MD-SupContrast: Multi-Domain Contrastive Learning

• Multi-Domain Supervised Contrastive Learning is a paradigm that aligns same-class samples across diverse domains to enhance intra-class compactness and inter-class separation.
• It integrates complementary feature domains—including texture, time-frequency cues, and statistical descriptors—to improve robustness in noisy and cluttered sensing environments.
• Empirical studies report that MD-SupContrast outperforms cross-entropy baselines, achieving significant accuracy gains on datasets like DomainNet, UAV RF, and radar detection tasks.

Multi-Domain Supervised Contrastive Learning (MD-SupContrast) denotes a family of supervised contrastive paradigms in which representation learning is driven by domain variation rather than treated as a nuisance to be ignored. One explicit definition describes MD-SupContrast as the application of Supervised Contrastive Learning to a source dataset composed of multiple related domains with a shared label space, so that images of the same class, regardless of domain, act as positives and images of different classes act as negatives (Tan et al., 2023). In later task-specific systems, the term is also used for architectures that combine multiple feature domains—such as texture, time-frequency positional cues, or shallow statistical descriptors—before or during supervised contrastive optimization, with the aim of improving intra-class compactness, inter-class separation, transferability, and robustness in open-set or cluttered sensing environments (Gao et al., 18 Aug 2025, Wang et al., 2024).

1. Definition, scope, and recurring design pattern

A direct formulation of MD-SupContrast appears in work on DomainNet transfer, where multi-domain training means combining data from several visual domains into a single pre-training corpus and treating all same-class samples across domains as positives in the supervised contrastive loss (Tan et al., 2023). In this sense, the method is a domain-invariant representation learner: it explicitly aligns class structure across domain boundaries rather than relying only on per-sample classification. A closely related formulation appears in multi-domain text classification, where a robust contrastive alignment method aligns text classification features of various domains in the same feature space by supervised contrastive learning and replaces the usual private-shared design with two universal feature extractors (Li et al., 2022).

The term is also used in signal-processing settings where “multi-domain” refers to complementary representational domains. In UAV radio-frequency recognition, texture features and time-frequency position features are fused and then optimized by supervised contrastive learning; in radar target detection, time-domain, frequency-domain, and time-frequency statistical features guide deep supervised contrastive learning through an auxiliary alignment mechanism (Gao et al., 18 Aug 2025, Wang et al., 2024). This suggests that MD-SupContrast has acquired two closely related meanings: cross-domain class alignment across datasets or sources, and multi-domain feature integration within a sensing pipeline.

Paper	Domain notion	Core mechanism
(Li et al., 2022)	Text domains	Two universal feature extractors aligned in one feature space
(Tan et al., 2023)	Six visual domains in DomainNet	Same-class cross-domain positives in SupCon pre-training
(Gao et al., 18 Aug 2025)	Texture and time-frequency position features	ResNet–TransformerEncoder fusion with SupCon
(Wang et al., 2024)	Time-, frequency-, and time-frequency features	Shallow-to-deep guidance plus supervised contrastive loss

2. Objective functions and representation geometry

The canonical MD-SupContrast objective follows the supervised contrastive formulation over an augmented mini-batch. For a batch of $N$ original samples augmented twice into $2N$ views, the per-sample loss is defined as

$\mathcal{L}^{\mathrm{SupCon}}_i = -\frac{1}{2N_{\tilde y_i}-1} \sum_{j=1}^{2N} \mathds{1}_{i\neq j}\,\mathds{1}_{\tilde y_i=\tilde y_j} \log \frac{ \exp\bigl(\mathrm{sim}(z_i,z_j)/\tau\bigr) }{ \sum_{k=1}^{2N}\mathds{1}_{i\neq k}\exp\bigl(\mathrm{sim}(z_i,z_k)/\tau\bigr) }, \qquad \mathcal{L}^{\mathrm{SupCon}} = \sum_{i=1}^{2N}\mathcal{L}^{\mathrm{SupCon}}_i .$

Here $z_i=h(g(x_i))$ is the $\ell_2$-normalized projection vector, $\tau$ is the temperature hyperparameter, and all other augmented views of the same class—including those from other domains—are positives (Tan et al., 2023). The geometric effect is explicit: same-class points are pulled together across domains, while all remaining samples in the batch appear in the denominator as negatives.

Task-specific MD-SupContrast systems retain this basic geometry but adapt it to domain structure. In UAV radio-frequency recognition, a projection head $h(\cdot)$ maps fused features $\tilde z$ to $z\in\mathbb R^D$, two random augmentations of each STFT sample form a batch of $2N$ embeddings, and positives are all samples sharing the same label across both texture and positional domains (Gao et al., 18 Aug 2025). In radar target detection, the supervised contrastive term is supplemented by a matching loss that projects weighted shallow features and deep features into a common unit sphere and aligns them bidirectionally; the total loss is $2N$0 (Wang et al., 2024). In multi-source domain adaptation, the contrastive objective is further interpolated under MixUp so that source class-conditional distributions are aligned while target examples are handled through consistency regularization and hard pseudo-labeling (Scalbert et al., 2021).

3. Architectural realizations

In multi-domain text classification, the robust contrastive alignment approach is defined against the private-shared paradigm. The abstract states that advanced baselines use a shared encoder for domain-shared features and a private encoder for each domain, but this becomes inefficient when new domains constantly emerge. The proposed method instead uses exactly two universal feature extractors, denoted $2N$1 and $2N$2, runs every input sentence through both encoders, aligns the paired representations in a joint feature space through supervised contrastive learning, and feeds a combination of $2N$3 and $2N$4 to a single classifier head at inference time (Li et al., 2022). The same abstract does not specify the detailed backbone type, hidden sizes, optimizer, batching strategy, datasets, metrics, or ablation results.

In UAV radio-frequency open-set recognition, the MD-SupContrast encoder consists of three modules. A ResNet-18 branch operating on normalized STFT power spectra $2N$5 extracts texture features. A TransformerEncoder branch extracts time-frequency positional features through two streams, one applied to $2N$6 and one to $2N$7, with preliminary Multi-Layer Nonlinear layers, sinusoidal positional encoding, and $2N$8 stacked TransformerEncoder layers. The resulting vectors $2N$9, $\mathcal{L}^{\mathrm{SupCon}}_i = -\frac{1}{2N_{\tilde y_i}-1} \sum_{j=1}^{2N} \mathds{1}_{i\neq j}\,\mathds{1}_{\tilde y_i=\tilde y_j} \log \frac{ \exp\bigl(\mathrm{sim}(z_i,z_j)/\tau\bigr) }{ \sum_{k=1}^{2N}\mathds{1}_{i\neq k}\exp\bigl(\mathrm{sim}(z_i,z_k)/\tau\bigr) }, \qquad \mathcal{L}^{\mathrm{SupCon}} = \sum_{i=1}^{2N}\mathcal{L}^{\mathrm{SupCon}}_i .$0, and $\mathcal{L}^{\mathrm{SupCon}}_i = -\frac{1}{2N_{\tilde y_i}-1} \sum_{j=1}^{2N} \mathds{1}_{i\neq j}\,\mathds{1}_{\tilde y_i=\tilde y_j} \log \frac{ \exp\bigl(\mathrm{sim}(z_i,z_j)/\tau\bigr) }{ \sum_{k=1}^{2N}\mathds{1}_{i\neq k}\exp\bigl(\mathrm{sim}(z_i,z_k)/\tau\bigr) }, \qquad \mathcal{L}^{\mathrm{SupCon}} = \sum_{i=1}^{2N}\mathcal{L}^{\mathrm{SupCon}}_i .$1 are concatenated as $\mathcal{L}^{\mathrm{SupCon}}_i = -\frac{1}{2N_{\tilde y_i}-1} \sum_{j=1}^{2N} \mathds{1}_{i\neq j}\,\mathds{1}_{\tilde y_i=\tilde y_j} \log \frac{ \exp\bigl(\mathrm{sim}(z_i,z_j)/\tau\bigr) }{ \sum_{k=1}^{2N}\mathds{1}_{i\neq k}\exp\bigl(\mathrm{sim}(z_i,z_k)/\tau\bigr) }, \qquad \mathcal{L}^{\mathrm{SupCon}} = \sum_{i=1}^{2N}\mathcal{L}^{\mathrm{SupCon}}_i .$2 and passed through a final MNL stack to form a unified embedding $\mathcal{L}^{\mathrm{SupCon}}_i = -\frac{1}{2N_{\tilde y_i}-1} \sum_{j=1}^{2N} \mathds{1}_{i\neq j}\,\mathds{1}_{\tilde y_i=\tilde y_j} \log \frac{ \exp\bigl(\mathrm{sim}(z_i,z_j)/\tau\bigr) }{ \sum_{k=1}^{2N}\mathds{1}_{i\neq k}\exp\bigl(\mathrm{sim}(z_i,z_k)/\tau\bigr) }, \qquad \mathcal{L}^{\mathrm{SupCon}} = \sum_{i=1}^{2N}\mathcal{L}^{\mathrm{SupCon}}_i .$3; after supervised contrastive optimization, an improved generative OpenMax algorithm yields the open-set model Open-RFNet (Gao et al., 18 Aug 2025).

In radar target detection, the architecture is deliberately shallow-to-deep. Each raw complex-valued radar echo segment of fixed length is augmented to generate three views, after which the complex modulus is taken to yield a real-valued signal. Six classic statistical features are extracted from the original echo segment: Relative Average Amplitude, Relative Doppler Peak Height, Relative Doppler Vector Entropy, Ridge Integral, Number of Connected Regions, and Maximum Size of Connected Regions. These are normalized across the batch and re-weighted by their Gini-index gain to produce $\mathcal{L}^{\mathrm{SupCon}}_i = -\frac{1}{2N_{\tilde y_i}-1} \sum_{j=1}^{2N} \mathds{1}_{i\neq j}\,\mathds{1}_{\tilde y_i=\tilde y_j} \log \frac{ \exp\bigl(\mathrm{sim}(z_i,z_j)/\tau\bigr) }{ \sum_{k=1}^{2N}\mathds{1}_{i\neq k}\exp\bigl(\mathrm{sim}(z_i,z_k)/\tau\bigr) }, \qquad \mathcal{L}^{\mathrm{SupCon}} = \sum_{i=1}^{2N}\mathcal{L}^{\mathrm{SupCon}}_i .$4. Deep features are obtained with a one-dimensional ResNet-50 encoder $\mathcal{L}^{\mathrm{SupCon}}_i = -\frac{1}{2N_{\tilde y_i}-1} \sum_{j=1}^{2N} \mathds{1}_{i\neq j}\,\mathds{1}_{\tilde y_i=\tilde y_j} \log \frac{ \exp\bigl(\mathrm{sim}(z_i,z_j)/\tau\bigr) }{ \sum_{k=1}^{2N}\mathds{1}_{i\neq k}\exp\bigl(\mathrm{sim}(z_i,z_k)/\tau\bigr) }, \qquad \mathcal{L}^{\mathrm{SupCon}} = \sum_{i=1}^{2N}\mathcal{L}^{\mathrm{SupCon}}_i .$5 and a two-layer MLP projection head $\mathcal{L}^{\mathrm{SupCon}}_i = -\frac{1}{2N_{\tilde y_i}-1} \sum_{j=1}^{2N} \mathds{1}_{i\neq j}\,\mathds{1}_{\tilde y_i=\tilde y_j} \log \frac{ \exp\bigl(\mathrm{sim}(z_i,z_j)/\tau\bigr) }{ \sum_{k=1}^{2N}\mathds{1}_{i\neq k}\exp\bigl(\mathrm{sim}(z_i,z_k)/\tau\bigr) }, \qquad \mathcal{L}^{\mathrm{SupCon}} = \sum_{i=1}^{2N}\mathcal{L}^{\mathrm{SupCon}}_i .$6; the method does not explicitly concatenate shallow and deep features but instead guides the deep encoder by aligning projected shallow and deep embeddings through $\mathcal{L}^{\mathrm{SupCon}}_i = -\frac{1}{2N_{\tilde y_i}-1} \sum_{j=1}^{2N} \mathds{1}_{i\neq j}\,\mathds{1}_{\tilde y_i=\tilde y_j} \log \frac{ \exp\bigl(\mathrm{sim}(z_i,z_j)/\tau\bigr) }{ \sum_{k=1}^{2N}\mathds{1}_{i\neq k}\exp\bigl(\mathrm{sim}(z_i,z_k)/\tau\bigr) }, \qquad \mathcal{L}^{\mathrm{SupCon}} = \sum_{i=1}^{2N}\mathcal{L}^{\mathrm{SupCon}}_i .$7 (Wang et al., 2024).

4. Training regimes and optimization patterns

A prototypical large-scale MD-SupContrast training pipeline is reported for DomainNet. The source dataset is the cleaned train split of DomainNet with six domains—real, clipart, painting, sketch, infograph, and quickdraw—covering the same 345 object classes and totaling 409,832 combined training images. Images are resized to $\mathcal{L}^{\mathrm{SupCon}}_i = -\frac{1}{2N_{\tilde y_i}-1} \sum_{j=1}^{2N} \mathds{1}_{i\neq j}\,\mathds{1}_{\tilde y_i=\tilde y_j} \log \frac{ \exp\bigl(\mathrm{sim}(z_i,z_j)/\tau\bigr) }{ \sum_{k=1}^{2N}\mathds{1}_{i\neq k}\exp\bigl(\mathrm{sim}(z_i,z_k)/\tau\bigr) }, \qquad \mathcal{L}^{\mathrm{SupCon}} = \sum_{i=1}^{2N}\mathcal{L}^{\mathrm{SupCon}}_i .$8, the encoder is ResNet-50, and the projection head is a two-layer MLP with FC $\mathcal{L}^{\mathrm{SupCon}}_i = -\frac{1}{2N_{\tilde y_i}-1} \sum_{j=1}^{2N} \mathds{1}_{i\neq j}\,\mathds{1}_{\tilde y_i=\tilde y_j} \log \frac{ \exp\bigl(\mathrm{sim}(z_i,z_j)/\tau\bigr) }{ \sum_{k=1}^{2N}\mathds{1}_{i\neq k}\exp\bigl(\mathrm{sim}(z_i,z_k)/\tau\bigr) }, \qquad \mathcal{L}^{\mathrm{SupCon}} = \sum_{i=1}^{2N}\mathcal{L}^{\mathrm{SupCon}}_i .$9 ReLU $z_i=h(g(x_i))$0 FC $z_i=h(g(x_i))$1. Pre-training uses SGD with momentum $z_i=h(g(x_i))$2, weight decay $z_i=h(g(x_i))$3, batch size $z_i=h(g(x_i))$4, initial learning rate $z_i=h(g(x_i))$5, linear warmup over the first 10 epochs, learning-rate decays by $z_i=h(g(x_i))$6 at epochs 250 and 350, total training length of 400 epochs, and temperature $z_i=h(g(x_i))$7. Linear evaluation freezes the encoder, discards the projection head, and trains only a linear classifier for 50 epochs (Tan et al., 2023).

The UAV radio-frequency system follows a two-stage regime. Raw I/Q streams are sliced at 3 ms windows, low-power slices below a noise threshold are discarded, the remaining samples are concatenated, and STFT is applied before $z_i=h(g(x_i))$8-scaling and min-max normalization. Training uses batch size $z_i=h(g(x_i))$9, Adam, and a cosine-annealing schedule over epochs. Contrastive pre-training minimizes $\ell_2$0 for 30 epochs; classification fine-tuning attaches a linear head and trains with cross-entropy loss for 10 epochs. The projection head dimension is approximately $\ell_2$1, the temperature is tuned in $\ell_2$2 with $\ell_2$3 given as an example, and augmentations include random time-frequency cropping, amplitude jitter, and small additive noise (Gao et al., 18 Aug 2025).

The radar detector is trained with batch size $\ell_2$4, 100 pre-training epochs, SGD with learning rate $\ell_2$5 and weight decay $\ell_2$6, temperature $\ell_2$7, and alignment weight $\ell_2$8 selected by ablation, with the best result at $\ell_2$9. Validation uses only clutter samples to set the detection threshold for a target false-alarm rate $\tau$0 (Wang et al., 2024).

Adjacent multi-domain contrastive systems broaden the optimization landscape. MDCL inserts two plug-and-play contrastive losses into shared-private backbones: inter-domain semantic alignment on the shared space and intra-domain contrastive clustering on the private or classifier-output space, with no additional model parameters (He et al., 2023). CMSDA combines interpolated supervised contrastive learning on mixed source examples with cross-entropy minimization, target consistency regularization, and hard pseudo-labeling in multi-source unsupervised domain adaptation (Scalbert et al., 2021). SUPMOCO integrates supervised and instance-discriminative contrastive learning through a momentum-updated key encoder and a queue of past representations, and in multi-domain training it uses impure batches drawn from the union of all domains rather than domain-stratified batches (Majumder et al., 2021).

5. Empirical results and documented advantages

The clearest transfer study reports that supervised contrastive learning on the multi-domain DomainNet dataset outperforms a cross-entropy baseline on all seven downstream datasets considered under fixed-feature linear evaluation. Averaged over five runs, the SupCon model reaches a mean score of $\tau$1 versus $\tau$2 for the cross-entropy model, corresponding to an average improvement of $\tau$3. The gains are especially large on Aircraft, where performance rises from $\tau$4 to $\tau$5, and on Flowers102, where it rises from $\tau$6 to $\tau$7 (Tan et al., 2023). The stated interpretation is that cross-domain positives force the network to ignore domain-specific artifacts and focus on shared class-defining features.

In UAV radio-frequency recognition, the reported outcome is $\tau$8 in closed-set recognition and $\tau$9 in open-set recognition under 25 UAV types; the detailed summary specifies 20 known UAV types for the closed-set task and 5 unknown types for the open-set task. The method is reported to outperform ResNet-only, Transformer-only, and cross-entropy training, and the gap between closed-set and open-set accuracy is described as less than $h(\cdot)$0 (Gao et al., 18 Aug 2025). The proposed explanation is that multi-domain fusion balances local texture invariants and global positional patterns, while supervised contrastive learning reduces feature imbalance and sharpens class separation.

In radar target detection, MDFG_SCL is reported to outperform tri-feature and TF-tri-feature shallow detectors, MDCCNN, standard supervised learning, unsupervised contrastive learning, and off-the-shelf supervised contrastive learning. At $h(\cdot)$1, the method is described as yielding mIoU gains of 1–2 percentage points over vanilla supervised contrastive across HH, HV, VH, and VV. In false-alarm-rate comparisons, MDFG_SCL with $h(\cdot)$2 reaches $h(\cdot)$3, compared with $h(\cdot)$4 for SupCon and $h(\cdot)$5 for supervised learning. In an HH-polarization ablation averaged over ten datasets, supervised learning attains $h(\cdot)$6 accuracy, $h(\cdot)$7 recall, and $h(\cdot)$8 mIoU; SupCon reaches $h(\cdot)$9, $\tilde z$0, and $\tilde z$1; MDFG_SCL without Gini weights reaches $\tilde z$2, $\tilde z$3, and $\tilde z$4; and MDFG_SCL with $\tilde z$5 and weighted features reaches $\tilde z$6, $\tilde z$7, and $\tilde z$8 (Wang et al., 2024).

Related multi-domain contrastive evidence points in the same direction under different supervision regimes. MDCL improves several shared-private backbones when annotations are scarce, with gains at 1% labels such as $\tilde z$9 on Amazon, $z\in\mathbb R^D$0 on MNIST-USPS, and $z\in\mathbb R^D$1 on Office-Home; its ablation study shows that combining inter-domain and intra-domain contrastive terms produces the best overall results (He et al., 2023). SUPMOCO reports that multi-domain pre-training on all datasets yields $z\in\mathbb R^D$2 average performance versus $z\in\mathbb R^D$3 for SupCon and that impure mixed-domain batches outperform pure one-domain batches by approximately $z\in\mathbb R^D$4 on average (Majumder et al., 2021). These results do not define MD-SupContrast identically, but they support the broader proposition that supervised contrastive structure can improve transfer and robustness in multi-domain settings.

6. Interpretation, boundaries, and common points of confusion

The cited literature suggests that “multi-domain” should not be treated as a single fixed formalism. In one line of work it denotes datasets composed of several domains sharing the same label space, as in DomainNet pre-training; in another it denotes complementary feature domains inside the model, as in UAV radio-frequency recognition and radar target detection (Tan et al., 2023, Gao et al., 18 Aug 2025, Wang et al., 2024). A plausible implication is that the central invariant is not the specific domain definition but the use of label-aware contrastive attraction across heterogeneous views that would otherwise remain weakly aligned.

MD-SupContrast is also not synonymous with the private-shared paradigm. The text-classification paper explicitly frames its contribution against private-shared methods, arguing that training one private encoder for each domain is inefficient when new domains are constantly emerging and proposing two universal feature extractors instead (Li et al., 2022). Conversely, MDCL remains deliberately compatible with shared-private models and adds inter-domain and intra-domain contrastive losses without extra parameters (He et al., 2023). The design space therefore includes both universal-encoder and shared-private instantiations.

Another common confusion concerns task scope. Open-set recognition, few-shot learning, active learning, and unsupervised domain adaptation are not intrinsic parts of MD-SupContrast; they are extensions that combine supervised contrastive learning with additional mechanisms such as IG-OpenMax, queue-based momentum contrast, Best-vs-Second-Best sampling, or FixMatch-style consistency regularization (Gao et al., 18 Aug 2025, Majumder et al., 2021, He et al., 2023, Scalbert et al., 2021). Similarly, not every implementation concatenates heterogeneous features: the radar detector uses alignment rather than explicit concatenation (Wang et al., 2024).

Finally, the evidentiary depth is uneven across the literature summarized here. For the 2022 text-classification contribution, the available abstract states the high-level motivation, the use of supervised contrastive alignment, the replacement of private-shared modeling by two universal feature extractors, and the claim of state-of-the-art parity or improvement, but it does not specify the exact loss formula, batch-construction strategy, optimizer, hyperparameters, named datasets, evaluation metrics, or ablation studies (Li et al., 2022). By contrast, the UAV, radar, DomainNet, CMSDA, MDCL, and SUPMOCO reports provide explicit objectives, training details, and quantitative comparisons, making them the main sources for the technical characterization of MD-SupContrast as a research area (Gao et al., 18 Aug 2025, Wang et al., 2024, Tan et al., 2023, Scalbert et al., 2021, He et al., 2023, Majumder et al., 2021).