Source-linked AI summary
What Price Fairness? Evaluating Energy - Fairness - Accuracy Trade-off in Recommender Systems
Abhirup Mitra, Oleg Lesota, Antonela Tommasel
TL;DR
Fairness evaluations of recommenders rarely quantify the computational costs of fairness interventions. This paper compares provider-side interventions and finds that their green cost is not universal.
Problem
Fairness-oriented recommender evaluations rarely quantify the computational costs of fairness interventions, although these methods may add computation at different pipeline stages.
Source: Introduction · p. 1Method
The study compares in-processing, graph-level reweighting, and post-processing interventions using recommendation quality, provider-exposure fairness, and energy measures.
Source: Methodology · pp. 1–2Results
The green cost of fairness is not universal, varying with intervention type, model, dataset, hardware, and measuring scope.
Source: Conclusions · p. 2Takeaways & Limitations
The findings support evaluating provider-side fairness through a joint accuracy–fairness–computational-cost lens.
Source: Conclusions · p. 2The study is limited by two datasets, a single popularity-based proxy for protected items, fixed fairness parameters, and approximate energy measures.
Source: Conclusions · p. 2
Abstract
from arXiv · showhide
Fairness-aware recommender systems aim to mitigate systematic imbalances in recommendation outcomes, including how visibility, relevance, and opportunities are distributed among users, items, and providers. However, these systems are usually evaluated in terms of accuracy and fairness alone, while their computational and environmental costs remain largely invisible. This omission matters because fairness interventions may affect the cost of recommendation in different ways. Training-time methods modify model optimization, post-processing methods add computation at inference time, and both may depend on the model, dataset, hardware, and deployment setting. We examine whether provider-side fairness in recommendation comes with a measurable green cost. We compare in-processing, graph-level reweighting and post-processing interventions across multiple models, two datasets, and two hardware settings. We measure recommendation quality, provider-side exposure, and energy consumption separately across training and inference stages. Our results show that the green cost of fairness is not uniform, post-processing shifts cost to repeated serving, while in-processing and graph-level methods avoid re-ranking overhead but vary substantially across models, datasets, and hardware. Findings call for evaluating fairness-aware recommendation as a three-way trade-off between accuracy, fairness, and computational cost.
Introduction.
Fairness-aware recommendation is usually evaluated for accuracy and fairness, leaving its computational and environmental costs largely unmeasured. This study examines provider-side fairness as a trade-off among recommendation quality, exposure-oriented fairness, and energy consumption.
Introduction
- Prior fairness evaluations emphasize accuracy–fairness trade-offs, while green recommender evaluations have mainly examined quality and energy rather than fairness-intervention costs.The study focuses on provider-side fairness: reducing disparities in visibility among item-provider groups.Source: Introduction · p. 1
- Fairness costs may vary by pipeline: training-time methods modify optimization, whereas post-processing adds inference computation, making a single green-cost estimate inadequate.The same intervention may be negligible in one pipeline and substantial in another.Source: Introduction · p. 1
- The study compares recommender configurations across quality, provider exposure fairness, and energy consumption to locate fairness costs and relate them to observed gains.It evaluates multiple recommenders and fairness strategies.Source: Introduction · p. 1
- Experiments use RecBole with fixed data splits, a random seed, and an evaluation protocol, measuring accuracy, exposure diagnostics, and energy after generating full-ranking top-K lists.Each configuration trains or loads a recommender and may apply a fairness intervention.Source: Methodology · p. 1
Methodology.
The study compares provider-exposure interventions across models, datasets, and hardware while measuring recommendation quality, exposure, and energy across training and inference. Its results show that cost and accuracy–fairness outcomes vary by intervention and setting.
Datasets and protected groups
- Experiments use ML-10M and Amazon Books with a popularity-based item-group proxy, five recommenders plus ML-10M ItemKNN, and chronological per-user 70/10/20 train/validation/test splits.Ratings are binarized at ≥4 and filtered to ten-core; the less-popular half is the protected long-tail group, an experimental proxy rather than a standard protected-status definition. Models use a fixed 10-epoch budget and Recall@10 validation.Source: Methodology · pp. 1–2
Fairness interventions
- The study compares a loss-level group-score-parity regularizer, LightGCN graph-level inverse-propensity edge reweighting, and FA*IR reranking with a protected-group prefix minimum of 0.3.The interventions target provider exposure but optimize different objectives and are treated as alternatives, not equivalent methods; FA*IR reranks a top-500 candidate pool.Source: Methodology · p. 2
Evaluation metrics
- Recommendation quality is nDCG@10, while fairness is assessed with catalog coverage and protected long-tail coverage to compare interventions beyond their own optimization objectives.Source: Methodology · p. 2
Energy measurement
- Energy is measured in kWh with CodeCarbon over three averaged runs, separately for fixed-budget training and inference, with reranking counted as additional serving-stage cost.Experiments run on a server and laptop with the same code, seed, split, and configuration; the study is limited by fixed fairness parameters, two datasets, one popularity-based proxy, and approximate energy measures.Source: Methodology · p. 2, Conclusions · p. 2
Inference-time fairness costs accumulate with repeated serving
- +6.5% (NCF) to +162.2% (MultiVAE) on ML-10M laptop and +1.1% (LightGCN) to +201.6% (MultiVAE) on server: FA*IR raised inference energy; ABs laptop increases ranged +39.2% (NCF) to +394.9% (LightGCN).Its full-pipeline increase was +0.1% (LightGCN on ML-10M server) to +36.2% (ItemKnn on ML-10M server), which does not capture recurring serving costs; training and inference energy should be reported separately.Source: Methodology · p. 2
In-processing fairness has model-specific energy costs
- Regularization raised ML-10M server full-pipeline energy +2.1% (BPR) to +5.1% (NCF), but LightGCN rose +192.9% on ABs laptop; graph IPW rose +79.4% on ABs laptop versus +0.9% on server.LightGCN regularization also increased energy +46.8% on ML-10M laptop and +72.7% on ML-10M server; IPW had little or no energy effect on ML-10M.Source: Methodology · p. 2
Energy costs must be interpreted together with fairness gains
- Interventions produced model- and dataset-dependent accuracy–exposure–energy trade-offs: some improved quality and exposure jointly, while others added energy or reduced ranking quality without clear gains.On ABs, in-processing improved exposure and nDCG@10 jointly for LightGCN at high energy cost and for MultiVAE at low cost; IPW added energy on ML-10M without clear metric gains.Source: Methodology · p. 2
Conclusions.
The results support including computational cost in fairness evaluation by examining where costs arise, how they accumulate, and what trade-offs accompany exposure gains. Fairness should not be assumed computationally free, but cost should not be used to dismiss fairness goals.
- Fairness-aware recommenders should be evaluated for where computational costs arise, how they accumulate, and what trade-offs accompany exposure gains.Source: Conclusions · p. 3
- Computational cost should neither be assumed absent nor used to dismiss fairness goals.Source: Conclusions · p. 3