GraphGenerator

package
license
support

GraphGenerator is an open-source tool for generating Graphs structure. It is based on BTER model and provides fair and representative graphs. This graphs can be useful in Deep Graph Learning problems, e.g. you can use it in GNN benchmarking.

Core features

The overall scheme of the generator is presented on the picture:

• Degrees of nodes are generated and divided on in-degree and out-degree values to keep the desired assortativity - making_degree_dist()
• These values are separated for blocks to keep the desired number of labels - making_clusters() or making_clusters_with_sizes()
• For each group of nodes with the same label, the edges are generated according to BTER model on in-degree nodes - bter_model_edges()
• At the end, edges on out-degrees are generated for all nodes at ones

Attributes are generated from normal distribution on node clusters adding noise from normal distribution:

Examples

The usage is presented in BTER_tuning.ipynb: As we are aim at graphs with given four graph characteristics, so we tune all input hyperparameters of generator so that the generated graph corresponds to the specified characteristics

License

This project is distributed under the 3-Clause BSD license.

Contacts

• egorshikov@itmo.ru write us to email and we will answer at your questions

Reference Paper

Polina Andreeva, Egor Shikov and Claudie Bocheninа "Attributed Labeled BTER-Based Generative Model for Benchmarking of Graph Neural Networks" Proceedings of the 17th International Workshop on Mining and Learning with Graphs (MLG) 2022:

@inproceedings{"mlg2022_5068",

title={Attributed Labeled BTER-Based Generative Model for Benchmarking of Graph Neural Networks}, author={Polina Andreeva, Egor Shikov and Claudie Bocheninа}, booktitle={Proceedings of the 17th International Workshop on Mining and Learning with Graphs (MLG)}, year={2022}}