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

TMLR Paper8888 Authors

Abstract:

Achieving effective data augmentation (DA) in time series (TS)
classification is challenging due to the complex nature of
temporal data. While state-of-the-art generative models for DA,
based on generative adversarial networks, diffusion models and variational autoencoders (VAEs), demonstrate potential,
they often fail to yield consistent performance gains across
diverse domains (e.g., ECG, power, vibration). To overcome this,
we propose $\textbf{ASCENSION}$ ($\textbf{A}$utoencoder-based latent
$\textbf{S}$pace $\textbf{C}$lass $\textbf{ExpaNSION}$), a novel
generative framework that leverages the probabilistic nature of a
VAE latent space together with a contrastive loss to promote
intra-class compactness and inter-class separability. Its key
innovation is an $\alpha$-scaling mechanism that progressively
expands per-class posterior covariances while preserving class
identity, populating under-represented neighbourhoods beyond the
training distribution. We evaluate ASCENSION on 102 univariate datasets
from the UCR benchmark using two established deep TS
classifiers and a recent TS foundation model, comparing
it against eight state-of-the-art DA methods. Empirical results
demonstrate that ASCENSION increases average classification
accuracy by approximately $2$%, while the strongest baseline
method results in a $-1.7$% decrease. Notably, ASCENSION
delivers non-negative performance gains on $73.5$% of datasets
(averaged over the three classifiers), compared to $50.0$%% for
the second best-performing baseline. An ablation study
further highlights the significant impact of the $\alpha$-scaling
mechanism on these gains. These findings position ASCENSION as
the only DA method in our benchmark that delivers consistent
positive gains across all three classifier families on the
102-dataset UCR archive, without requiring prior knowledge of
method suitability.

Submission Type: Long submission (more than 12 pages of main content)

Assigned Action Editor: ~Markus_Lange-Hegermann1

Submission Number: 8888