PAS: Estimating the target accuracy before domain adaptation

Raphaella Diniz; Jackson de Faria; Martin Ester

cs.CVApr 10, 2026

PAS: Estimating the target accuracy before domain adaptation

Authors: Raphaella Diniz, Jackson de Faria, Martin Ester

Organizations: School of Computing Science Simon Fraser University Canada

Abstract

The goal of domain adaptation is to make predictions for unlabeled samples from a target domain with the help of labeled samples from a different but related source domain. The performance of domain adaptation methods is highly influenced by the choice of source domain and pre-trained feature extractor. However, the selection of source data and pre-trained model is not trivial due to the absence of a labeled validation set for the target domain and the large number of available pre-trained models. In this work, we propose PAS, a novel score designed to estimate the transferability of a source domain set and a pre-trained feature extractor to a target classification task before actually performing domain adaptation. PAS leverages the generalization power of pre-trained models and assesses source-target compatibility based on the pre-trained feature embeddings. We integrate PAS into a framework that indicates the most relevant pre-trained model and source domain among multiple candidates, thus improving target accuracy while reducing the computational overhead. Extensive experiments on image classification benchmarks demonstrate that PAS correlates strongly with actual target accuracy and consistently guides the selection of the best-performing pre-trained model and source domain for adaptation.

Figures & tables

Figure 1: The Potential Adaptability Score (PAS) estimates the performance of adapting to an unlabeled target domain given a pre-trained feature extractor and a labeled source domain. It helps in the selection of the best pre-trained model and best source domain among many candidates and is highly correlated with the final target accuracy after domain adaptation.

Figure 2: Source and target samples in the embedding space of a pre-trained model. (top left) Ideally, a target sample from a given class should be more similar, and hence closer in the embedding space of the pre-trained model, to a source sample from the same class. (top right) If new discriminative features need to be learned, the chances of overfitting on the source domain during adaptation increase. (bottom) Illustration of the distances from a target sample to all source class centroids. Our PAS score considers the relationship between distances $d_{1}$ and $d_{2}$ , which correspond to the shortest and second shortest distances, respectively.

Figure 3: The correlation between the PAS score value and the target accuracy after the domain adaptation. Each box summarizes the target accuracy of different domain adaptation methods for a given source-target pair and a pre-trained feature extractor. Higher values for the PAS score are strongly correlated with higher target accuracy.

Figure 4: The PAS value and target accuracy for the DANN and MCC methods using different pre-trained feature extractors. The PAS score can help to select the feature extractor that leads to higher accuracy. (left) A→C adaptation in the Office-Home benchmark. (right) W→A adaptation in the Office-31 benchmark.

Table 1: Average target accuracy of domain adaptation methods and transferability scores for different image classification benchmarks. The highest values are highlighted. Our PAS has a high correlation with the target accuracy and, for each target domain, attributes the highest value for the source domain that leads to the highest target accuracy in most scenarios. * Oracle baseline that considers the target labels.

Figure 5: The PAS value varying with the number of samples for the Office-Home . The PAS values are quite robust to varying numbers of samples. Most importantly, the relative order of PAS values for different source domains remains unchanged.

PAS	\cellcolor gray!20 0.76	0.63	\cellcolor gray!20 0.44		\cellcolor gray!20 0.58
	Office-Home	Office-31	ImageCLEF	DomainNet	Total
Euclidean distance	0.70	\cellcolor gray!20 0.69	0.27	0.54	0.68
Average cosine distance	0.66	0.52	0.12	0.48	0.66

Table 5: Pearson correlation between the target accuracy and the PAS score, which considers the cosine distance to the cluster centroid, and modifications using the Euclidean distance to the centroid and the average cosine distance to the source cluster samples. The maximum correlation value for each benchmark is highlighted. The design choices of PAS lead to the higher overall correlation between the score and the target accuracy.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

\cellcolor white	Target	A		D		W		Correlation with PAS
\cellcolor white	Source	D	W	A	W	A	D	Pearson	Spearman
\cellcolor white	DANN	73.3	70.4	83.6	100.0	91.4	97.9	0.78	0.66
\cellcolor white	ADDA	69.6	72.5	90.0	99.7	94.6	97.5	0.67	0.60
\cellcolor white	BSP	74.1	73.8	88.2	100.0	92.7	97.9	0.75	0.66
\cellcolor white	DAN	66.9	65.2	87.3	100.0	84.2	98.4	0.83	0.83
\cellcolor white	JAN	69.2	71.0	89.4	100.0	93.7	98.4	0.70	0.60

Appendix

Table 7: Target accuracy of domain adaptation methods and transferability scores for the Office-31 dataset. The highest values are highlighted. * Oracle baseline that considers the target labels.

\cellcolor white	Target	C		I		P		Correlation with PAS
\cellcolor white	Source	I	P	C	P	C	I	Pearson	Spearman
\cellcolor white	RTN	95.3	92.2	86.9	86.8	72.7	75.6	0.29	0.49
\cellcolor white	MADA	96.0	92.2	88.8	87.9	75.2	75.0	0.20	0.26
\cellcolor white	iCAN	94.7	92	89.9	89.7	78.5	79.5	0.23	0.49
\cellcolor white	CDAN-E	97.7	94.3	91.3	90.7	74.2	77.7	0.27	0.49
\cellcolor white	SymNets	97.0	96.4	93.4	93.6	78.7	80.2	0.17	0.60

Appendix

Table 8: Target accuracy of domain adaptation methods and transferability scores for the ImageCLEF dataset. The highest values are highlighted. * Oracle baseline that considers the target labels.

Panel of Figure 6 — Figure 6: Examples of images misclassified by the domain adaptation method DANN in the dataset P ( Pascal VOC 2012 ) of the ImageCLEF benchmark

Domain Adaptation

Domain adaptation transfers machine-learning models across changing data distributions, often using unlabeled or limited target-domain examples.

See all 252 papers

Trust-Aware Joint Feature-Prediction Discrepancy for Robust Domain Adaptation
May 24, 2026Xi Ding, Lei Wang, Syuan-Hao Li +1Distribution Shift Robustness Unsupervised Domain Adaptation
School of Engineering and Built Environment, Griffith University, Australia
Domain-Division based Progressive Learning for Source-Free Domain Adaptation
Jul 31, 2026Pan Liu, Jing Li, Meng Zhao +3Source-Free Domain Adaptation Unsupervised Domain Adaptation
School of Computer Science and Engineering, Tianjin University of Technology, Tianjin 300384, China
Language-Induced Priors for Domain Adaptation
May 14, 2026Qiyuan Chen, Jiayu Zhou, Raed Al KontarMulti-Source Domain Adaptation Domain Adaptation
Industrial and Operations Engineering University of Michigan Ann Arbor, MI 48109 · School of Information University of Michigan Ann Arbor, MI 48109

May 24, 2026J/K move · Enter open · S save

Trust-Aware Joint Feature-Prediction Discrepancy for Robust Domain Adaptation

Xi Ding, Lei Wang, Syuan-Hao Li, Yongsheng Gao

School of Engineering and Built Environment, Griffith University, Australia

Distribution Shift Robustness Unsupervised Domain Adaptation Domain Adaptation

Open paper AskPDF arXiv

Domain adaptation aims to mitigate performance degradation caused by distribution shifts between a labeled source domain and an unlabeled or sparsely labeled target domain. Most existing approaches estimate domain discrepancy either in feature space or in prediction space. However, these single-perspective strategies overlook a critical problem under domain shift: the reliability of the signals used for alignment. In practice, both learned representations and semantic predictions may become unreliable, and treating all target samples equally can lead to misleading alignment and suboptimal transfer. We introduce trust-aware domain adaptation, a principled framework that models domain discrepancy through the reliability of feature and prediction signals. Central to our approach is the Joint Feature-Prediction Discrepancy (JFPD), a unified formulation that jointly captures representation divergence and prediction divergence while weighting their contributions by sample-specific trust. Trust is quantified via two complementary mechanisms: uncertainty-aware trust, derived from prediction entropy to suppress unreliable predictions, and semantic-alignment trust, computed from prototype similarity in feature space to emphasize well-aligned representations. By prioritizing confident and semantically consistent samples while down-weighting noisy or ambiguous ones, JFPD provides a reliability-aware estimate of domain discrepancy. We further integrate JFPD into a training objective that guides adaptation toward trustworthy regions of the target domain. Experiments on standard benchmarks demonstrate that the proposed framework consistently achieves superior adaptation performance and yields discrepancy estimates that correlate with target-domain error. This work addresses, for the first time, the importance of modeling trust in the interaction between features and predictions for domain adaptation.

Abstract

Domain Adaptation

Explore similar work

Trust-Aware Joint Feature-Prediction Discrepancy for Robust Domain Adaptation

Domain-Division based Progressive Learning for Source-Free Domain Adaptation

Language-Induced Priors for Domain Adaptation