Dynamic classification approach using scalable ensemble of autoencoders to classify data with drift

Авторы: Gurina Anastasiya O., Eliseev Vladimir L., Kolpinskiy Sergey V.
Журнал: Journal of Physics: Conference Series
Том: 2134
Номер: 1
Год издания: 2021
Издательство: IOP Publishing
Местоположение издательства: [Bristol, UK], England
Первая страница: 012009
DOI: 10.1088/1742-6596/2134/1/012009
Аннотация: The problem of classification under concept drift conditions is investigated. The importance of anomaly detection is emphasized as a key feature of successful approach to operate with adversarial attacks and data poisoning. An approach to classification in the context of both drift and anomalies is introduced. It is based on ensemble of one-class classifiers, implemented by neural network autoencoders. Numeric parameters and supplementary logic are also supposed to distinguish between different classification cases. The quality of classifiers is estimated by original characteristics (EDCA), which examine both training set area and the area around it. The proposed approach is evaluated on synthetic data to highlight its properties in various conditions including normal, drift, new class and anomaly cases.
Добавил в систему: Елисеев Владимир Леонидович

	ИСТИНА	Войти в систему Регистрация
	ИПМех РАН
	Главная Поиск Статистика О проекте Помощь

ИСТИНА