varying the parameters $(a, b, c, d)$, this single formula can perform five distinct operations—the "Five Stars": translation, rotation, homothety (scaling), inversion, and reflection.

Model Formulation

Entities are embedded into a $d$-dimensional complex projective space ($\mathbb{CP}^d$). Each relation is represented by $2 \times 2$ complex matrices that define the Möbius transformation for each dimension. The score function measures the plausibility of a triple $(h, r, t)$ by applying the relation's transformation to the head entity $h$ and calculating its similarity to the tail entity $t$ using a complex inner product.

Key Insights: Beyond Euclidean Constraints

$5^*E$ serves as a mathematical generalization of several prominent KGE models. By constraining its parameters, it reduces to TransE, RotatE, ComplEx, or DistMult. Furthermore, Möbius transformations are conformal, meaning they preserve local angles. This property is crucial for maintaining the semantic clustering of entities during transformation. The authors also provide a theoretical proof that $5^*E$ is "fully expressive," capable of representing any set of ground-truth triples given sufficient dimensionality.

Results and Discussion

The model was evaluated on standard benchmarks including FB15k-237, WN18RR, and NELL. $5^*E$ achieved state-of-the-art performance, particularly on WN18RR with an MRR of 0.50. It also demonstrated superior robustness to hierarchical density in the NELL datasets compared to specialized models like MuRP.

Figure 5: Visualizing the warping of the complex plane for hypernym and hyponym relations.

Visualizations of the model's behavior (Figure 5) illustrate how it warps the underlying grid to fit the graph's topology. Unlike models that only rotate or translate, $5^*E$ can bend, stretch, and invert the space, providing a much richer representational capacity for complex relations like hypernyms and their inverses.

Conclusion

By moving from Euclidean to projective geometry, $5^*E$ introduces a highly flexible framework for Knowledge Graph Embeddings. Its ability to handle multiple geometric transformations simultaneously allows it to capture the intricate structural variety of large-scale knowledge graphs more effectively than previous methods.

Summary updated

Abstract

Performing link prediction using knowledge graph embedding models has become a popular approach for knowledge graph completion. Such models employ a transformation function that maps nodes via edges into a vector space in order to measure the likelihood of the links. While mapping the individual nodes, the structure of subgraphs is also transformed. Most of the embedding models designed in Euclidean geometry usually support a single transformation type - often translation or rotation, which is suitable for learning on graphs with small differences in neighboring subgraphs. However, multi-relational knowledge graphs often include multiple sub-graph structures in a neighborhood (e.g. combinations of path and loop structures), which current embedding models do not capture well. To tackle this problem, we propose a novel KGE model (5*E) in projective geometry, which supports multiple simultaneous transformations - specifically inversion, reflection, translation, rotation, and homothety. The model has several favorable theoretical properties and subsumes the existing approaches. It outperforms them on the most widely used link prediction benchmarks
BibTeX
@inproceedings{nayyeri2021knowledge,
  author = {Mojtaba Nayyeri and Sahar Vahdati and Can Aykul and Jens Lehmann},
  title = {{5* Knowledge Graph Embeddings with Projective Transformations}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/aaai.v35i10.17095},
}
Figure