On Understanding Denoising Capability in Hypergraph Representation Learning

Detian Zhang, Chengqiang Zhang, Jingsong Lv, Li Qing, Chunjiang Zhu

Keywords: Graph Learning; Noise; Denoising Capability; Hypergraph Neural Networks

Abstract:

Hypergraphs are a powerful model for high-order relations and group interactions among entities. While many real-world network instances modeled by hypergraphs, e.g., social networks, brain connectome networks, and online question-answering communities, are rich in noise and error-prone, existing hypergraph representation learning methods often assume that hypergraphs contain limited or no noise. We reveal that even a small amount of Gaussian noise can deteriorate the performance in node classification and hyperedge prediction.
In this paper, we study the problem of alleviating the impact of noises present in node features on hypergraph representation learning. We first establish the connection between receptive fields and denoising capabilities, showing increasing receptive fields may enhance the denoising ability and robustness. We then develop a four-stage message-passing method that can increase the receptive fields within a single neural network layer, which is applicable to any existing two-stage SOTA methods. We demonstrate the increase in receptive fields both theoretically and empirically.
We have performed extensive experiments, including analysis of convergence time, an ablation study, and visualization of node embeddings to verify that our four-stage enhanced models achieve superior performance in node classification and hyperedge prediction under various noise settings.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 23234