Publikation
Improving Time Series Augmentation for Self-Supervised Learning via Instance-Conditioned Contrasting
Philipp Peter Engler; Nico Müller; Ludger van Elst; Sheraz Ahmed; Andreas Dengel
In: Lecture Notes in Computer Science. Annual Symposium of the German Association for Pattern Recognition (DAGM-2025), DAGM German Conference on Pattern Recognition 2025, September 23-26, Freiburg, Germany, Springer, 2025.
Zusammenfassung
Data sparsity is a major limiting factor for machine learning applications in industry. In order to make such applications feasible and also cost-effective, data synthesis strategies and self-supervised approaches can be employed, addressing the issue in different ways. While data synthesis aims to generate more data from the same distribution as existing data, self-supervised approaches leverage unlabeled data, which often is available in larger quantities than annotated data. Many industrial applications are based on time series data. We aim to advance both strategies by combining generative approaches for time series synthesis and self-supervised learning techniques, yielding benefits for both directions. In this paper we propose a self-supervised contrasting module for instance-conditioned generative adversarial networks (IC-GAN) and apply it to multiple time series datasets. This module improves the quality and diversity of generated time series without the need for data annotations. Simultaneously, the contrasting module can act as a self-supervised pre-training method of its own. We achieve significant improvements over the standalone IC-GAN and compete with self-supervised methods in the time series domain, such as TS2Vec. We further show advantages of our improved generator as an augmentation method, increasing accuracy by 1.8 p.p. over IC-GAN and 2.6 p.p. over hand-crafted transformations in a downstream task on UCR classification datasets. The code is available at