Answers are grounded in this paper and cite their sources.
We consider binary classification of spatiotemporal sequences under a severe shortage of examples from the “anomaly” class. Structure-oriented synthetic augmentation is proposed as a mechanism for regularizing the training distribution: a stochastic perturbation operator is introduced that generates semantically coherent, localized distortions while keeping the perturbation norm bounded. A theoretical justification based on regularization and Rademacher complexity explains why controlled expansion of the training set reduces the effective complexity of the decision function and stabilizes learning under strong class imbalance. Experimental validation is performed in two heterogeneous domains: functional magnetic resonance imaging (fMRI) and surveillance video. For fMRI we conduct cross-domain evaluation on independent open datasets with a subject-level protocol, including both a multidomain setting and a strict cross-domain assessment. Structure-oriented augmentation consistently outperforms unstructured noise perturbations in terms of ranking ability: in the multidomain setting AUC exceeds 0.95, whereas with unstructured augmentation the average AUC is around 0.70. In the strict cross-domain scenario the structural variant shows high ranking power and surpasses an autoencoder baseline, for which the AUC drops to ~0.20. A control experiment with a matched experimental task (resting state) is additionally performed to quantify the effect of changing the protocol and domain. In this scenario we observe a degradation in quality, which delineates the method’s applicability limits under a strong cross-domain and protocol shift. On surveillance video, structure-oriented augmentation also leads to improved performance compared to training without augmentation, demonstrating that the proposed principle applies to spatiotemporal data of different natures. Code, data loading scripts, and instructions for reproducing the experiments are available in an open repository: https://github.com/AnnaZverev/Structured-Synthetic-Augmentation-as-Data-Regularization .
No takes yet. Share an insight, caveat, or question.
Zvereva et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: