Normalizing flow-enhanced Gaussian embedding for few-shot knowledge graph completion

Abstract

Objectives:

Few-shot knowledge graph completion (FKGC) aims to infer missing facts of the query triples based on few-shot reference entity pairs. However, existing FKGC approaches often overlook the inherent uncertainty of relations in KGs, as deterministic semantic representations derived from sparse samples may be unreliable. Meanwhile, they neglect both noisy neighbor aggregation and inter-neighbor interactions, as well as the handling of complex relations, which largely limits model performance. This paper aims to overcome these limitations and enhance FKGC performance.

Methods:

This paper proposes a method incorporated normalizing flow with Gaussian Network for FKGC, namely NFGN. Specifically, we combine normalizing flow-enhanced Gaussian distribution to model the few-shot settings and multi-semantics uncertainty of relations, which learns the uncertain semantics of entity features based on limited data. Then, we introduce the GD-TransE decoder, which incorporates relation uncertainty to handle complex relations. To improve the model’s effectiveness, a gated neighbor encoder is designed to model semantic interactions among neighbors, and control the activation of noisy neighbors through gating thresholds.

Novelty:

This paper presents the first study that integrates normalizing flows with Gaussian embeddings for FKGC, offering a more robust representation of uncertainty in relations. The proposed method further introduces the gated neighbor encoder and GD-TransE decoder to handle neighborhood noise and complex relationships, thereby overcoming the limitations of existing FKGC methods.

Findings:

Extensive experiments conducted on three diverse benchmark datasets demonstrate that our method significantly outperforms state-of-the-art performance, achieving improvements of 5.7%, 2.4%, 10.6%, and 6.7% in MRR, Hits@1, Hits@5, and Hits@10, respectively.