Latent Pseudo-Anomaly Diversity Bottleneck
Background: Time-series anomaly detection often relies on generating pseudo-anomalies to train robust models, but current methods frequently produce synthetic data with insufficient diversity.
Question / Future Work: There is a need to develop mechanisms for generating high-diversity pseudo-anomalies in latent spaces, as current perturbation methods often yield limited variants that fail to capture the full spectrum of complex, real-world anomalies.
Why It Matters: The robustness of unsupervised anomaly classifiers is intrinsically linked to the coverage of the synthetic anomaly distribution; improving this is central to model performance and generalization.
Metadata & Links
- created_at
- 2026-04-18T05:34:26Z