Robust Link Prediction under Bilateral Edge Noise via Robust Graph Information Bottleneck (RGIB)

Link prediction is a fundamental graph learning task that aims to infer whether two nodes in a graph should be connected. While Graph Neural Networks (GNNs) have significantly improved performance, their robustness degrades sharply when the graph structure is corrupted by bilateral edge noise, which simultaneously perturbs the input topology and the target edge labels.

Existing noise‑robust methods mainly address node‑label noise or random edge removal, and they provide only marginal gains for link prediction under bilateral noise. The key difficulty is that the noise affects both inputs and outputs, creating a two‑way disturbance that is common in real‑world scenarios such as click‑through‑rate prediction and recommendation.

To tackle this challenge, the authors propose a new learning framework called Robust Graph Information Bottleneck (RGIB). RGIB extends the classic Graph Information Bottleneck (GIB) by explicitly decoupling and balancing the mutual information among graph topology, edge labels, and learned representations, thereby extracting reliable supervisory signals and preventing representation collapse.

Two concrete instantiations are introduced: RGIB‑SSL, which incorporates self‑supervised regularization (uniformity and alignment) to encourage informative and well‑distributed embeddings; and RGIB‑REP, which re‑parameterizes the noisy topology and label spaces to separate clean signals from noise via latent variables.

Extensive experiments on six benchmark datasets under varying noise ratios demonstrate that RGIB consistently outperforms state‑of‑the‑art baselines, achieving up to 12.9% AUC improvement on Cora and Citeseer. Ablation studies confirm that RGIB mitigates both input and label noise, restores representation uniformity, and remains effective across single‑side and bilateral noise settings.

The RGIB‑SSL variant has been deployed in Alibaba’s display advertising system. By constructing user‑ad interaction graphs and applying RGIB constraints during pre‑training, the method improves click‑through‑rate estimation robustness, leading to an estimated 5% increase in overall platform revenue.

The paper concludes by suggesting extensions of RGIB to node classification, graph classification, and knowledge‑graph reasoning, and highlights the importance of addressing both structural and feature noise in future work.