Universal Domain Adaptation Benchmark for Time Series Data Representation

TL;DR

The paper tackles the underexplored problem of Universal Domain Adaptation for time-series data by introducing a standardized benchmark and protocol to evaluate TS backbones and UniDA methods under domain shifts. It systematically compares six UniDA methods across four backbones (including CNN and Fourier-based FNO) using a Bayesian-model-selection approach guided by the H-score $(2A_C A_U)/(A_C + A_U)$. Key findings show backbone choice profoundly influences UniDA performance, with CNN and FNO delivering the strongest results, while newer architectures like TSLANet and S3 often underperform in this setting; UniJDOT emerges as the most robust method across datasets and backbones. The work provides a practical framework and insights to drive future development of TS-tailored UniDA backbones and evaluation protocols, aiding researchers and practitioners in assessing robustness and generalization for time-series adaptation tasks.

Abstract

Deep learning models have significantly improved the ability to detect novelties in time series (TS) data. This success is attributed to their strong representation capabilities. However, due to the inherent variability in TS data, these models often struggle with generalization and robustness. To address this, a common approach is to perform Unsupervised Domain Adaptation, particularly Universal Domain Adaptation (UniDA), to handle domain shifts and emerging novel classes. While extensively studied in computer vision, UniDA remains underexplored for TS data. This work provides a comprehensive implementation and comparison of state-of-the-art TS backbones in a UniDA framework. We propose a reliable protocol to evaluate their robustness and generalization across different domains. The goal is to provide practitioners with a framework that can be easily extended to incorporate future advancements in UniDA and TS architectures. Our results highlight the critical influence of backbone selection in UniDA performance and enable a robustness analysis across various datasets and architectures.

Figures (2)

• Figure 1: Illustration of Universal Domain Adaptation (UniDA).Top: A classifier trained on the source domain fails to generalize to the target domain due to domain shift. Bottom: UniDA aligns shared classes across domains while detecting unknown target samples, assigning them to a rejection zone.
• Figure 2: Overview of the benchmark:Domain Adaptation: UniDA models extract features using various architectures, learn the task with a classification loss ($\mathcal{L}_{cls}$), align domains via discrepancy/adversarial methods, and detect unknown samples with an OOD detector. Model Selection: Hyperparameters are optimized using Bayesian search and H-score to select the best model for final evaluation.