This repository contains the analysis code for evaluating the systemic impact of various content recommender systems within a simulated social network environment populated by LLM agents. The analysis focuses on measuring structural and behavioral outcomes such as topic diffusion, conformity, engagement momentum, and network evolution.
The input data for this analysis is generated using YSocial — a digital twin framework designed for simulating agent-driven social interactions. YSocial outputs simulation traces and interaction logs in SQLite format, which are used here to compute metrics and generate evaluation figures.
- Install dependencies:
pip install -r requirements.txt
- Execute to merge the SQLite outputs of interest into concatenated CSV files:
jupyter notebook notebooks/compute.ipynb- Execute any notebook of interest. For example:
- ananlysis_conformity: Measures the degree to which agents conform to the behavior of their peers.
- analysis_follow: Analyzes changes in the follower/following social network structure over time.
- analysis_topic_diffusion: Evaluates how topics spread through the network.
- analysis_EM: Assesses the persistence of engagement over time.
- The notebooks will generate figures and tables that summarize the results of the analysis.
├── notebooks/ # Jupyter notebooks for analysis
│ ├── compute.ipynb # Merge SQLite outputs into CSV
│ ├── analysis_conformity.ipynb # Conformity analysis
│ ├── analysis_follow.ipynb # Follower network analysis
│ ├── ...
├── src/ # Python modules for metric computation and utilities
│ ├── db_destroyer/
│ │ ├── conformity.py # Conformity score computation
│ │ ├── db_parser.py # Read and parse SQLite outputs
│ │ ├── df_utils.py # DataFrame manipulation tools
│ │ ├── ... # Other utility modules
├── requirements.txt # Python dependencies
├── setup.py # Setup file (optional packaging)
└── README.md