The code will coming soon!

Heterogeneous Diffusion Graph Learning

In recent years, research on graph-structured data modeling based on Graph Neural Networks (GNNs) has achieved significant breakthroughs. However, traditional GNNs are primarily designed for homogenous graphs or specific bipartite graphs, limiting their ability to accommodate the rich heterogeneity in structural information. While recent studies have suggested capturing dependencies between different types of heterogeneous interactions, few have explored two critical challenges: 1) The presence of noisy data in heterogeneous structures, which can contaminate the learning of embeddings and adversely affect the performance of graph learning task; 2) The sparse supervisory signals encountered in specific predictive tasks, such as recommendation tasks. To address these challenges, we propose a novel framework, termed Heterogeneous Graph Diffusion Learning (HDL). Our framework incorporates a simple yet effective latent space diffusion mechanism aimed at eliminating noise in the representation space of compressed and dense multi-type heterogeneous graph relations. We conducted extensive experiments on real-world data to evaluate the efficacy of our proposed framework. The results demonstrate that HDL exhibits superiority in both recommendation tasks and node classification tasks.

• python=3.8
• torch=1.12.1
• numpy=1.23.1
• scipy=1.9.1
• dgl=1.0.2+cu113