Rare Events in Deep Learning Models
Programmnaya Ingeneria · 18.05.2026 · 17 (5) · 239-252
DOI: 10.17587/prin.17.239-252 ↗Kurz gefasst
Fachartikel in der Zeitschrift Programmnaya Ingeneria, erschienen am 18.05.2026. Verfasst von F. M. Baryshnikov und 1 weiteren Autor:innen.
Abstract (Originalsprache)
Rare event detection is the task of identifying events that occur with very low frequency but have high practical or societal impact. The identification and prediction of such events present a fundamental challenge due to the scarcity of positive samples, high dimensionality of observational data, and the presence of complex temporal dependencies. Examples of rare events include certain types of rare diseases, equipment failures in manufacturing that lead to unplanned downtime, and complex cyberattacks. The frequency and nature of rare events vary substantially across different domains. In recent years, numerous methods have been developed for effective detection and classification of rare events, including modern deep learning models that achieve strong performance even when only a limited amount of labeled data is available. This article presents a systematic review and proposes а unified conceptual framework for rare event detection, distinguishing it from related tasks and clarifying the terminology used in the literature. The primary sources of rarity are analysed, including intrinsic rarity, sampling biases, annotation subjectivity, and measurement limitations. Particular attention is given to challenges arising in multivariate time series data, such as spatial correlations, short- and long-term temporal dependencies, unknown lag structures, and combined spatio-temporal effects. These factors substantially complicate modelling and often render classical balancing and augmentation techniques ineffective. The article reviews state-of-the-art methods for data preprocessing, feature selection, and feature engineering tailored to rare-event scenarios, as well as advanced deep learning models capable of capturing complex dependencies, including recurrent architectures, attention-based models, transformers, and deep reinforcement learning approaches. Specialized performance metrics designed to address extreme class imbalance are also discussed. The survey consolidates current methodological advances and identifies key challenges that remain open for future research. It provides a foundation for developing robust diagnostic, monitoring, and early-warning systems operating under conditions of rare and underrepresented events.
Zitieren
APA
Baryshnikov, F. M., & Namiot, D. E. (2026). Rare Events in Deep Learning Models. Programmnaya Ingeneria, 17(5), 239-252. https://doi.org/10.17587/prin.17.239-252
Vancouver
Baryshnikov FM, Namiot DE. Rare Events in Deep Learning Models. Programmnaya Ingeneria. 2026;17(5):239-252. doi:10.17587/prin.17.239-252
BibTeX
@article{Baryshnikov2026,
author = {Baryshnikov, F. M. and Namiot, D. E.},
title = {Rare Events in Deep Learning Models},
journal = {Programmnaya Ingeneria},
year = {2026},
volume = {17},
number = {5},
pages = {239--252},
doi = {10.17587/prin.17.239-252}
}
EUVIMED