Rethinking Generalization in Knowledge Graph Embedding: An Empirical Study and Sparse Peak Regularization
The study investigates regularization for Knowledge Graph Embedding (KGE) models across various architectures, finding that overfitting is common. The authors propose Sparse Peak Regularization (SPR), a model-agnostic method that selectively penalizes only the most important embedding components, leading to better generalization and superior performance compared to existing regularizers.