@InProceedings{10.1007/978-3-031-05936-0_28,
author="Akar, Tolga
and Werner, Thorben
and Yalavarthi, Vijaya Krishna
and Schmidt-Thieme, Lars",
editor="Gama, Jo{\~a}o
and Li, Tianrui
and Yu, Yang
and Chen, Enhong
and Zheng, Yu
and Teng, Fei",
title="Open Set Recognition for Time Series Classification",
booktitle="Advances in Knowledge Discovery and Data Mining",
year="2022",
publisher="Springer International Publishing",
address="Cham",
pages="354--366",
abstract="Traditional classification problems often assume that the number of classes present in the data is finite. This may hold true for the training data, but in real life, the risk of encountering unknown samples is ubiquitous. Classifying these unknown samples into one of the target classes can have drastic effects in some situations like security systems or body sensors. To address this problem, recently, open set recognition models that can correctly classify the known samples and detect the unknowns simultaneously, are proposed. In contrast to the existing models where unknown detection depends on the classification model, we propose, to the best of our knowledge, an open set recognition model for time series classification that works independent of the classifier by employing class-specific barycenters. Specifically, DTW distance, and the cross-correlation between the class-specific barycenters, and the input are used for detecting the unknown classes during testing. Our extensive experimental evaluation on the UEA multivariate time series archive with 30 datasets shows that the proposed open set recognition architecture deployed on top of the InceptionTime outperforms the state-of-the-art open set recognition models by an average of {\$}{\$}22{\backslash}{\%}{\$}{\$}22{\%}in terms of macro F1 score.",
isbn="978-3-031-05936-0"
}