Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

Can Machine Learning Support the Selection of Studies for Systematic Literature Review Updates?

T0 review · 3 major / 4 minor · reviewed 2026-08-08 · deepseek-v4-flash

Pith's one-line read Machine learning can trim a systematic-review update's screening load by a third, but cannot replace human reviewers.

desk verdict Honest negative result with open artifacts, but headline numbers are best-of-grid selected on the test set, so treat them as optimistic. read the letter →

arxiv 2502.08050 v1 pith:WGC4PSLN submitted 2025-02-12 cs.SE

classification cs.SE
keywords systematicliteraturereviewSLRupdatestudyselectionmachinelearningtextclassificationRandomForestSupportVectoreffortreduction
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

The paper asks whether machine-learning text classifiers can take over study selection when a systematic literature review is updated. Using a real update and a consensus oracle from three experienced reviewers, it finds the best model scores only 0.33 F-score, too low for automation. The same models, tuned for recall, can discard about a third of candidate studies without missing any included study, cutting manual screening effort. Treating the model as an additional reviewer worsens agreement with the final result relative to human pairs. The authors conclude that ML can support effort reduction but that experienced human reviewers remain necessary for rigorous selection.

What carries the argument

The central object is a supervised text-classification pipeline: candidate studies are converted with TF-IDF into numerical feature vectors, the most informative 1,200 features are selected by statistical tests (Chi-squared, Pearson correlation, or ANOVA F), and Random Forest or SVM classifiers are tuned with cross-validated hyperparameter search and then applied to a holdout testing set. The oracle is the consensus of three experienced researchers after full-text reading; for reviewer-pair analysis, the model's inclusion probabilities are mapped to the same 0/1/2 scale used by human reviewers and agreement is measured by Euclidean distance.

What would settle it

Train the same SVM and Random Forest pipelines on a second SLR update for which the original review's full list of excluded papers is available, using those true exclusions as the negative training set; if the 100%-recall discard fraction is not close to 33.9% or the F-score is not close to 0.33 on the holdout, the reported figures are an artifact of the surrogate negatives rather than a general property of ML for SLR updates.

Watch

Extended reading notes

Core claim

The paper establishes that, for the studied SLR update, supervised ML text classifiers trained on the original review's data cannot reliably automate the study-selection step: the best configuration (Random Forest with ANOVA-F feature selection and 1,200 features) reaches an F-score of 0.33 at the default inclusion threshold. It also establishes that a recall-optimized SVM can act as a safe pre-filter, excluding 33.9% of the candidate pool while keeping recall at 100% against the three-reviewer oracle. When the model is paired with a human as if it were an additional reviewer, agreement with the final curated result is worse than for any human-human pair, so ML support does not substitute for a human reviewer in the initial screening.

Load-bearing premise

The negative training examples are backward-snowballed references from the original SLR rather than the studies the original review actually excluded, so the measured F-score and the 33.9% safe-discard figure assume those negatives represent the real excluded pool.

Editorial extensions

If this is right

  • A recall-tuned ML classifier can be used as a pre-filter that removes roughly one-third of candidate studies while keeping recall at 100% relative to the final curated list.
  • Using the F-score-tuned model as a replacement reviewer degrades agreement with the final consensus; human-only pairs gave Euclidean distances around 8.9–9.2 versus 11.5–11.8 for human-ML pairs.
  • Raising the inclusion threshold trades recall for effort: at 97.4% recall the model could discard 48.3% of studies, but one included study would be lost.
  • For serious SLR updates, at least two experienced human reviewers should perform the initial screening.
  • The reported results come from a holdout testing set with no data leakage, unlike earlier work that reported much higher F-scores but applied feature selection before the train-test split.

Reading between the lines

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

  • Because the negative training examples were drawn by backward snowballing rather than from the original SLR's actual exclusion list, the 33.9% safe-discard figure and the 0.33 F-score could shift if real exclusions differ systematically from snowballed references.
  • The safe-discard fraction and F-score are measured on one update with only 38 included papers among 551 candidates; on reviews with different inclusion rates or topic breadth, the same pipeline would likely produce different numbers.
  • The human-ML pairing result may depend on how the model's probability output is converted to the reviewer scale; using the model only as a flag to discard low-probability papers, rather than averaging its score with a human's, might preserve more of the human signal.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper reports an empirical evaluation of supervised machine learning models (Random Forest and Support Vector Machines) for supporting study selection in an update of a specific systematic literature review in software engineering. Using 128 training studies (45 included, 83 excluded via backward snowballing) and a holdout test set of 551 studies (38 included, 513 excluded) from the update, the authors train classifiers with various feature-selection methods and feature counts, reporting a best F-score of 0.33 for RF (RQ1) and an SVM configuration achieving 100% recall while excluding 33.9% of studies (RQ2). For RQ3, they compare human-only and human-ML reviewer pairs using Euclidean distance to the final consensus. The main conclusions are that ML cannot reliably automate selection but can reduce screening effort without loss of evidence, and that human-only pairs outperform human-ML pairs.

Significance. The study provides a rare, openly available dataset and a reproducible pipeline, and its holdout evaluation is a useful corrective to earlier work (e.g., Watanabe et al.) that used leakage-prone cross-validation. If the quantitative results were unbiased, the 33.9% effort-reduction figure would be a practically meaningful benchmark. However, the strength of the conclusions is limited by the single-dataset design and by the selection-bias and representativeness issues detailed below.

major comments (3)
  1. [IV-B, V-A, V-B] The reported best results are selected from a large grid of configurations evaluated on the same holdout test set. The paper experiments with two algorithms, three feature-selection methods, feature counts from 900 to 1500, and two tuning targets, and then reports 'the best result' for RQ1 (RF, Anova-F, 1200 features, F-score 0.33) and RQ2 (SVM, Pearson, 1200 features, recall 1.0, 33.9% reduction). Because the test set is used for this outer model selection, these numbers are maxima over the grid rather than unbiased estimates for a pre-specified model. With only 38 positive test instances, many configurations can attain 100% recall, and the configuration that discards the most negatives is likely an outlier. This threatens the internal validity of the headline quantitative claims. The authors should either select the configuration using only the training set (e.g., via nested cross-validation) or report the complete distribution of results across all configurations and provide a corrected (e.g., best-vs-typical) analysis.
  2. [IV-A] The 83 negative training examples were obtained by backward snowballing from the original SLR's included references, rather than from the original SLR's actually excluded studies. This is a non-representative sample: backward-snowballed references may be more topically related to included studies, older, or otherwise different from the true negatives in the update. If these differences affect classification, the measured F-score and the 33.9% safe-discard figure may not generalize, even to the studied update. The authors acknowledge the lack of access to the original excluded list, but this is a load-bearing assumption for the training data. Please provide evidence of representativeness (e.g., topic distributions, temporal overlap, or a sensitivity analysis retraining on different negative samples) or accordingly temper the quantitative conclusions.
  3. [V-B] The claim of a 33.9% effort reduction 'without loss of evidence' rests on a single test set with only 38 positive instances. With 38/38 recalled, the lower 95% confidence bound on recall is about 90.7%, so the sample size provides limited evidence for the 'no evidence loss' guarantee. Moreover, the 33.9% figure is a point estimate with no measure of uncertainty; the paper should report confidence intervals or a sensitivity analysis over thresholds (Table III already shows the trade-off). As written, the guarantee-like phrasing overstates what can be concluded from one dataset.
minor comments (4)
  1. [IV-B] The preliminary tests that led to selecting SVM and RF out of the five classifiers are not reported. Since classifier selection is part of the pipeline, please provide the results of these initial tests or state that they were performed on the training set only.
  2. [V-C] The mapping of RF probabilities to the 0-2 scale uses thresholds (0.50, 0.60) that are adjusted to match the reviewers' assessment distribution. This adjustment appears to be done on the test set; please clarify and discuss the potential for bias.
  3. [Table IV] The Euclidean distances are small (e.g., 8 to 17 on a scale where the maximum distance per study is 2), and the claim that human-only pairs are 'much better aligned' would benefit from an effect-size measure or a test of whether the differences exceed what would be expected given that the oracle is derived from the trio itself.
  4. [Figure 1] In Figure 1, the label 'ML Training Set' appears twice, which may confuse readers about the direction of the training and testing splits; please correct the labeling.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: holdout evaluation against a manually curated oracle; reported 'best' figures carry selection-bias caveats but are not definitional.

full rationale

The derivation chain is empirical rather than definitional. The ML models are trained on the original SLR's included studies and backward-snowballed negatives (Section IV-A), then scored on a holdout set of 551 update studies against a manually curated oracle (Section V). The reported F-score of 0.33 and the 33.9% effort reduction are measured performance metrics of the trained models on that holdout set, not quantities reconstructed from the training labels or from the oracle by construction. The effort-reduction definition (Section III, RQ2) is a metric, and its value is an observed specificity at 100% recall, not a fitted parameter renamed as a prediction. Self-citations ([13], [25]) motivate algorithm choice and supply the dataset, but the conclusions do not reduce to those citations: the oracle is the authors' own manual curation, and the model outputs are compared with it directly. The main caveats are that the paper reports the best configuration among many feature-selection, algorithm, and tuning-target combinations evaluated on the test set (Sections IV-B and V), which makes the headline numbers optimistic, and that the training negatives were obtained by backward snowballing rather than sampled from the original SLR's exclusion decisions (Section IV-A); both are validity threats, not circularity. Therefore no circular step is exhibited.

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

All substantive inputs are empirical; no new theoretical entities are introduced.

free parameters (5)
  • Number of features (TF-IDF) = 1200
    Selected from a 900-1500 range as the best on the training set; a modeling choice, not a physical constant.
  • Inclusion probability threshold (RQ2) = 0.5 (default); 0.75-0.85 explored
    The 33.9% effort-reduction figure uses the default 0.5 threshold; higher thresholds trade recall for effort.
  • RQ3 vote-mapping thresholds = 0.51-0.60 as 'unsure'
    Chosen to match the human reviewers' vote distribution on the test set; a post hoc calibration.
  • ML hyperparameters (GridSearch) = not listed in text
    Tuned via GridSearch with k-fold CV on the training set; specific values are in the repository.
  • Feature selection method = Anova-F (RQ1), Pearson (RQ2)
    Chosen based on best F-score/recall during experiments; a configuration choice.
assumptions (4)
  • domain assumption The final consensus list of included/excluded studies produced by the three researchers is correct ground truth.
    Used as the oracle for all RQs; any errors in the manual curation propagate to the evaluation (Section IV-A).
  • domain assumption Title, abstract, and keywords contain sufficient information for study selection.
    Both human initial screening and ML use the same fields; final decisions used full text, which may disadvantage abstract-only screening (Section IV-A).
  • ad hoc to paper The backward-snowballed excluded studies are representative of the excluded studies in the SLR update.
    This training set construction is unique to this paper and is not grounded in the original screening; it affects what the model learns (Section IV-A).
  • domain assumption The 551 papers in the testing set are representative of the SLR update population.
    A single SLR update case study; generalization is not claimed (Section VII).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Can Machine Learning Support the Selection of Studies for Systematic Literature Review Updates?." pith.science (2026). https://pith.science/paper/WGC4PSLN

@misc{pith2026250208050,
  author       = {Pith},
  title        = {Pith review of: Can Machine Learning Support the Selection of Studies for Systematic Literature Review Updates?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WGC4PSLN}},
  note         = {Machine review of arXiv:2502.08050}
}
read the original abstract

[Background] Systematic literature reviews (SLRs) are essential for synthesizing evidence in Software Engineering (SE), but keeping them up-to-date requires substantial effort. Study selection, one of the most labor-intensive steps, involves reviewing numerous studies and requires multiple reviewers to minimize bias and avoid loss of evidence. [Objective] This study aims to evaluate if Machine Learning (ML) text classification models can support reviewers in the study selection for SLR updates. [Method] We reproduce the study selection of an SLR update performed by three SE researchers. We trained two supervised ML models (Random Forest and Support Vector Machines) with different configurations using data from the original SLR. We calculated the study selection effectiveness of the ML models for the SLR update in terms of precision, recall, and F-measure. We also compared the performance of human-ML pairs with human-only pairs when selecting studies. [Results] The ML models achieved a modest F-score of 0.33, which is insufficient for reliable automation. However, we found that such models can reduce the study selection effort by 33.9% without loss of evidence (keeping a 100% recall). Our analysis also showed that the initial screening by pairs of human reviewers produces results that are much better aligned with the final SLR update result. [Conclusion] Based on our results, we conclude that although ML models can help reduce the effort involved in SLR updates, achieving rigorous and reliable outcomes still requires the expertise of experienced human reviewers for the initial screening phase.

Figures

Figures reproduced from arXiv: 2502.08050 by the authors.

Figure 1
Figure 1. Data collection process Metrics Analysis (F-score, Recall) Number of studies excluded by ML Models Similarity Analysis (Euclidean Distance of aggregated outcomes to Final Result) RQ1 RQ2 RQ3 File Validation Text Filtering (NLP) Text Vectorization (TF-IDF) Best Features Selection ML Models Tuning ML Models Predictions Included Excluded Testing Set .bib .bib Included Excluded Training Set .bib .bib ML Algorithm 1 SVM … view at source ↗
Figure 2
Figure 2. Study design pipeline filtering techniques with Natural Language Processing (NLP) [16], such as Lemmatization and Tokenization, to remove irrelevant characters. Thirdly, we applied Text Vectorization on the filtered texts using Term-frequency/Inverse-Document￾Frequency (TF/IDF), a technique that transforms text data into a numerical matrix of features. Fourthly, we used statistical methods to compute and select the … view at source ↗
Figure 3
Figure 3. illustrates the distribution of the ML predictions’ inclusion probabilities made by RF with this configuration. The parameters tested and selected by GridSearch, as well as the predictions made by our ML model using these parameters, can be found in our online repository. Table II shows that by increasing the inclusion probability threshold, our RF model was able to achieve more accurate results with an F-score of 0… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Reviewers assessment distribution. =>Exclude Unsure Include V [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Predictions Distribution considering ”uncertain” range. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. A Methodological Framework for LLM-Based Mining of Software Repositories

    cs.SE 2025-08 conditional novelty 5.0 of 10

    A rapid review and survey of LLM-based repository mining yield a threat-mitigation map and the six-stage PRIMES 2.0 framework for conducting such studies.

Reference graph

Works this paper leans on

29 extracted references · 28 canonical work pages · cited by 1 Pith paper

  1. [1]

    Identifying, categorizing and mitigating threats to validity in software engineering secondary studies,

    A. Ampatzoglou, S. Bibi, P. Avgeriou, M. Verbeek, and A. Chatzi- georgiou, “Identifying, categorizing and mitigating threats to validity in software engineering secondary studies,” Information and Software Technology, vol. 106, pp. 201 – 230, 2019

  2. [2]

    Artifacts: Can Machine Learning Support the Selection of Studies for Systematic Literature Review Updates?

    M. Costalonga, B. M. N. ao, M. T. Baldassarre, K. R. Felizardo, I. Steinmacher, and M. Kalinowski, “Artifacts: Can Machine Learning Support the Selection of Studies for Systematic Literature Review Updates?” 2024. [Online]. Available: https://doi.org/10.5281/zenodo. 14063837

  3. [3]

    Formalizing a systematic review updating process,

    O. Dieste, M. L ´opez, and F. Ramos, “Formalizing a systematic review updating process,” in 6th Int. Conference on Software Engineering Research, Management and Applications (SERA’08), 2008, pp. 143–150

  4. [4]

    Externalising tacit knowledge of the system- atic review process,

    S. Fabbri, K. Felizardo, F. Ferrari, E. Hernandes, F. Octaviano, E. Naka- gawa, and J. Maldonado, “Externalising tacit knowledge of the system- atic review process,” IET Software, vol. 7, no. 6, pp. 298–307, 2013

  5. [5]

    Using forward snowballing to update systematic reviews in software engineering,

    K. Felizardo, E. Mendes, M. Kalinowski, E. F. Souza, and N. Vi- jaykumar, “Using forward snowballing to update systematic reviews in software engineering,” in International Symposium on Empirical Software Engineering and Measurement (ESEM) , 2016

  6. [6]

    Knowledge management for promoting update of systematic literature reviews: An experience report,

    K. R. Felizardo, E. F. de Souza, T. Malacrida, B. M. Napoleao, F. Petrillo, S. Hall´e, N. L. Vijaykumar, and E. Y . Nakagawa, “Knowledge management for promoting update of systematic literature reviews: An experience report,” in 2020 46th Euromicro Conference on Software Engineering and Advanced Applications (SEAA) , 2020, pp. 471–478

  7. [7]

    Chatgpt application in systematic literature reviews in software engineering: an evaluation of its accuracy to support the selection ac- tivity,

    K. R. Felizardo, M. S. Lima, A. Deizepe, T. U. Conte, and I. Stein- macher, “Chatgpt application in systematic literature reviews in software engineering: an evaluation of its accuracy to support the selection ac- tivity,” in International Symposium on Empirical Software Engineering and Measurement (ESEM) , 2024, p. 25–36

  8. [8]

    A visual analysis approach to update systematic reviews,

    K. Felizardo, E. Nakwgawa, S. MacDonell, and J. Maldonado, “A visual analysis approach to update systematic reviews,” in International Conference on Evaluation and Assessment in Software Engineering (EASE). ACM, 2014, pp. 1–10

Show all 29 references
  1. [9]

    An experience report on update of systematic literature reviews,

    L. Garc ´es, K. Felizardo, L. Oliveira, and E. Nakagawa, “An experience report on update of systematic literature reviews,” in Proceedings of the International Conference on Software Engineering and Knowledge Engineering (SEKE), 2017, pp. 91–96

  2. [10]

    Procedures for performing systematic reviews,

    B. Kitchenham, “Procedures for performing systematic reviews,” Soft- ware Engineering Group - Department of Computer Science - Keele University and Empirical SE - National ICT Australia Ltd, Joint Tech- nical Report TR/SE-0401 (Keele) - 0400011T.1 (NICTA), 2004

  3. [11]

    Kitchenham, D

    B. Kitchenham, D. Budgen, and P. Brereton, Evidence-Based Software Engineering and Systematic Reviews , ser. Chapman & Hall/CRC In- novations in Software Engineering and Software Development Series. Chapman & Hall/CRC, 2015

  4. [12]

    When to update systematic literature reviews in software engineering,

    E. Mendes, C. Wohlin, K. Felizardo, and M. Kalinowski, “When to update systematic literature reviews in software engineering,” Journal of Systems and Software , vol. 167, pp. 110–167, 2020

  5. [13]

    Automated support for searching and selecting evidence in software engineering: A cross- domain systematic mapping,

    B. M. Napole ˜ao, F. Petrillo, and S. Hall ´e, “Automated support for searching and selecting evidence in software engineering: A cross- domain systematic mapping,” in 47th Euromicro Conference on Software Engineering and Advanced Applications (SEAA) , 2021

  6. [14]

    Establishing a search string to detect secondary studies in software engineering,

    B. M. Napole ˜ao, K. R. Felizardo, E. F. d. Souza, F. Petrillo, S. Hall ´e, N. L. Vijaykumar, and E. Y . Nakagawa, “Establishing a search string to detect secondary studies in software engineering,” in 2021 47th Euromi- cro Conference on Software Engineering and Advanced Appli...

  7. [15]

    Emerging results on automated support for searching and selecting evidence for systematic literature review updates,

    B. M. Napole ˜ao, R. Sarkar, S. Hall ´e, F. Petrillo, and M. Kalinowski, “Emerging results on automated support for searching and selecting evidence for systematic literature review updates,” in Proceedings of the 1st IEEE/ACM International Workshop on Methodological Issues wi...

  8. [16]

    Natural Language Toolkit,

    NLTK Team, “Natural Language Toolkit,” https://pypi.org/project/nltk, online; accessed 21 April 2024

  9. [17]

    Feature selection methods for text classification: a systematic literature review,

    J. T. Pintas, L. A. Fernandes, and A. C. B. Garcia, “Feature selection methods for text classification: a systematic literature review,” Artificial Intelligence Review, vol. 54, no. 8, pp. 6149–6200, 2021

  10. [18]

    Runeson, M

    P. Runeson, M. Host, and A. Rainer, Case Study Research in Software Engineering: Guidelines and Examples . John Wiley & Sons, 2012

  11. [19]

    Sklearn Feature Selection ANOV A F,

    Scikit-learn, “Sklearn Feature Selection ANOV A F,” https://scikit-learn. org/stable/modules/generated/sklearn.feature selection.f classif.html, online; accessed 21 April 2024

  12. [20]

    Sklearn Feature Selection Chi2,

    ——, “Sklearn Feature Selection Chi2,” https://scikit-learn.org/stable/ modules/generated/sklearn.feature selection.chi2.html, online; accessed 21 April 2024

  13. [21]

    Sklearn Feature Selection Pearson’s r,

    ——, “Sklearn Feature Selection Pearson’s r,” https://scikit-learn.org/ stable/modules/generated/sklearn.feature selection.r regression.html, online; accessed 21 April 2024

  14. [22]

    A holistic overview of software engineering research strategies,

    K.-J. Stol and B. Fitzgerald, “A holistic overview of software engineering research strategies,” in CESI. IEEE Press, 2015, p. 47–54

  15. [23]

    Reducing efforts of software engineering systematic literature reviews updates using text classification,

    W. M. Watanabe, K. R. Felizardo, A. Candido, E. F. de Souza, J. ao Ede de Campos Neto, and N. L. Vijaykumar, “Reducing efforts of software engineering systematic literature reviews updates using text classification,” Information and Software Technology , vol. 128, 2020

  16. [24]

    A snowballing procedure for systematic literature studies and a replication,

    C. Wohlin, “A snowballing procedure for systematic literature studies and a replication,” in International Conference on Evaluation and Assessment in Software Engineering (EASE) , 2014, pp. 321–330

  17. [25]

    Suc- cessful combination of database search and snowballing for identification of primary studies in systematic literature studies,

    C. Wohlin, M. Kalinowski, K. Romero Felizardo, and E. Mendes, “Suc- cessful combination of database search and snowballing for identification of primary studies in systematic literature studies,” Information and Software Technology, vol. 147, 2022

  18. [26]

    Guidelines for the search strategy to update systematic literature reviews in software engineering,

    C. Wohlin, E. Mendes, K. R. Felizardo, and M. Kalinowski, “Guidelines for the search strategy to update systematic literature reviews in software engineering,” Information and Software Technology , vol. 127, 2020

  19. [27]

    Is it a case study? a critical analysis and guidance,

    C. Wohlin and A. Rainer, “Is it a case study? a critical analysis and guidance,” Journal of Systems and Software , vol. 192, 2022

  20. [28]

    Empirical research in software engineering — a literature survey,

    L. Zhang, J.-H. Tian, J. Jiang, Y . Liu, M.-Y . Pu, and T. Yue, “Empirical research in software engineering — a literature survey,” Journal of Computer Science and Technology , vol. 33, pp. 876–899, 2018

  21. [29]

    Quality assessment of systematic reviews in software engineering: A tertiary study,

    Y . Zhou, H. Zhang, X. Huang, S. Yang, M. A. Babar, and H. Tang, “Quality assessment of systematic reviews in software engineering: A tertiary study,” in International Conference on Evaluation and Assess- ment in Software Engineering (EASE) , 2015, pp. 1–14

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.