IGNP applies epidemic SIR model to boost multi-relational graph link prediction

Link prediction in multi-relational graphs—where edges have types and meanings—is critical for applications like social network analysis and knowledge graph completion. Traditional methods often rely on local features, but capturing global context is essential for accuracy. A new paper, accepted in Pattern Recognition and released on arXiv, tackles this by framing link existence and type as a form of node influence. The authors—Zidu Yin, Yuankai Qi, Dong Gong, Ehsan Abbasnejad, Kun Yue, and Javen Qinfeng Shi—argue that whether influence propagates between a node pair, and what type of influence travels, directly indicates whether and what kind of edge exists.

To operationalize this, they extend the Susceptible-Infectious-Recovered (SIR) epidemic model to simulate influence propagation across large-scale graphs via sub-graph structures. These sub-graphs are then compressed using virtual edges, drastically reducing the computational cost of leveraging global graph information. The result is IGNP (Influential Graph Neural Predictor), a framework that integrates influence propagation signals into a graph neural network. Extensive experiments on widely used real-world datasets demonstrate that IGNP outperforms strong baselines by a large margin, confirming that epidemic-inspired propagation provides a powerful inductive bias for relational prediction. This work opens a new direction for combining epidemiological models with graph deep learning for scalable, accurate link prediction.