Paper deep dive
SGN: A Similarity-based Generative Network for Data Generation under Distribution Shift
Jiaqi Zhu, Xincheng Chen, Yuncheng Wu, Zhaojing Luo, Beng Chin Ooi
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 91%
Last extracted: 7/21/2026, 6:14:28 AM
Summary
The paper introduces the Similarity-based Generative Network (SGN), a reusable framework for generating data in target domains with distribution shifts without parameter updates. SGN learns a latent space structured by label-induced pairwise similarities (zero-one matrix) using an encoder-decoder architecture. It allows generation of target-specific samples by encoding a small labeled representative set from the target domain and combining them in the learned latent space, maintaining class consistency.
Entities (5)
Relation Signals (5)
SGN → addresses → Distribution Shift
confidence 95% · SGN is proposed to mitigate the distribution shift problem where data distribution in the source domain is different from that in the target domain.
SGN → uses → Latent Space
confidence 92% · SGN learns a latent space structured by label-induced pairwise similarities.
SGN → implements → Encoder-Decoder Architecture
confidence 90% · SGN preserves reconstructive information through an encoder-decoder architecture.
SGN → defines → Zero-One Similarity Matrix
confidence 88% · We therefore design a zero-one similarity matrix: the similarity score of two samples is one if they have the same label, and zero otherwise.
SGN → enables → Target-Guided Data Augmentation
confidence 85% · Experiments ... demonstrate the effectiveness of SGN for target-guided data augmentation under source-to-target distribution shifts.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Generative models trained on a source domain often produce samples that are poorly aligned with shifted target domains, limiting their effectiveness for target-domain data augmentation. Although target-specific adaptation can reduce this mismatch, it typically requires additional optimization and domain-specific parameters. We propose a Similarity-based Generative Network (SGN), a reusable framework that is trained once on labeled source data and applied to new target domains without parameter updates. SGN learns a latent space structured by label-induced pairwise similarities while preserving reconstructive information through an encoder-decoder architecture. At generation time, a small labeled representative set from the target domain is encoded and combined in the learned latent space, allowing the generated samples to inherit target-specific characteristics while maintaining class consistency. We further analyze the realizability and dimensionality requirements of the proposed similarity structure. Experiments on image and tabular datasets demonstrate the effectiveness of SGN for target-guided data augmentation under source-to-target distribution shifts.
Tags
Links
- Source: https://arxiv.org/abs/2607.18072v1
- Canonical: https://arxiv.org/abs/2607.18072v1
PDF not stored locally. Use the link above to view on the source site.
Full Text
54,089 characters extracted from source content.
Expand or collapse full text
SGN: A Similarity-based Generative Network for Data Generation under Distribution Shift Jiaqi Zhu National University of Singapore jiaqi77@nus.edu.sg &Xincheng Chen National University of Singapore e0838447@u.nus.edu.sg &Yuncheng Wu Renmin University of China wuyuncheng@ruc.edu.cn &Zhaojing Luo Beijing Institute of Technology zjluo@bit.edu.cn &Beng Chin Ooi Zhejiang University ooibc@zju.edu.cn Abstract Generative models trained on a source domain often produce samples that are poorly aligned with shifted target domains, limiting their effectiveness for target-domain data augmentation. Although target-specific adaptation can reduce this mismatch, it typically requires additional optimization and domain-specific parameters. We propose a Similarity-based Generative Network (SGN), a reusable framework that is trained once on labeled source data and applied to new target domains without parameter updates. SGN learns a latent space structured by label-induced pairwise similarities while preserving reconstructive information through an encoder-decoder architecture. At generation time, a small labeled representative set from the target domain is encoded and combined in the learned latent space, allowing the generated samples to inherit target-specific characteristics while maintaining class consistency. We further analyze the realizability and dimensionality requirements of the proposed similarity structure. Experiments on image and tabular datasets demonstrate the effectiveness of SGN for target-guided data augmentation under source-to-target distribution shifts. 1 Introduction Generative models (GMs) provide a principled way to expand limited datasets by synthesizing new samples that preserve the structure of observed data. Representative paradigms, such as generative adversarial networks (GANs) (Goodfellow et al., 2014), variational autoencoders (VAEs) (Kingma and Welling, 2014), and more recently, diffusion-based models (Ho et al., 2020; Yang et al., 2024), have attracted increasing attention for data generation in various applications (Deldjoo et al., 2021; Zhu et al., 2026; Yi et al., 2017; Shaham et al., 2019). Despite their architectural differences, these methods are typically trained to approximate the data distribution observed during training. Conditional generative models further incorporate class labels, semantic attributes, or domain information to control the characteristics of generated samples (Mirza and Osindero, 2014; Miyato and Koyama, 2018; Chou et al., 2023; Zhu et al., 2025). When the training and deployment data follow similar distributions, such models can generate realistic and useful samples for data augmentation and downstream prediction. However, in many practical settings, a generative model trained on one source domain is reused across multiple target domains with different data distributions. Such shifts may arise from changes in acquisition devices, environments, populations, data-collection protocols, or class proportions (Quiñonero-Candela et al., 2008; Zhu et al., 2023). As a result, samples generated from the source distribution may be poorly aligned with the target domain and provide limited, or even negative, utility for target-domain augmentation. Although target-specific fine-tuning can mitigate this mismatch, it is often costly and unreliable when only limited target data are available. We therefore consider a reusable generation setting in which a model is trained once on a labeled source dataset and applied to multiple target domains without parameter updates, using only a small labeled representative set at generation time. The key challenge is to transfer class semantics learned from the source domain while preserving target-specific characteristics. Existing approaches only partially address this challenge. Source-trained conditional generators use class labels or semantic attributes to control generation, but the learned class-conditional distributions remain largely tied to the source domain (Miyato and Koyama, 2018; Hou et al., 2022a; Peebles and Xie, 2023; Sadat et al., 2025). Target-specific adaptation methods can align a pretrained generator with a new domain through fine-tuning or additional optimization, but require maintaining or updating a separate model for each target domain (Wang et al., 2020; Bahram et al., 2025). Sample-based augmentation methods, such as input-space mixing and latent interpolation, can directly incorporate target-domain characteristics from limited observations, yet they typically lack an explicitly learned semantic structure to preserve class consistency during generation (Zhang et al., 2018; Verma et al., 2019; Nguyen et al., 2026). These limitations motivate a reusable framework that combines transferable class structure with target-domain information during generation. In this paper, we propose a novel evaluation mechanism that forms a basis for an alternative paradigm for GMs. Specifically, we consider a generated sample to satisfy the evaluation criteria if its similarity scores with other data samples are close to the similarity scores of original data samples; otherwise, the quality of the generated data is poor. The rationale behind this mechanism is that the original data points definitely satisfy the evaluation criteria; therefore, if a generated data point results in the same performance as the original data points, it can be considered a satisfactory sample. The key to this mechanism is to develop an appropriate similarity measurement, which is challenging. If the measurement is tight, the generated samples will only be accepted if they are extremely close to the original data points. On the contrary, if the measurement is loose, the generated samples may not be acceptable even if they satisfy the evaluation criteria. To address this challenge, we utilize data labels to construct the similarity measurement because data labels provide a natural balance between the generalization and the specification of measurement, where sufficient numbers of data are categorized into each label. We therefore design a zero-one similarity matrix: the similarity score of two samples is one if they have the same label, and zero otherwise. Based on the evaluation mechanism, we propose a new generative network, called similarity-based generative network (SGN). Specifically, our evaluation mechanism establishes a latent space where each data sample has a set of colinear latent representations and a unique special latent representation. The similarity score between any two data points is reflected by the dot product of their special latent representations. Consequently, we train a neural network model to find the latent space and also the transformation between the latent and the original spaces. Essentially, SGN learns the special latent representations of the data points in the training dataset without being aware of its distribution. After SGN is trained, we utilize these latent representations to generate desired samples according to the need of downstream tasks. Compared to existing GMs, SGN is more amenable to generating samples that follow various distributions and is capable of addressing the distribution shift problem, where the data distribution in the source domain is different from that in the target domain. Mathematically, SGN can be viewed as a conditional generative paradigm. However, unlike existing conditional generative works that utilize the label information explicitly and represent each label with an embedding, SGN implicitly adopts label information, which could better alleviate the distribution shift problem because it provides a stable label-data relationship. In particular, in our design, the inner product of any two data representations always satisfies the same zero-one similarity requirement across different domains. In existing works, a data sample is constantly represented as a one-hot label representation, while the data representations vary across different domains.Thus, the relationship between the label and data representations exhibits significant variations across different domains. Our contributions are summarized as follows: • We propose a novel evaluation mechanism based on similarity measurements between different data points. To the best of our knowledge, this is the first work that designs a distribution-agnostic evaluation mechanism for GMs. • We present a similarity-based generative network SGN to generate desirable samples for target domains, mitigating the distribution shift problem. We further analyze the lower bound of the latent space dimension and provide the convergence analysis of SGN training. • We conduct extensive experiments on eleven real-world datasets including both image and tabular datasets, demonstrating the effectiveness of SGN for generating high-quality data and mitigating the distribution shift problem, compared to six state-of-the-art baselines. 2 Related Work Conditional Generative Model. Conditional generative models synthesize data according to auxiliary conditions, such as class labels, semantic attributes, text descriptions, or domain information. Early representative approaches include conditional generative adversarial networks (CGANs) (Mirza and Osindero, 2014), conditional variational autoencoders (CVAEs) (Sohn et al., 2015), and projection-based conditional GANs (Miyato and Koyama, 2018). Subsequent studies improved class-conditional generation through contrastive objectives, auxiliary classifiers, and more expressive conditioning mechanisms (Kang and Park, 2020; Kang et al., 2021). More recently, diffusion models have become a prevalent paradigm for conditional generation. Classifier and classifier-free guidance enable pretrained diffusion models to generate samples aligned with class labels, textual descriptions, or other conditions (Dhariwal and Nichol, 2021; Ho and Salimans, 2022). Recent studies further investigate the quality-diversity trade-off of guidance and develop more flexible conditioning mechanisms (Sadat et al., 2025; Li et al., 2025). Nevertheless, the conditional distributions learned by these models remain primarily shaped by their training data. When a generator trained on a source domain is directly applied to a shifted target domain, specifying the desired class does not by itself capture how the class-conditional data characteristics change across domains. Target-specific Generative Adaptation. Another line of research adapts a pretrained generator to a target domain using a limited number of target examples. Earlier GAN-based methods transfer source-domain generators through selective fine-tuning, regularization, or cross-domain correspondence preservation Wang et al. (2020); Bahram et al. (2025). Similar ideas have been extended to diffusion models, where a pretrained model is adapted using target-domain fine-tuning, domain-specific regularization, parameter-efficient updates, or additional guidance objectives Zhu et al. (2022). For example, Domain Guidance (Zhong et al., 2025) interprets diffusion-model transfer as conditional generation and introduces a target-domain guidance mechanism, whereas SaRA (Hu et al., 2025) performs parameter-efficient adaptation through sparse low-rank updates. These approaches can produce high-quality samples in a new domain, but still involve target-specific optimization and typically require storing either a separately adapted model or domain-specific parameters. In contrast, SGN keeps the source-trained model fixed and incorporates target-domain characteristics through representative samples only during generation. Sample-based Data Augmentation. Sample-based augmentation constructs new training examples directly from observed samples or their representations. Classical methods include Mixup (Zhang et al., 2018), which linearly combines training examples and labels, and Manifold Mixup (Verma et al., 2019), which performs interpolation in hidden representation spaces. Recent diffusion-based augmentation methods leverage strong pretrained generative priors to improve the fidelity and diversity of synthetic samples. Trabucco et al. (2024) edits real images using pretrained text-to-image models, while Wang and Chen (2025) performs class-wise interpolation between diffusion inversions and reconstructs the interpolated representations through a two-stage denoising process. More recent work studies which synthetic samples are most useful rather than generating them indiscriminately. In particular, targeted image augmentation selectively allocates diffusion-based augmentation to training examples for which synthetic data are expected to provide greater utility (Nguyen et al., 2026). These approaches demonstrate the effectiveness of sample-conditioned and representation-space augmentation. However, conventional interpolation rules are generally defined by geometric proximity in the input or latent space. Although modern pretrained representations may already encode semantic information, they do not explicitly impose a transferable pairwise class geometry learned from the source domain. SGN instead organizes the latent space using label-induced similarity constraints and uses this structure to guide the combination of limited target-domain representations without updating the model parameters. 3 Similarity-Based Generative Networks In this section, we present our similarity-based generative networks SGN that can generate desirable samples for the need of target domains. Generally speaking, SGN isolates the source domain training and target domain generation, where SGN is first trained in the source domain without knowing the information of the target domain, and then SGN generates required data samples of the target domain. Specifically, we first propose the similarity-based evaluation mechanism and then analyze the latent space it determines in Section 3.1. Subsequently, we describe the training paradigm of SGN in Section 3.2, and present the convergence analysis of the training process in Section 3.3. Finally, we describe how to generate desirable samples in the target domain in Section 3.4. We consider a labeled dataset =(i,yi)|i∈1,⋯,NX=\( x_i,y_i)|i∈\1,·s,N\\, where N is the number of samples in X. Let K be the number of distinct labels and =1,⋯,K a=\ a_1,·s, a_K\ be the label value set. Moreover, we denote the ithi^th original data and its latent representations as iori x_i^ori and ilat x_i^lat, respectively. 3.1 Similarity-Based Evaluation Mechanism We utilize the similarity score between samples to construct the evaluation mechanism, where the similarity scores are determined by data labels. Specifically, we require the similarity score of two samples to be one if they have the same label, and otherwise zero. Let S denote the zero-one similarity matrix defined by the evaluation mechanism. In essence, this similarity-based evaluation mechanism determines a latent linear space. In this latent space, we can utilize the inner product of two data points’ representations as their similarity score, where each data point has a special latent representation and a set of other latent representations. The inner product of the special latent representations is equivalent to the similarity score, and all latent representations for one data sample are co-linear. Formally, given a latent space with D dimensions, S is encoded through i,j=⟨ilat_S,jlat_S⟩=1yi=yj0yi≠yj, S_i,j= x_i^lat\_S, x_j^lat\_S = cases1&y_i=y_j\\ 0&y_i≠ y_j, cases (1) where ilat_S x_i^lat\_S and jlat_S x_j^lat\_S are the special latent representations of data points i and j, respectively, and ⟨ilat_S,jlat_S⟩ x_i^lat\_S, x_j^lat\_S is the inner product of ilat_S x_i^lat\_S and jlat_S x_j^lat\_S. Figure 1: An illustration of the SGN paradigm, which consists of two stages. In the training stage, the encoder and decoder learn the transformation between the original data space and latent space, where the latent space satisfies the similarity requirement defined in Equation 1. In the data generation stage, the encoder takes two data points with the same label and produces two latent representations; then, the decoder generates a data point by a linear combination of these latent representations as inputs. The above design has two properties. First, it ensures that the latent representations of different labels are orthogonal, i.e., latent representations of all labels essentially form an orthogonal set of the latent space, and each sample’s latent representation is a linear combination of its label’s corresponding orthogonal basis. Let =d|d∈1,…,D V=\ V_d|d∈\1,...,D\\ denote the aforementioned orthogonal set of the latent space, where different labels have different d V_d and each label may have multiple orthogonal basis. Let mkm_k be the number of orthogonal basis for the kthk^th label k a_k, then D=∑k=1KmkD= _k=1^Km_k. We use an indicator γd,k _d,k to represent the relationship between d V_d and k a_k: if d V_d is an orthogonal basis of k a_k, then γd,k=1 _d,k=1; otherwise γd,k=0 _d,k=0. As a consequence, ilat x_i^lat could be represented by V as ilat=∑d=1D∑k=1Kβi,dαi,kγd,kd, x_i^lat= _d=1^D _k=1^K _i,d _i,k _d,k V_d, (2) where βi,d _i,d is the linear coefficient, and αi,k _i,k indicates whether yiy_i equals to the label value k a_k, i.e., if k=yi a_k=y_i, αi,k=1 _i,k=1; otherwise, αi,k=0 _i,k=0. Second, the co-linear setting supports a flexible generative paradigm design, enabling the special latent representation to be set in a variety of ways. For example, we could use the normalized representation as the special latent representation due to the co-linear setting. Another option is to adopt a neural network to infer the special latent representation, where the network’s input is the original representation and the output can be set to the special latent representation. 3.2 SGN Training Notably, SGN refers to a family of deep neural networks trained based on the similarity-based evaluation mechanism. It should satisfy two requirements: (1) find the latent space determined by the similarity-based evaluation mechanism, i.e., the inner product of any two data points’ special latent representations is equivalent to the similarity of these two data points; (2) imitate the transformation between the latent space and the original space, i.e., given a latent representation of a data point, SGN could obtain its original representation. Therefore, the loss function is defined as follows: ℒ=ηN2∑i=1N∑j=1Nℒi,j(r1)+1−ηN∑i=1Nℒi(r2),L= ηN^2 _i=1^N _j=1^NL_i,j^(r1)+ 1-ηN _i=1^NL_i^(r2), (3) where ℒi,j(r1)L_i,j^(r1) corresponds to the regularization loss, which is the first requirement for any two data points, and ℒi(r2)L_i^(r2) corresponds to the reconstruction loss, which is the second requirement for any specific training data point. η∈[0,1]η∈[0,1] is a trade-off parameter which represents the relative importance between two requirements. Figure 1 illustrates the general pipeline of SGN, which follows the autoencoder structure: iori x_i^ori and ilat x_i^lat are the input and output of the encoder, respectively. The input of the decoder is a scalar multiplication of ilat x_i^lat that supports co-linear relationships among a data point’s latent representations, i.e., cilat‖ilat‖c x_i^lat\| x_i^lat\|, where c is sampled from the uniform distribution U(‖ilat‖−C,‖ilat‖+C)U(\| x_i^lat\|-C,\| x_i^lat\|+C). The constant C is the maximum length variation of the latent representation. Let irec x_i^rec denote the output of the decoder. Then the reconstruction loss ℒi(r2)L_i^(r2) can be defined as ℒi(r2)=‖irec−iori‖L_i^(r2)=\| x_i^rec- x_i^ori\|. For ℒi,j(r1)L_i,j^(r1), we define the special latent representation as follows: ℒi,j(r1)=|⟨f(ilat),f(jlat)⟩|yi≠yj|⟨f(ilat),f(jlat)⟩−1|yi=yj,L_i,j^(r1)= cases | f( x_i^lat),f( x_j^lat) |&y_i≠ y_j\\ | f( x_i^lat),f( x_j^lat) -1 |&y_i=y_j, cases (4) where f(⋅)f(·) is a post-processing function of the encoder’s output and ilat_S=f(ilat) x_i^lat\_S=f( x_i^lat). For example, it can be the identity function f()f( x) or the normalized function f()=‖f( x)= x\| x\|. We shall discuss the choice of this function for different types of datasets in Section 4. Determining the dimension of the special latent representation. Since the number of samples N is usually large, we utilize a batch of samples to calculate ℒL in each epoch. Let D be the dimension of the latent space. The batch size can be used to derive the lower bound of D. Let bX_b be a batch of the training data and B be the batch size. Suppose bX_b has Q different label values (Q≤KQ≤ K), and Nq(b)N_q^(b) is the number of data points with label value q a_q in bX_b, and apparently ∑q=1QNq(b)=B _q=1^QN_q^(b)=B. Moreover, to facilitate our analysis, we assume data in bX_b has different special latent representations. 111The situation we consider here is the worst case. If we have more different special latent representations in one batch, it would be more difficult to make ∑i=1B∑j=1Bℒi,j(r1) _i=1^B _j=1^BL_i,j^(r1) = 0. Therefore, we consider the worst case for training, where all the data in this batch have different special latent representations. If SGN is trained close to the optimal solution, according to Equation 1, for any other data sample whose special latent representation is otherlat_S x_other^lat\_S and label is yothery_other, it should satisfy the following conditions established by the data in bX_b, ⟨otherlat_S,ilat_S⟩=1,ifyi=yother⟨otherlat_S,ilat_S⟩=0,ifyi≠yother. cases x_other^lat\_S, x_i^lat\_S =1,if\,y_i=y_other\\ x_other^lat\_S, x_i^lat\_S =0,if\,y_i≠ y_other. cases (5) Apparently, the resulted coefficient matrix of Equation 5 is the special latent representations of bX_b, denoted as lat_SX^lat\_S. Let qlat_SX_q^lat\_S be the subset of lat_SX^lat\_S, which consists of special latent representations with label value q a_q. In essence, the dimension of latent representations is the length of variables in the equation system, which can be determined by the relationship between the rank of coefficient matrix and the number of independent solutions. Propositions 3.1 and 3.2 give the lower bound of the rank of a coefficient sub-matrix and the lower bound of the rank of the whole coefficient matrix, respectively. Based on which, we decide the dimension requirement of latent representations in Theorem 3.3. Proposition 3.1. If the dimension D is sufficiently large, then r(qlat_S)≥Nq(b)−1r(X_q^lat\_S)≥ N_q^(b)-1, where r(qlat_S)r(X_q^lat\_S) is the rank of qlat_SX_q^lat\_S. Proof. Since D is sufficiently large, the main idea of the proof is to construct a group of linearly independent vectors from qlat_SX_q^lat\_S, and the number of vectors in the constructed group is Nq(b)−1N_q^(b)-1. We can construct the group by randomly selecting a special latent representation ilat_S x^lat\_S_i from qlat_SX_q^lat\_S and generating the constructed group by qlat_S−ilat_SX_q^lat\_S- x^lat\_S_i, which has Nq(b)−1N_q^(b)-1 non-zero vectors. Note that the non-zero vectors in the constructed group are pairwise orthogonal, because for any two non-zero vectors jlat_S−ilat_S x^lat\_S_j- x^lat\_S_i and hlat_S−ilat_S x^lat\_S_h- x^lat\_S_i in the constructed group satisfy ⟨jlat_S−ilat_S,hlat_S−ilat_S⟩ x^lat\_S_j- x^lat\_S_i, x^lat\_S_h- x^lat\_S_i = 0. Hence, these vectors are linear independent. ∎ Proposition 3.2. If the dimension D is sufficiently large, then r(lat_S)≥B−Qr(X^lat\_S)≥ B-Q. Proof. It is apparent that any two vectors from different constructed groups are orthogonal. Therefore, these constructed groups form a large group with ∑q=1QNq(b)−Q=B−Q _q=1^QN_q^(b)-Q=B-Q independent vectors. ∎ Theorem 3.3. The dimension of latent representations needs to satisfy: D≥B+K−2QD≥ B+K-2Q. Proof. According to Proposition 3.2, there exist at most D−B+QD-B+Q numbers of independent solutions to Equation (1). Moreover, we have Q≤KQ≤ K, and thus K−QK-Q label values are not included in bX_b, which means there must exist at least K−QK-Q independent solutions to Equation (1). Therefore, we can obtain: D−B+Q≥K−QD-B+Q≥ K-Q. ∎ 3.3 Convergence Analysis Recall that any design of ℒi,j(r1)L_i,j^(r1) and ℒi(r2)L_i^(r2) is proposed to satisfy the aforementioned two requirements, and thus the global optimality essentially satisfies the two requirements, as stated in theorem˜3.4. Theorem 3.4. lat_S∗(lat_S∗)T=X^lat\_S^* (X^lat\_S^* )^T= S, where lat_S∗X^lat\_S^* represents the optimal solution of special latent representations lat_SX^lat\_S. Based on the global optimality, we could then analyze the convergence of SGN training. Assume that the original data representation space is discrete, we can then enumerate all the data points in such space. Accordingly, we can obtain their special latent representations, i.e., Wlat_SX_W^lat\_S. Thus, we have the following theorem. Theorem 3.5. There always exists a well-trained SGN such that Wlat_S∗(Wlat_S∗)T=WX_W^lat\_S^* (X_W^lat\_S^* )^T= S_W, where W S_W is the similarity among the enumerated data points and Wlat_S∗X_W^lat\_S^* is the optimal solution of Wlat_SX_W^lat\_S. Proof. We can transform the problem into: for any W S_W, there exists a matrix X that satisfies ()T= X( X)^T= S_W. The reason is that S_ W is a real symmetric matrix, such X will always exist. Consequently, if SGN has enough model capacity, we can finally find such a latent space. ∎ 3.4 Data Generation In this section, we present how to generate a data point according to the trained SGN and a representative dataset RepX_Rep in the target domain. Algorithm 1 summarizes the data generation process. First, we decide the label value k a_k of the generated sample. Subsequently, we select a set of NkN_k data points from RepX_Rep whose labels equal k a_k, and obtain equal=(i,yi=k)|i∈1,…,NkX_equal=\( x_i,y_i= a_k)|i∈\1,...,N_k\\; similarly, we select another NkN_k data points whose labels are not k a_k to construct unequal=(i,yi≠k)|i∈1,…,NkX_unequal=\( x_i,y_i≠ a_k)|i∈\1,...,N_k\\. Let select=equal∪unequalX_select=X_equal _unequal be the selected dataset. Next, we construct a latent representation newlat x_new^lat that satisfies Equation (1) for the selected dataset selectX_select. Notice that there exist multiple ways to construct newlat x_new^lat, i.e., solving the equation system Equation (1). In the experiments, we utilize a simple but effective way that linearly mixes the data points in equalX_equal. After that, we input newlat x_new^lat to the trained SGN, and the output is the generated data point new x_new. Finally, we can obtain a generated data point (new,ynew)( x_new,y_new), where ynew=ky_new= a_k. A noteworthy aspect is that our data generation method can address the distribution shift problem. Specifically, the way of generating data samples should meet the needed distribution of target domain, which is challenging because of the existence of the distribution shift. Typically, the data distribution p(x)p(x) is determined by the label distribution p(y)p(y) and the label-conditioned distribution p(x|y)p(x|y), as shown in the Equation (6), where different p(y)p(y) (i.e., prior probability shift) and p(x|y)p(x|y) (i.e., covariate shift) could lead to different p(x)p(x) (i.e., distribution shift). p(x)=∑k=1Kp(x|y=k)p(y=k).p(x)= _k=1^Kp(x|y= a_k)p(y= a_k). (6) Algorithm 1 Data generation based on the trained SGN model Input: The representative dataset RepX_Rep, the trained SGN, the number of data NkN_k to be sampled. Output: Generated data: (new,ynew) ( x_new,y_new ). Step 1: Select a label value k a_k. Step 2: Select datasets equal=(i,yi=k)|i∈1,…,NkX_equal=\( x_i,y_i= a_k)|i∈\1,...,N_k\\ and unequal=(i,yi≠k)|i∈1,…,NkX_unequal=\( x_i,y_i [rgb]0,0,0≠ a_k)|i∈\1,...,N_k\\ from RepX_Rep, and set select=equal∪unequalX_select=X_equal _unequal. Step 3: Compute a latent representation newlat x_new^lat that satisfies Equation (1) for selectX_select. Step 4: Input newlat x_new^lat to the trained SGN, and the output is the generated data new x_new. Step 5: Set ynew=ky_new= a_k and return (new,ynew) ( x_new,y_new ). In SGN, since we search for a latent space instead of learning the distribution during the training stage, we can use the needed distribution in the target domain to generate the data samples. In particular, in Step 1 and Step 2 of Algorithm 1, we can sample the label value k a_k and the equalX_equal dataset according to p(y)p(y) and p(x|y=k)p(x|y= a_k), respectively, where p(y)p(y) and p(x|y=k)p(x|y= a_k) are determined by the representative dataset RepX_Rep. This way, we can tackle the prior probability shift and covariate shift to mitigate the distribution shift problem. Note that in the training stage, the neural network learns the reflection between the data original representations and latent representations. When moving into the target domain, we use the learned reflection to calculate the latent representations of the data used in the target domain. These latent representations inherently consist of the distribution information of the target domain. With the mixture of the latent representations, we can make the generated data satisfy the target domain distribution. 4 Experiments 4.1 Experimental Setup We evaluate SGN on both image datasets and tabular datasets. For image datasets, we use two implementations: one is NeurSpec whose f(⋅)f(·) is an identity function, and the other is NormSpec whose f(⋅)f(·) is the normalized function. For the implementation on tabular datasets, we employ f(⋅)f(·) as the normalized function simply denoted SGN. All the experiments are run on a server equipped with an I9-11900K CPU and GEFORCE RTX 3090 Ti HOF OC LAB Edition * 2 GPUs. Datasets. We use six image datasets and five tabular datasets for the evaluation. For image datasets, we use BloodMNIST (Acevedo et al., 2020), OctMNIST (Kermany et al., 2018), TissueMNIST (Woloshuk et al., 2021), RetinaMNIST (DeepDRiD, 2020), SVHN (Netzer et al., 2011) and FashionMNIST (Xiao et al., 2017). We create different domains via the dataset rotation (Nguyen et al., 2021). For tabular datasets, we utilize Insect (de Souza et al., 2020), Adult (Kohavi, 1996), Avila (Stefano et al., 2011), Dry Bean (Koklu and Özkan, 2020), and Electrical Grid Stability (Arzamasov et al., 2018) datasets. For the Insect dataset, we utilize sub-datasets that already contain distribution shifts, namely Insects_\_Abr (denoted InsectsA), Insects_\_Incr (denoted InsectsI), Insects_\_IncrGrd (denoted InsectsIG). For the other four datasets, we divide each of them into two subsets to create distribution shifts. The details are explained in Appendix A. Baselines. For image datasets, we compare SGN with four state-of-the-art CGMs, namely, ProjGAN (Miyato and Koyama, 2018), ContraGAN (Kang and Park, 2020), ReACGAN (Kang et al., 2021), and ADCGAN (Hou et al., 2022b). We adopt the studioGAN library to implement these baseline models. The ProjGAN’s architecture is based on ResNet, and other baselines’ architectures are based on bigGAN. For tabular datasets, we compare SGN with the two conditional tabular dataset generation methods: CTGAN and TVAE (Xu et al., 2019), where we adopt the SDV library to implement them with the defaulted structures stated in the library. Metric. For image datasets, we measure the effectiveness of generated data in the target domain by the downstream classification tasks. The generated data is considered effective if it can improve the classifier’s performance. For tabular datasets, we measure the generated data quality by the benchmark provided by (Xu et al., 2019). We train regressor/classifier models on the generated data and test the models on the target domain. We use Linear Regression (LR) and Multilayer Perceptron (MLP) regression as regressors, and Decision Tree (DT) and MLP classification as classifiers. We utilize the F1 score and R2R^2 to evaluate the classifiers and regressors, respectively. More details on the experimental setup are provided in Appendix B. Table 1: Classification accuracy Methods BloodMNIST OctMNIST TissueMNIST RetinaMNIST SVHN FashionMNIST Original 0.63 ± 0.01 0.33 ± 0.03 0.38 ± 0.03 0.44 ± 0.03 0.47 ± 0.05 0.41 ± 0.03 ProjGAN 0.67 ± 0.02 0.32 ± 0.02 0.35 ± 0.03 0.40 ± 0.04 0.49 ± 0.02 0.42 ± 0.01 ContraGAN 0.68 ± 0.02 0.27 ± 0.03 0.39 ± 0.02 0.41 ± 0.04 0.45 ± 0.01 0.40 ± 0.02 ReACGAN 0.65 ± 0.03 0.31 ± 0.03 0.40 ± 0.02 0.42 ± 0.03 0.52 ± 0.03 0.45 ± 0.03 ADCGAN 0.68 ± 0.02 0.34 ± 0.04 0.42 ± 0.03 0.44 ± 0.03 0.51 ± 0.02 0.43 ± 0.03 NormSpec 0.68 ± 0.01 0.41 ± 0.03 0.42 ± 0.02 0.47 ± 0.03 0.52 ± 0.02 0.49 ± 0.03 NeurSpec 0.75 ± 0.01 0.43 ± 0.03 0.44 ± 0.03 0.50 ± 0.03 0.55 ± 0.01 0.53 ± 0.02 Figure 2: Visualization results of the RetinaMNIST and TissueMNIST datasets. The Original Data 1 and Original Data 2 are in the source domain for training the SGN. The Rotated Data 1 and Rotated Data 2 are in the target domain. The Generated Data is obtained in the data generation stage. 4.2 Evaluation on Image Datasets Downstream task evaluation. In this paper, we conduct classification as the downstream tasks, and the classifier is ResNet50 (He et al., 2016). The target domain is created by rotating the original dataset for 90 degrees. For each generative paradigm, we generate 200 images per label, the same number as that in the rotated training dataset. Table 1 summarizes the classification results over three runs evaluated on five generative paradigms. Note that the ‘Original’ column refers to the classification accuracy directly on the rotated training dataset (without generated data). The accuracy is low due to the limited number of training images. There are three main observations from Table 1. First, the new data generated by the baselines, i.e., ProjGAN, ContraGAN, RcACGAN, and ADCGAN, is ineffective for downstream tasks in the target domain. Specifically, except for the BloodMNIST dataset, the performances of the classifiers trained with generated data on the other five datasets are similar to or even lower than those trained without generated data. This is because the baselines are trained in the source domain, i.e., they learn the source domain’s distributions for generating new data. When the distribution shifts exist between the source domain and target domain, the generated new data is not helpful to the classification tasks in the target domain. For the BloodMNIST dataset, the four baselines can improve the classifiers’ performances because the key features of the label information are more robust. Note that BloodMNIST results in the highest ‘Original’ classification accuracy; thus, data rotation has little impact on the key features. Second, the data generated by SGN greatly improves the classification accuracy of the downstream tasks w.r.t. the six datasets. Even for the BloodMNIST dataset, where all the generative paradigms can improve the accuracy, SGN still outperforms the baselines. The reason is that SGN essentially learns a latent space during training and incorporates the target domain’s distribution into the data generation process. Therefore, it can generate data that follows the target domain’s distribution, leading to better performance. Third, the performance of the SGN’s NeurSpec realization is generally better than that of the NormSpec realization. For example, NeurSpec achieves about 2%∼6%2\% 6\% accuracy improvement compared to NormSpec. The reason is that NeurSpec employs more orthogonal bases to construct the latent representations, and thus, it has a higher capability to provide high-quality latent representations for generating better data. Visualization results. Figure 2 shows the visualization results of data generation in NeurSpec as it provides better accuracy on the downstream classification tasks. Figure 2 consists of three parts: original data, rotated data, and generated data, where the original data is used for NeurSpec training, the rotated data is the input of generation, and the generated data is the output. Moreover, the original data is in the source domain, while the rotated and generated data are in the target domain. We can observe from Figure 2 that, even though NeurSpec is not trained in the target domain, the generated data is similar to but different from the rotated data, demonstrating that NeurSpec is able to generate acceptable data by the linear combination of rotated data. For example, the third generated image of TissueMNIST, i.e., the fifth image in the fifth row highlighted in the red rectangle, consists of two circles, where the yellow circle inherits from rotated 1, and the blue circle inherits from rotated data 2. More visualization results of the generated data are provided in Appendix C. Table 2: Benchmark results (R2R^2 for regressors) on tabular datasets Dataset Source → Target SGN CTGAN TVAE LR MLP LR MLP LR MLP Insect InsectA → InsectI 0.5028 0.0422 -0.8304 -0.2169 0.4923 -0.2887 InsectA → InsectIG 0.4620 -0.1132 -1.3399 -0.4884 -0.0043 -0.9879 InsectA → InsectIR 0.4040 -0.0600 -0.8327 -0.2939 0.2641 -0.2887 InsectI → InsectA 0.5055 -0.0525 -0.0950 -0.4944 0.4964 -0.0928 InsectI → InsectIG 0.5108 -0.0984 -0.9219 -0.9173 0.2754 -0.1894 InsectI → InsectIR 0.5153 -0.1894 -0.2775 -0.5936 0.4581 -0.1164 InsectIG → InsectA 0.5723 -0.0837 -0.3835 -0.5262 0.2108 -0.3124 InsectIG → InsectI 0.2833 -0.1305 -0.4712 -0.4354 0.2261 -0.2591 InsectIG → InsectIR 0.3754 0.0116 -0.4423 -0.4035 0.1576 -0.0977 InsectIR → InsectA 0.5229 -0.0694 0.2952 -0.4751 0.2813 -0.0267 InsectIR → InsectI 0.6332 0.0477 0.3664 -0.4710 0.2457 0.0253 InsectIR → InsectIG 0.4970 -0.0905 0.3193 -0.4689 0.1291 -0.0010 Electrial Electrical0 → Electrical1 0.0987 -0.4460 -0.0518 -1.4251 -0.6375 -8.0482 Electrical1 → Electrical0 -0.1539 -1.5070 -1.1412 -2.4765 -1.8348 -7.2222 Table 3: Benchmark results (F1 score for classifiers) on tabular datasets Dataset Source → Target SGN CTGAN TVAE DT MLP DT MLP DT MLP Adult Adult0 → Adult1 0.5560 0.5723 0.3888 0.5707 0.4291 0.5711 Adult1 → Adult0 0.6834 0.8810 0.4348 0.8754 0.2409 0.1181 DryBean DryBean0 → DryBean1 0.0717 0.2306 0.0563 0.0759 0.0699 0.0704 DryBean1 → DryBean0 0.2198 0.2313 0.1717 0.2156 0.0606 0.1487 Avila Avila0 → Avila1 0.3016 0.2975 0.0692 0.1706 0.2714 0.2621 Avila1 → Avila0 0.4625 0.4323 0.0918 0.0713 0.3459 0.3844 4.3 Evaluation on Tabular Datasets Tables 2 and 3 summarize the comparison results for regressors and classifiers, respectively. There are two observations. First, SGN outperforms CTGAN and TVAE in most cases. This is because SGN is designed to learn a universal linear space and transmission across the target and source domains. However, CTGAN and TVAE are designed to learn the distributions of the source domain, and thus, cannot perform well in the target domain. Second, we observe that the results are not symmetric, meaning that the performance of SGN from one dataset to another may not be the same in reverse. For example, SGN performs well from Avila0 to Avia1, but it does not perform well from Avila1 to Avila0. This lack of symmetry can be attributed to differences between the datasets. While GMs trained on one dataset may find it easier to perform well on another, this does not necessarily guarantee the same for models trained on a different dataset. We further conduct an ablation study of the effect of the parameter η in the loss function in Appendix D. 5 Conclusion In this paper, we propose a similarity-based generative network SGN to generate desirable samples for downstream tasks. The core of SGN is a novel similarity-based evaluation mechanism, which utilizes the data labels to construct a zero-one similarity matrix. In the training stage, instead of learning the probability distribution of the training dataset in the source domain, SGN searches for a latent space that satisfies the similarity requirement. In the data generation stage, SGN incorporates the needed distribution information of the representative dataset to generate the desirable samples in the target domain, mitigating the distribution shift problem. Extensive experiments on various real-world datasets, including both image and tabular datasets, demonstrate the effectiveness of SGN and its superiority over the baselines. References A. Acevedo, A. Merino, S. Alférez, and et al. (2020) A dataset of microscopic peripheral blood cell images for development of automatic recognition systems. Data in Brief 30, p. 105474. Cited by: §4.1. V. Arzamasov, K. Böhm, and P. Jochem (2018) Towards concise models of grid stability. In SmartGridComm, p. 1–6. Cited by: §4.1. Y. Bahram, M. Desbos, M. Shateri, and E. Granger (2025) Uni-dad: unified distillation and adaptation of diffusion models for few-step few-shot image generation. CoRR abs/2511.18281. External Links: Link, Document, 2511.18281 Cited by: §1, §2. G. Chou, Y. Bahat, and F. Heide (2023) Diffusion-sdf: conditional generative modeling of signed distance functions. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, p. 2262–2272. External Links: Link, Document Cited by: §1. V. M. A. de Souza, D. M. dos Reis, A. G. Maletzke, and G. E. A. P. A. Batista (2020) Challenges in benchmarking stream learning algorithms with real-world data. Data Min. Knowl. Discov. 34 (6), p. 1805–1858. Cited by: §4.1. DeepDRiD (2020) The 2nd diabetic retinopathy – grading and image quality estimation challenge. Cited by: §4.1. Y. Deldjoo, T. D. Noia, and F. A. Merra (2021) A survey on adversarial recommender systems: from attack/defense strategies to generative adversarial networks. ACM Computing Surveys 54 (2), p. 1–38. Cited by: §1. P. Dhariwal and A. Q. Nichol (2021) Diffusion models beat gans on image synthesis. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, M. Ranzato, A. Beygelzimer, Y. N. Dauphin, P. Liang, and J. W. Vaughan (Eds.), p. 8780–8794. External Links: Link Cited by: §2. I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio (2014) Generative adversarial nets. NIPS 27. Cited by: §1. K. He, X. Zhang, S. Ren, and J. Sun (2016) Deep residual learning for image recognition. In CVPR, p. 770–778. Cited by: §4.2. J. Ho, A. Jain, and P. Abbeel (2020) Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, H. Larochelle, M. Ranzato, R. Hadsell, M. Balcan, and H. Lin (Eds.), External Links: Link Cited by: §1. J. Ho and T. Salimans (2022) Classifier-free diffusion guidance. CoRR abs/2207.12598. External Links: Link, Document, 2207.12598 Cited by: §2. L. Hou, Q. Cao, H. Shen, S. Pan, X. Li, and X. Cheng (2022a) Conditional gans with auxiliary discriminative classifier. In International Conference on Machine Learning, ICML 2022, 17-23 July 2022, Baltimore, Maryland, USA, K. Chaudhuri, S. Jegelka, L. Song, C. Szepesvári, G. Niu, and S. Sabato (Eds.), Proceedings of Machine Learning Research, Vol. 162, p. 8888–8902. External Links: Link Cited by: §1. L. Hou, Q. Cao, H. Shen, S. Pan, X. Li, and X. Cheng (2022b) Conditional GANs with auxiliary discriminative classifier. In ICML, Vol. 162, p. 8888–8902. Cited by: §4.1. T. Hu, J. Zhang, R. Yi, H. Huang, Y. Wang, and L. Ma (2025) SaRA: high-efficient diffusion model fine-tuning with progressive sparse low-rank adaptation. In The Thirteenth International Conference on Learning Representations, ICLR 2025, Singapore, April 24-28, 2025, External Links: Link Cited by: §2. M. Kang and J. Park (2020) Contragan: contrastive learning for conditional image generation. NeurIPS. Cited by: §2, §4.1. M. Kang, W. Shim, M. Cho, and J. Park (2021) Rebooting acgan: auxiliary classifier gans with stable training. NeurIPS 34. Cited by: §2, §4.1. D. S. Kermany, M. Goldbaum, W. Cai, C. C. Valentim, H. Liang, S. L. Baxter, A. McKeown, G. Yang, X. Wu, F. Yan, et al. (2018) Identifying medical diagnoses and treatable diseases by image-based deep learning. Cell 172 (5), p. 1122–1131. Cited by: §4.1. D. P. Kingma and M. Welling (2014) Auto-encoding variational bayes. In ICLR, Cited by: §1. R. Kohavi (1996) Scaling up the accuracy of naive-bayes classifiers: A decision-tree hybrid. In SIGKDD, p. 202–207. Cited by: §4.1. M. Koklu and I. A. Özkan (2020) Multiclass classification of dry beans using computer vision and machine learning techniques. Comput. Electron. Agric. 174, p. 105507. Cited by: §4.1. X. Li, C. Herrmann, K. C. K. Chan, Y. Li, D. Sun, C. Ma, and M. Yang (2025) A simple approach to unifying diffusion-based conditional generation. In The Thirteenth International Conference on Learning Representations, ICLR 2025, Singapore, April 24-28, 2025, External Links: Link Cited by: §2. M. Mirza and S. Osindero (2014) Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784. Cited by: §1, §2. T. Miyato and M. Koyama (2018) CGANs with projection discriminator. In ICLR, Cited by: §1, §1, §2, §4.1. Y. Netzer, T. Wang, A. Coates, A. Bissacco, B. Wu, and A. Y. Ng (2011) Reading digits in natural images with unsupervised feature learning. Cited by: §4.1. A. T. Nguyen, T. Tran, Y. Gal, and A. G. Baydin (2021) Domain invariant representation learning with domain density transformations. NeurIPS 34. Cited by: §4.1. D. Nguyen, J. Li, J. Zheng, and B. Mirzasoleiman (2026) Do we need all the synthetic data? targeted image augmentation via diffusion models. In International Conference on Learning Representations, Cited by: §1, §2. W. Peebles and S. Xie (2023) Scalable diffusion models with transformers. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, p. 4172–4182. External Links: Link, Document Cited by: §1. J. Quiñonero-Candela, M. Sugiyama, A. Schwaighofer, and N. D. Lawrence (2008) Dataset shift in machine learning. Cited by: §1. S. Sadat, M. Kansy, O. Hilliges, and R. M. Weber (2025) No training, no problem: rethinking classifier-free guidance for diffusion models. In The Thirteenth International Conference on Learning Representations, ICLR 2025, Singapore, April 24-28, 2025, External Links: Link Cited by: §1, §2. T. R. Shaham, T. Dekel, and T. Michaeli (2019) Singan: learning a generative model from a single natural image. In ICCV, p. 4570–4580. Cited by: §1. K. Sohn, H. Lee, and X. Yan (2015) Learning structured output representation using deep conditional generative models. NIPS 28, p. 3483–3491. Cited by: §2. C. D. Stefano, F. Fontanella, M. Maniaci, and A. S. di Freca (2011) A method for scribe distinction in medieval manuscripts using page layout features. In ICIAP, Vol. 6978, p. 393–402. Cited by: §4.1. B. Trabucco, K. Doherty, M. Gurinas, and R. Salakhutdinov (2024) Effective data augmentation with diffusion models. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024, External Links: Link Cited by: §2. V. Verma, A. Lamb, C. Beckham, A. Najafi, I. Mitliagkas, D. Lopez-Paz, and Y. Bengio (2019) Manifold mixup: better representations by interpolating hidden states. In Proceedings of the 36th International Conference on Machine Learning, ICML 2019, 9-15 June 2019, Long Beach, California, USA, K. Chaudhuri and R. Salakhutdinov (Eds.), Proceedings of Machine Learning Research, Vol. 97, p. 6438–6447. External Links: Link Cited by: §1, §2. Y. Wang and L. Chen (2025) Inversion circle interpolation: diffusion-based image augmentation for data-scarce classification. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2025, Nashville, TN, USA, June 11-15, 2025, p. 25560–25569. External Links: Link, Document Cited by: §2. Y. Wang, A. Gonzalez-Garcia, D. Berga, L. Herranz, F. S. Khan, and J. van de Weijer (2020) MineGAN: effective knowledge transfer from gans to target domains with few images. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, p. 9329–9338. External Links: Link, Document Cited by: §1, §2. A. Woloshuk, S. Khochare, A. F. Almulhim, A. T. McNutt, D. Dean, D. Barwinska, M. J. Ferkowicz, M. T. Eadon, K. J. Kelly, K. W. Dunn, et al. (2021) In situ classification of cell types in human kidney tissue using 3d nuclear staining. Cytometry Part A 99 (7), p. 707–721. Cited by: §4.1. H. Xiao, K. Rasul, and R. Vollgraf (2017) Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms. CoRR abs/1708.07747. Cited by: §4.1. L. Xu, M. Skoularidou, A. Cuesta-Infante, and K. Veeramachaneni (2019) Modeling tabular data using conditional GAN. In NeurIPS, p. 7333–7343. Cited by: §4.1, §4.1. L. Yang, Z. Zhang, Y. Song, S. Hong, R. Xu, Y. Zhao, W. Zhang, B. Cui, and M. Yang (2024) Diffusion models: A comprehensive survey of methods and applications. ACM Comput. Surv. 56 (4), p. 105:1–105:39. External Links: Link, Document Cited by: §1. Z. Yi, H. Zhang, P. Tan, and M. Gong (2017) Dualgan: unsupervised dual learning for image-to-image translation. In ICCV, p. 2849–2857. Cited by: §1. H. Zhang, M. Cissé, Y. N. Dauphin, and D. Lopez-Paz (2018) Mixup: beyond empirical risk minimization. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings, External Links: Link Cited by: §1, §2. J. Zhong, X. Zhang, J. Wang, and M. Long (2025) Domain guidance: A simple transfer approach for a pre-trained diffusion model. In The Thirteenth International Conference on Learning Representations, ICLR 2025, Singapore, April 24-28, 2025, External Links: Link Cited by: §2. J. Zhu, S. Cai, F. Deng, B. C. Ooi, and W. Zhang (2023) METER: a dynamic concept adaptation framework for online anomaly detection. Proceedings of the VLDB Endowment 17 (4), p. 794–807. Cited by: §1. J. Zhu, S. Cai, Y. Shen, G. Chen, F. Deng, and B. C. Ooi (2025) In-context adaptation to concept drift for learned database operations. In International Conference on Machine Learning, p. 79699–79726. Cited by: §1. J. Zhu, Y. Fan, G. Han, X. Shi, F. Deng, and J. Chen (2026) Generative anomaly detection: a comprehensive review of modeling principles, advances, and future opportunities. Artificial Intelligence Review. Cited by: §1. J. Zhu, H. Ma, J. Chen, and J. Yuan (2022) Few-shot image generation with diffusion models. CoRR abs/2211.03264. External Links: Link, Document, 2211.03264 Cited by: §2.