Revised classification of the CHIME fast radio bursts with machine learning

Hai-Nan Lin; Liang Liu; Li Tang

REVIEW 4 major objections 5 minor 47 references

Revised classification of the CHIME fast radio bursts with machine learning

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read The CHIME catalog splits naturally into two FRB groups; a straight line in rest-frame frequency width versus peak frequency separates repeaters from non-repeaters, implying that more than half of FRB sources may be intrinsic repeaters.

desk verdict Useful two-feature rule for FRB classification, but the 'natural two-cluster' claim is compromised by label-informed tuning and in-sample evaluation. read the letter →

arxiv 2509.02645 v1 pith:E3FM2J6H submitted 2025-09-02 astro-ph.IM astro-ph.HE

Liang Liu , Hai-Nan Lin , Li Tang This is my paper · ORCID

classification astro-ph.IMastro-ph.HE

keywords fastradioburstsCHIME/FRBcatalogunsupervisedmachinelearningt-SNEHDBSCANrepeaterclassificationrest-framefrequencywidthpeak

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 tries to establish that the observed split between repeating and non-repeating fast radio bursts reflects two genuinely distinct populations, and that the labels have been systematically misassigned. Re-analyzing 505 CHIME sources with one burst per source and letting an unsupervised algorithm — t-SNE dimensionality reduction followed by HDBSCAN clustering — find the structure, the authors recover two natural clusters: one holds nearly all known repeaters plus 230 sources previously labeled non-repeaters. They argue this implies that more than half of FRB sources may be intrinsic repeaters whose repetition has simply not been caught yet. The novel, testable result is that repeaters and non-repeaters are linearly separable using just two rest-frame spectral features, peak frequency and frequency width, with the boundary Δν = 0.95νp + 1.30; this two-variable rule alone reclassifies known repeaters with 95% recall. If correct, classifying an FRB becomes a quick spectral measurement instead of a long monitoring campaign.

What carries the argument

The load-bearing object is the ordered pair (νp, Δν) — peak frequency and rest-frame frequency width, defined as Δν = (νmax − νmin)(1+z) — which mutual-information analysis identifies as the two features that dominate the cluster structure. Around this pair the machinery is a three-stage pipeline: t-SNE projects the initially ten-dimensional feature space onto a two-dimensional embedding; HDBSCAN then finds two density-based clusters without a preset cluster count; and a support vector machine draws the maximum-margin straight line Δν = 0.95νp + 1.30 that separates the clusters in the original physical plane. The force of the pipeline is that cutting the input from ten features to just these

What would settle it

Re-run the two-feature clustering and the SVM boundary using only FRBs with direct host-galaxy redshifts (two in this sample; about twenty localized FRBs overall). If the line Δν = 0.95νp + 1.30 no longer separates repeaters from non-repeaters when real redshifts replace DM-inferred ones, the separation is an artifact of the DM budget. A complementary check: monitor a sample of the 230 candidate repeaters with exposure comparable to the first CHIME year; repetition rates no higher than the confirmed non-repeaters would refute the intrinsic-repeater claim.

Watch

Extended reading notes

Core claim

The central claim is that the FRB population is dichotomous at the intrinsic level and that the two classes are already visible in two radio features. Working from the first CHIME/FRB catalog plus the 2023 repeater catalog, with exactly one burst kept per source, the authors apply t-SNE and HDBSCAN and recover two stable clusters across ten-, six-, and two-feature inputs. The repeater cluster contains 41 confirmed repeaters and 230 candidate repeaters, from which the authors conclude that more than half of all FRB sources may be intrinsic repeaters; five of the six sources confirmed as repeaters after the first catalog fell naturally in this cluster. Mutual-information analysis singles out r

Load-bearing premise

The analysis assumes the redshifts inferred from dispersion measure — using fixed host-galaxy and Milky-Way halo corrections (DM_host = 70 and DM_halo = 30 pc cm^-3) and a fixed intergalactic baryon fraction — are accurate enough that the rest-frame frequency widths of the two classes are not systematically pushed apart by the inference itself.

Editorial extensions

If this is right

New CHIME bursts can be assigned to a repeater-like or non-repeater-like class from two measured quantities — peak frequency and rest-frame frequency width — with 95% recall on known repeaters, removing the need to wait for a second burst.
The 230 sources that the clusters place with repeaters become concrete targets for monitoring; confirmations would raise the known repeater fraction far above the currently confirmed count.
If more than half of FRB sources are intrinsic repeaters, the counts used to constrain progenitor models and the cosmic FRB rate must be reweighted, since the confirmed repeater fraction was an observational lower bound.
Contrary to earlier unsupervised analyses that reported several FRB subgroups, this source-level setup recovers exactly two major clusters, supporting a dichotomous population.
The two misclassified repeaters — one borderline (FRB20181017A) and one extreme (FRB20180910A) — mark the rule's failure modes and indicate where a distinct emission channel might hide.

Reading between the lines

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

A direct test of the 'more than half' claim would monitor a few dozen of the 230 candidate repeaters with exposure comparable to CHIME's first year; reproduction rates no higher than the confirmed non-repeater class would refute the candidate label.
Because Δν is built from (1+z) and z comes from dispersion measure for 503 of 505 sources, adopting other host-galaxy or halo DM values would shift the two classes differentially; re-deriving the SVM line across a range of DM budgets would show how much of the separation is scaffolded by the assumed DM budget.
The boundary is expressed in rest-frame coordinates, so it predicts that narrow-band emission marks repeating behavior at any observing frequency; checking the same Δν–νp plane with bursts detected at other radio bands would test the rule's universality.

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.

Desk Editor's Note

read the letter

Here is my take. The paper's most practical contribution is a simple boundary in rest-frame peak frequency vs width: Delta-nu = 0.95 nu_p + 1.30, with repeaters having systematically narrower widths. That is genuinely useful for follow-up targeting, and the decision to use one burst per source is a real improvement over treating all bursts as independent. The mutual information analysis showing that these two features carry most of the signal is clean and worth knowing. The recall on known repeaters is also decent, which is what you want when you are trying not to miss repeaters. The problem is that the central claim about a natural, label-free two-cluster structure does not hold up. Section 2.3 says the HDBSCAN hyperparameters were chosen by grid search to maximize classification performance. The only meaningful classification target here is repeater vs non-repeater. So the clustering is not independent of the labels: you have tuned the cluster definition using the very labels you say you are not relying on. And with min_cluster_size=150 out of 505 sources, any cluster containing most of the 43 repeaters is necessarily going to include at least 107 non-repeaters. So the '230 candidate repeaters' is partly an artifact of the parameter choice, not an unforced discovery. The SVM boundary is also fitted and evaluated on the same 505 points, so the 0.95 recall is in-sample. Given that the non-repeater class is ten times larger, labeling 230 non-repeaters as repeaters means precision is around 15%: acceptable for prioritization, but not evidence that more than half of FRBs are intrinsic repeaters. There are also two narrower issues worth noting. One, the rest-frame width depends on DM-derived redshifts; the fixed DM_halo=30, DM_host=70, f_IGM=0.83 are plausible but no sensitivity analysis is given, and systematic errors in (1+z) could shift repeaters and non-repeaters differently. Two, t-SNE is stochastic and only one random seed is used, so stability is not demonstrated. Bottom line: I would send this to a careful referee, but only with the expectation that they demand label-blind hyperparameter selection, cross-validation, precision, and DM-systematics checks. Without those, treat it as hypothesis-generating. The simple Delta-nu vs nu_p cut may still survive as an empirical rule, but the 'most FRBs are repeaters' conclusion is not yet supported.

Referee Report

4 major / 5 minor

Summary. The paper applies t-SNE dimensionality reduction and HDBSCAN clustering to 505 CHIME FRB sources (one burst per source) using ten observed/derived features, aiming to find intrinsic FRB groupings without using the repeater/non-repeater labels. It reports that the data split into two principal clusters, one containing 41 of 43 known repeaters plus 230 sources labeled as non-repeaters. Mutual-information analysis identifies rest-frame frequency width Δν and peak frequency νp as the most informative features, and the paper claims the same two-cluster separation persists when only these two features are used, with an SVM decision boundary Δν = 0.95νp + 1.30 and repeater recall 0.95.

Significance. If the central claim were established, it would be an important result: it would imply that a large fraction of FRB sources are intrinsically repeating, that the repeater/non-repeater dichotomy is primarily an observational artifact, and that a simple two-feature linear boundary could classify FRBs without expensive follow-up. The paper also has useful methodological elements: it deliberately keeps only one burst per source to reduce autocorrelation bias, makes its feature definitions explicit, and provides an interpretable boundary. However, these strengths cannot compensate for the two load-bearing weaknesses identified below: the clustering is not label-free as claimed, and the reported classification performance is entirely in-sample. The population-inference claims are therefore not supported by the analysis as presented.

major comments (4)

[§2.3, Table 2; §3.1] The paper states that HDBSCAN hyperparameters were chosen by 'a grid search approach to maximize classification performance.' Since the only classification target in this paper is repeater vs. non-repeater, this means the cluster definition was selected using the very labels the analysis claims to avoid. Choosing min_cluster_size=150 and min_samples=2 to maximize repeater/non-repeater separation effectively forces the 43 repeaters into one large cluster with many non-repeaters, so the finding of a 'repeater cluster' containing 230 candidate repeaters is not an unsupervised discovery. A label-blind analysis is required: for example, select hyperparameters by cluster stability or silhouette score on the unlabeled features, and show the cluster structure across a grid of min_cluster_size/min_samples rather than only the configuration that maximizes label-based recall.
[§3.3, Fig. 7] The SVM decision boundary Δν = 0.95νp + 1.30 and the reported recall of 0.95 are computed on the same 505 sources used to fit the boundary and to derive the HDBSCAN clusters. There is no train/test split, cross-validation, or bootstrap, so the recall is in-sample and cannot establish predictive linear separability. The paper needs out-of-sample evaluation, such as k-fold cross-validation or a held-out set, with confidence intervals on the boundary parameters and a confusion matrix for repeaters vs. non-repeaters. Without this, the claim that two features 'successfully recover the majority of known repeaters' is not a test of generalization.
[§2.1, Eqs. (1), (6), (7)] The rest-frame frequency width Δν = (νmax − νmin)(1+z) is one of the two features driving the clustering, and z is inferred from DM using fixed DM_halo=30 pc cm−3, DM_host=70 pc cm−3, and f_IGM=0.83 for 503 of 505 sources. Repeaters and non-repeaters have different DM distributions, so any systematic error in these assumed DM components will shift Δν differently for the two classes and could manufacture or exaggerate the separation in the Δν–νp plane. The authors should test robustness by varying DM_host, DM_halo, and f_IGM over plausible ranges and re-running the clustering and SVM analysis, and should also check the two sources with spectroscopic redshifts as a sanity check.
[§3.2, Figs. 3 and 5] The mutual-information analysis is presented as identifying νp and Δν as the features 'governing' the clustering, but the MI scores are computed against a single t-SNE embedding obtained with random_state=22. t-SNE is stochastic and the embedding can vary substantially across random seeds; without repeating the t-SNE/HDBSCAN pipeline over multiple random states and reporting the spread in MI values and cluster assignments, the robustness of the feature-importance claim is not established.

minor comments (5)

[Throughout (figures and equations)] Many figure labels and parts of equations appear as corrupted character sequences (e.g., 'uni00000017' strings in Fig. 1 and the axis labels in Figs. 2–7), making the figures unreadable in the submitted version. These need to be regenerated in a clean format.
[§2.3] The grid search is described only as 'maximize classification performance'; the exact objective, search ranges, and the fact that it used the repeater/non-repeater labels should be stated explicitly. This is essential for reproducibility and for assessing the claim of unsupervised analysis.
[§3.3] The statement that 'reducing the number of features from ten to two does not reduce recall' is based on in-sample comparison; it should be rephrased or replaced by a cross-validated comparison. Also, precision is discussed but never reported; if only recall is used for hyperparameter selection, this should be justified more carefully.
[§2.2, Table 1] The t-SNE perplexity is set to sqrt(505) ≈ 22.47 but reported as 22; the choice of random_state=22 is not justified, and the stability of the embedding with respect to perplexity is not explored.
[References] Reference [28] appears to duplicate reference [8] in content; also the in-text citation of the CHIME/FRB 2023 catalog should include the catalog's official designation to match the reference list.

Circularity Check

2 steps flagged · score 7.0 of 10

The central 'unsupervised' two-cluster discovery is label-informed: HDBSCAN hyperparameters are tuned by grid search to maximize classification performance against the very repeater/non-repeater labels the analysis claims to avoid, and the reported SVM recall is in-sample.

fitted input called prediction [Sec. 2.3 (HDBSCAN hyperparameters) and Sec. 3.1 (interpretation of the repeater cluster)]
"To determine the optimal HDBSCAN hyperparameters, we employ a grid search approach to maximize classification performance. ... The 'repeater cluster' contains 41 confirmed repeaters and 230 candidate repeaters, which implies that more than half FRB sources may be intrinsic repeaters."

The paper advertises unsupervised clustering 'without relying on preassigned labels', but the HDBSCAN hyperparameters (min_cluster_size=150, min_samples=2) are explicitly selected by a grid search that maximizes classification performance—i.e., performance against the known repeater/non-repeater labels. The resulting cluster partition is therefore not label-free; it is the partition that best reproduces the conventional dichotomy. Counting the 230 non-repeaters that fall into the repeater-rich cluster and presenting them as 'candidate repeaters' is then a consequence of label-informed tuning, not an independent unsupervised discovery. The inference that more than half of FRB sources may be intrinsic repeaters reduces to the fact that the clustering was optimized to put repeaters together w
fitted input called prediction [Sec. 3.2 (SVM boundary) and Sec. 3.3 (recall evaluation)]
"The support vector machine (SVM) algorithm was implemented to construct the maximum-margin hyperplane that optimally separates the two classes through margin maximization. The resulting decision boundary ... is given by Δν = 0.95νp + 1.30. ... The recall value is therefore 41/(41 + 2) = 0.95, indicating that even with only two input features, the model successfully recovers the majority of known repeaters."

The SVM is fit on all 505 sources and the reported recall is computed on the same training set; no train/test split or external validation is described. The 'linear separability' of repeaters and non-repeaters is asserted from a boundary that was chosen to separate them, and the recall of 0.95 is an in-sample training metric, not an independent prediction. This presents a fitted classifier as evidence of predictive separability, overstating the strength of the claim.

full rationale

The paper's central claim—that unsupervised ML 'without relying on preassigned labels' reveals two natural FRB populations and that the repeater cluster contains 230 candidate repeaters—is substantially undermined by the explicit statement in Sec. 2.3 that HDBSCAN hyperparameters were chosen by a grid search maximizing classification performance. Since the only meaningful classification target in this context is repeater vs. non-repeater, the cluster geometry is label-informed, and the subsequent counting of non-repeaters inside the repeater cluster as 'candidate repeaters' is a loop: labels are used to tune the clustering, and the tuned clustering is then used to assign new labels. The SVM boundary and recall are also in-sample, adding a fitted-input-called-prediction issue. No self-citation chain, uniqueness import, or ansatz-smuggling is present; the DM-based redshift inference is a physically motivated approximation and a potential systematic, but not a circular step. Overall, the claimed unsupervised discovery is partially constructed by the label-tuned hyperparameters, warranting a score of 7.

Assumptions & free parameters 9 free parameters · 5 assumptions · 0 invented entities

The analysis relies on a chain of modeling choices: DM-based redshift inference with fixed host/halo terms, t-SNE embedding as a faithful representation, and HDBSCAN parameters that were tuned using the very labels the clustering is supposed to discover. The fitted SVM boundary is an additional fitted quantity. No new physical entities are introduced.

free parameters (9)

HDBSCAN min_cluster_size = 150
Chosen by grid search 'to maximize classification performance' using known repeater/non-repeater labels; controls cluster granularity and effectively forces a small number of large clusters.
HDBSCAN min_samples = 2
Selected in the same grid search; affects density estimation.
t-SNE perplexity = 22
Set to sqrt(N)=22 by heuristic (Oskolkov 2019); influences local vs. global structure of embedding.
t-SNE random_state = 22
Arbitrary seed; t-SNE is stochastic, and no stability check across seeds is reported.
DM_host = 70 pc cm^-3
Fixed host-galaxy DM contribution adopted from Zhu-Ge et al.; enters every DM-based redshift estimate and hence all rest-frame features.
DM_halo = 30 pc cm^-3
Fixed Milky Way halo DM contribution adopted from Zhu-Ge et al.; same load-bearing role.
f_IGM = 0.83
Baryon fraction in the IGM taken from Fukugita et al.; directly scales the DM-z integral.
SVM decision boundary slope/intercept = 0.95 / 1.30
Fitted maximum-margin line in the Δν-νp plane using the labeled sample; reported without error bars or validation.
z_min cutoff = 0.002248
Minimum redshift assigned to sources with formally zero/negative inferred redshift; arbitrary, affects low-DM sources (mostly repeaters).

assumptions (5)

domain assumption DM decomposition DM = DM_MW + DM_halo + DM_IGM + DM_host/(1+z) with fixed halo and host terms
Invoked in Section 2.1 Eq. (6) to convert DM into redshift for every unlocalized FRB; host DM is known to vary source-to-source.
domain assumption Macquart-style DM-z relation with f_IGM=0.83 and Planck 2018 cosmology
Section 2.1 Eq. (7); the accuracy of inferred redshifts depends on this relation and on the assumed baryon fraction.
domain assumption The t-SNE 2D embedding preserves the intrinsic population structure
Section 2.2; all clustering and MI analysis is done on the embedding, not the original space.
domain assumption The first detected burst per source is representative of that source
Section 2.1; this selection is intended to remove burst-to-burst correlations, but the first burst may be biased by detection thresholds.
ad hoc to paper HDBSCAN with grid-searched parameters recovers 'natural' groupings
Section 2.3; because parameters were chosen using labels, the grouping is not independent of the conventional repeater/non-repeater classification.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Revised classification of the CHIME fast radio bursts with machine learning." pith.science (2026). https://pith.science/paper/E3FM2J6H

@misc{pith2026250902645,
  author       = {Pith},
  title        = {Pith review of: Revised classification of the CHIME fast radio bursts with machine learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/E3FM2J6H}},
  note         = {Machine review of arXiv:2509.02645}
}

read the original abstract

Fast radio bursts (FRBs) are short-duration and energetic radio transients of unknown origin. Observationally, they are commonly categorized into repeaters and non-repeaters. However, this binary classification may be influenced by observational limitations such as sensitivity and time coverage of telescopes. In this work, we employ unsupervised machine learning techniques to re-examine the CHIME/FRB catalog, with the goal of identifying intrinsic groupings in the FRB population without relying on preassigned labels. Using t-distributed stochastic neighbor embedding (t-SNE) for dimensionality reduction and hierarchical density-based spatial clustering of applications with noise (HDBSCAN) for clustering, we find that the FRB sample separates naturally into two major clusters. One cluster contains nearly all known repeaters but is contaminated by some apparently non-repeaters, while the other cluster is dominated by non-repeaters. This suggests that certain FRBs previously labeled as non-repeaters may share intrinsic similarities with repeaters. The mutual information analysis reveals that rest-frame frequency width and peak frequency are the most informative features governing the clustering structure. Even when reducing the input space to just these two features, the classification remains robust.

Figures

Figures reproduced from arXiv: 2509.02645 by the authors.

**Figure 2.** Dimensionality reduction and clustering results with ten features of FRBs. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

**Figure 3.** Feature correlation of t-SNE + HDBSCAN with ten features. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Dimensionality reduction and clustering results with six features of FRBs. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

**Figure 5.** Figure 5: Feature correlation of t-SNE + HDBSCAN with six features. [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

**Figure 6.** Figure 6: Dimensionality reduction and clustering results with two features of FRBs. [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]

**Figure 7.** Figure 7: The 2D plane of peak frequency νp and rest-frame frequency width ∆ν. bias against true but observationally incomplete repeaters. Recall thus offers a more meaningful measure of classifier effectiveness for FRB population studies. The recall is defined as the ratio of t…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

47 extracted references · 46 canonical work pages

[1]

, " * write output.state after.block = add.period write newline

ENTRY address archive author booktitle chapter collaboration edition editor eid eprint howpublished institution journal key month note number numpages organization pages publisher school series title type url volume year label INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.all := #1 'mid.sent...

work page
[2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

work page
[3]

D. R. Lorimer, M. Bailes, M. A. McLaughlin et al. Science, 318: 777 (2007)

work page 2007
[4]

Xu et al

J. Xu et al. Universe, 9 (7): 330 (2023)

work page 2023
[5]

E. F. Keane et al. Nature, 530: 453--456 (2016)

work page 2016
[6]

Chatterjee et al

S. Chatterjee et al. Nature, 541: 58 (2017)

work page 2017
[7]

Scholz et al

P. Scholz et al. Astrophys. J., 833 (2): 177 (2016)

work page 2016
[8]

L. G. Spitler et al. Nature, 531: 202 (2016)

work page 2016

Show all 47 references

[9]

Platts, A

E. Platts, A. Weltman, A. Walters et al. Phys. Rept., 821: 1--27 (2019)

2019
[10]

B. Zhang. Nature, 587: 45--53 (2020)

2020
[11]

Li et al

D. Li et al. Nature, 598 (7880): 267--271 (2021)

2021
[12]

Xu et al

H. Xu et al. Nature, 609 (7928): 685--688 (2022)

2022
[13]

C. H. Niu et al. Nature, 606 (7916): 873--877 (2022). [Erratum: Nature 611, E10 (2022)]

2022
[14]

B. H. Chen, T. Hashimoto, T. Goto et al. Mon. Not. Roy. Astron. Soc., 509 (1): 1227--1236 (2021)

2021
[15]

X. Yang, S. B. Zhang, J. S. Wang et al. Mon. Not. Roy. Astron. Soc., 522 (3): 4342--4351 (2023)

2023
[16]

Sun, J.-G

W.-P. Sun, J.-G. Zhang, Y. Li et al. Astrophys. J., 980 (2): 185 (2025)

2025
[17]

Luo, J.-M

J.-W. Luo, J.-M. Zhu-Ge, and B. Zhang. Mon. Not. Roy. Astron. Soc., 518 (2): 1629--1641 (2022)

2022
[18]

Zhu-Ge, J.-W

J.-M. Zhu-Ge, J.-W. Luo, and B. Zhang. Mon. Not. Roy. Astron. Soc., 519 (2): 1823--1836 (2022)

2022
[19]

B. C. Andersen et al. (CHIME/FRB Collaboration). Astrophys. J. Lett., 885 (1): L24 (2019)

2019
[20]

Fonseca et al

E. Fonseca et al. Astrophys. J. Lett., 891 (1): L6 (2020)

2020
[21]

Amiri et al

M. Amiri et al. (CHIME/FRB Collaboration). Astrophys. J. Supp., 257 (2): 59 (2021)

2021
[22]

Zhong, W.-J

S.-Q. Zhong, W.-J. Xie, C.-M. Deng et al. Astrophys. J., 926 (2): 206 (2022)

2022
[23]

B. J. R. Raquel, T. Hashimoto, T. Goto et al. Mon. Not. Roy. Astron. Soc., 524 (2): 1668--1691 (2023)

2023
[24]

C. R. Garc\' a, D. F. Torres, J.-M. Zhu-Ge et al. Astrophys. J., 977 (2): 273 (2024)

2024
[25]

Qiang, J

D.-C. Qiang, J. Zheng, Z.-Q. You et al. Astrophys. J., 982 (1): 16 (2025)

2025
[26]

Luo, J.-R

J.-W. Luo, J.-R. Niu, W.-Y. Wang et al. The Astrophysical Journal, 988 (1): 62 (2025)

2025
[27]

B. C. Andersen et al. (CHIME/FRB Collaboration). Astrophys. J., 947 (2): 83 (2023)

2023
[28]

S. P. Tendulkar et al. Astrophys. J. Lett., 834 (2): L7 (2017)

2017
[29]

Marcote et al

B. Marcote et al. Nature, 577 (7789): 190--194 (2020)

2020
[30]

Luo, J.-M

J.-W. Luo, J.-M. Zhu-Ge, and B. Zhang. Monthly Notices of the Royal Astronomical Society, 518 (2): 1629--1641 (2022)

2022
[31]

J. P. Macquart et al. Nature, 581 (7809): 391--395 (2020)

2020
[32]

Deng and B

W. Deng and B. Zhang. Astrophys. J. Lett., 783: L35 (2014)

2014
[33]

H. Gao, Z. Li, and B. Zhang. Astrophys. J., 788: 189 (2014)

2014
[34]

J. M. Cordes and T. J. W. Lazio. arXiv: astro-ph/0207156 (2002)

2002 arXiv
[35]

Fukugita, C

M. Fukugita, C. J. Hogan, and P. J. E. Peebles. Astrophys. J., 503: 518 (1998)

1998
[36]

Aghanim et al

N. Aghanim et al. (Planck Collaboration). Astron. Astrophys., 641: A6 (2020). [Erratum: Astron.Astrophys. 652, C4 (2021)]

2020
[37]

L. Cayton. Univ. of California at San Diego Tech. Rep, 12 (1-17): 1 (2005)

2005
[38]

Hinton and S

G. Hinton and S. Roweis. In Proceedings of the 16th International Conference on Neural Information Processing Systems, NIPS'02, pp. 857--864, (Cambridge, MA, USA: MIT Press2002)
[39]

van der Maaten and G

L. van der Maaten and G. E. Hinton. Journal of Machine Learning Research, 9: 2579--2605 (2008)

2008
[40]

van der Maaten

L. van der Maaten. J. Mach. Learn. Res., 15: 3221--3245 (2014)

2014
[41]

Oskolkov

N. Oskolkov. How to tune hyperparameters of tSNE. Towards Data Science, US (2019)

2019
[42]

Ester, H.-P

M. Ester, H.-P. Kriegel, J. Sander et al. In Proceedings of the Second International Conference on Knowledge Discovery and Data Mining, KDD'96, p. 226–231, (Portland, Oregon: AAAI Press1996)
[43]

McInnes and J

L. McInnes and J. Healy. In 2017 IEEE International Conference on Data Mining Workshops (ICDMW), pp. 33--42, (New Orleans, LA, USA: IEEE Press2017)

2017
[44]

R. J. G. B. Campello, D. Moulavi, and J. Sander. In J. Pei, V. S. Tseng, L. Cao et al., editors, Advances in Knowledge Discovery and Data Mining, pp. 160--172, (Berlin, Heidelberg: Springer Berlin Heidelberg2013)
[45]

McInnes, J

L. McInnes, J. Healy, and S. Astels. The Journal of Open Source Software, 2 (11): 205 (2017)

2017
[46]

R. Battiti. IEEE Transactions on Neural Networks , 5 (4): 537 (1994)

1994
[47]

Rovira, K

M. Rovira, K. Engvall, and C. Duwig. Chemical Engineering Journal, 438: 135250 (2022)

2022

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

, " * write output.state after.block = add.period write newline

ENTRY address archive author booktitle chapter collaboration edition editor eid eprint howpublished institution journal key month note number numpages organization pages publisher school series title type url volume year label INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.all := #1 'mid.sent...

work page

[2] [2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

work page

[3] [3]

D. R. Lorimer, M. Bailes, M. A. McLaughlin et al. Science, 318: 777 (2007)

work page 2007

[4] [4]

Xu et al

J. Xu et al. Universe, 9 (7): 330 (2023)

work page 2023

[5] [5]

E. F. Keane et al. Nature, 530: 453--456 (2016)

work page 2016

[6] [6]

Chatterjee et al

S. Chatterjee et al. Nature, 541: 58 (2017)

work page 2017

[7] [7]

Scholz et al

P. Scholz et al. Astrophys. J., 833 (2): 177 (2016)

work page 2016

[8] [8]

L. G. Spitler et al. Nature, 531: 202 (2016)

work page 2016