Supplementary Code and Data for "Network Formation and Dynamics among Multi-LLMs"

Installing Dependencies

The installation on a normal desktop computer should take a few minutes. To install the required packages use

pip install -r requirements.txt

Configuring the LLM API

Our experiments use GPT-3.5 and GPT-4 provided by OpenAI's API. To configure your environment please create a file named params.json structured as follows

{
    "OPENAI_API_KEY" : "YOUR_API_KEY",
    "OPENAI_ORG" : "YOUR_ORG_KEY"
}

Running the Simulations

The simulation files are located in principle_X.ipynb, where

• principle_1.ipynb corresponds to the simulations regarding preferential attachment
• principle_2.ipynb corresponds to the simulations regarding triadic closure
• principle_3.ipynb corresponds to the experiments regarding homophily and their community structure
• principle_5.ipynb corresponds to the experiments regardign small-world properties

Also, combined_model.ipynb corresponds to the experiments concerning the real-world datasets (located in datasets/).

The output figures are saved in figures/. The output tables are located in tables/.

The (cached) outputs of the simulations are located in outputs/. If you wish to rerun the experiments, erase the contents of this directory.

Simulation outputs for the real-world datasets

The simulation outputs for the real-world datasets can be found here. To include them in the project, download them and place them in the outputs/ folder.

Running the experiments on your own data

In order to run the software on your own data, you need to create a dataloader method in the dataloader.py file, similar to load_facebook100 and change the dataloader in combined_model.ipynb. The dataloader function should return a dictionary networks which has the labels of (potentially) multiple networks as keys, and the corresponding nx.Graph objects as values. Each nx.graph object can have features, which should be set using the nx..set_node_attributes(G, feat_dict, 'features') command where feat_dict is the dictionary of features (see also here for more information).

Expected runtimes

The experiments in principle_1.ipynb, principle_2.ipynb, principle_3.ipynb and principle_5.ipynb should take at most an hour each to run using the gpt-3.5-turbo model on a normal laptop computer.

The experiments in combined_model.ipynb take several hours (~4-5 hours) each to run using the gpt-4-1106-preview model for the datasets in question on a normal laptop computer.

Test system specifications

The experiments have been run at a MacBook M2 Pro with an Apple M2 Max chip, 32 GB of RAM, running macOS 13.0. The version of python used is Python 3.10.9.

Citation

If you use this code, please cite our work as follows

@article{papachristou2024network,
  title={Network Formation and Dynamics Among Multi-LLMs},
  author={Papachristou, Marios and Yuan, Yuan},
  journal={arXiv preprint arXiv:2402.10659},
  year={2024}
}