Carbon-aware re-ranking cuts e-commerce emissions at minimal engagement cost
New research uses LLMs to estimate product carbon footprints and re-rank recommendations for sustainability.
A new paper from researchers at the University of Oslo tackles the sustainability blind spot of e-commerce recommender systems. The authors propose a two-stage method: first, they estimate missing Product Carbon Footprint (PCF) labels using a retrieval-augmented pipeline that combines semantic similarity search, few-shot LLM prompting, and a nearest-neighbor fallback, transferring supervision from the Carbon Catalogue to a large unlabeled Amazon catalog. This allows them to assign carbon scores to products lacking life-cycle assessment data, a critical enabler for real-world deployment.
In the second stage, a post-hoc re-ranking strategy adjusts the final recommendation list by balancing predicted user engagement (derived from Amazon review interactions) against the estimated carbon footprint. The trade-off is controlled by a single parameter, lambda. Evaluating on three product categories—Home and Kitchen, Sports and Outdoors, Electronics—and three classic recommendation models (BPR, NeuMF, LightGCN), the team constructed Pareto frontiers. The findings reveal that substantial carbon reductions are possible with minimal engagement cost, though the exact trade-off varies by model and category, highlighting the need for domain-aware deployment.
- PCF labels inferred using retrieval-augmented generation (RAG) with few-shot LLM prompting and a Carbon Catalogue training set.
- Tested on three recommendation models (BPR, NeuMF, LightGCN) across three Amazon product categories (Home & Kitchen, Sports, Electronics).
- Pareto analysis shows substantial carbon reductions achievable with minimal engagement drop via a single tunable parameter (lambda).
Why It Matters
E-commerce platforms can reduce environmental impact without sacrificing user engagement or revenue.