Graph Contrastive Learning with Cohesive Subgraph Awareness

Our research paper has been accepted to The Web Conference 2024.

arXiv: http://arxiv.org/abs/2401.17580

Catalog of This Code Repository

• Graph Contrastive Learning with Cohesive Subgraph Awareness
  • CTAug Framework
  • Requirements
  • File Directory
  • Arguments Description
    • get_sub_counts.py
    • main.py
    • Methods List
  • Parameter Setting in Our Experiments
    • get_sub_counts.py
    • main.py (--factor)
  • Running Examples

CTAug Framework

Requirements

• python 3.8
• torch 1.12.0
• torch-geometric 2.0.4
• PyGCL 0.1.2
• networkx 2.8.4
• graph-tool (only used for get_sub_counts.py, not necessary for main.py; installation refers to https://git.skewed.de/count0/graph-tool/-/wikis/installation-instructions)

File Directory

.
├─ framework.png
├─ get_sub_counts.py
├─ main.py
├─ README.md
│
├─ CTAug
│  ├─ evaluate.py
│  ├─ model.py
│  ├─ preprocess.py
│  ├─ utils.py
│  ├─ __init__.py
│  │
│  └─ methods
│     ├─ GraphCL.py
│     ├─ GraphCL_AUG.py
│     ├─ InfoGraph.py
│     ├─ JOAO.py
│     ├─ JOAOv2.py
│     ├─ JOAO_AUG.py
│     ├─ MVGRL.py
│     ├─ MVGRL_AUG.py
│     ├─ __init__.py
│     │
│     └─ node_cls
│        ├─ CTAug_node.py
│        ├─ utils_ctaug.py
│        ├─ utils_sp.py
│        ├─ __init__.py
│        ├─ param
│        └─ pGRACE
│   
├─ data
│  ├─ ENZYMES_global_cycle_graph_5.pt
│  ├─ IMDB-BINARY_global_complete_graph_5.pt
│  ├─ IMDB-MULTI_global_complete_graph_5.pt
│  ├─ PROTEINS_global_cycle_graph_5.pt
│  ├─ ENZYMES
│  ├─ IMDB-BINARY
│  ├─ IMDB-MULTI
│  └─ PROTEINS
│
└─ log
    └─ GraphCL_AUG_OGSN_sub_IMDB-BINARY
         └─ IMDB-BINARY_global_complete_graph_5_0.2.log

• get_sub_counts.py is used to count the isomorphic subgraphs, the output results will be saved under data folder. This code mainly refers to https://github.com/gbouritsas/GSN.

• main.py is the running interface of CTAug, we can use different methods and settings by passing arguments.

• CTAug package implements our CTAug method and some baseline methods used in our experiment.

  • methods includes (1) Graph classification baseline methods: GraphCL, InfoGraph, JOAO, and JOAOv2 implemented by PyGCL(https://github.com/PyGCL/PyGCL); MVGRL refers to https://github.com/kavehhassani/mvgrl. (2) Cohesive topology augmentation GCL methods: CTAug-GraphCL, CTAug-JOAO, CTAug-MVGRL. (3) Node classification baseline methods: GRACE and GCA, which are borrowed from https://github.com/CRIPAC-DIG/GCA, and (4) corresponding cohesive topology augmentation GCL method: CTAug-GRACE and CTAug-GCA.
  • evaluate.py, model.py, preprocess.py and utils.py store a series of functions and classes used for evaluating graph embedding, building contrastive learning model, preprocessing datasets, extracting cohesive subgraph, etc.

• data folder stores raw and processed data files, including graph classification datasets (IMDB-BINARY, IMDB-MULTI, ENZYMES, and PROTEINS), and corresponding substructure counts datasets (IMDB-BINARY_global_complete_graph_5.pt, IMDB-MULTI_global_complete_graph_5.pt, ENZYMES_global_cycle_graph_5.pt, and PROTEINS_global_cycle_graph_5.pt).

  NB: Due to file size limitations we don't upload other datasets and their corresponding substructure counts datasets, but these datasets can be obtained by get_sub_counts.py easily.

• log folder records experiment output, e.g., loss of each epoch, and mean and standard deviation of accuracies. When running main.py, it will create a subfolder in the log folder automatically, generate a log file and save the embedding.

  • GraphCL_AUG_OGSN_sub_IMDB-BINARY/IMDB-BINARY_global_complete_graph_5_0.2.log is an output example. It is a log record for CTAug-GraphCL method, which use IMDB-BINARY dataset, use 3,4,5-clique isomorphic subgraphs counts as node feature, and set probability decay factor at 0.2.

Arguments Description

get_sub_counts.py

Name	Default value	Description
dataset	IMDB-MULTI	Dataset name, can be chosen from: graph classification: {'IMDB-BINARY', 'IMDB-MULTI', 'COLLAB', REDDIT-BINARY', 'reddit_threads', 'ENZYMES', 'PROTEINS'}, node classification: {'Coauthor-CS', 'Coauthor-Phy', 'Amazon-Computers'}.
k	5	Count all the subgraphs of the family that have size up to k.
id_scope	global	'local' vs 'global' --> GSN-e vs GSN-v (in O-GSN the default value is 'global').
id_type	complete_graph	Subgraphs family, can be chosen from {'cycle_graph', 'path_graph', 'complete_graph', 'binomial_tree', 'star_graph', 'nonisomorphic_trees'}.
level	graph	The dataset is used for 'graph' classification or 'node' classification.

main.py

Name	Default value	Description
method	GraphCL_AUG_OGSN	Use different contrastive learning methods. Can be chosen from: graph-level CTAug: {'GraphCL_AUG_OGSN', 'JOAO_AUG_OGSN', 'MVGRL_AUG_OGSN'}, graph-level baselines: {'GraphCL', 'MVGRL', 'JOAO', 'JOAOv2', 'InfoGraph'}, node-level CTAug: {'GCA_AUG_OGSN', 'GRACE_AUG_OGSN'}, node-level baselines: {'GCA', 'GRACE'}.
dataset	IMDB-BINARY	Dataset name, can be chosen from graph classification: {'IMDB-BINARY', 'IMDB-MULTI', 'COLLAB', REDDIT-BINARY', 'reddit_threads', 'ENZYMES', 'PROTEINS'}, node classification: {'Coauthor-CS', 'Coauthor-Phy', 'Amazon-Computers'}..
feature	sub	Use input features generated by different methods, can be chosen from {'one': use vector [1] as the feature of each node, 'sub': use substructure counts as node feature (which is needed in O-GSN), 'deg': use one-hot degree as node feature, 'sub_deg': concatenate substructure counts and one-hot degree as node feature}.
dataset_file	None	Dataset file name (containing substructure counting), the dataset should be placed under 'data' folder.
seed	42	Random seed.
times	5	The number of repetitions of the experiment.
save_path	None	The name of the folder to save log. The default path is a subfolder named as method name under the folder 'log'.
device	None	Running environment, 'cpu' or 'cuda'.
pn	0.2	The probability of dropping node, removing edge, or sampling subgraph.
factor	0.8	The decay factor of dropping probability in CTAug for probabilistic GCL methods, or the factor considering cohesive property in CTAug for deterministic GCL methods.
cal_weight	node	Choose the edge weight calculation strategy from {'node', 'edge'}. Only used in CTAug-MVGRL method.
core	both	Subgraph property, can be chosen from {'kcore', 'ktruss', 'both'}. Only used in CTAug-GRACE and CTAug-GCA methods.
epoch	None	Training epoch.
interval	None	Interval epoch to test.
batch_size	64	Batch size of dataset partition.
shuffle	True	Shuffle the graphs in the dataset or not.
hid_units	128	Dimension of hidden layers and embedding.
num_layer	2	Number of GConv layers.
save_model	False	Whether to save the model or not.
save_embed	True	Whether to save the embedding or not.
eval_model	True	Whether to evaluate the embedding or not.
norm	None	Whether to normalize embedding before logistic regression test or not.

NB: If some arguments are set to "None", they will be assigned default values which are used in our experiments automatically. The specific assignment can be seen in main.py.

Methods List

We can choose different methods by passing value to the "method" argument of main.py. For different methods, the corresponding "method" argument values are as follows:

Task	Category	Method	"Method" argument value
graph classification	CTAug	CTAug-GraphCL	GraphCL_AUG_OGSN
graph classification	CTAug	CTAug-JOAO	JOAO_AUG_OGSN
graph classification	CTAug	CTAug-MVGRL	MVGRL_AUG_OGSN
graph classification	baseline	GraphCL	GraphCL
graph classification	baseline	JOAO	JOAO
graph classification	baseline	MVGRL	MVGRL
graph classification	baseline	InfoGraph	InfoGraph
graph classification	baseline	JOAOv2	JOAOv2
node classification	CTAug	CTAug-GRACE	GRACE_AUG_OGSN
node classification	CTAug	CTAug-GCA	GCA_AUG_OGSN
node classification	baseline	GRACE	GRACE
node classification	baseline	GCA	GCA

NB1: "AUG" means use module 1 (Topology Augmentation Enhancement) of CTAug, and "OGSN" means use module 2 (Graph Learning Enhancement). So if you want to only use one module of CTAug, just modify the "method" argument value. E.g. "GraphCL_AUG" for only using module 1 of CTAug-GraphCL, and "GraphCL_OGSN" for only using module 2 of CTAug-GraphCL.

NB2: If using module 2, the "feature" arguments must be set to "sub" (because isomorphic subgraphs counts is necessary to O-GSN).

Parameter Setting in Our Experiments

get_sub_counts.py

Datasets	--id_type	--k
IMDB-BINARY	complete_graph	5
IMDB-MULTI	complete_graph	5
COLLAB	complete_graph	3
REDDIT-BINARY	complete_graph	5
reddit_threads	complete_graph	5
ENZYMES	cycle_graph	5
PROTEINS	cycle_graph	5

main.py (--factor)

Datasets	$\epsilon$ (CTAug-GraphCL)	$\epsilon$ (CTAug-JOAO)	$\eta$ (CTAug-MVGRL)
IMDB-BINARY	0.2	0.2	0.4
IMDB-MULTI	0.4	0.2	0.4
COLLAB	0.2	0.2	0.2
REDDIT-BINARY	0.4	0.2	/
reddit_threads	0.2	0.2	/
ENZYMES	0.4	0.2	0.6
PROTEINS	0.8	1.0	0.8

Running Examples

• Example 1: Run CTAug-GraphCL on IMDB-BINARY dataset.

  python main.py --method=GraphCL_AUG_OGSN --dataset=IMDB-BINARY --feature=sub --factor=0.2 --dataset_file=IMDB-BINARY_global_complete_graph_5.pt

  • The output file is log/GraphCL_AUG_OGSN_sub_IMDB-BINARY/IMDB-BINARY_global_complete_graph_5_0.2.log, means we use CTAug-GraphCL method, choose IMDB-BINARY dataset, use 3,4,5-clique isomorphic subgraphs counts as node feature, and set probability decay factor $\epsilon$ at 0.2.
  • The log file will record the loss every epoch, calculate the prediction accuracy per 20 epoch (the total training epoch is 100), and get the mean and standard deviation for repeated experiments.

• Example 2: Run CTAug-MVGRL on IMDB-MULTI dataset.

  python main.py --method=MVGRL_AUG_OGSN --dataset=IMDB-MULTI --feature=sub --factor=0.4 --dataset_file=IMDB-MULTI_global_complete_graph_5.pt

  • The output file is log/MVGRL_AUG_OGSN_sub_IMDB-MULTI/IMDB-MULTI_global_complete_graph_5_0.4.log, means we use CTAug-MVGRL method, choose IMDB-MULTI dataset, use 3,4,5-clique isomorphic subgraphs counts as node feature, and set factor $\eta$ at 0.4.
  • The log file will record the loss of every epoch, calculate the prediction accuracy per 10 epoch (the total training epoch is 50), and get the mean and standard deviation for repeated experiments.

• Example 3: Get isomorphic subgraphs counts of IMDB-BINARY dataset.

  python get_sub_counts.py --dataset=IMDB-BINARY --k=5 --id_scope=global --id_type=complete_graph --level=graph

  • This example use 3,4,5-clique numbers as node features, and the output dataset will be saved under data folder, named IMDB-BINARY_global_complete_graph_5.pt.