Pith. sign in

REVIEW 2 major objections 5 minor 66 references

Optimal Dataset Size for Recommender Systems: Evaluating Algorithms' Performance via Downsampling

T0 review · 2 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Training recommender systems on a fraction of the data can cut runtime and carbon emissions by up to half while retaining most top-10 ranking accuracy.

desk verdict A useful empirical benchmark on downsampling for green recommender systems, with a load-bearing caveat: the User-Subset outperformance claim is not supported because test sets change across portions. read the letter →

arxiv 2502.08845 v2 pith:HHVW6X3U submitted 2025-02-12 cs.IR

Ardalan Arabzadeh This is my paper · ORCID
classification cs.IR
keywords GreenRecommenderSystemsDatasetDownsamplingUser-BasedUser-SubsetnDCG@10CarbonEmissionsRuntimeReductionEnergyEfficiency
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This thesis asks whether recommender systems can be trained on substantially less data while keeping most of their recommendation quality, and answers that the trade-off is often favorable but scenario-dependent. Across seven datasets, twelve algorithms, and two downsampling strategies, it finds that reducing the training set to 30% can cut runtime by roughly a third to a half and cut estimated carbon emissions by tens of kilograms of CO2e per algorithm-dataset run, measured by nDCG@10, a standard top-10 ranking quality metric. On average, simpler algorithms keep about 81% of their full-size nDCG@10 with only half the training data. The result matters because data volume drives the energy cost of AI, so knowing how much data is actually needed allows practitioners to choose sustainability without assuming a large accuracy penalty.

What carries the argument

The load-bearing mechanism is the training-set downsampling protocol, implemented in two variants. User-Based downsampling randomly drops a fraction of each user's training interactions while keeping all users present in every split, so the number of users stays fixed and the test set composition is stable. User-Subset downsampling randomly selects a subset of users at each portion and adjusts each user's train/test ratio to keep the validation and test sets fixed in size, which means the actual users in those sets change across portions. These two protocols produce the different runtime/carbon savings and the different nDCG@10 trajectories that carry the paper's conclusions.

What would settle it

Re-run the User-Subset experiments with a fixed test set (same users and interactions) across all downsampling portions; if the relative performance values above 100% disappear or shrink to near 100%, the paper's 'beats full dataset' result is an artifact of changing test-set composition.

Watch

Extended reading notes

Core claim

The central claim is that dataset downsampling can make recommender-system training substantially cheaper without proportionally sacrificing ranking quality. Concretely, the thesis reports that keeping only 30% of the training data reduces runtime by about 36% (User-Based) to 52% (User-Subset), cutting estimated carbon emissions by up to 51.02 KgCO2e for a single algorithm-dataset training run; at 50% of the data, simpler algorithms retain about 81% of their full-dataset nDCG@10 on average. The paper also reports that in User-Subset configurations, some algorithms score higher in nDCG@10 than they do with the full dataset, and it attributes this to the way this downsampling method changes per-user train/test ratios and test-set composition rather than to data reduction alone.

Load-bearing premise

The User-Subset comparison assumes that holding validation and test sets fixed in size while changing which users they contain does not, by itself, change how hard the evaluation is; if the shifted test-set composition is what produces the apparent gains above 100%, the claim that smaller datasets can beat full-dataset performance is not supported.

Editorial extensions

If this is right

  • Downsampling to 30% of the training set cuts training-phase runtime by about 36% with User-Based and 52% with User-Subset downsampling, corresponding to estimated savings of roughly 35.32 and 51.02 KgCO2e per algorithm-dataset under the paper's assumptions.
  • With 50% of the training data, simpler algorithms retain on average about 81% of their full-dataset nDCG@10 under User-Based downsampling, while the User-Subset configuration averages about 110% relative performance.
  • Algorithms with lower absolute nDCG@10 scores, such as popularity-based and bias-based models, are less sensitive to training-data size, making them the most energy-efficient choices when peak accuracy is not required.
  • 30-core pruning improves absolute nDCG@10 scores, but its effect on downsampling efficiency depends on the strategy: it helps under User-Subset and slightly reduces efficiency under User-Based.
  • The User-Subset method is not a universal replacement for User-Based downsampling; its main consistent advantage is greater runtime and carbon reduction, while its accuracy comparisons across portions are complicated by changing test-set composition.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The carbon-savings figures assume a linear relation between runtime, energy, and CO2e on the same hardware; if deployed systems use dynamic power management or different energy mixes, real savings could deviate from these estimates.
  • A useful reading of the >100% results is that evaluation-set construction, not data quantity alone, drives the apparent accuracy gains; an immediate follow-up would fix the test users across all downsampling portions and measure how the relative-performance curves shift.
  • Because runtime falls roughly with data volume, downsampling could serve as a cheap screening step: run hyperparameter or algorithm selection on 30-50% of the data, then retrain the chosen configuration on the full dataset when the final model must be tuned.
  • The results imply practical data-rationing guidance that the thesis only sketches: teams with tight energy budgets can often halve training data for simpler algorithms with minimal nDCG@10 loss, but they should re-check with other metrics such as precision, recall, diversity, or novelty, since nDCG@10 alone does not capture all recommendation-quality dimensions.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The thesis empirically studies dataset downsampling as a Green Recommender Systems strategy. Seven datasets, twelve algorithms, and two core-pruning levels are evaluated under two downsampling schemes (User-Based, where each user's interaction history is subsampled with a fixed user population, and User-Subset, where progressively more users are included while the validation/test interaction count is held fixed). The reported results are that downsampling to 30–50% of the training set yields large runtime reductions (roughly 36–52%) and estimated carbon savings (up to 51.02 KgCO2e per algorithm-dataset), while retaining, on average, about 64% (User-Based) or about 110% (User-Subset) of full-dataset nDCG@10 at 50% downsampling. The abstract and conclusion also claim that some User-Subset configurations achieve higher nDCG@10 scores than the full dataset. The paper argues that the User-Subset trends are driven by the changing per-user train/test ratio, and it proposes scenario-dependent guidelines for selecting downsampling configurations.

Significance. If the core claims hold, the paper would provide a practically useful empirical benchmark for energy-efficient recommender-system research: it covers a broad matrix of algorithms, datasets, pruning levels, and two downsampling strategies, and it ships openly available code and detailed experimental configurations. The measured runtime reductions and the transparent carbon-estimation framework are genuine strengths. However, the most novel headline claim — that User-Subset downsampling can yield nDCG@10 scores above the full-dataset baseline — rests on comparing scores computed on different test sets at each downsampling portion, which undermines that specific claim as currently stated. The paper itself acknowledges the varying test-set composition in its limitations section, but the magnitude of the problem for the central comparison is not fully reflected in the abstract and conclusion.

major comments (2)
  1. [Section 3.3; Section 4.1; Section 4.2; Abstract] The central performance comparison for User-Subset is compromised because the test set changes at each downsampling portion; the claim of outperforming the full dataset is not supported as stated.
  2. [Section 5, Figure 5.1] The causal claim about train/test ratio being the primary driver is supported by only one algorithm and two datasets, and the experiment does not match the full User-Subset design.
minor comments (5)
  1. [Abstract; Section 4.2] Clarify whether the 81% retention figure is User-Based only or a combined average across the two approaches.
  2. [Section 4.4, Equations (4.1)–(4.6)] Present the carbon-savings results as order-of-magnitude estimates, not measured quantities.
  3. [Section 3.3] Use 'fixed interaction count' instead of 'fixed test set' to avoid the implication of a stable evaluation set.
  4. [Section 4.1, Figure 4.2] Add a disclaimer that the two approaches use different test sets, so their relative-performance curves should not be compared directly.
  5. [Section 5; Figure 5.1] Fix the indicated typos and the citation formatting in Section 5.

Circularity Check

0 steps flagged · score 1.0 of 10

Empirical benchmark with independent measurements; self-citations are not load-bearing.

full rationale

This is an empirical benchmark, not a derivation. No core claim is obtained by fitting a parameter and then relabeling it as a prediction; no uniqueness theorem or ansatz is imported from the authors' prior work to force a conclusion. The two self-references are [10] (a prior workshop version of the same experiments, used to situate the goal) and [13] (a pruning study by the supervisor, cited only to align with the paper's measured pruning observations). Neither supplies the runtime or nDCG@10 numbers, which come from direct experiments with five seeds and released code. The headline User-Subset 'higher nDCG@10 than full dataset' comparisons are not circular: relative performance is defined as the measured ratio to the full-size score, and values above 100% are arithmetic consequences of the measurements, not built into the definition. The paper itself flags the associated validity threat in Section 8 ('the approach involves varying groups of users in the validation and test sets across different downsampling portions, which, while statistically consistent in maintaining fixed sizes for these sets, may influence observed performance') and attempts to isolate it in Section 5. That is a correctness/validity concern, not a self-referential reduction, so it does not raise the circularity score.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

The central empirical claims rest on the experimental protocol, not on a mathematical derivation. The main unverified inputs are the proportionality assumptions used to convert runtime reductions into CO2e savings, and the comparability assumption for User-Subset test sets. No free parameters were fitted to data in this paper; external constants from prior work are treated as inputs.

assumptions (3)
  • domain assumption Runtime reduction can be used as a proxy for energy and CO2e savings, with a near-linear relationship.
    Section 4.4 computes KgCO2e savings as (100% - runtime%) x 0.51 kWh x 10 x 481 gCO2e/kWh x 40, explicitly assuming linearity. This is not validated against measured power on the experiment hardware.
  • ad hoc to paper nDCG@10 at fixed-size but composition-varying test sets is a valid comparator across User-Subset downsampling portions.
    Section 3.3 holds validation and test interaction counts fixed but changes the set of users at each portion. The analysis in Sections 4.1, 4.2, and 5 compares these scores as if they are comparable.
  • domain assumption Core pruning (10-core, 30-core) produces datasets representative enough for generalizable conclusions.
    Section 3.1 uses 10-core as primary and 30-core as supplementary; Section 8 concedes that 30-core excludes three Amazon datasets, limiting generalization.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Optimal Dataset Size for Recommender Systems: Evaluating Algorithms' Performance via Downsampling." pith.science (2026). https://pith.science/paper/HHVW6X3U

@misc{pith2026250208845,
  author       = {Pith},
  title        = {Pith review of: Optimal Dataset Size for Recommender Systems: Evaluating Algorithms' Performance via Downsampling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HHVW6X3U}},
  note         = {Machine review of arXiv:2502.08845}
}
read the original abstract

This thesis investigates dataset downsampling as a strategy to optimize energy efficiency in recommender systems while maintaining competitive performance. With increasing dataset sizes posing computational and environmental challenges, this study explores the trade-offs between energy efficiency and recommendation quality in Green Recommender Systems, which aim to reduce environmental impact. By applying two downsampling approaches to seven datasets, 12 algorithms, and two levels of core pruning, the research demonstrates significant reductions in runtime and carbon emissions. For example, a 30% downsampling portion can reduce runtime by 52% compared to the full dataset, leading to a carbon emission reduction of up to 51.02 KgCO2e during the training of a single algorithm on a single dataset. The analysis reveals that algorithm performance under different downsampling portions depends on factors like dataset characteristics, algorithm complexity, and the specific downsampling configuration (scenario dependent). Some algorithms, which showed lower nDCG@10 scores compared to higher-performing ones, exhibited lower sensitivity to the amount of training data, offering greater potential for efficiency in lower downsampling portions. On average, these algorithms retained 81% of full-size performance using only 50% of the training set. In certain downsampling configurations, where more users were progressively included while keeping the test set size fixed, they even showed higher nDCG@10 scores than when using the full dataset. These findings highlight the feasibility of balancing sustainability and effectiveness, providing insights for designing energy-efficient recommender systems and promoting sustainable AI practices.

Figures

Figures reproduced from arXiv: 2502.08845 by the authors.

Figure 4.1
Figure 4.1. Normalized and Averaged nDCG@10 Scores of Algorithms Across All Datasets for User-Based (upper plot) and User-Subset (lower plot) Downsampling Ap￾proaches. While figures 4.1 focus on normalized algorithm performance, the next step evaluates relative performance efficiency compared to full-size training data [PITH_FULL_IMAGE:figures/full_fig_p024_4_1.png] view at source ↗
Figure 4.2
Figure 4.2. Averaged Relative Performance (% of Full Dataset Performance) of All Al [PITH_FULL_IMAGE:figures/full_fig_p025_4_2.png] view at source ↗
Figure 4
Figure 4. presents the averaged and normalized [PITH_FULL_IMAGE:figures/full_fig_p026_4.png] view at source ↗
Figures from the paper (10 more)
Figure 4.3
Figure 4.3. Figure 4.3: Normalized and Averaged nDCG@10 Scores Across All Algorithms and Datasets: Comparison of 10-Core vs. 30-Core Pruning for User-Based (Upper Plot) and User-Subset (Lower Plot) Downsampling Approaches. 19 [PITH_FULL_IMAGE:figures/full_fig_p027_4_3.png]
Figure 4.4
Figure 4.4. Figure 4.4: Averaged Relative Performance (% of Full Dataset Performance) Across All [PITH_FULL_IMAGE:figures/full_fig_p029_4_4.png]
Figure 4.5
Figure 4.5. Figure 4.5: Algorithm-Specific Heat maps of Normalized Average [PITH_FULL_IMAGE:figures/full_fig_p031_4_5.png]
Figure 4.6
Figure 4.6. Figure 4.6: Average Relative Performance (% of Full Dataset Baseline) for Group 1 and [PITH_FULL_IMAGE:figures/full_fig_p033_4_6.png]
Figure 4.7
Figure 4.7. Figure 4.7: Dataset-Specific Heat maps of Normalized Average [PITH_FULL_IMAGE:figures/full_fig_p035_4_7.png]
Figure 4
Figure 4. Figure 4: illustrates these results with confidence bands, where applicable. The [PITH_FULL_IMAGE:figures/full_fig_p036_4.png]
Figure 4.8
Figure 4.8. Figure 4.8: Average Relative Performance (% of Full Dataset Baseline) for Dataset [PITH_FULL_IMAGE:figures/full_fig_p036_4_8.png]
Figure 4
Figure 4. Figure 4: presents the average runtime percentages for both User-Based downsampling [PITH_FULL_IMAGE:figures/full_fig_p038_4.png]
Figure 4.9
Figure 4.9. Figure 4.9: Runtime Reduction Analysis: User-Based vs. User-Subset Downsampling [PITH_FULL_IMAGE:figures/full_fig_p040_4_9.png]
Figure 5.1
Figure 5.1. Figure 5.1: Relative Performance (% of Full Dataset Baseline) of [PITH_FULL_IMAGE:figures/full_fig_p045_5_1.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

66 extracted references · 45 canonical work pages

  1. [1]

    Adomavicius and J

    G. Adomavicius and J. Zhang. Impact of data characteristics on recommender sys- tems performance. ACM Transactions on Management Information Systems (ACM Trans. Manag. Inform. Syst.), 3(1):1–17, April 2012. URL: http://doi.acm. org/10.1145/2151163.2151166, doi:10.1145/2151163.2151166

    arXiv 2012

  2. [2]

    O. Y . Al-Jarrah, P. D. Yoob, S. Muhaidat, G. K. Karagiannidis, and K. Taha. Efficient machine learning for big data: A review. Big Data Research , 2(3):87–99, 2015. doi:10.1016/j.bdr.2015.04.002

    work page doi:10.1016/j.bdr.2015.04.002 2015

  3. [3]

    M. T. Alam, S. Ubaid, S. Shakil, S. S. Sohail, M. Nadeem, S. Hussain, and J. Sid- diqui. Comparative analysis of machine learning based filtering techniques using movielens dataset. Procedia Computer Science, 194:210–217, 2021. Department of Computer Science and Engineering, Jamia Hamdard, New Delhi, India; School of Computer Science, University of Petrol...

    work page 2021

  4. [4]

    Y . H. Alfaifi. Recommender systems applications: Data sources, features, and chal- lenges. Information, 15(10):660, 2024. doi:10.3390/info15100660

    work page doi:10.3390/info15100660 2024

  5. [5]

    Alibrahim and S

    H. Alibrahim and S. A. Ludwig. Hyperparameter optimization: Comparing genetic algorithm against grid search and bayesian optimization. In2021 IEEE Congress on Evolutionary Computation (CEC), pages 1551–1559. IEEE, 2021

    work page 2021

  6. [6]

    Y . I. Alzoubi and A. Mishra. Green artificial intelligence initiatives: Potentials and challenges. Journal of Cleaner Production , 468:143090,

    work page

  7. [7]

    Auto-Surprise: An Automated Recommender-System (AutoRecSys) Library with Tree of Parzens Estimator (TPE) Optimization

    R. Anand and J. Beel. Auto-surprise: An automated recommender-system (au- torecsys) library with tree of parzens estimator (tpe) optimization. In 14th ACM Conference on Recommender Systems (RecSys) , pages 1–4, 2020. URL: https: //arxiv.org/abs/2008.13532

    work page Pith review arXiv 2020

  8. [8]

    D. A. Anggoro and S. S. Mukti. Performance comparison of grid search and random search methods for hyperparameter tuning in extreme gradient boosting algorithm to predict chronic kidney failure. International Journal of Intelligent Engineering & Systems, 14(6), 2021

    work page 2021

Show all 66 references
  1. [9]

    Arabzadeh

    A. Arabzadeh. Master thesis github repository, 2025. URL: https://github. com/Ardalan224/Master-Thesis/

    2025

  2. [10]

    Arabzadeh, T

    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 (RecSoGood) at the 18th ACM Conference on Recommender Systems...

    2024

  3. [11]

    Basaran, E

    D. Basaran, E. Ntoutsi, and A. Zimek. Redundancies in data and their effect on the evaluation of recommendation systems: A case study on the amazon reviews datasets. In Proceedings of the 2017 SIAM International Conference on Data Mining (SDM), pages 390–398. SIAM, 2017

    2017

  4. [12]

    J. Beel. Our use of ai-tools for writing research papers, 2024. In: Intelligent Sys- tems Group, Blog. URL: https://isg.beel.org/blog/2024/08/19/ our-use-of-ai-tools-for-writing-research-papers/

    2024

  5. [13]

    Beel and V

    J. Beel and V . Brunel. Data pruning in recommender systems research: Best-practice or malpractice? In ACM RecSys 2019 Late-breaking Results , Copenhagen, Den- mark, September 2019. 16th–20th September 2019

    2019

  6. [14]

    J. Beel, A. Said, T. Vente, and L. Wegmeth. Green recommender systems – a call for attention. Recommender-Systems.com Blog , 2024. URL: https: //isg.beel.org/pubs/2024_Green_Recommender_Systems-A_ Call_for_Attention.pdf, doi:10.31219/osf.io/5ru2g. 51

    2024 doi

  7. [15]

    Bentzer and H

    C. Bentzer and H. Thulin. Recommender systems using limited dataset sizes, June 8

  8. [16]

    S. A. Budennyy, V . D. Lazarev, N. N. Zakharenko, A. N. Korovin, O. A. Plosskaya, D. V . Dimitrov, V . S. Akhripkin, I. V . Pavlov, I. V . Oseledets, I. S. Barsola, et al. Eco2ai: carbon emissions tracking of machine learning models as the first step to- wards sustainable ai. ...

    2022

  9. [17]

    R. Burke. Hybrid recommender systems: Survey and experiments. User Mod- eling and User-Adapted Interaction , 12:331–370, 2002. doi:10.1023/A: 1021240730564

    2002 doi

  10. [18]

    Castellanos-Nieves and L

    D. Castellanos-Nieves and L. Garc´ıa-Forte. Improving automated machine-learning systems through green ai. Applied Sciences, 13(20), 2023. URL: https://www. mdpi.com/2076-3417/13/20/11583, doi:10.3390/app132011583

    2023 doi

  11. [19]

    Castellanos-Nieves and L

    D. Castellanos-Nieves and L. Garc ´ıa-Forte. Strategies of automated machine learn- ing for energy sustainability in green artificial intelligence. Applied Sciences (2076- 3417), 14(14), 2024

    2024

  12. [20]

    Ca ˜namares, P

    R. Ca ˜namares, P. Castells, and A. Moffat. Offline evaluation options for rec- ommender systems. Information Retrieval Journal , 23:387–410, 2020. doi: 10.1007/s10791-020-09371-3

    2020 doi

  13. [21]

    C. Chen, P. Zhang, H. Zhang, J. Dai, Y . Yi, H. Zhang, and Y . Zhang. Deep learning on computational-resource-limited platforms: A survey.Advances in Artificial Intel- ligence, 2020. First published: 01 March 2020.doi:10.1155/2020/8454327

    2020 doi

  14. [22]

    E. Cho, S. A. Myers, and J. Leskovec. Friendship and mobility: User movement in location-based social networks. In Proceedings of the 17th ACM SIGKDD Inter- national Conference on Knowledge Discovery and Data Mining (KDD ’11) , pages 1082–1090, San Diego, California, USA, 2011...

    2011

  15. [23]

    Cremonesi, Y

    P. Cremonesi, Y . Koren, and R. Turrin. Performance of recommender algorithms on top-n recommendation tasks. In RecSys ’10: Proceedings of the Fourth ACM 52 Conference on Recommender Systems , pages 39–46, New York, NY , USA, 2010. ACM. doi:10.1145/1864708.1864721

    2010

  16. [24]

    M. D. Ekstrand. Lenskit for python: Next-generation software for recommender systems experiments. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management , pages 2999–3006, Virtual Event, Ireland,

  17. [25]

    Carbon intensity of electricity generation – em- ber and energy institute, 2024

    Ember and Energy Institute. Carbon intensity of electricity generation – em- ber and energy institute, 2024. Yearly Electricity Data by Ember; Sta- tistical Review of World Energy by Energy Institute. Dataset processed by Our World in Data. URL: https://ourworldindata.org/grap...

    2024

  18. [26]

    A. A. Gowda, H.-K. Su, and W.-K. Kuo. Personalized e-commerce: Enhancing customer experience through machine learning-driven personalization. In 2024 IEEE International Conference on Information Technology, Electronics and Intel- ligent Communication Systems (ICITEICS) , pages...

    2024

  19. [28]

    F. M. Harper and J. A. Konstan. The movielens datasets: History and context. ACM Transactions on Interactive Intelligent Systems (TiiS), 5(4):1–19, 2015. doi: 10.1145/2827872

    2015 doi

  20. [29]

    Hennig, T

    L. Hennig, T. Tornede, and M. Lindauer. Towards leveraging automl for sus- tainable deep learning: A multi-objective hpo approach on deep shift neural net- works. In arXiv, 2024. URL: https://arxiv.org/abs/2404.01965, arXiv:2404.01965

    2024 arXiv

  21. [30]

    J. L. Herlocker, J. A. Konstan, and J. Riedl. Explaining collaborative filtering recom- mendations. In CSCW ’00: Proceedings of the 2000 ACM Conference on Computer Supported Cooperative Work, pages 241–250, New York, NY , USA, 2000. ACM. doi:10.1145/358916.358995. 53

  22. [31]

    doi:10.31219/osf.io/4znmd

    doi

  23. [32]

    S. KJ, S. N. BV , C. C. S. Balne, V . K. Sunkara, S. Bhaduri, V . Jain, and A. Chadha. Advancements in modern recommender systems: Industrial applications in so- cial media, e-commerce, entertainment, and beyond, 2024. hal-04711099. URL: https://hal.archives-ouvertes.fr/hal-04711099

    2024

  24. [33]

    Ku ˙zelewska

    U. Ku ˙zelewska. Effect of dataset size on efficiency of collaborative filtering rec- ommender systems with multi-clustering as a neighbourhood identification strat- egy. In V . V . Krzhizhanovskaya et al., editors, Computational Science – ICCS 2020, volume 12139 of Lecture No...

    2020 doi

  25. [34]

    Lacoste, A

    A. Lacoste, A. Luccioni, V . Schmidt, and T. Dandres. Quantifying the carbon emis- sions of machine learning. arXiv preprint, 2019. arXiv:1910.09700 [cs.CY]. URL: https://mlco2.github.io/impact/

    2019 arXiv

  26. [35]

    Jain and R

    K. Jain and R. Jindal. Sampling and noise filtering methods for recommender systems: A literature review. Engineering Applications of Artificial Intelligence , 122:106129, 2023

    2023

  27. [36]

    F. Lin, X. Zhu, Z. Zhao, D. Huang, Y . Yu, X. Li, Z. Zheng, T. Xu, and E. Chen. Knowledge graph pruning for recommendation. arXiv preprint arXiv:2405.11531, 2024

    2024 arXiv

  28. [37]

    Maslowska, E

    E. Maslowska, E. C. Malthouse, and L. D. Hollebeek. The role of recommender systems in fostering consumers’ long-term platform engagement.Journal of Service Management, 2022. Article publication date: 9 May 2022, Issue publication date: 8 July 2022

    2022

  29. [38]

    Z. Meng, R. McCreadie, C. Macdonald, and I. Ounis. Exploring data splitting strate- gies for the evaluation of recommendation models. In Proceedings of RecSys ’20: The 14th ACM Recommender Systems Conference (RecSys ’20), page 8, New York, NY , USA, 2020. ACM.doi:10.1145/11224...

    2020

  30. [39]

    Lannelongue, J

    L. Lannelongue, J. Grealey, and M. Inouye. Green algorithms: Quantify- ing the carbon footprint of computation. Advanced Science , 8(12):2100707,

  31. [40]

    W. Ng, B. Minasny, W. D. S. Mendes, and J. A. M. Dematt ˆe. The influence of training sample size on the accuracy of deep learning models for the prediction of soil properties with near-infrared spectroscopy data. SOIL, 6:565–578, 2020. doi: 10.5194/soil-6-565-2020

    2020 doi

  32. [41]

    J. Ni, J. Li, and J. McAuley. Justifying recommendations using distantly-labeled reviews and fine-grained aspects. In Proceedings of the 2019 Conference on Em- pirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Process...

    2019

  33. [42]

    Y . Qu, L. Qu, T. Chen, X. Zhao, Q. V . H. Nguyen, and H. Yin. Scalable dynamic embedding size search for streaming recommendation. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, pages 1941–1950, 2024

    1941

  34. [43]

    Roy and S

    C. Roy and S. S. Rautaray. Challenges and issues of recommender system for big data applications. In S. S. Rautaray, P. Pemmaraju, and H. Mohanty, editors,Trends of Data Science and Applications, volume 954 of Studies in Computational Intelli- gence. Springer, Singapore, 2021....

    2021 doi

  35. [44]

    Michiels, R

    L. Michiels, R. Verachtert, and B. Goethals. Recpack: An(other) experimentation toolkit for top-n recommendation using implicit feedback data. InProceedings of the 16th ACM Conference on Recommender Systems, RecSys ’22, page 648–651, New York, NY , USA, 2022. Association for C...

    2022

  36. [45]

    J. B. Schafer, D. Frankowski, J. Herlocker, and S. Sen. Collaborative filtering rec- ommender systems. Computer Science, 2007. 55

    2007

  37. [46]

    J. B. Schafer, J. A. Konstan, and J. Riedl. E-commerce recommendation applica- tions. GroupLens Research Project, Department of Computer Science and Engi- neering, University of Minnesota, 2001

    2001

  38. [47]

    Schmidt, J

    M. Schmidt, J. Nitschke, and T. Prinz. Evaluating the performance-deviation of itemknn in recbole and lenskit. arXiv preprint arXiv:2407.13531, 2024

    2024 arXiv

  39. [48]

    Silveira, M

    T. Silveira, M. Zhang, X. Lin, W. Ma, H.-F. Lam, and W. Guo. How good your recommender system is? a survey on evaluations in recommendation. Interna- tional Journal of Machine Learning and Cybernetics , 10:813–831, 2019. doi: 10.1007/s13042-017-0762-9

    2019 doi

  40. [49]

    S. O. S. Santos, A. Skiarski, D. Garc ´ıa-N´u˜nez, V . Lazzarini, R. De Andrade Moral, E. Galvan, A. L. C. Ottoni, and E. Nepomuceno. Green machine learning: Analysing the energy efficiency of machine learning models. In 2024 35th Irish Signals and Systems Conference (ISSC) , ...

    2024

  41. [50]

    Spillo, A

    G. Spillo, A. De Filippo, M. Milano, C. Musto, and G. Semeraro. Towards sustainability-aware recommender systems: Analyzing the trade-off between algo- rithms performance and carbon footprint. In Proceedings of the ACM Conference, page 7, Singapore, Singapore, 2023. doi:10.114...

    2023

  42. [51]

    Spillo, A

    G. Spillo, A. De Filippo, C. Musto, M. Milano, and G. Semeraro. Towards green rec- ommender systems: Investigating the impact of data reduction on carbon footprint and algorithm performances. In 18th ACM Conference on Recommender Systems , 2024

    2024

  43. [52]

    C. Sun, A. Shrivastava, S. Singh, and A. Gupta. Revisiting unreasonable effec- tiveness of data in deep learning era. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 843–852, 2017

    2017

  44. [53]

    Sundberg

    N. Sundberg. Tackling ai’s climate change problem. MIT Sloan Management Re- view, 65(2):38–41, 2024

    2024

  45. [54]

    M. Singh. Scalability and sparsity issues in recommender datasets: a sur- vey. Knowledge and Information Systems , 62:1–43, 2020. doi:10.1007/ s10115-018-1254-2

    2020

  46. [55]

    Tornede, A

    T. Tornede, A. Tornede, J. Hanselle, F. Mohr, M. Wever, and E. H ¨ullermeier. To- wards green automated machine learning: Status quo and future directions. arXiv / Journal of Artificial Intelligence Research, 77:427–457, 2021 / 2023

    2021

  47. [56]

    Van Meteren and M

    R. Van Meteren and M. Van Someren. Using content-based filtering for recom- mendation. In Proceedings of the machine learning in the new information age: MLnet/ECML2000 workshop, volume 30, pages 47–56. Barcelona, 2000

    2000

  48. [57]

    Vente, M

    T. Vente, M. Ekstrand, and J. Beel. Introducing lenskit-auto, an experimental auto- mated recommender system (autorecsys) toolkit. In Proceedings of the 17th ACM Conference on Recommender Systems , pages 1212–1216, 2023. URL: https: //dl.acm.org/doi/10.1145/3604915.3610656

    2023

  49. [58]

    Vente, L

    T. Vente, L. Wegmeth, A. Said, and J. Beel. From clicks to carbon: The environmen- tal toll of recommender systems. In Proceedings of the 18th ACM Conference on Recommender Systems, RecSys ’24, page 580–590, New York, NY , USA, 2024. As- sociation for Computing Machinery. URL:...

    2024

  50. [59]

    P. B. Thorat, R. M. Goudar, and S. Barve. Survey on collaborative filtering, content- based filtering and hybrid recommendation system. International Journal of Com- puter Applications, 110(4), January 2015. Poonam B. Thorat, R. M. Goudar, Sunita Barve, Computer Engineering, M...

    2015

  51. [60]

    L. Xu, Z. Tian, G. Zhang, J. Zhang, L. Wang, B. Zheng, Y . Li, J. Tang, Z. Zhang, Y . Hou, et al. Towards a more user-friendly and easy-to-use benchmark library for recommender systems. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development i...

    2023

  52. [61]

    X. Yang, Y . Wang, C. Chen, Q. Tan, C. Yu, J. Xu, and X. Zhu. Computation resource allocation solution in recommender systems.arXiv, 2103.02259, March 2021. URL: https://arxiv.org/abs/2103.02259, arXiv:2103.02259. 57

    2021 arXiv

  53. [62]

    Zogaj, J

    F. Zogaj, J. P. Cambronero, M. C. Rinard, and J. Cito. Doing more with less: Char- acterizing dataset downsampling for automl. Proceedings of the VLDB Endowment (PVLDB), 14(11):2059–2072, 2021. doi:10.14778/3476249.3476262. 58 Declaration of Authorship I hereby confirm that th...

    2021

  54. [64]

    Y . Wang, L. Wang, Y . Li, D. He, and T.-Y . Liu. A theoretical analysis of ndcg type ranking measures. In Shai Shalev-Shwartz and Ingo Steinwart, editors, Proceedings of the 26th Annual Conference on Learning Theory , volume 30 of Proceedings of Machine Learning Research, pag...

    2013

  55. [2020]

    doi:10.1145/3340531.3412778

  56. [2021]

    1002/advs.202100707, doi:10.1002/advs.202100707

    arXiv:https://onlinelibrary.wiley.com/doi/pdf/10. 1002/advs.202100707, doi:10.1002/advs.202100707

    doi

  57. [2023]

    Degree Project in Computer Science and Engineering, First cycle, 15 credits, KTH Royal Institute of Technology

  58. [2024]

    URL: https://www.sciencedirect.com/science/article/ pii/S0959652624025393, doi:10.1016/j.jclepro.2024.143090. 50

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.