ASCENSION: Autoencoder-Based Latent Space Class Expansion for Time Series Data Augmentation

Matthieu OLEKHNOVITCH, Dorian Joubaud, Adrien Bolling, Evgeny Zotov, Sylvain KUBLER, Maxime Cordy, Mike Papadakis, YVES LE TRAON

Abstract: Achieving effective data augmentation (DA) in time series classification is challenging due to the diverse nature of temporal data. While state-of-the-art generative models for DA -- based on GANs, diffusion models, or Variational Autoencoders (VAEs) -- demonstrate potential, they often fail to deliver consistent improvements across various datasets and application domains (e.g., ECG, power consumption, vibration sensor data), as confirmed in this study. To address this limitation, we introduce ASCENSION ($\textbf{A}$utoencoder-based latent $\textbf{s}$pace $\textbf{c}$lass $\textbf{e}$xpa$\textbf{nsion}$), a novel generative approach that harnesses the probabilistic structure of the VAE's latent space alongside an innovative controlled and progressive class expansion mechanism. It promotes compact intra-class representations while maximizing inter-class separability, thereby reducing the likelihood of class overlap during latent space exploration. We evaluate ASCENSION on 102~datasets from the UCR benchmark and compare it against six state-of-the-art DA methods. Empirical results show that ASCENSION improves average classification accuracy by approximately 1\%, whereas the strongest competing method leads to an average accuracy change of -0.3\%. ASCENSION achieves a non-negative improvement in classifier performance for 66.2\% of the 102 datasets — a 16.4\% improvement over the previous best method. These results establish ASCENSION as the first DA method that can be reliably applied in real-world scenarios where prior knowledge of method suitability is uncertain. Our study further explores the key factors driving its superior performance.

Primary Area: Deep Learning->Sequential Models, Time series

Keywords: data augmentation, classification, time series, variational auto-encoder, latent space representation, benchmark

Submission Number: 13820