Agent4Rec

On Generative Agents in Recommendation

Agent4Rec, a recommender system simulator with 1,000 LLM-empowered generative agents. These agents are initialized from the MovieLens-1M dataset, embodying varied social traits and preferences. Each agent interacts with personalized movie recommendations in a page-by-page manner and undertakes various actions such as watching, rating, evaluating, exiting, and interviewing. With Agent4Rec, we would like to explore the potential of LLM-empowered generative agents in simulating the behavior of genuine, independent humans in recommendation environments.

📋 Catalogue

• Catalogue

• Preparations

• Simulation

  • Quick Start
  • Explore Various Recommender Settings
  • See the result

• Simulation Cost

⚙️ Preparations

Step 1. Install requirements.txt

Set up a virtualenv and install the pytorch manually. After that, install all the dependencies listed in the requirements.txt file by running the following command:

pip install -r requirements.txt

Our experiments have been tested on Python 3.9.12 with PyTorch 1.13.1+cu117. Python version over 3.10 may lead to some bugs in the package 'reckit'.

Step 2. Set up necessary environments

Make sure you are in the directory of recommenders/ (where setup.py can be found), and run the following code.

python setup.py build_ext --inplace

The command will install necessary tools for accelerating recommender evaluation.

⌛️ Simulation

Make sure you are in the main directory (where main.py can be found).

Export your OpenAI API key first:

export OPENAI_API_KEY=

Replace \ with your own OpenAI API key.

Quick Start

By running the following command, you will start a toy simulation with 3 agents.

python main.py

The response of agents to recommended items will be printed in the terminal. This simulation will take around 3 minutes to finish.

Explore Various Recommender Settings

Agent4Rec supports various recommendation systems and different simulation configurations.

python main.py --simulation_name MyExp --modeltype MF --n_avatars 10 --max_pages 5 --items_per_page 4 --execution_mode parallel

By running this code, you will start a simulation named MyExp with 10 agents, each agent will browse max to 5 pages with 4 items on a single page. The recommender used in this example is Matrix Factorization (short for MF). And the experiment will be executed in parallel model to speed up the simulation.

You can choose the employed recommender by modifying --modeltype in the command. You can replace `` with the following supported recommenders:

• Random: Randomly recommend items to users.
• Pop: Randomly recommend popular items to users.
• MF: Pretrained Matrix Factorization model with BPR loss.
• MultVAE: Pretrained MultVAE model.
• LightGCN: Pretrained LightGCN model with BPR loss.

See the Results

The results of the simulation will be saved in storage/ml-1m// directory. As for the command in Section Explore Various Recommender Settings, the results will be saved in storage/ml-1m/MyExp directory. All the interaction history of agent 0 is documented in storage/ml-1m/MF/MyExp/running_logs/0.txt

💰 Simulation Cost

🛎️ Note that all the experiments are powered by ChatGPT-3.5, and a complete simulation involving 1000 users would cost approximately $16. ($0.016/User)