GIDN: A Lightweight Graph Inception Diffusion Network for High-efficient Link Prediction

TL;DR

This work tackles efficient link prediction on large knowledge graphs by introducing GIDN, a diffusion-based model that avoids deep, compute-heavy architectures. It generalizes graph diffusion across multiple feature spaces and uses an Inception module to capture multi-scale proximity signals with reduced cost. On the ogbl-collab dataset, GIDN achieves 0.7096 ± 0.0055 Hits@50, surpassing AGDN and PLNLP baselines and demonstrating both high accuracy and computational efficiency. Overall, GIDN shows that lightweight, diffusion-oriented architectures can deliver strong link-prediction performance on large-scale graphs.

Abstract

In this paper, we propose a Graph Inception Diffusion Networks(GIDN) model. This model generalizes graph diffusion in different feature spaces, and uses the inception module to avoid the large amount of computations caused by complex network structures. We evaluate GIDN model on Open Graph Benchmark(OGB) datasets, reached an 11% higher performance than AGDN on ogbl-collab dataset.