Is Architectural Complexity Overrated? Competitive and Interpretable Knowledge Graph Completion with RelatE
Abhijit Chakraborty, et al.
RelatE decomposes knowledge graph embeddings into modulus and phase components, preserving predictive performance while providing transparent semantics. The paper demonstrates that lightweight architectures can rival deeper models on benchmark datasets.
Abstract
We introduce RelatE, a knowledge graph embedding model with separated modulus and phase terms. This design yields interpretable factors for relation strength and direction while maintaining competitive link prediction accuracy. Experiments on FB15k-237, WN18RR, and real enterprise graphs show RelatE outperforms heavier models when interpretability constraints are applied.
Key Contributions
- Factorized embedding representation enabling direct inspection of relation importance and phase alignment.
- Demonstrates competitive performance with state-of-the-art models while reducing architectural complexity.
- Includes open-source implementation and evaluation pipeline for reproducibility.