Abstract
In this paper, we generalize single-relation pseudo-Riemannian graph embedding models to multi-relational networks. We demonstrate that the standard approach of encoding relations as manifold transformations translates effectively from the Riemannian to the pseudo-Riemannian case. Furthermore, we introduce a novel view of relations as separate spacetime submanifolds within multi-time manifolds and propose an interpolation between pseudo-Riemannian models and their Wick-rotated Riemannian counterparts. We validate these extensions through link prediction tasks on flat Lorentzian manifolds, demonstrating their efficacy in knowledge graph completion and biological knowledge discovery.
Introduction
Knowledge Graphs (KGs) represent entities and their diverse relationships as nodes and edges, serving as a foundation for applications ranging from semantic search to drug discovery. Traditionally, these graphs are embedded into Euclidean space. However, Euclidean space is inherently metric, meaning it adheres to strict geometric constraints like the triangle inequality.
Real-world relational data often violates these constraints. For instance, directed hierarchies and non-transitive relationships do not map cleanly to metric spaces. While Hyperbolic geometry has gained traction for modeling hierarchical structures, it remains a Riemannian (metric) space. This research explores pseudo-Riemannian embeddings—specifically Minkowski space—which utilize an indefinite metric. By incorporating "time" dimensions and "lightcone" structures, these models naturally capture directed, non-metric relationships. This paper extends these concepts from single-relation graphs to complex, multi-relational KGs.
The PseudoE Framework
The authors introduce PseudoE, a framework designed to scale pseudo-Riemannian embeddings to multi-relational environments using three primary geometric strategies:
1. Relations as Endomorphisms
Following the logic of established models like TransE, relations can be modeled as transformations (such as translations or scaling) that shift node embeddings within the manifold. This allows the model to represent how different relations alter the relative positioning of entities in the "spacetime" of the graph.
2. Relations as Spacetime Submanifolds
In a more novel approach, the authors define a manifold with multiple time dimensions. Each specific relation is treated as a projection onto a unique 1-time submanifold. In this view, different relations essentially "perceive" different versions of spacetime, allowing the model to isolate the specific temporal/directional dynamics of each relation type.
3. Wick-Rotation Interpolation
To accommodate relations that exhibit metric-like behavior (e.g., symmetric similarity), the framework introduces a mixing parameter ($\beta$). This allows for a "Wick-rotation," smoothly interpolating between a pseudo-Riemannian lightcone structure and a standard Riemannian (Euclidean) isotropic structure. This flexibility ensures the model can adapt to both directed and undirected relations within the same graph.
4. Learned Biases and Cones of Influence
The model incorporates node-specific and relation-specific bias terms. In pseudo-Riemannian geometry, these biases function as "cones of influence," defining the reach of a node. A node with a larger bias has a broader lightcone, indicating a higher probability of connection to other entities.
Experimental Evaluation
The PseudoE model was evaluated on three benchmark datasets:
- FB15K-237: General world facts from Freebase.
- WN18RR: Lexical relations from WordNet.
- Hetionet (small): A biological knowledge graph linking compounds, diseases, and genes.
Key Results
PseudoE demonstrated competitive performance, frequently ranking among the top models for link prediction across all datasets. Notably, on the Hetionet dataset, the version of PseudoE combining both endomorphism and submanifold strategies outperformed all baselines, achieving a Mean Reciprocal Rank (MRR) of 0.544 compared to 0.527 for the next best model.
Knowledge Discovery and De-biasing
A significant finding of this research is the correlation between node biases and node degree (popularity). In biological contexts, high-degree nodes often represent well-studied entities (e.g., common genes like TNF). By adjusting these learned biases at test time, the researchers demonstrated that the model could be "de-biased" to uncover novel, understudied connections. In a case study on breast cancer, this approach shifted predictions from common targets to biologically plausible but less-studied candidates like DTX3, highlighting the model's utility in scientific discovery.
Conclusion and Future Work
This work successfully extends pseudo-Riemannian embeddings to multi-relational contexts, providing a robust framework for modeling non-metric data. While this study focused on flat Minkowski space, future research will explore curved pseudo-Riemannian manifolds (such as Anti-de Sitter or de Sitter spaces) to capture even more complex hierarchical structures. Additionally, exploring non-linear projections for the spacetime submanifold approach may further enhance performance on graphs with high relational complexity.
Summary updated