Green Recommender Systems

• Green Recommender Systems are recommendation frameworks that minimize energy use, reduce carbon emissions, and steer users towards sustainable choices aligned with the UN SDGs.
• They employ multi-objective optimization, downsampling, and architectural innovations to balance predictive accuracy with significant environmental benefits.
• By integrating metrics for energy, carbon, and fairness, these systems ensure transparency, reproducibility, and accountability in both technical operations and outcome impacts.

Green recommender systems are recommender systems that integrate sustainability into both recommendation outcomes and recommender-system operation. In the narrow sense, they seek to minimize the energy consumption, carbon footprint, and broader computational impact of training, evaluation, deployment, and inference. In the broader sense, they also steer users toward environmentally sustainable, socially responsible, and economically sustainable decisions, often in explicit relation to the UN Sustainable Development Goals (SDGs) (Said, 10 Jan 2025, Felfernig et al., 30 Jul 2025, Zhou et al., 2024).

1. Conceptual scope and SDG alignment

The contemporary literature treats green recommender systems as a dual project. First, there is the environmental sustainability of the recommender system itself: the energy use, carbon emissions, and infrastructure burden created by model development and deployment. Second, there is the sustainability of what is recommended: whether recommendations induce lower-impact consumption, reduced waste, improved reuse, lower emissions, or broader social good. In a social-good framing, recommender systems are described not merely as personalization tools but as socio-technical infrastructures that can either support or hinder SDG 10 (Reduced Inequalities), SDG 12 (Responsible Consumption and Production), SDG 13 (Climate Action), and SDG 16 (Peace, Justice, and Strong Institutions) (Said, 10 Jan 2025).

This scope is broader than environmental accounting alone. Sustainability-oriented evaluation has been organized around three pillars—environmental, social, and economic—so that a recommender is not considered “green” merely because it is computationally efficient. A highly accurate system that promotes high-carbon products, amplifies harmful content, or entrenches producer concentration may be efficient in a narrow engineering sense while remaining unsustainable in outcome terms. Surveys of sustainable recommender systems therefore place green recommender systems alongside sustainable travel, food, buildings, industrial symbiosis, and social-impact applications, and treat sustainability as a property of both recommendation content and recommendation process (Felfernig et al., 30 Jul 2025, Felfernig et al., 2024).

Accountability is a further defining component. In this literature, transparency, explainability, reproducibility, auditability, and governance are not ancillary features; they are conditions under which sustainability claims become inspectable and contestable. This is especially salient in high-stakes domains, where fairness notions are contextual rather than universal, and where public perceptions of fair recommendation differ between entertainment, healthcare, finance, and justice (Said, 10 Jan 2025).

2. Sustainability objectives and formalization

A recurrent mathematical pattern in green recommender systems is the replacement of single-objective accuracy optimization with multi-objective or constrained optimization. In one conceptual formulation, a natural mathematical counterpart to the qualitative arguments is

$$L = \alpha L_{\text{accuracy}} + \beta L_{\text{carbon}} + \gamma L_{\text{fairness}},$$

or, equivalently, accuracy optimization under explicit energy or fairness constraints. This formulation is presented not as an explicit model from the source paper but as a natural mathematical counterpart to its qualitative trade-offs; its significance lies in making environmental and social costs first-class optimization targets rather than post hoc diagnostics (Said, 10 Jan 2025).

Concrete systems instantiate this pattern in domain-specific form. In sustainable tourism, SmartSustain ranks destinations with a weighted linear score

$$\text{score}(d,u,m) = w_E f_E(d) + w_P f_P(d) + w_S f_S(d,m) + w_I f_I(u,d) + w_A f_A(d),$$

where the terms encode transport emissions, destination popularity, seasonality, interest alignment, and personalization attributes such as walkability, air quality, or climate vulnerability. The design is explicitly “less is better” across dimensions, so that lower values correspond simultaneously to lower impact and better fit (Banerjee et al., 20 Oct 2025).

In sustainable food recommendation, GRAPE combines a standard recommendation loss with a sustainability-oriented auxiliary loss:

$$\mathcal{L} = \alpha \mathcal{L}_n + (1-\alpha)\mathcal{L}_g.$$

Here, $\mathcal{L}_n$ is a Bayesian Personalized Ranking objective, while $\mathcal{L}_g$ is a “Green Loss” that reorders items according to green indicators such as Environmental Impact Score, Nutritional Impact Score, and Healthy Meal Index. GRAPE further distinguishes between a non-prioritized Green Loss, where indicators are treated uniformly, and a prioritized Green Loss, where indicator-specific thresholds and ordering encode lexicographic trade-offs between sustainability criteria (Jing et al., 19 Aug 2025).

These formulations imply a family resemblance across domains. Travel systems combine emissions, crowding, and user interest; food systems combine taste, health, and environmental impact; industrial systems combine service quality and eco-efficiency; computationally green systems combine predictive quality and energy cost. A plausible implication is that green recommender systems are best understood not as a separate algorithmic class, but as a design regime in which sustainability enters through objectives, constraints, re-ranking, explanation, and governance.

3. Algorithmic and systems strategies for lowering computational footprint

The strongest empirical result in the area is that model choice matters enormously. One measurement study estimated that papers using deep learning algorithms emit approximately 42 times more CO$_2$ equivalents than papers using traditional methods, and reported an average of 3,297 kilograms of CO$_2$ equivalents for a single deep-learning-based paper (Vente et al., 2024). A later replication-style study retained the approximately 42-fold comparison and reported an average of 2,909 kilograms of CO$_2$ equivalents per deep-learning-based paper (Wegmeth et al., 16 Sep 2025). In parallel, a social-good analysis argued that a single deep learning recommender paper can exceed 3,000 kg CO$_2$e, and that the scale of deep-learning recommender publication in 2023 put the field’s emissions “on par with those of smaller nations” (Said, 10 Jan 2025).

This has led to explicit design advice against defaulting to deep learning. Neighborhood-based collaborative filtering and other “good old-fashioned AI” methods are repeatedly identified as competitive in many domains while consuming far less energy than large neural recommenders. The same literature argues for context-aware algorithm choice: in low-impact domains such as music or movies, marginal accuracy gains may not justify the carbon cost of complex models, whereas in high-impact domains such as fashion e-commerce, higher accuracy may be justified if it materially reduces shipping and returns (Said, 10 Jan 2025).

Dataset size has emerged as a separate green control variable. One line of work shows that downsampling can preserve substantial recommendation quality while materially reducing runtime and emissions. On sparse and skewed data such as Amazon Toys and Games, FunkSVD and BiasedMF maintained nDCG@10 within approximately 13% of full-data performance with up to a 50% reduction in training data, and a 50% downsampling scenario was associated with an estimated 27.4 kgCO$_2$e saving per algorithm–dataset workflow under the paper’s assumptions (Arabzadeh et al., 2024). A later, broader study reported that a 30% downsampling portion can reduce runtime by 52% compared to the full dataset and reduce carbon emissions by up to 51.02 KgCO2e during the training of a single algorithm on a single dataset, while some lower-complexity algorithms retained 81% of full-size performance using only 50% of the training set (Arabzadeh, 12 Feb 2025).

Benchmark work in content-heavy recommendation has shown that architectural reorganization can be as important as model family. In news recommendation, the only-encode-once (OLEO) paradigm caches content representations and avoids repeatedly encoding the same article, achieving competitive accuracy relative to end-to-end paradigms while delivering up to a 2992% improvement in sustainability metrics (Liu et al., 2024). By contrast, ensemble methods illustrate the opposite pattern: modest accuracy gains can be bought at disproportionate environmental cost. A systematic study of ensemble recommenders found 0.3–5.7% accuracy improvements at 19–2,549% more energy, with selective “Top Performers” ensembles consistently more efficient than exhaustive averaging strategies (Nitschke, 10 Nov 2025).

The same trade-off appears in model families often treated as intrinsically modern. In eco-aware graph recommenders, emissions rise with embedding size across NGCF, LightGCN, SimGCL, and LightGCL, and contrastive GNN variants frequently incur higher carbon cost than simpler graph propagation schemes for modest gains (Purificato et al., 2024). At industrial inference scale, the relevant question becomes not only which model is trained, but how much computation is allocated per request. GreenFlow addresses this by selecting model instances and item counts per stage of a cascade recommender through dynamic primal-dual optimization, and reports a 41% reduction in computation consumption in deployment, with approximately 5000 kWh of electricity saved and 3 tons of carbon emissions reduced per day while maintaining commercial revenue (Lu et al., 2023).

4. Measurement, benchmarks, and sustainability metrics

Green recommender systems depend on measurement infrastructure. EMERS, described as the first software library that simplifies measuring, monitoring, recording, and sharing the energy consumption of recommender systems experiments, uses smart power plugs to capture whole-machine power and provides both a Python interface and a Flask-based monitoring UI (Wegmeth et al., 2024). This whole-system approach is important because it measures CPU, GPU, memory, storage, and cooling together rather than relying only on component-level software counters.

A second development is the emergence of benchmark-level green metrics. In news recommendation, GreenRec defines carbon emissions as

$$\text{score}(d,u,m) = w_E f_E(d) + w_P f_P(d) + w_S f_S(d,m) + w_I f_I(u,d) + w_A f_A(d),$$0

where $$\text{score}(d,u,m) = w_E f_E(d) + w_P f_P(d) + w_S f_S(d,m) + w_I f_I(u,d) + w_A f_A(d),$$1 is device power, $$\text{score}(d,u,m) = w_E f_E(d) + w_P f_P(d) + w_S f_S(d,m) + w_I f_I(u,d) + w_A f_A(d),$$2 is runtime, and $$\text{score}(d,u,m) = w_E f_E(d) + w_P f_P(d) + w_S f_S(d,m) + w_I f_I(u,d) + w_A f_A(d),$$3 is carbon efficiency, and then combines recommendation performance and carbon into

$$\text{score}(d,u,m) = w_E f_E(d) + w_P f_P(d) + w_S f_S(d,m) + w_I f_I(u,d) + w_A f_A(d),$$4

This “AUC per Carbon Emission” metric formalizes the trade-off between predictive quality and carbon cost and underpins comparisons between end-to-end and OLEO paradigms (Liu et al., 2024). In industrial cascade recommenders, PFEC evaluation extends this logic by jointly tracking Performance, FLOPs, Energy, and Carbon rather than reducing efficiency to a single systems metric (Lu et al., 2023).

A representative metric vocabulary has now been proposed for sustainability-aware evaluation. The following metrics are presented as basic sustainability evaluation metrics for recommender systems (Felfernig et al., 30 Jul 2025).

Pillar	Representative metrics	What they quantify
Environmental	AvgCarFI, GIRec, ECRec, ECTrain, ECPDat, ESTRec, RTR, AvgLCI	Carbon intensity of recommended items, share of green items, inference/training energy, downstream energy savings, reuse, lifecycle impact
Social	Demographic parity, ListD$$\text{score}(d,u,m) = w_E f_E(d) + w_P f_P(d) + w_S f_S(d,m) + w_I f_I(u,d) + w_A f_A(d),$$5, SER$$\text{score}(d,u,m) = w_E f_E(d) + w_P f_P(d) + w_S f_S(d,m) + w_I f_I(u,d) + w_A f_A(d),$$6, ACC$$\text{score}(d,u,m) = w_E f_E(d) + w_P f_P(d) + w_S f_S(d,m) + w_I f_I(u,d) + w_A f_A(d),$$7, HIER, HIRec	Fairness, diversity, serendipity, accessibility, harmful exposure, health improvement
Economic	LBPR, AvgLoyalty, PEF	Local business support, long-term customer loyalty, producer exposure fairness

These metrics shift evaluation from a purely behavioral or predictive frame to a consequence-oriented frame. AvgCarFI measures the average carbon footprint of recommended items; GIRec measures the share of recommended items labeled as green; ECRec and ECTrain measure recommender energy per recommendation and per training epoch; ESTRec captures downstream energy savings caused by recommendations; and AvgLCI aggregates life-cycle impact. The same work argues that such metrics should complement rather than replace accuracy, precision, recall, NDCG, or satisfaction (Felfernig et al., 30 Jul 2025).

Despite this progress, standardization remains incomplete. Several papers explicitly note the absence of widely adopted protocols for measuring recommender energy use, carbon footprint, and SDG contributions, which makes cross-paper comparison difficult and leaves large “measurement gaps” in the field (Said, 10 Jan 2025).

5. Application domains and domain-specific designs

Tourism has become a central testbed for green recommendation because sustainability trade-offs are tangible and multi-dimensional. SmartSustain rethinks city-trip recommendation around CO$$\text{score}(d,u,m) = w_E f_E(d) + w_P f_P(d) + w_S f_S(d,m) + w_I f_I(u,d) + w_A f_A(d),$$8e emissions, destination popularity, seasonality, and personalized interests, and uses interactive city cards, dynamic trade-off banners, and real-time impact feedback to make trade-offs visible rather than hidden. In a preliminary user study with 21 participants, 45% rated the system “Good,” 40% rated it “Excellent,” and 50% considered it “Highly effective” at promoting eco-friendly choices (Banerjee et al., 20 Oct 2025). The domain significance is not merely that tourism has a large footprint, but that recommender interfaces can operationalize “gentle steering” without becoming prescriptive.

Food recommendation provides the clearest example of multi-dimensional green optimization. GRAPE models user preference sequences together with environmental and health indicators and introduces Green Loss functions to prioritize sustainable food options aligned with users’ evolving preferences. On a real-world dataset of 6,290 users, 74,324 recipes, and 316,116 interactions, it reported HR@20 of 0.0472 compared with 0.0437 for MSSR and NDCG@20 of 0.0279 compared with 0.0266, while also improving environmental and health-related recommendation characteristics (Jing et al., 19 Aug 2025). More broadly, sustainable food recommendation has been framed as simultaneously addressing taste, health, and environmental impact rather than treating them as separate systems (Zhou et al., 2024).

Sustainable e-commerce has recently incorporated LLMs and multi-agent architectures. LLMGreenRec combines a cross-encoder reranker that filters a 100-item candidate set down to 20 items with a six-agent LLM pipeline for evaluation, error detection, prompt refinement, augmentation, and prompt selection. In a sustainable product recommendation setting on the Bundle dataset, it reported HR@1 of 0.3950, HR@5 of 0.5504, NDCG@1 of 0.3950, and NDCG@5 of 0.4715 for sessions where the ground-truth target was a sustainable product (Nguyen et al., 11 Mar 2026). The paper explicitly frames the system as “green” both because it prioritizes eco-friendly products and because its two-stage design aims to reduce unnecessary interactions and digital carbon footprint.

Buildings and energy systems represent a different class of green recommender system, one in which the recommendation target is often an action, schedule, or operational strategy rather than a conventional catalog item. A survey of energy-efficiency recommendation in buildings organizes systems by strategy versus action recommendation, collaborative filtering, context awareness, Rasch-based persuasion, multi-agent systems, and deep reinforcement learning, all with the objective of reducing energy use, costs, and CO$$\text{score}(d,u,m) = w_E f_E(d) + w_P f_P(d) + w_S f_S(d,m) + w_I f_I(u,d) + w_A f_A(d),$$9 emissions while preserving comfort (Himeur et al., 2021). In renewable energetic communities, recommender logic has even been used for user admission and profiling: candidate users are clustered by consumption profiles so that community composition better matches renewable production and storage, thereby maximizing locally shared energy and minimizing management complexity (Guzzi et al., 2022). These applications show that green recommendation is not restricted to “eco-label reranking”; it can structure infrastructure, operations, and collective resource allocation.

6. Accountability, fairness, and open research problems

Green recommender systems inherit classical recommender concerns, but under sustainability constraints they become sharper. Fairness is repeatedly described as contextual: there is no single universally correct definition, and notions such as demographic parity or equalized odds may be appropriate in some domains but not in others. In high-stakes settings, recommenders should be audited for demographic disparities, and fairness interventions must be adapted to the domain and to users’ perceptions of what counts as fair (Said, 10 Jan 2025).

Accountability mechanisms are correspondingly central. Reproducibility is treated as a form of accountability because it enables validation, critique, and audit; transparency and explainability allow stakeholders to understand why an option was recommended and what its implications are; and governance structures are needed to monitor both bias and environmental impact over time. This includes versioning datasets, code, and configurations; documenting experimental and deployment settings; providing accessible explanations; and enabling corrective action when harm is identified (Said, 10 Jan 2025).

Several open problems recur across the literature. One is data availability: many catalogs lack carbon, lifecycle, repairability, or sustainability labels, which constrains metrics such as AvgCarFI, GIRec, or AvgLCI and can introduce systematic bias through low “LabelCoverage” (Felfernig et al., 30 Jul 2025). A second is scenario dependency: the best downsampling strategy, model family, or fairness intervention depends on sparsity, domain impact, user population, and evaluation protocol, which makes universal green heuristics difficult (Arabzadeh, 12 Feb 2025). A third is research culture and incentives: publication pressure and novelty incentives still favor larger, more complex, higher-emission models, while institutional incentives for low-carbon experimentation remain weak (Said, 10 Jan 2025).

Future directions are increasingly convergent. Proposed agendas include longitudinal studies of behavioral change rather than short-term lab effects; integration with live transport, accommodation, and product APIs to compute sustainability signals dynamically; explicit Pareto optimization for accuracy, utility, and sustainability; cost-aware and uncertainty-aware ensemble selection; smaller or more specialized LLMs for sustainable recommendation; and community standards for energy and carbon reporting comparable to what Green AI has begun to establish in other fields (Banerjee et al., 20 Oct 2025, Nguyen et al., 11 Mar 2026, Wegmeth et al., 16 Sep 2025). Taken together, these directions suggest that green recommender systems are evolving from a set of isolated interventions into a general framework for designing recommendation ecosystems that are computationally efficient, outcome-aware, and accountable to environmental and social consequences.