Welcome to the WWW'25 AgentSociety Challenge! This repository provides the tools and framework needed to participate in a competition that focuses on building LLM Agents for user behavior simulation and recommendation systems based on open source datasets.
Participants are tasked with developing intelligent agents that interact with a simulated environment and perform specific tasks in two competition tracks:
- User Behavior Simulation Track: Agents simulate user behavior, including generating reviews and ratings.
- Recommendation Track: Agents generate recommendations based on provided contextual data.
This repository includes:
- The core library
websocietysimulatorfor environment simulation. - Scripts for dataset processing and analysis.
- Example usage for creating and evaluating agents.
This is the core library containing all source code required for the competition.
agents/: Contains base agent classes (SimulationAgent,RecommendationAgent) and their abstractions. Participants must extend these classes for their implementations.task/: Defines task structures for each track (SimulationTask,RecommendationTask).llm/: Contains base LLM client classes (DeepseekLLM,OpenAILLM).tools/: Includes utility tools:InteractionTool: A utility for interacting with the Yelp dataset during simulations.EvaluationTool: Provides comprehensive metrics for both recommendation (HR@1/3/5) and simulation tasks (RMSE, sentiment analysis).
simulator.py: The main simulation framework, which handles task and groundtruth setting, evaluation and agent execution.
Contains usage examples of the websocietysimulator library. Includes sample agents and scripts to demonstrate how to load scenarios, set agents, and evaluate them.
A script to process the raw Yelp dataset into the required format for use with the websocietysimulator library. This script ensures the dataset is cleaned and structured correctly for simulations.
The repository is organized using Python Poetry. Follow these steps to install the library:
-
Clone the repository:
git clone <this_repo> cd websocietysimulator
-
Install dependencies:
- Option 1: Install dependencies using Poetry: (Recommended)
poetry install && \ poetry shell - Option 2: Install dependencies using pip(COMING SOON):
pip install websocietysimulator
- Option 3: Install dependencies using conda:
conda create -n websocietysimulator python=3.11 && \ conda activate websocietysimulator && \ pip install -r requirements.txt && \ pip install .
- Verify the installation:
import websocietysimulator
- Download the raw dataset from the Yelp[1], Amazon[2] or Goodreads[3].
- Run the
data_process.pyscript to process the dataset:python data_process.py --input <path_to_raw_dataset> --output <path_to_processed_dataset>
- Check out the Data Preparation Guide for more information.
- NOTICE: You Need at least 16GB RAM to process the dataset.
Ensure the dataset is organized in a directory structure similar to this:
<your_dataset_directory>/
├── item.json
├── review.json
├── user.json
You can name the dataset directory whatever you prefer (e.g., dataset/).
Create a custom agent by extending either SimulationAgent or RecommendationAgent. Refer to the examples in the example/ directory. Here's a quick template:
from yelpsimulator.agents.simulation_agent import SimulationAgent
class MySimulationAgent(SimulationAgent):
def workflow(self):
# The simulator will automatically set the task for your agent. You can access the task by `self.task` to get task information.
print(self.task)
# You can also use the `interaction_tool` to get data from the dataset.
# For example, you can get the user information by `interaction_tool.get_user(user_id="example_user_id")`.
# You can also get the item information by `interaction_tool.get_item(item_id="example_item_id")`.
# You can also get the reviews by `interaction_tool.get_reviews(review_id="example_review_id")`.
user_info = interaction_tool.get_user(user_id="example_user_id")
# Implement your logic here
# Finally, you need to return the result in the format of `stars` and `review`.
# For recommendation track, you need to return a candidate list of items, in which the first item is the most recommended item.
stars = 4.0
review = "Great experience!"
return stars, review- Check out the Tutorial for Agent Development.
- Check out the Baseline User Behavior Simulation Agent for the baseline implementation (Track 1).
- Check out the Baseline Recommendation Agent for the baseline implementation (Track 2).
Run the simulation using the provided Simulator class:
from websocietysimulator import Simulator
from my_agent import MySimulationAgent
# Initialize Simulator
simulator = Simulator(data_dir="path/to/your/dataset", device="auto", cache=False)
# The cache parameter controls whether to use cache for interaction tool.
# If you want to use cache, you can set cache=True. When using cache, the simulator will only load data into memory when it is needed, which saves a lot of memory.
# If you want to use normal interaction tool, you can set cache=False. Notice that, normal interaction tool will load all data into memory at the beginning, which needs a lot of memory (20GB+).
# Load scenarios
simulator.set_task_and_groundtruth(task_dir="path/to/task_directory", groundtruth_dir="path/to/groundtruth_directory")
# Set your custom agent
simulator.set_agent(MySimulationAgent)
# Set LLM client
simulator.set_llm(DeepseekLLM(api_key="Your API Key"))
# Run evaluation
# If you don't set the number of tasks, the simulator will run all tasks.
agent_outputs = simulator.run_simulation(number_of_tasks=None, enable_threading=True, max_workers=10)
# Evaluate the agent
evaluation_results = simulator.evaluate()- If you want to use your own LLMClient, you can easily implement it by inheriting the
LLMBaseclass. Refer to the Tutorial for more information.
- You should register your team firstly in the competition homepage (Homepage).
- Submit your solution through the submission button at the specific track page. (the submission button is at the top right corner of the page)
- User Modeling Track
- Recommendation Track
- Please register your team first.
- When you submit your agent, please carefully SELECT the TRACK you want to submit to.
- The content of your submission should be a zip file containing your agent (Only one
{your_agent}.pyfile without evaluation code). - Example submissions:
- For Track 1: submission_1
- For Track 2: submission_2
The InteractionTool is the core utility for interacting with the dataset. It provides an interface for querying user, item, and review data.
-
Get User Information: Retrieve user data by user ID or current scenario context.
user_info = interaction_tool.get_user(user_id="example_user_id")
-
Get Item Information: Retrieve item data by item ID or current scenario context.
item_info = interaction_tool.get_item(item_id="example_item_id")
-
Get Reviews: Fetch reviews related to a specific item or user, filtered by time.
reviews = interaction_tool.get_reviews(review_id="example_review_id") # Fetch a specific review reviews = interaction_tool.get_reviews(item_id="example_item_id") # Fetch all reviews for a specific item reviews = interaction_tool.get_reviews(user_id="example_user_id") # Fetch all reviews for a specific user
This project is licensed under the MIT License. See the LICENSE file for details.
[1] Yelp Dataset: https://www.yelp.com/dataset
[2] Amazon Dataset: https://amazon-reviews-2023.github.io/
[3] Goodreads Dataset: https://sites.google.com/eng.ucsd.edu/ucsdbookgraph/home
