Domain Transfer Score (DTS)

• Domain Transfer Score (DTS) is a family of metrics that quantify how useful a source domain, model, or sample is for a target domain under distribution shift.
• It encompasses discrepancy-based, model-conditioned, and decision-oriented formulations, each balancing domain differences, task variances, and adaptation costs.
• DTS evaluations use empirical measures such as Wasserstein distances, Kendall’s tau, and optimal transport to rank source-target transfers and guide practical source selection.

Searching arXiv for recent and foundational papers on transferability metrics and related uses of “DTS”. Domain Transfer Score (DTS) is best understood, in the current arXiv literature, as a family of quantities that estimate how useful a source domain, source model, or source sample will be for a target domain under distribution shift. The exact term is often absent: some papers instead define a transferability metric, a domain discrepancy score, an information-gain signal, or a learned source-selection proxy, while other papers use the acronym “DTS” for unrelated concepts such as “Diffusion-based Target Sampler” or “Digital Twin Synchronization” (Zhan et al., 2023, Kazemi et al., 28 Apr 2025, Zhang et al., 2023, Sammartino, 2 Jun 2026). In this broad technical sense, a DTS-like quantity is used to rank candidate sources, predict whether transfer should be attempted, decompose transferability into domain and task components, or modulate how much source information flows into target learning.

1. Terminological status and conceptual scope

Several arXiv papers treat the underlying problem of DTS without standardizing the name. In "Benchmarking Transferability: A Framework for Fair and Robust Evaluation" (Kazemi et al., 28 Apr 2025), the exact term “Domain Transfer Score” is not used; the closest equivalent is the paper’s proposed weight-based transferability score, defined as the Wasserstein distance between original pre-trained weights and weights after short pseudo-label fine-tuning on the target domain. In "Model-based Transfer Learning for Automatic Optical Inspection based on domain discrepancy" (Salgado et al., 2023), the DTS-like object is the domain discrepancy score used for source-domain selection. In "A Collaborative Transfer Learning Framework for Cross-domain Recommendation" (Zhang et al., 2023), the closest domain-level analogue is the information gain computed by the Symmetric Companion Network, while the sample-level analogue is the selector weight $p_i^s$. In "Fast and Accurate Transferability Measurement for Heterogeneous Multivariate Data" (Park et al., 2019), the paper instead proposes Transmeter, a learned transferability measurement used to rank heterogeneous source datasets.

This plurality of terminology reflects a substantive distinction. Some papers define a DTS-like quantity as a pure discrepancy between source and target distributions; others define it as a proxy for expected post-transfer performance; still others define it as a decision statistic that explicitly compares transfer against non-transfer. "To transfer or not transfer: Unified transferability metric and analysis" (Zhan et al., 2023) is explicit on this point: transferability should answer not only how good transfer may be, but whether transfer is preferable to training on the target alone.

A further complication is acronym collision. In "Diffusion-based Target Sampler for Unsupervised Domain Adaptation" (Zhang et al., 2023), DTS means Diffusion-based Target Sampler, a generative module for UDA rather than a transferability score. In "SA-DTS: Semantic-Aware Digital Twin Synchronization over 6G Networks" (Sammartino, 2 Jun 2026), DTS means Digital Twin Synchronization, and the paper’s actual score is the Semantic Fidelity Score rather than any domain-transfer metric. Any encyclopedia treatment of DTS therefore has to separate the general transferability notion from paper-specific acronym usage.

2. Principal mathematical formulations

One major DTS lineage is risk-bound based. WDJE in (Zhan et al., 2023) defines a transferability statistic by taking an upper bound on target risk after transfer and subtracting the target risk without transfer:

$$Tr_{s,t} = \mathcal{R}_{\mathcal{D}^S}\left(h(\varnothing,\mathcal{D}^S),f^S\right) +k\lambda W\left[p^S(x),p^T(x)\right] +W\left[p^{S1}(y),p^T(y)\right] +\mathbb{E}_{y^S\sim p^{S2}(y)}\left[\|y^S\|^p\right]^{1/p} +kM\phi(\lambda) -\mathcal{R}_{\mathcal{D}^T}\left[h(\varnothing,\mathcal{D}^T),f^T\right].$$

Its interpretation is directional and decision-oriented: if $Tr_{s,t}<0$, transfer is predicted to help; if $Tr_{s,t}\ge 0$, transfer is not recommended. The decomposition makes the score a joint function of source performance, domain difference, task difference, and a probabilistic transfer penalty. The same paper gives the corresponding target-risk bound

$$\mathcal R_{\mathcal D^T}(h,f^T) \le \mathcal R_{\mathcal D^S}(h,f^S) +k\lambda W[p^S(x),p^T(x)] +W[p^S(y),p^T(y)] +kM\phi(\lambda),$$

and extends it to limited-label and unsupervised settings (Zhan et al., 2023).

Another formulation treats DTS as a source-selection discrepancy score. In the AOI paper (Salgado et al., 2023), the score is

$$disc\!\left(D^{(T)},D^{(i)}\right) = \frac{EMD(S_i,S_T)\cdot g^{(T)}}{|Y'^{(T)}\cap Y'^{(i)}|+\delta},$$

where the numerator combines Earth Mover’s Distance between dataset signatures and a Gini-based imbalance term, and the denominator uses label-space overlap. The intended use is operational rather than merely descriptive: candidate source datasets are ranked by increasing discrepancy, the three smallest scores are retained, and the one with more samples is chosen (Salgado et al., 2023).

A third formulation is OTCE, introduced as a transferability metric for cross-domain, cross-task supervised classification (Tan et al., 2021). OTCE characterizes transferability as a combination of domain difference and task difference, uses optimal transport to estimate domain difference and the optimal coupling between source and target distributions, and derives conditional entropy from that coupling. On DomainNet and Office31, OTCE shows an average of 21% gain in the correlation with the ground truth transfer accuracy compared to state-of-the-art methods, and the paper further investigates source model selection and multi-source feature fusion (Tan et al., 2021). The supplied material does not provide the exact OTCE formula, so the metric is most faithfully described at this level of abstraction.

3. Model-conditioned and learned transfer scores

Not all DTS-like quantities are distribution-only. A substantial strand of the literature defines transferability through the model’s response to the target domain. In the benchmarking framework of (Kazemi et al., 28 Apr 2025), the closest DTS-like metric is the proposed weight-based Wasserstein score:

$$\hat{Y}=\arg\max \phi_l(X), \qquad T_l^{W}=-W_1(\mathcal{P}_{\theta_l},\mathcal{P}_{\theta_l'}) \approx -W(\theta_l,\theta_l'),$$

where $\theta_l$ are pre-trained parameters and $\theta_l'$ are the parameters after two epochs of pseudo-label fine-tuning on target inputs. The negative sign is essential: smaller adaptation distance implies better transferability. This is a model-conditioned score rather than a pure source-target discrepancy, because it measures how much the model must move in parameter space to accommodate the target domain (Kazemi et al., 28 Apr 2025).

CCTL (Zhang et al., 2023) pushes this model-conditioned view further by separating domain-level and sample-level transfer signals. The Symmetric Companion Network computes batch-level information gain

$r = Loss_{pure} - Loss_{tgt},$

which evaluates whether adding weighted source data improves prediction on the target samples. Positive $$Tr_{s,t} = \mathcal{R}_{\mathcal{D}^S}\left(h(\varnothing,\mathcal{D}^S),f^S\right) +k\lambda W\left[p^S(x),p^T(x)\right] +W\left[p^{S1}(y),p^T(y)\right] +\mathbb{E}_{y^S\sim p^{S2}(y)}\left[\|y^S\|^p\right]^{1/p} +kM\phi(\lambda) -\mathcal{R}_{\mathcal{D}^T}\left[h(\varnothing,\mathcal{D}^T),f^T\right].$$0 indicates beneficial transfer; $$Tr_{s,t} = \mathcal{R}_{\mathcal{D}^S}\left(h(\varnothing,\mathcal{D}^S),f^S\right) +k\lambda W\left[p^S(x),p^T(x)\right] +W\left[p^{S1}(y),p^T(y)\right] +\mathbb{E}_{y^S\sim p^{S2}(y)}\left[\|y^S\|^p\right]^{1/p} +kM\phi(\lambda) -\mathcal{R}_{\mathcal{D}^T}\left[h(\varnothing,\mathcal{D}^T),f^T\right].$$1 indicates negative transfer. At the finer granularity, the Information Flow Network outputs a selector weight $$Tr_{s,t} = \mathcal{R}_{\mathcal{D}^S}\left(h(\varnothing,\mathcal{D}^S),f^S\right) +k\lambda W\left[p^S(x),p^T(x)\right] +W\left[p^{S1}(y),p^T(y)\right] +\mathbb{E}_{y^S\sim p^{S2}(y)}\left[\|y^S\|^p\right]^{1/p} +kM\phi(\lambda) -\mathcal{R}_{\mathcal{D}^T}\left[h(\varnothing,\mathcal{D}^T),f^T\right].$$2 for each source sample, and this weight directly scales the source loss contribution in mixed-tower training. The paper explicitly states that SCN evaluates the total benefit of the source domain, while IFN predicts the benefit of a single source sample to the target domain (Zhang et al., 2023).

Transmeter (Park et al., 2019) offers a related but supervised formulation for heterogeneous multivariate data. The paper’s conceptual definition of transferability is the relative target-accuracy improvement after transfer,

$$Tr_{s,t} = \mathcal{R}_{\mathcal{D}^S}\left(h(\varnothing,\mathcal{D}^S),f^S\right) +k\lambda W\left[p^S(x),p^T(x)\right] +W\left[p^{S1}(y),p^T(y)\right] +\mathbb{E}_{y^S\sim p^{S2}(y)}\left[\|y^S\|^p\right]^{1/p} +kM\phi(\lambda) -\mathcal{R}_{\mathcal{D}^T}\left[h(\varnothing,\mathcal{D}^T),f^T\right].$$3

but the practical score is not a closed-form discrepancy. Instead, the method trains a source-target pairwise model consisting of a target encoder, target decoder, label predictor, and domain classifier, and uses the resulting target predictive performance as the transferability estimate. The architecture is asymmetric: source features are left in their own space, while target features are encoded into a homogeneous space compatible with the pre-trained source classifier. Domain adversarial training reduces the domain gap, and reconstruction prevents trivial alignment (Park et al., 2019).

These model-conditioned formulations differ from pure discrepancy scores in a consistent way. They do not ask only whether source and target distributions are close, but whether the source model, source samples, or source representation are actually useful once exposed to target data. That distinction is central to later benchmarking results.

4. Evaluation protocols and empirical behavior

The literature evaluates DTS-like quantities primarily by their agreement with downstream transfer outcomes. In (Kazemi et al., 28 Apr 2025), the ranking criterion is Kendall’s tau,

$$Tr_{s,t} = \mathcal{R}_{\mathcal{D}^S}\left(h(\varnothing,\mathcal{D}^S),f^S\right) +k\lambda W\left[p^S(x),p^T(x)\right] +W\left[p^{S1}(y),p^T(y)\right] +\mathbb{E}_{y^S\sim p^{S2}(y)}\left[\|y^S\|^p\right]^{1/p} +kM\phi(\lambda) -\mathcal{R}_{\mathcal{D}^T}\left[h(\varnothing,\mathcal{D}^T),f^T\right].$$4

with weighted Kendall’s tau used in experiments to emphasize correct ordering among top-performing models. The benchmark varies source dataset, model-hub complexity, and fine-tuning strategy across 12 visual target datasets, 11 supervised architectures, and 10 self-supervised models. Its central conclusion is explicitly non-universal: the results do not support the existence of a single universally dominant score. For example, ETran drops from an average weighted Kendall tau of 0.562 on ImageNet to 0.143 on CIFAR-100 source models, while the proposed Wasserstein metric has lower standard deviation $$Tr_{s,t} = \mathcal{R}_{\mathcal{D}^S}\left(h(\varnothing,\mathcal{D}^S),f^S\right) +k\lambda W\left[p^S(x),p^T(x)\right] +W\left[p^{S1}(y),p^T(y)\right] +\mathbb{E}_{y^S\sim p^{S2}(y)}\left[\|y^S\|^p\right]^{1/p} +kM\phi(\lambda) -\mathcal{R}_{\mathcal{D}^T}\left[h(\varnothing,\mathcal{D}^T),f^T\right].$$5 across supervised model-hub subsets and achieves a relative $$Tr_{s,t} = \mathcal{R}_{\mathcal{D}^S}\left(h(\varnothing,\mathcal{D}^S),f^S\right) +k\lambda W\left[p^S(x),p^T(x)\right] +W\left[p^{S1}(y),p^T(y)\right] +\mathbb{E}_{y^S\sim p^{S2}(y)}\left[\|y^S\|^p\right]^{1/p} +kM\phi(\lambda) -\mathcal{R}_{\mathcal{D}^T}\left[h(\varnothing,\mathcal{D}^T),f^T\right].$$6 improvement over SFDA in the head-training setup, with aggregated average weighted Kendall tau $$Tr_{s,t} = \mathcal{R}_{\mathcal{D}^S}\left(h(\varnothing,\mathcal{D}^S),f^S\right) +k\lambda W\left[p^S(x),p^T(x)\right] +W\left[p^{S1}(y),p^T(y)\right] +\mathbb{E}_{y^S\sim p^{S2}(y)}\left[\|y^S\|^p\right]^{1/p} +kM\phi(\lambda) -\mathcal{R}_{\mathcal{D}^T}\left[h(\varnothing,\mathcal{D}^T),f^T\right].$$7 versus $$Tr_{s,t} = \mathcal{R}_{\mathcal{D}^S}\left(h(\varnothing,\mathcal{D}^S),f^S\right) +k\lambda W\left[p^S(x),p^T(x)\right] +W\left[p^{S1}(y),p^T(y)\right] +\mathbb{E}_{y^S\sim p^{S2}(y)}\left[\|y^S\|^p\right]^{1/p} +kM\phi(\lambda) -\mathcal{R}_{\mathcal{D}^T}\left[h(\varnothing,\mathcal{D}^T),f^T\right].$$8 (Kazemi et al., 28 Apr 2025).

OTCE is evaluated in a more conventional transferability-prediction setting. Its abstract reports experiments on DomainNet and Office31 and states that OTCE shows an average of 21% gain in the correlation with the ground truth transfer accuracy compared to state-of-the-art methods. The same paper also studies source model selection and multi-source feature fusion, which are canonical DTS use cases because they require transferability estimation before costly adaptation (Tan et al., 2021).

Transmeter is assessed through source-identification accuracy rather than correlation. On 10 heterogeneous multivariate binary-classification datasets, it achieves top-1 best-source identification accuracy $$Tr_{s,t} = \mathcal{R}_{\mathcal{D}^S}\left(h(\varnothing,\mathcal{D}^S),f^S\right) +k\lambda W\left[p^S(x),p^T(x)\right] +W\left[p^{S1}(y),p^T(y)\right] +\mathbb{E}_{y^S\sim p^{S2}(y)}\left[\|y^S\|^p\right]^{1/p} +kM\phi(\lambda) -\mathcal{R}_{\mathcal{D}^T}\left[h(\varnothing,\mathcal{D}^T),f^T\right].$$9, top-2 $Tr_{s,t}<0$0, top-3 $Tr_{s,t}<0$1, and top-4 $Tr_{s,t}<0$2, compared with HeMap-t’s $Tr_{s,t}<0$3, $Tr_{s,t}<0$4, $Tr_{s,t}<0$5, and $Tr_{s,t}<0$6. The paper also reports up to $Tr_{s,t}<0$7 faster training time than HeMap-t, and that using Transmeter for source selection followed by full transfer is up to $Tr_{s,t}<0$8 faster than exhaustive HeMap while giving better or equal accuracy in 8 out of 10 cases (Park et al., 2019).

Across these evaluations, two empirical regularities recur. First, DTS-like scores are typically assessed as ranking devices rather than as absolute predictors of target accuracy. Second, the strongest metrics are context-sensitive. Architecture scale, pre-training source, label availability, and adaptation regime materially alter which score works best.

5. Theoretical constraints on what a score should measure

A theoretical limitation of many DTS-like quantities is that marginal matching alone does not guarantee content-aligned transfer. "Domain Transfer Becomes Identifiable via a Single Alignment" (Shrestha et al., 18 May 2026) makes this point explicit. The paper studies unsupervised domain transfer under the data model

$Tr_{s,t}<0$9

with $Tr_{s,t}\ge 0$0, and shows that classical distribution matching

$Tr_{s,t}\ge 0$1

is non-identifiable because measure-preserving automorphisms can preserve marginals while altering cross-domain correspondences. Its identifiable criterion is

$Tr_{s,t}\ge 0$2

and Theorem 1 states that under structural sparsity, a regularity condition, and $Tr_{s,t}\ge 0$3, one anchor sample suffices to identify the ground-truth transfer almost everywhere (Shrestha et al., 18 May 2026).

This result has a direct implication for DTS design. A score that uses only distribution fit can rank a semantically wrong transfer map highly. A theoretically better score should incorporate, at minimum, distribution matching, anchor consistency, and some proxy for structural simplicity. The same paper provides a scalable practical surrogate for Jacobian sparsity,

$Tr_{s,t}\ge 0$4

based on randomized masked finite differences rather than explicit Jacobian evaluation (Shrestha et al., 18 May 2026). Although the paper does not define a scalar DTS, it supplies a principled criterion for distinguishing true transferability from MPA-like ambiguity.

Theoretical caution also appears in WDJE. There, the transferability decision depends on loss assumptions—triangle inequality, symmetry, boundedness, and $Tr_{s,t}\ge 0$5-Lipschitzness—and on the notion of $Tr_{s,t}\ge 0$6-Lipschitz transferability under a coupling $Tr_{s,t}\ge 0$7 (Zhan et al., 2023). These assumptions explain why WDJE is a bound-based decision score rather than an exact oracle.

6. Application domains, limitations, and acronym ambiguity

DTS-like quantities appear across a wide range of application domains. In AOI for magnetic tile inspection, the domain discrepancy score is used to select an industrial source dataset whose transferred backbone outperforms benchmark-source transfer, with the paper reporting increases in the F1 score and the PR curve up to 20% compared with transfer learning using benchmark datasets (Salgado et al., 2023). In cross-domain recommendation, CCTL uses information gain $Tr_{s,t}\ge 0$8 and source-sample weights $Tr_{s,t}\ge 0$9 to avoid negative migration under different CTR distributions, schema sizes, and data quantities; the abstract reports that the method achieved SOTA score on offline metrics and was deployed in Meituan, bringing 4.37% CTR and 5.43% GMV lift (Zhang et al., 2023). In dialogue state tracking, no explicit DTS is defined, but post-transfer target-domain turn-level accuracy and gain over zero-shot baseline act as transfer-effectiveness proxies; policy-gradient adaptation improves nearly every cross-domain source-target pair and sometimes approaches in-domain supervised performance (Bingel et al., 2019).

The main limitations are equally consistent across papers. Some scores are label-dependent or require labeled target samples, as in feature-based methods such as LogME and SFDA, and in the conceptual structure of OTCE’s task-difference term (Kazemi et al., 28 Apr 2025, Tan et al., 2021). Weight-based scores can operate with unlabeled target samples, but they require a short adaptation loop for every candidate model and are therefore more expensive than purely static feature-based metrics (Kazemi et al., 28 Apr 2025). Learned scores such as Transmeter are supervised on the target dataset and involve pairwise training overhead (Park et al., 2019). CCTL’s information-gain signals are model-dependent and batch-sensitive rather than static domain similarities (Zhang et al., 2023). More broadly, the benchmarking evidence emphasizes that transferability estimation is scenario-dependent rather than universally reliable (Kazemi et al., 28 Apr 2025).

A final source of confusion is lexical rather than methodological. In (Zhang et al., 2023), DTS is not a score at all but a diffusion-based generator for pseudo target samples in UDA. In (Sammartino, 2 Jun 2026), DTS refers to Digital Twin Synchronization, and the relevant metric is the Semantic Fidelity Score

$$\mathcal R_{\mathcal D^T}(h,f^T) \le \mathcal R_{\mathcal D^S}(h,f^S) +k\lambda W[p^S(x),p^T(x)] +W[p^S(y),p^T(y)] +kM\phi(\lambda),$$0

which measures downstream task preservation under semantic communication rather than source-to-target transferability (Sammartino, 2 Jun 2026). Any use of “DTS” therefore requires immediate contextual disambiguation.

Taken together, the arXiv literature does not present a single canonical Domain Transfer Score. Instead, it presents a technically coherent design space. At one end are discrepancy-based scores, which compare source and target marginals or label distributions; at another are model-conditioned scores, which measure adaptation distance or target loss reduction; and at the most decision-oriented end are bound-based statistics such as WDJE, which compare predicted transfer risk against the no-transfer baseline. The common objective is unchanged: estimate, before exhaustive fine-tuning, whether a particular source-to-target transfer path is likely to be beneficial.