A Scalable Inter-edge Correlation Modeling in CopulaGNN for Link Sign Prediction

A research team led by Jinkyu Sung has published a significant advancement in graph neural networks with their paper 'A Scalable Inter-edge Correlation Modeling in CopulaGNN for Link Sign Prediction,' accepted for ICLR 2026. The work tackles the challenging problem of link sign prediction—determining whether relationships in networks (like social media connections) are positive or negative. Traditional graph methods struggle with negative edges because they violate the homophily assumption that connected nodes are similar. The researchers extend the 2021 CopulaGNN framework by directly modeling statistical dependencies between edges using Gaussian copula theory, but they faced a major scalability hurdle: naive modeling of edge-edge relations becomes computationally impossible even for moderately sized graphs.

The breakthrough comes from two clever optimizations that make the method practical. First, they represent the massive correlation matrix as a Gramian of lower-dimensional edge embeddings, drastically reducing parameters. Second, they mathematically reformulate the conditional probability distribution to slash inference costs. The team provides theoretical proof of their method's linear convergence rate. In extensive experiments, their model demonstrates significantly faster convergence than existing baselines while maintaining prediction performance competitive with state-of-the-art models. This work opens the door to applying sophisticated correlation-aware GNNs to real-world, large-scale signed networks like online review platforms, political affiliation networks, or trust/distrust systems where relationship polarity is crucial.