WSDM'2025: DiffGraph: Heterogeneous Graph Diffusion Model

Framework

Abstract

Recent advances in Graph Neural Networks (GNNs) have signifi- cantly improved modeling of graph-structured data. However, tradi- tional GNNs face challenges in dealing with complex heterogeneous structures commonly found in real-world applications. While re- cent studies have addressed dependencies among heterogeneous interactions, two major challenges persist: 1) Noisy data within heterogeneous structures can impair the learning of embeddings and negatively affect graph learning tasks; 2) Existing methods fail to capture complex semantic transitions among heterogeneous re- lations, impacting downstream predictions. To address these issues, we introduce a novel framework, Heterogeneous Graph Diffusion Model (DiffGraph), which incorporates a cross-view denoising strat- egy. This strategy effectively transforms auxiliary heterogeneous data into the target semantic space to distill task-relevant infor- mation. Our approach features a latent heterogeneous graph dif- fusion mechanism, which manages noise through an innovative forward and backward diffusion process. This method simulta- neously achieves heterogeneous graph denoising and cross-type transition, and also eases the challenges of graph generation by leveraging its latent-space diffusion process. We validated our pro- posed framework through comprehensive experiments on both public and industrial datasets. The evaluation results demonstrate that DiffGraph outperforms existing methods in both link predic- tion and node classification tasks, showcasing its robustness and efficiency in processing heterogeneous graphs.

Environment

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

Code Structures

The folder DiffGraph-Rec presents the code and datasets for link prediction(Recommendation), while DiffGraph_NC contains the code and datasets for the node classification task.

.
├──DiffGraph-Rec
 ├── DataHandler.py
 ├── main.py
 ├── param.py
 ├── Utils                    
 │   ├── TimeLogger.py            
 │   ├── Utils.py                             
 ├── Model.py
├──DiffGraph_NC
 ├──DataHandler.py
 ├── main.py
 ├── param.py
 ├── Utils                    
 │   ├── TimeLogger.py            
 │   ├── Utils.py
 ├── Model.py
└── README

Datasets

We evaluate HGDM on both the link prediction and node classification tasks. For link prediction, we utilize three publicly available datasets collected from real-world commercial platforms: Tmall, Retailrocket, and IJCAI. For the node classifi- cation task, we use two public datasets, DBLP and AMiner, which focus on publications and academic social ties, as well as an Indus- try dataset for user classification collected from a popular game platform. Statistics of these datasets are in Table 1. Below is the detailed descriptions for the experimental datasets.

Dataset	User #	Item #	Link #	Interaction Types
Tmall	31882	31232	1,451,29	View, Favorite, Cart, Purchase
Retail Rocket	2174	30113	97,381	View, Cart, Transaction
IJCAI	17435	35920	799,368	View, Favorite, Cart, Purchase
Industry	1M	361	23,890,445	Purchase, Friend, Complete Task

	Node	Metapath		Node	Metapath
DBLP	Author:4057	APA	AMiner	paper:6564	PAP
	Paper:14328	APCPA		author:13329	PRP
	Conference:20	APTPA		Reference:35890	POS
	Term:7723