Evaluating machine-learning models for multiclass classification of LIGO gravitational-wave glitches
Journal of Physics Communications · 01.11.2026 · 10 (11) · 115001
DOI: 10.1088/2399-6528/aeb8df ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Physics Communications, online erschienen am 04.11.2026. Verfasst von Rudhresh Manoharan und 1 weiteren Autor:innen. Die Arbeit nennt 30 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Gravitational-wave detectors are affected by short-duration non-Gaussian noise transients, or glitches, which can obscure astrophysical signals and complicate downstream analyses. We present a systematic benchmark of XGBoost and nine deep tabular architectures for 24-class glitch classification using structured metadata from the Gravity Spy O3 catalogue. The original comparison evaluates predictive performance, random-seed robustness, computational cost, data-scaling behavior, and model-native feature-attribution rankings. We additionally reconstruct a duplicate-aware cohort of 437,753 events and evaluate representative tree-based, fully connected, and attention-based models under frozen grouped-random and chronological protocols. XGBoost achieves the highest grouped-random mean weighted F 1 among the three audited representative models, reaching 0.8297 ± 0.0002 with the original nine features. Chronological testing is materially more difficult for all three architectures, and removing absolute timestamp variables does not eliminate the random-to-chronological performance gap; the effects of timestamp removal are also model dependent. Targeted ablations show substantial reliance on signal-strength descriptors and quality factor, whereas restoring bandwidth to the timestamp-free representation produces only modest, architecture- and split-dependent changes. Model-native TreeSHAP and Integrated Gradients rankings exhibit partially aligned but heterogeneous feature orderings. A common permutation-importance analysis supports interpreting this agreement as a protocol-dependent comparative diagnostic rather than as evidence of explanation equivalence or causal feature importance. Controlled operational measurements further show that XGBoost provides the strongest grouped-random predictive performance, while the custom multilayer perceptron has the smallest reconstructed deployment bundle and the lowest measured single-event CPU latency. These results show that random-split performance can overestimate temporal generalization and that model selection depends on the relative importance of predictive performance, deployment footprint, latency, temporal robustness, and interpretability stability.
Zitieren
APA
Manoharan, R., & Cleaver, G. (2026). Evaluating machine-learning models for multiclass classification of LIGO gravitational-wave glitches. Journal of Physics Communications, 10(11), 115001. https://doi.org/10.1088/2399-6528/aeb8df
Vancouver
Manoharan R, Cleaver G. Evaluating machine-learning models for multiclass classification of LIGO gravitational-wave glitches. Journal of Physics Communications. 2026;10(11):115001. doi:10.1088/2399-6528/aeb8df
BibTeX
@article{Manoharan2026,
author = {Manoharan, Rudhresh and Cleaver, Gerald},
title = {Evaluating machine-learning models for multiclass classification of LIGO gravitational-wave glitches},
journal = {Journal of Physics Communications},
year = {2026},
volume = {10},
number = {11},
pages = {115001},
doi = {10.1088/2399-6528/aeb8df}
}
EUVIMED