Trading Engagement for Sustainability: Carbon-Aware Re-ranking for E-commerce Recommendations

Researchers demonstrate a carbon-aware recommendation system for e-commerce that infers missing Product Carbon Footprint data and applies post-hoc re-ranking to balance user engagement against sustainability. The framework achieves substantial carbon reductions with minimal engagement cost across multiple product categories and recommendation models.

This research addresses a critical gap in e-commerce sustainability: while recommender systems powerfully shape consumer purchasing behavior, carbon footprint data remains largely unavailable at scale. The authors tackle this through a clever two-stage approach combining data inference with algorithmic adjustment. Their retrieval-augmented pipeline leverages semantic similarity and few-shot LLM prompting to estimate PCF values for unlabeled products, transforming a data scarcity problem into a tractable optimization challenge. The subsequent re-ranking strategy operates as a lever rather than a constraint, allowing platforms to modulate environmental impact without abandoning relevance-driven recommendations entirely. Testing across three major product categories using Amazon review data reveals encouraging Pareto frontiers: meaningful carbon reductions emerge at minimal engagement sacrifice, suggesting sustainability and profitability need not be zero-sum propositions. However, the variation in achievable carbon headroom across models and categories highlights an underappreciated reality—sustainability outcomes depend heavily on both technical architecture and domain-specific constraints. For e-commerce platforms, this work translates to actionable optimization opportunities within existing recommendation pipelines. The practical significance lies not in theoretical elegance but in demonstrating that environmental considerations can integrate seamlessly into production systems without requiring fundamental redesigns. As consumer pressure for sustainable commerce intensifies and regulatory frameworks evolve, platforms that efficiently balance engagement and carbon footprint gain competitive advantage.

• →Carbon footprint inference via semantic similarity and LLM prompting enables sustainability scoring across unlabeled e-commerce catalogs at scale.
• →Post-hoc re-ranking with a tunable lambda parameter achieves measurable carbon reductions while preserving most user engagement across recommendation models.
• →Sustainability-engagement trade-offs vary significantly by product category and model choice, requiring context-specific optimization rather than one-size-fits-all approaches.
• →The framework demonstrates that environmental considerations can integrate into production recommendation systems without catastrophic engagement loss.
• →Offline evaluation using implicit feedback provides a practical methodology for testing sustainability-aware recommendation strategies before deployment.