In this repo, you can train a GRL or GNN model on graph datasets (TU Dataset format), perform inference on your own graphs, conduct robustness analysis, or analyze your computed graph embeddings.
- Graph2Vec
- NetLSD
- DeepWalk
- GCN
- GAT
- GIN
- PNA
For embedding analysis we also include code for t-SNE and Isomap visualizations.
Note: UMAP support is available but requires manual installation due to dependency conflicts with
karateclub. See the visualization module for details.
For downstream classification on graph embeddings:
- SVM (Support Vector Machine with RBF/Linear kernels)
- MLP (Multi-Layer Perceptron)
For unsupervised analysis of graph embeddings:
- K-Means
- Spectral Clustering
- t-SNE manifold visualization
- Isomap manifold visualization
- Cluster scatter plots
- Graph structure visualization (with node/edge labels)
Tools for testing model robustness under graph perturbations:
- Random edge addition
- Random edge removal
- Node attribute shuffling
All training runs report:
- Accuracy
- AUROC (Area Under ROC Curve)
- F1 Score (macro-averaged for multiclass)
- Precision
- Recall
- Specificity
- Confusion Matrix
- Peak Memory Usage
- Training Time
This package supports datasets in the TU Dataset format (e.g., MUTAG, ENZYMES, PROTEINS, IMDB-MULTI). The dataset folder should contain:
*_A.txt- Edge list*_graph_indicator.txt- Graph membership for each node*_graph_labels.txt- Graph class labels*_node_labels.txt(optional) - Node labels*_node_attributes.txt(optional) - Node feature vectors*_edge_labels.txt(optional) - Edge labels
Example structure: data/MUTAG/ ├── MUTAG_A.txt ├── MUTAG_graph_indicator.txt ├── MUTAG_graph_labels.txt ├── MUTAG_node_labels.txt └── MUTAG_edge_labels.txt
To install as a pypi package, in the root directory, just do:
pip3 install .Package name is: information_systems, so make sure to check if this exists.
We implemented a CLI so you don't have to write code every time you need to train, perform inference or even perform simple analysis. You can do information_systems --help for more information, but we will list some examples below:
information_systems train --model graph2vec \
--dataset_dir data/MUTAG \
--test_size 0.25 \
--classifier SVM \
--out_channels 256 \
--epochs 100 \
--model_name graph2vec_model.pkl \
--device cpu \information_systems train --model gat \
--dataset_dir data/ENZYMES \
--test_size 0.25 \
--classifier SVM \
--hidden_channels 64 \
--out_channels 128 \
--dropout 0.5 \
--batch_size 2 \
--epochs 1000 \
--patience 100 \
--model_name gat_best.pth \information_systems inference --model graph2vec \
--out_channels 128 \
--dataset_dir data/MUTAG \
--model_weights graph2vec_model.pkl \
--out_json graph2vec_inference.json \information_systems inference --model gat \
--num_layers 2 \
--hidden_channels 64 \
--out_channels 128 \
--dropout 0.5 \
--dataset_dir data/ENZYMES \
--model_weights gat_best.pth \
--out_json gat_inference.json \# Add 20% random edges
information_systems inference --model gin \
--dataset_dir data/ENZYMES \
--model_weights gin_best.pth \
--out_json gin_perturbed.json \
--add_random_edges 0.2
# Remove 15% random edges
information_systems inference --model graph2vec \
--out_channels 128 \
--dataset_dir data/MUTAG \
--model_weights graph2vec_model.pkl \
--out_json graph2vec_perturbed.json \
--remove_random_edges 0.15
# Shuffle node attributes
information_systems inference --model gat \
--num_layers 2 \
--hidden_channels 64 \
--out_channels 128 \
--dataset_dir data/ENZYMES \
--model_weights gat_best.pth \
--out_json gat_shuffled.json \
--shuffle_node_attributesinformation_systems analysis --in_jsons gat_inference.json \
--manifold TSNEinformation_systems analysis --in_jsons gat_inference.json \
--clustering kmeans
information_systems analysis --in_jsons gat_inference.json \
--clustering spectral