CORE: Data Augmentation for Link Prediction via Information Bottleneck

Key findings from this study

• The study found that CORE simultaneously recovers missing edges and removes noise from graph structures to enhance link prediction performance.
• The authors report that the method outperforms state-of-the-art approaches across multiple benchmark datasets.
• The researchers demonstrate that Information Bottleneck principles can be operationalized through data augmentation to improve model robustness in link prediction tasks.

Overview

Link prediction represents a fundamental task in graph representation learning with applications across multiple domains. Current models suffer from compromised generalizability due to noisy or spurious graph information and inherent incompleteness of graph data. The authors introduce CORE, a data augmentation method grounded in Information Bottleneck principles. CORE simultaneously recovers missing edges while removing noise from graph structures to enhance model robustness and predictive performance.

Methods and approach

CORE operates as a data augmentation framework designed to learn compact and predictive augmentations for link prediction tasks. The method integrates the Information Bottleneck principle to balance information preservation with noise reduction. By completing missing edges and reducing spurious structural information, CORE produces augmented graphs that improve downstream model performance. The approach was evaluated across multiple benchmark datasets against state-of-the-art baseline methods.

Results

Extensive experiments on benchmark datasets demonstrate CORE's superiority over existing methods for link prediction. The augmentation approach effectively enhances model robustness when applied to graphs containing noisy or spurious information. CORE consistently outperforms state-of-the-art approaches, establishing its utility as a leading technique for robust link prediction in graph representation learning contexts.

Implications

The work addresses a critical limitation in graph representation learning: the vulnerability of link prediction models to graph incompleteness and noise. By leveraging Information Bottleneck principles, CORE provides a theoretically grounded approach to mitigating these issues through strategic data augmentation. This contribution extends the methodological toolkit for practitioners working with incomplete or noisy graph data across domains including social networks, knowledge graphs, and recommendation systems.

Scope and limitations

This summary is based on the study abstract and available metadata. It does not include a full analysis of the complete paper, supplementary materials, or underlying datasets unless explicitly stated. Findings should be interpreted in the context of the original publication.