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

The paper introduces LLMGreenRec, a novel multi-agent framework leveraging Large Language Models to deduce user green intents and prioritize sustainable product recommendations, thereby bridging the gap between eco-friendly intentions and actions while minimizing the system's own digital carbon footprint.

The Gist

Imagine you are walking into a massive, chaotic supermarket. The shelves are stacked to the ceiling with millions of products. You have a specific goal: you want to buy something eco-friendly (like a reusable water bottle or a shirt made from recycled plastic).

But here's the problem:

1. The Store is Overwhelming: There are so many options that you get tired (decision fatigue) and just grab the familiar, cheap brand you've seen a thousand times, even though you wanted to be green.
2. The Store is Wasteful: Every time you walk down an aisle, pick up an item, put it back, and walk to another aisle, the store's lights and computers are burning energy. The more you wander, the more "digital carbon" you create.

LLMGreenRec is a new, super-smart shopping assistant designed to fix both problems. Think of it as a personal shopper who is also a green activist and an energy saver.

Here is how it works, broken down into simple steps:

1. The "Smart Filter" (The Reranker)

Imagine you walk into the store, and instead of showing you all 100,000 items, this assistant instantly pulls out a small basket containing only the top 20 items that match what you were just looking at.

• How it works: It looks at your recent clicks (like "I looked at running shoes") and uses a smart filter to throw away everything irrelevant.
• The Benefit: You don't waste time scrolling through junk. This saves energy because the computer doesn't have to process millions of useless options.

2. The "Six-Person Dream Team" (The Multi-Agent System)

This is the magic part. Instead of one robot trying to guess what you want, the system uses six specialized AI agents who act like a team of experts sitting around a table, debating and refining the best recommendation.

Here is what each "person" on the team does:

• 🧐 The Evaluator (The Judge): "Okay, based on what the user just clicked, here is my list of top 20 products. Let's see how good this list is."
• 🚨 The Error Detector (The Safety Net): "Wait! The user actually wanted a sustainable option, but my list put a cheap plastic one at #1. That's a mistake!"
• 🕵️ The Reasoner (The Detective): "Why did we make that mistake? Ah, I see. The instructions we gave the AI were too vague. We didn't emphasize that the user cares about 'recycled materials' enough."
• ✍️ The Refiner (The Editor): "Got it. Let's rewrite the instructions to the AI. Instead of saying 'Show me shoes,' let's say 'Show me shoes made from recycled rubber that are good for running.'"
• 🎭 The Augmenter (The Creative): "Let's try saying that in three different ways to see which one works best. Maybe 'Eco-friendly running gear' sounds better than 'Recycled rubber shoes'?"
• 🏆 The Selector (The Coach): "I've tested all these different instructions. The one that says 'Eco-friendly running gear' got the best results. Let's use that one for the next round!"

3. The "Self-Correcting Loop"

The team doesn't just stop after one try. They keep looping:

1. They make a guess.
2. They check if they were right.
3. If they were wrong, they figure out why, rewrite their instructions, and try again.
4. They keep doing this until they find the perfect prompt (the perfect set of instructions) to give the AI.

Why is this a Big Deal?

1. It Bridges the "Say-Do" Gap
Many people say they want to be green, but they do buy the cheap, non-green stuff because it's easier to find. This system acts like a helpful friend who gently nudges you toward the eco-friendly choice before you get tired and give up. It makes the "green" choice the "easy" choice.

2. It Saves Energy (The Digital Carbon Footprint)
Traditional shopping bots make you click, scroll, and search through hundreds of items to find what you want. That uses a lot of electricity in data centers.
LLMGreenRec is like a high-speed elevator instead of a staircase. It predicts exactly what you need with fewer clicks. Fewer clicks = less energy burned = a smaller carbon footprint for the internet itself.

The Bottom Line

LLMGreenRec is a smart, self-improving team of AI agents that helps you find sustainable products faster. It saves you time, saves the planet from digital pollution, and helps eco-friendly small businesses get the attention they deserve, rather than getting lost in a sea of big brands.

It's not just about selling you a product; it's about making the whole shopping experience smarter, greener, and less wasteful.

Technical Summary

Here is a detailed technical summary of the paper "LLMGreenRec: LLM-Based Multi-Agent Recommender System for Sustainable E-Commerce."

1. Problem Statement

The paper addresses a dual challenge in modern e-commerce:

• The "Say-Do" Gap in Sustainability: While consumer surveys indicate a strong intent to purchase sustainable products (e.g., 54% of Vietnamese consumers willing to pay a premium), actual purchasing behavior rarely reflects this due to decision fatigue, information overload, and the dominance of conventional products on platforms.
• Digital Carbon Footprint: Traditional session-based recommender systems are optimized for short-term conversion rates using popularity data. This often leads to redundant user interactions (scrolls, searches) and fails to surface niche eco-friendly products, thereby increasing the energy consumption of data centers and failing to support a green economy.

Core Objective: To develop a recommender system that accurately identifies user intent to promote sustainable consumption while simultaneously minimizing the computational energy required to generate recommendations.

2. Methodology: LLMGreenRec Architecture

The proposed LLMGreenRec is a two-stage framework designed to filter candidates and optimize recommendations through a collaborative multi-agent system.

Stage 1: Candidate Filtering (Reranker Model)

• Input: A current session sequence ($S$) and an initial candidate set ($C_{in}$) of 100 random items (ensuring the ground-truth target is included).
• Mechanism: A Cross-encoder Reranker (based on Sentence Transformers and a Classifier) evaluates the relevance of every candidate item against the session history.
• Process:
  1. Pairs are formed between session items ($s$) and candidate items ($c$).
  2. A joint embedding is generated to capture semantic relationships.
  3. A relevance score is computed for each candidate by averaging pairwise scores across the session.
• Output: The top 20 most relevant items are selected to form a filtered candidate set ($C_{filter}$), significantly reducing the search space for the next stage.

Stage 2: Multi-Agent Prompt Optimization

This stage leverages Large Language Models (LLMs) and a Multi-Agent System to refine prompts and select the best sustainable recommendations. Six specialized agents operate in a closed-loop feedback cycle:

1. Evaluate Agent: Generates an initial ranked list of products based on the current prompt and session data. It serves as the baseline for performance testing.
2. DetectError Agent: Checks if the ground-truth target item is ranked within a specific threshold (e.g., top 10). If not, it flags the session as an "error case."
3. InferReason Agent: Analyzes the error to deduce the root cause (e.g., "prompt lacked context" or "ignored recent behavior") using the LLM's self-reflection capabilities.
4. RefinePrompt Agent: Modifies the original prompt based on the inferred reasons, adding specific context or adjusting wording to address the identified failure.
5. Augment Agent: Generates 3–5 linguistic variations of the refined prompt to explore different phrasings while preserving the core meaning.
6. Select Agent: Uses the Upper Confidence Bound (UCB) algorithm to balance exploitation (choosing high-performing prompts) and exploration (testing new variations). It selects the optimal prompt for the next iteration.

Sustainability Integration: The agents are explicitly tasked with prioritizing eco-friendly attributes and bridging the gap between user intent and green product availability.

3. Key Contributions

• Novel Multi-Agent Framework: Introduces a self-correcting, iterative prompt optimization system where agents collaboratively diagnose failures and refine LLM prompts, moving beyond static prompting methods.
• Dual-Objective Optimization: Successfully balances the trade-off between recommendation accuracy (hitting the user's intent) and sustainability (promoting green products), while inherently reducing energy consumption by shortening the user journey and filtering candidates early.
• Bridging the Intent-Action Gap: The system actively interprets nuanced user behaviors to surface sustainable alternatives that traditional popularity-based models often overlook.
• Efficiency: By reducing the candidate pool from 100 to 20 before LLM processing and using a targeted multi-agent loop, the system minimizes unnecessary LLM inference calls.

4. Experimental Results

The system was evaluated on three real-world datasets: MovieLens-1M, Amazon Games, and Amazon Bundle (containing Electronic, Clothing, and Food categories).

• General Recommendation Performance (RQ1):
  • LLMGreenRec significantly outperformed baselines, including traditional methods (MostPop, SKNN), deep learning models (NARM, STAMP, GCE-GNN), and other LLM-based approaches (NIR, PO4ISR).
  • Key Metrics: On the Games dataset, it achieved a 40.7% relative improvement in NDCG@5 over the strong PO4ISR baseline. On the Bundle dataset, it showed a 26.6% improvement in HR@5.
• Sustainable Recommendation Performance (RQ2):
  • Tested specifically on the Bundle dataset for sustainable product selection.
  • Results: Achieved an HR@1 of 0.3950 (correct sustainable product ranked #1 in ~40% of cases) and an HR@5 of 0.5504 (appearing in top 5 over 55% of the time).
  • This demonstrates the system's ability to effectively integrate environmental criteria without sacrificing recommendation relevance.

5. Significance and Impact

• Strategic Shift for E-Commerce: The paper proposes a paradigm shift from purely conversion-driven algorithms to value-driven, responsible AI. It offers a technical solution to the "green gap" where consumer intent fails to translate into action.
• Environmental Impact: By optimizing the search process and reducing redundant interactions, the system directly contributes to lowering the digital carbon footprint of online retail.
• Scalability: The modular multi-agent design allows for continuous improvement and adaptation to new sustainability criteria without requiring full model retraining, making it a viable path for future ethical personalization systems.

In conclusion, LLMGreenRec demonstrates that leveraging the reasoning capabilities of LLMs within a structured multi-agent framework can create a recommender system that is not only more accurate but also environmentally conscious, effectively fostering a sustainable digital economy.