Bayesian Optimization of Functions over Node Subsets in Graphs [NeurIPS2024]

This is the repo for GraphComBO, a Bayesian optimization tool for black-box and expensive functions defined over node subsets in graphs, where the goal is to find the optimal subset within a limited number of queries (evaluations of the underlying function).

Search Animation

Recursive Combo-subgraph Sampling - Demo

Installation

1. Create a virtual environment for GraphComBO

conda create -n GraphComBO
conda activate GraphComBO

2. Install Common Packages (skip if satisfied)

The following common packages are assumed to be installed beforehand:

numpy pandas matplotlib scipy networkx math functools itertools future multiprocessing

Note that scikit-learn and PyTorch are also required, which can be installed in the links.

3. Install BoTorch

We use BoTorch and GPyTorch to implement our BO algorithm.

To install BoTorch with conda:

conda install botorch -c pytorch -c gpytorch -c conda-forge

or alternatively install with pip

pip install botorch

4. Install other Packages for Experiments

Lastly, the following packages are used for different experiment setups.

pip install torch_geometric
pip install ndlib
pip install osmnx
pip install ogb

Run

Use the following line in a bash shell to test the algorithm with the default config from configurations/testing.yaml, and enable --plot_result True to visualize the result.

python main.py --plot_result True

To reproduce the experiment results, use the following code with pre-specified configurations:

python main.py --problem BA
python main.py --problem WS
python main.py --problem GRID
python main.py --problem SBM
python main.py --problem Transitivity
python main.py --problem GNN_attack
python main.py --problem Coauthor_IC
python main.py --problem SIR
python main.py --problem Patient_Zero