An unsupervised multi-view clustering method for clustering information noisy problem

Abstract

The application discloses an unsupervised multi-view clustering method for a clustering information noisy problem, performs feature processing on multi-view data, and keeps consistent feature vectors in a feature space; a decoupling variational autoencoder is established, and a loss function with a regularization constraint is used to perform feature extraction on each view to obtain private information and shared information of each view; a reconstruction loss function of the decoupling variational autoencoder is defined, a corresponding relationship between the private information and the shared information of each view output by the decoupling variational autoencoder is utilized, consistency of separation of the private information and the shared information is ensured, and double consistency constraints of the reconstruction loss and a contrast loss are set, and different latent feature distinguishability is enhanced through contrast learning; and finally, unsupervised multi-view clustering is performed by using the reconstructed and optimized feature information. The application combines the private information and the shared information of each view, designs a double consistency constraint optimization DVAE model, and improves the accuracy and robustness of the clustering effect.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, relates to image recognition, and particularly to an unsupervised multi-view clustering method for addressing the problem of noisy clustering information. Background Technology

[0002] In the rich information flow of the internet, the complexity and diversity of data often lead to information noise, especially in multi-view clustering analysis. The diversity and consistency of multi-view data are often overlooked, making direct clustering of this data difficult and resulting in a large amount of redundant information. This redundancy can cause chaotic clustering results, affecting clustering quality and further complicating the clustering task.

[0003] Current clustering methods mainly include K-means-based multi-view clustering algorithms and subspace learning methods. K-means clustering, by concatenating multi-view data, ignores the correlation and consistency between views, making it susceptible to noise and resulting in poor clustering performance. Subspace learning methods, such as CCA and MCCA, align multi-view data by learning common subspaces, but they only address view alignment to a certain extent, and their ability to handle noisy information remains limited.

[0004] In addition, there are contrastive learning and other methods that extract latent features from multi-view data through deep network models. Although they perform well in feature extraction, existing deep clustering methods focus on exploring the collectively shared information of multiple views, ignoring the importance of individual private information in a single view. Summary of the Invention

[0005] In order to overcome the shortcomings of the prior art, the present invention aims to provide an unsupervised multi-view clustering method for the problem of noisy clustering information, which takes into account both private and shared information and achieves dual consistency of private and shared information, thereby solving the problem of poor view clustering effect under noisy information.

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

[0007] An unsupervised multi-view clustering method for addressing the problem of noisy clustering information includes the following steps:

[0008] Step 1: Perform feature processing on the multi-view data to maintain consistency of the feature vectors in the feature space;

[0009] Step 2: Establish a decoupled variational autoencoder and use a loss function with regularization constraints to extract features from each view after processing in Step 1, obtaining the private information and shared information of each view, i.e., the latent feature vector; the private information is the feature unique to a certain view, and the shared information is the feature that exists in all views.

[0010] Step 3: Define the reconstruction loss function of the decoupled variational autodecoder. Utilize the correspondence between the private and shared information of each view output by the decoupled variational autoencoder to ensure the consistency of the separation between private and shared information. In conjunction with avoiding trivial solutions, set a dual consistency constraint between the reconstruction loss and the contrastive loss, and enhance the discriminative power of different latent features through contrastive learning.

[0011] Step 4: Use the decoupled variational autodecoder to reconstruct the optimized feature information for unsupervised multi-view clustering.

[0012] In one embodiment, the multi-view data consists of different objects captured from multiple different perspectives; the multi-view data undergoes data feature processing, including size normalization and noise removal.

[0013] In one embodiment, the hidden layer of the decoupled variational autoencoder incorporates a ReLU activation function to convert the input view features into latent feature vectors, and performs latent space decoupling in the decoupled variational autoencoder to capture an independent generative factor for each view, i.e., unique private or shared information in the same view.

[0014] In one embodiment, both the decoupled variational autoencoder and the decoupled variational autodecoder follow a Gaussian distribution, forming a DAVE model. The decoupled variational autoencoder uses a prior distribution p(z) as input and outputs q(z|x), where z is the latent feature vector and x is the view. The regularization constraint is implemented using KL divergence, expressed as:

[0015]

[0016] Where, μ i and σ i q(z|x) represents the mean and standard deviation of the i-th latent feature vector, respectively, to ensure that the posterior distribution of private and shared information, i.e., q(z|x), is close to the prior distribution p(z). N is the number of views. The goal of the constraint is to make the information of each feature as independent as possible, i.e., to reduce the correlation between features and improve the decoupling ability of the model.

[0017] In one embodiment, step 3, the reconstruction loss function, is expressed as follows:

[0018]

[0019] Where E v and D v These represent encoding and decoding the v-th view using a decoupled variational autoencoder and a decoupled variational decoder, respectively. (v) The set of private and shared information representing the characteristics of a view, where N is the number of views.

[0020] In one embodiment, step 3 involves extracting latent feature vectors for each view from the feature variables of all views input into the DAVE model by combining a decoupled variational autoencoder with an adversarial network. The latent feature vectors of different views are learned by maximizing the lower bound of mutual information, while simultaneously minimizing the similarity between private and shared information. This can be expressed as:

[0021]

[0022] In the formula, I represents mutual information, describing the correlation between two latent feature variables; a higher value indicates a stronger correlation. H represents entropy, describing the uncertainty of a random feature variable; a higher value indicates greater uncertainty in the value of the random variable. α is a weighting parameter, and Z... (1) and Z (2) These represent the latent feature variables of the two views respectively; the goal of contrastive learning is to maximize the mutual information between the two latent feature variables while minimizing their respective entropies.

[0023] In one embodiment, step 3 uses a PoE expert production model to induce inference on shared information in the DVAE. By optimizing the shared information of a view, it can better capture and reconstruct shared information between other views. In the case of missing views and multiple views, the DAVE model combines PoE technology with local consensus regularization and local embedding consensus regularization to reconstruct shared information in the DAVE encoder, thereby achieving alignment of network information consistency. The formula of the PoE expert production model is as follows:

[0024]

[0025] In the POE model, shared inference is defined, where ∞ represents separation, and V... (i) Let V represent the i-th view, and Z represent the view set. s p(Z) represents shared information. i ) represents the posterior probability of the information. This is equivalent to viewing all private information. For all views to share information, a posterior distribution of the i-th view is p(z). i |V (i) ).

[0026] In one embodiment, the objective function loss of the POE technique is as follows:

[0027]

[0028] Let be the objective function of POE, expressed as the local consensus regularization of different view sets. and Local Embedded Consensus Regularization The total expected value of the probability is minimized by the objective function loss, enabling the DAVE model to learn feature representations that both preserve the local structure of each view and ensure consistent information sharing across different views.

[0029] In one embodiment, step 4 involves decoupling the variational autodecoder to perform unsupervised multi-view clustering by taking reconstructed independent latent feature representations from the latent space optimized by POE technology and contrastive learning.

[0030] Compared with existing technologies, this invention extracts and separates features by decoupling variational autoencoders to obtain private and shared information for each view. It designs a dual consistency constraint to optimize the DVAE model and can further promote information consistency by utilizing PoE technology to infer shared network information. Finally, cluster analysis is performed. This solves the ambiguity and missing information problems caused by noisy clustering information, improving the accuracy and robustness of clustering results. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the main process of the present invention.

[0032] Figure 2 This is a flowchart illustrating the method implementation in an embodiment of the present invention.

[0033] Figure 3 This is a schematic diagram of the PoE technology principle.

[0034] Figure 4 This refers to the clustering performance metrics of the embodiments of this invention.

[0035] Figure 5 This is the clustering effect after 200 rounds of training in Embodiment 2 of the present invention. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0037] like Figure 1 and Figure 2 As shown, the present invention provides an unsupervised multi-view clustering method for addressing the problem of noisy clustering information, comprising the following steps:

[0038] Step 1: To improve the accuracy of important feature recognition and classification, feature processing is performed on the multi-view data, including size standardization and noise removal, in order to maintain the consistency of feature vectors in the feature space.

[0039] This invention utilizes the DAVE model, with the dataset sourced from a social media platform (Instagram). The platform includes user information views (user ID, username, bio, number of followers, location), text comment views (comment content, time, number of likes, sentiment tags), image and video views (media content links, descriptions, number of likes, metadata), interaction information views (interactions between users), and hashtag views (hashtags, popularity, relevance).

[0040] In this embodiment, the multi-view data consists of different objects captured from multiple different perspectives. Specifically, it comprises 14,000 images collected across 7 categories and 5 perspectives. The 7 categories include airplanes, cars, cats, dogs, chairs, tables, and vases. The 5 perspectives are front, side, top, oblique, and bottom views. Each image in each category contains latent feature information. The extracted latent features include shared and private information.

[0041] Given the nature of multi-view datasets, data acquisition and cleaning are necessary. The data used in this invention is designed for clustering noisy network information, and an appropriate API is selected to access the data. This invention uses the Instagram Graph API to obtain the required images and videos. Python code is written to construct a module for obtaining API keys and access tokens to acquire the data. Then, the acquired multi-view data is cleaned and preprocessed, including deduplication, handling missing values, and standardization. Furthermore, for the data in this invention, bilinear interpolation is used to fill in missing values ​​in the multi-view image data, considering both dimensions of the image and using a weighted average of four adjacent pixels. The specific formula is given below:

[0042]

[0043] (x,y) are the coordinates of the point to be interpolated, and the vector form better illustrates how to combine the function values ​​of four known points to estimate f(p). For example, a 3×3 image matrix is ​​given by the following formula, where there is one missing value.

[0044]

[0045] This embodiment performs data feature processing on multi-view data, including size normalization and noise removal. All views are resized to the same size to eliminate the effects of size differences. A filter (median filtering) is then used to remove noise from the image.

[0046] Specifically, size standardization adjusts features in each view to a uniform size range, eliminating size inconsistencies caused by differences in shooting angle, distance, or resolution. Size standardization makes feature extraction and alignment of the model more accurate and reliable. By standardizing the size to 256×256, the model's generalization ability is improved.

[0047] For a pixel P in the image of the view, a 3×3 median filter window N(P) is applied. If the pixel values ​​within N(P) are arranged in row-major order as [10, 20, 30, 40, 50, 60, 70, 80, 90], then the median-filtered value of pixel P will be updated to 50, which is the median of these nine values. This process is repeated for every pixel in every view, ultimately resulting in a smoothed version of the image with salt-and-pepper noise removed.

[0048] Step 2: Establish the DAVE model, which consists of a decoupled variational autoencoder and a decoupled variational autodecoder, using a Gaussian distribution to output features from a fully connected network. The decoupled variational autoencoder compresses the original features to the same dimension to obtain the latent feature variables of the view. The decoupled variational autodecoder is responsible for mapping the input data to the latent space, sampling a point from the latent space, and mapping it back to the original data space. In DAVE, the latent information feature variables are sampled from a Gaussian distribution.

[0049] The decoupled variational autoencoder uses a prior distribution p(z) as input and outputs q(z|x), where z is the latent feature vector and x is the view. For the decoupled variational autoencoder, a loss function with regularization constraints is used to extract features from each view after step 1, obtaining the private and shared information of each view, i.e., the latent feature vector. Private information consists of features unique to a particular view, which are difficult to find corresponding examples in other views, such as a specific angle of an object. Shared information consists of features present in all views, which exhibit similar behavior or can be mapped to the same concept across different views. Applying the decoupled variational autoencoder can improve the reconstruction quality and consistency of view information, effectively extracting view features.

[0050] To ensure the network can thoroughly learn the original feature information, this invention adds a ReLU activation function to the hidden layer of the decoupled variational autoencoder, thereby converting the input view features into latent feature vectors. The ReLU activation function is given by the following formula:

[0051] f(x) = max(0,x)

[0052] For any input view feature x: if x>0, then f(x) = x; if x≤0, then f(x) = 0.

[0053] The decoupled variational autoencoder reconstructs the original input from the information distribution of the view, thereby extracting the latent feature vectors in the view. In order to make the latent feature vectors as independent as possible, the latent space is decoupled in the decoupled variational autoencoder, and only one independent generating factor of the data is captured for each view, that is, the unique private information or shared information in the same view.

[0054] The decoupled variational autoencoder and decoupled variational autodecoder of this invention both follow a Gaussian distribution and together constitute a DAVE model. In the DAVE model, a loss function constrained by a regularization term is used to make each different feature information as independent of other information as possible.

[0055] Regularization constraints can enhance the decoupling capability of DVAEs. This constraint can be achieved using KL divergence, which measures the difference between two probability distributions, i.e., penalizing the deviation between q(z|x) and the prior p(z). KL divergence can be simplified as:

[0056]

[0057] Where, μ i and σ i q(z|x) represents the mean and standard deviation of the i-th latent feature vector, respectively, ensuring that the posterior distribution of private and shared information, i.e., q(z|x), approximates the prior distribution p(z). N is the number of views. The goal of the constraint is to make the information of each feature as independent as possible, i.e., to reduce the correlation between features and improve the decoupling ability of the model.

[0058] Through the steps described above, the decoupled variational autoencoder maps data from different views to a shared latent space. The data for each view is represented as the distribution parameters of the latent variables, namely the mean and standard deviation. Mapping the feature representations in existing views to similar representations can be used to infer information from missing views.

[0059] Step 3: For the decoupled variational autodecoder, define its reconstruction loss function to force the decoupled variational autoencoder to output the feature correspondence between the private information and shared information of each view, ensuring the consistency of the separation between private information and shared information, that is, the model can accurately distinguish which information is private information specific to a single view and which information is shared information.

[0060] This step, based on the DAVE model, ensures similarity between internal data points in different views and high correlation between data points within the same view. Each view is passed to a decoupled variational autoencoder, which then uses a custom loss function. Specifically, during DAVE model training, based on the existing training, adjustments are made to optimize the constraints on the decoupled variational autoencoder. Figure 1To ensure consistency, a reconstruction loss function is introduced. The goal of the reconstruction function is typically to make the reconstructed information as close as possible to the original information, achieved through the Adam optimizer's reconstruction loss. To ensure that the latent feature variables encompass information about the structure and clustering of the data within the same view, the reconstruction loss function is expressed as follows:

[0061]

[0062] Where E v and D v These represent encoding and decoding the v-th view using a decoupled variational autoencoder and a decoupled variational decoder, respectively. (v) The set of private and shared information representing the characteristics of a view, where N is the number of views.

[0063] Through the steps described above, the decoupled variational autodecoder can remap latent variables back to the original data space, attempting to reconstruct the data for all views. This allows for the generation of predictions or estimates of the features of samples in missing views.

[0064] Step 4: Combining the avoidance of trivial solutions, a dual consistency constraint is set for reconstruction loss and contrastive loss. Contrastive learning enhances the discriminative power of different latent features. For example, the private and shared information of two views are compared, and view feature representations are learned by narrowing the distance between shared information and widening the distance between private information.

[0065] The feature representations learned by the DVAE model are often the same or nearly the same for all views. To avoid trivial solutions, contrast loss is introduced.

[0066] In the DAVE model, among the different feature variables of all views input, latent feature vectors of each view are extracted by decoupling variational autoencoders and combining them with adversarial networks. The latent feature vectors of different views are learned by maximizing the lower bound of mutual information, while simultaneously minimizing the similarity between private and shared information, as expressed below:

[0067]

[0068] In the formula, I represents mutual information, describing the correlation between two latent feature variables; a higher value indicates a stronger correlation. H represents entropy, describing the uncertainty of a random feature variable; a higher value indicates greater uncertainty in the value of the random variable. α is a weighting parameter, and Z... (1) and Z (2) These represent the latent feature variables of the two views respectively; the goal of contrastive learning is to maximize the mutual information between the two latent feature variables while minimizing their respective entropies.

[0069] To perform comparative learning, this invention also needs to construct positive and negative samples of information. First, positive sample pairs are constructed by selecting a pair of samples from the dataset. If this pair of samples has similar representations in multiple views or shared views, these samples are considered to belong to the same category. Conversely, negative sample pairs are constructed by selecting another pair of samples from the dataset, classifying those with significant differences in views as different categories. Positive sample pairs are... Negative sample pairs The view is represented by the following features:

[0070]

[0071] The cosine similarity formula is given below:

[0072]

[0073] By optimizing through contrastive learning, the loss is calculated, and the model is optimized to minimize the distance between positive sample pairs and maximize the distance between negative sample pairs. This promotes the learning of more discriminative and representative features in the latent or feature space. Specifically, the similarity between shared information and positive sample pairs is first calculated, then the similarity between private information and negative sample pairs is calculated, and the similarity values ​​are concatenated. Finally, the InfoNCE loss is used to achieve the goal of contrastive learning: to shorten the distance between positive sample pairs and widen the distance between negative sample pairs. This process learns common feature representations between different views by maximizing the lower bound of mutual information, and combines the latent feature representation optimized by POE technology for unsupervised clustering, ultimately completing unsupervised clustering for noisy network information. The contrastive loss is calculated using the following formula.

[0074]

[0075] Among them, z i It is shared information x i The representation features, It is a positive sample Feature representation, private information z j Negative sample pairs Feature representation, sim(z) i ,z j ) represents sample z i and z j The similarity between them can be expressed using cosine similarity. τ represents the smoothness of the temperature coefficient adjustment distribution, and N is the number of views in the batch.

[0076] This step allows the information between two views to be as similar as possible while retaining sufficient diversity to distinguish potential features. Therefore, by reconstructing the loss function and introducing a double consistency constraint to avoid trivial solutions, the final clustering result can be improved.

[0077] Step 5: Use POE technology to infer shared information and align view information for consistency. Finally, use a decoupled variational autodecoder to reconstruct the optimized feature information for unsupervised multi-view clustering.

[0078] To improve the latent feature alignment of the DAVE model, in cases of missing views and multiple views, private and shared information are separated by combining the latent feature vectors from the input views. Specifically, this step uses a Proof-of-Earth (POE) expert production model to induce inference on shared information in the DVAE. By optimizing the shared information of a view, it can better capture and reconstruct shared information between other different views. This allows for the combination of other views to maintain consistency of shared information, aligning view network information and achieving model reconstruction optimization.

[0079] Through Proof-of-Effect (POE) technology, the model can learn features shared across different views and reconstruct this shared information in the latent space. This helps the model infer features of the missing view by utilizing information from other views. (Reference) Figure 3 As shown, Proof-of-Effect (POE) technology helps improve the clustering accuracy of models on data by optimizing the latent feature vector representation and reconstructing shared information. This is because a more comprehensive and consistent feature representation can better distinguish and cluster different data points. For example, given two views X... (1) ,X (2) When in view X (2) In the case of missing information, q(Z) can be retrieved from the remaining shared information. skered |X) Inference View X (1) The information is as follows, and vice versa. The specific formula is given below:

[0080]

[0081] Where X is the set of views, p(X|Z) is the posterior distribution of each private or shared network, and the expectation is inferred from this formula.

[0082] Similarly, in the cases of missing views and multiple views, the DAVE model combines PoE technology with local consensus regularization and local embedding consensus regularization to reconstruct the shared information in the DAVE encoder, thereby achieving consistency in the aligned network information. The formula for the PoE expert production model is as follows:

[0083]

[0084] In the POE model, shared inference is defined, where ∞ represents separation, and V... (i) Let V represent the i-th view, and Z represent the view set. s p(Z) represents shared information. i ) represents the posterior probability of the information. This is equivalent to viewing all private information. For all views to share information, a posterior distribution of the i-th view is p(z). i |V (i) From a shared network q(Z) s Private and shared information networks separated in |V)

[0085] Specifically, if network noise exists in both single-view and multi-view aspects of the DAVE model, for each view V, the mean μ of the latent variable Z is obtained through the model. (v) and variance δ (v)2 The specific formula is given below:

[0086] q(z∣x (v) )=η(z∣μ (v) ,δ (v)2 )

[0087] By combining the mean and variance of multiple views, the mean and variance of the joint distribution are calculated. The specific calculation formula is given below:

[0088]

[0089] In this step, the objective function loss of the POE technique is defined as follows:

[0090]

[0091] The objective function of POE is expressed as local consensus regularization for different view sets. and Local Embedded Consensus Regularization The expected total probability is minimized by the objective function loss L, enabling the DAVE model to learn feature representations that preserve the local structure of each view and ensure consistent information sharing across different views. This helps improve the performance of clustering, classification, and other downstream tasks in multi-view learning tasks.

[0092] Through the above steps, the decoupled variational autodecoder of this invention reconstructs independent latent feature representations, i.e., private or shared information, from the latent space optimized by POE technology and contrastive learning. Based on this, unsupervised multi-view clustering can effectively reduce noisy information in the views and improve the accuracy and consistency of clustering.

[0093] The training process of the model in this invention is as follows: First, the model underwent 200 rounds of pre-training, including reconstruction loss and dual constraints. Then, it was fine-tuned for 50 rounds. Finally, the Adam optimizer in the PyTorch framework was used to optimize the total loss, with a batch size of 128 and a learning rate of 0.0001. For the dataset of this invention, the missing data rate was fixed at 0.5. For methods that can only process complete view data, samples were randomly deleted while retaining at least one view, and then the mean of the same view was used to fill in the missing data. During training, the total loss in the DVAE model was calculated and backpropagation and optimization were performed to ensure that DVAE could effectively reconstruct the input view data and make the distribution of latent variable features close to a Gaussian distribution. After optimizing the model through contrastive learning and POE techniques, the DVAE decoder outputs the reconstructed view data for unsupervised clustering.

[0094] Accuracy (ACC), Normalized Mutual Information (NMI), and Purity (PUR) are three commonly used metrics to evaluate the clustering results. In this embodiment, the evaluated accuracy is 0.895, normalized mutual information is 0.732, and purity is 0.843. This invention has produced excellent results in clustering noisy network information. Figure 4 and Figure 5 As shown.

[0095] Therefore, this invention achieves consistency in structural features and shared information in multi-view network information, and solves the problem of semantic noise in multi-view networks with private and shared information.

Claims

1. An unsupervised multi-view clustering method for the problem of cluster information noise, characterized in that, Comprising the following steps: Step 1, feature processing is performed on the multi-view data to keep the feature vectors of the feature space consistent; the multi-view data is images of different objects captured from multiple different angles; Step 2, a decoupled variational autoencoder is established, and a loss function with regularization constraints is used to extract features from each view processed in step 1, obtaining private information and shared information, i.e., latent feature vectors, of each view; the private information is a feature unique to a view, and the shared information is a feature present in all views; Step 3, define the reconstruction loss function of the decoupled variational auto-decoder, use the decoupled variational autoencoder to output the correspondence between the private information and the shared information of each view, and ensure the consistency of the separation of the private information and the shared information; and set double consistency constraints of reconstruction loss and contrast loss to avoid trivial solutions, and enhance the discriminability of different latent features through contrast learning; Step 4, use the decoupled variational auto-decoder to reconstruct the optimized feature information for unsupervised multi-view clustering; Wherein: The decoupled variational autoencoder and the decoupled variational auto-decoder both follow Gaussian distribution, and the two constitute a DAVE model, the decoupled variational autoencoder uses a prior distribution As input, the output is where z is a latent feature vector, x is a view, the regularization term is constrained by KL divergence, and is expressed as: where, and are the mean and standard deviation of the i-th latent feature vector corresponding to the private information and shared information, respectively, ensuring that the posterior distribution of the private information and shared information, i.e. is close to the prior distribution , is the number of views, and the goal of the constraint is to make the information of each feature as independent as possible, i.e., to reduce the correlation between features and improve the decoupling ability of the model; The reconstruction loss function of step 3 is as follows: in and These represent the use of a decoupled variational autoencoder and a decoupled variational decoder to process the first... Each view is encoded and decoded. The set of private and shared information representing the characteristics of a view, where N is the number of views; In step 3, among all the feature variables of the input views in the DAVE model, the latent feature vectors of each view are extracted by the decoupled variational autoencoder combined with the adversarial network, the latent feature vectors of different views are represented by maximizing the lower bound of mutual information, and the similarity of private information and shared information is minimized, represented as: where, represents mutual information, describing the degree of correlation between two latent feature variables, the higher the value, the more relevant the two are, represents entropy, describing the uncertainty of a random feature variable, the higher the value, the more uncertain the value of the random variable is, is a weight parameter, and represent the latent feature variables of the two views respectively; the goal of contrastive learning is to maximize the mutual information between the two latent feature variables while minimizing their respective entropies.

2. The unsupervised multi-view clustering method for the cluster information noisy problem according to claim 1, wherein, The multi-view data is subjected to data feature processing, including size standardization and noise removal.

3. The unsupervised multi-view clustering method for the cluster information noisy problem according to claim 1, wherein, The hidden layer of the decoupled variational autoencoder adds a Relu activation function to convert the input view features into latent feature vectors, and performs latent space decoupling in the decoupled variational autoencoder to capture one independent generating factor for each view data, i.e., unique private information or shared information in the same view.

4. The unsupervised multi-view clustering method for the cluster information noisy problem according to claim 1, wherein, In step 3, the POE expert production model is used to induce inference on the shared information in DVAE, so that the shared information of one view can better capture and reconstruct the shared information between other views; in the case of missing views and multi-views, the DAVE model combines POE technology to use local consensus regularization and local embedding consensus regularization to reconstruct the shared information in the DAVE encoder, thereby completing the alignment of network information consistency, and the formula of the POE expert production model is as follows: shared inference is defined in the POE model, is separated, represents the ith view, V represents a view set, represents shared information, represents the information posterior probability, is equal to the total private information of the view, is the total shared information of the view, and the posterior distribution of the ith view is .

5. The unsupervised multi-view clustering method for the cluster information noisy problem according to claim 4, characterized in that, The objective function loss of the POE technology is as follows: The objective function for POE, denoted as local consensus regularization of different view sets and local embedding consensus regularization The total expectation of the probability of the objective function loss by minimizing the DAVE model can learn to maintain the local structure of each view and ensure consistent feature representation between different views that share information.

6. The unsupervised multi-view clustering method for the cluster information noisy problem according to claim 5, wherein, In step 4, the decoupled variational auto-decoder adopts the reconstructed independent latent feature representation from the latent space optimized by the POE technology and contrast learning to perform unsupervised multi-view clustering.

Images