Publication
Improving Time Series Augmentation and Self-Supervised Learning via Instance-Conditioned Contrasting
Philipp Peter Engler; Nico Müller; Ludger van Elst; Sheraz Ahmed; Andreas Dengel
In: International Joint Conference on Neural Networks. International Joint Conference on Neural Networks (IJCNN-2026), located at WCCI 2026, June 21-26, Maastricht, Netherlands, IEEE, 2026.
Abstract
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. We aim to advance both directions by combining generative approaches for time series synthesis and self-supervised learning techniques, yielding benefits for both. In this paper, we propose a self-supervised contrastive 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 contrastive module can act as a self-supervised pre-training method of its own. Generated samples can also be used as augmentation in self-supervised techniques by injecting noise into the condition. We achieve significant improvements over the standalone IC-GAN, increasing the classification accuracy of a classifier trained on synthetic data by 5.5 percentage points on average across the 22 shortest time series datasets from the UCR classification archive. Even being trained without data annotations, our GAN is competitive with recent class-conditional time series generation models and generates samples orders of magnitudes faster than a diffusion model at comparable sample quality.