Reproduce “BiGCN: A Bi-directional Low-Pass Filtering Graph Neural Network” on OpenResearch

About this paper

Zhixian Chen, Tengfei Ma, Zhihua Jin, Yangqiu Song, Yang Wang

BiGCN introduces a bi-directional low-pass filtering Graph Neural Network that processes node features along both the explicit node graph and a latent feature-correlation graph. This approach enhances robustness to node-wise noise, feature-wise noise, and structural errors, consistently outperforming baselines in node classification and link prediction on various datasets, particularly under noisy conditions.

Key insights

• Combining smoothing along both node and feature dimensions effectively acts as a comprehensive denoising mechanism, addressing both node-wise and feature-wise noise that traditional GCNs struggle with.
• Leveraging a latent feature-connection graph provides resilience against structural inaccuracies in the explicit node graph, reducing dependency on perfect graph topology.
• The ADMM algorithm offers a practical and efficient way to integrate bi-directional filtering operations within a GNN layer, enabling scalable application.

Summary by alphaXiv.