Contrastive incomplete data clustering method fusing multiple filling and attention mechanism

Abstract

The disclosure provides a contrastive incomplete data clustering method fusing multiple filling and attention mechanism, comprising: filling the original incomplete data set by adopting a chain equation multiple filling algorithm to obtain a filled data set with multi-view data; inputting the filled data set into a self-encoder to obtain the latent features of each view based on the encoder; fusing the latent features of each view by adopting an attention mechanism to obtain fused latent features; inputting the fused latent features into a decoder to obtain reconstructed features; calculating a reconstruction loss according to the latent features and the reconstructed features; performing contrastive learning on the positive sample pairs and the negative sample pairs in the fused latent features based on mapping to calculate a contrastive learning loss; performing clustering on the fused latent features based on a generative adversarial network to calculate a clustering loss; and training the self-encoder according to a loss function. The accuracy of incomplete data filling is improved, the shared feature information between different views is effectively mined, and the clustering accuracy is improved.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, and in particular to a comparative incomplete data clustering method that integrates multiple imputation and attention mechanisms. Background Technology

[0002] With the rapid rise of cloud computing and artificial intelligence, massive amounts of data have become crucial assets, and extracting valuable information from these vast datasets has become key to enhancing competitiveness across industries. Clustering, as an effective data mining technique, is widely used in feature selection, image segmentation, and content retrieval. Its basic idea is to classify data samples into different categories based on their correlation without labels. However, in practical applications, data gaps due to factors such as sensor failure, measurement bias, and instability are common. Therefore, research and discussion on incomplete data clustering algorithms have shown a significant upward trend in recent years.

[0003] The inventors discovered that existing clustering methods generally use single-impact techniques to handle incomplete data, failing to adequately account for the uncertainty of missing values ​​and resulting in poor imputation performance. Multiple imputation algorithms, however, precisely overcome the shortcomings of single-impact methods, cleverly transforming single-view data into multi-view data that provides different aspects or perspectives of the data. Furthermore, clustering methods for multi-view data generally suffer from two problems: First, deep multi-view clustering models only use simple methods such as matrix averaging to fuse the latent features of the views, failing to obtain optimal fused features. Second, the private information of the views may dominate feature fusion, leading to poor clustering results. Summary of the Invention

[0004] To address the aforementioned issues, this invention proposes a contrastive incomplete data clustering method that integrates multiple imputation and attention mechanisms. By fusing the multiple imputation algorithm with deep multi-view clustering technology, the accuracy of incomplete data imputation is improved. Furthermore, applying a heterogeneous graph attention network to the latent features of each view enables more refined integration of the unique information provided by each view, thereby significantly enhancing the overall model performance. In addition, to further reduce the impact of view-specific information on feature fusion and further improve clustering results, a multi-view contrastive learning algorithm is employed, which effectively extracts and discovers shared feature information among different views.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] In a first aspect, the present invention provides a method for clustering contrastive incomplete data that integrates multiple padding and attention mechanisms, comprising:

[0007] The original incomplete dataset is obtained and filled using a chain equation multiple fill algorithm to obtain a filled dataset with multiple views.

[0008] The filled dataset is input into the autoencoder to obtain the latent features of each view based on the encoder; the latent features of each view are fused using an attention mechanism to obtain fused latent features; the fused latent features are input into the decoder to obtain reconstructed features;

[0009] Calculate the reconstruction loss based on the latent features and reconstruction features;

[0010] The fusion latent features are mapped to obtain the mapped fusion latent features; comparative learning is performed based on the positive and negative sample pairs in the mapped fusion latent features, and the comparative learning loss is calculated.

[0011] Clustering of fused potential features is performed based on generative adversarial networks, and clustering loss is calculated.

[0012] The autoencoder is trained using reconstruction loss, contrastive learning loss, and clustering loss.

[0013] Preferably, the chain equation multiple imputation algorithm is used to imput the dataset to obtain the imputed dataset. The specific process is as follows:

[0014] The single-view variables in the original incomplete dataset contain missing values;

[0015] Obtain the set of variables without missing variables;

[0016] Extract single-view variables from the original incomplete dataset, calculate the statistical relationship between the single-view variables and the set of variables without missing data, imput the single-view variables based on the statistical relationship to obtain imputed multi-view variables, and add the imputed multi-view variables to the set of variables without missing data.

[0017] Other single-view variables are extracted sequentially from the original incomplete dataset, the statistical relationship between them and the updated set of variables without missing values ​​is calculated, and missing values ​​are imputed to obtain the filled dataset.

[0018] Preferably, the process of fusing the latent features of each view using an attention mechanism to obtain fused latent features is as follows:

[0019] Calculate the mean of the latent features for each sample across different views;

[0020] The weights of the latent features of different views for each sample relative to the overall latent features are calculated using an adaptive attention mechanism.

[0021] For each sample, the latent features of different views are summed based on the weights to obtain the fused latent features.

[0022] Preferably, the adaptive attention mechanism specifically involves randomly transforming the weights multiple times and averaging them to obtain the final weights.

[0023] Preferably, the step of performing comparative learning on positive and negative sample pairs based on the mapping-fused latent features and calculating the comparative learning loss is as follows:

[0024] The similarity between positive and negative sample pairs is calculated using cosine distance:

[0025]

[0026] Where <·, ·> represent dot product operations; any two views H m With H n The feature contrast loss between them is expressed as:

[0027]

[0028] Where τ F Given the temperature parameter, the cumulative multi-view feature contrast loss across all views is:

[0029]

[0030] Preferably, the clustering of fusion potential features based on generative adversarial networks specifically involves applying a prior distribution to the membership distribution Q.

[0031] Preferably, the step of training the autoencoder based on reconstruction loss, contrastive learning loss, and clustering loss specifically involves:

[0032]

[0033] Among them, L Re L represents the reconstruction loss. Ad-G L represents the clustering loss. Co The contrast loss is represented by α and β, which are regularization parameters that balance the clustering loss and the contrast loss.

[0034] Secondly, the present invention provides a contrastive incomplete data clustering system that integrates multiple padding and attention mechanisms, comprising:

[0035] The multi-fill module is used to obtain the original incomplete dataset and fill it with the chain equation multi-fill algorithm to obtain a filled dataset with multiple view data.

[0036] The feature fusion module is used to input the padding dataset into the autoencoder, obtain the latent features of each view based on the encoder, fuse the latent features of each view using an attention mechanism to obtain fused latent features, and input the fused latent features into the decoder to obtain reconstructed features.

[0037] The loss calculation module is used to calculate the reconstruction loss based on the latent features and the reconstruction features; map the fused latent features to obtain the mapped fused latent features; perform comparative learning based on the positive and negative sample pairs in the mapped fused latent features and calculate the comparative learning loss; cluster the fused latent features based on the generative adversarial network and calculate the clustering loss; and train the autoencoder based on the reconstruction loss, comparative learning loss and clustering loss.

[0038] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the comparative incomplete data clustering method that integrates multiple padding and attention mechanisms as described in the first aspect.

[0039] Fourthly, the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in the comparative incomplete data clustering method that integrates multiple padding and attention mechanisms as described in the first aspect.

[0040] Compared with the prior art, the beneficial effects of this disclosure are as follows:

[0041] This invention introduces the Chain Equation Multiple Imputation (MICE) algorithm, which improves upon the traditional single imputation method's failure to adequately consider the uncertainty of missing data. It cleverly combines the results of multiple imputations with a deep multi-view clustering model. This not only overcomes the challenges of processing high-dimensional data but also significantly enhances clustering performance. Furthermore, to avoid the limitations of using simple matrix averaging to fuse features, we extend the heterogeneous graph attention network and apply it to integrate the latent features of multiple views, thereby enhancing the overall performance of the model. In addition, a contrastive learning algorithm is used to operate on the latent features of each view, aiming to extract common feature information shared across views, thus mitigating the impact of excessive private information specific to a single view on subsequent task results.

[0042] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0043] The accompanying drawings, which form part of this disclosure, are used to provide a further understanding of this disclosure. The illustrative embodiments of this disclosure and their descriptions are used to explain this disclosure and do not constitute a limitation thereof.

[0044] Figure 1 The main flowchart of a comparative incomplete data clustering method that integrates multiple imputation and attention mechanisms provided in this disclosure;

[0045] Figure 2 This disclosure provides a framework diagram for a comparative incomplete data clustering algorithm that integrates multiple padding and attention mechanisms. Detailed Implementation

[0046] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0047] Example 1

[0048] like Figure 1 As shown, this embodiment discloses a method for clustering contrastive incomplete data that integrates multiple padding and attention mechanisms, including the following steps:

[0049] S1: Obtain the original incomplete dataset, and use the chain equation multiple imputation algorithm to imput it, so as to obtain an imputed dataset with multiple view data;

[0050] S2: Input the filled dataset into the autoencoder to obtain the latent features of each view based on the encoder; use an attention mechanism to fuse the latent features of each view to obtain fused latent features; input the fused latent features into the decoder to obtain reconstructed features;

[0051] S3: Calculate the reconstruction loss based on the latent features and reconstruction features;

[0052] S4: Map the fusion latent features to obtain the mapped fusion latent features; perform comparative learning based on the positive and negative sample pairs in the mapped fusion latent features, and calculate the comparative learning loss.

[0053] S5: Cluster the fused potential features based on generative adversarial networks and calculate the clustering loss;

[0054] S6: Train the autoencoder based on reconstruction loss, contrastive learning loss, and clustering loss.

[0055] Clustering algorithm framework such as Figure 2 As shown. To overcome the limitations of a single filling algorithm when processing incomplete data, this invention, for the first time, integrates a multiple filling algorithm with deep multi-view clustering technology to solve the problem of incomplete data clustering.

[0056] (I) Multiple Filling

[0057] Multiple imputation is one of the most commonly used methods for handling missing values. The basic idea is to provide m imputation values ​​for each missing value, forming m complete datasets.

[0058] This embodiment employs the Chain Equation Multiple Imputation (MICE) algorithm. The essential difference between this method and other multiple imputation algorithms is that it does not consider the joint distribution of the imputed variable and covariates during imputation. In other words, it does not prioritize the joint distribution during operation, but instead uses the conditional distribution of individual variables for imputation one by one. The chain equation is a descriptive term; the actual algorithm consists of a series of regression models that characterize the conditional distributions of the imputed variable and covariates. These conditional distributions can be of various types.

[0059] Specifically, let X be a set of variables without missing values, Y1, Y2, ..., Y... h There are h variables with missing values.

[0060] For Y1, Y2, ..., Y h The interpolation process requires a total of n rounds of iteration. The algorithm flow is as follows:

[0061] 1. In the first iteration, first perform a regression of Y1 on X, and then impute the missing values ​​of Y1 based on this regression. Imputation should follow the principle of multiple imputation, that is, simultaneously consider the random variability of model parameters and the random variability of model residuals;

[0062] 2. Then perform regression of Y2 on Y1 (including imputed values) and X, and impute the missing values ​​of Y2 based on this regression;

[0063] 3. Continue in this manner until the final Y test is performed. h In Y1,..,Y h-1 Regression on X, and based on this regression, on Y h The missing values ​​are imputed.

[0064] The imputation rounds from the 2nd to the nth round will follow the process of the first round, except that each regression now includes all variables except the original variable, and each round should use the latest imputed value. After all n rounds are completed, the imputation result of the nth round is taken as the final result, forming a complete dataset. To obtain m datasets, the above n rounds need to be performed independently m times.

[0065] (II) Adaptive Attention Mechanism

[0066] Deep multi-view clustering model. This embodiment employs a novel deep multi-view learning framework that intuitively integrates latent features into an autoencoder (AE) framework.

[0067] Specifically, after obtaining the shared latent representation Z = {z1, z2, ..., z} N This data is then individually input into multiple decoders for decoding. This network structure not only preserves the local characteristics specific to each view but also captures the global characteristics of all views.

[0068] Given incomplete data X = {x1, x2, ..., x...} N There are N samples. After M imputations, M complete datasets are obtained. This data is fed into a deep multi-view clustering model, which first uses multiple autoencoders to embed the multi-view data into different latent spaces. Then, we fuse the obtained latent features to obtain the shared latent features Z = {z1, z2, ..., z}. N},

[0069] Specifically, a heterogeneous graph attention mechanism is introduced to achieve adaptive fusion of latent features, enabling the model to better capture the relationships and information between features, thereby effectively improving the overall performance of the model.

[0070] Specifically, we use an attention mechanism to process the latent features of each view. We calculate the correlation weight of each view's feature with the overall mean of latent features, and then use this weight to multiply each feature with its own feature for weighted integration. Simultaneously, we employ a multi-head attention mechanism to perform deep learning and comprehensive evaluation on these fused features from multiple perspectives in parallel, to arrive at the final optimized result.

[0071] The module's calculation process is as follows: First, the mean of the latent representation of different views for each sample is calculated:

[0072]

[0073] Secondly, the attention mechanism is used to calculate the weights of the latent representations of different views for each sample relative to the overall latent features:

[0074]

[0075] Specifically, the attention mechanism is implemented using a single-feedforward neural network, where the weight vector 'a' serves as its core parameter, and the feature activation is enhanced through the non-linear activation function LeakyReLU. Mathematically, "||" represents a concatenation operation. This represents the latent features of the l-th sample in the m-th view. This corresponds to the specific weight coefficients of each sample in different views. We use normalized attention coefficients to calculate the weighted linear combination of the latent features from multiple views. For each sample, its final fused representation can be expressed as:

[0076]

[0077] Furthermore, to improve the model's stability and learning efficiency, we employ a multi-head attention mechanism. Specifically, for the calculation process in the above formula, we perform U independent attention transformations, each using a different weight vector 'a'. Finally, we average the results of these U transformations to obtain the final sample feature representation, thereby achieving a more comprehensive and stable adaptive fusion of latent features from different views.

[0078]

[0079] Where u∈[1,...,U]. The loss function is expressed as follows, which consists of two parts: reconstruction loss and clustering loss.

[0080]

[0081] Among them, L Re L represents the reconstruction loss. Ad-G L represents the clustering loss. Co The contrast loss is represented by α and β, which are regularization parameters that balance the clustering loss and the contrast loss.

[0082] This model uses multiple imputation and deep multi-view clustering to solve the problem of incomplete data clustering. It unifies the clustering process and the data imputation process into a single framework, which is an end-to-end process. This means that the results of data imputation directly and efficiently serve the subsequent clustering process, ensuring that the imputed data can more accurately and appropriately assist clustering analysis, thereby significantly improving the overall clustering effect and accuracy.

[0083] This approach combines multiple padding and deep multi-view clustering models to address the problem of incomplete data clustering. It also incorporates a heterogeneous graph attention network to fuse latent features, enabling the model to better capture the relationships and information between features, thereby effectively improving the overall performance of the model.

[0084] (III) Multi-view contrastive learning algorithm

[0085] To avoid conflicts between reconstruction loss and consistency objectives when optimizing latent features, an innovative strategy is adopted. Specifically, based on the acquired original latent features, a multilayer perceptron structure is superimposed to learn a new set of features in another space. The aim is to obtain general and consistent semantic information from all views. Primarily through comparative learning of the same samples across different views, we ensure that the model achieves cross-view consistency on these newly learned features, effectively avoiding the conflict between reconstruction loss and consistency goals, and promoting improved model performance.

[0086] The specific calculation process is as follows: First, for each feature... There are (MN-1) feature pairs in There are (M-1) positive feature pairs and M(N-1) negative feature pairs. We use cosine distance to measure the similarity between two features:

[0087]

[0088] Where <·, ·> represent the dot product operation. Any two views H m With H n The feature contrast loss between them is expressed as:

[0089]

[0090] Where τ F Given the temperature parameter, the cumulative multi-view feature contrast loss across all views is:

[0091]

[0092] The final loss function is expressed as the following formula, which consists of three parts: reconstruction loss, clustering loss, and contrastive loss.

[0093]

[0094] Among them, L Re L represents the reconstruction loss. Ad-G L represents the clustering loss. Co The contrast loss is represented by α and β, which are regularization parameters that balance the clustering loss and the contrast loss.

[0095] (iv) Clustering Algorithm

[0096] KL divergence only sharpens the soft distribution of data, but it lacks sufficient discriminative power to distinguish the data distributions of different clusters. This can lead to degeneracy problems, where all data samples are assigned to a few clusters beyond expectations. Existing algorithms, drawing on the ideas of generative adversarial learning and the model architecture of WGAN-GP, design a discriminative network to replace the traditional KL clustering terms for class division. A prior distribution is applied to the membership distribution Q to make the class information closer to the one-hot form, thus making the classes more clearly distinguishable. The generator loss and adversarial loss of the discriminative network are defined by the following formulas:

[0097] L Ad-G =-E q～Q [D(q)]

[0098]

[0099] Where r=òp+(1-ò)q follows ò~U[0,1], it is a representation of uniform sampling along the straight line between the classification distribution P and the soft assignment q. It is a single-center gradient penalty that restricts the gradient of the adversarial network to around 1, E q～Q [D(q)]-E p～P [D(p)] represents the Wasserstein distance between the membership distribution Q and the prior distribution P. The purpose of these two terms is to minimize the gap between the membership distribution and the prior distribution. E q～Q [D(q)] is the score given by the discriminator for membership degree; a higher score indicates a clearer membership degree distribution. We aim to obtain a higher score to help the proposed model obtain a clearer cluster distribution, thereby improving the model's discriminative ability.

[0100] This embodiment employs an adversarial network approach for data clustering. Specifically, the current KL divergence only sharpens the soft distribution of the data but lacks sufficient discriminative power to distinguish the data distributions of different clusters, which may lead to degeneracy problems. This embodiment, drawing on the ideas of generative adversarial learning and the model architecture of WGAN-GP, designs a discriminative network to replace the previous KL clustering items for category division. A prior distribution is applied to the membership distribution Q to make the category information closer to the one-hot form, that is, a prior distribution P ~ Cat(K, p = 1 / K) is applied to the membership distribution Q to make the category information closer to the one-hot form, thereby making the categories clearer and more distinguishable.

[0101] This embodiment improves upon traditional single-impact methods by introducing the Chain Equation Multiple Imputation (MICE) algorithm, which fails to adequately account for the uncertainty of missing data. It cleverly combines the results of multiple imputations with a deep multi-view clustering model. This not only overcomes the challenges of processing high-dimensional data but also significantly enhances clustering performance. Furthermore, to avoid the limitations of simple matrix averaging methods for feature fusion, we extend the heterogeneous graph attention network and apply it to integrate latent features from multiple views, thereby enhancing the overall model performance. In addition, a contrastive learning algorithm is used to operate on the latent features of each view, aiming to extract shared common feature information across views, thus mitigating the impact of excessive private information specific to a single view on subsequent task results.

[0102] Example 2

[0103] This embodiment provides a contrastive incomplete data clustering system that integrates multiple padding and attention mechanisms, including:

[0104] The multi-fill module is used to obtain the original incomplete dataset and fill it with the chain equation multi-fill algorithm to obtain a filled dataset with multiple view data.

[0105] The feature fusion module is used to input the padding dataset into the autoencoder, obtain the latent features of each view based on the encoder, fuse the latent features of each view using an attention mechanism to obtain fused latent features, and input the fused latent features into the decoder to obtain reconstructed features.

[0106] The loss calculation module is used to calculate the reconstruction loss based on the latent features and the reconstruction features; map the fused latent features to obtain the mapped fused latent features; perform comparative learning based on the positive and negative sample pairs in the mapped fused latent features and calculate the comparative learning loss; cluster the fused latent features based on the generative adversarial network and calculate the clustering loss; and train the autoencoder based on the reconstruction loss, comparative learning loss and clustering loss.

[0107] Example 3

[0108] This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps in the contrastive incomplete data clustering method that integrates multiple padding and attention mechanisms as described in Embodiment 1 above.

[0109] Example 4

[0110] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the comparative incomplete data clustering method that integrates multiple padding and attention mechanisms as described in Embodiment 1 above.

[0111] The steps or modules involved in Embodiments 2 to 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0112] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A clustering method for contrastive incomplete data that integrates multiple imputation and attention mechanisms, characterized in that, The method is used for image segmentation and includes: Obtain the original incomplete dataset X= have A set of samples was filled using a chain equation multiple fill algorithm to obtain a filled dataset with multiple views. The specific process is as follows: the single-view variables in the original incomplete dataset contain missing values; obtain the set of variables without missing values; extract the single-view variables from the original incomplete dataset, calculate the statistical relationship between the single-view variables and the set of variables without missing values, imput the single-view variables based on the statistical relationship to obtain imputed multi-view variables, and add the imputed multi-view variables to the set of variables without missing values; extract other single-view variables from the original incomplete dataset in sequence, calculate the statistical relationship between them and the updated set of variables without missing values, imput missing values, and obtain the filled dataset; The filled dataset is input into the autoencoder to obtain the latent features of each view; the latent features of each view are fused using an attention mechanism to obtain fused latent features; the fused latent features are input into the decoder to obtain reconstructed features; Calculate the reconstruction loss based on the latent features and reconstruction features; The fusion latent features are mapped to obtain the mapped fusion latent features; comparative learning is performed based on the positive and negative sample pairs in the mapped fusion latent features, and the comparative learning loss is calculated. Clustering of fused potential features is performed based on generative adversarial networks, and clustering loss is calculated. The autoencoder is trained using reconstruction loss, contrastive learning loss, and clustering loss.

2. The contrastive incomplete data clustering method integrating multiple imputation and attention mechanisms as described in claim 1, characterized in that, The process of fusing the latent features of each view using an attention mechanism to obtain fused latent features is as follows: Calculate the mean of the latent features for each sample across different views; The weights of the latent features of different views for each sample relative to the overall latent features are calculated using an adaptive attention mechanism. For each sample, the latent features of different views are summed based on the weights to obtain the fused latent features.

3. The contrastive incomplete data clustering method integrating multiple imputation and attention mechanisms as described in claim 2, characterized in that, The adaptive attention mechanism specifically involves randomly transforming the weights multiple times and averaging them to obtain the final weights.

4. The contrastive incomplete data clustering method integrating multiple imputation and attention mechanisms as described in claim 1, characterized in that, The process involves comparative learning of positive and negative sample pairs based on the latent features fused by mapping, and calculating the comparative learning loss. Specifically, cosine distance is used to calculate the similarity between positive and negative sample pairs. Among them, < > indicates dot product operation; any two views and The feature contrast loss between them is expressed as: in Given the temperature parameter, the cumulative multi-view feature contrast loss across all views is: 。 5. The contrastive incomplete data clustering method integrating multiple imputation and attention mechanisms as described in claim 1, characterized in that, The clustering of fused latent features based on generative adversarial networks specifically involves: clustering the membership distribution... Q Apply a prior distribution.

6. The contrastive incomplete data clustering method integrating multiple imputation and attention mechanisms as described in claim 1, characterized in that, The autoencoder is trained using reconstruction loss, contrastive learning loss, and clustering loss, and the final loss function is as follows: ； in, Indicates the reconstruction loss; Represents the clustering loss. Indicates comparative loss, and Regularization parameters are used to balance clustering loss and contrastive loss.

7. A contrastive incomplete data clustering system integrating multiple imputation and attention mechanisms, characterized in that, The system is used for image segmentation and includes: The multiple padding module is used to obtain the original incomplete dataset. X= have A set of samples was filled using a chain equation multiple fill algorithm to obtain a filled dataset with multiple views. The specific process is as follows: the single-view variables in the original incomplete dataset contain missing values; obtain the set of variables without missing values; extract the single-view variables from the original incomplete dataset, calculate the statistical relationship between the single-view variables and the set of variables without missing values, imput the single-view variables based on the statistical relationship to obtain imputed multi-view variables, and add the imputed multi-view variables to the set of variables without missing values; extract other single-view variables from the original incomplete dataset in sequence, calculate the statistical relationship between them and the updated set of variables without missing values, imput missing values, and obtain the filled dataset; The feature fusion module is used to input the padding dataset into the autoencoder, obtain the latent features of each view based on the autoencoder, fuse the latent features of each view using an attention mechanism to obtain fused latent features, and input the fused latent features into the decoder to obtain reconstructed features. The loss calculation module is used to calculate the reconstruction loss based on the latent features and the reconstruction features; map the fused latent features to obtain the mapped fused latent features; perform comparative learning based on the positive and negative sample pairs in the mapped fused latent features and calculate the comparative learning loss; cluster the fused latent features based on the generative adversarial network and calculate the clustering loss; and train the autoencoder based on the reconstruction loss, comparative learning loss and clustering loss.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the contrastive incomplete data clustering method that integrates multiple padding and attention mechanisms as described in any one of claims 1-6.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the contrastive incomplete data clustering method that integrates multiple padding and attention mechanisms as described in any one of claims 1-6.

Images