Pith. sign in

REVIEW 1 major objections 1 minor 71 references

Evaluating Machine Learning Models for Post-Wildfire Debris-Flow Prediction

T0 review · 1 major / 1 minor · reviewed 2026-08-08 · deepseek-v4-flash

Pith's one-line read This paper claims that post-wildfire debris-flow prediction has a practical top tier rather than a single best model, with the pre-trained tabular foundation model TabPFN at 0.637 threat score and tuned tree ensembles within 0.017, all…

desk verdict A careful, honest benchmark of TabPFN and tree models for debris-flow prediction, undermined mainly by spatial leakage in the cross-validation—likely fixable with a spatial holdout. read the letter →

arxiv 2608.05265 v1 pith:SRI5ZOP5 submitted 2026-08-05 cs.LG cs.CV

classification cs.LGcs.CV
keywords post-wildfiredebris-flowpredictionmachinelearningtabularfoundationmodelTabPFNSHAPsyntheticdataaugmentationthreatscoreclassimbalance
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

Post-wildfire debris flows are hard to predict because the positive class is rare, the two outcome classes overlap in feature space, and training data are small. This paper compares fifteen models on 1,550 basin-storm records using the same stratified 5-fold splits repeated ten times, and its central claim is that TabPFN—a transformer pre-trained on synthetic tabular tasks that predicts by in-context learning—reaches the highest unaugmented threat score (0.637), statistically tied with the best tree ensembles and far above the four-feature operational logistic baseline (0.152). It further claims that SHAP attributions place short-duration rainfall intensity and storm accumulation at the top of every explained model, and that TabPFN-generated synthetic rows improve all but one model, with the largest gains going to the most feature-restricted and capacity-limited models. A sympathetic reader would care because the result suggests that a tuning-free foundation model can serve operational hazard screening and that model families converge on the same physically sensible rainfall signals.

What carries the argument

The argument is carried by a shared evaluation protocol: fifty stratified 5-fold splits, precomputed once and reused identically across all fifteen models, with fold-fitted imputation and standardization and a common threat-score-optimized hyperparameter search, so that model architecture is the only thing that varies. Within that protocol, TabPFN acts as a pre-trained transformer that classifies in a single forward pass by in-context learning, needing no dataset-specific tuning, while SHAP (a per-observation attribution method that decomposes predictions additively) assigns feature contributions on the same held-out folds; TreeSHAP is used for the tree models and a permutation explainer for TabPFN. For the augmentation component, a class-conditional TabPFN generator produces synthetic rows per fold that are shared identically across all models, and generator fidelity is checked with univariate distribution distances and a multivariate energy distance. These pieces let the paper isolate model family, feature signal, and data augmentation on one small, imbalanced dataset.

What would settle it

Re-run the identical 15-model benchmark with leave-one-fire-out cross-validation, training on all basins from other fires and testing on each held-out fire in turn. If the mean threat score falls well below the reported 0.637, or the ordering within the top tier changes materially, the paper's top-tier ranking and absolute score depend on within-fire leakage rather than skill at new fires.

Watch

Extended reading notes

Core claim

The paper's central discovery is that the correct summary of the benchmark is a practical top tier, not a single dominant model. Under the same fifty stratified validation splits, TabPFN, ExtraTrees, CatBoost, XGBoost, and Random Forest land at threat scores between 0.620 and 0.637, a span smaller than each model's fold-to-fold standard deviation, so repeat-level comparisons find no durable ordering within the group; the same five models also trace the upper envelope of ROC curves, confirming the ranking is not threshold-dependent. The gap to the operational four-feature logistic baseline is large and stable (roughly +0.48 threat score on every fold), and the ten-feature logistic model sits halfway, showing that both feature breadth and nonlinear decision surfaces contribute. SHAP analysis of the four explained models finds that short-window peak rainfall intensity and storm accumulation dominate every ranking and are stable across folds, while burn-severity and terrain features rank lower and less stably. Synthetic augmentation from a class-conditional TabPFN generator is distributionally close to real data and raises threat score most for the operational baseline (+0.133) and the deep-learning models (up to +0.078), while moving the top nonlinear tier by at most +0.022, indicating that augmentation mainly relieves feature restriction and low model capacity.

Load-bearing premise

The performance claims assume that the 1,550 basin-storm rows are effectively independent across the random train-test splits, even though 61 percent of records come from a southern California cluster and neighboring basins within a fire share rainfall and burn patterns, so scores may be inflated by information leaking between similar rows.

Editorial extensions

If this is right

  • Operational warning programs can treat the five leading nonlinear models as interchangeable within sampling error and choose among them by tuning cost, inference speed, or interpretability.
  • A tuning-free tabular foundation model can match tuned tree ensembles on a small imbalanced hazard dataset, which makes it a candidate for rapid-response screening where per-region hyperparameter search is impractical.
  • Rainfall features, especially 15-minute peak intensity and storm accumulation, are the information that monitoring networks and warning thresholds should prioritize.
  • Synthetic augmentation is a tunable pipeline component rather than a universal fix: it most helps feature-restricted logistic models and weaker deep learners, so deployment should validate augmentation on real held-out folds.
  • Expanding beyond the four features of the operational logistic model and allowing nonlinearity accounts for most of the performance gap to the top tier.

Reading between the lines

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

  • Editorial inference: a strict fire-holdout evaluation would likely lower the reported 0.637 ceiling, since basins from the same fire are spatially autocorrelated; the tabular foundation model and tree top tier might still rank first, but the absolute scores for truly new fires would probably be lower.
  • Editorial inference: because TabPFN and ExtraTrees are statistically tied while TabPFN needs no per-dataset tuning, the practical advantage of the foundation model is operational speed and simplicity rather than predictive accuracy.
  • Editorial inference: the paper's augmentation result predicts where gains will appear elsewhere; models with restricted feature sets or limited capacity should benefit most from class-conditional synthetic rows, which is a testable hypothesis for other imbalanced hazard datasets.
  • Editorial inference: spatial holdouts could also change the augmentation conclusion, because synthetic rows that currently help models interpolate within a region might become more valuable when models must extrapolate to unseen fire regions.
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

1 major / 1 minor

Summary. The paper benchmarks 15 machine learning models, including the tabular foundation model TabPFN, for post-wildfire debris-flow prediction using the USGS basin-storm dataset of 1,550 observations. The evaluation protocol is repeated stratified 5-fold cross-validation with fold-fitted preprocessing, shared splits across all models, and per-fold hyperparameter tuning. The paper reports that TabPFN achieves the highest unaugmented threat score (0.637), closely followed by a top tier of tree-based models (ExtraTrees 0.636, CatBoost 0.622, XGBoost 0.621, Random Forest 0.620), and that this top tier is statistically tied rather than a stable ordering. SHAP analyses are used to claim that rainfall intensity and storm accumulation dominate feature importance, with burn severity and terrain features ranked lower. A final module uses TabPFN as a class-conditional synthetic data generator; augmentation improves most models, with the largest gains for Staley17 (+0.133) and the deep-learning models (LSTM +0.076, MLP +0.078), while TabPFN itself gains only +0.004.

Significance. If the findings are valid, the paper makes a meaningful contribution to post-wildfire debris-flow prediction: it demonstrates that a pre-trained tabular foundation model can match tuned tree ensembles on a small, imbalanced, partially overlapping tabular hazard dataset without dataset-specific hyperparameter search, and it provides a careful multi-model benchmark with shared splits, fold-fitted preprocessing, paired fold-level comparisons, and honest treatment of statistical ties within the top tier. The SHAP analysis is also methodical, including fold-stability checks, basin-level spatial maps, and agreement with physical expectations. The synthetic augmentation evaluation is thoughtfully designed with shared synthetic rows and per-model tuning, and the paper explicitly acknowledges the main limitation of its evaluation protocol. The principal weakness is that the central performance and augmentation claims rest on random stratified cross-validation of rows that are spatially and meteorologically clustered; the paper's own stated need for stricter spatial transfer tests is never met within the manuscript, which tempers the significance of the headline threat-score numbers.

major comments (1)
  1. [§3.3, Table 5] The synthetic augmentation conclusion that 'the benefit of augmentation is governed primarily by feature-set restriction and model capacity' (Conclusion) relies on the same random-split protocol. Under spatial leakage, the measured gains, especially Staley17's +0.133 threat score, could partly reflect the generator reproducing storm-scale patterns that already appear in the training folds, rather than a genuine improvement in predicting novel events. The paper's argument against generator-model bias (TabPFN gains only +0.004) is reasonable, but the augmentation recommendation itself would be more credible with a leave-one-fire-out version of the augmentation experiment or an explicit statement that the augmentation gains are only demonstrated under within-inventory random CV.
minor comments (1)
  1. [Figure 12] The synthetic sample counts shown in the legend (debris-flow synthetic n=972, no-debris-flow synthetic n=268) are consistent with a per-fold generator, but the caption and text should clarify that these counts refer to fold 0 training rows, not the full dataset, to avoid apparent inconsistency with the global class counts of 334 and 1,216.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the benchmark is measured on held-out real observations, and the TabPFN dual-role augmentation design is self-referential but explicitly tested and not load-bearing.

full rationale

The paper's central claims are empirical comparisons against held-out real data. Repeated stratified 5-fold CV uses precomputed splits shared by all models, with fold-fitted imputation and outer-fold hyperparameter tuning, so reported threat scores are measured on basins not seen during training or tuning. No equation in the paper defines a prediction in terms of the answer; the Staley17 reformulation R = PeakI15 × 0.25 h is a unit conversion that preserves the published equation and does not inject the outcome into the inputs. SHAP feature importances are computed on refit models and held-out rows, making the feature-ranking claims independent of the performance benchmark. The one structurally self-referential element is TabPFN acting both as an evaluated model and as the synthetic-data generator (Sections 2.10 and 2.13). This is a design risk, but the paper explicitly tests it: Section 4.3 reports that TabPFN gains only +0.004 threat score while several non-generator models gain substantially more, and the KDE and t-SNE fidelity checks show the synthetic rows are distributionally close to real rows. Because the augmentation conclusion is not forced by the generator choice—the generator model itself benefits least—this does not reduce to circularity. The acknowledged limitation that random stratified CV 'did not replace the need for stricter spatial transfer tests' (Section 5) is a generalizability and validity concern about the CV design, not a circularity; it does not make any reported quantity equal to an input by construction. The only self-citations (Xu et al., a and b) are contextual references to prior remote-sensing and tabular-synthesis work and are not load-bearing for any central claim. The paper is therefore self-contained against external held-out data, with no equation-level circularity; score 2 reflects the minor self-referential augmentation design and non-load-bearing self-citations, not actual circularity.

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

The central benchmarking claim depends on standard ML assumptions and on domain data assumptions about the USGS inventory; no invented physical entities are introduced. The main hand-chosen settings are the decision threshold, class weighting, and the augmentation rebalancing rule; these affect the magnitude of reported threat scores but are disclosed, and no scientific constants are fitted.

free parameters (4)
  • Decision threshold for threat score = Unstated, likely library default of 0.5
    Threat score is threshold-dependent, and threshold optimization was explicitly not performed (§2.4); the unspecified operating point affects all reported TS values.
  • Synthetic augmentation rebalancing rule = Real rows duplicated once; synthetic count per class equals real count of opposing class
    The augmentation outcome depends on this hand-chosen rule (§2.14); it drives the reported delta threat scores.
  • TabPFN generator sampling temperature = 1.0
    Figure 5d specifies temperature 1.0; it controls the dispersion of synthetic observations and therefore augmentation fidelity.
  • Class-weight rebalancing mode = Library default balanced mode
    Held fixed across models (§2.4) to handle imbalance; changing it shifts the decision surface and threat score.
assumptions (5)
  • domain assumption The USGS post-wildfire debris-flow dataset provides reliable labels and features for benchmarking.
    All model comparisons in §2.1 and §3 rest on the quality of the Staley et al. dataset; label noise and observation errors are not quantified.
  • domain assumption Basin-storm observations are exchangeable, so random stratified cross-validation estimates generalization.
    Section 2.2 uses 10x5 random CV on rows that are spatially clustered within fires and regions; this assumption is acknowledged as a limitation in §4.1 and §5.
  • standard math Missing peak rainfall intensities follow a power-law intensity-duration relationship used for imputation.
    Used in §2.3 to fill missing rainfall values; if this physical relationship does not hold for those rows, imputation changes feature values.
  • domain assumption TabPFN's pretrained prior and its unsupervised generator are appropriate for this hazard domain.
    TabPFN is both the top model and the synthetic data generator (§2.10, §2.13); the paper tests but cannot fully exclude coupling between generator and model.
  • standard math SHAP explainers faithfully approximate the models' Shapley values.
    Permutation SHAP for TabPFN uses a fixed evaluation budget and TreeSHAP is exact for trees (§2.12); approximation error is not quantified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Evaluating Machine Learning Models for Post-Wildfire Debris-Flow Prediction." pith.science (2026). https://pith.science/paper/SRI5ZOP5

@misc{pith2026260805265,
  author       = {Pith},
  title        = {Pith review of: Evaluating Machine Learning Models for Post-Wildfire Debris-Flow Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SRI5ZOP5}},
  note         = {Machine review of arXiv:2608.05265}
}
read the original abstract

Prediction of post-wildfire debris flows is critical for mitigating hazards to communities, infrastructure, and resources during intense rainfall in recently burned areas. However, identifying reliable machine learning models is complicated by overlapping debris-flow and non-debris-flow events in feature space, the need for model interpretability, and limited training data. This paper addresses these challenges through a systematic evaluation of machine learning models in terms of predictive performance, feature importance, and synthetic data augmentation. Using basin-scale observations of post-wildfire debris-flow events across the western United States, we compare 15 models, including the Tabular Prior-Data Fitted Network (TabPFN). Repeated stratified cross-validation shows that TabPFN achieves the highest unaugmented performance with a threat score of 0.637, closely followed by the best tree-based models. SHapley Additive exPlanations (SHAP) are used to identify the features driving predictions, revealing that short-duration rainfall intensity and storm accumulation consistently rank highest, while burn severity and terrain features contribute less. We further evaluate synthetic data augmentation using TabPFN-generated samples to address the scarcity of debris-flow observations. Synthetic augmentation improves the performance of all models except CNN, with the largest mean threat score increase of +0.041 among the deep learning models. By combining rigorous model benchmarking, interpretable feature analysis, and synthetic data augmentation, this work provides a comprehensive framework for improving post-wildfire debris-flow prediction.

Figures

Figures reproduced from arXiv: 2608.05265 by the authors.

Figure 1
Figure 1. Study-area overview of the 34 fire events in the USGS post-wildfire debris-flow dataset (2000–2013, western United States). Panel (a) shows the regional extent with labeled fire centroids on a digital elevation model (DEM), panel (b) zooms to the southern California cluster, and panel (c) shows a United States locator. Panels (d)–(h) show delineated basins for the Station, Grand Prix–Old, Schultz, Little Bear, and W… view at source ↗
Figure 2
Figure 2. Conceptual overview of the experimental pipeline. The left column lists the sequential steps applied within each outer training fold: data sources, features and response, cross-validation splitting, preprocessing, and hyperparameter tuning. The right column shows the three analysis modules: the 15-model benchmark, SHAP feature importance evaluation, and TabPFN-based synthetic data augmentation evaluation. features a… view at source ↗
Figure 3
Figure 3. Upper-triangle Spearman rank correlation matrix for the ten benchmark features and the binary debris-flow response. Each cell gives a rank correlation value. Dashed lines separate the Response row and column from the feature– feature entries. et al., c). Watershed and burn-severity features are ContribArea, PropHM23, and dNBR (Key and Benson). The soil feature KF is the USLE soil erodibility factor (𝐾) for the fine … view at source ↗
Figures from the paper (13 more)
Figure 4
Figure 4. Figure 4: Grand Prix–Old fire: spatial distribution of nine features across delineated basins for a representative storm. The top row shows burn- and soil-related features, and the lower rows show rainfall accumulation, duration, mean intensity, and peak 15-, 30-, and 60-minute …
Figure 5
Figure 5. Figure 5: Overview of TabPFN architecture, training, prediction, and synthetic data generation. Panel (a) shows the TabPFN layer block, repeated twelve times, combining feature attention, sample attention, and a multi-layer perceptron. Panel (b) depicts TabPFN pre-training, wher…
Figure 6
Figure 6. Figure 6: Mean ROC curves under unaugmented training, with one curve per model averaged over the same 50 stratified outer folds (10 repeats of 5 folds) and a dashed diagonal at random chance. The legend lists models in descending unaugmented threat score order, and the bottom-ri…
Figure 7
Figure 7. Figure 7: Per-model ROC traces under unaugmented training, arranged in a three-row by five-column grid sorted by descending mean AUC, with Staley17 in the bottom-right panel. Within each panel, thin green lines are the 50 individual outer-fold ROC curves, the navy line is the fo…
Figure 8
Figure 8. Figure 8: Two-panel t-SNE (Maaten and Hinton) projection (perplexity 30) of the full dataset, with class-wise marginal histograms along each axis. Panel (a) uses preprocessed raw features and panel (b) uses out-of-fold TabPFN internal representations. The fold-stability heatmap …
Figure 9
Figure 9. Figure 9: Combined SHAP summary for TabPFN, pooling held-out basin explanations across all 50 stratified outer folds. The left panel shows mean |SHAP| per feature as a horizontal bar chart, and the right panel shows the corresponding beeswarm in which each dot is one basin-level…
Figure 10
Figure 10. Figure 10: Fold-wise stability of TabPFN SHAP feature importance as a heatmap, with rows for the ten features and columns for the 50 stratified outer folds. Each cell gives the mean |SHAP| on the fold’s held-out basins [PITH_FULL_IMAGE:figures/full_fig_p016_10.png]
Figure 11
Figure 11. Figure 11: Agreement between SHAP-derived feature directions and physical expectations. Orange = positive direction (toward debris flow); teal = negative; gray = weak (|𝑟| < 0.15). Y/N indicates agreement or contradiction with physics; +/− is shown directly for features with amb…
Figure 12
Figure 12. Figure 12: Two-panel t-SNE (Maaten and Hinton) comparison of TabPFN-generated synthetic training rows against real rows for a single representative outer fold (fold 0), with class-wise marginal histograms along each axis. All four groups are projected under one global t-SNE fit …
Figure 13
Figure 13. Figure 13: Basin-level SHAP feature importance for the Station fire in a three-row, four-column layout for PeakI15, PropHM23, dNBR, and KF. The top row shows feature-value maps and the middle row shows SHAP-value maps under a shared diverging colormap. The bottom row shows depen…
Figure 14
Figure 14. Figure 14: Basin-level SHAP feature importance for the Grand Prix–Old fire under the same three-row, four-column layout as [PITH_FULL_IMAGE:figures/full_fig_p021_14.png]
Figure 15
Figure 15. Figure 15: Class-conditional kernel density estimates of the benchmark features in a nine-panel grid, with panels (a)–(i) shown separately for each feature and a shared bottom legend. Solid curves are imputed real training observations and dotted curves are fold 0 synthetic obse…
Figure 16
Figure 16. Figure 16: Per-model threat score for the unaugmented and augmented evaluations as overlaid box plots (left axis) and the mean change in threat score from unaugmented to augmented as bars (right axis). Models are ordered along the horizontal axis by descending unaugmented threat…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

71 extracted references · 36 canonical work pages

  1. [1]

    and Negri, Jacquelyn A

    Staley, Dennis M. and Negri, Jacquelyn A. and Kean, Jason W. and Laber, Jayme L. and Tillery, Anne C. and Youberg, Ann M. , urldate =. Prediction of spatially explicit rainfall intensity–duration thresholds for post-fire debris-flow generation in the western United States , volume =. doi:10.1016/j.geomorph.2016.10.019 , pages =

  2. [2]

    and Addison, Priscilla and Oommen, Thomas and Salazar, Sean E

    Kern, Ashley N. and Addison, Priscilla and Oommen, Thomas and Salazar, Sean E. and Coffman, Richard A. , urldate =. Machine Learning Based Predictive Modeling of Debris Flow Probability Following Wildfire in the Intermountain Western United States , volume =. doi:10.1007/s11004-017-9681-2 , abstract =

  3. [3]

    and Negri, Jacquelyn A

    Staley, Dennis M. and Negri, Jacquelyn A. and Kean, Jason W. and Laber, Jayme L. and Tillery, Anne C. and Youberg, Ann M. , urldate =. Updated logistic regression equations for the calculation of post-fire debris-flow likelihood in the western United States , issn =. doi:10.3133/ofr20161106 , abstract =

  4. [4]

    and Destro, Elisa and Bhuiyan, Md Abul Ehsan and Borga, Marco and Anagnostou, Emmanouil N

    Nikolopoulos, Efthymios I. and Destro, Elisa and Bhuiyan, Md Abul Ehsan and Borga, Marco and Anagnostou, Emmanouil N. , urldate =. Evaluation of predictive models for post-fire debris flow occurrence in the western United States , volume =. doi:10.5194/nhess-18-2331-2018 , abstract =

  5. [5]

    Machine Learning for Improved Post-fire Debris Flow Likelihood Prediction , url =

    Roten, Daniel and Block, Jessica and Crawl, Daniel and Lee, Jenny and Altintas, Ilkay , urldate =. Machine Learning for Improved Post-fire Debris Flow Likelihood Prediction , url =. 2022. doi:10.1109/BigData55660.2022.10020574 , abstract =

  6. [6]

    and Gartner, Joseph E

    Cannon, Susan H. and Gartner, Joseph E. and Rupert, Michael G. and Michael, John A. and Rea, Alan H. and Parrett, Charles , urldate =. Predicting the probability and volume of postwildfire debris flows in the intermountain western United States , volume =. doi:10.1130/B26459.1 , abstract =

  7. [7]

    doi:10.48550/arXiv.2511.08667 , shorttitle =

    Grinsztajn, Léo and Flöge, Klemens and Key, Oscar and Birkel, Felix and Jund, Philipp and Roof, Brendan and Jäger, Benjamin and Safaric, Dominik and Alessi, Simone and Hayler, Adrian and Manium, Mihir and Yu, Rosen and Jablonski, Felix and Hoo, Shi Bin and Garg, Anurag and Robertson, Jake and Bühler, Magnus and Moroshan, Vladyslav and Purucker, Lennart an...

  8. [8]

    Accurate predictions on small data with a tabular foundation model , volume =

    Hollmann, Noah and Müller, Samuel and Purucker, Lennart and Krishnakumar, Arjun and Körfer, Max and Hoo, Shi Bin and Schirrmeister, Robin Tibor and Hutter, Frank , urldate =. Accurate predictions on small data with a tabular foundation model , volume =. doi:10.1038/s41586-024-08328-6 , abstract =

Show all 71 references
  1. [9]

    doi:10.48550/arXiv.2207.01848 , shorttitle =

    Hollmann, Noah and Müller, Samuel and Eggensperger, Katharina and Hutter, Frank , urldate =. doi:10.48550/arXiv.2207.01848 , shorttitle =. 2207.01848 [cs] , keywords =

  2. [10]

    doi:10.48550/arXiv.2508.10053 , shorttitle =

    Beaglehole, Daniel and Holzmüller, David and Radhakrishnan, Adityanarayanan and Belkin, Mikhail , urldate =. doi:10.48550/arXiv.2508.10053 , shorttitle =. 2508.10053 [cs] , keywords =

  3. [11]

    Landslide Risk Assessment as a Reference for Disaster Prevention and Mitigation: A Case Study of the Renhe District, Panzhihua City, China , volume =

    Zhou, Yimeng and Xue, Lei and Ding, Hao and Wang, Haoyu and Huang, Kun and Li, Longfei and Li, Zhuan , urldate =. Landslide Risk Assessment as a Reference for Disaster Prevention and Mitigation: A Case Study of the Renhe District, Panzhihua City, China , volume =. doi:10.3390/...

  4. [12]

    Assessment of post-wildfire debris flow occurrence using classifier tree , volume =

    Addison, Priscilla and Oommen, Thomas and Sha, Qiuying , urldate =. Assessment of post-wildfire debris flow occurrence using classifier tree , volume =. doi:10.1080/19475705.2018.1530306 , abstract =

  5. [13]

    Interpretable Machine Learning for

    Rundel, David and Kobialka, Julius and Crailsheim, Constantin von and Feurer, Matthias and Nagler, Thomas and Rügamer, David , urldate =. Interpretable Machine Learning for. doi:10.1007/978-3-031-63797-1_23 , abstract =. 2403.10923 [cs] , keywords =

  6. [14]

    Transformers Can Do Bayesian Inference , url =

    Müller, Samuel and Hollmann, Noah and Arango, Sebastian Pineda and Grabocka, Josif and Hutter, Frank , urldate =. Transformers Can Do Bayesian Inference , url =. doi:10.48550/arXiv.2112.10510 , abstract =. 2112.10510 [cs] , keywords =

  7. [15]

    and Kean, Jason W

    Staley, Dennis M. and Kean, Jason W. and Cannon, Susan H. and Schmidt, Kevin M. and Laber, Jayme L. , urldate =. Objective definition of rainfall intensity–duration thresholds for the initiation of post-fire debris flows in southern California , volume =. doi:10.1007/s10346-01...

  8. [16]

    and Benson, Nathan C

    Key, Carl H. and Benson, Nathan C. , urldate =. Landscape Assessment (

  9. [17]

    A Unified Approach to Interpreting Model Predictions , url =

    Lundberg, Scott and Lee, Su-In , urldate =. A Unified Approach to Interpreting Model Predictions , url =. doi:10.48550/arXiv.1705.07874 , abstract =. 1705.07874 [cs] , keywords =

  10. [18]

    A method for better mapping of susceptibility to thaw hazards in data-scarce cold regions , volume =

    Zhu, Hualiang and Zhang, Xianwei and Wei, Gang and Wang, Qingzhi and Liu, Xinyu and Yan, Lei and Wang, Gang , urldate =. A method for better mapping of susceptibility to thaw hazards in data-scarce cold regions , volume =. doi:10.1016/j.rse.2026.115338 , abstract =

  11. [19]

    Reliability and effectiveness of early warning systems for natural hazards: Concept and application to debris flow warning , volume =

    Sättele, Martina and Bründl, Michael and Straub, Daniel , urldate =. Reliability and effectiveness of early warning systems for natural hazards: Concept and application to debris flow warning , volume =. doi:10.1016/j.ress.2015.05.003 , shorttitle =

  12. [20]

    Knowledge-Data Dually Driven Paradigm for Accurate Landslide Susceptibility Prediction under Data-Scarce Conditions Using Geomorphic Priors and Tabular Foundation Model , url =

    Yang, Yuting and Mei, Gang and Chen, Feng and Zhang, Yongshuang and Peng, Jianbing , urldate =. Knowledge-Data Dually Driven Paradigm for Accurate Landslide Susceptibility Prediction under Data-Scarce Conditions Using Geomorphic Priors and Tabular Foundation Model , url =

  13. [21]

    Susceptibility Prediction of Post-Fire Debris Flows in Xichang, China, Using a Logistic Regression Model from a Spatiotemporal Perspective , volume =

    Jin, Tao and Hu, Xiewen and Liu, Bo and Xi, Chuanjie and He, Kun and Cao, Xichao and Luo, Gang and Han, Mei and Ma, Guotao and Yang, Ying and Wang, Yan , urldate =. Susceptibility Prediction of Post-Fire Debris Flows in Xichang, China, Using a Logistic Regression Model from a ...

  14. [22]

    Exploring the Application of a Debris Flow Likelihood Regression Model in Mediterranean Post-Fire Environments, Using Field Observations-Based Validation , volume =

    Diakakis, Michalis and Mavroulis, Spyridon and Vassilakis, Emmanuel and Chalvatzi, Vassiliki , urldate =. Exploring the Application of a Debris Flow Likelihood Regression Model in Mediterranean Post-Fire Environments, Using Field Observations-Based Validation , volume =. doi:1...

  15. [23]

    Addressing class imbalance in soil movement predictions , volume =

    Kumar, Praveen and Priyanka, Priyanka and Uday, Kala Venkata and Dutt, Varun , urldate =. Addressing class imbalance in soil movement predictions , volume =. doi:10.5194/nhess-24-1913-2024 , abstract =

  16. [24]

    Optimizing the Predictive Ability of Machine Learning Methods for Landslide Susceptibility Mapping Using

    Wang, Yumiao and Wu, Xueling and Chen, Zhangjian and Ren, Fu and Feng, Luwei and Du, Qingyun , urldate =. Optimizing the Predictive Ability of Machine Learning Methods for Landslide Susceptibility Mapping Using. doi:10.3390/ijerph16030368 , abstract =

  17. [25]

    The meaning and use of the area under a receiver operating characteristic (

    Hanley, J A and. The meaning and use of the area under a receiver operating characteristic (. doi:10.1148/radiology.143.1.7063747 , abstract =

  18. [26]

    Proceedings of the 22nd

    Chen, Tianqi and Guestrin, Carlos , urldate =. Proceedings of the 22nd. doi:10.1145/2939672.2939785 , series =

  19. [27]

    A Scalable Framework for Post Fire Debris Flow Hazard Assessment Using Satellite Precipitation Data , volume =

    Orland, Elijah and Kirschbaum, Dalia and Stanley, Thomas , urldate =. A Scalable Framework for Post Fire Debris Flow Hazard Assessment Using Satellite Precipitation Data , volume =. doi:10.1029/2022GL099850 , abstract =

  20. [28]

    Understanding variable importances in forests of randomized trees , volume =

    Louppe, Gilles and Wehenkel, Louis and Sutera, Antonio and Geurts, Pierre , urldate =. Understanding variable importances in forests of randomized trees , volume =. Advances in Neural Information Processing Systems , publisher =

  21. [29]

    Random Forests , volume =

    Breiman, Leo , urldate =. Random Forests , volume =. doi:10.1023/A:1010933404324 , abstract =

  22. [30]

    A comparative study of different classification techniques for marine oil spill identification using

    Xu, Linlin and Li, Jonathan and Brenning, Alexander , urldate =. A comparative study of different classification techniques for marine oil spill identification using. doi:10.1016/j.rse.2013.10.012 , abstract =

  23. [31]

    Chawla, N. V. and Bowyer, K. W. and Hall, L. O. and Kegelmeyer, W. P. , urldate =. doi:10.1613/jair.953 , shorttitle =

  24. [32]

    , urldate =

    He, Haibo and Garcia, Edwardo A. , urldate =. Learning from Imbalanced Data , volume =. doi:10.1109/TKDE.2008.239 , abstract =

  25. [33]

    , urldate =

    Branco, Paula and Torgo, Luís and Ribeiro, Rita P. , urldate =. A Survey of Predictive Modeling on Imbalanced Domains , volume =. doi:10.1145/2907070 , abstract =

  26. [34]

    and Bottou, L

    Lecun, Y. and Bottou, L. and Bengio, Y. and Haffner, P. , urldate =. Gradient-based learning applied to document recognition , volume =. doi:10.1109/5.726791 , abstract =

  27. [35]

    Optuna: A Next-generation Hyperparameter Optimization Framework , isbn =

    Akiba, Takuya and Sano, Shotaro and Yanase, Toshihiko and Ohta, Takeru and Koyama, Masanori , urldate =. Optuna: A Next-generation Hyperparameter Optimization Framework , isbn =. Proceedings of the 25th. doi:10.1145/3292500.3330701 , series =

  28. [36]

    Hyperband: A Novel Bandit-Based Approach to Hyperparameter Optimization , volume =

    Li, Lisha and Jamieson, Kevin and. Hyperband: A Novel Bandit-Based Approach to Hyperparameter Optimization , volume =

  29. [37]

    Visualizing Data using t-

    Maaten, Laurens van der and Hinton, Geoffrey , urldate =. Visualizing Data using t-

  30. [38]

    Cox, D. R. , urldate =. The Regression Analysis of Binary Sequences , volume =. doi:10.1111/j.2517-6161.1958.tb00292.x , abstract =

  31. [39]

    The Wasserstein distances , isbn =

    Villani, Cédric , editor =. The Wasserstein distances , isbn =. Optimal Transport: Old and New , publisher =. doi:10.1007/978-3-540-71050-9_6 , abstract =

  32. [40]

    and Hart, P

    Cover, T. and Hart, P. , urldate =. Nearest neighbor pattern classification , volume =. doi:10.1109/TIT.1967.1053964 , abstract =

  33. [41]

    Support-vector networks , volume =

    Cortes, Corinna and Vapnik, Vladimir , urldate =. Support-vector networks , volume =. doi:10.1007/BF00994018 , abstract =

  34. [42]

    Extremely randomized trees , volume =

    Geurts, Pierre and Ernst, Damien and Wehenkel, Louis , urldate =. Extremely randomized trees , volume =. doi:10.1007/s10994-006-6226-1 , abstract =

  35. [43]

    and Hinton, Geoffrey E

    Rumelhart, David E. and Hinton, Geoffrey E. and Williams, Ronald J. , urldate =. Learning representations by back-propagating errors , volume =. doi:10.1038/323533a0 , abstract =

  36. [44]

    and Rizzo, Maria L

    Székely, Gábor J. and Rizzo, Maria L. , urldate =. Energy statistics: A class of statistics based on distances , volume =. doi:10.1016/j.jspi.2013.03.018 , shorttitle =

  37. [45]

    Why do tree-based models still outperform deep learning on tabular data? , url =

    Grinsztajn, Léo and Oyallon, Edouard and Varoquaux, Gaël , urldate =. Why do tree-based models still outperform deep learning on tabular data? , url =. doi:10.48550/arXiv.2207.08815 , abstract =. 2207.08815 [cs.LG] , keywords =

  38. [46]

    Deep Neural Networks and Tabular Data: A Survey , volume =

    Borisov, Vadim and Leemann, Tobias and Seßler, Kathrin and Haug, Johannes and Pawelczyk, Martin and Kasneci, Gjergji , urldate =. Deep Neural Networks and Tabular Data: A Survey , volume =. doi:10.1109/TNNLS.2022.3229161 , shorttitle =

  39. [47]

    Modeling Tabular data using Conditional

    Xu, Lei and Skoularidou, Maria and Cuesta-Infante, Alfredo and Veeramachaneni, Kalyan , urldate =. Modeling Tabular data using Conditional. doi:10.48550/arXiv.1907.00503 , abstract =. 1907.00503 [cs.LG] , keywords =

  40. [48]

    Machine-Learning-Based Prediction Modeling for Debris Flow Occurrence: A Meta-Analysis , volume =

    Yang, Lianbing and Ge, Yonggang and Chen, Baili and Wu, Yuhong and Fu, Runde , urldate =. Machine-Learning-Based Prediction Modeling for Debris Flow Occurrence: A Meta-Analysis , volume =. doi:10.3390/w16070923 , shorttitle =

  41. [49]

    doi:10.2113/gseegeosci.21.4.277 , abstract =

    The Timing of Susceptibility to Post-Fire Debris Flows in the Western United States , volume =. doi:10.2113/gseegeosci.21.4.277 , abstract =

  42. [50]

    Language Models are Few-Shot Learners , volume =

    Brown, Tom and Mann, Benjamin and Ryder, Nick and Subbiah, Melanie and Kaplan, Jared D and Dhariwal, Prafulla and Neelakantan, Arvind and Shyam, Pranav and Sastry, Girish and Askell, Amanda and Agarwal, Sandhini and Herbert-Voss, Ariel and Krueger, Gretchen and Henighan, Tom a...

  43. [51]

    The Elements of Statistical Learning , rights =

    Hastie, Trevor and Tibshirani, Robert and Friedman, Jerome , urldate =. The Elements of Statistical Learning , rights =. doi:10.1007/978-0-387-84858-7 , series =

  44. [52]

    Mamba: Linear-Time Sequence Modeling with Selective State Spaces , url =

    Gu, Albert and Dao, Tri , urldate =. Mamba: Linear-Time Sequence Modeling with Selective State Spaces , url =. doi:10.48550/arXiv.2312.00752 , shorttitle =. 2312.00752 [cs.LG] , keywords =

  45. [53]

    A survey of cross-validation procedures for model selection , volume =

    Arlot, Sylvain and Celisse, Alain , urldate =. A survey of cross-validation procedures for model selection , volume =. doi:10.1214/09-SS054 , issue =

  46. [54]

    A study of cross-validation and bootstrap for accuracy estimation and model selection , isbn =

    Kohavi, Ron , urldate =. A study of cross-validation and bootstrap for accuracy estimation and model selection , isbn =. Proceedings of the 14th international joint conference on Artificial intelligence - Volume 2 , publisher =

  47. [55]

    and Alexander, R

    Schwarz, Gregory E. and Alexander, R. B. , urldate =. State Soil Geographic (. doi:10.3133/ofr95449 , abstract =

  48. [56]

    and Talbot, Nicola L.C

    Cawley, Gavin C. and Talbot, Nicola L.C. , urldate =. On Over-fitting in Model Selection and Subsequent Selection Bias in Performance Evaluation , volume =

  49. [57]

    A mathematical framework for studying rainfall intensity-duration-frequency relationships , volume =

    Koutsoyiannis, Demetris and Kozonis, Demosthenes and Manetas, Alexandros , urldate =. A mathematical framework for studying rainfall intensity-duration-frequency relationships , volume =. doi:10.1016/S0022-1694(98)00097-3 , abstract =

  50. [58]

    Grinsztajn, Leo and Purucker, Lennart and Krishnakumar, Arjun and Körfer, Max and Hoo, Shi Bin and Schirrmeister, Robin Tibor and Bergman, Eddie and Flöge, Klemens and Key, Oscar and Safaric, Dominik and Hollmann, Noah and Müller, Samuel and Hutter, Frank , urldate =

  51. [59]

    , urldate =

    Schaefer, Joseph T. , urldate =. The Critical Success Index as an Indicator of Warning Skill , volume =. doi:10.1175/1520-0434(1990)005<0570:TCSIAA>2.0.CO;2 , abstract =

  52. [60]

    pfdf - Python library for postfire debris-flow hazard assessments and research, version 3.0.2 , url =

    Jonathan M King , urldate =. pfdf - Python library for postfire debris-flow hazard assessments and research, version 3.0.2 , url =. doi:10.5066/P1JJXSXD , abstract =

  53. [61]

    Why Should I Trust You?

    Ribeiro, Marco Tulio and Singh, Sameer and Guestrin, Carlos , urldate =. "Why Should I Trust You?": Explaining the Predictions of Any Classifier , url =. doi:10.48550/arXiv.1602.04938 , shorttitle =. 1602.04938 [cs.LG] , keywords =

  54. [62]

    and Erion, Gabriel and Chen, Hugh and

    Lundberg, Scott M. and Erion, Gabriel and Chen, Hugh and. Explainable. doi:10.48550/arXiv.1905.04610 , shorttitle =. 1905.04610 [cs.LG] , keywords =

  55. [63]

    Statistical Comparisons of Classifiers over Multiple Data Sets , volume =

    Demšar, Janez , urldate =. Statistical Comparisons of Classifiers over Multiple Data Sets , volume =

  56. [64]

    Long short-term memory

    Hochreiter, Sepp and Schmidhuber, Jurgen , urldate =. Long short-term memory. , volume =. doi:10.1162/neco.1997.9.8.1735 , abstract =

  57. [65]

    , urldate =

    Massey, Frank J. , urldate =. The Kolmogorov-Smirnov Test for Goodness of Fit , volume =. doi:10.2307/2280095 , abstract =

  58. [66]

    Cross-validation pitfalls when selecting and assessing regression and classification models , volume =

    Krstajic, Damjan and Buturovic, Ljubomir J and Leahy, David E and Thomas, Simon , urldate =. Cross-validation pitfalls when selecting and assessing regression and classification models , volume =. doi:10.1186/1758-2946-6-10 , abstract =

  59. [67]

    Attention is All you Need , volume =

    Vaswani, Ashish and Shazeer, Noam and Parmar, Niki and Uszkoreit, Jakob and Jones, Llion and Gomez, Aidan N and Kaiser,. Attention is All you Need , volume =. Advances in Neural Information Processing Systems , publisher =

  60. [68]

    Advances in Neural Information Processing Systems , publisher =

    Prokhorenkova, Liudmila and Gusev, Gleb and Vorobev, Aleksandr and Dorogush, Anna Veronika and Gulin, Andrey , urldate =. Advances in Neural Information Processing Systems , publisher =

  61. [69]

    , urldate =

    Dietterich, Thomas G. , urldate =. Approximate Statistical Tests for Comparing Supervised Classification Learning Algorithms. , volume =. doi:10.1162/089976698300017197 , abstract =

  62. [70]

    Time for a Change: a Tutorial for Comparing Multiple Classifiers Through Bayesian Analysis , volume =

    Benavoli, Alessio and Corani, Giorgio and Demšar, Janez and Zaffalon, Marco , urldate =. Time for a Change: a Tutorial for Comparing Multiple Classifiers Through Bayesian Analysis , volume =

  63. [71]

    Inference for the Generalization Error , volume =

    Nadeau, Claude and Bengio, Yoshua , urldate =. Inference for the Generalization Error , volume =. doi:10.1023/A:1024068626366 , abstract =

Pith tools

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