Agent4Rec is the implementation of the paper "On Generative Agents in Recommendation" (SIGIR 2024 perspective), a recommender system simulator with 1,000 LLM-empowered generative agents. These agents are initialized from the MovieLens-1M dataset and embody different social characteristics and preferences. Each agent interacts with personalized movie recommendations in a page-by-page manner and performs various actions such as watching, rating, reviewing, exiting, and being interviewed. Through Agent4Rec, one can explore the potential of LLM-empowered generative agents to simulate realistic, independent human behavior in recommendation environments.
Main capabilities:
- Supports simulation of 1,000 generative agents.
- Agents are initialized from the MovieLens-1M dataset and have diverse social characteristics and preferences.
- Each agent interacts with personalized movie recommendations page by page and performs actions such as watching, rating, reviewing, exiting, and being interviewed.
- Supports multiple recommender systems: Random (randomly recommend items), Pop (randomly recommend popular items), MF (pre-trained matrix factorization model with BPR loss), MultVAE (pre-trained MultVAE model), LightGCN (pre-trained LightGCN model with BPR loss).
- Supports different simulation configurations, such as the number of agents, items per page, maximum number of browsing pages, parallel mode, etc.
Usage:
- Prepare the environment:
- Install requirements.txt. It is recommended to set up a virtual environment and manually install PyTorch. The experiments were tested with Python 3.9.12 and PyTorch 1.13.1+cu117. Python versions higher than 3.10 may cause bugs in the 'reckit' package.
- Install the necessary environment: enter the recommenders/ directory (which contains setup.py) and run the relevant code to install the tools required to accelerate recommendation evaluation.
- Run the simulation:
- Make sure you are in the main directory (which contains main.py), export the OpenAI API key: export OPENAI_API_KEY=<Your OpenAI API key>, and replace <Your OpenAI API key> with your own OpenAI API key.
- Run the following command to start a toy simulation containing 3 agents. The agents' responses to recommended items will be printed in the terminal and it will finish in about 3 minutes.
- Supports multiple recommender systems and configurations: for example, run a simulation named MyExp, with 10 agents, each agent browsing at most 5 pages, 4 items per page, using MF as the recommender, and accelerating the experiment in parallel mode. You can choose the recommender by modifying --modeltype <model_name>.
- View the results:
- The simulation results are saved in the storage/ml-1m/<model_name>/<experiment_name> directory. For example, the results of the above command are saved in the storage/ml-1m/MyExp directory. All interaction history of agent 0 is recorded in storage/ml-1m/MF/MyExp/running_logs/0.txt.
Simulation cost: All experiments are driven by ChatGPT-3.5. The cost of a complete simulation involving 1,000 users is about $16 ($0.016 per user).
