arxiv: 2604.08796 · v1 · submitted 2026-04-09 · 🌀 gr-qc · astro-ph.IM

Recognition: unknown

Evaluating Deep Learning Models for Multiclass Classification of LIGO Gravitational-Wave Glitches

Rudhresh Manoharan (Baylor University) , Gerald Cleaver (Baylor University)

Authors on Pith no claims yet

Pith reviewed 2026-05-10 16:59 UTC · model grok-4.3

classification 🌀 gr-qc astro-ph.IM

keywords LIGO glitchesgravitational wavesdeep learningmulticlass classificationtabular datamodel benchmarkingfeature attribution

0 comments

The pith

Several deep learning models classify LIGO gravitational-wave glitches as well as tree-based methods but use far fewer parameters.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper evaluates various machine learning models for identifying different types of noise glitches in LIGO data using numerical features instead of images. It compares classical gradient-boosted trees to neural networks like multilayer perceptrons and attention models on metrics including accuracy, speed, and how performance changes with more data. The results indicate that some neural architectures reach similar accuracy levels with much smaller model sizes and show unique patterns in how they improve with additional training data. A further analysis of which features each model finds important reveals that different architectures often agree on the most relevant characteristics of the glitches. This work helps clarify when to choose simpler or more complex models for practical use in gravitational-wave detector maintenance.

Core claim

While gradient-boosted decision trees serve as strong baselines, several neural architectures achieve comparable multiclass classification accuracy on Gravity Spy tabular features with substantially lower parameter counts, display unique data-scaling patterns, and share partially overlapping feature attribution rankings.

What carries the argument

Benchmark of gradient-boosted trees against neural architectures (multilayer perceptrons, attention-based models, neural decision ensembles) on numerical glitch metadata, with evaluation of parameter efficiency, scaling behavior, and cross-model attribution consistency.

If this is right

Deep learning models become viable for resource-limited deployment in LIGO characterization pipelines.
Data efficiency varies by architecture, so training set size can be chosen to match the selected model.
Shared feature priorities across models suggest stable physical markers for certain glitch classes.
Interpretability tools can be applied more confidently when attributions align between architectures.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The approach could transfer to tabular anomaly detection in other large-scale physics experiments.
Consistent feature rankings may help isolate universal glitch signatures that do not depend on model choice.
Live-stream testing on current LIGO data would directly check whether the reported efficiencies hold in operations.

Load-bearing premise

The numerical features extracted from the Gravity Spy dataset are sufficiently informative and representative for the classification results to generalize to real LIGO detector operations.

What would settle it

A new test set of labeled glitches drawn from a later LIGO observing run where deep learning models fall below tree-based accuracy or lose their partial attribution consistency.

Figures

Figures reproduced from arXiv: 2604.08796 by Gerald Cleaver (Baylor University), Rudhresh Manoharan (Baylor University).

**Figure 1.** Figure 1: shows the class distribution of glitch types in the full Gravity Spy O3 dataset on a logarithmic scale, highlighting the pronounced class imbalance Scattered_Light Fast_Scattering Tomte Blip_Low_Frequency Low_Frequency_Burst Extremely_Loud Koi_Fish No_Glitch Low_Frequency_Lines WhistleBlip 1400Ripples Paired_Doves Air_Compressor Repeating_Blips Violin_Mode Power_Line 1080Lines Scratchy Light_ModulationHel… view at source ↗

**Figure 2.** Figure 2: provides a schematic overview of the end-toend workflow used in this study, from dataset construction through model training and evaluation. The figure summarizes the data-splitting strategy, the set of models evaluated, and the unified training and validation protocol, serving as a visual reference for the comparaGravity Spy O3 Dataset Sampled (5 × 104 ) Full (∼ 5 × 105 ) Stratified Train / Validatio… view at source ↗

**Figure 4.** Figure 4: compares the average weighted F1 score achieved by each model against the corresponding wallclock training time, shown on a logarithmic scale. This comparison addresses the practical question of how computationally expensive it is to reach a given level of classification performance under a fixed training protocol. Tree-based boosting achieves high performance with comparatively modest training cost, ref… view at source ↗

**Figure 5.** Figure 5: Weighted F1 score versus per-sample inference time for all models. Inference time is shown on a logarithmic scale to emphasize differences relevant for lowlatency deployment. passes per sample, while neural ensemble-style methods aggregate outputs across multiple learned components. These structural differences help explain why models achieving similar classification performance can nevertheless exhibit… view at source ↗

**Figure 7.** Figure 7: Spearman rank correlation between feature-importance vectors of each model and XGBoost, computed across glitch classes. Boxes indicate the distribution of correlations, with median values annotated. F. Cross-Model Interpretability Heatmap To further examine how different models encode feature-importance structure, we compute pairwise Spearman rank correlations between flattened per-class feature-importan… view at source ↗

**Figure 8.** Figure 8: Cross-model interpretability alignment measured via Spearman rank correlations between flattened per-class feature-importance vectors. Each entry quantifies the agreement between two models in how they rank tabular features across all glitch classes. Clusters of high correlation indicate models with similar inductive biases and attribution structure, while lower correlations highlight divergent feature-s… view at source ↗

**Figure 9.** Figure 9: Row-normalized confusion matrix for DANet on the 24-class glitch classification task. Each row is normalized to unity, so entries represent the fraction of samples from a given true class assigned to each predicted class. Strong diagonal structure indicates effective class separation, while off-diagonal entries reveal systematic confusions between morphologically similar glitch types. The normalization h… view at source ↗

**Figure 10.** Figure 10: Confusion matrix for DANet showing raw prediction counts across the 24 glitch classes. Each entry denotes the number of samples from a given true class assigned to a predicted class. The matrix reflects the underlying class distribution of the dataset, with dominant classes contributing larger counts along the diagonal. Consistent off-diagonal patterns corroborate systematic misclassifications observed … view at source ↗

**Figure 11.** Figure 11: Scaling behavior of models as training data increases from 50,000 to ∼ 500,000 examples. Each line connects the weighted F1 score achieved on the sampled dataset to that obtained on the full dataset. aggregation to ensure comparability across explainers and architectures [PITH_FULL_IMAGE:figures/full_fig_p011_11.png] view at source ↗

**Figure 13.** Figure 13: shows normalized feature importances for each model separately. Although broad qualitative agreement exists—particularly for dominant features such as peak_time and peak_frequency—substantial variation is evident in secondary features. For example, attention-based models tend to distribute importance more evenly across temporal and spectral descriptors, while tree-based and MLP models exhibit sharper con… view at source ↗

read the original abstract

Gravitational-wave detectors are affected by short-duration non-Gaussian noise transients, commonly referred to as glitches, which can obscure astrophysical signals and complicate downstream analyses. While recent work has demonstrated the effectiveness of deep learning models for glitch classification using image-based time-frequency representations, comparatively less attention has been given to systematic evaluations of machine-learning architectures operating directly on tabular glitch metadata. In this work, we present a comprehensive benchmark of classical and deep learning models for multiclass glitch classification using numerical features derived from the Gravity Spy dataset. We compare gradient-boosted decision trees with a diverse set of neural architectures, including multilayer perceptrons, attention-based models, and neural decision ensembles, and evaluate them in terms of classification performance, inference efficiency, parameter efficiency, data-scaling behavior, and cross-model interpretability alignment. We find that while tree-based methods remain strong baselines for tabular data, several deep learning models achieve competitive performance with substantially fewer parameters and exhibit distinct inductive biases and scaling behavior. A cross-model attribution analysis further reveals partially consistent feature-importance hierarchies across architectures, providing new insight into interpretability structure in tabular models. These results clarify trade-offs between performance, complexity, data efficiency, and interpretability in tabular gravitational-wave analyses and provide practical guidance for deploying machine-learning models in detector characterization pipelines.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

This is a competent tabular benchmark for LIGO glitch classification that shows some efficient neural models but stops short of testing real detector conditions.

read the letter

The paper runs a clean side-by-side comparison of gradient-boosted trees against several neural architectures on numerical features from the Gravity Spy dataset. It reports classification metrics, parameter counts, inference speed, scaling with training data size, and a cross-model feature-attribution check that finds partial agreement on which inputs matter most. That combination of results is the main new piece: prior glitch work has focused on spectrogram images, so this fills in the tabular side with explicit trade-off numbers rather than another image-only run.

Referee Report

2 major / 2 minor

Summary. The manuscript presents a benchmark comparing gradient-boosted decision trees against deep learning architectures (MLPs, attention-based models, neural decision ensembles) for multiclass classification of LIGO glitches using numerical features extracted from the Gravity Spy dataset. Models are evaluated on classification performance, parameter/inference efficiency, data-scaling behavior, and cross-model interpretability via attribution methods. The central claims are that selected deep learning models achieve competitive accuracy with substantially fewer parameters, exhibit distinct inductive biases and scaling, and display partially consistent feature-importance hierarchies across architectures, thereby clarifying trade-offs for tabular gravitational-wave analyses.

Significance. If the empirical results hold under scrutiny, the work offers practical value for LIGO detector-characterization pipelines by identifying efficient, interpretable alternatives to tree-based baselines on tabular glitch metadata. The inclusion of scaling curves and cross-architecture attribution analysis is a positive contribution beyond standard benchmarks, as it highlights inductive biases relevant to glitch morphology. The study is grounded in a public dataset and standard supervised-learning procedures.

major comments (2)

The claim that the results provide 'practical guidance for deploying machine-learning models in detector characterization pipelines' (abstract and conclusion) rests on the untested assumption that Gravity Spy numerical features and fixed splits generalize to non-stationary live LIGO data; no experiments on distribution shifts, unseen glitch morphologies, or integration with operational pipelines are reported, which directly affects the applicability of the performance and interpretability findings.
In the interpretability section, the statement that attribution analysis 'reveals partially consistent feature-importance hierarchies' lacks a quantitative measure of consistency (e.g., average rank correlation or overlap statistics across model pairs); without this, it is difficult to assess whether the observed alignment is statistically meaningful or merely qualitative.

minor comments (2)

The methods section should explicitly report the hyperparameter search procedure, random seeds, and any post-hoc feature selection or augmentation choices to enable full reproducibility of the reported performance gaps.
Tables comparing parameter counts and inference times would benefit from explicit units and confidence intervals; scaling plots should include error bands to reflect variability across data subsets.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. The comments highlight important limitations in the scope of our benchmark and the rigor of our interpretability claims. We address each point below and indicate the revisions we will make to the manuscript.

read point-by-point responses

Referee: The claim that the results provide 'practical guidance for deploying machine-learning models in detector characterization pipelines' (abstract and conclusion) rests on the untested assumption that Gravity Spy numerical features and fixed splits generalize to non-stationary live LIGO data; no experiments on distribution shifts, unseen glitch morphologies, or integration with operational pipelines are reported, which directly affects the applicability of the performance and interpretability findings.

Authors: We agree that the manuscript does not perform experiments on live LIGO streams, distribution shifts, or unseen morphologies, as the study is confined to the public Gravity Spy dataset with its fixed train/test splits. While Gravity Spy features are extracted from actual LIGO data and the dataset supports ongoing detector characterization, this does not substitute for direct validation on non-stationary operational data. In the revised manuscript we will qualify the practical-guidance language in both the abstract and conclusion to make clear that the recommendations are benchmark-derived rather than deployment-ready, and we will add a dedicated limitations paragraph discussing the need for future work on streaming data and morphological generalization. revision: yes
Referee: In the interpretability section, the statement that attribution analysis 'reveals partially consistent feature-importance hierarchies' lacks a quantitative measure of consistency (e.g., average rank correlation or overlap statistics across model pairs); without this, it is difficult to assess whether the observed alignment is statistically meaningful or merely qualitative.

Authors: The referee is correct that our description of consistency remained qualitative. We will strengthen the interpretability section by adding quantitative metrics: specifically, we will compute pairwise Spearman's rank correlation coefficients between the feature-attribution rankings produced by each model pair, report the mean correlation together with its standard deviation, and include p-values to indicate statistical significance. These results will be presented in a new table (or as an extension to the existing attribution figure) so that readers can directly evaluate the degree of alignment. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical benchmark on public dataset

full rationale

This paper performs a standard supervised-learning benchmark comparing gradient-boosted trees and neural architectures on numerical features extracted from the public Gravity Spy dataset. All reported metrics, scaling curves, and attribution rankings are computed directly from fixed data splits and model training runs. No derivations, fitted parameters renamed as predictions, self-citation chains, or ansatzes are present; the central claims rest on observable performance differences rather than any reduction to the inputs by construction.

Axiom & Free-Parameter Ledger

2 free parameters · 0 axioms · 0 invented entities

The work is an applied empirical benchmark. It inherits standard supervised-learning assumptions (i.i.d. samples, cross-entropy loss, etc.) but introduces no new theoretical axioms or entities. Free parameters consist of the usual neural-network hyperparameters and training choices that are fitted to the Gravity Spy data.

free parameters (2)

neural network hyperparameters
Learning rates, layer sizes, attention heads, and regularization strengths are tuned during training for each architecture.
data split ratios and augmentation choices
Train/validation/test partitions and any feature scaling or sampling decisions are chosen by the authors.

pith-pipeline@v0.9.0 · 5541 in / 1330 out tokens · 68155 ms · 2026-05-10T16:59:35.124330+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

30 extracted references · 7 canonical work pages · 1 internal anchor

[1]

B. P. Abbott, R. Abbott, T. D. Abbott, M. R. Aber- nathy, F. Acernese, K. Ackley, C. Adams, T. Adams, P. Addesso, R. X. Adhikari,et al., Physical review let- ters116, 061102 (2016)

2016
[2]

A. Abac, I. Abouelfettouh, F. Acernese, K. Ackley, C. Adamcewicz, S. Adhicary, D. Adhikari, N. Ad- hikari, R. Adhikari, V. Adkins,et al., arXiv preprint arXiv:2508.18082 (2025)

work page internal anchor Pith review arXiv 2025
[3]

Cuoco, J

E. Cuoco, J. Powell, M. Cavaglià, K. Ackley, M. Bejger, C. Chatterjee, M. Coughlin, S. Coughlin, P. Easter, R. Essick,et al., Machine Learning: Science and Tech- nology2, 011002 (2020)

2020
[4]

Cuoco, M

E. Cuoco, M. Cavaglià, I. S. Heng, D. Keitel, and C. Messenger, Living Reviews in Relativity28, 1 (2025)

2025
[5]

Zevin, S

M. Zevin, S. Coughlin, S. Bahaadini, E. Besler, N. Ro- hani, S. Allen, M. Cabero, K. Crowston, A. K. Kat- saggelos, S. L. Larson,et al., Classical and quantum gravity34, 064003 (2017)

2017
[6]

Glanzer, S

J. Glanzer, S. Banagiri, S. Coughlin, S. Soni, M. Zevin, 10 C. P. L. Berry, O. Patane, S. Bahaadini, N. Rohani, K. Crowston,et al., Classical and Quantum Gravity 40, 065004 (2023)

2023
[7]

Y. Wu, M. Zevin, C. P. Berry, K. Crowston, C. Østerlund, Z. Doctor, S. Banagiri, C. B. Jackson, V. Kalogera, and A. K. Katsaggelos, Classical and Quantum Gravity42, 165015 (2025)

2025
[8]

Zevin, C

M. Zevin, C. B. Jackson, Z. Doctor, Y. Wu, C. Øster- lund, L. C. Johnson, C. P. Berry, K. Crowston, S. B. Coughlin, V. Kalogera,et al., The European Physical Journal Plus139, 100 (2024)

2024
[9]

S. Soni, C. P. L. Berry, S. B. Coughlin, M. Harandi, C. B. Jackson, K. Crowston, C. Østerlund, O. Patane, A. K. Katsaggelos, L. Trouille,et al., Classical and Quantum Gravity38, 195016 (2021)

2021
[10]

George, H

D. George, H. Shen, and E. A. Huerta, Phys. Rev. D 97, 101501 (2018)

2018
[11]

Fernandes, S

T. Fernandes, S. Vieira, A. Onofre, J. C. Bustillo, A. Torres-Forné, and J. A. Font, Classical and Quan- tum Gravity40, 195018 (2023)

2023
[12]

Razzano and E

M. Razzano and E. Cuoco, Classical and Quantum Gravity35, 095016 (2018)

2018
[13]

Bahaadini, N

S. Bahaadini, N. Rohani, S. Coughlin, M. Zevin, V. Kalogera, and A. K. Katsaggelos, in2017 ieee in- ternational conference on acoustics, speech and signal processing (icassp)(IEEE, 2017) pp. 2931–2935

2017
[14]

J. Yan, A. P. Leung, and C. Hui, Monthly Notices of the Royal Astronomical Society515, 4606 (2022)

2022
[15]

Powell, L

J. Powell, L. Sun, K. Gereb, P. D. Lasky, and M. Doll- mann, Classical and Quantum Gravity40, 035006 (2023)

2023
[16]

Oshino, Y

S. Oshino, Y. Sakai, M. Meyer-Conde, T. Uchiyama, Y. Itoh, Y. Shikano, Y. Terada, and H. Takahashi, Physics Letters B , 139938 (2025)

2025
[17]

Y. Li, Y. Wu, and A. K. Katsaggelos, inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition(2024) pp. 6837–6846

2024
[18]

Srivastava and A

D. Srivastava and A. Niedzielski, arXiv preprint arXiv:2510.06273 (2025)

work page arXiv 2025
[19]

A.-K. Malz, G. Ashton, and N. Colombo, Physical Re- view D111, 084078 (2025)

2025
[20]

Meijer, M

Q. Meijer, M. Lopez, D. Tsuna, and S. Caudill, Phys. Rev. D109, 022006 (2024)

2024
[21]

Licciardi, D

A. Licciardi, D. Carbone, L. Rondoni, and A. Nagar, Physical Review D111, 084044 (2025)

2025
[22]

A closer look at deep learning methods on tabular datasets, 2025

H.-J. Ye, S.-Y. Liu, H.-R. Cai, Q.-L. Zhou, and D.-C. Zhan, arXiv preprint arXiv:2407.00956 (2024)

work page arXiv 2024
[23]

S. Ö. Arik and T. Pfister, inProceedings of the AAAI conference on artificial intelligence, Vol. 35 (2021) pp. 6679–6687

2021
[24]

TabTransformer: Tabular data modeling using contextual embeddings,

X. Huang, A. Khetan, M. Cvitkovic, and Z. Karnin, arXiv preprint arXiv:2012.06678 (2020)

work page arXiv 2012
[25]

Gorishniy, I

Y. Gorishniy, I. Rubachev, V. Khrulkov, and A. Babenko, Advances in neural information process- ing systems34, 18932 (2021)

2021
[26]

W. Song, C. Shi, Z. Xiao, Z. Duan, Y. Xu, M. Zhang, and J. Tang, inProceedings of the 28th ACM interna- tional conference on information and knowledge man- agement(2019) pp. 1161–1170

2019
[27]

M. S. Mia, A. A. B. Voban, A. B. H. Arnob, A. Naim, M. K. Ahmed, and M. S. Islam, in2023 International Conference on the Cognitive Computing and Complex Data (ICCD)(IEEE, 2023) pp. 51–62

2023
[28]

Popov, S

S. Popov, S. Morozov, and A. Babenko, arXiv preprint arXiv:1909.06312 (2019)

work page arXiv 1909
[29]

Joseph and H

M. Joseph and H. Raj, arXiv preprint arXiv:2207.08548 (2022)

work page arXiv 2022
[30]

Pytorch tabular: A framework for deep learning with tabular data,

M. Joseph, “Pytorch tabular: A framework for deep learning with tabular data,” (2021), arXiv:2104.13638 [cs.LG]. Appendix: Appendix A - Scaling with Data Volume To provide additional context on how different models respond to increased data availability, we examine per- formancechangeswhenscalingthetrainingsetsizefrom a sampled subset of approximately50,0...

work page arXiv 2021