Graph Neural Network Bandits

Code to our paper Graph Neural Network Bandits, NeurIPS 2022. This paper considers Bandit/Bayesian Optimization problems on finite but very large graph domains, and uses a GNN for construct confidence sets for the target function.

Setup

The code is light and self-explanatory. Follows a few pointers to get you started.

• 'requirements.txt' lists the needed packages/used versions.

• The main two files are run_phasedgp.py and run_gnnucb.py, which run the GNN-PE and GNN-UCB algorithms, given that the environment is set up.

• This repository does not include the synthetis datasets, but does include the code to generate them from scratch. graph_env sub-folder contains the code to the Graph and Reward class, which together make up the environment. Scripts generate_dataset.py and launch_datagen.pyinitialize the environment and create the synthetic data. The latter includes a scheduler that parallelizes the process and submits the code to a cluster.

• Alternatively, you can write a data-loader that maps any given dataset into the graph and reward classes defined in the repository.

• All launcher files are scripts that generate the data for the experiments in our paper and are included for the sake of reproducibility.

• Implimentation of different BO algorithms are in algorithms.py and net.py includes our implimentation of a toy GNN. For real-world applications, we suggest using more complex architectures.