Pith. sign in

REVIEW 5 major objections 6 minor 56 references

Eco-Friendly AI: Unleashing Data Power for Green Federated Learning

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Selecting a subset of federated nodes by carbon footprint and data quality, and trimming data volume, cuts carbon emissions by an average of 56% (peak about 90%) compared with baseline training while preserving or improving accuracy.

desk verdict A legitimate but incremental FL data-management system whose headline emissions reduction is currently an artifact of an unvalidated simulator energy model and early-stopping dynamics, not of data-volume reduction. read the letter →

arxiv 2507.17241 v1 pith:GTTH5Y3O submitted 2025-07-23 cs.LG cs.AIcs.DBcs.DC

Mattia Sabella , Monica Vitali This is my paper · ORCID
classification cs.LGcs.AIcs.DBcs.DC
keywords greenAIfederatedlearningdata-centriccarbonemissionsnodeselectiondataqualitytimeseriesclassificationenergyefficiency
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 paper sets out to establish that the environmental cost of federated learning can be cut sharply without sacrificing model quality by treating data volume and node choice as optimization levers. It analyzes how data volume and quality dimensions—label accuracy, consistency, and completeness—affect both accuracy and carbon emissions across heterogeneous fog nodes, then builds an interactive recommender that predicts a target data volume and ranks nodes by a weighted score of carbon footprint and data quality. Evaluated on time-series classification in a simulated federated setting, the recommender's Node Selection method reduces carbon emissions by an average of 56% compared with training on all nodes with full data, with a peak reduction of about 90%, while maintaining or slightly improving accuracy. If the simulator's energy model is faithful, this gives practitioners a concrete recipe for greener federated learning without changing the model or aggregation scheme.

What carries the argument

The central mechanism is the FL Configuration Recommender, which couples a per-node score with a predicted node count. Each node is scored as $Score_n = W_E(1 - \mathrm{CO}_{2}^{n}/(2 \cdot MaxCO_2)) + \sum_i W_i Q_n^i$, where $CO_2^n$ is the node's carbon footprint, $MaxCO_2$ is the footprint of the most polluting node, $Q_n^i$ are data-quality dimensions such as consistency and completeness, and the weights sum to 1. A Gradient Boosting regressor trained on dataset features predicts the number of nodes $\hat{N}$ needed to meet the accuracy threshold, and the recommender then applies one of three strategies—Node Selection, Minimal Smart Reduction, or Smart Reduction—to choose nodes and trim or clean their data. The FL Simulator supplies the raw material: controlled degradations of volume, label accuracy, consistency, and completeness produce accuracy-versus-volume and energy-versus-volume curves, summarized by logarithmic regressions, that the predictor learns from.

What would settle it

Measure actual energy consumption of the same ResNet-based federated training on physical fog devices with power meters; if real per-node kilowatt-hours and epoch counts do not reproduce the simulated accuracy-versus-energy relationship, the claimed average 56% emission reduction is an artifact of the simulator.

Watch

Extended reading notes

Core claim

The paper's central claim is that data-centric configuration choices—how much data each client trains on and which clients participate—are a first-order lever for the carbon footprint of federated learning, and that these choices can be automated without losing accuracy. In a simulated fog environment with heterogeneous nodes, the Node Selection method, which ranks nodes by a weighted score of $\mathrm{CO}_2$ footprint and data quality and keeps only the top-ranked nodes with data trimmed to a target volume, cut emissions by 56% on average (up to about 90%) relative to the baseline. The same method improved accuracy by about 10% over the baseline and met the researcher's accuracy threshold in 87% of the experiments; the variant that also cleans dirty data (Minimal Smart Reduction) improved accuracy by 8% and cut emissions by 45%, while the variant that adds nodes to preserve the target volume (Smart Reduction) produced the largest accuracy gain (12%) but the smallest emission cut (25%). The paper therefore claims that its FL Configuration Selection System can recommend greener configurations that meet a predefined accuracy target, and that vertical (node-level) data reduction is the preferred strategy because it preserves accuracy while lowering emissions.

Load-bearing premise

The load-bearing premise is that the simulator's energy and carbon accounting—synthetic per-node power values, simulated epoch counts, and regional carbon intensities—faithfully represents real federated training on fog devices.

Editorial extensions

If this is right

  • Federated learning can be made substantially greener without touching the model architecture, aggregation rule, or privacy properties—only data volume and participant choice change.
  • Training on less data is not an accuracy penalty in these experiments: the accuracy-first method (Smart Reduction) improved accuracy by 12% over the baseline while cutting emissions by 25%.
  • The one-time exploration cost of building the predictor—1.93 kg CO2e for volume experiments and 0.54 kg CO2e for quality experiments—can be amortized over many reuse workloads, making the methodology more sustainable over time.
  • Faced with an accuracy target, a practitioner can choose among an emission-first configuration (Node Selection), a balanced one (Minimal Smart Reduction), and an accuracy-first one (Smart Reduction), with the paper's recommendation being Node Selection.

Reading between the lines

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

  • One extension the authors leave implicit: the node-scoring formula depends only on each node's carbon footprint and data-quality metrics, so the same recommender logic could be retrained for other federated tasks, such as natural language or vision, rather than only time-series classification.
  • Their early-stopping result—clean data prevents premature stopping and therefore increases epochs and emissions—implies that data cleaning and stopping criteria should be co-optimized; the paper measures this effect but does not exploit it as a control knob.
  • A testable refinement would make the carbon-intensity term time-varying, so node rankings react to real-time grid mix and combine this data-centric approach with carbon-aware scheduling.
  • Because the validation rests on synthetic power profiles, the decisive next check is a physical deployment with power meters; that would also reveal whether communication costs, which the simulator may undercount, change the ranking of the three methods.
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

5 major / 6 minor

Summary. The paper proposes a data-centric methodology for reducing the carbon footprint of federated learning (FL). It combines an FL simulator that injects controlled data-quality degradations, a logarithmic curve extractor, a gradient-boosting predictor that maps dataset features and target accuracy to a recommended data volume/number of nodes, and a recomender that ranks nodes by a weighted score of carbon intensity and data quality. Three selection methods (Node Selection, Minimal Smart Reduction, and Smart Reduction) are compared against a no-optimization baseline in a Flower-based simulation on three time-series classification tasks. The central claim, stated in Sec. 6.4, is that Node Selection reduces carbon emissions by an average of 56% (up to ~90%) while maintaining or slightly improving accuracy. The paper also reports life-cycle emissions of the exploration phase and points to a public repository.

Significance. If the energy and carbon accounting were validated, the contribution would be practically relevant: it addresses an underexplored interaction between data quality/volume and FL sustainability, provides a reusable recommender, and ships code and data for reproducibility. The paper is also transparent about the number of simulations and repetitions (8 runs per validation configuration). However, the central claim is currently established only inside a simulator whose energy model is not validated against real hardware or existing green-FL baselines, so the quantitative headline (56%/90% reduction) should be treated as a simulation result pending further evidence.

major comments (5)
  1. [Sec. 5.1, Table 4] The energy model is the load-bearing component of the claimed 56%/90% carbon reductions, but it is never specified or validated. Table 4 lists static per-node 'Power (kWh)' values, yet no equation or algorithm description connects those values to local epochs, batch counts, communication rounds, node utilization, or training duration. Carbon emissions are then just energy times a location-specific intensity, so every downstream comparison inherits this unvalidated accounting. Please provide the full energy/emissions computation, calibrate it against measurements on real devices (e.g., energy profilers or published per-device power data), and report a sensitivity analysis over power assumptions.
  2. [Secs. 6.1-6.3] The paper's own early-stopping observations undermine the attribution of the reported savings to the data-volume reduction mechanism. The text reports that dirty data triggers early stopping (fewer epochs, lower emissions), while clean data avoids early stopping (more epochs, higher emissions), and that in Configuration 2 the SR method emits more than the Baseline. This means the sign and magnitude of an emissions difference can reverse depending on the simulator's early-stopping rule, so the headline reduction is not robustly attributable to node selection or data cleaning. Please report epoch counts and training durations per method, and show results with fixed epoch budgets as a sensitivity check.
  3. [Secs. 4.2 and 5.2] The FL Reduction System's predictor is central to the recommender, but its generalization error is not reported. The text states that Gradient Boosting had the lowest test error among the evaluated regressors, but no RMSE, MAE, or cross-validation details are given, and the held-out evaluation datasets (FreezerRegularTrain, TwoLeadECG, ElectricDevices) are used only in the validation phase. Please report the predictor's held-out error, the features used, and, ideally, a comparison of the predicted number of nodes with the number that would be optimal under an exhaustive search on the validation configurations.
  4. [Sec. 6, overall validation] The evaluation lacks a random-selection baseline and any comparison against existing green-FL methods. The three proposed methods are compared only against a no-optimization Baseline that uses all nodes and all data. Without a random node-selection baseline that selects the same number of nodes, the observed improvements cannot be attributed to the scoring function of Eq. (8) rather than to the trivial effect of training on fewer nodes. Please add a random-subset baseline with matched node counts and, if feasible, a comparison to a published carbon-aware FL method (e.g., FedZero or FedGreen) on the same simulated setup.
  5. [Sec. 5.2 and Sec. 5.3] The term 'vertical' is used inconsistently, which confuses the methodology. In Sec. 5.2, the 'vertical approach' refers to experiments where degradations affect only a subset of nodes and is selected as the baseline for later steps. In Sec. 3 and Sec. 5.3, 'vertical data reduction' is defined as selecting a subset of nodes for training. Please disambiguate these two uses, as the recommender's design depends on the intended meaning.
minor comments (6)
  1. [Throughout] There are several typos and formatting issues: 'th quality' in Sec. 3, 'Accuray Estimation' in Table 2, and broken math formatting around Eq. (8). A careful proofreading pass is needed.
  2. [Fig. 12] Figure 12 is captioned 'Configuration 3 Energy Results', but the text describes carbon emissions. Please make the metric labels consistent (energy in kWh vs. emissions in kg CO2e).
  3. [Table 4] The column header 'Power (kWh)' mixes power and energy units. If the values are power draws, they should be in kW or W; if they are per-training energy estimates, that should be stated explicitly and connected to the energy model.
  4. [Eq. (8)] The notation CO_n^2 is confusing because it resembles a squared quantity; it should be written as CO2_n or carbon_n. Also, the weights in Table 3 are fixed ad hoc, and a sensitivity analysis of the ranking to these weights would strengthen the recommendation component.
  5. [Sec. 6.4] The 'Life Cycle Emissions' paragraph reports 2.47 kg CO2e for the exploration phase but does not amortize this cost over the validation experiments or over the claimed reuse by multiple researchers. Please clarify how the one-time exploration cost compares with the per-task savings.
  6. [References] Reference [1] lacks an accessed date and a stable URL besides 'Electricity Maps - Data Portal'; please provide full citation details. Reference [4] and [24] are duplicates of the same archive and could be consolidated.

Circularity Check

0 steps flagged · score 2.0 of 10

No meaningful circularity in the derivation chain; the score reflects only a minor, non-load-bearing self-citation and a partly self-referential simulator evaluation, not a construction-level reduction.

full rationale

The paper's derivation chain is not circular by the definitions in the review criteria. The FL Reduction System is fitted to simulator-generated curves from five UCR datasets in Sec. 5.2 and then applied to three disjoint validation datasets in Sec. 6; the predicted quantity (data volume / number of nodes) is not the evaluation metric (accuracy and carbon emissions), and the validation configurations are not the training configurations. The carbon accounting depends on external carbon-intensity data from Electricity Maps [1] and per-node power values derived from hardware specifications, and the evaluation is not forced: in Configuration 2 the SR method emits more than the Baseline (Sec. 6.2), and Sec. 6.4 states that NS performs best 'without data removal,' showing that simulated early-stopping dynamics can reverse the outcome. The only self-citation, Anselmo and Vitali [3], is used as related work and as a prior expectation that the paper confirms with its own experiments; it is not load-bearing. The main fragility is that the simulator's internal energy model is not externally validated, but that is a correctness and generalizability concern, not a circularity of the kind requiring a quote-and-reduction demonstration. Accordingly, the appropriate finding is near-zero circularity, with score 2 reflecting the minor self-citation and the otherwise self-contained empirical evaluation.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim relies on four hand-chosen or fitted inputs: ranking weights, curve parameters, regressor hyperparameters, and per-node power profiles. It also depends on four domain assumptions about simulation fidelity, electricity intensity, data-quality availability, and regressor generalization. No new physical entities are postulated; the proposed system is software assembled from existing components.

free parameters (4)
  • Ranking weights WE, W_consistency, W_completeness (Eq. 8) = 0.7, 0.2, 0.1
    Chosen by hand in Table 3 to favor energy; no sensitivity analysis is reported even though the score determines all node selections.
  • Logarithmic curve parameters (Eq. 3) = not reported
    Fitted to simulator outputs for each metric and experiment; they are the basis for the data-volume predictor but the exact values are not listed.
  • Gradient Boosting regressor configuration = not reported
    Selected by hyperparameter search in Sec 5.2; the chosen hyperparameters are omitted, so the predictor is under-specified.
  • Per-node power values (Table 4) = 10 to 350, units ambiguous
    Assigned to the validation scenarios as scenario inputs; the column header says kWh but values plausibly represent watts, and no device mapping is given.
assumptions (4)
  • domain assumption A single-machine Flower simulator faithfully reproduces energy consumption and accuracy of FL across heterogeneous nodes.
    All phase-one curves and validation results come from the simulator; no physical or measured deployment is used (Sec 5.1, Sec 6).
  • domain assumption Electricity Maps carbon intensity values [1] are accurate and constant during training.
    Emissions are computed from energy multiplied by location-specific intensity; the actual values are not listed in the paper and no uncertainty is given.
  • domain assumption Researchers can supply reliable data-quality metrics and distinguish clean from dirty data for MSR and SR.
    Smart reduction methods filter low-quality data based on node quality metrics; no cleaning or detection algorithm is described (Sec 5.3).
  • domain assumption The data-volume regressor trained on five UCR datasets generalizes to unseen FL tasks.
    The predictor is trained on StarlightCurves, ChlorineConcentration, PhalangesOutlinesCorrect, Yoga, and ItalyPowerDemand, then applied to FreezerRegularTrain, TwoLeadECG, and ElectricDevices; no confidence intervals or transfer analysis are shown.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Eco-Friendly AI: Unleashing Data Power for Green Federated Learning." pith.science (2026). https://pith.science/paper/GTTH5Y3O

@misc{pith2026250717241,
  author       = {Pith},
  title        = {Pith review of: Eco-Friendly AI: Unleashing Data Power for Green Federated Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GTTH5Y3O}},
  note         = {Machine review of arXiv:2507.17241}
}
read the original abstract

The widespread adoption of Artificial Intelligence (AI) and Machine Learning (ML) comes with a significant environmental impact, particularly in terms of energy consumption and carbon emissions. This pressing issue highlights the need for innovative solutions to mitigate AI's ecological footprint. One of the key factors influencing the energy consumption of ML model training is the size of the training dataset. ML models are often trained on vast amounts of data continuously generated by sensors and devices distributed across multiple locations. To reduce data transmission costs and enhance privacy, Federated Learning (FL) enables model training without the need to move or share raw data. While FL offers these advantages, it also introduces challenges due to the heterogeneity of data sources (related to volume and quality), computational node capabilities, and environmental impact. This paper contributes to the advancement of Green AI by proposing a data-centric approach to Green Federated Learning. Specifically, we focus on reducing FL's environmental impact by minimizing the volume of training data. Our methodology involves the analysis of the characteristics of federated datasets, the selecting of an optimal subset of data based on quality metrics, and the choice of the federated nodes with the lowest environmental impact. We develop a comprehensive methodology that examines the influence of data-centric factors, such as data quality and volume, on FL training performance and carbon emissions. Building on these insights, we introduce an interactive recommendation system that optimizes FL configurations through data reduction, minimizing environmental impact during training. Applying this methodology to time series classification has demonstrated promising results in reducing the environmental impact of FL tasks.

Figures

Figures reproduced from arXiv: 2507.17241 by the authors.

Figure 2
Figure 2. FL Configuration Selection System Architecture. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. FL Simulator: Sub-experiment generation process [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 5
Figure 5. Accuracy and Data Volume trade-off. 5.3 FL Configuration Recommendation The FL Configuration Recommendation phase aims to minimize the carbon footprint while maintaining the required performance level for a specific training task submitted by a researcher. The researcher provides the FL configuration data as input, in￾cluding for each node: power profile, geographical location, dataset volume, and data quality. Base… view at source ↗
Figures from the paper (7 more)
Figure 6
Figure 6. Figure 6: Energy and Data Volume trade-off. assigned to each node, 𝑆𝑐𝑜𝑟𝑒𝑛, is computed as follows: 𝑆𝑐𝑜𝑟𝑒𝑛 = 𝑊𝐸 · (1 − 𝐶𝑂𝑛 2 𝑀𝑎𝑥𝐶𝑂2 ) + ∑︁𝐼 𝑖=1 𝑊𝑖 · 𝑄 𝑛 𝑖 (8) where𝐶𝑂𝑛 2 represents the environmental footprint of node 𝑛, while 𝑀𝑎𝑥𝐶𝑂2 denotes the footprint of the node with the high…
Figure 7
Figure 7. Figure 7: Configuration 1 Accuracy Results [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Configuration 1 Carbon Emissions Results. [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: Configuration 2 Accuracy Results [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]
Figure 10
Figure 10. Figure 10: Configuration 2 Carbon Emissions Results. [PITH_FULL_IMAGE:figures/full_fig_p009_10.png]
Figure 11
Figure 11. Figure 11: Configuration 3 Accuracy Results. to an improvement in performance in this FL environment. The NS and the MSR methods have selected only 2 nodes, while the SR method selected 3 clients to satisfy the data volume requirement. All the methods show improvements in accura…
Figure 12
Figure 12. Figure 12: Configuration 3 Energy Results. In Configuration 3, the Baseline carbon emissions is relevant, reaching 2.5 kg𝐶𝑂2e. The dataset used in Configuration 3 has more than 8000 samples to train, resulting in the most demanding FL training compared with the previous configur…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

56 extracted references · 44 canonical work pages

  1. [1]

    Electricity Maps - Data Portal

    2023, July 1. Electricity Maps - Data Portal. Technical Report. https://www. electricitymaps.com/data-portal

    work page 2023

  2. [2]

    Ali Abbasi, Fan Dong, Xin Wang, Henry Leung, Jiayu Zhou, and Steve Drew. 2024. FedGreen: Carbon-aware Federated Learning with Model Size Adaptation. In2024 IEEE International Conference on Communications Workshops (ICC Workshops) . IEEE, 1352–1358

    work page 2024

  3. [3]

    Martín Anselmo and Monica Vitali. 2023. A data-centric approach for reducing carbon emissions in deep learning. In International Conference on Advanced Information Systems Engineering. Springer, 123–138

    work page 2023

  4. [4]

    Anthony Bagnall, Hoang Anh Dau, Jason Lines, Michael Flynn, James Large, Aaron Bostrom, Paul Southam, and Eamonn Keogh. 2018. The UEA multivariate time series classification archive, 2018. arXiv:1811.00075 [cs.LG]

    arXiv 2018

  5. [5]

    Laure Berti-Equille. 2019. Learn2clean: Optimizing the Sequence of Tasks for Web Data Preparation. In The World Wide Web Conference. 2580–2586

    work page 2019

  6. [6]

    Daniel J Beutel, Taner Topal, Akhil Mathur, Xinchi Qiu, Javier Fernandez- Marques, Yan Gao, Lorenzo Sani, Hei Li Kwing, Titouan Parcollet, Pedro PB de Gusmão, and Nicholas D Lane. 2020. Flower: A Friendly Federated Learning Research Framework. arXiv preprint arXiv:2007.14390 (2020)

    arXiv 2020

  7. [7]

    Keith Bonawitz, Hubert Eichner, Wolfgang Grieskamp, Dzmitry Huba, Alex Ingerman, Vladimir Ivanov, Chloe Kiddon, Jakub Konečn`y, Stefano Mazzocchi, Brendan McMahan, et al . 2019. Towards federated learning at scale: System design. Proceedings of machine learning and systems 1 (2019), 374–388

    work page 2019

  8. [8]

    Lukas Budach, Moritz Feuerpfeil, Nina Ihde, Andrea Nathansen, Nele Noack, Hendrik Patzlaff, Felix Naumann, and Hazar Harmouch. 2022. The effects of data quality on machine learning performance. arXiv preprint arXiv:2207.14529 (2022)

    arXiv 2022

Show all 56 references
  1. [9]

    Roger Creus Castanyer, Silverio Martínez-Fernández, and Xavier Franch. 2024. Which design decisions in AI-enabled mobile applications contribute to greener AI? Empirical Software Engineering 29, 1 (2024), 2

    2024

  2. [10]

    Chengliang Chai, Kaisen Jin, Nan Tang, Ju Fan, Dongjing Miao, Jiayi Wang, Yuyu Luo, Guoliang Li, Ye Yuan, and Guoren Wang. 2025. Cost-effective Missing Value Imputation for Data-effective Machine Learning. ACM Transactions on Database Systems (2025)

    2025

  3. [11]

    Chengliang Chai, Jiayi Wang, Yuyu Luo, Zeping Niu, and Guoliang Li. 2023. Data Management for Machine Learning: A Survey. IEEE Transactions on Knowledge and Data Engineering 35, 5 (2023), 4646–4667. https://doi.org/10.1109/TKDE. 2022.3148237

    2023

  4. [12]

    Graham Cormode, Igor L Markov, and Harish Srinivas. 2024. Private and Efficient Federated Numerical Aggregation.. In EDBT. 734–742

    2024

  5. [13]

    Nathan C Frey, Dan Zhao, Simon Axelrod, Michael Jones, David Bestor, Vijay Gadepally, Rafael Gómez-Bombarelli, and Siddharth Samsi. 2022. Energy-aware Neural Architecture Selection and Hyperparameter Optimization. In 2022 IEEE International Parallel and Distributed Processing ...

    2022

  6. [14]

    Stefanos Georgiou et al . 2022. Green ai: Do deep learning frameworks have different costs?. In Proceedings of the 44th International Conference on Software Engineering. 1082–1094

    2022

  7. [15]

    Nitin Gupta, Shashank Mujumdar, Hima Patel, Satoshi Masuda, Naveen Panwar, Sambaran Bandyopadhyay, Sameep Mehta, Shanmukha Guttula, Shazia Afzal, Ruhi Sharma Mittal, et al . 2021. Data quality for machine learning tasks. In Proceedings of the 27th ACM SIGKDD conference on know...

    2021

  8. [16]

    Harris, K

    Charles R. Harris, K. Jarrod Millman, Stéfan J. van der Walt, et al. 2020. Array programming with NumPy. Nature 585, 7825 (sep 2020), 357–362. https://doi. 12 org/10.1038/s41586-020-2649-2

    2020 doi

  9. [17]

    Kaiming He et al . 2016. Deep Residual Learning for Image Recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition . 770–778

    2016

  10. [18]

    Anh-Tu Hoang, Ahmed Lekssays, Barbara Carminati, Elena Ferrari, et al. 2023. Privacy-preserving Decentralized Learning of Knowledge Graph Embeddings.. In EDBT/ICDT Workshops

    2023

  11. [19]

    Ting-Yun Hsiao et al. 2019. Filter-based Deep-compression with Global Average Pooling for Convolutional Networks. Journal of Systems Architecture 95 (2019), 9–18

    2019

  12. [20]

    J. D. Hunter. 2007. Matplotlib: A 2D graphics environment. Computing in Science & Engineering 9, 3 (2007), 90–95. https://doi.org/10.1109/MCSE.2007.55

    2007 doi

  13. [21]

    Streamlit Inc. 2024. Streamlit: The fastest way to build and share data apps. https://streamlit.io/

    2024

  14. [22]

    Abhinav Jain, Hima Patel, Lokesh Nagalapatti, Nitin Gupta, Sameep Mehta, Shanmukha Guttula, Shashank Mujumdar, Shazia Afzal, Ruhi Sharma Mittal, and Vitobha Munigala. 2020. Overview and importance of data quality for machine learning tasks. In Proceedings of the 26th ACM SIGKD...

    2020

  15. [23]

    Johannes Jakubik, Michael Vössing, Niklas Kühl, Jannis Walk, and Gerhard Satzger. 2024. Data-centric artificial intelligence. Business & Information Systems Engineering 66, 4 (2024), 507–515

    2024

  16. [24]

    Eamonn Keogh, Xiaopeng Xi, Li Wei, and Chotirat Ann Ratanamahatana. 2006. The UCR time series classification/clustering homepage. URL= http://www. cs. ucr. edu/˜ eamonn/time_series_data (2006)

    2006

  17. [25]

    Afsana Khan, Marijn ten Thij, and Anna Wilbik. 2025. Vertical federated learning: A structured literature review. Knowledge and Information Systems (2025), 1–39

    2025

  18. [26]

    Will Knight. 2020. AI can do great things - if it doesn’t burn the planet. Wired Magazine (2020)

    2020

  19. [27]

    Nikolaos Konstantinou and Norman W Paton. 2020. Feedback Driven Improve- ment of Data Preparation Pipelines. Information Systems 92 (2020), 101480

    2020

  20. [28]

    Qinbin Li, Yiqun Diao, Quan Chen, and Bingsheng He. 2022. Federated Learn- ing on Non-IID Data Silos: An Experimental Study. In 2022 IEEE 38th Interna- tional Conference on Data Engineering (ICDE) . 965–978. https://doi.org/10.1109/ ICDE53745.2022.00077

    2022

  21. [29]

    Qinbin Li, Zeyi Wen, Zhaomin Wu, Sixu Hu, Naibo Wang, Yuan Li, Xu Liu, and Bingsheng He. 2023. A Survey on Federated Learning Systems: Vision, Hype and Reality for Data Privacy and Protection.IEEE Transactions on Knowledge and Data Engineering 35, 4 (2023), 3347–3366. https://...

    2023

  22. [30]

    Yang Liu, Yan Kang, Tianyuan Zou, Yanhong Pu, Yuanqin He, Xiaozhou Ye, Ye Ouyang, Ya-Qin Zhang, and Qiang Yang. 2024. Vertical Federated Learning: Concepts, Advances, and Challenges. IEEE Transactions on Knowledge and Data Engineering 36, 7 (2024), 3615–3634. https://doi.org/1...

    2024

  23. [31]

    Federica Lucivero. 2020. Big data, big waste? A reflection on the environmental sustainability of big data initiatives. Science and Engineering Ethics 26, 2 (2020), 1009–1030

    2020

  24. [32]

    Antonio Maccioni and Riccardo Torlone. 2018. KAYAK: a Framework for just-in- time Data Preparation in a Data Lake. In Advanced Information Systems Engi- neering: 30th International Conference, CAiSE 2018, Tallinn, Estonia, June 11-15, 2018, Proceedings 30. Springer, 474–489

    2018

  25. [33]

    Yuyi Mao, Xianghao Yu, Kaibin Huang, Ying-Jun Angela Zhang, and Jun Zhang

  26. [34]

    Zhuqi Miao et al. 2023. A Data Preparation Framework for Cleaning Electronic Health Records and Assessing Cleaning Outcomes for Secondary Analysis. In- formation Systems 111 (2023), 102130

    2023

  27. [35]

    The pandas development team. 2020. pandas-dev/pandas: Pandas. https://doi. org/10.5281/zenodo.3509134

    2020 doi

  28. [36]

    Fabian Pedregosa, Gaël Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, et al. 2011. Scikit-learn: Machine learning in Python. Journal of machine learning research 12, Oct (2011), ...

    2011

  29. [37]

    Mark Pilgrim and Simon Willison. 2009. Dive Into Python 3 . Vol. 2. Springer

    2009

  30. [38]

    Neoklis Polyzotis, Sudip Roy, Steven Euijong Whang, and Martin Zinkevich

  31. [39]

    Xinchi Qiu, Titouan Parcollet, Javier Fernandez-Marques, Pedro PB Gusmao, Yan Gao, Daniel J Beutel, Taner Topal, Akhil Mathur, and Nicholas D Lane. 2023. A first look into the carbon footprint of federated learning. Journal of Machine Learning Research 24, 129 (2023), 1–23

    2023

  32. [40]

    David Rolnick et al . 2022. Tackling Climate Change with Machine Learning. ACM Computing Surveys (CSUR) 55, 2 (2022), 1–96

    2022

  33. [41]

    Stefano Savazzi, Vittorio Rampa, Sanaz Kianoush, and Mehdi Bennis. 2022. An energy and carbon footprint analysis of distributed and federated learning. IEEE Transactions on Green Communications and Networking 7, 1 (2022), 248–264

    2022

  34. [42]

    Roy Schwartz et al. 2020. Green AI. Commun. ACM 63, 12 (2020), 54–63

    2020

  35. [43]

    Yunjung Shin et al. 2020. Practical Methods of Image Data Preprocessing for Enhancing the Performance of Deep Learning Based Road Crack Detection. ICIC Express Letters, Part B: Applications 11, 4 (2020), 373–379

    2020

  36. [44]

    Emma Strubell, Ananya Ganesh, and Andrew McCallum. 2020. Energy and policy considerations for modern deep learning research. In Proceedings of the AAAI conference on artificial intelligence , Vol. 34. 13693–13696

    2020

  37. [45]

    Chen Sun et al. 2017. Revisiting Unreasonable Effectiveness of Data in Deep Learning Era. In Proceedings of the IEEE International Conference on Computer Vision. 843–852

    2017

  38. [46]

    Dipanwita Thakur, Antonella Guzzo, Giancarlo Fortino, and Francesco Piccialli

  39. [47]

    Roberto Verdecchia, Luis Cruz, June Sallou, Michelle Lin, James Wickenden, and Estelle Hotellier. 2022. Data-Centric Green AI An Exploratory Empirical Study. In 2022 International Conference on ICT for Sustainability (ICT4S) . IEEE, 35–45. https://doi.org/10.1109/ict4s55073.2022.00015

    2022

  40. [48]

    Yansheng Wang, Yongxin Tong, Dingyuan Shi, and Ke Xu. 2021. An Efficient Approach for Cross-Silo Federated Learning to Rank. In 2021 IEEE 37th Inter- national Conference on Data Engineering (ICDE) . 1128–1139. https://doi.org/10. 1109/ICDE51399.2021.00102

    2021

  41. [49]

    Vitor Werner de Vargas et al. 2023. Imbalanced Data Preprocessing Techniques for Machine Learning: a Systematic Mapping Study. Knowledge and Information Systems 65, 1 (2023), 31–57

    2023

  42. [50]

    Steven Euijong Whang, Yuji Roh, Hwanjun Song, and Jae-Gil Lee. 2023. Data collection and quality challenges in deep learning: A data-centric ai perspective. The VLDB Journal 32, 4 (2023), 791–813

    2023

  43. [51]

    Philipp Wiesner, Ramin Khalili, Dennis Grinwald, Pratik Agrawal, Lauritz Tham- sen, and Odej Kao. 2024. Fedzero: Leveraging renewable excess energy in feder- ated learning. In Proceedings of the 15th ACM International Conference on Future and Sustainable Energy Systems . 373–385

    2024

  44. [52]

    Yuncheng Wu, Naili Xing, Gang Chen, Tien Tuan Anh Dinh, Zhaojing Luo, Beng Chin Ooi, Xiaokui Xiao, and Meihui Zhang. 2023. Falcon: A privacy- preserving and interpretable vertical federated learning system. Proceedings of the VLDB Endowment 16, 10 (2023), 2471–2484

    2023

  45. [53]

    Daochen Zha, Zaid Pervaiz Bhat, Kwei-Herng Lai, Fan Yang, Zhimeng Jiang, Shaochen Zhong, and Xia Hu. 2025. Data-centric artificial intelligence: A survey. Comput. Surveys 57, 5 (2025), 1–42. 13

    2025

  46. [2018]

    ACM Sigmod Record 47, 2 (2018), 17–28

    Data lifecycle challenges in production machine learning: a survey. ACM Sigmod Record 47, 2 (2018), 17–28

    2018

  47. [2024]

    Green edge AI: A contemporary survey. Proc. IEEE (2024)

    2024

  48. [2025]

    Green Federated Learning: A new era of Green Aware AI. Comput. Surveys (2025)

    2025

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.