LLMGreenRec: LLM-Based Multi-Agent Recommender System for Sustainable E-Commerce

TL;DR

LLMGreenRec optimizes green product recommendations using a multi-agent system and large language models, reducing digital carbon footprint.

cs.MA 🔴 Advanced 2026-03-12 12 views
Hao N. Nguyen Hieu M. Nguyen Son Van Nguyen Nguyen Thi Hanh
recommender systems large language models sustainability e-commerce multi-agent systems

Key Findings

Methodology

LLMGreenRec employs a multi-agent framework integrating Large Language Models (LLMs) for user interaction analysis and prompt optimization. The system deduces green-oriented user intents through collaborative analysis and prioritizes eco-friendly product recommendations. It consists of two main stages: first, a Cross-encoder reranker filters the candidate product set; then, a multi-agent system assesses user relevance and sustainability attributes. A final LLM reasons over these outputs to select the optimal product set.

Key Results

  • Experiments on the Bundle dataset show that LLMGreenRec achieves an HR@1 of 0.3950 and an HR@5 of 0.5504, demonstrating high efficiency in recommending green products. This means the correct sustainable product is successfully ranked first in nearly 40% of cases and appears within the top five recommendations over 55% of the time.
  • Compared to other high-performing LLM methods like PO4ISR, LLMGreenRec improves HR@5 by 26.6% on the Games dataset and NDCG@5 by 40.7% on the Bundle dataset. This highlights the system's superior performance in capturing complex, dynamic user intents.
  • By reducing unnecessary user interactions, LLMGreenRec effectively shortens search times and lowers operational energy consumption, directly reducing data center energy usage.

Significance

The introduction of LLMGreenRec holds significant implications for both academia and industry. It addresses the shortcomings of traditional recommender systems in recognizing green consumer intentions and promotes a green economy by reducing digital carbon footprints. By optimizing user interaction processes, it shortens shopping paths and reduces unnecessary energy consumption, providing a sustainable solution for e-commerce platforms.

Technical Contribution

LLMGreenRec distinguishes itself from existing state-of-the-art methods through its multi-agent system that automatically identifies and optimizes failures in the recommendation process, offering new engineering possibilities. By incorporating large language models, the system captures user intents more accurately and naturally integrates sustainability criteria into the recommendation process.

Novelty

LLMGreenRec is the first system to combine large language models with a multi-agent system for sustainable product recommendations. Unlike existing session-based recommender systems, it not only identifies green consumer intentions but also enhances recommendation accuracy and efficiency through automated prompt optimization.

Limitations

  • LLMGreenRec may underperform when handling sessions with high randomness, as it relies on user interaction history to predict intents.
  • The system may incur high computational costs, especially when processing large datasets, due to its reliance on significant computational resources.
  • The lack of standardized green index data may hinder the system's ability to accurately assess product sustainability in certain scenarios.

Future Work

Future research directions include developing more efficient algorithms to reduce computational costs, expanding the system to support more types of green indices, and exploring ways to further reduce energy consumption without sacrificing performance. The authors also suggest investigating applications in other domains such as sustainable tourism and energy management.

AI Executive Summary

In the rapid evolution of e-commerce, every user click and search query not only represents a consumer action but also consumes energy resources to power data centers. When multiplied by billions of daily interactions globally, the carbon footprint of online shopping becomes a significant concern. Consequently, a dual challenge emerges: guiding users toward environmentally friendly products while simultaneously optimizing the interaction process itself to minimize energy waste.

Consumer awareness of environmental issues and demand for sustainable products is clearly growing. According to PwC's Voice of the Consumer Survey, 54% of Vietnamese consumers are willing to pay up to 10% more for products made from recycled or sustainable materials. However, a significant contradiction persists: the stated sustainable consumption intentions of most customers do not align with their actual purchasing behavior. A survey by BCG indicates that while up to 80% of consumers claim they consider sustainability in their daily purchases, only 1-7% alter their purchasing habits to reflect this.

LLMGreenRec introduces a novel framework combining large language models and a multi-agent system to promote sustainable consumption. Through collaborative analysis of user interactions and iterative prompt refinement, LLMGreenRec's specialized agents deduce green-oriented user intents and prioritize eco-friendly product recommendations. Notably, this intent-driven approach also reduces unnecessary interactions and energy consumption.

Extensive experiments on benchmark datasets validate LLMGreenRec's effectiveness in recommending sustainable products, demonstrating a robust solution that fosters a responsible digital economy. Compared to traditional session-based recommender systems, LLMGreenRec captures user green consumption intentions more accurately and naturally integrates sustainability criteria into the recommendation process.

The introduction of LLMGreenRec holds significant implications for both academia and industry. It addresses the shortcomings of traditional recommender systems in recognizing green consumer intentions and promotes a green economy by reducing digital carbon footprints. By optimizing user interaction processes, it shortens shopping paths and reduces unnecessary energy consumption, providing a sustainable solution for e-commerce platforms. Future research directions include developing more efficient algorithms to reduce computational costs, expanding the system to support more types of green indices, and exploring ways to further reduce energy consumption without sacrificing performance.

Deep Analysis

Background

In the rapid evolution of e-commerce, recommender systems have become crucial tools for helping users find suitable products amidst vast information. However, traditional recommender systems often focus on maximizing short-term conversion rates, neglecting deeper user intents, especially in sustainable consumption. As environmental awareness increases, consumer demand for green products is rising, yet existing systems fall short in identifying and recommending these products. Recently, with the development of deep learning and large language models, significant advancements have been made in capturing user intents and integrating sustainability criteria into recommender systems.

Core Problem

Traditional session-based recommender systems significantly lack in recognizing users' green consumption intentions. These systems typically rely on popularity data to predict the most likely next action, making it difficult for small-scale enterprises offering sustainable products to gain attention in a digital marketplace. Additionally, users face information overload on complex e-commerce platforms, making it challenging to find and evaluate eco-friendly alternatives. This not only leads to decision fatigue but also increases unnecessary energy consumption, posing a technical barrier to the development of a green economy.

Innovation

LLMGreenRec introduces a novel framework combining large language models and a multi-agent system to promote sustainable consumption. Its core innovations include: 1) Utilizing the complex reasoning capabilities of large language models to accurately predict user needs with fewer interactions; 2) Collaborative analysis through a multi-agent system to automatically identify and optimize failures in the recommendation process; 3) Naturally integrating sustainability criteria into the recommendation process, prioritizing eco-friendly products. These innovations not only improve recommendation accuracy and efficiency but also reduce digital carbon footprints.

Methodology

LLMGreenRec's methodology involves two main stages:


  • �� Cross-encoder reranker: Serves as a relevance filtering step, processing current session interactions and the initial candidate set, generating semantic relationship embeddings, and outputting relevance scores.

  • �� Multi-agent system: Comprises six agents—Evaluate, DetectError, InferReason, RefinePrompt, Augment, and Select—each responsible for assessing user relevance, detecting errors, reasoning causes, optimizing prompts, generating prompt variants, and selecting the best prompt. Through collaborative analysis and iterative optimization, the system accurately captures users' green consumption intentions and prioritizes eco-friendly product recommendations.

Experiments

The experimental design includes evaluating LLMGreenRec's performance on MovieLen-1M, Amazon Games, and Bundle datasets. Hit Rate (HR@K) and Normalized Discounted Cumulative Gain (NDCG@K) are used as evaluation metrics, with K set to 1 or 5. The experimental results show that LLMGreenRec significantly outperforms all other baseline methods across all datasets and evaluation metrics, particularly excelling in capturing complex, dynamic user intents.

Results

Experiments on the Bundle dataset show that LLMGreenRec achieves an HR@1 of 0.3950 and an HR@5 of 0.5504, demonstrating high efficiency in recommending green products. Compared to other high-performing LLM methods like PO4ISR, LLMGreenRec improves HR@5 by 26.6% on the Games dataset and NDCG@5 by 40.7% on the Bundle dataset. By reducing unnecessary user interactions, LLMGreenRec effectively shortens search times and lowers operational energy consumption, directly reducing data center energy usage.

Applications

LLMGreenRec can be directly applied to e-commerce platforms, helping users quickly find eco-friendly products and reduce energy consumption during the shopping process. Its collaborative analysis capabilities of the multi-agent system can also be extended to other fields such as sustainable tourism and energy management, promoting the development of a green economy.

Limitations & Outlook

Despite LLMGreenRec's outstanding performance in recommending green products, it may underperform when handling sessions with high randomness. Additionally, the system may incur high computational costs, especially when processing large datasets. The lack of standardized green index data may hinder the system's ability to accurately assess product sustainability in certain scenarios. Future research directions include developing more efficient algorithms to reduce computational costs, expanding the system to support more types of green indices, and exploring ways to further reduce energy consumption without sacrificing performance.

Plain Language Accessible to non-experts

Imagine you're in a massive shopping mall, trying to find the most eco-friendly products. This mall has countless stores and products, and you need a smart assistant to help you find the items that best match your green intentions. LLMGreenRec acts like this assistant, analyzing your shopping habits and preferences to help you quickly find eco-friendly products. It not only recognizes your green consumption intentions but also saves you time and effort by reducing unnecessary shopping steps.

LLMGreenRec uses a technology called large language models, which is like a super-smart shopping guide that can understand every little detail of your shopping journey. It also collaborates with other assistants to optimize your shopping path, ensuring you find the most suitable products in the shortest time.

This system acts like a green shopping navigator, helping you make more environmentally friendly choices during your shopping process while reducing the mall's energy consumption. In this way, LLMGreenRec not only makes your shopping experience more enjoyable but also contributes to environmental protection.

ELI14 Explained like you're 14

Hey there, buddy! Imagine you're in a super big mall, looking to buy some eco-friendly stuff. But this mall is so huge, and there are so many things, you don't know where to start. That's when a smart robot assistant shows up, and it's called LLMGreenRec!

This robot assistant is super cool. It can guess what eco-friendly products you might like by looking at what you've bought before and what you like. It's like a super-smart friend who always knows what you want.

Even cooler, this assistant can work with other robots to make sure the things it recommends are not only what you like but also good for the environment. So, you get to buy what you like and help protect the planet at the same time!

So next time you're shopping, imagine having such a smart robot assistant helping you, making your shopping fast and eco-friendly. Isn't that awesome?

Glossary

Large Language Model

A large language model is an AI model capable of understanding and generating natural language, typically used for natural language processing tasks.

In LLMGreenRec, used for analyzing user interactions and optimizing recommendation prompts.

Multi-Agent System

A multi-agent system consists of multiple agents that can collaborate to accomplish complex tasks.

In LLMGreenRec, used for collaborative analysis of user intents and optimizing the recommendation process.

Recommender System

A recommender system is a tool that helps users find content of interest among vast information, widely used in e-commerce and content platforms.

LLMGreenRec is a system focused on recommending sustainable products.

Sustainability

Sustainability refers to meeting current needs without compromising the ability of future generations to meet their own needs, often involving environmental protection and resource conservation.

LLMGreenRec promotes sustainable consumption by prioritizing eco-friendly products.

Cross-encoder reranker

A Cross-encoder reranker is a model used to assess the relevance of candidate items with respect to user session interactions.

In LLMGreenRec, used for filtering and ranking the candidate product set.

Hit Rate (HR@K)

Hit Rate is a metric for evaluating recommender system performance, indicating the proportion of target items included in the top K recommendations.

Used to evaluate LLMGreenRec's effectiveness in recommending green products.

Normalized Discounted Cumulative Gain (NDCG@K)

NDCG is a metric for evaluating the ranking quality of recommender systems, considering the relevance and position of recommended items.

Used to evaluate LLMGreenRec's ranking quality in recommending green products.

Collaborative Filtering

Collaborative filtering is a recommendation technique that suggests content by analyzing users' historical behaviors and preferences of similar users.

A commonly used method in traditional recommender systems.

Semantic Embedding

Semantic embedding is a method of converting text or other data into vector representations to capture semantic information.

In LLMGreenRec, used to generate semantic relationship embeddings for user sessions and candidate items.

Automated Prompt Optimization

Automated prompt optimization is a process of iteratively adjusting prompts to improve system performance.

In LLMGreenRec, used to optimize recommendation prompts for improved accuracy.

Open Questions Unanswered questions from this research

  • 1 How can LLMGreenRec's energy consumption be further reduced without sacrificing performance? Current methods are effective but may incur high computational costs when processing large datasets. More efficient algorithms need to be developed to reduce computational costs while maintaining recommendation accuracy.
  • 2 How can LLMGreenRec be expanded to support more types of green indices? The system may not accurately assess product sustainability in certain scenarios, especially when standardized green index data is lacking. Research is needed to introduce more types of green indices to enhance the system's evaluation capabilities.
  • 3 How can LLMGreenRec's performance be improved when handling sessions with high randomness? The system currently relies on user interaction history to predict intents, but may underperform in sessions with high randomness. New methods need to be developed to improve the system's performance in these scenarios.
  • 4 How can LLMGreenRec's multi-agent system be applied in other fields? While the system excels in recommending green products, its collaborative analysis capabilities can also be extended to other fields such as sustainable tourism and energy management. Research is needed to effectively apply the system in these areas.
  • 5 How can LLMGreenRec's adaptability be improved in different cultural and market environments? Consumer demand and preferences for green products may vary across regions. Research is needed to adjust the system to adapt to different cultural and market environments, enhancing its global applicability.

Applications

Immediate Applications

E-commerce Platforms

LLMGreenRec can help e-commerce platforms optimize green product recommendations, reduce energy consumption during the shopping process, and improve user satisfaction.

Sustainable Tourism

By analyzing tourists' preferences and behaviors, LLMGreenRec can help tourism platforms recommend eco-friendly travel plans, promoting sustainable tourism development.

Energy Management

LLMGreenRec can be used in energy management systems to recommend energy-saving solutions by analyzing users' energy usage habits, helping users reduce energy consumption.

Long-term Vision

Global Green Economy

By promoting green products and reducing digital carbon footprints, LLMGreenRec can foster the development of a green economy globally, promoting sustainable consumption.

Smart Cities

LLMGreenRec's multi-agent system can be used in building smart cities, optimizing resource allocation and reducing energy consumption to enhance sustainable urban development.

Abstract

Rising environmental awareness in e-commerce necessitates recommender systems that not only guide users to sustainable products but also minimize their own digital carbon footprints. Traditional session-based systems, optimized for short-term conversions, often fail to capture nuanced user intents for eco-friendly choices, perpetuating a gap between green intentions and actions. To tackle this, we introduce LLMGreenRec, a novel multi-agent framework that leverages Large Language Models (LLMs) to promote sustainable consumption. Through collaborative analysis of user interactions and iterative prompt refinement, LLMGreenRec's specialized agents deduce green-oriented user intents and prioritize eco-friendly product recommendations. Notably, this intent-driven approach also reduces unnecessary interactions and energy consumption. Extensive experiments on benchmark datasets validate LLMGreenRec's effectiveness in recommending sustainable products, demonstrating a robust solution that fosters a responsible digital economy.

cs.MA cs.IR

References (20)

Automatic Prompt Optimization with "Gradient Descent" and Beam Search

Reid Pryzant, Dan Iter, Jerry Li et al.

2023 583 citations View Analysis →

Improved Recurrent Neural Networks for Session-based Recommendations

Yong Kiam Tan, Xinxing Xu, Yong Liu

2016 818 citations View Analysis →

Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks

Nils Reimers, Iryna Gurevych

2019 16585 citations View Analysis →

Neural Attentive Session-based Recommendation

Jing Li, Pengjie Ren, Zhumin Chen et al.

2017 1545 citations View Analysis →

Large Language Model Can Interpret Latent Space of Sequential Recommender

Zhengyi Yang, Jiancan Wu, Yanchen Luo et al.

2023 33 citations View Analysis →

Advancing Sustainability via Recommender Systems: A Survey

Xin Zhou, Lei Zhang, Hong-lei Zhang et al.

2024 14 citations View Analysis →

MACRec: A Multi-Agent Collaboration Framework for Recommendation

Zhefan Wang, Yuanqing Yu, Wen-Xun Zheng et al.

2024 68 citations View Analysis →

A Bi-Step Grounding Paradigm for Large Language Models in Recommendation Systems

Keqin Bao, Jizhi Zhang, Wenjie Wang et al.

2023 154 citations View Analysis →

Global Context Enhanced Graph Neural Networks for Session-based Recommendation

Ziyang Wang, Wei Wei, G. Cong et al.

2020 583 citations View Analysis →

E-Commerce Recommendation Applications

J. Schafer, J. Konstan, J. Riedl et al.

2004 1826 citations

Application of Dimensionality Reduction in Recommender System - A Case Study

B. Sarwar, G. Karypis, J. Konstan et al.

2000 1569 citations

Zero-Shot Next-Item Recommendation using Large Pretrained Language Models

Lei Wang, Ee-Peng Lim

2023 86 citations View Analysis →

A Novel AI-Driven Recommendation System for Eco-Conscious Consumers

T. Dinesh, K. Ammaiyappan, G. Janav et al.

2024 13 citations

Spatiotemporal-aware Session-based Recommendation with Graph Neural Networks

Yinfeng Li, Chen Gao, Xiaoyi Du et al.

2022 16 citations

A graph model for E-commerce recommender systems

Zan Huang, Wingyan Chung, Hsinchun Chen

2004 255 citations

Enhancing Tourism Recommender Systems for Sustainable City Trips Using Retrieval-Augmented Generation

Ashmi Banerjee, Adithi Satish, Wolfgang Wörndl

2024 8 citations View Analysis →

Recommender systems in e-commerce

Sanjeevan Sivapalan, A. Sadeghian, H. Rahnama et al.

2014 689 citations

Large Language Models are Zero-Shot Rankers for Recommender Systems

Yupeng Hou, Junjie Zhang, Zihan Lin et al.

2023 492 citations View Analysis →

Revisiting Bundle Recommendation: Datasets, Tasks, Challenges and Opportunities for Intent-aware Product Bundling

Zhu Sun, Jie Yang, Kaidong Feng et al.

2022 39 citations

STAMP: Short-Term Attention/Memory Priority Model for Session-based Recommendation

Qiao Liu, Yifu Zeng, Refuoe Mokhosi et al.

2018 987 citations