Time-Series Open-set Recognition with Adaptive Local Outlier Synthesis and Exposure

Zahra Taghiyarrenani, Abdallah Alabdallah, Ali Etemad

Keywords: Open Set Recognition, Outlier Exposure, Time Series

TL;DR: ALSET improves open-set recognition in time-series by generating latent-space outliers via adaptive local Gaussians, boosting unknown detection while preserving known-class accuracy.

Abstract:

Open-set recognition (OSR) in time-series data presents a significant challenge due to the need to detect and reject unknown classes while maintaining robust classification of known classes. To address this, we introduce Adaptive Localized latent outlier Synthesis and Exposure for Time-series (ALSET), designed to operate within the latent space of any representation learning backbone, leveraging learned embeddings to enhance OSR performance for time-series.
ALSET is an outlier exposure mechanism that generates outliers by modeling the empty space around the samples from known classes using multiple Gaussians and sampling from them. It constructs local Gaussian distributions centered on known samples in the latent space, with learnable, per-dimension standard deviations, which are estimated by a neighborhood estimator. These distributions expand during training through a feedback loop between the classifier and the neighborhood estimator, which learns the structure of these Gaussians while preventing overlap with known samples.
For evaluation, we adapt several state-of-the-art OSR techniques, originally designed for computer vision, to the time-series domain for the first time, establishing a comprehensive baseline for this underexplored area. Extensive experiments on UCR, UEA, and HAR benchmarks demonstrate that ALSET consistently surpasses these baselines, achieving state-of-the-art OSR performance while preserving known-class F1. The code will be made publicly available upon acceptance.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 23449