This repository contains the official implementation of our SIGIR 2024 demo paper:

• Wang, Zhefan, et al. "Multi-Agent Collaboration Framework for Recommender Systems." arXiv preprint arXiv:2402.15235 (2024).

The video demo is available at Video Demo.

A demo of using MACRec:

• macrec/: The source folder.
  • agents/: All agent classes are defined here.
    • analyst.py: The Analyst agent class.
    • base.py: The base agent class and base tool agent class.
    • interpreter.py: The Task Interpreter agent class.
    • manager.py: The Manager agent class.
    • reflector.py: The Reflector agent class.
    • searcher.py: The Searcher agent class.
  • dataset/: All dataset preprocessing methods.
  • evaluation/: The basic evaluation method, including the ranking metrics and the rating metrics.
  • llms/: The wrapper for LLMs (both API and open source LLMs).
  • pages/: The web demo pages are defined here.
  • rl/: The datasets and reward function for the RLHF are defined here.
  • systems/: The multi-agent system classes are defined here.
    • base.py: The base system class.
    • collaboration.py: The collaboration system class. We recommend using this class for most of the tasks.
    • analyse.py: (Deprecated) The system with a Manager and an Analyst. Do not support the chat task.
    • chat.py: (Deprecated) The system with a Manager, a Searcher, and a Task Interpreter. Only support the chat task.** Support all the tasks and all the agents.
    • react.py: (Deprecated) The system with a single Manager. Do not support the chat task.
    • reflection.py: (Deprecated) The system with a Manager and a Reflector. Do not support the chat task.
  • tasks/: For external function calls (e.g. main.py). Note needs to be distinguished from recommended tasks.
    • base.py: The base task class.
    • calculate.py: The task for calculating the metrics.
    • chat.py: The task for chatting with the ChatSystem.
    • evaluate.py: The task for evaluating the system on the rating prediction or sequence recommendation tasks. The task is inherited from generation.py.
    • feedback.py: The task for selecting the feedback for the Reflector. The task is inherited from generation.py.
    • generation.py: The basic task for generating the answers from a dataset.
    • preprocess.py: The task for preprocessing the dataset.
    • pure_generation.py: The task for generating the answers from a dataset without any evaluation. The task is inherited from generation.py.
    • reward_update.py: The task for calculating the reward function for the RLHF.
    • rlhf.py: The task for training the Reflector with the PPO algorithm.
    • sample.py: The task for sampling from the dataset.
    • test.py: The task for evaluating the system on few-shot data samples. The task is inherited from evaluate.py.
  • utils/: Some useful functions are defined here.
• config/: The config folder.
  • api-config.json: Used for OpenAI-like APIs' configuration. We give an example for the configuration, named api-config-example.json.
  • agents/: The configuration for each agent.
  • prompts/: All the prompts used in the experiments.
    • agent_prompt/: The prompts for each agent.
    • data_prompt/: The prompts used to prepare the input data for each task.
    • manager_prompt/: The prompts for the Manager in the CollaborationSystem with different configurations.
    • old_system_prompt/: (Deprecated) The prompts for other systems' agents.
    • task_agent_prompt/: (Deprecated) The task-specific prompts for agents in other systems.
  • systems/: The configuration for each system. Every system has a configuration folder.
  • tools/: The configuration for each tool.
  • training/: Some configuration for the PPO or other RL algorithms training.
• ckpts/: The checkpoint folder for PPO training.
• data/: The dataset folder which contains both the raw and preprocessed data.
• log/: The log folder.
• run/: The evaluation result folder.
• scripts/: Some useful scripts.

0. Make sure the python version is greater than or equal to 3.10.13. We do not test the code on other versions.

1. Run the following commands to install PyTorch (Note: change the URL setting if using another version of CUDA):

   pip install torch --extra-index-url https://download.pytorch.org/whl/cu118

2. Run the following commands to install dependencies:

   pip install -r requirements.txt

3. Run the following commands to download and preprocess the dataset (including ml-100k and Amazon Beauty):

   bash ./scripts/preprocess.sh

Use the following to run specific tasks:

python main.py -m $task_name --verbose $verbose $extra_args

Then main.py will run the ${task_name}Task defined in macrec/tasks/*.py.

E.g., to evaluate the sequence recommendation task in MovieLens-100k dataset for the CollaborationSystem with Reflector, Analyst, and Searcher, just run:

python main.py --main Evaluate --data_file data/ml-100k/test.csv --system collaboration --system_config config/systems/collaboration/reflect_analyse_search.json --task sr

You can refer to the scripts/ folder for some useful scripts.

Use the following to run the web demo:

streamlit run web_demo.py

Then open the browser and visit https://localhost:8501/ to use the web demo.

Please note that the systems utilizing open-source LLMs or other language models may require a significant amount of memory. These systems have been disabled on machines without CUDA support.

If you find our work useful, please do not save your star and cite our work:

@article{wang2024multi,
  title={Multi-Agent Collaboration Framework for Recommender Systems},
  author={Wang, Zhefan and Yu, Yuanqing and Zheng, Wendi and Ma, Weizhi and Zhang, Min},
  journal={arXiv preprint arXiv:2402.15235},
  year={2024}
}