Pith. sign in

REVIEW 3 major objections 6 minor 55 references

GRACE claims that fine-tuning a pretrained recommender with a differentiable Gumbel-softmax green loss plus gradient projection yields greener top-K lists while preserving accuracy and adding no inference-time cost.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

A fine-tuning framework uses soft sorting and gradient projection to make pretrained food recommenders greener without retraining or reranking.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection The core differentiable ranking relaxation in GRACE is algebraically wrong, so the paper's central mechanism doesn't do what it claims; it should be desk rejected, not sent to review. the 3 major comments →

arxiv 2607.22341 v1 pith:HIIIU6TE submitted 2026-07-24 cs.IR

Bringing GRACE to Recommendation: Fine-Tuning for Sustainable and Accurate Personalization

Yibowen Zhao , Yinan Zhang , Ning Liu , Lizhen Cui , Chunyan Miao This is my paper
classification cs.IR
keywords sustainable recommendationgreen recommender systemsfine-tuningGumbel-softmaxdifferentiable sortinggradient projectionmulti-objective optimizationfood recommendation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that making environmental and health goals directly differentiable is the right way to turn an existing recommender into a 'green' one, without the cost of retraining from scratch or the latency of an extra reranking pass. It proposes GRACE, which relaxes discrete top-K sustainability labels with a Gumbel-softmax soft permutation matrix, defines a listwise green loss over the relaxed ranks, and uses gradient projection to strip from the green update any component that would push against the accuracy gradient. If GRACE is correct, platforms can upgrade a deployed model by fine-tuning only, with the green lists emitted naturally from the base model's own scores at serving. Experiments across four sequential backbones on two food datasets report higher environmental and health indices (or lower carbon footprints) with accuracy roughly preserved, and sometimes improved accuracy, which the paper attributes to the green signal acting as a regularizer. The practical appeal is simple: the green-accuracy trade-off is controlled by a single projection knob, so operators can dial sustainability up or down without retraining.

Core claim

The central claim is that the expected sustainability score over the top-N relaxed ranks, computed from a Gumbel-perturbed soft permutation matrix, is a valid differentiable proxy for the discrete green objective, and that the proxy is optimizable end-to-end with bounded gradients. Because the relaxation is consistent with the discrete permutation as temperature approaches zero, GRACE maintains that optimizing the relaxed green loss directly shifts the model's score distribution toward greener items at the true top-K positions. A gradient projection step then removes the component of the green update that aligns with the accuracy update, so first-order interference is avoided, and a norm cap

What carries the argument

The load-bearing object is a NeuralSort-style soft permutation matrix built from Gumbel-perturbed normalized scores. Eq. (4) defines an iterative score array and Eq. (5) turns it into N softmax rows intended to approximate the rank positions; this gives, for each rank n, a differentiable green score DGS(n) = Σ_i P̂_{n,i} e_i^r and a differentiable hit DH(n) = Σ_i P̂_{n,i} y_i. These two signals feed the green loss and an NDCG-style accuracy loss, and the gradient projection couples them. The projection formula g_proj = g_green - ((g_green^T g_rec)/(||g_rec||^2)) g_rec removes the component in the direction of the accuracy gradient, and the final update scales g_proj by min(1, α||g_rec||/||g_

Load-bearing premise

That the soft permutation matrix rows genuinely select the items ranked at the corresponding positions — if row n's highest-probability item is not the nth largest score, then the differentiable green loss is computing expected greenness at the wrong ranks and the method optimizes the wrong objective.

What would settle it

On a toy problem with three items and perturbed scores, say (3, 2, 1), and temperature 0.9, compute the soft permutation via Eq. (5) and check whether the argmax of row 2 is the item with score 2. A more thorough check is to compare the gradient of the relaxed green loss against the finite-difference gradient of the exact discrete top-K green score; any sign mismatch indicates the relaxation is not a faithful surrogate.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Inference-time complexity remains exactly that of the base recommender; no extra sorting pass, so large-catalog low-latency services incur zero added latency.
  • A single scalar projection ratio α lets operators sweep the sustainability-accuracy frontier on a deployed model, with no retraining and no changes to serving code.
  • The differentiable formulation extends to any item-side scalar signal, so the same framework can be applied to health, fairness, or diversity labels without paired supervision.
  • Because the green objective is optimized directly in parameter space, the resulting green lists come from the model's own scores and are stable per user, unlike heuristic reranking that can generalize poorly.
  • The paper's case study shows that top-ranked items remain stable while lower-ranked positions are swapped for greener items, indicating the preference anchor is doing its job.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same Gumbel-softmax relaxation could be reused to make any discrete list metric (e.g., exposure fairness, recency, or price ceilings) directly optimizable, and the projection mechanism would apply verbatim when the new objective conflicts with accuracy.
  • The reported accuracy improvements in several configurations suggest that the green signal may act as a form of regularization against overfitting to popularity or spurious signals; that hypothesis is testable by comparing validation loss trajectories of GRACE-tuned vs. pretrained models.
  • The consistency claim is the key thing to verify: the soft permutation rows must actually map to the true ranking positions, and a small-scale numeric check on the argmax of each row would settle whether the green loss optimizes the intended positions.
  • A practical extension is to make the projection ratio user-specific, e.g., derived from a user's prior interactions with green items, which the paper lists as future work; our inference is that this could flatten the global trade-off and improve both objectives on average.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper proposes GRACE, a fine-tuning framework that injects item-level sustainability labels (EIS/NIS/HMI on GreenRec, CO2-oriented greenness on RecipeEmission) into pretrained recommender models without retraining from scratch or adding inference-time reranking. The method has two components: a differentiable green loss built from a Gumbel-perturbed NeuralSort-style soft permutation matrix, and a gradient-projection mechanism that removes the component of the green gradient aligned with the recommendation gradient. Experiments on four backbones and two datasets report green-metric improvements with mostly preserved accuracy, plus ablations, hyperparameter analyses, and a significance diagram. The central claim is that GRACE enables direct optimization of top-K green outcomes while preserving personalization accuracy.

Significance. The problem is well motivated and practically important: integrating sustainability signals into existing recommenders without full retraining or inference-time overhead would be a useful contribution. The experimental design is a genuine strength: multiple backbones, two datasets, reranking and multi-objective baselines, ablations, and a statistical significance diagram. However, the paper's main mathematical mechanism — the differentiable soft permutation construction in Eqs. (4)–(5) — is incorrect. Because the green loss L_green in Eq. (13) is built on the resulting DGS(n) terms, the method does not optimize the stated average top-N green score. The reported gains therefore cannot be attributed to the claimed direct optimization of the green criterion. If the construction were repaired and the experiments re-run, the framework might be viable, but as written the central claim is unsupported.

major comments (3)
  1. [§III-B1, Eqs. (4)–(5), Proposition 1] The proposed soft permutation construction is not a valid NeuralSort-style relaxation. For V=3 and tilde{s}=(3,2,1), Eq. (4) gives Ahat=(-9,0,3). With Eq. (5), row 2 has logits (9,0,-3)/tau, so its argmax is the largest-score item, not the second-ranked item; rows 1 and 3 also peak at the top item. Consequently, the proof's assertion that 'the corresponding logit in row n is the unique maximum' is false. Proposition 1 does not hold, and DGS(n) in Eq. (7) is not the expected green score at rank n. Since L_green in Eq. (13) sums these DGS(n) terms, the differentiable green objective does not optimize the stated average top-N green score; it concentrates probability mass on the highest-scoring item. The central claim of direct green-criterion optimization is therefore unsupported.
  2. [§III-B1, Proposition 2] The bounded-gradient result bounds gradients with respect to the perturbed scores tilde{s}, not with respect to the model parameters theta. Training updates use nabla_theta L through the recommender f(theta); boundedness in score space does not imply bounded or stable parameter-space gradients without an additional control on the model Jacobian. As stated, the proposition does not justify the claimed stable fine-tuning behavior.
  3. [§IV-A4 and §IV-B2] Hyperparameters (tau, N, alpha) are tuned on validation NDCG@20 rather than on any sustainability metric. Since the paper's claim is about green improvements and a controllable accuracy-green trade-off, tuning only on accuracy leaves open whether the reported point is favorable on the green-accuracy Pareto frontier. A green-metric-validated hyperparameter sweep is needed to support the trade-off claims.
minor comments (6)
  1. [§III-B1] Eq. (4) is described as 'NeuralSort-style,' but it is not the standard NeuralSort formula. If the standard relaxation is intended, the paper should present it correctly and cite [39]; if a new formula is intended, its properties need to be established rather than assumed.
  2. [Eq. (13) and Eq. (14)] The notation Phat^{(b)}_{n,i} should explicitly indicate dependence on the sampled Gumbel noise; DGS and DH are single-sample stochastic estimates, not expectations over the Gumbel distribution. Clarify this in the text.
  3. [Table IV] The R+G" row is numerically identical to R, which is expected because a non-differentiable additive term has no gradient. The conclusion that 'differentiability is essential' is not supported by this ablation; a fair baseline would use a differentiable but non-relaxed surrogate or a straight-through estimator.
  4. [§IV-B4] The case study draws conclusions from a single user. Consider reporting aggregate statistics over a sample of users to support the claim that GRACE 'can improve both health alignment and relevance for some users.'
  5. [Eq. (18)] The projection in Eq. (18) is the PCGrad update (projecting one task gradient onto the normal plane of another). If this is intended, PCGrad should be cited and compared explicitly; otherwise the novelty of the gradient-projection component should be clarified.
  6. [§IV-B1] The CD diagram is said to summarize 'average ranks across datasets'; clarify whether the ranks are averaged over all metrics and datasets, and how ties are handled.

Circularity Check

0 steps flagged

No significant circularity: the differentiable green loss directly optimizes external sustainability labels, and self-citations are not load-bearing.

full rationale

GRACE's central derivation is transparent rather than circular. L_green (Eq. 13) maximizes the softmax-weighted average of item-level external sustainability scores (Eqs. 5, 7), and the evaluation metrics EIS/NIS/HMI and CO2 are averages of those same external labels over top-K outputs. That is direct objective optimization, not a fitted parameter renamed as a prediction; the paper does not fit a constant to test green outcomes and then report it as a discovered result. Hyperparameters are selected on validation NDCG rather than on green metrics, so the green improvements are not forced by tuning. The paper's self-citations, notably GRAPE [17] and GreenRec [34], supply an external dataset and a pretrained backbone and are not used to justify the core mathematical claims. The projection formula (Eq. 18) may resemble PCGrad and lack a citation, but that is an attribution issue, not circularity. The skeptic's concern about Eqs. (4)-(5)—whether each row of the soft permutation matrix peaks at the true n-th ranked item—is a mathematical correctness/validity objection, not a circularity, because the proposition's failure would not make the derivation equivalent to its inputs by construction. Under the stated circularity standard, no circular step is exhibited, so the score is 0.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The central method rests on one invalid mathematical premise (the soft-permutation identity in Eq. 4), one domain assumption about the validity of the green labels, and the standard PCGrad-style projection heuristic. Hyperparameters (τ, N, α, green-score weighting) are hand-set or tuned on validation NDCG, not fit to the reported green metrics.

free parameters (5)
  • Gumbel-softmax temperature τ = 0.9
    Hyperparameter chosen on validation NDCG@20 (§IV-A4); controls smoothness of the soft ranking.
  • Top-N positions in differentiable loss = 10
    Number of soft-permutation rows used in L_green and L_rec (§IV-A4); GPU memory grows rapidly with N (Fig. 4c).
  • Projection ratio α = not reported
    Caps the green update magnitude relative to the recommendation gradient (Eq. 19); the paper states α∈[0,1] but never gives the value used in the reported tables.
  • Scalar green score weights = uniform 1/3 each (EIS, NIS, HMI)
    On GreenRec the training target is the unweighted average of aligned EIS/NIS/HMI (§IV-A4); hand-chosen definition of the green objective.
  • Reranking baseline hyperparameters = FHFRS λ=0.1, γ=0.7; CFARS δ=0.8 (GreenRec); FHFRS λ=0.05, γ=0.95; CFARS δ=0.95 (RecipeEmission)
    Fixed for the reported comparisons (§IV-A4); influence the green/accuracy trade-off of baselines.
axioms (5)
  • ad hoc to paper The soft-permutation construction (Eqs. 4–5) has the argmax of row i at the i-th-ranked item
    Required for DGS(n) to mean 'expected green score at rank n'; false under the published algebra (row 2 of the softmax selects the top item for scores (3,2,1)). Invoked in Proposition 1.
  • domain assumption Item-level sustainability labels (EIS/NIS/HMI, CO2-derived greenness) are externally valid and suitable as training signals
    Taken from GreenRec and RecipeEmission; the paper treats them as fixed, objective item properties (§III-A).
  • domain assumption Orthogonal projection of the green gradient onto the recommendation gradient's normal plane reduces harmful interference
    Justifies Eq. (18); a first-order heuristic with no convergence guarantee and no theoretical support beyond the projection definition (§III-B2).
  • standard math L2 score normalization preserves item ordering and stabilizes gradients
    Positive common scaling leaves rank order unchanged; used in Eq. (1).
  • standard math Gumbel-perturbed scores are distinct almost surely, so the discrete permutation is unique
    Needed for Proposition 1; standard Gumbel-continuity argument.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Bringing GRACE to Recommendation: Fine-Tuning for Sustainable and Accurate Personalization." pith.science (2026). https://pith.science/paper/HIIIU6TE

@misc{pith2026260722341,
  author       = {Pith},
  title        = {Pith review of: Bringing GRACE to Recommendation: Fine-Tuning for Sustainable and Accurate Personalization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HIIIU6TE}},
  note         = {Machine review of arXiv:2607.22341}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Growing concern about environmental sustainability (e.g., reducing carbon emissions and resource use) and public health has motivated ``green'' recommender systems that steer users toward more eco-friendly and healthier choices. However, many existing green recommendation approaches require training new models from scratch, incurring substantial computational and energy costs. Reranking-based methods, meanwhile, introduce an additional sorting stage at inference, increasing latency and computational cost. In this work, we propose GRACE (Green Recommendation via Adaptive Conflict-rEsolution), a fine-tuning framework that integrates item-level sustainability signals (e.g., eco-scores or health indices) into pretrained recommendation models. Since these green values are usually discrete and non-differentiable, existing methods often rely on pairwise comparisons to promote greener items. GRACE instead introduces a differentiable approximation that enables direct optimization of the green criterion. To balance sustainability and personalization quality, GRACE further employs a gradient projection mechanism to mitigate conflicts between the green objective and the accuracy objective during fine-tuning. Experiments on real-world datasets demonstrate that GRACE improves sustainability-oriented recommendation outcomes while generally preserving recommendation accuracy through a controllable preference-anchored update mechanism.

Figures

Figures reproduced from arXiv: 2607.22341 by Chunyan Miao, Lizhen Cui, Ning Liu, Yibowen Zhao, Yinan Zhang.

Figure 1
Figure 1. Figure 1: The GRACE framework. GRACE fine-tunes pretrained recommendation models by jointly optimizing for accuracy [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Distribution of Sustainability Indicators in the Two [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Hyperparameter Analysis in GRACE Table IV: Ablation study of GRACE under different loss configurations. Cells shaded in greener tones indicate better performance, while redder tones denote worse outcomes. Cell colors indicate relative performance within each metric column (greener denotes better results and redder denotes worse ones). Loss Hit ↑ NDCG ↑ EIS ↓ NIS ↑ HMI↑ R 0.0572 0.0260 95.39 37.77 46.65 G 0… view at source ↗
Figure 5
Figure 5. Figure 5: Case study of GRACE on GreenRec. Top: Top-5 recipes from pretrained SASRec. Bottom: Top-5 recipes after GRACE fine-tuning. “Picked” indicates that the user interacted with the corresponding item. baselines across the evaluated recommendation models. The primary goal is to improve sustainability while avoiding accu￾racy loss. Across most settings, GRACE achieves this goal by improving sustainability with co… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

55 extracted references · 5 linked inside Pith

  1. [1]

    A novel healthy and time-aware food recommender system using attributed community detection,

    M. Rostami, V . Farrahi, S. Ahmadian, S. M. J. Jalali, and M. Oussalah, “A novel healthy and time-aware food recommender system using attributed community detection,”Expert Systems with Applications, vol. 221, p. 119719, 2023

    2023

  2. [2]

    Self-supervised learning for recommender systems: A survey,

    J. Yu, H. Yin, X. Xia, T. Chen, J. Li, and Z. Huang, “Self-supervised learning for recommender systems: A survey,”IEEE Transactions on Knowledge and Data Engineering, vol. 36, no. 1, pp. 335–355, 2023

    2023

  3. [3]

    E-commerce and consumer behavior: A review of ai- powered personalization and market trends,

    M. A. Raji, H. B. Olodo, T. T. Oke, W. A. Addy, O. C. Ofodile, and A. T. Oyewole, “E-commerce and consumer behavior: A review of ai- powered personalization and market trends,”GSC advanced research and reviews, vol. 18, no. 3, pp. 066–077, 2024

    2024

  4. [4]

    A survey on accuracy-oriented neural recommendation: From collaborative filtering to information-rich recommendation,

    L. Wu, X. He, X. Wang, K. Zhang, and M. Wang, “A survey on accuracy-oriented neural recommendation: From collaborative filtering to information-rich recommendation,”IEEE transactions on knowledge and data engineering, vol. 35, no. 5, pp. 4425–4445, 2022

    2022

  5. [5]

    Green recommender systems: Optimizing dataset size for energy-efficient algorithm performance,

    A. Arabzadeh, T. Vente, and J. Beel, “Green recommender systems: Optimizing dataset size for energy-efficient algorithm performance,” in International Workshop on Recommender Systems for Sustainability and Social Good. Springer, 2024, pp. 73–82

    2024

  6. [6]

    Green recommender systems: A call for attention,

    J. Beel, A. Said, T. Vente, and L. Wegmeth, “Green recommender systems: A call for attention,” inACM SIGIR Forum, vol. 58, no. 2. ACM New York, NY , USA, 2025, pp. 1–5

    2025

  7. [7]

    A novel healthy food recommendation to user groups based on a deep social community detection approach,

    M. Rostami, K. Berahmand, S. Forouzandeh, S. Ahmadian, V . Farrahi, and M. Oussalah, “A novel healthy food recommendation to user groups based on a deep social community detection approach,”Neurocomput- ing, vol. 576, p. 127326, 2024

    2024

  8. [8]

    A hybrid recommendation algorithm for green food based on review text and review time,

    H. Geng, W. Peng, X. G. Shan, and C. Song, “A hybrid recommendation algorithm for green food based on review text and review time,”CyTA- Journal of Food, vol. 21, no. 1, pp. 481–492, 2023

    2023

  9. [9]

    Reducing food’s environmental impacts through producers and consumers,

    J. Poore and T. Nemecek, “Reducing food’s environmental impacts through producers and consumers,”Science, vol. 360, no. 6392, pp. 987–992, 2018

    2018

  10. [10]

    Global epidemiological patterns in the burden of main non-communicable diseases, 1990–2019: Relationships with socio-demographic index,

    J. Bai, J. Cui, F. Shi, and C. Yu, “Global epidemiological patterns in the burden of main non-communicable diseases, 1990–2019: Relationships with socio-demographic index,”International journal of public health, vol. 68, p. 1605502, 2023

    1990

  11. [11]

    Reducing climate change impacts from the global food system through diet shifts,

    Y . Li, P. He, Y . Shan, Y . Li, Y . Hang, S. Shao, F. Ruzzenenti, and K. Hubacek, “Reducing climate change impacts from the global food system through diet shifts,”Nature Climate Change, vol. 14, no. 9, pp. 943–953, 2024

    2024

  12. [12]

    Simple dietary substitutions can reduce carbon footprints and improve dietary quality across diverse segments of the us population,

    A. H. Grummon, C. J. Lee, T. N. Robinson, E. B. Rimm, and D. Rose, “Simple dietary substitutions can reduce carbon footprints and improve dietary quality across diverse segments of the us population,”Nature food, vol. 4, no. 11, pp. 966–977, 2023

    2023

  13. [13]

    Incorporating price into recommendation with graph convolutional networks,

    Y . Zheng, C. Gao, X. He, D. Jin, and Y . Li, “Incorporating price into recommendation with graph convolutional networks,”IEEE Transactions on Knowledge and Data Engineering, vol. 35, no. 2, pp. 1609–1623, 2021

    2021

  14. [14]

    Cross-domain knowledge graph chiasmal embedding for multi-domain item-item recommenda- tion,

    J. Liu, W. Huang, T. Li, S. Ji, and J. Zhang, “Cross-domain knowledge graph chiasmal embedding for multi-domain item-item recommenda- tion,”IEEE Transactions on Knowledge and Data Engineering, vol. 35, no. 5, pp. 4621–4633, 2022

    2022

  15. [15]

    Recommender systems for sustainability: overview and research issues,

    A. Felfernig, M. Wundara, T. N. T. Tran, S. Polat-Erdeniz, S. Lubos, M. El Mansi, D. Garber, and V .-M. Le, “Recommender systems for sustainability: overview and research issues,”Frontiers in big Data, vol. 6, p. 1284511, 2023

    2023

  16. [16]

    Bundle recommendation and generation with graph neural networks,

    J. Chang, C. Gao, X. He, D. Jin, and Y . Li, “Bundle recommendation and generation with graph neural networks,”IEEE Transactions on Knowledge and Data Engineering, vol. 35, no. 3, pp. 2326–2340, 2021

    2021

  17. [17]

    Bites of tomorrow: Personalized recommendations for a healthier and greener plate,

    J. Jing, Y . Zhang, and C. Miao, “Bites of tomorrow: Personalized recommendations for a healthier and greener plate,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 11, 2025, pp. 11 897–11 905

    2025

  18. [18]

    Towards carbon footprint-aware recommender systems for greener item recommenda- tion,

    R. Kalisvaart, M. Mansoury, A. Hanjalic, and E. Isufi, “Towards carbon footprint-aware recommender systems for greener item recommenda- tion,”ACM Transactions on Recommender Systems, 2025

    2025

  19. [19]

    Towards health-aware fairness in food recipe recommendation,

    M. Rostami, M. Aliannejadi, and M. Oussalah, “Towards health-aware fairness in food recipe recommendation,” inProceedings of the 17th ACM Conference on Recommender Systems, 2023, pp. 1184–1189

    2023

  20. [20]

    Meta-learn to unlearn: Enhanced exact machine unlearning in recommendation systems with meta-learning,

    A. Alshabanah, K. Balasubramanian, and M. Annavaram, “Meta-learn to unlearn: Enhanced exact machine unlearning in recommendation systems with meta-learning,”Proceedings on Privacy Enhancing Tech- nologies, 2025

    2025

  21. [21]

    Scheduling on a budget: Avoiding stale recommendations with timely updates,

    R. Verachtert, O. Jeunen, and B. Goethals, “Scheduling on a budget: Avoiding stale recommendations with timely updates,”Machine Learn- ing with Applications, vol. 11, p. 100455, 2023

    2023

  22. [22]

    Preference-consistent knowledge distillation for recommender system,

    Z. Zhu and W. Zhang, “Preference-consistent knowledge distillation for recommender system,”IEEE Transactions on Knowledge and Data Engineering, vol. 37, no. 4, pp. 2071–2084, 2025

    2071

  23. [23]

    From clicks to carbon: the environmental toll of recommender systems,

    T. Vente, L. Wegmeth, A. Said, and J. Beel, “From clicks to carbon: the environmental toll of recommender systems,” inProceedings of the 18th ACM Conference on Recommender Systems, 2024, pp. 580–590

    2024

  24. [24]

    Recommen- dation systems: An insight into current development and future research challenges,

    M. Marcuzzo, A. Zangari, A. Albarelli, and A. Gasparetto, “Recommen- dation systems: An insight into current development and future research challenges,”IEEE Access, vol. 10, pp. 86 578–86 623, 2022

    2022

  25. [25]

    Differentiable ranking metric using relaxed sorting for top-k recommendation,

    H. Lee, S. Cho, Y . Jang, J. Kim, and H. Woo, “Differentiable ranking metric using relaxed sorting for top-k recommendation,”IEEE Access, vol. 9, pp. 114 649–114 658, 2021

    2021

  26. [26]

    Item: Improving training and evaluation of message-passing based gnns for top-k recommendation,

    Y . Karmim, E. Ramzi, R. Fournier-S’Niehotta, and N. Thome, “Item: Improving training and evaluation of message-passing based gnns for top-k recommendation,”arXiv preprint arXiv:2407.07912, 2024

    Pith/arXiv arXiv 2024

  27. [27]

    Bayesian negative sampling for recommendation,

    B. Liu and B. Wang, “Bayesian negative sampling for recommendation,” in2023 IEEE 39th International Conference on Data Engineering (ICDE). IEEE, 2023, pp. 749–761

    2023

  28. [28]

    Enhanced bayesian personalized ranking for robust hard negative sampling in recommender systems,

    K. Shi, J. Zhang, L. Fang, W. Wang, and B. Jing, “Enhanced bayesian personalized ranking for robust hard negative sampling in recommender systems,”arXiv preprint arXiv:2403.19276, 2024

    Pith/arXiv arXiv 2024

  29. [29]

    Meta auxiliary learning for top-k recommendation,

    X. Li, C. Ma, G. Li, P. Xu, C. H. Liu, Y . Yuan, and G. Wang, “Meta auxiliary learning for top-k recommendation,”IEEE Transactions on Knowledge and Data Engineering, vol. 35, no. 10, pp. 10 857–10 870, 2022

    2022

  30. [30]

    Advancing sustainability via recommender systems: A survey,

    X. Zhou, L. Zhang, H. Zhang, Y . Zhang, X. Zhang, J. Zhang, and Z. Shen, “Advancing sustainability via recommender systems: A survey,” arXiv preprint arXiv:2411.07658, 2024

    Pith/arXiv arXiv 2024

  31. [31]

    Xsimgcl: Towards extremely simple graph contrastive learning for recommenda- tion,

    J. Yu, X. Xia, T. Chen, L. Cui, N. Q. V . Hung, and H. Yin, “Xsimgcl: Towards extremely simple graph contrastive learning for recommenda- tion,”IEEE Transactions on Knowledge and Data Engineering, vol. 36, no. 2, pp. 913–926, 2023

    2023

  32. [32]

    One person, one model, one world: Learning continual user representation without forgetting,

    F. Yuan, G. Zhang, A. Karatzoglou, J. Jose, B. Kong, and Y . Li, “One person, one model, one world: Learning continual user representation without forgetting,” inProceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, 2021, pp. 696–705

    2021

  33. [33]

    Intervention strategies to promote healthy and sustainable food choices among parents with lower and higher socioeconomic status,

    M. V os, B. Deforche, A. Van Kerckhove, N. Michels, M. Geuens, and W. Van Lippevelde, “Intervention strategies to promote healthy and sustainable food choices among parents with lower and higher socioeconomic status,”BMC Public Health, vol. 22, no. 1, p. 2378, 2022

    2022

  34. [34]

    Greenrec: A large-scale dataset for green food recommendation,

    L. Zhang, Y . Zhang, X. Zhou, and Z. Shen, “Greenrec: A large-scale dataset for green food recommendation,” inCompanion Proceedings of the ACM Web Conference 2024, 2024, pp. 625–628

    2024

  35. [35]

    Bpr: Bayesian personalized ranking from implicit feedback,

    S. Rendle, C. Freudenthaler, Z. Gantner, and L. Schmidt-Thieme, “Bpr: Bayesian personalized ranking from implicit feedback,” inProceedings of the Twenty-Fifth Conference on Uncertainty in Artificial Intelligence, 2009, pp. 452–461

    2009

  36. [36]

    Top-n-rank: A scalable list-wise ranking method for recommender systems,

    J. Liang, J. Hu, S. Dong, and V . Honavar, “Top-n-rank: A scalable list-wise ranking method for recommender systems,” in2018 IEEE International Conference on Big Data (Big Data). IEEE, 2018, pp. 1052–1058

    2018

  37. [37]

    Setrank: A setwise bayesian approach for collaborative ranking in recommender system,

    C. Wang, H. Zhu, C. Zhu, C. Qin, E. Chen, and H. Xiong, “Setrank: A setwise bayesian approach for collaborative ranking in recommender system,”ACM Transactions on Information Systems, vol. 42, no. 2, pp. 1–32, 2023

    2023

  38. [38]

    Causal denoising framework for generalizable recommendation system using graph neural network,

    Y . Zhao, Y . Xu, N. Liu, Y . Zhang, W. Guo, X. Lu, and L. Cui, “Causal denoising framework for generalizable recommendation system using graph neural network,” in2024 IEEE International Conference on Multimedia and Expo (ICME). IEEE, 2024, pp. 1–6

    2024

  39. [39]

    Stochastic optimization of sorting networks via continuous relaxations,

    A. Grover, E. Wang, A. Zweig, and S. Ermon, “Stochastic optimization of sorting networks via continuous relaxations,” inInternational Conference on Learning Representations, 2019. [Online]. Available: https://openreview.net/forum?id=H1eSS3CcKX

    2019

  40. [40]

    Neuralndcg: Direct optimisation of a ranking metric via differentiable relaxation of sorting,

    P. Pobrotyn and R. Białobrzeski, “Neuralndcg: Direct optimisation of a ranking metric via differentiable relaxation of sorting,”arXiv preprint arXiv:2102.07831, 2021

    Pith/arXiv arXiv 2021

  41. [41]

    Optimizing preference alignment with differentiable ndcg ranking,

    J. Zhou, X. Wang, and J. Yu, “Optimizing preference alignment with differentiable ndcg ranking,”arXiv preprint arXiv:2410.18127, 2024

    arXiv 2024

  42. [42]

    Ensuring user-side fairness in dynamic recom- mender systems,

    H. Yoo, Z. Zeng, J. Kang, R. Qiu, D. Zhou, Z. Liu, F. Wang, C. Xu, E. Chan, and H. Tong, “Ensuring user-side fairness in dynamic recom- mender systems,” inProceedings of the ACM Web Conference 2024, 2024, pp. 3667–3678. IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING 12

    2024

  43. [43]

    Fast differentiable sorting and ranking,

    M. Blondel, O. Teboul, Q. Berthet, and J. Djolonga, “Fast differentiable sorting and ranking,” inInternational Conference on Machine Learning. PMLR, 2020, pp. 950–959

    2020

  44. [44]

    Softsort: A continuous relaxation for the argsort operator,

    S. Prillo and J. Eisenschlos, “Softsort: A continuous relaxation for the argsort operator,” inInternational Conference on Machine Learning. PMLR, 2020, pp. 7793–7802

    2020

  45. [45]

    A pareto-efficient algorithm for multiple objective optimization in e-commerce recommendation,

    X. Lin, H. Chen, C. Pei, F. Sun, X. Xiao, H. Sun, Y . Zhang, W. Ou, and P. Jiang, “A pareto-efficient algorithm for multiple objective optimization in e-commerce recommendation,” inProceedings of the 13th ACM Conference on recommender systems, 2019, pp. 20–28

    2019

  46. [46]

    Multi-gradient descent for multi-objective recommender systems,

    N. Milojkovic, D. Antognini, G. Bergamin, B. Faltings, and C. Musat, “Multi-gradient descent for multi-objective recommender systems,” arXiv preprint arXiv:2001.00846, 2020

    Pith/arXiv arXiv 2001

  47. [47]

    Personalized approximate pareto-efficient recommendation,

    R. Xie, Y . Liu, S. Zhang, R. Wang, F. Xia, and L. Lin, “Personalized approximate pareto-efficient recommendation,” inProceedings of the web conference 2021, 2021, pp. 3839–3849

    2021

  48. [48]

    Pareto- based multi-objective recommender system with forgetting curve,

    J. Jin, Z. Zhang, Z. Li, X. Gao, X. Yang, L. Xiao, and J. Jiang, “Pareto- based multi-objective recommender system with forgetting curve,” in Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, 2024, pp. 4603–4611

    2024

  49. [49]

    Pareto front approximation for multi-objective session-based recommender systems,

    T. Wilm, P. Normann, and F. Stepprath, “Pareto front approximation for multi-objective session-based recommender systems,” inProceedings of the 18th ACM Conference on Recommender Systems, 2024, pp. 809–812

    2024

  50. [50]

    Grad- craft: Elevating multi-task recommendations through holistic gradient crafting,

    Y . Bai, Y . Zhang, F. Feng, J. Lu, X. Zang, C. Lei, and Y . Song, “Grad- craft: Elevating multi-task recommendations through holistic gradient crafting,” inProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2024, pp. 4774–4783

    2024

  51. [51]

    Self-attentive sequential recommenda- tion,

    W.-C. Kang and J. McAuley, “Self-attentive sequential recommenda- tion,” in2018 IEEE international conference on data mining (ICDM). IEEE, 2018, pp. 197–206

    2018

  52. [52]

    Feature-level deeper self-attention network for sequential recom- mendation

    T. Zhang, P. Zhao, Y . Liu, V . S. Sheng, J. Xu, D. Wang, G. Liu, X. Zhou et al., “Feature-level deeper self-attention network for sequential recom- mendation.” inIJCAI, 2019, pp. 4320–4326

    2019

  53. [53]

    Frequency enhanced hybrid attention network for sequential recommendation,

    X. Du, H. Yuan, P. Zhao, J. Qu, F. Zhuang, G. Liu, Y . Liu, and V . S. Sheng, “Frequency enhanced hybrid attention network for sequential recommendation,” inProceedings of the 46th international ACM SIGIR conference on research and development in information retrieval, 2023, pp. 78–88

    2023

  54. [54]

    Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks,

    Z. Chen, V . Badrinarayanan, C.-Y . Lee, and A. Rabinovich, “Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks,” inInternational conference on machine learning. PMLR, 2018, pp. 794–803

    2018

  55. [55]

    Multi-task learning as multi-objective opti- mization,

    O. Sener and V . Koltun, “Multi-task learning as multi-objective opti- mization,”Advances in neural information processing systems, vol. 31, 2018

    2018

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.