[Paper Review] Uncertainty-Aware Relational Graph Neural Network for Few-Shot Knowledge Graph Completion

Qian Li, Shu Guo+4 more|arXiv (Cornell University)|Mar 7, 2024

Advanced Graph Neural Networks4 citations

TL;DR

This paper proposes UFKGC, a novel uncertainty-aware few-shot knowledge graph completion framework that models entity and relation uncertainty via Gaussian distribution representations and employs an uncertainty-aware relational GNN (UR-GNN) to aggregate neighborhood information under uncertainty. By using multiple random sampling within Gaussian distributions and uncertainty optimization, the model achieves state-of-the-art performance on benchmark datasets, demonstrating robustness to noisy, limited few-shot data.

ABSTRACT

Few-shot knowledge graph completion (FKGC) aims to query the unseen facts of a relation given its few-shot reference entity pairs. The side effect of noises due to the uncertainty of entities and triples may limit the few-shot learning, but existing FKGC works neglect such uncertainty, which leads them more susceptible to limited reference samples with noises. In this paper, we propose a novel uncertainty-aware few-shot KG completion framework (UFKGC) to model uncertainty for a better understanding of the limited data by learning representations under Gaussian distribution. Uncertainty representation is first designed for estimating the uncertainty scope of the entity pairs after transferring feature representations into a Gaussian distribution. Further, to better integrate the neighbors with uncertainty characteristics for entity features, we design an uncertainty-aware relational graph neural network (UR-GNN) to conduct convolution operations between the Gaussian distributions. Then, multiple random samplings are conducted for reference triples within the Gaussian distribution to generate smooth reference representations during the optimization. The final completion score for each query instance is measured by the designed uncertainty optimization to make our approach more robust to the noises in few-shot scenarios. Experimental results show that our approach achieves excellent performance on two benchmark datasets compared to its competitors.

Motivation & Objective

• Address the challenge of noise and uncertainty in few-shot knowledge graph completion (FKGC), where limited reference triples lead to unreliable representations.
• Overcome the limitation of existing FKGC methods that treat entities as deterministic vectors, ignoring inherent uncertainty in sparse data.
• Improve generalization in low-resource scenarios by modeling uncertainty explicitly through probabilistic representations.
• Develop a unified framework that integrates uncertainty modeling, uncertainty-aware message passing, and uncertainty-optimized scoring for improved robustness.

Proposed method

• Transform entity and triple representations from deterministic vectors into Gaussian distributions, where the variance encodes uncertainty.
• Design an uncertainty representation module to estimate the uncertainty scope of entity pairs based on their feature distributions.
• Propose an uncertainty-aware relational GNN (UR-GNN) that performs convolution operations between Gaussian-distributed node features, preserving uncertainty during aggregation.
• Apply multiple random sampling from the predicted Gaussian distributions to generate smooth, diverse reference representations during optimization.
• Introduce uncertainty optimization to measure the final completion score, ensuring robustness to noisy or uncertain reference triples.
• Maintain the original data distribution while enhancing generalization through stochastic sampling within the uncertainty space.

Experimental results

Research questions

• RQ1 How can uncertainty in few-shot knowledge graph completion be effectively modeled to improve robustness against noisy, limited reference triples?
• RQ2 Can uncertainty-aware message passing in graph neural networks enhance representation learning when training data is scarce?
• RQ3 What is the optimal number of random samples from the Gaussian distribution to balance performance and noise in uncertainty optimization?
• RQ4 How does the number of UR-GNN layers affect performance and over-smoothing in few-shot KG completion?

Key findings

• The proposed UFKGC framework achieves state-of-the-art performance on two benchmark datasets (NELL and Wiki) across multiple few-shot settings, outperforming existing SOTA methods.
• Using 10 random samples per Gaussian distribution yields optimal performance on the NELL dataset, while 8 samples are optimal for Wiki, indicating a trade-off between variance and noise.
• A three-layer UR-GNN configuration provides the best balance between accessing higher-order neighborhood information and avoiding over-smoothing.
• Increasing the number of random samples improves performance up to a threshold, after which performance slightly declines due to introduced noise.
• The uncertainty-aware framework demonstrates superior robustness to noisy reference triples, especially in low-shot scenarios with high uncertainty.
• The uncertainty optimization mechanism enables consistent knowledge learning from smooth, sampled representations without distorting the original data distribution.