Pith. sign in

REVIEW 4 major objections 6 minor 53 references

A Morphological Model to Separate Resolved-Unresolved Sources in the DESI Legacy Surveys: Application in the LS4 Alert Stream

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A machine-learning model trained on a single deep field separates stars from galaxies across the entire DESI Legacy Surveys, assigning scores to about 3.1 billion sources and feeding the LS4 transient alert stream.

desk verdict Solid applied ML contribution with a useful public catalog; the headline accuracy claim is only validated on bright, well-observed sources, so the 3-billion-source promise needs a caveat. read the letter →

arxiv 2505.17174 v2 pith:RTEFDNPB submitted 2025-05-22 astro-ph.IM

Chang Liu , Adam A. Miller , Joshua S. Bloom , Robert A. Knop , Peter E. Nugent This is my paper · ORCID
classification astro-ph.IM
keywords star-galaxyclassificationmorphologyXGBoostDESILegacySurveysresolvedandunresolvedsourcesLS4machine-learningcatalog
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 claims that a supervised machine-learning model can separate resolved (galaxy-like) from unresolved (star-like) sources in the DESI Legacy Surveys more accurately than the survey's own morphological typing. The central contribution is a "Hybrid" model that combines two XGBoost classifiers, one using blue-filter (g and r) aperture flux ratios and one using red-filter (i and z) ratios, so that predictions remain reliable when sources are missing one or more filters. The authors state this is the largest machine-learning catalog of resolved versus unresolved sources ever produced, covering roughly 3.1 billion LS DR10 objects to about 26th magnitude. A reader should care because the catalog is already embedded in the La Silla Schmidt Southern Survey (LS4) alert pipeline, where it helps reject foreground Galactic stars so that extragalactic transients can be found quickly.

What carries the argument

The load-bearing mechanism is the Hybrid model: a weighted linear combination of two independently trained XGBoost classifiers, score = (α_hyb·det_gr·score_gr + (1−α_hyb)·det_iz·score_iz)/(α_hyb·det_gr + (1−α_hyb)·det_iz), where det flags say whether the blue (gr) or red (iz) filter pair has data and α_hyb = 0.33. Each sub-model uses the same five morphological features—the normalized improvement in fit when replacing a PSF profile with REX, DEV, EXP, or Sersic profiles (delta dchisq) plus a mask flag for forced-point-source or not-fit cases—and seven aperture flux ratios between consecutive apertures from 0.5 to 7.0 arcseconds, computed on the S/N-weighted mean flux of the gr or iz pair. The weighted combination lets the model use any single available filter while avoiding the collapse in performance that the grz model suffers when a filter is missing.

What would settle it

Take a spectroscopic sample such as DESI or SDSS restricted to fields with 5-sigma detection limits in r fainter than 25 and seeing better than 1.2 arcseconds, restrict to white_mag between 22 and 25 where stars are rare, and compare the Hybrid model's scores against spectroscopic types; if the true positive rate for stars in that deep, good-seeing sample falls substantially below the 0.707 CV figure of merit with the paper's thresholds, the COSMOS transferability assumption fails.

Watch

Extended reading notes

Core claim

The paper establishes that the Hybrid model separates point sources from extended sources in DESI Legacy Surveys DR10 with higher accuracy and better robustness to missing filters than the LS morphological typing. On the HST COSMOS training set the Hybrid model reaches a figure of merit (true positive rate at a fixed 0.5% false positive rate) of 0.707 in cross-validation, 0.9778 on the DESI DR1 spectroscopic test set, and 0.9704 on the SDSS DR17 test set, and it outperforms both the "White" model and the "grz" model on the wide-field test sets. The model outputs a star score for roughly 3.1 billion LS sources, and the resulting catalog is used in the real-time LS4 pipeline to flag galactic versus extragalactic origins of transients.

Load-bearing premise

The HST COSMOS training labels, and the feature distributions in that single deep field, are representative of the whole LS footprint, including shallower, worse-seeing, and more crowded regions where the model is expected to degrade.

Editorial extensions

If this is right

  • LS4 transient alerts can be cross-matched to point-source scores for counterparts about two magnitudes deeper than the survey's typical detection limit, letting the pipeline reject stars before follow-up.
  • The catalog supplies classification scores for essentially the whole LS DR10 footprint, about 3.1 billion sources, which is the largest resolved–unresolved classification catalog ever produced.
  • Because the Hybrid model keeps nearly all galaxies (about 99.5% at the recommended 0.5 threshold) while rejecting most stars, it should substantially reduce Galactic false positives in searches for supernovae, kilonovae, and tidal disruption events.
  • The recommended threshold of 0.5 coincides with a 0.5% false positive rate on the HST training set, and per-magnitude thresholds are provided so users can tune purity against completeness.
  • Model scores degrade gracefully with worse seeing, shallower fields, and higher stellar density, and the false positive rate stays nearly constant with crowding, so a single threshold can be used across most of the sky.

Reading between the lines

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

  • The same blue/red split architecture could be transferred to other multi-band surveys with incomplete filter coverage, such as LSST, provided the aperture ratios are recomputed for that survey's PSF and aperture set.
  • Users working at low Galactic latitudes should expect lower true positive rates for stars than the COSMOS-based numbers suggest, and may want to apply density-dependent classification thresholds rather than the single global threshold.
  • The practice of adding 1 to the ML score for reliable Gaia stars, recovering about 5 million misclassified bright stars, points toward a general strategy of injecting high-purity astrometric prior information into photometric classifiers.
  • If the COSMOS training labels under-represent stars in shallow or crowded fields, the published catalog's star fraction of about 15% may be an underestimate at faint magnitudes and low Galactic latitudes.
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

4 major / 6 minor

Summary. The paper presents a supervised XGBoost classifier for separating resolved and unresolved sources in DESI Legacy Surveys DR10. It is trained on roughly 240,000 LS sources in the COSMOS field with HST morphological labels, using features derived from Tractor dchisq values, a new mask feature, and aperture-flux ratios. To handle incomplete filter coverage, the authors build a Hybrid model that linearly combines two XGBoost sub-models using blue (gr) and red (iz) aperture ratios, and they tune the relative weight alpha_hyb. Performance is reported via COSMOS cross-validation, on spectroscopically classified DESI DR1 (~15.4 million objects) and SDSS DR17 (~3.4 million objects) test sets, and on ~176 million Gaia-selected stars. The final LS-PSC catalog provides Hybrid model scores for ~3.15 billion LS DR10 sources and has been integrated into the LS4 alert stream.

Significance. If the claimed accuracy holds, LS-PSC is a major community resource: it is, to my knowledge, the largest machine-learning catalog of resolved versus unresolved classifications, it is accompanied by released code and catalog access, and it addresses the practically important problem of missing filters in the LS footprint. The independent spectroscopic test sets are unusually large, and the authors are candid about degradation with depth, seeing, and crowding, which is a genuine strength of the presentation. These strengths do not, however, remove the need to validate the model in the non-COSMOS regimes where the catalog will actually be applied; the referee's concern about unvalidated weak regimes is real and is the main reason the manuscript needs revision rather than acceptance.

major comments (4)
  1. [§5.2.2, Figures 7–8] Figures 7 and 8 show score distributions for DESI stars and galaxies as a function of detection limit and seeing, and the text interprets them as showing that the model "downgrades" in shallower and poorer-seeing fields. Because these panels report score quantiles rather than classification error against ground truth, they cannot bound the TPR/FPR that the catalog achieves in those regimes; the conclusion that "there is no evidence for significant systematics" does not follow from the displayed quantities. The same applies to Figure 9: the claim that FPR remains largely unaffected by crowding is not measured. Since the final catalog is applied to the full DR10 footprint, the absence of error-rate validation in shallow, poor-seeing, and crowded bricks is a load-bearing gap.
  2. [§2.1, Eq. (1), Table 3] The HST cross-validation metrics are not independent of the label-revision procedure: the polynomial boundary P(MAG_AUTO) is fit to the same HST COSMOS data that define the training labels, and the FoM/accuracy in Table 3 are evaluated on those relabeled targets. The spectroscopic checks shown in the right panel of Figure 1 are sparse at the faint end where the polynomial boundary matters most, so the COSMOS CV numbers partly certify consistency with the hand-tuned boundary rather than external truth. This does not invalidate the DESI/SDSS/Gaia evaluations, but it means Table 3 and Table 6 should be labeled as internal validation, not external validation, in the text.
  3. [§2.2.1–2.2.2, Tables 4–5] The DESI and SDSS test sets are heavily filtered (QSOs removed, deltachi2 < 50 removed, efftime_spec < 100 removed, z < 0.001 and z > 1.6 cuts, artifact rejection, and SGA-based label flips), and by construction they are dominated by relatively bright targets; the faint, shallow, poor-seeing, and crowded regimes highlighted in Figures 3, 7, 8, and 9 are far less represented. As a result, the high FoM/accuracy values in Tables 4 and 5 cannot by themselves support the paper's generalization claim across the whole DR10 footprint; an explicit demonstration that the test sets probe the weak regimes, or additional validation in those regimes, is needed.
  4. [§6.2, Table 6] The catalog adopts a fixed threshold of 0.5, and the paper states that this retains about 99.5% of galaxies, but the threshold and TPR values are calibrated exclusively on COSMOS CV (Table 6), and the paper itself warns that TPR is lower in crowded fields. No measurement of FPR is provided for the shallow or poor-seeing fields that occupy a non-negligible part of the DR10 footprint (§5.2.2), so the "almost all galaxies" statement is not established for the deployed catalog. The authors should either demonstrate that the threshold maintains a stable FPR across the footprint or publish spatially or threshold-dependent performance maps.
minor comments (6)
  1. [Abstract and §7] The training-set size is quoted as ~2×10^5 in the Abstract and ~2×10^6 in Section 7; the actual number after cuts is 240,671, so these quotes should be made consistent.
  2. [§4] The phrase "With conduct a 5-fold cross validation" should read "We conduct a 5-fold cross-validation."
  3. [Eq. (10)] The subscript det_iz is rendered as "detiz" in Equation (10); please fix the typesetting so it matches the definitions of det_gr and det_iz in the surrounding text.
  4. [Figure 6 caption] Percentages in the legend such as "Gaia stars (98.8%)" are TPR values for subsets, but the caption does not state this; clarify to avoid reading them as sample fractions.
  5. [§5.2.2] The concluding sentence "there is no evidence for significant systematics that will bias the fidelity of the Hybrid model as the detection limit or seeing varies" is stronger than what the box plots show; recommend rewording to something like "we do not detect a strong bias in the median scores, although classification error rates in these regimes remain unquantified."
  6. [§2.1] The sentence beginning "we expect the HST x LS dataset exhibits representative source distributions..." is an assumption rather than a demonstrated property; please present it as such and, if possible, support it with a quantitative comparison, for example against DESI or SDSS number counts in the same magnitude and color bins.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central accuracy claims rest on independent spectroscopic and astrometric test sets, not on the model's own inputs.

full rationale

The paper's central claim—that the Hybrid model separates resolved from unresolved LS sources with higher accuracy than LS morphological typing—is supported by evaluation on DESI DR1 (~15 million objects) and SDSS DR17 (~3 million objects) with spectroscopic labels, plus a Gaia DR3 astrometric sample. These test labels are independent of the model's morphological features and of the HST COSMOS training labels. The 5-fold CV on the HST training set is standard supervised evaluation, and the label revision in Section 2.1 is disclosed and cross-checked against DESI and zCOSMOS spectroscopy. The Hybrid weight αhyb is selected from training-set CV (αhyb = 0.33) and only checked on the test sets, not adopted from them, so no fitted parameter is renamed as a prediction. The mask feature in Equation (6) encodes LS fitbits including the Gaia point-source flag, and Section 6.2 transparently overrides scores for Gaia stars by adding 1; this is disclosed domain-knowledge post-processing rather than a disguised prediction, and the primary accuracy claims in Tables 4 and 5 do not depend on that override. Self-citations to TM18 and Miller et al. (2017) supply methodology, feature definitions, and the FPR threshold convention, but they are not load-bearing for the new catalog or the new performance results. Assumptions about COSMOS representativeness, and the degradation seen in shallow, poor-seeing, or crowded fields (Figures 7-9), are limitations and transferability risks, not circular steps. No derivation chain reduces to its own inputs by construction.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central claim relies on the transferability of HST COSMOS labels across the LS footprint, on the consistency of LS Tractor and aperture photometry, and on the accuracy of DESI and SDSS spectroscopic labels used for evaluation. These are standard domain assumptions for survey ML classifiers; none are ad hoc inventions.

free parameters (4)
  • alpha_hyb = 0.33
    Weight balancing the blue (gr) and red (iz) sub-models in the Hybrid score, tuned on training-set cross-validation (Section 4).
  • P(MAG_AUTO) polynomial coefficients = Not stated explicitly
    Third-order polynomial boundary between stars and galaxies in training labels, fit to KDE-derived boundary points (Section 2.1).
  • XGBoost hyperparameters = eta=0.02; n_estimator=850/1300/800; max_depth=12/18/12; min_child_weight=6/10/6; gamma=1.0/0.1/0.5…
    Tuned via grid search in an inner cross-validation loop (Section 4); the paper reports low sensitivity to these choices.
  • KDE kernel size and slope threshold = 0.075 and 0.02
    Hand-chosen parameters defining the revised star-galaxy label boundary in Section 2.1.
assumptions (3)
  • domain assumption HST COSMOS labels are reliable and transferable to the full LS footprint
    Training labels come from Leauthaud et al. 2007 with a modified boundary; the paper assumes these labels represent true star and galaxy classes across varying seeing, depth, and stellar density (Section 2.1).
  • domain assumption LS Tractor dchisq and aperture photometry are consistent across the survey
    Model features rely on LS DR10 catalog products; varying observing conditions shift the feature distributions, as shown in Section 5.2.2.
  • domain assumption DESI and SDSS spectroscopic classifications are correct after filtering
    Used as ground truth for evaluation; the paper applies several purity filters but assumes remaining labels are accurate (Section 2.2).

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Morphological Model to Separate Resolved-Unresolved Sources in the DESI Legacy Surveys: Application in the LS4 Alert Stream." pith.science (2026). https://pith.science/paper/RTEFDNPB

@misc{pith2026250517174,
  author       = {Pith},
  title        = {Pith review of: A Morphological Model to Separate Resolved-Unresolved Sources in the DESI Legacy Surveys: Application in the LS4 Alert Stream},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RTEFDNPB}},
  note         = {Machine review of arXiv:2505.17174}
}
abstract

Separating resolved and unresolved sources in large imaging surveys is a fundamental step to enable downstream science, such as searching for extragalactic transients in wide-field time-domain surveys. Here we present our method to effectively separate point sources from the resolved, extended sources in the Dark Energy Spectroscopic Instrument (DESI) Legacy Surveys (LS). We develop a supervised machine-learning model based on the Gradient Boosting algorithm $\texttt{XGBoost}$. The features input to the model are purely morphological and are derived from the tabulated LS data products. We train the model using $\sim$$2\times10^5$ LS sources in the COSMOS field with HST morphological labels and evaluate the model performance on LS sources with spectroscopic classification from the DESI Data Release 1 ($\sim$$2\times10^7$ objects) and the Sloan Digital Sky Survey Data Release 17 ($\sim$$3\times10^6$ objects), as well as on $\sim$$2\times10^8$ Gaia stars. A significant fraction of LS sources are not observed in every LS filter, and we therefore build a ''Hybrid'' model as a linear combination of two \texttt{XGBoost} models, each containing features combining aperture flux measurements from the ''blue'' ($gr$) and ''red'' ($iz$) filters. The Hybrid model shows a reasonable balance between sensitivity and robustness, and achieves higher accuracy and flexibility compared to the LS morphological typing. With the Hybrid model, we provide classification scores for $\sim$$3\times10^9$ LS sources, making this the largest ever machine-learning catalog separating resolved and unresolved sources. The catalog has been incorporated into the real-time pipeline of the La Silla Schmidt Southern Survey (LS4), enabling the identification of extragalactic transients within the LS4 alert stream.

Figures

Figures reproduced from arXiv: 2505.17174 by the authors.

Figure 1
Figure 1. The modified boundary in the MAG AUTO − MU MAX/MAG AUTO plane reasonably separate stars (orange) from galaxies (blue) in the HST training set. The original boundary in Leauthaud et al. (2007) is displayed as the dashed line. As illustrated in the text, for objects >20 mag we define a new boundary as a polynomial fit (dotted line) to the boundary in each MAG AUTO bin (asterisks), determined in a data-driven way. For … view at source ↗
Figure 2
Figure 2. Distributions of stars and galaxies suggest that, in the faintest magnitude bins with single band filters (e.g., r), the star fraction deviates significantly from the reality. Using the flux-weighted average magnitude white mag (defined in the text) alleviates the bias. Upper: the distribution of HST stars and galaxies in the COSMOS field among all cross-matches between LS and HST (dashed line) as well as in the tra… view at source ↗
Figure 3
Figure 3. ROC curves of the three XGBoost models (White, grz, and Hybrid) as evaluated in the training set by CVs in different magnitude bins. Models including aperture fluxes from individual filters (grz and Hybrid) outperforms the White model which only includes the average aperture flux of all filters, especially for faint sources. The thick lines show the performance of each model on the entire training set. The empty bla… view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: The FPR and FNR (the fraction of misclassi￾fied stars and galaxies) as a function of white mag of the LS model (dashed black lines) and our XGBoost models (solid blue, purple, and green lines). The bin widths are 0.5 mag. Each marker corresponds to the FPR/FNR evaluate…
Figure 5
Figure 5. Figure 5: Accuracy as a function of the white mag evaluated in the HST training set (using CV) and two independent test sets (DESI and SDSS). The bin widths are 0.5 mag, and in each bin, the number of stars and galaxies have been normalized to match their relative ratio in the f…
Figure 6
Figure 6. Figure 6: Score distribution of the Hybrid model for Gaia stars in the northern (left) and southern (right) LS footprint. The grey histograms correspond to the entire dataset, whereas the orange (purple) histograms present the subset of sources missing photometry in both g and r…
Figure 7
Figure 7. Figure 7: Box plots of model scores showing how the Hybrid model downgrades for fainter sources (16–23 mag; from left to right) in different fields with a variety of 5-σ detection limits (23–26 mag; from red to blue colors) in LS. Each box presents the distribution of scores for…
Figure 8
Figure 8. Figure 8: Similar as [PITH_FULL_IMAGE:figures/full_fig_p018_8.png]
Figure 9
Figure 9. Figure 9: Box plots showing the score distribution for DESI stars and galaxies as a function of stellar crowding. Model scores are lower for stars in fields with higher stellar density, because they are more likely to be blended with another star. The score distribution of galax…
Figure 10
Figure 10. Figure 10: The source density map of LS-PSC. The red boxes indicate the bricks containing sources labeled as bailout. The actual size of the bricks (∼0.25◦ × 0.25◦ ) is much smaller than the box size. The dash line marks the Galactic plane. The source density in the northern foo…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

53 extracted references · 4 canonical work pages

  1. [1]

    2022, ApJS, 259, 35, doi: 10.3847/1538-4365/ac4414

    Abdurro’uf, Accetta, K., Aerts, C., et al. 2022, ApJS, 259, 35, doi: 10.3847/1538-4365/ac4414

    doi:10.3847/1538-4365/ac4414 2022

  2. [2]

    M., Davis, T

    Alexander, D. M., Davis, T. M., Chaussidon, E., et al. 2023, AJ, 165, 124, doi: 10.3847/1538-3881/acacfc Astropy Collaboration, Robitaille, T. P., Tollerud, E. J., et al. 2013, A&A, 558, A33, doi: 10.1051/0004-6361/201322068 Astropy Collaboration, Price-Whelan, A. M., Sip˝ ocz, B. M., et al. 2018, AJ, 156, 123, doi: 10.3847/1538-3881/aabc4f 23

    doi:10.3847/1538-3881/acacfc 2023

  3. [3]

    M., Brunner, R

    Ball, N. M., Brunner, R. J., Myers, A. D., & Tcheng, D. 2006, ApJ, 650, 497, doi: 10.1086/507440

    doi:10.1086/507440 2006

  4. [4]

    C., & Szapudi, I

    Beck, R., Dodds, S. C., & Szapudi, I. 2022, MNRAS, 515, 4711, doi: 10.1093/mnras/stac1714

    doi:10.1093/mnras/stac1714 2022

  5. [5]

    C., Kulkarni, S

    Bellm, E. C., Kulkarni, S. R., Graham, M. J., et al. 2019, PASP, 131, 018002, doi: 10.1088/1538-3873/aaecbe

    doi:10.1088/1538-3873/aaecbe 2019

  6. [6]

    1996, A&AS, 117, 393, doi: 10.1051/aas:1996164

    Bertin, E., & Arnouts, S. 1996, A&AS, 117, 393, doi: 10.1051/aas:1996164

    doi:10.1051/aas:1996164 1996

  7. [7]

    C., Magnier, E

    Chambers, K. C., Magnier, E. A., Metcalfe, N., et al. 2016, arXiv e-prints, arXiv:1612.05560, doi: 10.48550/arXiv.1612.05560

    arXiv doi:10.48550/arxiv.1612.05560 2016

  8. [8]

    2023, ApJ, 944, 107, doi: 10.3847/1538-4357/acb3c2

    Chaussidon, E., Y` eche, C., Palanque-Delabrouille, N., et al. 2023, ApJ, 944, 107, doi: 10.3847/1538-4357/acb3c2

    doi:10.3847/1538-4357/acb3c2 2023

Show all 53 references
  1. [9]

    Chen, T., & Guestrin, C. 2016, in Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’16 (New York, NY, USA: Association for Computing Machinery), 785794, doi: 10.1145/2939672.2939785

    2016

  2. [10]

    M., Tzanidakis, A., et al

    De, K., Kasliwal, M. M., Tzanidakis, A., et al. 2020, ApJ, 905, 58, doi: 10.3847/1538-4357/abb45c

    2020 doi

  3. [11]

    M., Riddle, R., et al

    Dekany, R., Smith, R. M., Riddle, R., et al. 2020, PASP, 132, 038001, doi: 10.1088/1538-3873/ab4ca2 DESI Collaboration, Abdul-Karim, M., Adame, A. G., et al. 2025, arXiv e-prints, arXiv:2503.14745, doi: 10.48550/arXiv.2503.14745

    2020 doi arXiv

  4. [12]

    J., Lang, D., et al

    Dey, A., Schlegel, D. J., Lang, D., et al. 2019, AJ, 157, 168, doi: 10.3847/1538-3881/ab089d

    2019 doi

  5. [13]

    J., Olsen, K., Economou, F., et al

    Fitzpatrick, M. J., Olsen, K., Economou, F., et al. 2014, in Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, Vol. 9149, Observatory Operations: Strategies, Processes, and Systems V, ed. A. B. Peck, C. R. Benn, & R. L. Seaman, 91491T, doi: 10.1117/1...

    2014 doi

  6. [14]

    A., Sharma, Y., et al

    Fremling, C., Miller, A. A., Sharma, Y., et al. 2020, ApJ, 895, 32, doi: 10.3847/1538-4357/ab8943

    2020 doi

  7. [15]

    Friedman, J. H. 2001, The Annals of Statistics, 29, 1189 , doi: 10.1214/aos/1013203451

    2001

  8. [16]

    2024, ApJS, 271, 54, doi: 10.3847/1538-4365/ad2ae6 Gaia Collaboration, Prusti, T., de Bruijne, J

    Fu, Y., Wu, X.-B., Li, Y., et al. 2024, ApJS, 271, 54, doi: 10.3847/1538-4365/ad2ae6 Gaia Collaboration, Prusti, T., de Bruijne, J. H. J., et al. 2016, A&A, 595, A1, doi: 10.1051/0004-6361/201629272 Gaia Collaboration, Babusiaux, C., van Leeuwen, F., et al. 2018, A&A, 616, A10...

    2024 doi

  9. [17]

    J., Kulkarni, S

    Graham, M. J., Kulkarni, S. R., Bellm, E. C., et al. 2019, PASP, 131, 078001, doi: 10.1088/1538-3873/ab006c

    2019 doi

  10. [18]

    J., Bloemen, S., Vreeswijk, P

    Groot, P. J., Bloemen, S., Vreeswijk, P. M., et al. 2024, PASP, 136, 115003, doi: 10.1088/1538-3873/ad8b6a

    2024 doi

  11. [19]

    J., Ruiz-Macias, O., et al

    Hahn, C., Wilson, M. J., Ruiz-Macias, O., et al. 2023, AJ, 165, 253, doi: 10.3847/1538-3881/accff8

    2023 doi

  12. [20]

    R., Millman, K

    Harris, C. R., Millman, K. J., van der Walt, S. J., et al. 2020, Nature, 585, 357, doi: 10.1038/s41586-020-2649-2

    2020 doi

  13. [21]

    J., Hand, D

    Henrion, M., Mortlock, D. J., Hand, D. J., & Gandy, A. 2011, MNRAS, 412, 2286, doi: 10.1111/j.1365-2966.2010.18055.x

    2011

  14. [22]

    Hunter, J. D. 2007, Computing in Science and Engineering, 9, 90, doi: 10.1109/MCSE.2007.55 Ivezi´ c,ˇZ., Kahn, S. M., Tyson, J. A., et al. 2019, ApJ, 873, 111, doi: 10.3847/1538-4357/ab042c

    2007 doi

  15. [23]

    2019, MNRAS, 485, 4539, doi: 10.1093/mnras/stz680 Kov´ acs, A., & Szapudi, I

    Jin, X., Zhang, Y., Zhang, J., et al. 2019, MNRAS, 485, 4539, doi: 10.1093/mnras/stz680 Kov´ acs, A., & Szapudi, I. 2015, MNRAS, 448, 1305, doi: 10.1093/mnras/stv063

    2019 doi

  16. [24]

    Kron, R. G. 1980, ApJS, 43, 305, doi: 10.1086/190669

    1980 doi

  17. [25]

    2007, ApJS, 172, 219, doi: 10.1086/516598

    Leauthaud, A., Massey, R., Kneib, J.-P., et al. 2007, ApJS, 172, 219, doi: 10.1086/516598

    2007 doi

  18. [26]

    J., Le F` evre, O., Renzini, A., et al

    Lilly, S. J., Le F` evre, O., Renzini, A., et al. 2007, ApJS, 172, 70, doi: 10.1086/516589

    2007 doi

  19. [27]

    J., Le Brun, V., Maier, C., et al

    Lilly, S. J., Le Brun, V., Maier, C., et al. 2009, ApJS, 184, 218, doi: 10.1088/0067-0049/184/2/218

    2009 doi

  20. [28]

    2018, A&A, 616, A2, doi: 10.1051/0004-6361/201832727

    Lindegren, L., Hern´ andez, J., Bombrun, A., et al. 2018, A&A, 616, A2, doi: 10.1051/0004-6361/201832727

    2018 doi

  21. [30]

    2019, PASP, 131, 038002, doi: 10.1088/1538-3873/aaf3fa

    Mahabal, A., Rebbapragada, U., Walters, R., et al. 2019, PASP, 131, 038002, doi: 10.1088/1538-3873/aaf3fa

    2019 doi

  22. [31]

    J., Laher, R

    Masci, F. J., Laher, R. R., Rusholme, B., et al. 2019, PASP, 131, 018003, doi: 10.1088/1538-3873/aae8ac

    2019 doi

  23. [32]

    A., & Hall, X

    Miller, A. A., & Hall, X. J. 2021, PASP, 133, 054502, doi: 10.1088/1538-3873/abf038

    2021 doi

  24. [33]

    A., Kulkarni, M

    Miller, A. A., Kulkarni, M. K., Cao, Y., et al. 2017, AJ, 153, 73, doi: 10.3847/1538-3881/153/2/73

    2017 doi

  25. [34]

    A., Abrams, N

    Miller, A. A., Abrams, N. S., Aldering, G., et al. 2025, arXiv e-prints, arXiv:2503.14579, doi: 10.48550/arXiv.2503.14579

    2025 doi arXiv

  26. [35]

    2023, ApJS, 269, 3, doi: 10.3847/1538-4365/acfaa2

    Moustakas, J., Lang, D., Dey, A., et al. 2023, ApJS, 269, 3, doi: 10.3847/1538-4365/acfaa2

    2023 doi

  27. [36]

    J., Bilicki, M., Pollo, A., et al

    Nakoneczny, S. J., Bilicki, M., Pollo, A., et al. 2021, A&A, 649, A81, doi: 10.1051/0004-6361/202039684

    2021 doi

  28. [37]

    Nikutta, R., Fitzpatrick, M., Scott, A., & Weaver, B. A. 2020, Astronomy and Computing, 33, 100411, doi: 10.1016/j.ascom.2020.100411

    2020

  29. [38]

    2011, Journal of Machine Learning Research, 12, 2825

    Pedregosa, F., Varoquaux, G., Gramfort, A., et al. 2011, Journal of Machine Learning Research, 12, 2825

    2011

  30. [39]

    A., Fremling, C., Sollerman, J., et al

    Perley, D. A., Fremling, C., Sollerman, J., et al. 2020, ApJ, 904, 35, doi: 10.3847/1538-4357/abbd98 24

    2020 doi

  31. [40]

    S., Wadadekar, Y., Kembhavi, A., & Joseph, K

    Philip, N. S., Wadadekar, Y., Kembhavi, A., & Joseph, K. B. 2002, A&A, 385, 1119, doi: 10.1051/0004-6361:20020219

    2002 doi

  32. [41]

    A., et al

    Raichoor, A., Moustakas, J., Newman, J. A., et al. 2023, AJ, 165, 126, doi: 10.3847/1538-3881/acb213

    2023 doi

  33. [42]

    A., Jegou Du Laz, T., et al

    Rehemtulla, N., Miller, A. A., Jegou Du Laz, T., et al. 2024, ApJ, 972, 7, doi: 10.3847/1538-4357/ad5666

    2024 doi

  34. [43]

    P., Shanks, T., Cannon, R

    Ross, N. P., Shanks, T., Cannon, R. D., et al. 2008, MNRAS, 387, 1323, doi: 10.1111/j.1365-2966.2008.13332.x

    2008

  35. [44]

    J., et al

    Sevilla-Noarbe, I., Hoyle, B., March˜ a, M. J., et al. 2018, MNRAS, 481, 5451, doi: 10.1093/mnras/sty2579

    2018 doi

  36. [45]

    J., Prieto, J

    Shappee, B. J., Prieto, J. L., Grupe, D., et al. 2014, ApJ, 788, 48, doi: 10.1088/0004-637X/788/1/48

    2014 doi

  37. [46]

    2023, A&A, 680, A109, doi: 10.1051/0004-6361/202347576

    Stoppa, F., Bhattacharyya, S., Ruiz de Austri, R., et al. 2023, A&A, 680, A109, doi: 10.1051/0004-6361/202347576

    2023 doi

  38. [47]

    Tachibana, Y., & Miller, A. A. 2018, PASP, 130, 128001, doi: 10.1088/1538-3873/aae3d9

    2018 doi

  39. [48]

    Tonry, J. L. 2011, PASP, 123, 58, doi: 10.1086/657997

    2011 doi

  40. [49]

    L., Denneau, L., Heinze, A

    Tonry, J. L., Denneau, L., Heinze, A. N., et al. 2018, PASP, 130, 064505, doi: 10.1088/1538-3873/aabadf van Velzen, S., Gezari, S., Hammerstein, E., et al. 2021, ApJ, 908, 4, doi: 10.3847/1538-4357/abc258

    2018 doi

  41. [50]

    C., de Carvalho, R

    Vasconcellos, E. C., de Carvalho, R. R., Gal, R. R., et al. 2011, AJ, 141, 189, doi: 10.1088/0004-6256/141/6/189 von Marttens, R., Marra, V., Quartin, M., et al. 2024, MNRAS, 527, 3347, doi: 10.1093/mnras/stad3373

    2011 doi

  42. [51]

    Waskom, M. L. 2021, Journal of Open Source Software, 6, 3021, doi: 10.21105/joss.03021 Wes McKinney. 2010, in Proceedings of the 9th Python in Science Conference, ed. St´ efan van der Walt & Jarrod Millman, 56 – 61, doi: 10.25080/Majora-92bf1922-00a

    2021 doi

  43. [52]

    G., Adelman, J., Anderson, John E., J., et al

    York, D. G., Adelman, J., Anderson, John E., J., et al. 2000, AJ, 120, 1579, doi: 10.1086/301513

    2000 doi

  44. [53]

    Z., Hafezianzadeh, F., Zhang, Y., et al

    Zeraatgari, F. Z., Hafezianzadeh, F., Zhang, Y., et al. 2024, MNRAS, 527, 4677, doi: 10.1093/mnras/stad3436

    2024 doi

  45. [54]

    bailed out

    Zhou, R., Dey, B., Newman, J. A., et al. 2023, AJ, 165, 58, doi: 10.3847/1538-3881/aca5fb 25 APPENDIX A. FEATURE IMPORTANCE In this section we report the feature importance for each XGBoost model, defined as the average gain in accuracy when a feature is used in a split. The r...

    2023 doi

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.