Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

Evaluation of nearby young moving groups based on unsupervised machine learning

T0 review · 4 major / 4 minor · reviewed 2026-08-14 · deepseek-v4-flash

Pith's one-line read Unsupervised clustering of young stars recovers six known moving groups and merges three others into two, so standard group boundaries should be revised.

desk verdict An honest re-clustering of known moving-group members that recovers six groups but overreaches when it defines THC by post-hoc union of clusters the two algorithms keep separate. read the letter →

arxiv 1908.05922 v1 pith:52RY5EFV submitted 2019-08-16 astro-ph.SR

classification astro-ph.SR
keywords nearbyyoungmovinggroupsunsupervisedmachinelearningK-meansclusteringagglomerativestellarkinematicsgroupmembershipassociationssolarneighbourhood
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

This paper asks whether the nine nearby young moving groups — loose, unbound associations of young stars that share ages and space motions — are distinct real structures or artefacts of how they were discovered. The authors feed 652 previously confirmed members, described by Galactic position, space velocity, and a five-level age class, into two unsupervised clustering algorithms and search for groupings that recur across 16,000 randomised trials. Six of the nine groups (AB Doradus, Argus, $\beta$ Pic, TWA, Carina, and Volans-Carina) reappear as stable clusters, while ThOr, Columba, and TucHor do not; in their place the algorithms form a ThOr-Columba hybrid and a TucHor-Columba hybrid. If this is right, some standard moving-group boundaries in the solar neighbourhood should be redrawn.

What carries the argument

The load-bearing mechanism is a repeated-clustering stability test. K-means (which minimises within-cluster squared distances) and agglomerative clustering with Ward linkage are each run 1,000 times for every choice of cluster number from 3 to 10, with each run using a random 95 per cent subsample of the 652 input stars after the positions, velocities, and age classes have been rescaled to comparable ranges. A star is counted as a bona fide member of a newly recognised group only if it falls in the same cluster in more than 70 per cent of the runs of both algorithms. The conclusions are built from the recurrent grouping pattern across the 16,000 runs rather than from any single clustering.

What would settle it

Run the same two clustering algorithms on a sample of young nearby stars selected without prior membership labels—for example, all stars within 100 pc with high-precision astrometry, radial velocities, and lithium-based ages—without fixing the cluster number to the old nine-group count, and ask whether the six recovered groups persist while ThOr-Col and THC reappear as their own clusters. If the two hybrid groups dissolve into the field or recombine differently, the proposed redrawing of ThOr, Columba, and TucHor is not robust.

Watch

Extended reading notes

Core claim

The central claim is that unsupervised clustering in the seven-dimensional space of Galactic position ($X,Y,Z$), space velocity ($U,V,W$), and categorical age recovers six previously defined moving groups as robust entities but does not recover ThOr, Columba, and TucHor as separate clusters. With the cluster number set to nine, both K-means and agglomerative clustering place essentially all of ThOr together with a large subset of Columba in a group the authors name ThOr-Col, and place TucHor with most of the remaining Columba members in a group named THC. The authors conclude that those three groups cannot be cleanly separated with the available age-spatio-kinematic information, so their traditional division should be revised; they note this agrees with an earlier proposal to merge TucHor and Columba. They also report that 11 of 18 known planet-host stars remain bona fide members of the newly defined groups, while 7 are no longer consistently retained.

Load-bearing premise

The test depends on the authors' earlier membership catalog being an unbiased and complete picture of the true moving groups; the machine can only regroup the 652 stars that someone already accepted as members, so any group missing from that input or any selection bias in it would be inherited by the clusters.

Editorial extensions

If this is right

  • Six of the traditional nine groupings—AB Doradus, Argus, $\beta$ Pic, TWA, Carina, and Volans-Carina—are stable clusters in the repeated trials, so those moving-group labels appear to correspond to real spatio-kinematic structures.
  • The ThOr-Col and THC groupings imply that ThOr, Columba, and TucHor should not be treated as three independent associations in studies of young stars.
  • The THC result supports the earlier suggestion that TucHor and Columba members should be combined because they cannot be reliably separated.
  • The new bona fide member lists keep 11 of 18 planet-host stars and drop 7, so the redefined groups change which stars are used as young-planet host samples.
  • Adopting the new eight-group configuration provides a consistently defined input set for future, fully unbiased searches of nearby young stars.

Reading between the lines

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

  • Because the input stars were preselected as known members, the recovery of six groups does not by itself prove those groups are real; the genuinely new information is the recombination of ThOr, Columba, and TucHor.
  • Using continuous ages with measured uncertainties instead of five broad age classes could split the ThOr-Col hybrid into finer populations, a testable extension of this scheme.
  • Applying the same bootstrap-clustering protocol to an all-sky sample that includes field stars would test whether the eight proposed groups persist without any prior membership labels.
  • Requiring agreement among more than two algorithms, or a higher consensus threshold, would likely shrink the new member lists, so the group sizes should be read as sensitive to the chosen stability criterion.
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

4 major / 4 minor

Summary. The paper applies two unsupervised clustering algorithms (K-means and Agglomerative Clustering) to 652 previously identified bona fide members of nine nearby young moving groups from the authors' prior catalog (Lee & Song 2019), using XYZ, UVW, and an age-class variable. For each algorithm and for nclusters values from 3 to 10, the authors run 1000 trials with 95% random sampling. At nclusters=9, they report recovering six of the nine input groups (AB Doradus, Argus, beta Pic, Carina, TWA, Volans-Carina) and propose two new merged groups: ThOr-Col (ThOr plus part of Columba) and THC (TucHor plus part of Columba). The paper concludes that some traditionally defined moving group boundaries are not supported by the spatio-kinematic-age data.

Significance. If the central claim is validated, the finding that ThOr, Columba, and TucHor do not emerge as distinct clusters would be of real interest to the nearby-young-moving-group community and would lend support to earlier suggestions of mergers (e.g., Zuckerman et al. 2011). The paper is commendable for making its clustering exercise transparent via repeated bootstrap runs and for presenting detailed visualizations of run-to-run variability. However, the validity of the two new merged groups, and of the recovery claim for six groups, is weakened by internal inconsistencies with the paper's own recognition criteria, as detailed below.

major comments (4)
  1. [Section 3.2, Table 1] The paper's own reliability criterion is not met for Carina and VCA. In Table 1, Carina is recognized in 70% of K-means runs but only 35% of Agglomerative Clustering runs, and VCA likewise in 70% and 35% of runs, respectively. According to the criterion stated in Section 2.3.2 that members are recognized as bona fide only when consistently assessed in more than 70% of trial runs between the two algorithms, Carina and VCA would not be reliably recognized by the Agglomerative algorithm. The paper nevertheless lists them among the six recovered groups, which overstates the support for the recovery claim.
  2. [Section 3.2, Table 1] The definition of THC violates the paper's own group-recognition standard. The text states that in the remaining TucHor/Columba members, the two groups found by K-means and Agglomerative Clustering 'do not match well,' and that a single group is then defined enclosing these two group members. THC is therefore a residual set, not a cluster produced by either algorithm, and it cannot satisfy the >70% consistency criterion across algorithms. The claim that TucHor and Columba form a new combined group is unsupported by the clustering output. Appendix A reinforces this by showing that at nclusters=8 Agglomerative Clustering keeps Columba and TucHor separate, and at nclusters=10 TucHor is split into two groups; THC appears to be an artifact of the chosen nclusters=9 and the post-hoc merging of disagreeing clusters.
  3. [Section 2.1 and Section 1] The evaluation is not independent of the input labels. The clustering input is the authors' own Lee & Song (2019) bona fide member catalog, and the 'recovery' of six groups is measured against those same labels. Consequently, the recovery of six groups is partly a restatement of the input selection and cannot by itself validate the reality of those groups. The language in the abstract and Section 1 about evaluating NYMGs 'without relying on any previous knowledge' is misleading because, while the algorithms themselves are unsupervised, the data selection and the decision to use nclusters=9 (Section 3.2) reintroduce prior knowledge. The paper should explicitly limit the strength of the recovery claims.
  4. [Section 3.2, Appendix A] The choice of nclusters=9 is not robustly justified. The four-group test with nclusters=4 recovering the four input groups is suggestive but does not establish that nclusters=9 is the correct number of groups for the full dataset. Appendix A shows that results are qualitatively sensitive to nclusters: Carina and VCA are merged at nclusters=8 and separated only at nclusters=9-10, and TucHor/Columba separations change with nclusters. A quantitative cluster-validity index (e.g., silhouette width) or a formal stability analysis across nclusters should be provided before drawing strong conclusions about the number and composition of distinct moving groups.
minor comments (4)
  1. [Abstract] The sentence 'Three the other known groups are recognised as well; however, they are combined into two new separate groups' is grammatically awkward; it should be rephrased, for instance as 'The remaining three known groups are not recovered in their original form; instead, they are combined into two new groups.'
  2. [Appendix A] The table header 'Tabel A1' contains a typo and should read 'Table A1'.
  3. [Section 2.3.1] The transformation constants for the age class (subtract 3.125, multiply by 1.25) are stated without explanation; providing the rationale for these particular values would aid reproducibility.
  4. [Section 3.2, Table 1] The notation using primes (′, ″) to indicate partial membership is not defined in the table itself; it should be defined in the caption or in the text immediately preceding the table.

Circularity Check

1 steps flagged · score 6.0 of 10

The claimed THC group is a post-hoc union of leftover input members, not a cluster; the six recovered groups and ThOr-Col are genuine unsupervised outputs, so the circularity is partial.

  1. renaming known result [Section 3.2, definition of THC immediately after Table 1]
    "While each K-means and Agglomerative Clustering algorithm recognises two groups in the remaining members, the two groups from these algorithms do not match well (i.e., the specific content of two groups from both algorithms are different). Therefore, we define a single group enclosing these two group members. This group is called THC (TucHor-Col)."

    At nclusters=9, the two algorithms never agree on a TucHor+Columba cluster: Table 1 shows K-means producing Columba'+TucHor' (60%) and Columba'+TucHor'' (60%), while Agglomerative Clustering produces Columba' (80%) and TucHor' (50%). THC is therefore defined post hoc as the union of the leftover members of two input groups, and the abstract's statement that these three known groups 'are combined into two new separate groups (ThOr+Columba and TucHor+Columba)' is true by definition for THC, not by the clustering output. The paper's own >70% consistent-assignment criterion for bona fide groups is not met by THC.

full rationale

The core clustering procedure is not circular: the two unsupervised algorithms never see the input group labels, so the recovery of ABDor, Argus, BPMG, TWA, Carina, and VCA is a genuine consistency check, and ThOr-Col is also an emergent cluster rather than a relabeling of the input. The use of the authors' own Paper I membership catalog as input is a real selection-bias limitation, and the paper explicitly concedes that a fully unbiased test would need to include field stars and unknowns; but that is a limitation of experimental design, not a derivation that reduces to its own input. The one circular step is the THC group: it is not a cluster produced by either algorithm at nclusters=9 but a group defined by the authors around the disagreeing residual members of the input TucHor and Columba groups. Since the novel claim that TucHor and Columba merge into a new group rests on this definition, that component of the paper's central result is circular by construction, while the remaining claims retain independent content.

Assumptions & free parameters 3 free parameters · 3 assumptions · 2 invented entities

The central result rests on two invented merged groups (ThOr-Col and THC) that are defined from the clustering output, with no independent validation. The analysis also depends on several free parameters: the number of clusters, the bootstrapping threshold, and the age-scaling constants. The foundational assumption is that the input membership lists from the authors' prior paper are unbiased enough to serve as ground truth, which is the main source of circularity.

free parameters (3)
  • nclusters = 9 for the main result, range 3 to 10 tested
    The number of clusters is the key user-chosen parameter. The justification is a small test with four groups that says nclusters equal to the number of input groups works, which is not a guarantee for the full dataset.
  • Resilience threshold = 70 percent of runs across both algorithms
    Stars assigned to a group in more than 70 percent of trials are called bona fide members. This threshold is chosen by the authors and directly sets the final membership lists.
  • Age class transformation constants = scale by 1.25 and shift by 3.125
    The age class is transformed to match the scaled XYZ and UVW ranges; the constants are chosen manually to make ranges similar.
assumptions (3)
  • domain assumption The bona fide membership lists of Lee and Song (2019) are a sufficiently unbiased and complete representation of the true nearby young moving groups.
    This assumption is the foundation of the analysis, stated in Section 2.1 where the input data are defined. If the input lists contain selection biases or errors, the clustering will inherit them.
  • domain assumption The 7D spatio-kinematic plus age-class space is sufficient to characterize moving group membership.
    The paper excludes metallicity because it has large uncertainty, arguing that young nearby stars share similar abundances. This is a modeling choice that could hide group distinctions visible in chemistry.
  • domain assumption The two clustering algorithms and their underlying distance geometry are appropriate for identifying stellar moving groups.
    K-means assumes roughly spherical clusters in the scaled space, which is not necessarily true for real associations, and the paper does not test other algorithms such as density-based clustering.
invented entities (2)
  • ThOr-Col group
    purpose: A proposed merged nearby young moving group containing ThOr and part of Columba.
    The group is defined by the clustering output from the authors' own input catalog, and no independent kinematic, isochronal, or chemical evidence is presented beyond the clustering frequencies.
  • THC group
    purpose: A proposed merged nearby young moving group containing TucHor and the rest of Columba.
    The group is assembled from the remaining members after the first merger, with the authors explicitly noting that the two algorithms disagree on how to split these stars, and no independent evidence is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Evaluation of nearby young moving groups based on unsupervised machine learning." pith.science (2026). https://pith.science/paper/52RY5EFV

@misc{pith2026190805922,
  author       = {Pith},
  title        = {Pith review of: Evaluation of nearby young moving groups based on unsupervised machine learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/52RY5EFV}},
  note         = {Machine review of arXiv:1908.05922}
}
abstract

Nearby young stellar moving groups have been identified by many research groups with different methods and criteria giving rise to cautions on the reality of some groups. We aim to utilise moving groups in an unbiased way to create a list of unambiguously recognisable moving groups and their members. For the analysis, two unsupervised machine learning algorithms (K-means and Agglomerative Clustering) are applied to previously known bona fide members of nine moving groups from our previous study. As a result of this study, we recovered six previously known groups (AB Doradus, Argus, $\beta$-Pic, Carina, TWA, and Volans-Carina). Three the other known groups are recognised as well; however, they are combined into two new separate groups (ThOr+Columba and TucHor+Columba).

Figures

Figures reproduced from arXiv: 1908.05922 by the authors.

Figure 1
Figure 1. Distribution of input data. Raw and transformed scales are displayed along the bottom and the top X-axis, respectively. Units for raw data are pc (X, Y, and Z) and km s−1 (U, V, and W). Age (age class) has no unit. and an algorithm, we need to identify the most frequent pattern of grouping through the one thousand runs. The tested input parameter, nclusters, is in a range of 3 to 10 [PITH_FULL_IMAGE:figures/full_fi… view at source ↗
Figure 2
Figure 2. Results from one thousand runs with nclusters=9. Left and right panels present the results of K-means and Agglomerative Clustering algorithms, respectively. The X and Y-axis represent a run number and stellar index, respectively. The results are sorted along the X and Y-axis for a better display and interpretation. Stars (the Y-axis) are sorted based on previously known grouping. The “N run” (the X-axis) is sorted t… view at source ↗
Figure 3
Figure 3. Distribution of bona fide members of the new eight group (this study; top) and those of classical groups (input data; bottom). there is a difficulty assigning some of them into either Tu￾cHor of Columba. Our result supports the claim of Zucker￾man et al. (2011). The input data include eighteen planet-host stars. Among these 18 stars, 11 stars are retained as bona fide members via cluster analysis in this study, whil… view at source ↗

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. Peering through the disc of HD 98800 BaBb. Precise timing predictions for the HD 98800 AaAb occultation

    astro-ph.EP 2026-07 conditional novelty 5.0 of 10

    New RVs and multi-wavelength astrometry tighten the HD 98800 outer orbit by ~2×, yielding 5–15-day 1σ windows for the 2025–2031 disc occultation of AaAb.

Reference graph

Works this paper leans on

24 extracted references · 23 canonical work pages · cited by 1 Pith paper

  1. [1]

    Elliott P., Bayo A., Melo C. H. F., Torres C. A. O., Sterzik M., Quast G. R., Montes D., Brahm R., 2014, , 590, 13

  2. [2]

    Faherty J., Bochanski J., Gagn\' e J., Nelson O., Coker K., Smithka I., Desir D., Vasquez C., 2018, , 863, 91

  3. [3]

    Gagn\' e J., Lafreni\` e re D., Doyon R., Malo L., Artigau \' E ., 2014, , 783, 121

  4. [4]

    et al., 2018, , 856, 23

    Gagn\' e J. et al., 2018, , 856, 23

  5. [5]

    H., Zuckerman B., Weintraub D.A., Forveille T., 1997, Science, 277, 67

    Kastner J. H., Zuckerman B., Weintraub D.A., Forveille T., 1997, Science, 277, 67

  6. [6]

    Lee J., Song I., 2019, , 486, 3434

  7. [7]

    Liu M., 2015, Proceedings IAU Symposium, 314, 2015

  8. [8]

    V., Urban S., 2000, , 317, 298

    Makarov V. V., Urban S., 2000, , 317, 298

Show all 24 references
  1. [9]

    Malo L., Doyon R., Lafreni\` e re D., Artigau, \' E , Gagn\' e J., Baron, F., Riedel A., 2013, , 762, 88

  2. [10]

    E., 2007, in B.G

    Mamajek E. E., 2007, in B.G. Elmegreen, J. Palous, eds, Proc. IAU Symp. 237, Triggered Star Formation in a Turbulent ISM. Cambridge Univ. Press, p. 442

  3. [11]

    et al., 2011, J

    Pedregosa F. et al., 2011, J. Mach. Learn. Res., 12, 2825

  4. [12]

    Riedel et al., 2017, , 153, 95

  5. [13]

    Rodriguez et al., 2011, , 727, 62

  6. [14]

    2010, , 48, 581

    Soderblom D. 2010, , 48, 581

  7. [15]

    Torres C. A. O., da Silva L., Quast G. R., de la Reza R., Jilinski E., 2000, , 120, 1410

  8. [16]

    Torres C. A. O., Quast G. R., de La Reza R., da Silva L., Melo C. H. F., 2003, ASPC, 287, 439

  9. [17]

    R., Melo C

    Torres C.A.O., Quast G. R., Melo C. H. F., Sterzik M., 2008, in Handbook of Star Forming Regions, Volume II: The Southern Sky, ed. B. Reipurth (ASP Monograph Publ., Vol. 5; San Francisco, CA: ASP), 757

  10. [18]

    Viana Almeida P., Santos N.C., Melo C., Ammler-von Eiff M., Torres C.A.O., Quast G.R., Gameiro J.F., Sterzik M., 2009, , 501, 965

  11. [19]

    A., 2000, , 535, 959

    Zuckerman B, Webb R. A., 2000, , 535, 959

  12. [20]

    S., Webb R

    Zuckerman B, Song I., Bessell M. S., Webb R. A., 2001, , 562, 87

  13. [21]

    Zuckerman B., 2001, ASP conference, 244, 122

  14. [22]

    Zuckerman B, Song I., 2004, , 42, 685

  15. [23]

    S., 2011, , 732, 61

    Zuckerman B., Rhee J., Song I., Bessell M. S., 2011, , 732, 61

  16. [24]

    Zuckerman B., 2019, , 870, 27

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.