{"id":"78066008-2c40-4d3a-834c-6b4ee7f2c5d1","arxiv_id":"2411.18609","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"Clustering of Gaia variability features separates variable hot subdwarfs from cataclysmic variables and yields hundreds of new variable-star candidates.","lead":"This paper applies machine-learning clustering to Gaia multi-epoch photometry, sorting 1,576 hot subdwarf candidates into clear variables, dubious variables, and high-amplitude sources that are mostly cataclysmic variables. It reports dozens of new variable hot subdwarf and CV candidates from archival data alone, which could feed follow-up spectroscopy and population studies.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Cluster-2 purity is the load-bearing assumption: 152 CV candidates are labelled solely from amplitude-driven cluster membership, with no contamination test.","rationale":"The reader's weakest assumption correctly identifies cluster-2 purity as the main vulnerability, and my reading of the manuscript agrees: the paper validates cluster 2 only with known CVs and then extends that label to all remaining members, despite the clustering being dominated by amplitude features that are not CV-specific. My concern sharpens the reader's point by noting that (i) 8 of the 13 t-SNE/UMAP mismatched objects straddle the cluster-0/cluster-2 boundary at high amplitude, (ii) 70 of the 152 'candidate CVs' are SIMBAD hot-subdwarf candidates, and (iii) the authors themselves defer confirmation to spectroscopy. These facts make the purity assumption plainly untested rather than merely hypothetical. That said, the paper has independent support: two dimensionality-reduction methods agree on 99% of the cluster assignments, 140 known CVs do fall in cluster 2, and TESS period agreement supports many cluster-0 classifications. The concern does not overturn the existing CONDITIONAL verdict, but it does define the condition: the candidate-CV claims should be withheld or explicitly caveated until a contamination control is run. I therefore recommend no change to the reader's verdict, with the concrete test described above as the natural acceptance criterion.","tokens_in":21517,"tokens_out":3912,"duration_ms":138232,"concrete_test":"Re-run the 27-feature t-SNE/UMAP plus GMM clustering pipeline with a control sample of roughly 50 spectroscopically confirmed high-amplitude non-CV hot subdwarfs (HW Vir systems from Schaffenroth et al. 2022, reflection-effect systems, and BLAPs) added to the 1,576 objects, keeping feature construction identical. If more than about 10% of the injected non-CV controls are assigned to the cluster-2 GMM component, then G-band amplitude alone is insufficient to establish CV candidacy, and the 152 'candidate CVs' should be relabelled as unclassified high-amplitude variables. A complementary check is to query archival LAMOST/SDSS spectra for the 70 SIMBAD hot-subdwarf candidates among the 152; detection of Balmer absorption without CV emission lines would falsify the purity assumption.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.2 labels all 152 non-CV SIMBAD objects in cluster 2 as candidate CVs solely because the 140 spectroscopically confirmed objects in that cluster are CVs. This inference is load-bearing: the abstract's headline claim of 'an additional 152 objects that we consider candidates for new CVs' and the paper's claimed ability to separate CVs from hot subdwarfs both depend on cluster-2 purity. The paper's own feature analysis (Table A.3, Fig. 2) shows that cluster separation is driven almost entirely by G-band amplitude, range, and interquartile range, and §3.3 describes cluster 2 as containing 'high-amplitude ambiguous variables.' Amplitude is not a CV-specific diagnostic: deep-eclipse HW Vir systems, large-amplitude reflection/ellipsoidal systems, and BLAPs can reach comparable G-band amplitudes. The ambiguity is not hypothetical: among the 13 t-SNE/UMAP mismatches, 8 objects with peak-to-peak variations at least 0.5 mag are placed in cluster 2 by t-SNE but in cluster 0 by UMAP, showing that the high-amplitude boundary already produces disagreement about non-CV versus hot-subdwarf assignment. Moreover, 70 of the 152 'candidate CVs' are SIMBAD hot-subdwarf candidates and only 3 are SIMBAD CV candidates, so the cluster assignment directly overrides the available cross-identification. The concluding section itself lists the 152 objects as needing confirmation. No contamination model, spectroscopic check, or out-of-sample validation of cluster-2 purity is provided. Secondary count inconsistencies (156 vs 152 in §5; 89 vs 85 new variables in §5 vs the abstract) reinforce that the candidate lists are not yet reliable.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper applies t-SNE and UMAP dimensionality reduction followed by Gaussian mixture modeling to 1,576 candidate hot subdwarf variables extracted from Gaia DR3 epoch photometry, using 27 features selected via random-forest importance scores. The clustering yields three groups: cluster 0 (290 objects) interpreted as clear variable hot subdwarfs, cluster 1 (990 objects) as dubious or low-S/N objects, and cluster 2 (296 objects) as cataclysmic variables. The authors report 85 new hot subdwarf variables from Gaia and TESS light curves and 108 new variables from Gaia light curves alone, with subtypes including reflection-effect systems, HW Vir systems, ellipsoidal variables, and high-amplitude pulsators. They also label 152 objects in cluster 2 as candidate CVs.","tokens_in":21832,"tokens_out":9217,"duration_ms":71885,"significance":"The basic experimental design is sound and the paper is a useful contribution: it demonstrates that a relatively simple unsupervised clustering pipeline fed with sparse Gaia multi-epoch statistics can separate high-amplitude variables (including known CVs) from lower-amplitude variable hot subdwarfs, and it provides a large list of candidate variables for follow-up. The independent validation is a genuine strength: 140 known CVs fall in cluster 2, and TESS periods agree with Gaia-derived periods for many cluster 0 objects. The 99% agreement between t-SNE and UMAP also supports the stability of the three-cluster solution. If the claims hold, the method will be valuable for building target lists for 4MOST, SDSS-V, and wide-field time-domain surveys. However, the paper's headline count of 152 new CV candidates rests on a purity assumption for cluster 2 that is not demonstrated, and the subtype classifications are based on visual inspection without explicit criteria.","major_comments":[{"comment":"The inference that all 152 remaining members of cluster 2 are CV candidates is not supported by the evidence presented. The sentence \"We considered all of these objects as candidate CVs since all known objects in cluster 2 are CVs without contamination from other classes\" (§4.2) assumes that the known 140 CVs establish cluster purity, but the cluster separation is driven almost entirely by G-band amplitude, range, and interquartile range (Table A.3, Fig. 2), and §3.3 itself describes cluster 2 as containing \"high-amplitude ambiguous variables.\" Amplitude is not a CV-specific diagnostic: deep-eclipse HW Vir systems, large-amplitude reflection/ellipsoidal systems, and BLAPs can reach comparable G-band amplitudes. The paper's own t-SNE/UMAP comparison shows that 8 objects with peak-to-peak variations of at least 0.5 mag are placed in cluster 2 by t-SNE but in cluster 0 by UMAP (§3.3), demonstrating that the high-amplitude boundary is already ambiguous for non-CV objects. Moreover, 70 of the 152 \"candidate CVs\" are SIMBAD hot-subdwarf candidates and only 3 are SIMBAD CV candidates (§4.2). No contamination model, spectroscopic check, or out-of-sample validation of cluster-2 purity is provided; the conclusion itself lists these objects as needing confirmation (§5). I request either a quantitative contamination estimate (e.g., the expected number of high-amplitude non-CV variables in the input catalogue) or a validation subsample with TESS light curves or spectroscopy, before presenting the 152 objects as new CV candidates in the abstract.","section":"§4.1, §4.2, Table 1"},{"comment":"The classification into reflection-effect, HW Vir, ellipsoidal, and pulsating variables is performed by visual inspection of phase-folded light curves, but no explicit criteria are given for these subtypes (e.g., eclipse depth or width thresholds, presence of two maxima, period ratios, or amplitude ratios). Because the abstract's headline results include specific counts of these subtypes among the 85 and 108 new variables, the reliability of these counts depends on a classification scheme that is not described. I ask for a statement of the decision rules and a validation against a known sample, e.g., a confusion matrix showing how many spectroscopically confirmed systems of each type are recovered in cluster 0.","section":"§3.2.1, §3.2.2, §3.3"},{"comment":"The feature selection step uses random-forest importance scores computed from manually assigned labels (\"clear variability\" vs \"ambiguous variability\") to reduce the feature set from 49 to 27, and these same 27 features are then used for the unsupervised t-SNE/GMM clustering. This introduces a supervised component into the clustering pipeline, and the clusters are subsequently interpreted in terms of light-curve clarity and S/N. The circularity is not fatal, but the interpretation of the clusters as purely data-driven should be qualified. A robustness check using the full 49 features (or a purely unsupervised feature-selection method) would help confirm that the three clusters are not an artifact of the label-driven feature ranking; the silhouette scores for 49 vs 27 features are similar, but the label-based ranking is still used in both cases.","section":"§5 vs Abstract"}],"minor_comments":[{"comment":"The number of new hot subdwarf variables from Gaia and TESS is given as 85 in the abstract and Table 1, but as 89 in the conclusion bullet list; please reconcile.","section":"§4.2 and §5"},{"comment":"The number of candidate CVs in cluster 2 is given as 152 in §4.2 and the abstract, but as 156 in the conclusion bullet list; please reconcile. The conclusion also refers to \"153 candidate hot subdwarf\" new variables, which does not match any number in Table 1 or §4.1.2.","section":"§4.1.2"},{"comment":"In the text, \"we found 42 reflection-effect systems, 21 HW Vir systems, 3 pulsating variables, and 2 ellipsoidal variables\" for the 78 unconfirmed hot subdwarfs with TESS; Table 1 lists 40 new reflection candidates, 14 new HW Vir candidates, and 1 new pulsating candidate. The relationship between the text numbers and the table numbers is unclear; please clarify whether the text numbers include previously known variables.","section":"§4.1.1"},{"comment":"The sentence about the 15 systems classified from Gaia alone says \"we found 5 sinusoidal-like light curves, 5 eclipsing binaries, and 2 HW Vir systems,\" which sums to 12, but the preceding sentence says 15 systems. The missing three objects appear as \"Others/Unclear\" in Table 1; the text should state this explicitly.","section":"Fig. 7"},{"comment":"The caption states \"candidate hot subdwarfs (1682) from Culpan et al. (2022) with Gaia light curves,\" but the analysis sample after quality cuts is 1,576 objects. Please clarify what is plotted.","section":"§3.2.1"},{"comment":"The optimization of t-SNE hyperparameters (perplexity = 50, learning rate = 600) is described only by the final values; a brief statement of the search range and the metric used for optimization would aid reproducibility.","section":"§2.1"},{"comment":"The RUWE<7 cutoff is justified by reference to Dawson et al. (2024), but the choice of exactly 7 is presented without a sensitivity test; a short robustness statement (e.g., how the clustering changes for RUWE<3) would strengthen the analysis.","section":"§4.1.3"},{"comment":"The text says \"nine already known pulsating variables\" and then states that TIC 178626010 is a new pulsating variable; this is contradictory. Please rephrase to distinguish previously known pulsators from the newly identified one.","section":"§4.1.4"},{"comment":"For the two high-amplitude pulsating candidates, the text correctly uses the word \"candidate BLAPs,\" but the abstract and conclusion call them \"high-amplitude pulsating variables\" without the candidate qualifier; please ensure the speculative nature is clear in the summary sections.","section":"Table 1"},{"comment":"The Table 1 layout is difficult to parse because rows mix confirmed and candidate hot subdwarfs, and the \"Confirmed Variables\" and \"New Variables\" columns are not repeated for the right-hand block. A clearer layout with explicit subheadings would make the new-variable counts easier to verify.","section":"§4.2"},{"comment":"In the period distribution of known and candidate CVs (Fig. 6), the period for candidate CVs is derived from the same cluster membership that is under scrutiny; if some candidates are not CVs, the period distribution is not meaningful. This should at least be acknowledged in the text.","section":"§5"},{"comment":"The conclusion states that the algorithm \"efficiently identifies CVs without the need for expensive follow-up spectroscopic observations.\" Given the concerns about cluster-2 purity, this claim should be softened to refer to candidate selection pending confirmation.","section":"Conclusion"}],"recommendation":"major_revision","confidential_remarks":"The paper comes from a strong group and the core demonstration—that unsupervised clustering on Gaia statistics separates high-amplitude variables from lower-amplitude hot subdwarf variables, with known CVs validating the high-amplitude cluster—is likely correct. The main risk is over-claiming the CV candidate list; the authors should either provide a contamination estimate or re-label these as 'candidate variables of unknown nature' pending follow-up. The numerical inconsistencies between the abstract, conclusion, and Table 1 suggest the manuscript needs a careful consistency pass before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read of Ranaivomanana et al. (arXiv:2411.18609). The paper does something genuinely useful: it applies t-SNE/UMAP plus GMM clustering to Gaia DR3 multi-epoch photometry of 1,576 candidate hot subdwarfs, and produces candidate lists that roughly double the known sample of variable hot subdwarfs and add ~150 CV candidates. The cluster-0 validation is solid: 140 known CVs land in cluster 2, and TESS periods agree with Gaia for many cluster-0 objects. The two dimensionality-reduction methods agree on 99% of objects. The bespoke features for sparse Gaia sampling (e.g., the 95th percentile of the top 100 peaks) are sensible and clearly described. This is a legitimate application of existing ML tools to a new dataset, not a paradigm shift.\n\nThe soft spot is exactly where the stress-test note lands: cluster-2 purity. The paper labels all 152 non-CV objects in cluster 2 as candidate CVs solely because the 140 known objects in that cluster are CVs. But the clustering is dominated by G-band amplitude, and the paper itself calls cluster 2 'high-amplitude ambiguous variables.' Amplitude is not CV-specific; deep-eclipse HW Vir systems, large reflection/ellipsoidal systems, and BLAPs can all reach those amplitudes. The 13 t-SNE/UMAP mismatches include 8 high-amplitude objects that t-SNE puts in cluster 2 and UMAP puts in cluster 0; that is direct evidence that the high-amplitude boundary is not clean. And 70 of the 152 'candidate CVs' are SIMBAD hot-subdwarf candidates, while only 3 are CV candidates. The paper overrides those cross-identifications on cluster membership alone, with no contamination model or spectroscopic check. The conclusion does list these objects as needing confirmation, which is honest, but the abstract's headline claim treats them as CV candidates without the same caveat.\n\nThe minor issues are real but fixable: the conclusion says 156 CV candidates while the abstract says 152, and the new-variable counts disagree (89 vs 85 in the text vs the abstract). That sloppiness does not undermine the main result but needs cleaning.\n\nThe physical subclassifications (reflection, HW Vir, ellipsoidal) are by visual inspection, not by the ML method; that is fine for a catalogue paper but should be labeled as such.\n\nWho is this for? The hot-subdwarf and CV communities will use these lists, and the method may transfer to other sparse time-domain surveys. The central claim about cluster-2 purity is not yet established, so the paper needs a contamination test (e.g., check the 152 candidates against ZTF or ASAS light curves, or spectroscopically follow a subset) and a revision that softens the CV-candidate claim.\n\nMy recommendation: send it to peer review. It deserves referee time despite the load-bearing assumption, because the candidate lists are valuable and the method is transparent. I would ask for major revision.\n\nWould I cite it? Probably not until the CV candidates are vetted, but I would use it for the hot-subdwarf variable lists.","headline":"Useful candidate lists and solid cluster-0 validation, but the CV candidacy rests on an untested purity assumption; referee it.","tokens_in":22438,"tokens_out":2792,"would_cite":false,"duration_ms":31559,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that clustering photometric statistics from Gaia DR3 separates variable hot subdwarfs from cataclysmic variables, yielding 85 new TESS-confirmed variables, 108 Gaia-only variables, and 152 candidate CVs.","keywords":["hot subdwarf stars","variable stars","cataclysmic variables","Gaia DR3","machine learning","t-SNE","UMAP","photometric variability"],"falsifier":"Spectroscopically observe, or check existing spectra of, the 152 objects classified as candidate cataclysmic variables in cluster 2; if a significant fraction show hot-subdwarf or other non-CV signatures such as Balmer absorption typical of sdB stars rather than CV emission lines, the claim that cluster membership alone identifies CVs is falsified. A cheaper test is to run known high-amplitude non-CV variables, such as RR Lyrae stars or large-amplitude ellipsoidal binaries, through the same feature pipeline and see whether they land in cluster 2.","tokens_in":21311,"feed_emoji":"⭐","tokens_out":6073,"duration_ms":51470,"temperature":0.7,"pith_summary":"This paper claims that unsupervised machine-learning clustering of statistical features extracted from sparse Gaia multi-epoch photometry can separate genuinely variable hot subdwarf stars from noisy light curves and from cataclysmic variables, without needing spectroscopy. Applied to 1,576 candidate hot subdwarfs, the method reports 108 newly identified variables from Gaia alone and 85 confirmed with TESS, mostly reflection-effect and HW Vir binaries, plus two new high-amplitude pulsators consistent with blue large amplitude pulsators. The same cluster that contains 140 known cataclysmic variables holds 152 objects the paper labels as candidate CVs. The result matters because it offers a generalisable, photometry-only route to building large samples of variable subdwarfs and CVs for binary-evolution and asteroseismic studies.","feed_headline":"Clustering exposes 108 new variable hot subdwarfs","feed_subtitle":"Gaia light curves alone yield 108 new variables, plus 152 new cataclysmic variable candidates.","key_machinery":"The machinery is an unsupervised clustering pipeline: a hybrid Ψ-periodogram (generalised Lomb-Scargle plus Lafler-Kinman statistics) for dominant-frequency searches with Monte Carlo uncertainties; 84 features drawn from the Gaia DR3 variability statistics table, custom periodogram-peak percentiles, and Gaia source parameters; random-forest feature ranking and Pearson-correlation pruning to 27 features; t-SNE and UMAP embedding into a 2D feature space; Gaussian mixture model clustering with silhouette-score validation; and independent TESS light-curve comparison to confirm periods and refine classifications. The G-band amplitude is the single most important feature, and the paper places a lower bound of roughly 20 millimagnitudes on clear variability detectable in this sample.","core_discovery":"On its own terms, the paper's central finding is that a feature space built from Gaia DR3 epoch-photometry statistics and custom periodogram-peak statistics, reduced to 2D with t-SNE and UMAP and clustered with a Gaussian mixture model, yields three sharply separated groups: cluster 0 (290 objects) with clear, high signal-to-noise variability; cluster 1 (990 objects) with dubious low signal-to-noise variability; and cluster 2 (296 objects) with high-amplitude, ambiguous variability. Because all 140 previously known objects in cluster 2 are cataclysmic variables, the paper treats the remaining 152 members as candidate CVs. Within cluster 0 it classifies 78 known and 212 candidate hot subdwarfs; using TESS light curves for consistency it reports 85 new variables from Gaia and TESS and 108 new variables from Gaia alone, adding reflection-effect systems, HW Vir binaries, ellipsoidal variables, and two candidate blue large amplitude pulsators.","pith_inferences":["Because the G-band amplitude dominates the clustering, cluster 2 may be capturing an amplitude regime as much as a physical class; the CV purity of the 152 candidates is the least secured part of the pipeline and needs external validation.","The same pipeline could probably be used to search for high-amplitude pulsating subdwarfs and other blue variables in surveys without TESS coverage, but the absence of TESS confirmation for the 108 Gaia-only variables leaves their classifications provisional.","If the candidate CVs are spectroscopically confirmed, they would substantially increase the known CV population in this colour-magnitude region, changing estimates of CV space density and of contamination in hot-subdwarf selections.","The roughly 20-millimagnitude amplitude floor for clear variability is an implicit selection effect: low-amplitude sdB pulsators will be missed, so the method complements rather than replaces dedicated asteroseismic searches."],"forward_implications":["The method can be applied as-is to Gaia DR4 epoch photometry, which the paper estimates would extend the analysis to the remaining roughly 59,000 candidate hot subdwarfs without retraining.","The newly identified reflection-effect, HW Vir, and ellipsoidal systems enlarge the known sample of variable hot subdwarfs, supporting population-level binary-evolution and asteroseismic studies.","The two new high-amplitude pulsators, if confirmed as blue large amplitude pulsators, add to a rare class of hot pulsators that can be found in Gaia data alone.","The feature set and clustering approach are survey-agnostic and transfer to other time-domain surveys such as BlackGEM, ZTF, GOTO, and LSST, as the paper argues.","The cluster-2 result adds 152 cataclysmic-variable candidates and 56 new candidate orbital periods for known CVs, expanding the sample available for CV population studies."],"supporting_citations":[{"why":"Supplies the 61,585-candidate hot-subdwarf catalogue whose Gaia epoch-photometry subset is the input sample for this work.","marker":"Culpan et al. 2022"},{"why":"Provides the Gaia DR3 variability statistics table and epoch-photometry products that supply most of the clustering features.","marker":"Eyer et al. 2023"},{"why":"Defines the reflection-effect and HW Vir classification references and period trends used to interpret the cluster 0 variables.","marker":"Schaffenroth et al. 2022"},{"why":"One of the catalogues of known variable hot subdwarfs and cataclysmic variables used to validate the cluster assignments.","marker":"Barlow et al. 2022"},{"why":"Additional catalogue of known cataclysmic variables used to establish that every known object in cluster 2 is a CV.","marker":"Hou et al. 2023"},{"why":"Provides the known CV sample and its periods; 127 of the 140 confirmed CVs in cluster 2 cross-match to this catalogue.","marker":"Canbay et al. 2023"},{"why":"Supplies the known TESS pulsating hot subdwarfs used for period comparison and for distinguishing p-mode from g-mode pulsators.","marker":"Krzesinski & Balona 2022"},{"why":"Provides the t-SNE dimensionality-reduction algorithm that produces the primary 2D embedding used for clustering.","marker":"van der Maaten & Hinton 2008"},{"why":"Provides UMAP, the second dimensionality-reduction algorithm whose matching clusters corroborate the t-SNE result.","marker":"McInnes et al. 2018"}],"fun_headline_variants":["Machine learning finds 108 new variable hot subdwarfs","Gaia DR3 machine learning reveals 108 new variable subdwarfs","Clustering Gaia photometry uncovers 108 new hot subdwarf variables","Machine learning exposes 152 new cataclysmic variable candidates","t-SNE and UMAP identify 108 new variable hot subdwarfs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Cluster 2 is labelled as cataclysmic variables purely from cluster membership: every known object in it happens to be a CV, and the 152 remaining members are assigned CV status with no spectroscopic confirmation, while the clustering is dominated by G-band amplitude so high-amplitude non-CV variables could in principle fall in the same group.","fun_headline_variants_meta":{"raw":{"variants":["Machine learning finds 108 new variable hot subdwarfs","Gaia DR3 machine learning reveals 108 new variable subdwarfs","Clustering Gaia photometry uncovers 108 new hot subdwarf variables","Machine learning exposes 152 new cataclysmic variable candidates","t-SNE and UMAP identify 108 new variable hot subdwarfs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000825,"raw_usage":{"total_tokens":3668,"prompt_tokens":1070,"completion_tokens":2598,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":686,"completion_tokens_details":{"reasoning_tokens":2502}},"tokens_in":686,"tokens_out":2598,"duration_ms":17493,"temperature":1.0,"reasoning_tokens":2502,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:00:06.902424+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Spectroscopically observe, or check existing spectra of, the 152 objects classified as candidate cataclysmic variables in cluster 2; if a significant fraction show hot-subdwarf or other non-CV signatures such as Balmer absorption typical of sdB stars rather than CV emission lines, the claim that cluster membership alone identifies CVs is falsified. A cheaper test is to run known high-amplitude non-CV variables, such as RR Lyrae stars or large-amplitude ellipsoidal binaries, through the same feature pipeline and see whether they land in cluster 2.","supporting_citations":[{"cited_title":"2022, , 662, A40","cited_arxiv_id":null,"evidence_quote":"Supplies the 61,585-candidate hot-subdwarf catalogue whose Gaia epoch-photometry subset is the input sample for this work."},{"cited_title":"2023, , 674, A13","cited_arxiv_id":null,"evidence_quote":"Provides the Gaia DR3 variability statistics table and epoch-photometry products that supply most of the clustering features."},{"cited_title":"N., Geier , S., & Kupfer , T","cited_arxiv_id":null,"evidence_quote":"Defines the reflection-effect and HW Vir classification references and period trends used to interpret the cluster 0 variables."},{"cited_title":"L., Dong , Y.-Q., Chen , X.-L., & Bai , Z.-R","cited_arxiv_id":null,"evidence_quote":"Additional catalogue of known cataclysmic variables used to establish that every known object in cluster 2 is a CV."},{"cited_title":"& Balona , L","cited_arxiv_id":null,"evidence_quote":"Supplies the known TESS pulsating hot subdwarfs used for period comparison and for distinguishing p-mode from g-mode pulsators."},{"cited_title":"& Hinton, G","cited_arxiv_id":null,"evidence_quote":"Provides the t-SNE dimensionality-reduction algorithm that produces the primary 2D embedding used for clustering."}],"review_version":1}