Incomplete multi-view clustering method and system, storage medium and equipment

Abstract

The invention relates to the technical field of multi-view clustering analysis, and discloses an incomplete multi-view clustering method and system, a storage medium and equipment. In order to solve the problems that a small amount of easy-to-obtain supervision information is not utilized in an existing method, and the clustering performance is deteriorated due to low view interpolation quality under high missing degree, the invention provides a technical scheme of combining paired constraint weighted interpolation and double-level feature fusion. The method comprises the following steps: firstly, screening similar samples by using pairwise constraint information, and reconstructing a missing view through similarity weighting; then extracting feature mean values and standard deviations of all views based on an encoder, aligning features through view hierarchy self-adaptive comparison learning and reserving private information, and optimizing feature distribution in combination with sample hierarchy semi-supervised loss; and finally, completing clustering through a Gaussian mixture model. According to the method, the pairwise constraint information is effectively utilized, the view recovery quality and the feature complementarity under the high-missing scene are improved, the clustering performance is remarkably improved, and the method is suitable for scenes such as data analysis of multi-view data missing and the like.

Description

Technical Field

[0001] This invention relates to the field of multi-view clustering technology, and in particular to an incomplete multi-view clustering method, system, storage medium, and device. Background Technology

[0002] Currently, multi-view data is prevalent in real-world applications, such as images of objects from multiple perspectives or data collected by different sensors. Over the past decade, multi-view clustering has been extensively studied for data analysis. However, in practical applications, multi-view datasets often suffer data loss due to sensor or storage device failures. If traditional multi-view clustering methods are directly applied to process this type of incomplete data, their performance typically degrades significantly. To address this challenge, many incomplete multi-view clustering methods have been proposed. These methods are specifically designed to mitigate the negative impact of missing views.

[0003] Incomplete multi-view clustering methods have garnered significant attention in recent years due to their ability to handle real-world scenarios with missing multi-view data. Traditional incomplete multi-view clustering methods aim to learn common representations or multiple latent representations through methods such as nonnegative matrix factorization, multi-kernel learning, and graph learning. For example, the earliest incomplete multi-view clustering methods used nonnegative matrix factorization to obtain a shared latent representation for clustering. Multi-kernel learning-based methods aim to recover the kernel matrix and learn a common representation from each view. Similarly, graph learning-based methods recover information from missing views or learn shared representations by reconstructing incomplete graphs. However, these traditional methods are limited by shallow models and struggle to handle more complex data. Deep neural networks, with their powerful nonlinear fitting capabilities, can more effectively handle these incomplete multi-view data, and recent research has begun exploring deep learning-based incomplete multi-view clustering methods to address the limitations of shallow models.

[0004] Deep incomplete multi-view clustering methods can be mainly divided into two categories: imputation-based methods and non-imputation-based methods. Non-imputation-based methods directly learn latent feature representations on the available views, while minimizing the differences between the latent representations of each view and learning a shared feature to ensure cross-view clustering. Figure 1Consistency is a challenge. However, when the semantic differences between views are significant, these methods struggle to learn an effective and consistent representation. This problem is exacerbated by missing views, leading to a decline in clustering performance. Imputation-based methods, on the other hand, reconstruct the missing views before clustering. View reconstruction typically uses generative models such as generative adversarial networks to generate the missing views, or infers features from the missing views using graph convolutional neural networks, and then treats this as a complete multi-view dataset for clustering. In scenarios with high missingness, if high-quality view reconstruction can be achieved, imputation-based methods have a clear advantage over non-imputation-based methods. Therefore, imputation-based incomplete multi-view clustering methods have surged in recent years.

[0005] These two types of incomplete multi-view clustering methods have shortcomings. First, the existing methods are unsupervised learning and do not utilize a small amount of easily obtainable supervisory information to guide clustering. Existing contrastive learning algorithms force all positive sample pairs to be similar, which may lead to the loss of private information of the view and ignore the complementarity of multi-view data. Second, in the case of high missing data, the imputation-based methods cannot achieve high-quality view restoration, which leads to a sharp deterioration in clustering performance or even complete failure. Summary of the Invention

[0006] To address the problems in the prior art where the limited amount of readily available supervisory information is not fully utilized and where low-quality interpolation of views with high missing values ​​leads to deterioration in clustering performance, this invention provides an incomplete multi-view clustering method, system, storage medium, and device.

[0007] In a first aspect, the present invention provides an incomplete multi-view clustering method, comprising the following steps:

[0008] S1. Obtain an incomplete multi-view dataset. The incomplete multi-view dataset contains N samples, each of which has V views. If the vth view of the i-th sample is missing, then the i-th sample is a sample with missing views, and the vth view is a missing view. S2. The missing views are weighted and interpolated based on pairwise constraint information to obtain the complete incomplete multi-view dataset. S3. Input each view of each sample in the completed incomplete multi-view dataset into the corresponding encoder to extract the view-level features corresponding to each view; S4. Fuse the view-level features of all views for each sample to obtain the sample-level features of that sample; based on the sample-level features, reconstruct the original data of each view using the decoder corresponding to each view; S5. Construct a total loss function to train the encoder and the decoder to obtain a trained encoder and decoder; the total loss function includes an evidence lower bound loss function, a contrastive learning loss function, a consistency loss function, and a semi-supervised loss function based on Euclidean distance between samples; S6. Extract the latent features of all samples using the trained encoder, calculate the clustering probability of each sample based on the latent features of all samples, and form the final clustering result based on the clustering probability of each sample.

[0009] Secondly, the present invention provides an incomplete multi-view clustering system, comprising the following modules: The data acquisition module is used to acquire an incomplete multi-view dataset, which contains N samples, each sample has V views, and if the vth view of the i-th sample is missing, then the i-th sample is a sample with missing views, and the vth view is a missing view. The pairwise constraint weighted interpolation module is used to perform weighted interpolation on the missing views based on pairwise constraint information to obtain a completed incomplete multi-view dataset. The multi-view feature extraction module is used to input each view of each sample in the completed incomplete multi-view dataset into the corresponding encoder to extract the view-level features corresponding to each view. The feature fusion and reconstruction module is used to fuse the view-level features of all views of each sample to obtain the sample-level features of that sample; based on the sample-level features, the original data of each view is reconstructed through the decoder corresponding to each view. The model joint training module is used to construct a total loss function to train the encoder and the decoder, thereby obtaining a trained encoder and decoder; the total loss function includes an evidence lower bound loss function, a contrastive learning loss function, a consistency loss function, and a semi-supervised loss function based on Euclidean distance between samples; The clustering output and label assignment module is used to extract the latent features of all samples through the trained encoder, calculate the clustering assignment probability of each sample based on the latent features of all samples, and form the final clustering result according to the clustering assignment probability of each sample.

[0010] Thirdly, the present invention provides a storage medium storing a computer program, which, when executed by a processor, can implement the incomplete multi-view clustering method described above.

[0011] Fourthly, the present invention provides a computer device including a memory and a processor, wherein the memory stores computer-readable instructions, and when executed by the processor, the computer-readable instructions cause the processor to perform the incomplete multi-view clustering method described above.

[0012] Compared with the prior art, the advantages of this invention are as follows: This invention overcomes the limitations of traditional imputation-based incomplete multi-view clustering methods by employing pairwise constraint-weighted imputation. By introducing a small amount of supervisory information, it reconstructs views within the same view layer, avoiding cross-view data recovery. Through pairwise constraint-weighted imputation, this invention achieves high-quality view imputation even with high missing data, significantly improving clustering performance. Furthermore, the encoder and decoder loss functions employ an evidence lower bound loss function, a contrastive learning loss function, a consistency loss function, and a semi-supervised loss function. Contrastive learning helps align multiple views of samples in a multi-view dataset to optimize the clustering structure. The contrastive learning loss function is constructed based on view pair similarity and learnable weights, assigning high weights to highly similar views to strengthen shared representations, while assigning lower weights to views with lower similarity that may contain category discrimination information to preserve this private information. Simultaneously, the use of a semi-supervised loss function based on Euclidean distance between samples yields a more reasonable clustering structure and reduces the negative impact of view-level adaptive contrastive learning defining similar samples as negative pairs. Attached Figure Description

[0013] Figure 1 This is a flowchart of an incomplete multi-view clustering method provided by the present invention; Figure 2 This is a schematic diagram of the process of an incomplete multi-view clustering method provided by the present invention. Detailed Implementation

[0014] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0015] Example 1 like Figure 1 As shown, this embodiment provides an incomplete multi-view clustering method, the steps of which are as follows: S1. Obtain an incomplete multi-view dataset. The incomplete multi-view dataset contains N samples, each of which has V views. If the vth view of the i-th sample is missing, then the i-th sample is a sample with missing views, and the vth view is a missing view. S2. Perform weighted interpolation on the missing views based on pairwise constraint information to obtain the completed incomplete multi-view dataset; S3. Input each view of each sample in the completed incomplete multi-view dataset into the corresponding encoder to extract the view-level features corresponding to each view; S4. Fuse the view-level features of all views for each sample to obtain the sample-level features; based on the sample-level features, reconstruct the original data of each view through the decoder corresponding to each view. S5. Construct the total loss function to train the encoder and decoder, and obtain the trained encoder and decoder. The total loss function includes the evidence lower bound loss function, the contrastive learning loss function, the consistency loss function, and the semi-supervised loss function based on the Euclidean distance between samples. S6. Extract the latent features of all samples using the trained encoder, calculate the clustering probability of each sample based on the latent features of all samples, and form the final clustering result based on the clustering probability of each sample.

[0016] Step S1 obtains an incomplete multi-view dataset. The incomplete multi-view dataset contains N samples, each with V views. If the v-th view of the i-th sample is missing, then the i-th sample is a sample with missing views, and the v-th view is a missing view. The expression for the obtained incomplete multi-view dataset is:

[0017] in, This is a missing view.

[0018] Step S2 performs weighted interpolation on the missing views based on pairwise constraint information to obtain the completed incomplete multi-view dataset, including the following steps: S2.1 Filter out all samples in the incomplete multi-view dataset that form a Must-Link relationship with the missing view sample as ML samples; S2.2 Calculate the similarity between each ML sample and the missing sample in the view. The expression for the similarity is:

[0019] in, for and The number of non-zero view similarities; x i For the missing sample of the view, x j The ML sample; For the missing view, For the ML sample of the first One view; Indicates sample and The Element-wise multiplication is performed on each of the non-missing views. In the missing samples of the views, if, except for the first... If there are missing values ​​outside of the first view, the mean of the corresponding views from all ML samples will be used as the replacement. .

[0020] S2.3. In the ML samples, select the L samples with the highest similarity to participate in imputation; if the number of ML samples is less than the set value... Then all ML samples are used for imputation; S2.4. Calculate the weight of each ML sample participating in imputation based on similarity, and perform weighted imputation on the missing views based on these weights to obtain the completed incomplete multi-view dataset. The weights are calculated as follows:

[0021] in, For the aforementioned similarity, and Representing samples respectively and The Data for each view.

[0022] Weighted interpolation is performed on the missing view based on weights. The expression for weighted interpolation is:

[0023] in, Indicates sample One of the L most similar samples in the ML sample The A view.

[0024] Furthermore, in cases of high missing views, where only a small subset of samples have all views available (e.g., 10%), and the remaining samples have one or more unavailable views, some ML sample pairs may lack corresponding views for imputation. In such cases, step S2 is not executed. Not restoring the views avoids introducing erroneous estimates that could lead to incorrect model learning. Through the pairwise constraint-weighted imputation described above, an incomplete multi-view dataset with a significantly reduced number of missing views is obtained.

[0025] Step S3 inputs each view of each sample in the completed incomplete multi-view dataset into the corresponding encoder to extract view-level features for each view. View-level features include the view feature mean and the view sample difference. The encoder includes a variational autoencoder; in this embodiment, a variational autoencoder is assigned to each view for specific view feature extraction. The variational autoencoder typically uses the lower bound of evidence (ELBO) as the loss function.

[0026] Step S4 fuses the view-level features of all views for each sample to obtain the sample-level features; based on the sample-level features, the original data of each view is reconstructed through the decoder corresponding to each view.

[0027] In practice, view-level features are multiplied and fused into sample-level features. Sample-level features include the sample mean and sample standard deviation. The expression for sample-level features is:

[0028] in, The mean of the sample features; The standard deviation of the sample; The average value of the view features for each view; This represents the view sample difference for each view.

[0029] Step S5 constructs the total loss function to train the encoder and decoder, resulting in a trained encoder and decoder. The total loss function includes the evidence lower bound loss function, the contrastive learning loss function, the consistency loss function, and the semi-supervised loss function based on the Euclidean distance between samples.

[0030] In practical implementation, the expression for the evidence lower bound loss function is:

[0031] in, Let the lower bound of the evidence be the loss function; This represents the latent features after fusion and reparameterization; For categorical variables; Indicates the calculation of the posterior distribution The expected value below; Indicates category as Time-latent features The prior distribution of is a Gaussian distribution in this embodiment; The prior distribution of the categorical variable itself is represented by a uniform distribution in this embodiment; and They represent the samples respectively The obtained latent features The distribution and latent features The distribution of the reconstructed view is inferred by fitting a view-specific encoder and decoder, which are structurally symmetric multilayer perceptrons. The last two terms are KL divergence terms. By optimizing the corresponding loss, the extracted sample features are constrained to conform to a Gaussian distribution, and the prior class distribution is set to a uniform distribution. Since the sample features are modeled using a Gaussian distribution, the extracted features can be represented as the mean of the view features corresponding to each view. and view standard deviation .

[0032] In the specific implementation process, a contrastive learning loss function is constructed based on view pair similarity and learnable weights. View pair similarity is calculated through view-level features. The calculation of view pair similarity is as follows:

[0033] in, The similarity of the view pairs; The mean of the view features of the missing view; For the remaining samples The mean of view features for each view; , For view pairs; A contrastive learning loss function is constructed based on view similarity and learnable weights. The expression for the contrastive learning loss function is as follows:

[0034]

[0035] in, To learn the loss function by comparison; and For learnable weights, The view can learn the weights of positive samples. The view learnable negative sample weights are randomly initialized before training and maintained during training. .

[0036] In practical implementation, when views are missing, the feature fusion strategy tends to exacerbate the problem of the fused features being dominated by views with smaller variance. We introduce a consistency loss function to force the features of each view to be similar to the fused features, mitigating this imbalance and ensuring that each view can provide meaningful information for the fused features. The expression for the consistency loss function is:

[0037] in, Let z be the consistency loss function; z is the latent feature, generated from sample-level features through reparameterization. , This indicates element-wise multiplication. It is random noise with values ​​in the range [0,1].

[0038] In the specific implementation process, the clustering structure is further optimized using pairwise constraint information, and the negative impact of defining similar samples as negative sample pairs is reduced by considering the view hierarchy. A matrix is ​​defined based on the pairwise constraint information. Based on the matrix Construct a semi-supervised loss function to reduce the Euclidean distance between ML sample pairs and separate CL sample pairs. Matrix The expression is:

[0039] The expression for the semi-supervised loss function is:

[0040] in, Let be the semi-supervised loss function; , These are latent features of the sample, generated from sample-level features through reparameterization.

[0041] The total loss function is constructed based on the evidence lower bound loss function, the adaptive contrastive learning loss function, the consistency loss function, and the semi-supervised loss function. The expression for the total loss function is as follows:

[0042] in, Let the lower bound of the evidence be the loss function; Let be the consistency loss function; Let be the contrastive learning loss function; Let be the semi-supervised loss function; α The weight hyperparameters are used to compare the learning loss. The weight hyperparameters are for the semi-supervised loss.

[0043] The encoder, decoder, and learnable weights are optimized by minimizing the total loss function and then using backpropagation. Parameter updates are performed using an Adam optimizer with a learning rate of 0.0005, and the weight hyperparameters are compared between the learned loss function and the semi-supervised loss function. and Based on rigorous empirical analysis, the values ​​are set to 1 and 10 respectively in this embodiment. The network architecture of the encoder and decoder is D-500-500-2000-10 (10-2000-500-500-D), where D is the number of input sample features, and ReLU is used as the activation function.

[0044] Step S6 extracts the latent features of all samples using the trained encoder, calculates the clustering probability of each sample based on the latent features of all samples, and forms the final clustering result based on the clustering probability of each sample.

[0045] In practice, Gaussian mixture models are used for clustering, and the latent features follow a Gaussian mixture distribution. Each cluster corresponds to A Gaussian distribution is used. The latent features of all samples are input into a Gaussian mixture model to obtain the clustering probability of each sample. The expression for the Gaussian mixture model is:

[0046]

[0047] in, Represents the identity matrix. For the first The prior probabilities of each cluster, and satisfying ; Indicates the first Potential characteristics of each category The distribution of , i.e., the Gaussian distribution; Represents all potential features The distribution is a mixture Gaussian distribution.

[0048] The cluster assignment probability of each sample is calculated based on the Gaussian mixture model, and the expression for the cluster assignment probability is as follows: .

[0049] Finally, the category with the highest clustering probability is taken as the label of the sample.

[0050] like Figure 2 As shown, the main technical solution of the incomplete multi-view clustering method provided in this embodiment is pairwise constraint weighted interpolation and two-level feature fusion.

[0051] For pairwise constraint-weighted imputation, this embodiment does not use available views of the same sample to reconstruct the missing view across views. Instead, it utilizes the same views of other samples for reconstruction. In multi-view datasets, different views of the same sample contain different information, especially when the modalities of different views are different (such as text and images) or when there is a high degree of missing data, the information imbalance between views becomes more pronounced. In this case, the available view data for a single sample is small and the missing data of each sample is different, making it difficult for the generative model to obtain a reliable reconstructed view. The recovered view is often introduced as noise, leading to a decrease in clustering performance. To solve this problem, this embodiment proposes a pairwise constraint-weighted imputation method to reconstruct the missing view. This method first uses pairwise constraint information to select labeled samples of the same category as the missing sample from all samples; then, it calculates the cosine similarity between the missing sample and each sample of the same category on the available views to determine the corresponding weights; finally, it performs weighted imputation on the data of multiple samples of the same category on the target view based on these weights, thereby realizing the reconstruction and recovery of the missing view.

[0052] For two-level feature fusion, contrastive learning has been widely applied in unsupervised learning fields such as clustering, helping to align multiple views of samples in a multi-view dataset to optimize the clustering structure. However, most current contrastive learning methods applied to multi-view clustering force all positive sample pairs to be similar and negative sample pairs to be separated. This embodiment considers that the information contained in each view in a multi-view dataset is different, and a certain view may contain information that is beneficial for category judgment. Retaining this information can improve clustering performance. Based on this, an adaptive contrastive learning is proposed at the view level. A dynamic weight is assigned to each view. By optimizing the adaptive contrastive learning loss function, high weights are assigned to highly similar views to strengthen shared representations, while lower similar views may contain category discrimination information and are assigned lower weights to retain this private information. The sample level after fusion through the product method takes into account both the consistency and complementarity of multiple view data. To further utilize pairwise constraint information, we designed a semi-supervised loss function based on the Euclidean distance between samples at the sample level. By optimizing the loss function, the Euclidean distance between all ML sample pairs is shortened, while all CL sample pairs are forced to separate, resulting in a more reasonable clustering structure. This also reduces the negative impact of view-level adaptive contrastive learning defining similar samples as negative sample pairs.

[0053] Example 2 This embodiment proposes an incomplete multi-view clustering system, including the following modules: The data acquisition module is used to acquire an incomplete multi-view dataset. The incomplete multi-view dataset contains N samples, each sample has V views. If the vth view of the i-th sample is missing, then the i-th sample is a sample with missing views, and the vth view is a missing view.

[0054] The pair constraint weighted interpolation module is used to perform weighted interpolation on missing views based on pair constraint information to obtain a complete incomplete multi-view dataset.

[0055] The multi-view feature extraction module is used to input each view of each sample in the completed incomplete multi-view dataset into the corresponding encoder to extract the view-level features corresponding to each view. The feature fusion and reconstruction module is used to fuse the view-level features of all views of each sample to obtain the sample-level features of that sample; based on the sample-level features, the original data of each view is reconstructed through the decoder corresponding to each view. The model joint training module is used to construct the total loss function to train the encoder and decoder, resulting in a trained encoder and decoder. The total loss function includes the evidence lower bound loss function, the contrastive learning loss function, the consistency loss function, and the semi-supervised loss function based on the Euclidean distance between samples. The clustering output and label assignment module is used to extract the latent features of all samples through the trained encoder, calculate the cluster assignment probability of each sample based on the latent features of all samples, and form the final clustering result according to the cluster assignment probability of each sample.

[0056] Example 3 This embodiment proposes a storage medium storing a computer program. When the computer program is executed by a processor, it can implement the incomplete multi-view clustering method described in Embodiment 1.

[0057] Example 4 This embodiment proposes a computer device, including a memory and a processor. The memory stores computer-readable instructions, which, when executed by the processor, cause the processor to perform an incomplete multi-view clustering method as described in Embodiment 1.

[0058] In summary, this invention proposes a pairwise constraint-weighted imputation method for clustering incomplete multi-view data with high missing values. Pairwise constraint information is easier to obtain than label information. Pairwise constraint information describes the relationship between two samples and is divided into two categories: Must-Link (ML) and Cannot-Link (CL), representing whether the two samples belong to the same or different classes, respectively. Specifically, after determining the samples for view restoration using ML constraint information, this invention weights the same view data of these samples based on similarity to restore the missing view, achieving intra-view reconstruction. Secondly, adaptive contrastive learning is used to align view-level features. Adaptive contrastive learning does not force all positive sample pairs to be similar, thus preserving some private information of the view. After fusion, the view-level features are further aligned using pairwise constraint information at the sample-level features to optimize the latent feature space distribution. Finally, a Gaussian mixture model is used to cluster the latent representations of the samples.

[0059] This invention overcomes the limitations of traditional imputation-based incomplete multi-view clustering methods, which may fail to obtain high-quality reconstructed views when there is high missing value or significant differences in information between views. To address these limitations, this invention employs pairwise constraint-weighted imputation. By introducing a small amount of supervisory information, view reconstruction is performed at the same view layer, avoiding cross-view data recovery. Through pairwise constraint-weighted imputation, this invention can still obtain high-quality view imputation even with high missing value, improving clustering performance. Existing contrastive learning algorithms force all positive sample pairs to be similar, which may lead to the loss of private information of views and ignore the complementarity of multi-view data. This invention explicitly preserves the private information of views by introducing adaptive weights, balancing the consistency and complementarity of multi-view data. This invention fully utilizes a small amount of readily available pairwise constraint information to improve clustering performance. Using it for view reconstruction significantly improves the view recovery capability under high missing value scenarios. Furthermore, in the sample feature fusion stage, these pairwise constraint information are further utilized to optimize the clustering structure and reduce the negative impact of spurious negative samples in contrastive learning, significantly improving clustering performance.

[0060] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and substitutions can be made without departing from the technical principles of the present invention, and these improvements and substitutions should also be considered within the scope of protection of the present invention.

Claims

1. An incomplete multi-view clustering method, characterized in that, Includes the following steps: S1. Obtain an incomplete multi-view dataset. The incomplete multi-view dataset contains N samples, each of which has V views. If the vth view of the i-th sample is missing, then the i-th sample is a sample with missing views, and the vth view is a missing view. S2. The missing views are weighted and interpolated based on pairwise constraint information to obtain the complete incomplete multi-view dataset. S3. Input each view of each sample in the completed incomplete multi-view dataset into the corresponding encoder to extract the view-level features corresponding to each view; S4. Fuse the view-level features of all views for each sample to obtain the sample-level features; Based on the sample-level features, the original data of each view is reconstructed through the decoder corresponding to each view; S5. Construct a total loss function to train the encoder and the decoder to obtain a trained encoder and decoder; the total loss function includes an evidence lower bound loss function, a contrastive learning loss function, a consistency loss function, and a semi-supervised loss function based on Euclidean distance between samples; S6. Extract the latent features of all samples using the trained encoder, calculate the clustering probability of each sample based on the latent features of all samples, and form the final clustering result based on the clustering probability of each sample.

2. The method according to claim 1, characterized in that, Step S2 includes the following steps: S2.

1. Select all samples in the incomplete multi-view dataset that form a Must-Link relationship with the missing view sample as ML samples; S2.2 Calculate the similarity between each ML sample and the missing sample in the view; S2.

3. From the ML samples, select the L samples with the highest similarity to participate in imputation; if the number of ML samples is less than a set value... Then all the ML samples mentioned above are used for imputation; S2.4 Calculate the weight of each ML sample participating in the imputation based on the similarity, and perform weighted imputation on the missing view based on the weight to obtain the completed incomplete multi-view dataset.

3. The method according to claim 2, characterized in that, In step S2.4, the weight of each sample participating in the imputation is calculated based on the similarity, and the weight is calculated as follows: in, For the aforementioned similarity, and Representing samples respectively and The Data for each view; The missing view is imputed using weighted interpolation based on the weights, and the expression for the weighted interpolation is: in, Indicates sample One of the L most similar samples in the ML sample The A view.

4. The method according to claim 1, characterized in that, In step S4, the view-level features of all views for each sample are fused into the sample-level features through multiplication. The view-level features include the view feature mean and view sample difference, and the sample-level features include the sample feature mean and sample standard deviation. The expression for the sample-level features is: in, The mean of the sample features; The standard deviation of the sample; The average value of the view features for each view; This represents the view sample difference for each view.

5. The method according to claim 1, characterized in that, In step S5, the contrastive learning loss function is constructed based on view pair similarity and learnable weights. The view pair similarity is calculated using the view-level features, including the mean of view features. The calculation of the view pair similarity is as follows: in, The similarity of the view pairs; The mean of the view features of the missing view; For the remaining samples The mean of view features for each view; , For view pairs; The contrastive learning loss function is constructed based on the similarity of the view pairs and the learnable weights. The expression of the contrastive learning loss function is as follows: in, Let be the contrastive learning loss function; and For the learnable weights, The view can learn the weights of positive samples. The view learnable negative sample weights are randomly initialized before training and maintained during training. .

6. The method according to claim 1, characterized in that, In step S5, a total loss function is constructed based on the evidence lower bound loss function, the contrastive learning loss function, the consistency loss function, and the semi-supervised loss function. The expression for the total loss function is: in, Let the lower bound of the evidence be the loss function; Let be the consistency loss function; Let be the contrastive learning loss function; Let be the semi-supervised loss function; α The weight hyperparameters are used to compare the learning loss. The weight hyperparameters are for the semi-supervised loss.

7. The method according to claim 1, characterized in that, In step S6, the latent features of all samples are input into a Gaussian mixture model to obtain the clustering assignment probability of each sample. The expression of the Gaussian mixture model is: in, Represents the identity matrix. For the first The prior probabilities of each cluster, and satisfying ; Indicates the first Potential characteristics of each category The distribution of , i.e., the Gaussian distribution; Represents all potential features The distribution of, i.e., the mixture Gaussian distribution; The cluster assignment probability of each sample is calculated based on the Gaussian mixture model, and the expression for the cluster assignment probability is: 。 8. An incomplete multi-view clustering system, characterized in that, Includes the following modules: The data acquisition module is used to acquire an incomplete multi-view dataset, which contains N samples, each sample has V views, and if the vth view of the i-th sample is missing, then the i-th sample is a sample with missing views, and the vth view is a missing view. The pairwise constraint weighted interpolation module is used to perform weighted interpolation on the missing views based on pairwise constraint information to obtain a completed incomplete multi-view dataset. The multi-view feature extraction module is used to input each view of each sample in the completed incomplete multi-view dataset into the corresponding encoder to extract the view-level features corresponding to each view. The feature fusion and reconstruction module is used to fuse the view-level features of all views of each sample to obtain the sample-level features of that sample. Based on the sample-level features, the original data of each view is reconstructed through the decoder corresponding to each view; The model joint training module is used to construct a total loss function to train the encoder and the decoder, thereby obtaining a trained encoder and decoder; the total loss function includes an evidence lower bound loss function, a contrastive learning loss function, a consistency loss function, and a semi-supervised loss function based on Euclidean distance between samples; The clustering output and label assignment module is used to extract the latent features of all samples through the trained encoder, calculate the clustering assignment probability of each sample based on the latent features of all samples, and form the final clustering result according to the clustering assignment probability of each sample.

9. A storage medium, characterized in that, The storage medium stores a computer program that, when executed by a processor, can implement the incomplete multi-view clustering method according to any one of claims 1-7.

10. A computer device, characterized in that, The system includes a memory and a processor, wherein the memory stores computer-readable instructions, and the computer-readable instructions, when executed by the processor, cause the processor to perform an incomplete multi-view clustering method as described in any one of claims 1-7.

Images