Pith. sign in

REVIEW 5 major objections 6 minor 26 references

AAPA: An Archetype-Aware Predictive Autoscaler with Uncertainty Quantification for Serverless Workloads on Kubernetes

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 This paper argues that modeling serverless workload archetypes and prediction confidence improves autoscaling outcomes, trading higher resource usage for large gains in SLO compliance and latency.

desk verdict Useful dataset and sensible design, but the paper's headline numbers contradict its own figures. read the letter →

arxiv 2507.05653 v3 pith:7QONCYP2 submitted 2025-07-08 cs.DC

Guilin Zhang , Srinivas Vippagunta , Raghavendra Nandagopal , Suchitra Raman , Jeff Xu , Marcus Pfeiffer , Shreeshankar Chatterjee , Ziqi Tan
show 2 more authors
Wulan Guo Hailong Jiang
This is my paper
classification cs.DC
keywords serverlesscomputingautoscalingworkloadarchetypesuncertaintyquantificationweaksupervisionKubernetesSLOviolationsresourceefficiencyindex
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

AAPA is a predictive autoscaler for serverless workloads that first sorts each load window into one of four behavioral archetypes—SPIKE, PERIODIC, RAMP, STATIONARY—and then applies a scaling policy tailored to that archetype, with the policy's aggressiveness modulated by a calibrated confidence score. Using a weakly labeled dataset built from production serverless traces, the paper shows that this archetype-aware, uncertainty-aware approach cuts SLO violations by up to 50% and lowers latency by 40% relative to the default Kubernetes autoscaler. The gains are not free: under spike-dominated workloads, AAPA consumes 2–8× more resources, mostly to keep warm pools ready. The paper also contributes a unified Resource Efficiency Index (REI) for scoring autoscalers on performance, cost, and stability simultaneously.

What carries the argument

The load-bearing mechanism is a calibrated workload classifier coupled to per-archetype control policies. The classifier takes 37 statistical, time-domain, and frequency-domain features extracted from each 60-minute window and outputs both a label and a confidence score $c\in[0,1]$; $\beta$ calibration makes the score trustworthy. Each archetype maps to a distinct strategy: SPIKE uses warm pools with a low 30% CPU target and long cooldowns, PERIODIC uses predictive Holt-Winters scaling with short cooldowns and a 75% target, RAMP follows trend extrapolation, and STATIONARY uses conservative scaling. Algorithm 1 converts confidence into action through a margin multiplier $m=1+0.5(1-c)$, lowering the CPU target, lengthening cooldowns, and adding replicas as confidence drops. The Resource Efficiency Index, $\mathrm{REI} = \alpha S_{\mathrm{SLO}} + \beta S_{\mathrm{eff}} + \gamma S_{\mathrm{stab}}$, is the evaluation instrument that lets the paper weigh these gains against resource cost.

What would settle it

Replay the same production traces on a real container-orchestration cluster and compare AAPA against the default reactive autoscaler under realistic nonuniform arrivals, node contention, and cold-start cascades; if AAPA's SLO advantage narrows to statistical noise, the central claim that archetype-aware uncertainty-based scaling beats reactive scaling in practice is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that autoscaling decisions improve when the controller knows what kind of workload it is facing and how sure it is about that classification. AAPA operationalizes this by classifying 60-minute windows into four archetypes with a gradient-boosted classifier calibrated by beta calibration; each archetype has its own target CPU, cooldown, and provisioning strategy, and low-confidence predictions automatically become more conservative. In the simulated evaluation, the largest performance gains are on stationary-noisy workloads, where the SLO violation rate drops from 3.6% to 1.8%, and on spike workloads, where response time improves by roughly 47%, while periodic workloads show no significant performance difference and thus no reason to pay the extra resource cost. The paper concludes that machine-learning-based autoscaling shifts rather than eliminates the cost-performance tradeoff, and recommends AAPA for latency-critical services and hybrid deployments.

Load-bearing premise

The evaluation rests on a discrete-event simulator that is assumed to reproduce real Kubernetes autoscaling behavior, including 2-second pod startup latency, 1-minute metric aggregation, FIFO request queuing, and uniform request arrival within each minute; if actual cluster dynamics differ, the measured SLO and latency improvements may not carry over.

Editorial extensions

If this is right

  • For latency-sensitive, spike-prone services, AAPA can meet SLO targets that the default reactive autoscaler misses, at the price of pre-warmed idle capacity.
  • For periodic or cost-sensitive workloads, the results imply simple reactive scaling remains sufficient, so the right deployment is a hybrid mix of autoscalers rather than one global policy.
  • The REI metric gives operators a single score for comparing autoscalers across performance, efficiency, and stability, and the reported sensitivity analysis indicates rankings are stable to small weight changes.
  • The release of AAPAset provides a reproducible, weakly labeled benchmark for training and comparing workload-archetype classifiers and autoscalers.
  • The general conclusion that ML shifts rather than eliminates the autoscaling cost-performance tradeoff implies that any predictive autoscaler should be judged by its full cost curve, not by SLO gains alone.

Reading between the lines

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

  • A natural extension is to make the confidence-based margin a separate control knob on top of any predictive autoscaler, not just the four archetype policies; one could test whether calibrated confidence alone, without archetype selection, explains most of the SLO gains.
  • The weak-supervision labeling pipeline suggests a broader recipe: domain heuristics over statistical features can generate labeled training data for other systems behaviors where manual labels are scarce, such as contention episodes or failure modes.
  • The strongest tests would be online: run AAPA with the feedback loop enabled under workload drift, and check whether REI-triggered recalibration keeps the SLO advantage stable over weeks; the paper's static temporal split does not exercise this.
  • The 500ms SLO and the REI weights are choices; a testable cross-check is to re-run the comparison with different SLO thresholds and see whether the archetype ordering and the 'shift not eliminate tradeoff' conclusion persist.
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 presents AAPA, an archetype-aware predictive autoscaler for Kubernetes serverless workloads. AAPA classifies workload windows into four archetypes (SPIKE, PERIODIC, RAMP, STATIONARY) using a LightGBM classifier trained on AAPAset, a weakly labeled dataset of 300,000 windows derived from Azure Functions traces. The autoscaler then applies archetype-specific scaling policies with confidence-based adjustments, and is evaluated in a SimPy discrete-event simulation against Kubernetes HPA and a generic predictive autoscaler. The paper claims 99.8% classification accuracy, SLO violation reductions up to 50%, latency reductions of 40%, and proposes the Resource Efficiency Index (REI) as a unified evaluation metric.

Significance. If the headline claims were supportable, the paper would offer a useful contribution: a released simulator and dataset, a no-per-application-training classifier, archetype-specific scaling policies, and a multi-objective evaluation metric. The reproducibility aspects (code and dataset links, repeated trials, Wilcoxon tests) are commendable. However, several central quantitative claims are contradicted by the paper's own reported results: the 40% latency reduction is not present in Fig. 2d, the REI claims are contradicted by Table VI Part B, and the 99.8% accuracy is partly circular because labels are generated by deterministic rules over the same features the classifier consumes. The current significance is therefore lower than advertised, and the paper needs substantial revision before the stated contributions can be accepted.

major comments (5)
  1. [Abstract and Fig. 2d] The abstract's claim that AAPA "lowers latency by 40%" is not supported by Fig. 2d, where the response-time improvement percentages relative to HPA are -47% for SPIKE, -5% for PERIODIC, -3% for RAMP, and only +21% for STATIONARY. No P95/P99 response-latency tables are reported even though Section IV.D lists them as evaluation metrics. The central latency claim must be removed or substantiated with the missing tail-latency data.
  2. [Section V.B and Table VI Part A] For SPIKE workloads, Section V.B reports AAPA's SLO violation rate as 17.5% versus HPA's 16.3%, meaning AAPA is slightly worse, yet Table VI Part A reports p=0.008 for SLO violations and the text characterizes the statistical tests as confirming "key performance improvements." A statistically significant difference in the wrong direction is being presented as an improvement; the direction of each comparison must be reported and interpreted correctly.
  3. [Table VI Part B and Section III.C] Section III.C states that REI uses default weights α=0.5, β=0.3, γ=0.2, but Table VI Part B uses a baseline of α=0.4, β=0.3, γ=0.3. More importantly, every row of Table VI Part B reports HPA>AAPA for REI, which directly contradicts the Introduction's claim that AAPA achieves "steady gains on our proposed Resource Efficiency Index." The REI sensitivity analysis therefore undermines the paper's stated conclusions and needs to be reconciled with the narrative.
  4. [Sections III.A.2 and V.A] The 99.8% classification accuracy is circular as a measure of archetype recognition: the labels are generated by deterministic labeling functions over the same 37 features that the classifier consumes, so the accuracy primarily measures how well the classifier reproduces the hand-written rules. The paper should explicitly acknowledge this and avoid claiming that the accuracy validates recognition of real workload archetypes.
  5. [Section IV.A and Table IV] Section IV.A states that the test set has 35% SPIKE, 30% STATIONARY, 25% PERIODIC, and 10% RAMP. The confusion matrix in Table IV, however, contains 70.2% PERIODIC, 17.6% SPIKE, 12.0% STATIONARY, and 0.2% RAMP. This discrepancy changes the interpretation of the reported accuracy and must be reconciled, either by explaining the sampling procedure or correcting the stated test-set distribution.
minor comments (6)
  1. [Section III.C, Algorithm 1] The text says lower-confidence predictions yield "higher CPU thresholds," but Algorithm 1 line 2 reduces the CPU target when confidence is low (cpu_adj = cpu_target · (1 - 0.2(1-c))). The wording should be corrected to say lower CPU target or the algorithm should be changed.
  2. [Section V.A and Table IV] The text reports RAMP precision of 95.0% and recall of 96.6%. From the confusion matrix, RAMP precision is 57/57 = 100% and recall is 57/60 = 95.0%, so the reported numbers are inconsistent with the table.
  3. [Section IV.B] Section IV.B states that results are reported with 95% confidence intervals, but no confidence intervals appear in the text, tables, or figures. Please either add them or remove the statement.
  4. [Section V.A] The overall dataset distribution (70.2% PERIODIC, 0.2% RAMP) differs sharply from the test-set distribution stated in Section IV.A (25% PERIODIC, 10% RAMP). The discrepancy is not explained by the temporal split and needs clarification.
  5. [Section III.B.1] Beta calibration is originally defined for binary classifiers; the paper should briefly explain how it is extended to the four-class probability vector produced by LightGBM.
  6. [Fig. 2d] The figure caption should state explicitly which row corresponds to SLO violations, response time, and resource usage, and should define the sign convention for "improvement" (positive or green meaning AAPA is better). The current caption is ambiguous.

Circularity Check

1 steps flagged · score 5.0 of 10

The headline '99.8% accurate classifier' claim is circular by construction: AAPAset labels are deterministic Boolean rules over the same 37 features the LightGBM classifier consumes, so test accuracy measures rule-reproduction consistency, not independently validated archetype recognition. The simulated SLO/latency results themselves are not circular.

  1. self definitional [Section III.A.2 (Weak Supervision) and Section V.A (Workload Classification Performance).]
    "Each LF is a Boolean rule over extracted features: •SPIKE: high kurtosis (>10) and max-to-median ratio (>20) •PERIODIC: low spectral entropy (<0.5) and high autocorrelation (>0.6) •RAMP: consistent linear slope (R2 >0.8) •STATIONARY: low standard deviation and low spectral energy variation. ... Label quality is validated by training a LightGBM classifier, which achieves 99.8% test accuracy, indicating high intra-class consistency."

    The ground-truth labels are generated by deterministic Boolean rules over the exact 37 features (kurtosis, spectral entropy, autocorrelation, slope, etc.) that the LightGBM classifier is trained on. The reported 99.8% test accuracy therefore measures how well the classifier learned to invert the labeling functions, i.e., agreement between two functions of the same inputs. With expressive models and deterministic label rules, near-perfect held-out accuracy is expected by construction; no independent human-annotated ground truth validates that archetypes are real semantic classes. Presenting this accuracy as 'validated' label quality and later claiming the weak supervision approach 'successfully labeled the dataset with high fidelity' is self-referential.

full rationale

The paper's central autoscaling claim is not circular: AAPA's archetype-specific parameters (Table III) are fixed design choices, and Section IV-V performance is measured by replaying real Azure invocation traces through a SimPy discrete-event simulator with pod-startup latency and FIFO queuing. SLO violation rates and response times are emergent quantities, not fitted to the reported metric. No parameter is fitted to the test-set outcome. The REI metric is a defined weighted sum (Eq. 1) used to compare results, not a derivation. Self-citations [8], [9], [26] appear only in contextual citation clusters or the reference list and carry no load-bearing uniqueness or ansatz claims, so they do not raise the score. The circularity that does exist is the dataset-validation loop: AAPAset labels are Boolean rules over the classifier's own features, so the headline 99.8% accuracy (Conclusion: 'guided by a 99.8%-accurate classifier') and the calibrated confidence attached to each archetype prediction reduce by construction to rule-reproduction consistency. This is a genuine self-definitional step in a prominently marketed component, but the simulated SLO/latency outcomes remain independently measured, giving partial rather than total circularity (score 5). Separately, the Abstract's 'lowers latency by 40%' is not supported by Fig. 2d (response-time improvement is -47% SPIKE, -5% PERIODIC, -3% RAMP, +21% STATIONARY), and Table VI Part B shows HPA>AAPA on REI under every weight configuration; these are internal-consistency and correctness concerns, not circularity, and per the review rules they are noted but do not further raise the circularity score.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim relies on hand-set labeling thresholds, archetype policy parameters, REI weights, and simulation fidelity. No new physical entities are introduced; REI is a composite metric rather than an invented entity.

free parameters (5)
  • LF thresholds: kurtosis >10, max-to-median >20, spectral entropy <0.5, autocorrelation >0.6, R^2 >0.8 = 10, 20, 0.5, 0.6, 0.8
    Hand-chosen thresholds in the labeling functions (Sec. III.A.2) that determine archetype labels; changing them changes the dataset and classifier behavior.
  • Archetype scaling parameters (Table III) = CPU targets 30/75/60/55%; cooldowns 20/3/7/12 min
    Hand-selected policy parameters per archetype; no sensitivity analysis is shown for these values.
  • Uncertainty adjustment coefficients = 0.5 and 0.2 in Algorithm 1
    Arbitrary constants that set how strongly confidence modulates margin, cooldown, and CPU target.
  • REI weights = alpha=0.5, beta=0.3, gamma=0.2 (Sec. III.C) but alpha=0.4, beta=0.3, gamma=0.3 in Table VI
    Chosen weighting of SLO, efficiency, stability; inconsistent across the paper and changes the REI ranking.
  • Window and stride sizes = 60 min, 10 min
    Fixed analysis parameters that affect the number of windows and feature values.
assumptions (4)
  • domain assumption The 2019 Azure Functions trace is representative of current serverless workloads
    Dataset used to build AAPAset and simulate workloads; generalization to other platforms is acknowledged as a threat.
  • domain assumption The weak supervision labeling functions correctly capture the four behavioral archetypes
    If LFs mislabel, the classifier and strategy selection inherit the error; the paper only validates intra-class consistency.
  • domain assumption The SimPy simulation accurately models Kubernetes autoscaling dynamics
    Pod startup 2s, 1-min aggregation, FIFO queuing, uniform within-minute arrivals; simplifies real networking and contention.
  • standard math Statistical/spectral features used are valid discriminators of archetype
    Autocorrelation, spectral entropy, kurtosis are standard time-series tools; their adequacy for this task is assumed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AAPA: An Archetype-Aware Predictive Autoscaler with Uncertainty Quantification for Serverless Workloads on Kubernetes." pith.science (2026). https://pith.science/paper/7QONCYP2

@misc{pith2026250705653,
  author       = {Pith},
  title        = {Pith review of: AAPA: An Archetype-Aware Predictive Autoscaler with Uncertainty Quantification for Serverless Workloads on Kubernetes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7QONCYP2}},
  note         = {Machine review of arXiv:2507.05653}
}
read the original abstract

Serverless platforms such as Kubernetes are increasingly adopted in high-performance computing, yet autoscaling remains challenging under highly dynamic and heterogeneous workloads. Existing approaches often rely on uniform reactive policies or unconditioned predictive models, ignoring both workload semantics and prediction uncertainty. We present AAPA, an archetype-aware predictive autoscaler that classifies workloads into four behavioral patterns -- SPIKE, PERIODIC, RAMP, and STATIONARY -- and applies tailored scaling strategies with confidence-based adjustments. To support reproducible evaluation, we release AAPAset, a weakly labeled dataset of 300,000 Azure Functions workload windows spanning diverse patterns. AAPA reduces SLO violations by up to 50% and lowers latency by 40% compared to Kubernetes HPA, albeit at 2-8x higher resource usage under spike-dominated conditions. To assess trade-offs, we propose the Resource Efficiency Index (REI), a unified metric balancing performance, cost, and scaling smoothness. Our results demonstrate the importance of modeling workload heterogeneity and uncertainty in autoscaling design.

Figures

Figures reproduced from arXiv: 2507.05653 by the authors.

Figure 1
Figure 1. The overview of AAPA architecture: (A) AAPAset Construction, (B) AAPA, and (C) Evaluation and Feedback. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Comprehensive analysis of autoscaling strategies: (a) Performance scores focusing on user experience, (b) Cost-performance scatter plot, (c) Balanced [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

26 extracted references · 23 canonical work pages

  1. [1]

    Serverless in the wild: Characterizing and optimizing the serverless workload at a large cloud provider,

    M. Shahrad, R. Fonseca, I. Goiri, G. Chaudhry, P. Batum, J. Cooke, E. Laureano, C. Tresness, M. Russinovich, and R. Bianchini, “Serverless in the wild: Characterizing and optimizing the serverless workload at a large cloud provider,” in2020 USENIX Annual Technical Conference (USENIX ATC 20), 2020, pp. 205–218

    work page 2020

  2. [2]

    The state of serverless appli- cations: Collection, characterization, and community consensus,

    S. Eismann, J. Scheuner, E. Van Eyk, M. Schwinger, J. Grohmann, N. Herbst, C. L. Abad, and A. Iosup, “The state of serverless appli- cations: Collection, characterization, and community consensus,”IEEE Transactions on Software Engineering, vol. 48, no. 10, pp. 4152–4166, 2021

    work page 2021

  3. [3]

    Dynamic resource allocation in serverless architechtures using ai-based forecasting

    A. A. Soni and J. A. Soni, “Dynamic resource allocation in serverless architechtures using ai-based forecasting.”

    work page

  4. [4]

    Subspace structural constraint-based discriminative feature learning via nonnegative low rank representation,

    A. Li, X. Liu, Y . Wang, D. Chen, K. Lin, G. Sun, and H. Jiang, “Subspace structural constraint-based discriminative feature learning via nonnegative low rank representation,”PloS one, vol. 14, no. 5, p. e0215450, 2019

    work page 2019

  5. [5]

    Multi-level ml based burst-aware autoscaling for slo assurance and cost efficiency,

    C. Meng, H. Tong, T. Wu, M. Pan, and Y . Yu, “Multi-level ml based burst-aware autoscaling for slo assurance and cost efficiency,”arXiv preprint arXiv:2402.12962, 2024

    arXiv 2024

  6. [6]

    Harmonizing efficiency and practicability: optimizing resource utilization in serverless computing with jiagu,

    Q. Liu, Y . Yang, D. Du, Y . Xia, P. Zhang, J. Feng, J. R. Larus, and H. Chen, “Harmonizing efficiency and practicability: optimizing resource utilization in serverless computing with jiagu,” in2024 USENIX Annual Technical Conference (USENIX ATC 24), 2024, pp. 1–17

    work page 2024

  7. [7]

    With great freedom comes great opportunity: Rethinking resource allocation for serverless functions,

    M. Bilal, M. Canini, R. Fonseca, and R. Rodrigues, “With great freedom comes great opportunity: Rethinking resource allocation for serverless functions,” inProceedings of the Eighteenth European Conference on Computer Systems, 2023, pp. 381–397

    work page 2023

  8. [8]

    Amp4ec: Adaptive model partitioning framework for efficient deep learning inference in edge computing environments,

    G. Zhang, W. Guo, Z. Tan, and H. Jiang, “Amp4ec: Adaptive model partitioning framework for efficient deep learning inference in edge computing environments,”arXiv preprint arXiv:2504.00407, 2025

    arXiv 2025

Show all 26 references
  1. [9]

    KIS-S: A GPU- Aware Kubernetes Inference Simulator with RL-Based Auto-Scaling,

    G. Zhang, W. Guo, Z. Tan, Q. Guan, and H. Jiang, “KIS-S: A GPU- Aware Kubernetes Inference Simulator with RL-Based Auto-Scaling,” arXiv preprint arXiv:2507.07932, 2025, submitted to IPCCC 2025. [Online]. Available: https://arxiv.org/abs/2507.07932

    2025 arXiv

  2. [10]

    Data pipeline approaches in serverless computing: a taxonomy, review, and research trends,

    A. Mampage, S. Karunasekera, and R. Buyya, “Data pipeline approaches in serverless computing: a taxonomy, review, and research trends,” Journal of Big Data, vol. 11, no. 1, pp. 1–45, 2024

    2024

  3. [11]

    Serverless comput- ing: State-of-the-art and performance challenges,

    P. Raith, T. Rausch, A. Furutanpey, and S. Dustdar, “Serverless comput- ing: State-of-the-art and performance challenges,” inCompanion of the 2023 ACM/SPEC International Conference on Performance Engineer- ing, 2023, pp. 57–63

    2023

  4. [12]

    (2023) Horizontal pod autoscaler

    Kubernetes Documentation. (2023) Horizontal pod autoscaler. [Online]. Available: https://kubernetes.io/docs/tasks/run-application/ horizontal-pod-autoscale/

    2023

  5. [13]

    Thomperoo

    J. Thomperoo. (2023) Predictive horizontal pod au- toscaler. [Online]. Available: https://github.com/jthomperoo/ predictive-horizontal-pod-autoscaler

    2023

  6. [14]

    Cycle-stealing in load- imbalanced hpc applications,

    P. H. Chen, A. Bali, S. Yang, P. Haghi, C. Knox, B. Li, A. A. Abouelmagd, A. Skjellum, and M. Herbordt, “Cycle-stealing in load- imbalanced hpc applications,” in2024 IEEE High Performance Extreme Computing Conference (HPEC). IEEE, 2024, pp. 1–8

    2024

  7. [15]

    Syndeo: Portable ray clusters with secure containerization,

    W. Li, E. Gregori, A. Reuther, and J. Kepner, “Syndeo: Portable ray clusters with secure containerization,” inIEEE High Performance Extreme Computing Conference (HPEC), 2024, pp. 1–7

    2024

  8. [16]

    Firm: An intelligent fine-grained resource management framework for slo-oriented microservices,

    H. Qiu, S. S. Banerjee, S. Jha, Z. T. Kalbarczyk, and R. K. Iyer, “Firm: An intelligent fine-grained resource management framework for slo-oriented microservices,” in14th USENIX Symposium on Operating Systems Design and Implementation (OSDI 20), 2020, pp. 805–825

    2020

  9. [17]

    Aware: Automate workload autoscaling with reinforcement learning in production cloud systems,

    H. Qiu, W. Mao, A. Patke, C. Wang, H. Franke, Z. T. Kalbarczyk, T. Bas ¸ar, and R. K. Iyer, “Aware: Automate workload autoscaling with reinforcement learning in production cloud systems,” in2023 USENIX Annual Technical Conference (USENIX ATC 23), 2023, pp. 315–330

    2023

  10. [18]

    Magicscaler: Uncertainty-aware, predictive autoscaling,

    Z. Yanget al., “Magicscaler: Uncertainty-aware, predictive autoscaling,” inProceedings of the VLDB Endowment, vol. 16, no. 12, 2023, pp. 3808–3821

    2023

  11. [19]

    Rl-based serverless container autoscaler,

    L. Ninget al., “Rl-based serverless container autoscaler,” MIT PRIMES- AT, Tech. Rep., 2023

    2023

  12. [20]

    Snorkel: Rapid training data creation with weak supervision,

    A. Ratner, S. H. Bach, H. Ehrenberg, J. Fries, S. Wu, and C. R ´e, “Snorkel: Rapid training data creation with weak supervision,” in Proceedings of the VLDB Endowment, vol. 11, no. 3, 2017, pp. 269–282

    2017

  13. [21]

    Enhancing machine learning- based autoscaling for cloud resource orchestration,

    I. Pintye, J. Kov ´acs, and R. Lovas, “Enhancing machine learning- based autoscaling for cloud resource orchestration,”Journal of Grid Computing, vol. 22, no. 4, pp. 1–31, 2024

    2024

  14. [22]

    Collecting a large scale dataset for classifying fake news tweets using weak supervision,

    S. Helmstetter and H. Paulheim, “Collecting a large scale dataset for classifying fake news tweets using weak supervision,”Future Internet, vol. 13, no. 5, p. 114, 2021

    2021

  15. [23]

    Beta calibration: a well-founded and easily implemented improvement on logistic calibration for binary classifiers,

    M. Kull, T. Silva Filho, and P. Flach, “Beta calibration: a well-founded and easily implemented improvement on logistic calibration for binary classifiers,”Proceedings of Machine Learning Research, vol. 54, pp. 623–631, 2017

    2017

  16. [24]

    Machine learning for predictive resource scaling of micro-services,

    M. Fern ´andezet al., “Machine learning for predictive resource scaling of micro-services,” inProceedings of the 24th International Middleware Conference, 2023, pp. 226–239

    2023

  17. [25]

    Ai-powered cloud resource management: Machine learning for dynamic autoscaling and cost optimization,

    L. Emma, “Ai-powered cloud resource management: Machine learning for dynamic autoscaling and cost optimization,” 2025

    2025

  18. [26]

    Semi- supervised subspace learning for pattern classification via robust low rank constraint,

    A. Li, R. An, D. Chen, G. Sun, X. Liu, Q. Wu, and H. Jiang, “Semi- supervised subspace learning for pattern classification via robust low rank constraint,”Mobile Networks and Applications, vol. 25, pp. 2258– 2269, 2020

    2020

Pith tools

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