Code for the paper :

"Contingency-Aware Influence Maximization: A Reinforcement Learning Approach"

Authors : Haipeng Chen¹, Wei Qiu², Han-Ching Ou¹, Bo An², Milind Tambe¹

¹Harvard University     ²Nanyang Technological University

Accepted by UAI-2021.

Installation

First, install libraries via requirements.txt:

pip install -r requirements.txt 

Then install pytorch_sparse via wheel files. Go to this site: https://github.com/rusty1s/pytorch_sparse to install or download wheel files to speedup installation via this link: https://pytorch-geometric.com/whl/.

Run the code

RL training

1. Default command line:

python main.py --config=rl4im --env-config=basic_env --results-dir=results with lr=1e-3

All the default environment and method-related parameters are specified in src/tasks/config. You can set customized values of hyperparameters after with as demonstrated in the above command.

2. Batch run

There are three ways to run tasks in batch, running tasks with bash scripts on the current machine, running tasks on a distributed system via SLURM or running tasks with Docker. You can run jobs using sh multi_run.sh. Example settings in the paper are specified there which could be used to reproduce the results.

2.1 Running tasks with bash scripts on the current machine.

For example running tasks on GPU 0:

bash multi_run.sh --platform normal --gpuid 0

The normal means running tasks in the Python environment of current machine.

2,2 Running tasks on a distributed system via SLURM.

For example running tasks via SLURM on GPU 0:

bash multi_run.sh --platform slurm --gpuid 0

2.3 Running tasks with Docker.

Follow the instructions in https://github.com/Haipeng-Chen/RL4IM-with-Contingency/blob/main/docker/README.md For example running tasks with Docker on GPU 0:

bash multi_run.sh --platform docker --gpuid 0

RL validation (find the best checkpoint)

The training results will be stored under the directory results/rl4im/sacred/xx where xx is the id of one running of the experiment. This step is to find the optimal checkpoint from the validation results. You can use the script validation.ipynb, where you need to specify the runs id xx. We will try to upgrade in the next version with an end-to-end pipeline.

RL test

After finding the best check point via the validation process above, you are ready to run the test! An example of RL test is also given in multi_run.sh

It will load the model with the max step by default. If you want to load the specified model, add load_step=#OPT, where #OPT is the optimal checkpoint you find using the validation process above.

Run CHANGE baseline

An example code of it is also included in multi_run.sh

Cite

@article{chen2021contingency,
  title={Contingency-Aware Influence Maximization: A Reinforcement Learning Approach},
  author={Chen, Haipeng and Qiu, Wei and Ou, Han-Ching and An, Bo and Tambe, Milind},
  journal={arXiv preprint arXiv:2106.07039},
  year={2021}
}