Noise-Resilient Unsupervised Graph Representation Learning via Multi-Hop Feature Quality Estimation

What is it about?

This paper explores the impact of noise features on unsupervised graph representation learning. The analysis reveals the “double-edged sword” nature of propagation in dealing with noisy features, and a new method, MQE, is designed to make full use of the advantages of propagation while mitigating the disadvantages of propagation. Experimental results demonstrate the effectiveness of MQE.

Featured Image

Why is it important?

In this paper, we undertake the first endeavor in unsupervised graph representation learning (URGL) on graph data with noisy features — a challenge that remains largely unexplored in real-world scenarios, despite its inevitability. Through empirical analysis, we reveal the strengths and weaknesses of message propagation in solving the noise feature problem, and recognize the importance of assessing the quality of propagated information. Based on these insights, we propose a simple and efficient, called MQE, which learns reliable node representations by estimating the quality of multi-hop propagation features using conditional Gaussian models.