Smooth Diffusion Models

This is the code for the submitted paper "[SDMG: Smoothing Your Diffusion Models for Powerful Graph Representation Learning]".

SDMG bridges the gap between generative and representation learning by aligning low-frequency reconstruction, enhancing performance across graph tasks. We evaluated SDMG on 11 public graph datasets.

Datasets

These datasets will be downloaded automatically when executing the code.

Graph classification datasets

• IMDB-B
• IMDB-M
• COLLAB
• PROTEINS
• MUTAG

Node classification datasets

• Cora
• Citeseer
• PubMed
• Ogbn-arxiv
• Amazon-Computer
• Amazon-Photo

Framework

Usage

conda create -n SDMG python=3.10
conda activate SDMG
pip install -r requirements.txt

Change directory to the specific example folder ( MUTAG for a graph-level experiment )

cd GraphExp
python main_graph.py --yaml_dir ./yamls/MUTAG.yaml

Investigation

Classification accuracy across varying fractions of low-frequency components reconstruction on two datasets.

Results

Node classification performance (%)

Category	Method	Cora	CiteSeer	PubMed	Ogbn-arxiv	Computer	Photo
Supervised	GCN	81.5 ±0.5	70.3 ±0.6	79.0 ±0.4	71.7 ±3.0	86.5 ±0.5	92.4 ±0.2
	GAT	83.0 ±0.7	72.5 ±0.5	79.0 ±0.3	72.1 ±0.1	86.9 ±0.3	92.6 ±0.4
Random Walk	node2vec	74.8	52.3	80.3	-	84.39	89.67
	DeepWalk	75.7	50.5	80.5	-	85.68	89.44
Self-Supervised	DGI	82.3 ±0.6	71.8 ±0.7	76.8 ±0.6	70.3 ±0.2	84.0 ±0.5	91.6 ±0.2
	MVGRL	83.5 ±0.6	73.3 ±0.5	80.1 ±0.7	70.3 ±0.5	87.5 ±0.1	91.7 ±0.1
	BGRL	82.7 ±0.6	71.1 ±0.8	79.6 ±0.5	71.6 ±0.1	89.7 ±0.3	92.9 ±0.3
	InfoGCL	83.5 ±0.3	73.5 ±0.4	79.1 ±0.2	71.2 ±0.2	88.7 ±0.4	93.1 ±0.1
	CCA-SSG	84.0 ±0.4	73.1 ±0.3	81.0 ±0.4	71.2 ±0.2	88.7 ±0.3	93.1 ±0.1
	GPT-GNN	80.1 ±1.0	68.4 ±1.6	76.3 ±0.8	-	-	-
	GraphMAE	84.2 ±0.4	73.4 ±0.4	81.1 ±0.4	71.8 ±0.2	88.6 ±0.2	93.6 ±0.2
	GraphTCM	81.5 ±0.5	72.8 ±0.6	77.2 ±0.5	54.7 ±0.2	84.9 ±0.3	92.1 ±0.2
	VGAE	76.3 ±0.2	66.8 ±0.2	75.8 ±0.4	66.4 ±0.2	85.8 ±0.3	91.5 ±0.2
	SP-GCL	83.2 ±0.1	71.9 ±0.4	79.2 ±0.7	68.3 ±0.2	89.7 ±0.2	92.5 ±0.3
	GraphACL	84.2 ±0.3	73.6 ±0.2	82.0 ±0.2	71.7 ±0.3	89.8 ±0.3	93.3 ±0.2
	DSSL	83.5 ±0.4	73.2 ±0.5	81.3 ±0.3	69.9 ±0.4	89.2 ±0.2	93.1 ±0.3
	DDM	83.1 ±0.3	72.1 ±0.4	79.6 ±0.9	71.3 ±0.3	89.8 ±0.2	93.8 ±0.2
Our	SDMG	84.3 ±0.5	73.9 ±0.4	80.0 ±0.5	72.1 ±0.3	91.6 ±0.2	94.7 ±0.2

Graph classification performance (%)

Category	Method	IMDB-B	IMDB-M	PROTEINS	COLLAB	MUTAG
Supervised	GIN	75.1 ±5.1	52.3 ±2.8	76.2 ±2.8	80.2 ±1.9	89.4 ±5.6
	DiffPool	72.6 ±3.9	-	75.1 ±3.5	78.9 ±2.3	85.0 ±10.3
Random Walk	node2vec	50.20 ±0.90	36.0 ±0.70	57.49 ±3.57	-	72.63 ±10.20
	Sub2Vec	55.26 ±1.54	36.70 ±0.80	53.03 ±5.55	-	61.05 ±15.80
	graph2vec	71.10 ±0.54	50.44 ±0.87	73.30 ±0.05	-	83.15 ±9.25
Self-supervised	InfoGraph	73.03 ±0.87	49.69 ±0.53	74.44 ±0.31	70.65 ±1.13	89.01 ±1.13
	GraphCL	71.14 ±0.44	48.58 ±0.67	74.39 ±0.45	71.36 ±1.15	86.80 ±1.34
	JOAO	70.21 ±3.08	49.20 ±0.77	74.55 ±0.41	69.50 ±0.36	87.35 ±1.02
	GCC	72.0	49.4	-	78.9	-
	MVGRL	74.20 ±0.70	51.20 ±0.50	-	-	89.70 ±1.10
	GraphMAE	75.52 ±0.66	51.63 ±0.52	75.30 ±0.39	80.32 ±0.46	88.19 ±1.26
	InfoGCL	75.10 ±0.90	51.40 ±0.80	-	80.00 ±1.30	91.20 ±1.30
	SimGRACE	71.30 ±0.77	-	75.35 ±0.09	71.72 ±0.82	89.01 ±1.31
	DDM	74.05 ±0.17	52.02 ±0.29	71.61 ±0.56	80.70 ±0.18	90.15 ±0.46
Our	SDMG	76.03 ±0.53	52.5 ±0.42	73.17 ±0.16	82.23 ±0.35	91.58 ±0.28