BiGDN

Implementation of "BiGDN: An End-To-End Influence Maximization Framework Based on Deep Reinforcement Learning and Graph Neural Networks", view

Run the code

Train BiGDN-Teacher model

python main.py 

Train BiGDN-Student model

python main_s.py 

Test BiGDN-Teacher model

python test.py --test_graph_path test_data/Wiki-2.txt \
                 --teacher\
                 --model q_net.pth \
                 --k 10 
                 --selection_strategy o

Test BiGDN-Student model

python test.py --test_graph_path test_data/Wiki-2.txt \
                 --noteacher \
                 --model q_net_s.pth \
                 --k 10\
                 --selection_strategy o

Dependency requirement

• Python 3.9.18
• NumPy 1.23.5
• PyTorch 1.9.1+cu111
• PyG (PyTorch Geometric) 2.4.0
• PyTorch Scatter 2.0.7
• Tqdm 4.66.1
• networkx 3.2.1

Code files

• main.py: load program arguments, graphs and set up RL agent and environment and conduct simulation, train RL agent(teacher).
• main_s.py: load program arguments, graphs and set up RL agent and environment and conduct simulation, train RL agent(student).
• test.py: test RL agent.
• models.py: define parameters and structures of BiGDN and BiGDNs.
• agent.py: define agents to follow reinforcement learning procedure.
• environment.py: store the process of simulation.
• utils.py: obtain graph information function and use Monte Carlo simulation of propagation range under IC model and ReplayBuffer.
• node_encoder/train.ipynb: implementation and Pre training of Node Encoder.
• graph_process.ipynb: process the graph by converting nodes into consecutive integers numbered from 0.

Please refer to the code for specific parameter settings and implementation.

Folders

• test_graphs: there are 8 test graphs, the YouTube graph file is too large to upload.
• train_graphs: there are 15 training graphs, consistent with the training graphs of ToupleGDD.
• node_encoder: contains training graphs and training code used by NE.

License

This project is licensed under the terms of the MIT license.

Cite

If the code is userful for you, you can cite it:

@article{BiGDN,
title = {BiGDN: An end-to-end influence maximization framework based on deep reinforcement learning and graph neural networks},
journal = {Expert Systems with Applications},
volume = {270},
pages = {126384},
year = {2025},
issn = {0957-4174},
doi = {https://doi.org/10.1016/j.eswa.2025.126384},
url = {https://www.sciencedirect.com/science/article/pii/S0957417425000065},
author = {Wenlong Zhu and Kaijing Zhang and Jiahui Zhong and Chengle Hou and Jie Ji},
}