Review Papers Claim by Claim

Quick visual summaries (figures first, text second)

Critical findings (visual first, then concise evidence-backed commentary)

1. Performance vs predictors: The Random Forest model reported R2≈0.685 and RMSE≈10.59 on the held-out test set; those values indicate the model explains ~68.5% of variance using ClimSST, SSTA, depth, latitude, and longitude as predictors — a useful, but not mechanistic, ecological predictor set. Paper performance and methods
2. Parallel computing claim: Authors used scikit-learn's n_jobs to vary parallelism (1, 4, -1) on an Intel i7-1165G7 (4 physical/4 logical cores) and measured reduced wall-clock times for training and grid-search, but speedup was non-linear due to overhead and hyperthreading limits — consistent with known scikit-learn/joblib behavior and Amdahl-like overheads for CPU-bound tasks. Paper parallel timing details
3. Data-processing choices that affect validity: The authors removed negative SSTA rows and limited years to 1997–2019; both choices change the training distribution and can bias the model toward heat-induced bleaching events (positive SSTA) but reduce the ability to generalize across non-heat drivers. This selective filtering should be justified with sensitivity analyses and ideally compared with models trained on unfiltered data. Preprocessing & filtering choices
4. Reproducibility and transparency: Paper cites GCBD (BCO‑DMO) and NOAA CRW data sources, but does not publish code or a packaged dataset; this reduces reproducibility. For ML work that claims speed and performance, sharing code + environment (scikit-learn versions, random seeds, exact train/test splits) is essential. Data links and reproducibility statement
5. Comparison with prior literature: Authors contrast their R2 (0.685) with Madireddy et al. (R2≈0.25 reported in that reference) to argue improvement — but direct comparability depends on features, preprocessing, and evaluation protocol; caution is warranted. Madireddy et al. on coral-bleaching ML

Deeper methodological critique (short bullets, evidence-linked)

• Choice of features: Using ClimSST, SSTA, depth, lat/lon is defensible — these are known correlates of bleaching — but absence of species-level, local stressors (pollution, turbidity, overfishing) means omitted-variable bias; authors acknowledge but do not quantify the blindspot. Feature selection and limitations
• Evaluation metrics & 'accuracy' label: The paper reports R2 and also an 'Akurasi' 95.19% — for regression targets 'Percent_Bleaching', percent-accuracy is ambiguous (thresholded?) and should be clarified; high 'accuracy' alongside R2≈0.685 suggests a possible nonstandard definition. Authors should provide predicted vs observed scatter, calibration plots, and error distributions. Metrics reporting ambiguity
• Parallelism implementation limits: Using scikit-learn/joblib (n_jobs) parallelism parallelizes tree building but is still limited by Python process/thread overhead and GIL I/O in parts; for true large-scale speedup one would evaluate distributed frameworks (Spark/MLlib) or native C/C++ GPU implementations. The paper demonstrates practical gains on commodity laptop hardware, which is useful for practitioners, but not a general benchmark for large clusters. Parallel RF in Spark (big-data context)

Practical reproducibility checklist (what the authors should publish next)

1. Full training code (notebook + script), exact scikit-learn version (they list v1.6), Python environment (requirements.txt / conda env), and joblib config.
2. Preprocessing script showing how rows were filtered (exact logic: year filtering, SSTA negative removal) and the final train/test indices and random seed.
3. Predicted vs observed scatter, residual histogram, and calibration curve for Percent_Bleaching; SHAP or permutation importance to quantify feature contributions beyond correlation claims.
4. Raw timing logs for training/grid-search/inference with wall-clock numbers and CPU utilization (so others can reproduce speedups on different hardware).

Conclusions, confidence, and how to falsify the paper's central claims

Bottom-line: the paper demonstrates an applied pipeline (GCBD+NOAA CRW → RF → parallel training) that yields usable predictive skill for coral bleaching at global scales (R2≈0.685). The parallel-computing contribution is practical (multi-core speedups on a laptop) but unsurprising; speedups are limited by overhead and hyperthreading. Confidence in ecological generality is moderate but constrained by preprocessing choices and lack of independent external validation. To falsify or materially change the paper's claims one could:

• Re-train the same model on the full GCBD (without SSTA negative removal or with different year windows) and show that R2 drops substantially (e.g., to prior 0.25 values), demonstrating sensitivity to preprocessing.
• Evaluate the trained model on independent, recent bleaching observations (not used in training) and show poor generalization (much lower R2 / higher RMSE).
• Reproduce the timing experiments on a different machine and show no meaningful wall-clock speedup with n_jobs (contradicting the claim of practical time gain).

Reference to the paper's own description (DOI 10.34010/komputika.v14i1.14982) is provided throughout for traceability. Paper core content

Actionable next steps for the authors (concise)

• Publish code + exact seeds and a smaller reproducible subset of data (or a data-processing container) so others can reproduce timing and metrics.
• Provide sensitivity analyses: show performance with/without SSTA-negative-filter, different year windows, and with alternative features (e.g., local stressors if available).
• Report calibration plots and prediction intervals (RF quantile or conformal methods) to make outputs actionable for conservation decision-making.
• Consider distributed or GPU RF (or TreeSHAP acceleration) if scaling beyond laptop is a future aim — cite distributed RF literature for context. Distributed RF (Spark) for big data