A Method Integrating Enhanced Hinge Loss Function for Few-Shot Knowledge Graph Completion

Yingqi Guo,Ying Zhao,Bo Wang

DOI: https://doi.org/10.1145/3672758.3672845

2024-01-01

Abstract:Few-shot knowledge graph completion (FKGC) is a fundamental task to supply missing triples for knowledge graphs. Many recently proposed few-shot relational learning methods exhibit excellent performance. However, they rely solely on the angle-based scoring function that is valid only for some relations. The distance-based scoring function has difficulty adapting to embedding distribution where positive candidates follow a loose distribution embracing the negatives, which is common in the FKGC dataset. Here, we propose a novel approach integrating enhanced hinge loss function for FKGC. It is based on an adaptive Ensemble Learning FrameworK, namely ELFK, which contains two base modules with different types of scoring functions to improve the model's generalization across different relations. Specially, we utilize a modified margin hinge loss function by incorporating an upper bound loss function for the base module with an Euclidean-distance-based scoring function to improve the embedding distribution and reduce the chance that positive candidates are distributed outside the negatives. In addition, two base module are adaptive integrated by a trainable weight learned in an ensemble module. Experiments on the NELL and wiki datasets demonstrate that our approach achieves state-of-the-art performance.