Incomplete multi-view clustering method and equipment based on domain condition diffusion

Abstract

The invention relates to an incomplete multi-view clustering method and equipment based on domain condition diffusion. The method comprises the following steps: acquiring a multi-view data set; constructing an auto-encoder network for each view to obtain a potential representation feature corresponding to each view; according to the potential representation feature corresponding to each view, obtaining a neighborhood condition of each view; constructing a conditional diffusion generation model; taking neighborhood conditions as condition input, and training a condition diffusion generation model; generating and complementing potential representations corresponding to the missing views by adopting a classifier free guiding method; and fusing the potential representation of each view and obtaining a final clustering result by adopting a K-means clustering algorithm. According to the method, the cross-view stable neighborhood structure is introduced into the potential representation space to serve as condition information, the condition diffusion generation model is guided to conduct robust recovery on the potential representation of the missing view, and the problems that in an existing method, structural constraint is insufficient and representation is unstable under the condition that the heterogeneous view is missing are effectively solved.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and unsupervised learning technology, and specifically relates to an incomplete multi-view clustering method and device based on domain conditional diffusion. Background Technology

[0002] In practical applications, with the continuous advancement of sensing and data acquisition technologies, various application scenarios can be represented using heterogeneous multi-view features, making multi-view learning a core topic in unsupervised representation learning. Multi-view clustering (MVC) leverages cross-view features... Figure 1 While methods aim to reveal the shared structure of data through consistency, most existing methods still assume that all views can be fully observed. However, in practical applications, this assumption often fails to hold due to missing views. This problem has directly driven the research and development of Incomplete Multi-View Clustering (IMVC) methods.

[0003] Existing incomplete multi-view clustering methods are mainly divided into the unfilled paradigm and the filled paradigm. Unfilled methods circumvent the explicit missing view recovery process by directly aggregating available views, exhibiting high computational efficiency when processing incomplete observation data. However, due to the lack of key cross-view information, these methods often face severe information imbalance during aggregation and struggle to effectively capture shared semantics between views when the view missing rate is high. To address this issue, filled methods explicitly complete multi-view information by reconstructing missing views, aiming to better preserve cross-view information. Figure 1 This consistency improves the representational quality of downstream clustering tasks, thus providing higher-quality representation support.

[0004] In imputation methods, generative models have become an important technological direction due to their ability to model complex data distributions and recover latent representations of missing views. Early Generative Adversarial Networks (GANs) synthesized missing features through adversarial learning, but faced risks of training instability and model collapse. Subsequent flow-based generative methods, while able to recover missing items through cross-view modeling, were limited by their reliance on strict Gaussian prior assumptions, restricting their ability to capture complex non-Gaussian data structures. Recently, diffusion-based generative methods have shown strong robustness in handling severe data gaps by iteratively optimizing noisy representations, such as using conditional diffusion to generate missing views or combining contrastive learning to enhance clustering representations. However, existing diffusion-based methods often neglect stable cross-view neighborhood structures, resulting in weak structural constraints in heterogeneous missing scenarios and instability in the missing item completion process. Summary of the Invention

[0005] The purpose of this invention is to provide an incomplete multi-view clustering method and device based on domain conditional diffusion, which can achieve structurally consistent and semantically aligned latent representation learning in heterogeneous missing patterns and high missing rate scenarios, thereby improving the accuracy and robustness of incomplete multi-view data clustering results.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] An incomplete multi-view clustering method and device based on domain conditional diffusion includes the following steps:

[0008] Obtain a multi-view dataset where different views in the dataset have missing samples, and each sample contains at least one observable view feature;

[0009] An autoencoder network is constructed for each view to obtain the latent representation features corresponding to each view;

[0010] Based on the latent representation features corresponding to each view and the observation state information of samples under each conditional view, the neighborhood information in the latent representation space of each observable conditional view is aggregated to obtain the neighborhood conditional components of the corresponding conditional view. According to the preset view order, all neighborhood conditional components are processed. By splicing, we obtain the first... The unified condition vector of each sample The unified condition vectors of all samples are stacked row by row to form the target view. Neighborhood conditions ;

[0011] Construct a conditional diffusion generation model;

[0012] target view Neighborhood conditions As a conditional input, the conditional diffusion generation model is trained;

[0013] A classifier-guided approach is used to generate and complete the latent representations corresponding to missing views;

[0014] The final clustering result is obtained by fusing the potential representations of each view and using the K-means clustering algorithm.

[0015] Preferably, the autoencoder network includes an encoder and a decoder, wherein the encoder maps view features to a latent representation space of uniform dimension; the decoder reconstructs the corresponding view features based on the latent representation; the autoencoder network is trained by minimizing the view reconstruction error; and the view reconstruction error loss function... as follows:

[0016]

[0017] in, For the number of views, For the first Incomplete dataset for each view For the first The output of the view reconstruction model for each view.

[0018] Preferably, based on the observable state of the sample in the conditional view;

[0019] When a sample is observable in the conditional view, a k-nearest neighbor search method is used for all valid samples in the latent representation space of the conditional view to determine the k nearest neighbor samples to the latent representation of sample i, and their latent representations are aggregated to obtain the neighborhood conditional components of the corresponding view. The calculation method is as follows:

[0020]

[0021] in, Indicates the first Each view in the potential representation space is related to the sample The potential representation of the nearest A set of neighbor sample indexes;

[0022] When a sample is missing in all conditional views, the neighborhood conditional component of the corresponding view... They are all uniformly set as zero vectors to form unconditional inputs;

[0023] When a sample is observable only in a partial conditional view, a k-nearest neighbor search is performed for each observable view, and the neighborhood aggregation result is calculated. Meanwhile, for the neighborhood conditional components corresponding to the missing views... Set as the zero vector, the first Neighborhood condition components corresponding to each view The definition is as follows:

[0024]

[0025] in, Indicates sample In the The observation status under each view.

[0026] Preferably, the conditional diffusion generation model is trained using both conditional and unconditional generation methods, with the training objective being to minimize the noise prediction error. The loss function is as follows:

[0027]

[0028] in, This represents the potential representation after adding noise. Represents the target view conditions, This is the noise prediction function for the conditional diffusion generation model. This is real noise.

[0029] Preferably, the classifier-guided method guides the generation of latent representations of missing views by combining conditional and unconditional denoising results, and its guided noise prediction is defined as follows:

[0030]

[0031] in, This indicates the conditional noise reduction result. This represents the result of unconditional denoising. For guiding strength parameters.

[0032] Preferably, after obtaining the guided noise prediction result, it is substituted into the mean calculation formula of the reverse diffusion process in the conditional diffusion generation model to obtain the mean expression under the guided conditions:

[0033]

[0034] Subsequently, at each time step, the denoising update is performed as follows:

[0035]

[0036] in, The standard deviation parameter corresponds to the time step. This represents Gaussian noise sampled independently at each step;

[0037] After completion After several denoising iterations, the final result will be As a potential representation for the recovery of missing samples in the target view.

[0038] In another aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the steps in the aforementioned incomplete multi-view clustering method based on domain conditional diffusion.

[0039] In another aspect, the present invention provides an incomplete multi-view clustering device based on domain conditional diffusion, comprising:

[0040] Memory, used to store software applications.

[0041] A processor is configured to execute the software application, wherein each program of the software application correspondingly performs the steps in the aforementioned incomplete multi-view clustering method based on domain conditional diffusion.

[0042] This invention introduces cross-view stable neighborhood structures as conditional information into the latent representation space, guiding a conditional diffusion generative model to robustly recover latent representations of missing views. This effectively alleviates the problems of insufficient structural constraints and unstable representations in existing methods when heterogeneous views are missing. During the training phase, this invention enhances the model's adaptability to different missing patterns through joint conditional and unconditional generative learning. During the inference phase, it employs a classifier-guided approach to enhance the responsiveness of the generated results to neighborhood conditions, thereby obtaining structurally consistent and semantically aligned multi-view latent representations. Clustering analysis based on these latent representations significantly improves the accuracy and robustness of clustering results under high missing rates and complex missing patterns. Attached Figure Description

[0043] Figure 1 This is a structural diagram of the model of the present invention. Detailed Implementation

[0044] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0045] This invention discloses an incomplete multi-view clustering method and apparatus based on domain conditional diffusion, comprising the following steps:

[0046] Step 1: Obtain a multi-view dataset, in which different views in the multi-view dataset have missing samples, and each sample contains at least one observable view feature;

[0047] In this embodiment, incomplete multi-view data is represented as ,in This indicates the number of views. For any given view, its corresponding data is represented as... ,in, Indicates the number of samples. Indicates the first The dimensions of features in each view.

[0048] To represent the observation state of samples in different views, an observation mask set is defined. ,in, In order to be with the first The mask vector corresponding to each view is used to indicate the observation status of each sample in that view.

[0049] More specifically, mask elements The definition is as follows:

[0050]

[0051] Step 2: Construct an autoencoder network for each view to obtain the latent representation features corresponding to each view;

[0052] Specifically, in incomplete multi-view clustering tasks, autoencoder networks can be used for unsupervised latent representation learning of multi-view data. Through nonlinear mapping and reconstruction-based learning, autoencoders can effectively model heterogeneous, high-dimensional multi-view feature data and generate stable latent representations, thus providing a reliable foundation for subsequent neighborhood condition construction and latent representation completion based on generative models.

[0053] In this invention, the autoencoder network is configured in a view-independent manner to accommodate the heterogeneity between different view feature spaces. For any view... Through the corresponding encoder Input sample Mapping to a latent representation space of uniform dimension, we obtain the sample at the [missing value]. Potential representations under each view:

[0054]

[0055] in, The dimension representing the potential representation. Indicates the first The set of latent representations corresponding to all samples in each view.

[0056] The encoder employs a multi-layer neural network structure and introduces a non-linear activation function to extract complex structural features within each view and map the features of different views to a unified latent representation space. Correspondingly, the decoder... The input features are reconstructed based on the latent representation, and the reconstruction process is as follows:

[0057]

[0058] The autoencoder network is trained by jointly optimizing the reconstruction process of each view, and its reconstruction loss function is defined as:

[0059]

[0060] By minimizing the reconstruction loss function, the autoencoder network can learn a compact and structurally consistent multi-view latent representation, providing the necessary prior representation foundation for subsequent domain condition construction and conditional diffusion-based latent representation completion processes.

[0061] Step 3, Domain Condition Construction: When applying the diffusion generation model to the incomplete multi-view clustering task, a key technical challenge lies in designing discriminative and reliable conditional signals to effectively guide the denoising generation process of the diffusion model. Unlike tasks with explicit semantic conditions, such as image generation or text generation, incomplete multi-view data typically exhibits heterogeneous feature spaces and irregular view missing patterns. Therefore, it is necessary to construct conditional representations that can both characterize cross-view structural information and remain robust to view missingness. Without effective conditional constraints, the latent representation completion process based on the diffusion model will be weakly constrained, easily leading to unstable generation results and a lack of consistency between different views.

[0062] To address the aforementioned issues, this embodiment proposes a neighborhood condition construction module, denoted as NCC, which converts the latent representations in the observable views into a unified condition vector to guide the diffusion generative model in completing the latent representations of missing views. The basic idea is that the local neighborhood structure of samples in the latent space can encode stable cross-view semantic information and, compared to single-point features, is more robust to noise perturbations and view-specific biases. Based on this understanding, the NCC module aggregates the neighborhood information in each observable view and represents the missing view by introducing zero-valued conditional components, thereby constructing a structurally stable multi-view condition vector.

[0063] Specifically, for the target view First, obtain its corresponding latent representation set. And simultaneously extract all conditional views latent representation set And the corresponding view observation state information. For any sample The NCC module constructs corresponding view condition components based on their observability in different condition views, and adopts corresponding condition construction strategies under different availability states, thereby forming a neighborhood condition representation to guide the generation process. Different neighborhood condition construction methods are adopted according to the availability state of the condition view, including but not limited to the following:

[0064] Scenario 1: When the conditional view is fully available

[0065] When sample In Conditions View When the middle is an observable state, the neighborhood condition construction module, in the latent representation space of this view, performs a neighborhood condition construction on all valid samples (including sample 0) (It itself) adopts The k-Nearest Neighbors (kNN) method determines the nearest neighbor to the sample. Potential representation of nearest The neighborhood conditional components of the corresponding view are obtained by aggregating the neighbor samples and their latent representations:

[0066]

[0067] in, Indicates the first Each view in the potential representation space is related to the sample The closest A set of neighbor sample indexes.

[0068] Scenario 2: All conditional views are missing

[0069] When sample In all condition views When all conditional components are missing, the neighborhood conditional construction module cannot compute the neighborhood aggregation result because there are no available potential representations of the conditional views. In this case, the neighborhood conditional construction module uniformly sets the conditional components corresponding to all conditional views to zero vectors, thereby forming zero-valued conditional inputs to guide the conditional diffusion generative model to learn and infer under unconditional constraints.

[0070] By introducing the aforementioned unconditional learning approach, the conditional diffusion generative model can avoid over-reliance on specific conditional patterns during training, reduce training bias, and enable the model to learn the latent structural features of the target view itself. During model operation, even in the absence of cross-view auxiliary information, the conditional diffusion generative model can still generate and complete latent representations based on the inherent latent structure of the target view, thereby improving the robustness of the conditional diffusion generative model under different missing patterns.

[0071] In addition, the aforementioned unconditional learning method plays an implicit regularization role in the model training process, which helps to enhance the generalization ability and stability of the generative model in diverse missing scenarios.

[0072] Scenario 3: When the conditional view is partially available

[0073] When sample When a partial conditional view is observable, the neighborhood condition building module applies a condition to each observable view. Nearest neighbor search is performed, and neighborhood aggregation results are calculated. Conditional components corresponding to missing views are set as zero vectors to maintain consistency in the dimension and structure of the conditional vectors. Specifically, the first... The conditional components corresponding to each view are defined as follows:

[0074]

[0075] In this way, the generative model is exposed to conditional inputs with both partially sufficient and partially missing information during training. The observable view provides cross-view structural cues for the missing view, while the zero-valued conditional component is used to suppress the model's over-reliance on complete conditions, thereby improving the model's generalization ability and stability under incomplete conditions.

[0076] After obtaining the conditional components corresponding to each conditional view, all conditional components are spliced ​​together according to the preset view order to form the target view. medium sample Unified neighborhood condition vector:

[0077]

[0078] Then, the uniform condition vectors of all samples are stacked row by row to form the target view. Neighborhood conditions ;

[0079]

[0080] During model training, a conditional random missing value strategy is further employed, i.e., missing values ​​are assigned a certain probability. The target view Neighborhood conditions The zero vector is replaced, allowing the generative model to learn both conditional and unconditional denoising generation processes during the training phase, thereby enhancing the model's adaptability to different missing modes.

[0081] The neighborhood condition construction method described above can transform heterogeneous and incomplete multi-view latent representations into structurally stable and semantically aligned conditional signals, providing reliable guidance for the latent representation completion process based on the diffusion generation model, thereby effectively improving the robustness of missing view recovery and clustering results in incomplete multi-view clustering tasks.

[0082] Step 4, Training the Conditional Diffusion Generative Model: Obtaining the target view output by the neighborhood conditional building module. Neighborhood conditions Subsequently, this invention employs a conditional denoising diffusion probability model to model and learn the missing latent representations in the target view. This conditional diffusion generation model, based on the standard denoising diffusion model, incorporates the neighborhood conditions into each denoising step, thereby guiding the model to complete a structurally consistent generation process in the latent representation space.

[0083] Step 4.1, Forward Diffusion Process (Noise Injection Process): During model training, a forward diffusion process is first performed on the latent representation of the target view. This involves progressively injecting Gaussian noise into the original latent representation over multiple time steps, gradually degenerating it into a random variable that approximates a standard Gaussian distribution. Specifically, at time steps... At that time, the forward diffusion process is defined as:

[0084]

[0085] in, Indicates a Gaussian distribution. and Defined respectively as well as ,in , This represents the total number of steps in the diffusion process. Used to control noise variance.

[0086] Step 4.2, Backdiffusion Process (Denoising Generation Process): In the backdiffusion process, the model, in the manner of a Markov chain, transitions from the noisy state... Gradually restore to the original latent representation At any time step The reverse denoising process is modeled as a conditional Gaussian distribution:

[0087]

[0088] in, and Let represent the mean and standard deviation of the Gaussian distribution, respectively, and their definitions are as follows:

[0089]

[0090]

[0091] in, This represents a noise prediction network, specifically a U-Net denoising network, used to predict noise levels at time steps [1-3]. The injected noise component is the target view that the network constructs during the prediction process based on neighborhood conditions. conditions As a conditional input Indicates model parameters.

[0092] Step 4.3, Model Training Objective Function: Through the above diffusion process, the conditional diffusion generation model is trained to minimize the noise prediction error. Its training objective function is defined as:

[0093]

[0094] in, During training, the target view conditions Provided directly by the NCC module, it can naturally adapt to different situations such as fully available, partially available, and completely missing conditional views, thereby guiding the diffusion generative model to learn the latent representation recovery rules under different conditional constraints.

[0095] Step 5, Generation of latent representations for missing views based on classifier-guided free thinking: During model operation, for samples with missing data in the target view, this invention uses a conditional diffusion inference method based on classifier-guided free thinking to generate and complete the latent representations corresponding to the missing views.

[0096] Step 5.1, Latent Representation Initialization and Iterative Denoising: Initialize the latent representation corresponding to the target view as a standard Gaussian distributed random variable:

[0097]

[0098] and from time step Iterate step by step to At each time step, a denoising update operation is performed to progressively recover the underlying representation of the target view.

[0099] Step 5.2, Classifier Free-Guided Noise Prediction: To fully utilize the target view Neighborhood conditions To improve generation stability under weak or noisy conditions, a classifier-guided strategy is adopted during inference, which weights and combines conditional noise prediction results with unconditional noise prediction results. The guided noise prediction is defined as follows:

[0100]

[0101] in, This represents the conditional noise prediction result based on neighborhood conditions. This represents the result of unconditional denoising. For guiding strength parameters.

[0102] Step 5.3, Guided Mean Calculation and Denoising Update: After obtaining the guided noise prediction result, substitute it into the mean calculation formula for the back diffusion process to obtain the mean expression under the guided conditions:

[0103]

[0104] Subsequently, at each time step, the denoising update is performed as follows:

[0105]

[0106] in, The standard deviation parameter corresponds to the time step. This represents Gaussian noise sampled independently at each step;

[0107] Step 5.4, Obtain the latent representation of the missing view: upon completion After several denoising iterations, the final result will be As a recovery potential representation of missing samples in the target view, it is used for subsequent multi-view potential representation fusion and cluster analysis.

[0108] Step 6, Clustering: After completing model optimization and restoring the latent representation of the missing view, the restored latent representation of the missing view is concatenated with the latent representation of the observed view in the corresponding sample to form a unified latent representation of the sample. Based on the unified latent representation, the K-means clustering algorithm is used to cluster the sample to obtain the final clustering result.

[0109] In this embodiment, three general datasets, Fashion, Aloi_deep, and Scene_15, were selected. Under different view missing rates, the method of the present invention was compared with a variety of existing incomplete multi-view clustering algorithms. The experimental results are shown in Table 1, which shows that the method of the present invention has advantages in clustering performance on the above datasets.

[0110] Table 1 shows the performance of different algorithms on the Fashion, Aloi_deep, and Scene_15 datasets.

[0111]

[0112] In summary, this invention provides an incomplete multi-view clustering method based on neighborhood conditional diffusion. By introducing cross-view stable neighborhood structures as conditional information into the latent representation space, it constructs neighborhood conditions for the target view and combines them with a conditional diffusion generation model to achieve robust generation and completion of latent representations for missing views. During the training phase, this invention enhances the model's adaptability to different view missing patterns through joint conditional and unconditional learning. During the model runtime phase, it introduces a classifier free guidance mechanism to improve the stability of the generated results' response to conditional information, thereby obtaining structurally consistent and semantically aligned multi-view latent representations. Clustering analysis based on these latent representations yields relatively stable clustering results even under conditions of high missing rates and complex missing patterns.

[0113] In another aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the steps in the aforementioned incomplete multi-view clustering method based on domain conditional diffusion.

[0114] In another aspect, the present invention provides an incomplete multi-view clustering device based on domain conditional diffusion, comprising:

[0115] Memory, used to store software applications.

[0116] A processor is configured to execute the software application, wherein each program of the software application correspondingly performs the steps in the aforementioned incomplete multi-view clustering method based on domain conditional diffusion.

Claims

1. An incomplete multi-view clustering method based on domain conditional diffusion, characterized in that, Includes the following steps: Obtain a multi-view dataset where different views in the dataset have missing samples, and each sample contains at least one observable view feature; An autoencoder network is constructed for each view to obtain the latent representation features corresponding to each view; Based on the latent representation features corresponding to each view and the observation state information of samples under each conditional view, the neighborhood information in the latent representation space of each observable conditional view is aggregated to obtain the neighborhood conditional components of the corresponding conditional view. According to the preset view order, all neighborhood conditional components are processed. By splicing, we obtain the first... The unified condition vector of each sample The unified condition vectors of all samples are stacked row by row to form the target view. Neighborhood conditions ; Construct a conditional diffusion generation model; target view Neighborhood conditions As a conditional input, the conditional diffusion generation model is trained; A classifier-guided approach is used to generate and complete the latent representations corresponding to missing views; The final clustering result is obtained by fusing the potential representations of each view and using the K-means clustering algorithm.

2. The incomplete multi-view clustering method based on domain conditional diffusion according to claim 1, characterized in that, The autoencoder network includes an encoder and a decoder. The encoder maps view features to a unified-dimensional latent representation space. The decoder reconstructs the corresponding view features based on the latent representation. The autoencoder network is trained by minimizing the view reconstruction error. The view reconstruction error loss function... as follows: in, For the number of views, For the first Incomplete dataset for each view For the first The output of the view reconstruction model for each view.

3. The incomplete multi-view clustering method based on domain conditional diffusion according to claim 1, characterized in that, Based on the observable state of the sample in the conditional view; When a sample is observable in the conditional view, a k-nearest neighbor search method is used for all valid samples in the latent representation space of the conditional view to determine the k nearest neighbor samples to the latent representation of sample i, and their latent representations are aggregated to obtain the neighborhood conditional components of the corresponding view. The calculation method is as follows: in, Indicates the first Each view in the potential representation space is related to the sample The potential representation of the nearest A set of neighbor sample indexes Indicates the first Samples in each view Potential representation of . When a sample is missing in all conditional views, the neighborhood conditional component of the corresponding view... They are all uniformly set as zero vectors to form unconditional inputs; When a sample is observable only in a partial conditional view, a k-nearest neighbor search is performed for each observable view, and the neighborhood aggregation result is calculated. Meanwhile, for the neighborhood conditional components corresponding to the missing views... Set as the zero vector, the first Neighborhood condition components corresponding to each view The definition is as follows: in, Indicates sample In the The observation status under each view.

4. The incomplete multi-view clustering method based on domain conditional diffusion according to claim 1, characterized in that, The conditional diffusion generation model is trained using both conditional and unconditional generation methods. Its training objective is to minimize the noise prediction error, and the loss function is as follows: in, This represents the potential representation after adding noise. Represents the target view conditions, This is the noise prediction function for the conditional diffusion generation model. This is real noise.

5. The incomplete multi-view clustering method based on domain conditional diffusion according to claim 1, characterized in that, The classifier-guided method guides the generation of latent representations for missing views by combining conditional and unconditional denoising results. Its guided noise prediction is defined as follows: in, This indicates the conditional noise reduction result. This represents the result of unconditional denoising. For guiding strength parameters.

6. The incomplete multi-view clustering method based on domain conditional diffusion according to claim 5, characterized in that, After obtaining the guided noise prediction results, substitute them into the mean calculation formula for the reverse diffusion process of the conditional diffusion generation model to obtain the mean expression under the guided conditions: Subsequently, at each time step, the denoising update is performed as follows: in, In the first After the noise reduction update, the target view The potential for recovery, The standard deviation parameter corresponds to the time step. This represents Gaussian noise sampled independently at each step; After completion After several denoising iterations, the final result will be As a potential representation for the recovery of missing samples in the target view.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it performs the steps in the incomplete multi-view clustering method based on domain conditional diffusion as described in any one of claims 1 to 6.

8. An incomplete multi-view clustering device based on domain conditional diffusion, characterized in that, include: Memory, used to store software applications. A processor for executing the software application, wherein each program of the software application correspondingly performs the steps in the incomplete multi-view clustering method based on domain conditional diffusion as described in any one of claims 1 to 6.

Images