One-Class Graph Neural Networks for Anomaly Detection in Attributed Networks (OCGNN)

The Pytorch and DGL implement of the paper.

Details of our dataset

The Cora dataset has 7 categories of machine learning papers: "Case Based", "Genetic Algorithms", "Neural Networks (Class label = 2 in the DGL dataloader)", "Probabilistic Methods", "Reinforcement Learning", "Rule Learning", "Theory";

The Citeseer dataset consists of 6 paper classes: "Agents", "AI", "DB", "IR (Class label = 3)", "M"L, "HCI";

Each publication in the Pubmed dataset is classified into one of three classes ("Diabetes Mellitus, Experimental", "Diabetes Mellitus Type 1", "Diabetes Mellitus Type 2 (Class label = 2)").

In our experiments, classes in bold are defined as the normal classes, while the other classes are anomalous classes.

GNN based methods

Example:

python main.py --dataset [cora/citeseer/pubmed] --module [GCN/GAT/GraphSAGE/GAE] --nu 0.1 --lr 0.001 --n-hidden 32 --n-layers 2 --weight-decay 0.0005 --n-epochs 4000 --early-stop

Requirements:

pytorch>=1.4 DGL>=0.4.2 sklearn>=0.20.1 numpy>=1.16 networkx>=2.1

Two-stage mixture methods

Example:

python twostage.py --dataset [cora / citeseer / pubmed] --mode [A/X/AX] --emb-method [DeepWalk / Node2Vec / LINE / SDNE / Struc2Vec] --ad-method [PCA / OCSVM / IF / AE]

Requirements:

Pyod>=0.7.6 tensorflow>=1.4.0,<=1.14.0 gensim==3.6.0 DGL>=0.4.2 sklearn>=0.20.1 numpy>=1.16 networkx>=2.1