Unsupervised multi-view clustering method for problem of noisy clustering information

The private information and shared information of multi-view data are extracted by decoupling variational autoencoder and decoupling variational autodecoder, and combined with POE technology to perform information alignment, the view correlation caused by noisy information in multi-view clustering is solved and Consistency problem achieves a more accurate and robust clustering effect.

CN119068224AActive Publication Date: 2024-12-03INNER MONGOLIA UNIV OF TECH

Patent Information

Application Number
CN202411137936.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-19
Publication Date
2024-12-03
Estimated Expiration
2044-08-19

AI Technical Summary

Technical Problem

Existing multi-view clustering methods are difficult to maintain correlation and consistency between views when processing noisy information, resulting in poor clustering and ignoring the importance of private information.

Method used

Using decoupled variational autoencoder and decoupled variational autodecoder, through regularization constraints and contrast learning, private information and shared information of each view are extracted, and dual consistency constraints are designed to ensure the private information and shared information. Separate consistency, combine POE technology to perform information alignment, and realize unsupervised multi-view clustering.

Benefits of technology

It improves the accuracy and robustness of the clustering effect, reduces ambiguity and missingness, and enhances the distinction and information consistency of feature extraction.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119068224A_ABST
    Figure CN119068224A_ABST
Patent Text Reader

Abstract

The invention discloses an unsupervised multi-view clustering method for solving the problem of noisy clustering information, and the method comprises the steps: carrying out the feature processing of multi-view data, and maintaining the consistency of feature vectors of a feature space; establishing a decoupling variational auto-encoder, and performing feature extraction on each view by using a regularized constrained loss function to obtain private information and shared information of each view; defining a reconstruction loss function of the decoupling variational self-decoder, outputting a corresponding relation between private information and shared information of each view by using the decoupling variational self-decoder, ensuring separation consistency of the private information and the shared information, and setting dual consistency constraints of reconstruction loss and comparison loss, enhancing the distinction degree of different potential features through comparative learning; and finally, performing unsupervised multi-view clustering by using the feature information after reconstruction optimization. According to the method, the private information and the shared information of each view are combined, the dual consistency constraint optimization DVAE model is designed, and the accuracy and the robustness of the clustering effect are improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

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

[0002] In the rich information flow of the Internet, due to the complexity and diversity of data, there is often a problem of noisy information, especially in multi-view clustering analysis. The diversity and consistency of multi-view data are usually ignored, which makes it difficult to cluster these data directly, and there is a lot of redundant information, which may cause confusion in the clustering results and affect the clustering quality, making the clustering task more complicated.

[0003] Current clustering methods mainly include K-means-based multi-view clustering algorithms and subspace learning methods. The K-means clustering algorithm ignores the correlation and consistency between views by splicing multi-view data, and is easily affected by noisy information noise, resulting in poor clustering effect. Subspace learning methods such as CCA and MCCA align multi-view data by learning a common subspace, but can only solve the view alignment problem to a certain extent, and the ability to process noisy information is still limited.

[0004] In addition, there are methods such as contrastive learning, which extract latent features of multi-view data through deep network models. Although they perform well in feature extraction, existing deep clustering methods focus on exploring the collective shared information of multiple views and ignore the importance of personal private information in a single view. Summary of the invention

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

[0006] In order to achieve the above object, the technical solution adopted by the present invention is:

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

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

[0009] Step 2: Establish a decoupled variational autoencoder and use a regularized constrained loss function to extract features from each view processed in step 1 to obtain private information and shared information, i.e., a potential feature vector, of each view; the private information is a feature unique to a certain view, and the shared information is a feature that exists in all views;

[0010] Step 3: Define the reconstruction loss function of the decoupled variational autodecoder, 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 private information and shared information; and set the dual consistency constraints of reconstruction loss and contrast loss in combination with avoiding trivial solutions, and enhance the discrimination of different potential features through contrastive learning;

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

[0012] In one embodiment, the multi-view data are different objects captured from multiple different viewing angles; and data feature processing is performed on the multi-view data, including size normalization and noise removal.

[0013] In one embodiment, a Relu activation function is added to the hidden layer of the decoupled variational autoencoder to convert the input view features into a latent feature vector, and latent space decoupling is performed in the decoupled variational autoencoder to capture an independent generating factor of the data for each view, that is, unique private information or shared information in the same view.

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

[0015]

[0016] Among them, μ i and σ i are the mean and standard deviation of q(z|x) corresponding to the i-th potential feature vector, ensuring that the posterior distribution of private information 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, that is, to reduce the correlation between features and improve the decoupling ability of the model.

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

[0018]

[0019] Where E v and D v Denotes encoding and decoding of the v-th view using the decoupled variational autoencoder and decoupled variational decoder, respectively, X (v) Represents the set of private and shared information of view features, where N is the number of views.

[0020] In one embodiment, in step 3, among the feature variables of all views input into the DAVE model, the potential feature vector of each view is extracted by combining the decoupled variational autoencoder with the adversarial network, and the potential feature vectors of different views are compared and learned by maximizing the lower bound representation of the mutual information, while minimizing the similarity between private information and shared information, which is expressed as:

[0021]

[0022] In the formula, I represents mutual information, which describes the correlation between two potential feature variables. The higher the value, the more correlated they are. H represents entropy, which describes the uncertainty of a random feature variable. The higher the value, the more uncertain the value of the random variable is. α is a weight parameter, and Z (1) and Z (2) They 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 entropy.

[0023] In one embodiment, the step 3 uses the POE expert production model to induce inference of the shared information in the DVAE, and optimizes the shared information of a view so that it can better capture and reconstruct the shared information between other views; in the case of missing views and multiple views, the DAVE model combines the 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. The formula of the POE expert production model is as follows:

[0024]

[0025] In the POE model, shared inference is defined, ∞ is separation, V (i) represents the i-th view, V represents the view set, Z s represents shared information, p(Z i ) represents the information posterior probability, Equal to view all private information, For all views sharing information, the posterior distribution of the i-th view is p(z i |V (i) ).

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

[0027]

[0028] is the objective function of POE, expressed as the local consensus regularization of different view sets and local embedding consensus regularization By minimizing the objective function loss, the DAVE model can learn feature representations that can maintain the local structure of each view and ensure consistent shared information between different views.

[0029] In one embodiment, in step 4, the decoupled variational autodecoder takes the reconstructed independent latent feature representation from the latent space optimized by POE technology and contrastive learning, and performs unsupervised multi-view clustering.

[0030] Compared with the prior art, the present invention extracts separate features through decoupling variational autoencoders, obtains private information and shared information of each view, designs a dual consistency constraint optimization DVAE model, and can use POE technology to infer shared network information to further promote information consistency, and finally perform cluster analysis. It can solve the ambiguity and missing degree problems caused by noisy clustering information and improve the accuracy and robustness of clustering effects. BRIEF DESCRIPTION OF THE DRAWINGS

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

[0032] Figure 2 It is a flowchart of 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 It is the index performance of clustering in the embodiment of the present invention.

[0035] Figure 5 This is the clustering effect after 200 rounds of training in the embodiment of the present invention. DETAILED DESCRIPTION

[0036] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention are further described in 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 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 multi-view data, including size standardization and noise removal, to keep the feature vectors in the feature space consistent.

[0039] The present invention uses the DAVE model, and the data set comes from a social media platform (Instagram). The platform includes user information view (user ID, user name, personal profile, number of followers, location), text comment view (comment content, time, number of likes, emotional tags), picture and video view (media content link, description, number of likes, metadata), interaction information view (interaction information between users) and topic tag view (topic tags, popularity, relevance).

[0040] In this embodiment, multi-view data refers to different objects captured from multiple different perspectives. Specifically, it is 14,000 images of 7 categories and 5 perspectives collected. The 7 categories include airplanes, cars, cats, dogs, chairs, tables, and vases. The 5 perspectives are front, side, top, oblique, and bottom perspectives. Among them, each perspective picture of each category contains potential feature information. The extracted potential features are shared information and private information.

[0041] In view of the nature of the multi-view data set, data acquisition and cleaning are required. The data used in the present invention is for the task of clustering noisy information on the network, and an appropriate API is selected to access its data. The present invention uses Instagram GraphAPI to obtain the required pictures and videos, wherein a python code is written to construct a module for obtaining API keys and access tokens to obtain data. Then, the acquired multi-view data is cleaned and preprocessed, including steps such as deduplication, processing missing values, and standardization. And for the data of the present invention, the bilinear interpolation method is used for missing multi-view image data. Considering the two dimensions of the image, the missing values ​​are filled by the weighted average of the four adjacent pixels. The specific formula is given below:

[0042]

[0043] (x,y) are the coordinates of the point you want to interpolate, and the vector form better shows how to combine the function values ​​at the four known points to estimate f(p). For example, the 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 standardization and noise removal. All views are adjusted to the same size to eliminate the impact of size differences. A filter (median filter) is used to remove noise in the image.

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

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

[0048] Step 2: Establish a DAVE model, which consists of a decoupled variational autoencoder and a decoupled variational self-decoder, and uses 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 self-decoder 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 the 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 processed in step 1 to obtain private information and shared information, i.e., latent feature vectors, of each view. Private information is a feature unique to a certain view, which is difficult to find a corresponding feature in other views, such as a specific angle of an object. Shared information is a feature that exists in all views, which behaves similarly in different views or can be mapped to the same concept. The application of the decoupled variational autoencoder can improve the effectiveness of the reconstruction quality and consistency in the view information and effectively extract view features.

[0050] In order to ensure that the network can thoroughly learn the original feature information, the present invention adds a ReLU activation function to the hidden layer of the decoupled variational autoencoder, so that the input view features can be converted into potential 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 to extract the latent feature vector in the view. In order to make the latent feature vector as independent as possible, latent space decoupling is performed in the decoupled variational autoencoder to capture only one independent generating factor of the data for each view, that is, the unique private information or shared information in the same view.

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

[0055] The decoupling ability of DVAE can be enhanced by regularization constraints. This constraint can be implemented through KL divergence, which measures the difference between two probability distributions, that is, the deviation between the penalty q(z|x) and the prior p(z). KL divergence can be simplified as:

[0056]

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

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

[0059] Step 3: For the disentangled variational autodecoder, define its reconstruction loss function to force the disentangled variational autoencoder to output the feature correspondence between the private information and the shared information of each view, and ensure the consistency of the separation of 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 is based on the DAVE model, ensuring that the internal data points of different views are similar and the data points of the same view have a high correlation, and passing each view to the decoupled variational autoencoder to use a custom loss function. When training the DAVE model, based on the original training, in order to optimize the decoupled variational autoencoder constrained view Figure 1Therefore, the reconstruction loss function is introduced. The goal of the reconstruction function is usually to make the reconstructed information as close to the original information as possible, which is achieved through the Adam optimizer reconstruction loss. In order for the potential feature variables to contain the structure and clustering information of the data in the same view, the reconstruction loss function is expressed as follows:

[0061]

[0062] Where E v and D v Denotes encoding and decoding of the v-th view using the decoupled variational autoencoder and decoupled variational decoder, respectively, X (v) Represents the set of private and shared information of view features, where N is the number of views.

[0063] Through the above steps, the disentangled variational autodecoder is able to remap the latent variables back to the original data space and try to reconstruct the data of all views. This can generate predictions or estimates of the features of samples in the missing views.

[0064] Step 4: In order to avoid trivial solutions, dual consistency constraints of reconstruction loss and contrast loss are set to enhance the discrimination of different latent features through contrastive learning. For example, the private information and shared information of two views are compared to learn the view feature representation by shortening the distance of shared information and increasing the distance of private information.

[0065] The feature representations learned by the DVAE model are the same or almost the same for all views. In order to avoid trivial solutions, contrast loss is introduced.

[0066] Among the different feature variables of all views input into the DAVE model, the potential feature vectors of each view are extracted by combining the decoupled variational autoencoder with the adversarial network. The potential feature vectors of different views are compared and learned by maximizing the lower bound of the mutual information while minimizing the similarity between private information and shared information, which is expressed as:

[0067]

[0068] In the formula, I represents mutual information, which describes the correlation between two potential feature variables. The higher the value, the more correlated they are. H represents entropy, which describes the uncertainty of a random feature variable. The higher the value, the more uncertain the value of the random variable is. α is a weight parameter, and Z (1) and Z (2) They 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 entropy.

[0069] In order to perform contrastive learning, the present invention also needs to construct positive and negative samples of information. First, construct a positive sample pair, select a pair of samples from the data set, if the pair of samples have similar representations in multiple views or shared views, these samples are considered to belong to the same category. Conversely, construct a negative sample pair, select another pair of samples from the data set, and the samples with large differences in views belong to different categories. The positive sample pair is The negative sample pair is Given a view represented by the following features:

[0070]

[0071] The cosine similarity formula is given by:

[0072]

[0073] Through contrastive learning optimization, loss is calculated, and the model is optimized. The distance between positive sample pairs is as small as possible, while the distance between negative sample pairs is as large as possible. This can promote the learning of more discriminative and expressive features in the latent space or feature space. Specifically, the similarity between shared information and positive sample pairs is first calculated, and then the similarity between private information and negative sample pairs is calculated, and the similarity values ​​are spliced ​​together. Finally, the calculation is implemented by adopting InfoNCE loss to achieve the goal of contrastive learning: shorten the distance between positive sample pairs and increase the distance between negative sample pairs. This process learns the common feature representation between different views by maximizing the lower bound of mutual information, and combines the latent feature representation optimized by POE technology for unsupervised clustering, and finally completes the unsupervised clustering of noisy network information. The contrast loss calculation is given by the following formula.

[0074]

[0075] Among them, z i Is shared information x i The characteristics of It is a positive sample The feature representation of private information z j is a negative sample pair The characteristic representation of sim(z i ,z j ) represents sample z i and z j The similarity between them can be measured by cosine similarity. τ is the temperature coefficient that adjusts the smoothness of the distribution, and N is the number of views in the batch.

[0076] Through this step, the information between the two views can be made as similar as possible, but retain enough diversity to distinguish the potential features. Therefore, by reconstructing the loss function and introducing double consistency constraints to avoid trivial solutions, the final clustering effect can be improved.

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

[0078] In order to improve the potential feature alignment of the DAVE model, in the case of missing views and multiple views, the potential feature vectors in the input view are combined to separate private information from shared information. This step specifically uses the POE expert production model to induce inference of shared information in DVAE. By optimizing the shared information of a view, it can better capture and reconstruct the shared information between other different views, so that it can be combined with other views to maintain the consistency of shared information, align the view network information, and achieve reconstruction optimization of the model.

[0079] Through POE technology, the model can learn the common features between different views and reconstruct this shared information in the latent space. This helps the model to use the information of other views to infer the missing view features when a view is missing. Figure 3 As shown in Figure 1, POE technology helps improve the accuracy of the model's clustering of 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, there are two views X (1) ,X (2) , when in view X (2) In the case of missing, q(Z skered |X) Infer view X (1) and vice versa. The specific formula is given as follows:

[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 case of missing views and multiple 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 network information consistency. The formula of the POE expert production model is as follows:

[0083]

[0084] In the POE model, shared inference is defined, ∞ is separation, V (i) represents the i-th view, V represents the view set, Z s represents shared information, p(Z i ) represents the information posterior probability, Equal to view all private information, For all views sharing information, the posterior distribution of the i-th view is p(z i |V (i) ), from a shared network q(Z s | V) separate private information and shared information networks

[0085] Specifically, if there is network noisy information in the single view and multiple views in the DAVE model, for each view V, the mean μ of the latent variable Z is obtained by the model (v) and variance δ (v)2 The specific formula is given by:

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

[0087] Combine the means and variances of multiple views to calculate the mean and variance of the joint distribution. The specific calculation formula is given as follows:

[0088]

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

[0090]

[0091] The objective function of POE is expressed as the local consensus regularization of different view sets and local embedding consensus regularization By minimizing this objective function loss L, the DAVE model can learn feature representations that can maintain the local structure of each view and ensure consistent shared information between different views. This helps improve the performance of clustering, classification, and other downstream tasks in multi-view learning tasks.

[0092] After the above steps, the decoupled variational autodecoder of the present invention takes the reconstructed independent potential feature representation, i.e., private information or shared information, from the potential space optimized by POE technology and contrastive learning. On this basis, unsupervised multi-view clustering is performed, which can effectively reduce the noisy information in the view and improve the accuracy and consistency of clustering.

[0093] The process of model training of the present invention is as follows: First, the model was pre-trained for 200 rounds, 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 data set of the present invention, the fixed missing rate is 0.5. For methods that can only process complete view data, samples are randomly deleted while retaining at least one view, and then the mean of the same view is used to fill the missing data. During the training process, the total loss in the DVAE model is calculated and back-propagated and optimized, so that DAVE can not only effectively reconstruct the input view data, but also make the distribution of the latent variable features close to the Gaussian distribution. Among them, after optimizing the model by contrastive learning and POE technology, the DVAE decoder outputs the reconstructed view data for unsupervised clustering.

[0094] The three commonly used indicators, accuracy (ACC), normalized mutual information (NMI) and purity (PUR), are used to evaluate the clustering results. In this embodiment, the evaluated accuracy is 0.895, the normalized information is 0.732, and the purity is 0.843. The present invention has produced excellent results in clustering noisy network information, such as Figure 4 and Figure 5 shown.

[0095] It can be seen that the present invention realizes the consistency of structural features and shared information in multi-view network information, and solves the problem of multi-view semantic noisiness that occurs in view networks of private information and shared information.

Claims

1. An unsupervised multi-view clustering method for the problem of noisy clustering information, characterized in that: The steps include: Step 1: Perform feature processing on multi-view data to keep the feature vectors of the feature space consistent; Step 2: Establish a decoupled variational autoencoder and use a regularized constrained loss function to extract features from each view processed in step 1 to obtain private information and shared information, i.e., a potential feature vector, of each view; the private information is a feature unique to a certain view, and the shared information is a feature that exists in all views; Step 3: Define the reconstruction loss function of the decoupled variational autodecoder, 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 private information and shared information; and set the dual consistency constraints of reconstruction loss and contrast loss in combination with avoiding trivial solutions, and enhance the discrimination of different potential features through contrastive learning; Step 4: Use the decoupled variational self-decoder to reconstruct the optimized feature information for unsupervised multi-view clustering.

2. The unsupervised multi-view clustering method for the noisy clustering information problem according to claim 1, characterized in that: The multi-view data is different objects captured from multiple different viewing angles; 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 noisy clustering information problem according to claim 1, characterized in that: The hidden layer of the decoupled variational autoencoder adds a Relu activation function to convert the input view features into a latent feature vector, and performs latent space decoupling in the decoupled variational autoencoder to capture an independent generation factor of the data for each view, that is, the unique private information or shared information in the same view.

4. The unsupervised multi-view clustering method for the noisy clustering information problem according to claim 1, characterized in that: The decoupled variational autoencoder and the decoupled variational autodecoder both follow Gaussian distribution, and the two constitute a DAVE model. The decoupled variational autoencoder uses the prior distribution p(z) as input and outputs q(z|x), where z is the potential feature vector and x is the view. The regularization term constraint is implemented through KL divergence, which is expressed as: Among them, μ i and σ i are the mean and standard deviation of q(z|x) corresponding to the i-th potential feature vector, ensuring that the posterior distribution of private information 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, that is, to reduce the correlation between features and improve the decoupling ability of the model.

5. The unsupervised multi-view clustering method for the noisy clustering information problem according to claim 4, characterized in that: In step 3, the reconstruction loss function is expressed as follows: Where E v and D v Denotes encoding and decoding of the v-th view using the decoupled variational autoencoder and decoupled variational decoder, respectively, X (v) Represents the set of private and shared information of view features, where N is the number of views.

6. The unsupervised multi-view clustering method for the noisy clustering information problem according to claim 5, characterized in that: In step 3, among the feature variables of all views input into the DAVE model, the potential feature vectors of each view are extracted by combining the decoupled variational autoencoder with the adversarial network, and the potential feature vectors of different views are compared and learned by maximizing the lower bound of the mutual information while minimizing the similarity between private information and shared information, which is expressed as: In the formula, I represents mutual information, which describes the correlation between two potential feature variables. The higher the value, the more correlated they are. H represents entropy, which describes the uncertainty of a random feature variable. The higher the value, the more uncertain the value of the random variable is. α is a weight parameter, and Z (1) and Z (2) Represent the latent feature variables of the two views respectively; The goal of contrastive learning is to maximize the mutual information between two latent feature variables while minimizing their respective entropy.

7. The unsupervised multi-view clustering method for the noisy clustering information problem according to claim 6, characterized in that: In step 3, the POE expert production model is used to induce inference of the shared information in DVAE, and by optimizing the shared information of one view, it can better capture and reconstruct the shared information between other views; in the case of missing views and multiple views, the DAVE model combines the 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. The formula of the POE expert production model is as follows: In the POE model, shared inference is defined, ∞ is separation, V (i) represents the i-th view, V represents the view set, Z s represents shared information, p(Z i ) represents the information posterior probability, Equal to view all private information, For all views sharing information, the posterior distribution of the i-th view is p(z i |V (i) ).

8. The unsupervised multi-view clustering method for the noisy clustering information problem according to claim 7, characterized in that: The objective function loss of the POE technology is as follows: is the objective function of POE, expressed as the local consensus regularization of different view sets and local embedding consensus regularization By minimizing the objective function loss, the DAVE model can learn feature representations that can maintain the local structure of each view and ensure consistent shared information between different views.

9. The unsupervised multi-view clustering method for the noisy clustering information problem according to claim 8, characterized in that: In step 4, the decoupled variational autodecoder takes the reconstructed independent latent feature representation from the latent space optimized by the POE technology and contrastive learning, and performs unsupervised multi-view clustering.

Citation Information

Patent Citations

  • Neuron coverage guiding type testing method for automatic driving system

    CN117152550A

  • Self-supervised multi-view clustering method based on confrontation attention mechanism and decoupling representation

    CN117611855A

  • Medical image segmentation method based on weak supervision

    CN117830332A

  • Multi-modal adaptive fusion deep clustering model and method based on auto-encoder

    US20240095501A1

Cited By

  • Label guide representation learning method for incomplete multi-view multi-label classification

    CN122471174A

  • Label-guided representation learning method for incomplete multi-view multi-label classification

    CN122471174B