Source-linked AI summary
MACRec: a Multi-Agent Collaboration Framework for Recommendation
Zhefan Wang, Yuanqing Yu, Wendi Zheng, Weizhi Ma, Min Zhang
TL;DR
Recommendation research has used agents to simulate users and items or build single-agent recommenders, while multi-agent recommendation frameworks have had limited roles or fixed collaboration. MACRec applies specialized, adaptable agents directly to recommendation tasks and presents examples across four scenarios, alongside a web interface visualizing collaboration.
Problem
Agent-based recommendation has focused on user/item simulation or single-agent recommenders, while existing multi-agent work has limited agent types and fixed collaboration modes.
Source: §1 · p. 1Method
MACRec uses specialized LLM-powered agents, including a Manager, Analysts, Reflector, Searcher, and Task Interpreter, to collaborate directly on recommendation tasks.
Source: §1 · p. 1Results
The paper presents MACRec applications for rating prediction, sequential recommendation, conversational recommendation, and explanation generation, plus an online interface that visualizes agent collaboration.
Source: §1 · pp. 1–2, §6 · p. 4Takeaways & Limitations
MACRec demonstrates adaptable multi-agent collaboration for diverse recommendation scenarios, with agents selected and customized to scenario requirements.
Source: §2.1 · p. 2, §1 · p. 2
Abstract
from arXiv · showhide
LLM-based agents have gained considerable attention for their decision-making skills and ability to handle complex tasks. Recognizing the current gap in leveraging agent capabilities for multi-agent collaboration in recommendation systems, we introduce MACRec, a novel framework designed to enhance recommendation systems through multi-agent collaboration. Unlike existing work on using agents for user/item simulation, we aim to deploy multi-agents to tackle recommendation tasks directly. In our framework, recommendation tasks are addressed through the collaborative efforts of various specialized agents, including Manager, User/Item Analyst, Reflector, Searcher, and Task Interpreter, with different working flows. Furthermore, we provide application examples of how developers can easily use MACRec on various recommendation tasks, including rating prediction, sequential recommendation, conversational recommendation, and explanation generation of recommendation results. The framework and demonstration video are publicly available at https://github.com/wzf2000/MACRec.
1 INTRODUCTION
MACRec addresses the gap between agent-based recommendation simulation and systems that directly perform recommendation, using specialized agents that collaborate on tasks. It also presents examples across recommendation scenarios and an online interface for visualizing collaboration.
1 INTRODUCTION
- Prior agent-based recommendation work either simulates user or item behavior, uses a single recommender agent, or offers multi-agent collaboration with limited roles and a fixed mode.The paper identifies complex recommendation decisions as tasks on which single-agent instances are unable to perform well.Source: §1 · p. 1
- MACRec directly tackles recommendation tasks through collaboration among LLM-powered agents with distinct roles, rather than using agents only to simulate users or items.Roles include Manager, Reflector, User/Item Analysts, Searcher, and Task Interpreter.Source: §1 · p. 1
- MACRec provides application examples for rating prediction, sequential recommendation, conversational recommendation, and recommendation explanation generation.The framework also demonstrates selecting and customizing agents for different scenario requirements.Source: §1 · pp. 1–2
- An online web interface visualizes how agents collaboratively tackle recommendation tasks.Source: §1 · p. 2
2 RELATED WORK
Prior agent-based recommendation research includes behavior simulation and recommender-oriented systems, while broader multi-agent collaboration research had not investigated recommendation scenarios. MACRec is presented as supporting adaptable, multi-type agent collaboration for diverse recommendation uses.
2.1 Agents-based Recommendation
- Simulation-oriented systems use agents to model user behavior, item characteristics, or user-item interactions, but the paper says this work falls short of integration into recommender systems.RecAgent [23] and Agent4Rec [17] use single-type user agents; AgentCF [25] uses user and item agents with simple interactions.Source: §2.1 · p. 2
- Recommender-oriented work includes single-agent systems focused on planning and reflection, while RAH [11] supports different agent types only in a fixed collaboration mode.The paper states that RAH lacks publicly accessible code or demos.Source: §2.1 · p. 2
- MACRec enables adaptable collaboration and is described as the first open-source framework supporting multi-type agents for diverse recommendation scenarios.Source: §2.1 · p. 2
2.2 Multi-agent Collaboration
- Prior multi-agent collaboration research includes dialogue-based systems such as CAMEL [6] and AutoGen [20], but had not investigated its potential in recommendation scenarios.Source: §2.2 · p. 2
3 THE MACREC FRAMEWORK
MACRec coordinates specialized agents to solve recommendation tasks through a Manager-led workflow. Agents translate tasks, gather user and item analyses or search results, and reflect on proposed answers.
3.1 Framework Overview
- In the sequential recommendation example, the Task Interpreter clarifies the task, then the Manager requests analyses and search results before producing a candidate ranking.The Reflector reviews the answer, and the Manager reattempts the task using its suggestions.Source: §3.1 · pp. 2–3
3.2 Agent Roles
- The Manager assigns subtasks, oversees collaboration, and alternates Thought, Action, and Observation phases while deciding whether to answer or seek help.Source: §3.2.1 · p. 3
- The Reflector judges the Manager’s answer on later runs and either ends the process or identifies possible improvements.One example is reminding the Manager to consider highly rated items in the user’s history.Source: §3.2.2 · p. 3
- User/Item Analysts examine user preferences and item attributes using information and interaction-history tools, while the Searcher retrieves and summarizes relevant text for the Manager.The Task Interpreter translates dialogs into executable recommendation tasks and can summarize long conversation histories for its prompt.Source: §3.2.3 · p. 3, §3.2.4 · p. 3, §3.2.5 · p. 3
4 APPLICATIONS ON RECOMMENDATION SCENARIOS
MACRec presents four recommendation applications and describes how agent roles can be selected for each. The examples tailor collaboration to the needs of rating, sequential, explanation, and conversational tasks.
4 APPLICATIONS ON RECOMMENDATION SCENARIOS
- Table 2 summarizes which agents are selected for each of MACRec’s four recommendation applications.Source: Table 2 · p. 3, §4 · p. 3
4.1 Rating Prediction (RP)
- For rating prediction, User and Item Analysts provide preference and item-characteristic analyses that help the Manager predict a user’s rating.The analyses include the user’s historical interactions and preferences, and the target item’s recent ratings.Source: §4.1 · p. 4
4.2 Sequential Recommendation (SR)
- For sequential recommendation, the User Analyst examines short- and long-term interests, while the Reflector can address omissions in behavioral analysis and ranking-format issues.The passage notes that the sequence can contain many relevant items and that the task requires ranking candidates.Source: §4.2 · p. 4
4.3 Explanation Generation (EG)
- For explanation generation, User/Item analysis is needed, and the Searcher can retrieve extra item information that may help the Manager interpret user behavior.The example is information about a movie director that is absent from the dataset.Source: §4.3 · p. 4
4.4 Conversational Recommendation (CR)
- For conversational recommendation, the Task Interpreter converts potentially non-instructive dialog into a clear task, and the Searcher can find information the Manager does not know.MACRec can support other scenarios by customizing configurations.Source: §4.4 · p. 4
5 INTERFACE DEMONSTRATION
The interface demonstrates how users configure tasks and systems, then view agents collaborating on a recommendation. A conversational case shows the workflow from interpreting a request to producing a recommendation.
5 INTERFACE DEMONSTRATION
- Figure 2 presents the web interface and a case study of three agents collaborating on conversational recommendation.Source: §5 · p. 4
- Figure 2 shows a configuration panel for selecting tasks and customizing systems, alongside an interaction panel for the agents’ collaboration.Users can select tasks such as rating prediction.Source: §5 · p. 4
- In the conversational case, the Interpreter clarifies a request for historical films, and the Manager uses two Searcher rounds before recommending Amistad.The search covers historical movies and movies similar to Schindler’s List.Source: §5 · p. 4
6 CONCLUSION
MACRec is an LLM-based multi-agent framework that directly addresses recommendation tasks through agent collaboration. The paper presents applications across four recommendation tasks and an online interface visualizing collaboration.
6 CONCLUSION
- MACRec directly tackles recommendation tasks through collaboration among LLM-based agents, rather than using agents only to simulate users or items.Source: §6 · p. 4
- The paper presents MACRec applications for rating prediction, sequential recommendation, conversational recommendation, and explanation generation.Source: §6 · p. 4
- An online web interface visualizes how agents collaborate in MACRec.Source: §6 · p. 4