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

Gravitational-wave detectors are affected by short-duration non-Gaussian noise transients, or glitches, which can obscure astrophysical signals and complicate downstream analyses. This work presents a systematic benchmarking study of machine-learning models for multiclass classification of gravitational-wave detector glitches using structured metadata derived from the Gravity Spy dataset. While deep learning approaches based on time–frequency representations have shown strong performance for glitch classification, comparatively less attention has been given to controlled comparisons of models operating directly on tabular features. In this work, we present a comprehensive benchmark of classical and deep learning models for multiclass 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 across multiple axes: classification performance, inference efficiency, parameter efficiency, data-scaling behavior, and cross-model interpretability alignment. We find that tree-based methods remain strong baselines for tabular data, while several deep learning models achieve competitive performance with fewer parameters and exhibit distinct inductive biases and scaling behavior. A cross-model attribution analysis reveals partially consistent feature-importance hierarchies across architectures, introducing cross-model interpretability consistency as an additional evaluation axis for detector-characterization models operating on tabular metadata. We further observe clustering of feature-attribution structure across model families, providing empirical evidence that inductive biases shape not only predictive performance but also interpretability behavior. These results clarify trade-offs between performance, complexity, data efficiency, and interpretability in machine-learning models for gravitational-wave detector characterization, and provide practical guidance for model selection and deployment in low-latency analysis pipelines. This is a preprint version of a manuscript submitted to Journal of Physics Communications for peer review. An earlier version of this work is available on arXiv:https://arxiv.org/abs/2604.08796 Associated code, configuration files, processed datasets, and figure-generation scripts:https://github.com/rudhresh1997/gw-glitch-tabular Versioned repository archive:https://doi.org/10.5281/zenodo.19475319 Original public dataset:https://doi.org/10.5281/zenodo.5649212