A Comprehensive Benchmark of Source-Free Universal Domain Adaptation on Time Series Representations

Romain Mussard; Fannia Pacheco; Maxime Berar; Paul Honeine; Gilles Gasso

cs.LGSep 30, 2026

A Comprehensive Benchmark of Source-Free Universal Domain Adaptation on Time Series Representations

Authors: Romain Mussard, Fannia Pacheco, Maxime Berar, Paul Honeine, Gilles Gasso

Organizations: Univ Rouen Normandie, INSA Rouen Normandie, Universit´e Le Havre Normandie, Normandie Univ, LITIS UR 4108, F-76000 Rouen, France

Abstract

Source-Free Universal Domain Adaptation (SF-UniDA) extends Universal Domain Adaptation by removing access to source data at adaptation time while still handling label-set mismatches between domains. Despite growing interest in this setting for image data, no benchmark exists for time series, which are more challenging. We present the first SF-UniDA benchmark on time series. In addition, we provide the first study of pretrained foundation models as feature extractors for time series domain adaptation. In this context, we identify a critical and previously underexplored limitation of all existing SF-UniDA methods: the inference threshold for unknown-sample rejection is highly sensitive. We address this by proposing a plug-in auto-thresholding module that can be integrated into any SF-UniDA method. Experiments on three well-known time series datasets confirm the suitability of this module. They also highlight that foundation models do not systematically outperform classical backbones and that SF-UniDA tailored for time series is yet to be developed.

Figures & tables

Fig. 1 : Overview of the SF-UniDA benchmark for time series. Task-specific backbone architectures and pretrained foundation models are evaluated as feature extractors under the SF-UniDA methods. Both fixed thresholding and auto-thresholding are applied to each SF-UniDA method. The performance is measured with H-score over 3 time series datasets.

Model	Type	Architecture	Param.	Input	Pre-training objective	Task
MOMENT [ 5 ]	Encoder	T5	40M	Patches	Reconstruction	General-purpose
Mantis [ 3 ]	Encoder	ViT	8M	Tokens	Contrastive learning	Classification
Chronos [ 1 ]	Seq2Seq	T5 and LLM	8M	Scalar quantized	Autoregression	Forecasting

Table 1 : Comparison of MOMENT, Mantis, and Chronos time series foundation models. General-purpose tasks include classification, forecasting, anomaly detection, and imputation.

Method	Key idea	Unknown detection	Source-free	Threshold-free (Inference)
UMAD [ 9 ]	Dual-head classifier consistency	Consistency + MixUP	Weak	✓
GLC [ 14 ]	Global One Vs All + local k-NN	One Vs All clustering	Full	✗
GLC++ [ 15 ]	GLC + contrastive loss	One Vs All clustering	Full	✗
LEAD [ 13 ]	Orthogonal feature decomposition	Gaussian Mixture Model	Full	✗

Table 2 : Comparison of SF-UniDA methods. Threshold-free refers to inference only. “Weak” source-free indicates UMAD requires a dual-head architecture and auxiliary orthogonal loss during source pre-training.

Fig. 2 : Source model confidence distributions on target data (no adaptation) for known and unknown samples. Left: CNN / EDF (Time Series). Right: ResNet50 / Office-31 (Images).

Panel of Figure 5 — Fig. 3 : Threshold sensitivity. Left: Mean H-score as a function of $\tau$ for GLC on time series (HAR) and image (Office) data with a CNN backbone and two foundation models, showing that the optimal threshold range is far narrower for time series. Right: Per-scenario H-score as a function of $\tau$ for GLC with a CNN backbone on HAR, illustrating that the optimal threshold also varies across adaptation scenarios.

Datasets	Methods	Backbones				Foundation Models
Datasets	Methods	CNN	TFE	S3	TSLANet	Mantis	Moment	Chronos
HAR	UniJDOT	61.0 ∗ (–)	64.6 ∗ (–)	54.1 ∗ (–)	59.7 ∗ (–)	82.5 (–)	30.1 (–)	72.2 (–)
	UMAD	52.7 ( $\downarrow$ 2.5)	64.6 ( $\uparrow$ 29.7)	36.1 ( $\downarrow$ 7.5)	43.1 ( $\uparrow$ 11.5)	51.2 ( $\uparrow$ 16.2)	24.2 ( $\uparrow$ 11.4)	46.0 ( $\uparrow$ 13.3)
	GLC	46.2 ( $\uparrow$ 4.2)	42.2 ( $\uparrow$ 29.0)	52.3 ( $\uparrow$ 11.2)	54.0 ( $\uparrow$ 5.8)	53.4 ( $\uparrow$ 15.8)	16.7 ( $\uparrow$ 16.6)	50.6 ( $\uparrow$ 28.9)
	GLC++	44.7 ( $\uparrow$ 5.4)	36.4 ( $\uparrow$ 20.8)	50.7 ( $\uparrow$ 5.0)	54.1 ( $\uparrow$ 17.8)	59.6 ( $\uparrow$ 51.1)	13.6 ( $\uparrow$ 10.8)	50.9 ( $\uparrow$ 33.8)
	LEAD	43.8 ( $\uparrow$ 7.2)	42.3 ( $\uparrow$ 28.1)	50.4 ( $\uparrow$ 4.4)	53.1 ( $\uparrow$ 15.5)	57.8 ( $\uparrow$ 13.3)	13.7 ( $\uparrow$ 11.5)	50.6 ( $\uparrow$ 30.7)

Table 3 : H-score (%) with Auto-Thresholding (Yen’s criterion) over 10 seeds. Parentheses show gain vs. fixed thresholding: (

\uparrow

) gain, (

\downarrow

) loss, (

=

) no change. Bold and italic indicate the best and second-best architectures for each method, respectively. UniJDOT in gray is a UniDA method for which backbone results marked with ∗ are reported directly from [ 11 ] .

Threshold	Backbones			Foundation M.
Threshold	CNN	TFE	S3Layer	Mantis	Chronos
yen	46.2	42.2	52.3	53.4	50.6
otsu	44.5	42.9	49.8	45.7	42.9
li	42.0	41.3	55.3	48.0	40.3

Table 4 : H-score (%) over 10 seeds for GLC on HAR across backbones and thresholding methods

Explore similar work

Temporal-Spectral Alignment with Frequency Adaptation for Source-Free Time-Series Adaptation
Jun 22, 2026Shichang Meng, Linquan Wu, Xuan Ai +1Domain Adaptation Frequency-Aware Regularization
City University of Hong Kong Hong Kong, China
LFM: Leveraging Foundation Models for Source-Free Universal Domain Adaptation
Jul 20, 2026Jing Li, Pan Liu, Meng Zhao +7Domain Adaptation Wireless Foundation Models
School of Computer Science and Engineering, Tianjin University of Technology, Tianjin 300384, China · Engineering Research Center of Learning-Based Intelligent System, Ministry of Education of the People’s Republic of China, Tianjin University of Technology, Tianjin 300384, China · School of Artificial Intelligence, Tianjin University, Tianjin 300350, China +1
CPDA: Class-Conditional Path Distribution Alignment for Unsupervised Time-Series Domain Adaptation
Aug 10, 2026Felix Ott, Christopher MutschlerDomain Adaptation Time-Series Classification
Fraunhofer Institute for Integrated Circuits IIS, 90411 Nürnberg, Germany · Machine Learning and Positioning Systems Department, University of Technology Nürnberg (UTN), 90461 Nürnberg

Jun 22, 2026, cs.LGJ/K move · Enter open · S save

Temporal-Spectral Alignment with Frequency Adaptation for Source-Free Time-Series Adaptation

Shichang Meng, Linquan Wu, Xuan Ai, Linqi Song

City University of Hong Kong Hong Kong, China

Domain Adaptation Frequency-Aware Regularization

Open paper AskPDF arXiv

The goal of source-free domain adaptation (SFDA) for time-series data is to transfer knowledge from a pre-trained source model to an unlabeled target domain without requiring access to source data, while addressing feature shift and temporal drift inherent in the signals. Although existing approaches have explored temporal dynamics in unsupervised source-free adaptation, they largely overlook spectral shifts in time-series data. Towards this end, we propose a novel approach termed temporal-Spectral Alignment with Frequency Adaptation (SAFA) for source-free time-series domain adaptation. Specifically, we first model the source domain at multiple scales by jointly capturing temporal dependencies and spectral characteristics. To adapt time-series data in the target domain, we introduce a trainable frequency adaptation module that modulates the phase and amplitude of target signals in the frequency domain to align them with the source distribution. Extensive experiments on multiple benchmark datasets demonstrate the efficacy and robustness of SAFA.

Abstract

Explore similar work

Temporal-Spectral Alignment with Frequency Adaptation for Source-Free Time-Series Adaptation

LFM: Leveraging Foundation Models for Source-Free Universal Domain Adaptation

CPDA: Class-Conditional Path Distribution Alignment for Unsupervised Time-Series Domain Adaptation