Robust incomplete multi-view clustering method with pseudo-label guided universal negative set learning
By employing pseudo-label guidance and general negative class learning, the false negative problem caused by incompleteness in multi-view clustering is solved, enabling efficient ad delivery under missing data conditions and improving the robustness and accuracy of clustering.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- BEIJING NORMAL UNIVERSITY
- Filing Date
- 2025-09-19
- Publication Date
- 2026-05-26
AI Technical Summary
Existing multi-view clustering methods are prone to false negatives when dealing with incomplete data, which leads to a decrease in the accuracy and stability of ad delivery.
A general negative class learning method guided by pseudo-labels is adopted. Pseudo-labels are generated through preliminary representation learning to fill in missing view samples, and the negative class sample mechanism is used to enhance the cluster boundary and improve the cluster robustness during the fine-tuning stage.
It effectively alleviates the false negative problem in multi-view clustering, improves the accuracy and stability of ad delivery, and can maintain good clustering performance even in the presence of missing data.
Smart Images
Figure CN121190801B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of advertising push clustering analysis technology, and more specifically, to a robust incomplete multi-view clustering method based on pseudo-label-guided general negative class learning. Background Technology
[0002] Multi-view data is widely used in real-world applications, such as text and image information, video multimodal information, and remote sensing multichannel imagery. Multi-view clustering (MVC) aims to use information from multiple views to more accurately cluster samples and is an important research direction in multimodal learning.
[0003] Multi-view clustering (MVC) has been widely used in personalized recommendation scenarios such as ad delivery in recent years due to its ability to integrate heterogeneous data from multiple sources and mine potential user interest categories. Combining contrastive learning to enhance the consistency of the same user across different views helps to build a more refined user segmentation structure and improve ad matching effectiveness. In recent years, contrastive learning has been introduced into MVC in the field of ad delivery to enhance clustering performance by maximizing the consistency of the same instance across different views, thereby enabling the delivery of appropriate ad types to target users. However, existing methods generally adopt a one-to-many contrast mechanism, which is prone to mistaking similar samples as negative samples, causing false negatives and impairing intra-class compactness and discrimination boundaries. In addition, multi-view data in reality often has incompleteness, i.e., some views are missing, which further exacerbates the above problems and poses a greater challenge to existing contrastive learning frameworks. Traditional multi-view clustering methods generally rely on the completeness of all view information, making it difficult to directly adapt to such incomplete data situations in reality, which may lead to incorrect classification of target users and affect the accuracy of ad delivery.
[0004] Therefore, it is necessary to optimize multi-view clustering to solve the false negative problem of incomplete multi-view clustering, thereby improving the accuracy of ad push to target users. Summary of the Invention
[0005] The purpose of this invention is to provide a robust incomplete multi-view clustering method for general negative class learning guided by pseudo-labels, which solves the false negative problem in incomplete multi-view clustering.
[0006] This invention is achieved through the following technical solution:
[0007] A robust, incomplete multi-view clustering method for general negative class learning guided by pseudo-labels includes the following steps:
[0008] Obtain multi-view data of the target user;
[0009] During the warm-up phase, preliminary representation learning is performed based on multi-view data to generate pseudo-labels;
[0010] Use pseudo-labels to complete missing view samples;
[0011] During the fine-tuning phase, the general negative sample mechanism enhances cluster boundary modeling and expands the discriminative boundary between categories, thereby improving the robustness of clustering for multi-view data and enabling matching of ad push types to target users.
[0012] Preferably, the method for preliminary representation learning based on multi-view data during the warm-up phase is as follows:
[0013] The original input sample is reconstructed based on the multi-view data using an encoder and decoder, and the reconstruction loss is obtained.
[0014] Perform in-view contrast learning and out-of-view contrast learning separately to obtain in-view contrast loss and out-of-view contrast loss;
[0015] The total loss during the warm-up phase is obtained by weighted summation of reconstruction loss, in-view contrast loss, and out-of-view contrast loss.
[0016] By minimizing the total loss during the warm-up phase, the parameters of the encoder and decoder are jointly optimized to improve the initial representation quality of multi-view data.
[0017] Preferably, the method for obtaining the reconstruction loss is as follows:
[0018]
[0019] in, For reconstruction loss, v represents the number of the view data, V is the total number of view data, and f v (·) represents the encoder for the v-th view data, g v (·) is the decoder for the v-th view data. For the i-th sample of the v-th view data, N v This represents the total number of samples in the v-th view data.
[0020] Preferably, the method for performing intra-view comparison learning and inter-view comparison learning respectively is as follows:
[0021] Using Dropout to represent the latent features of the i-th sample in the v-th view Apply random perturbations to generate enhanced representations. Used to construct positive contrast samples within a view, v = 1, 2, ..., V, i = 1, 2, ..., N, where N is the total number of samples within a view's data;
[0022] Perform in-view comparison:
[0023]
[0024] in, S is the in-view contrast loss, S(.,.) is the function for calculating cosine similarity, τ is the coefficient controlling the steepness, and exp is the exponential function.
[0025] Compare the first view data and the second view data between views:
[0026]
[0027] in, Loss due to comparison between first view data and second view data;
[0028] All data in the p-th view and the q-th view are processed in the same way as above. The possible combinations of p and q values include p = h and q = h + 1, h = 1, 2, ..., V-1, and p = V and q = 1.
[0029] The sum of the inter-view contrast losses for all p-th view data and q-th view data is used as the inter-view contrast loss.
[0030] Preferably, the method for obtaining the pseudo-tags is as follows:
[0031] Multiple view data are merged and concatenated, and k-means clustering is used to obtain pseudo-labels;
[0032] The method for completing missing view samples using pseudo-labels includes:
[0033] For samples with missing views, the pseudo-labels generated in other views are used to return to the missing view. Based on the pseudo-labels, a set of samples with the same pseudo-labels and complete features in the missing view is retrieved. The complete feature representation of the missing view is constructed based on the feature mean of the set.
[0034] Define a set of category labels c represents the category label, c = 1, 2, ..., C;
[0035] Let the first The i-th sample is missing in each view of data, based on the corresponding mean feature. For the missing data Fill in:
[0036]
[0037] in, This represents the set of all pseudo-labeled samples within the view data that belong to the missing view samples. These are pseudo-labels that are filled in for the missing sample using other views. Representing the Latent features of the j-th sample in each view;
[0038] For all partially missing view samples, the confidence level is estimated by calculating the similarity between their completion features and the class center to which the pseudo-label belongs. Only completed samples with a confidence level higher than a set threshold are retained for subsequent training.
[0039] Preferably, the boundary modeling and robust clustering based on the negative class sample mechanism in the fine-tuning stage is as follows:
[0040] Negative class samples are generated based on the multi-view data;
[0041] Construct a general negative class contrast loss between views and a general negative class contrast loss within views based on negative class samples;
[0042] The total loss during the fine-tuning stage is obtained by weighting and summing the reconstruction loss, the general negative class contrast loss between views, and the general negative class contrast loss within views.
[0043] By minimizing the total loss during the fine-tuning phase, the model parameters are optimized to improve the clustering performance and robustness of multi-view data.
[0044] Preferably, in a scenario involving comparison between views, the method for generating negative class samples based on the multi-view data is as follows:
[0045] Using the first view as the main view, identify all data in the first view that have pseudo-labels. Potential characteristics The negative class of the view, where i is the sample index. This represents the latent features of the i-th sample in the first view;
[0046] Define a set of category labels c represents the category label, c = 1, 2, ..., C;
[0047] Using the second view as an auxiliary view, the second view data defines a sample index set with pseudo-labels.
[0048]
[0049] Where N is the total number of samples within a view of data. For the j-th latent feature of the second view data pseudo-tags;
[0050] Based on sample index set Obtain the negative class centroid of the second view data with category label c.
[0051]
[0052] For all category labels, if the following conditions are met And c is not equal to Category labels will With the corresponding negative class centroid The corresponding initial negative class sample u is generated by mixing. i，c :
[0053]
[0054] Where λ is the mixing coefficient, λ∈(0,1);
[0055] Based on all initial negative class samples u i，c The average value is obtained. negative class sample U i , i = 1, 2, ..., N:
[0056]
[0057] Collect all negative class samples U i This yields the set of negative class samples between the first view data and the second view data.
[0058] The same method is applied to all data in the p-th and q-th views. The possible combinations of p and q include p = h and q = h+1, where h = 1, 2, ..., V-1, and V is the total number of views, as well as p = V and q = 1. This yields the set of negative class samples between all data in the p-th and q-th views.
[0059] Preferably, in a scenario involving in-view comparison, the method for generating negative class samples based on the multi-view data is as follows:
[0060]
[0061] in, Let v be the negative class sample within the v-th view data, where v = 1, 2, ..., V. Augmentation data for the latent features of the j-th sample within the v-th view data.
[0062] Preferably, the method for constructing the general negative class contrast loss between views is as follows:
[0063] The view boundary term is defined based on the set of negative class samples from the comparison loss between the first view data and the second view data.
[0064]
[0065] Where τ is a coefficient controlling the steepness, exp is an exponential function, and S(.,.) is a function for calculating cosine similarity;
[0066] Obtain the negative class contrast loss between the first view data and the second view data.
[0067] in, and They are respectively The negative sample set and the positive sample set;
[0068] To achieve cross-view representation alignment, for each pair of adjacent view pairs (p,q), the corresponding cross-view contrast loss is calculated, where the combinations of p and q include p = h and q = h+1, h = 1, 2, ..., V-1, and p = V and q = 1;
[0069] The average loss among all view pairs is used as the general negative class contrast loss among views.
[0070] Preferably, the method for obtaining the general negative class contrast loss within the view is as follows:
[0071] Based on the set of general and sample data, define the view-in-view boundary item of the v-th view data.
[0072]
[0073] Obtain the general negative class contrast loss within the view.
[0074]
[0075] The technical solution of the present invention has at least the following advantages and beneficial effects:
[0076] In the clustering process, this invention considers the inherent incompleteness of multi-view user data in advertising push scenarios and the tendency of traditional contrastive learning to misclassify categories. Therefore, it proposes a robust clustering strategy based on pseudo-label guidance and a general negative class generation mechanism, which effectively alleviates the problem of push mismatch and improves the accuracy and stability of advertising push.
[0077] This invention introduces cross-class generalization of negative class samples based on Universum Learning, and indirectly improves inter-class discriminability through boundary modeling, thereby effectively reducing the occurrence of false negatives;
[0078] This invention supplements missing view data, maintaining good performance even with missing data. It achieves low-cost completion of missing views through pseudo-label estimation and feature imputation strategies, thereby enabling compatibility with incomplete data scenarios. Attached Figure Description
[0079] Figure 1 This is a flowchart illustrating the robust incomplete multi-view clustering method for general negative class learning provided in Embodiment 1 of the present invention. Detailed Implementation
[0080] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.
[0081] Example 1
[0082] This embodiment provides a robust incomplete multi-view clustering method for general negative class learning guided by pseudo-labels. See [link to documentation]. Figure 1 This includes the following steps:
[0083] Step S1: Obtain multi-view data for the target user.
[0084] Multi-view data can be obtained from various life scenarios and behavioral records of users, such as browsing history, transaction history, and social interaction data, to form a multimodal feature set of user profile.
[0085] As an implementation example, the target user's multi-view data may include:
[0086] See Figure 1 User behavior view, input is X 1 Each sample X i 1 The input vector represents the behavioral statistical features of user i, such as recent click sequence, browsing history, dwell time, and click frequency. After embedding or normalization, it forms the original input vector.
[0087] View 2: Demographics view, input is X 2 Each sample X i 2 The static structured attributes of user i, such as age, gender, region, and device type, are represented by one-hot encoding or low-dimensional embedding and then concatenated to form the original feature representation.
[0088] View 3: Transformation Behavior View, Input is X 3 Each sample X i 3 This represents the historical conversion behavior characteristics of user i, including numerical characteristics such as the frequency, amount statistics, and conversion cycle of conversion events such as adding to cart, placing orders, and making payments.
[0089] Content Interest View, input is X 4 Each sample X 4 i The semantic representation of the ad content that user i recently clicked or interacted with, including the Transformer embedding representation of text content (such as title and description) and the visual features of image / video content (such as CNN extraction or frame-level average pooling vectors), are concatenated or uniformly mapped to form the original input of the view.
[0090] Step S2: In the warm-up phase, perform preliminary representation learning based on multi-view data to generate pseudo-labels.
[0091] In this embodiment, the method for preliminary representation learning based on multi-view data during the warm-up phase is as follows:
[0092] The original input sample is reconstructed based on the multi-view data using an encoder and decoder, and the reconstruction loss is obtained.
[0093] Perform in-view contrast learning and out-of-view contrast learning separately to obtain in-view contrast loss and out-of-view contrast loss;
[0094] The total loss during the warm-up phase is obtained by weighted summation of reconstruction loss, in-view contrast loss, and out-of-view contrast loss.
[0095] By minimizing the total loss during the warm-up phase, the parameters of the encoder and decoder are jointly optimized to improve the initial representation quality of multi-view data.
[0096] Based on this, the method for obtaining the reconstruction loss is as follows:
[0097]
[0098] in, For reconstruction loss, v represents the number of the view data, V is the total number of view data, and f v (·) represents the encoder for the v-th view data, g v (·) is the decoder for the v-th view data. For the i-th sample of the v-th view data, N v This represents the total number of samples in the v-th view data.
[0099] Furthermore, the methods for performing intra-view contrast learning and inter-view contrast learning are as follows:
[0100] Using Dropout to represent the latent features of the i-th sample in the v-th view Apply random perturbations to generate enhanced representations. Used to construct positive contrast samples within a view, where v = 1, 2, ..., V, i = 1, 2, ..., N, and N is the total number of samples within a view's data. Specifically, a latent feature is typically selected from the sample set as an anchor point, and its positive and negative sample sets are constructed to bring positive samples closer and push negative samples further away. Generally, this can be done by using the original feature... Encoder processing is performed to obtain potential features.
[0101] Perform in-view comparison:
[0102]
[0103] in, S is the in-view contrast loss, S(.,.) is the function for calculating cosine similarity, τ is the coefficient controlling the steepness, and exp is the exponential function.
[0104] Compare the first view data and the second view data between views:
[0105]
[0106] in, Loss due to comparison between first view data and second view data;
[0107] All data in the p-th view and the q-th view are processed in the same way as above. The possible combinations of p and q values include p = h and q = h + 1, h = 1, 2, ..., V-1, and p = V and q = 1.
[0108] The sum of the inter-view contrast losses for all p-th view data and q-th view data is used as the inter-view contrast loss.
[0109] The final weighted sum yields the total loss during the warm-up phase. It can be:
[0110]
[0111] Where λ rec ω1 and ω2 are both weights.
[0112] In this step, multi-view data is reconstructed using an encoder-decoder structure, and reconstruction loss is calculated for optimization to maintain the integrity of feature information. Intra-view contrastive learning and inter-view contrastive learning are acquired separately, resulting in intra-view contrastive loss and inter-view contrastive loss. These losses help extract more accurate similar features and differential features across different views. The weighted sum of these three losses yields the total loss for the warm-up stage, which is used to optimize the initial feature representation.
[0113] Step S3: Use pseudo-labels to complete missing view samples.
[0114] As a preferred embodiment, the method for obtaining the pseudo-tag is as follows:
[0115] Multiple view data are merged and concatenated, and k-means clustering is used to obtain pseudo-labels;
[0116] The method for completing missing view samples using pseudo-labels includes:
[0117] For samples with missing views, the pseudo-labels generated in other views are used to return to the missing view. Based on the pseudo-labels, a set of samples with the same pseudo-labels and complete features in the missing view is retrieved. The complete feature representation of the missing view is constructed based on the feature mean of the set.
[0118] Define a set of category labels c represents the category label, c = 1, 2, ..., C;
[0119] Let the first The i-th sample is missing in each view of data, based on the corresponding mean feature. For the missing data Fill in:
[0120]
[0121] in, This represents the set of all pseudo-labeled samples within the view data that belong to the missing view samples. These are pseudo-labels that are filled in for the missing sample using other views. Representing the Latent features of the j-th sample in each view;
[0122] For all partially missing view samples, the confidence level is estimated by calculating the similarity between their completion features and the class center to which the pseudo-label belongs. Only completed samples with a confidence level higher than a set threshold are retained for subsequent training.
[0123] In this step, pseudo-labels are obtained based on the fused features. These pseudo-labels can be used to assist in subsequent missing view data completion. For a missing view, pseudo-labels are searched in the corresponding other views to determine the pseudo-label for the missing sample. Alternatively, data for that sample can be generated based on the mean feature and then filled in. The pseudo-labels and data can then be used to fill in the missing view. Finally, the feature similarity between the completed sample and its category center is calculated, and its confidence level is estimated. This allows for the retention of completed samples with confidence levels higher than a preset threshold, while eliminating completed samples with insufficient confidence levels. This reduces the negative impact of low-quality completed data on model training, achieving high-confidence completion for all missing view samples.
[0124] Step S4: In the fine-tuning stage, the general negative sample mechanism enhances the cluster boundary modeling, expands the discriminative boundary between categories, thereby improving the clustering robustness of multi-view data and achieving efficient matching of the target user's ad push type, such as pushing pet ads, sneaker ads, etc.
[0125] As a preferred embodiment, the boundary modeling and robust clustering based on the negative class sample mechanism during the fine-tuning stage are as follows:
[0126] Negative class samples are generated based on the multi-view data;
[0127] Construct a general negative class contrast loss between views and a general negative class contrast loss within views based on negative class samples;
[0128] The total loss during the fine-tuning stage is obtained by weighting and summing the reconstruction loss, the general negative class contrast loss between views, and the general negative class contrast loss within views.
[0129] By minimizing the total loss during the fine-tuning phase, the model parameters are optimized to improve the clustering performance and robustness of multi-view data.
[0130] Next, in the scenario of comparing views, the method for generating negative class samples based on the multi-view data is as follows:
[0131] Using the first view as the main view, identify all data in the first view that have pseudo-labels. Potential characteristics The negative class of the view, where i is the sample index. This represents the latent features of the i-th sample in the first view;
[0132] Define a set of category labels c represents the category label, c = 1, 2, ..., C;
[0133] Using the second view as an auxiliary view, the second view data defines a sample index set with pseudo-labels.
[0134]
[0135] Where N is the total number of samples within a view of data. For the j-th latent feature of the second view data pseudo-tags;
[0136] Based on sample index set Obtain the negative class centroid of the second view data with category label c.
[0137]
[0138] For all category labels, if the following conditions are met And c is not equal to Category labels will With the corresponding negative class centroid The corresponding initial negative class sample u is generated by mixing. i，c :
[0139]
[0140] Where λ is the mixing coefficient, λ∈(0,1);
[0141] Based on all initial negative class samples u i，c The average value is obtained. negative class sample U i , i = 1, 2, ..., N:
[0142]
[0143] Collect all negative class samples U i This yields the set of negative class samples between the first view data and the second view data.
[0144] The same method is applied to all data in the p-th and q-th views. The possible combinations of p and q include p = h and q = h+1, where h = 1, 2, ..., V-1, and V is the total number of views, as well as p = V and q = 1. This yields the set of negative class samples between all data in the p-th and q-th views.
[0145] Furthermore, in scenarios involving in-view comparison, the method for generating negative class samples based on the multi-view data is as follows:
[0146]
[0147] in, Let v be the negative class sample within the v-th view data, where v = 1, 2, ..., V. Augmentation data for the latent features of the j-th sample within the v-th view data.
[0148] Preferably, the method for constructing the general negative class contrast loss between views is as follows:
[0149] The view boundary term is defined based on the set of negative class samples of the contrast loss between the first view data and the second view data.
[0150]
[0151] Where τ is a coefficient controlling the steepness, exp is an exponential function, and S(.,.) is a function for calculating cosine similarity;
[0152] Obtain the negative class contrast loss between the first view data and the second view data.
[0153] in, and They are respectively The negative sample set and the positive sample set;
[0154] To achieve cross-view representation alignment, for each pair of adjacent view pairs (p,q), the corresponding cross-view contrast loss is calculated, where the combinations of p and q include p = h and q = h+1, h = 1, 2, ..., V-1, and p = V and q = 1;
[0155] The average loss among all view pairs is used as the general negative class contrast loss among views.
[0156] Based on this, the method for obtaining the general negative class contrast loss within the view is as follows:
[0157] Based on the set of general and sample data, define the view-in-view boundary item of the v-th view data.
[0158]
[0159] Obtain the general negative class contrast loss within the view.
[0160]
[0161] In step S4, a general negative class contrast loss within a view and a general negative class contrast loss between views are constructed based on the negative class mechanism. Combined with the reconstruction loss obtained in step S2, the clarity of cluster boundaries can be enhanced and the risk of false negatives can be reduced.
[0162] In summary, traditional clustering-based ad delivery typically only brings similar samples closer together and pushes different samples further apart. However, in real-world multimodal and multi-view scenarios, similar samples from different modalities may be incorrectly pushed apart. In other words, traditional clustering schemes are prone to false negatives when applied to advertising. Users with similar interests may be mistakenly identified as having different interests, leading to incorrect ad delivery by the recommendation algorithm. Furthermore, the model may treat similar users as negative samples, impacting the accuracy of feature learning and recommendations. This embodiment establishes a universal negative class mechanism, creating negative class contrast both between and within views. This mechanism expands the inter-class interval, enhancing the ability to distinguish class boundaries and reducing false negatives. Missing view samples are also filled in, improving data integrity. During the warm-up phase, reconstruction loss, intra-view contrast loss, and inter-view contrast loss are used to learn and optimize the basic representation. In the fine-tuning phase, a universal intra / inter-view negative class contrast loss is constructed based on negative class samples, and the representation is fine-tuned in conjunction with the reconstruction loss to enhance cluster boundaries and robustness. Based on the solution in this embodiment, user interests can be more accurately distinguished and erroneous recommendations can be reduced in ad delivery.
[0163] The above are merely preferred embodiments of the present invention and are not intended to limit the present 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 robust incomplete multi-view clustering method for general negative class learning guided by pseudo-labels, characterized in that, Includes the following steps: Obtain multi-view data for the target user. The multi-view data for the target user includes View 1, View 2, and View 3. View 1 is the user behavior view, and the input is X. 1 its samples View 1 represents the behavioral statistics of user i, including recent click sequence, browsing history, dwell time, and click frequency; View 2 is a demographic view, with X as the input. 2 its samples View 3 represents the static structured attributes of user i, including age, gender, region, and device type; View 4 is the conversion behavior view, with X as the input. 3 its samples This represents the historical conversion behavior characteristics of user i, including numerical characteristics of conversion events, such as adding to cart, placing an order, and making a payment. Numerical characteristics include frequency, amount statistics, and conversion cycle. During the warm-up phase, preliminary representation learning is performed based on multi-view data to generate pseudo-labels; Use pseudo-labels to complete missing view samples; During the fine-tuning phase, the general negative sample mechanism enhances cluster boundary modeling, expands the discriminative boundary between categories, thereby improving the robustness of clustering for multi-view data and enabling matching of ad push types to target users; The method for boundary modeling and robust clustering based on the negative class sample mechanism in the fine-tuning stage is as follows: Negative class samples are generated based on the multi-view data, and all negative class samples are collected. This yields the set of negative class samples between the first view data and the second view data. , Let v be the negative class sample within the data of the v-th view, where v = 1, 2, ..., V, and V is the total number of views. The latent features of the j-th sample within the v-th view data Enhanced data; Construct a general negative class contrast loss between views and a general negative class contrast loss within views based on negative class samples; The total loss during the fine-tuning stage is obtained by weighting and summing the reconstruction loss, the general negative class contrast loss between views, and the general negative class contrast loss within views. By minimizing the total loss during the fine-tuning phase, the model parameters are optimized to improve the clustering performance and robustness of multi-view data. The method for constructing the general negative class contrast loss between views is as follows: The view boundary term is defined based on the set of negative class samples of the contrast loss between the first view data and the second view data. : ； in, To control the steepness of the slope, exp is an exponential function. The function for calculating cosine similarity. The latent features of the i-th sample in the first view data The latent features of the i-th sample in the second view data; Obtain the negative class contrast loss between the first view data and the second view data. : ； in, and They are respectively The negative sample set and the positive sample set, N is the total number of samples within a view of data; To achieve cross-view representation alignment, for each pair of adjacent view pairs (p, q), the corresponding cross-view contrast loss is calculated, where the combinations of p and q include p=h and q=h+1, h=1,2,…,V-1, and p=V and q=1; The average loss among all view pairs is used as the general negative class contrast loss among the views. The method for obtaining the general negative class contrast loss within the view is as follows: Based on the set of general and sample data, define the view-in-view boundary item of the v-th view data. v=1,2,…,V: ； in, The latent features of the i-th sample in the v-th view Enhanced representation; Obtain the general negative class contrast loss within the view. : 。 2. The robust incomplete multi-view clustering method for pseudo-label-guided general negative class learning according to claim 1, characterized in that, The method for preliminary representation learning based on multi-view data during the warm-up phase is as follows: The original input sample is reconstructed based on the multi-view data using an encoder and decoder, and the reconstruction loss is obtained. Perform in-view contrast learning and out-of-view contrast learning separately to obtain in-view contrast loss and out-of-view contrast loss; The total loss during the warm-up phase is obtained by weighted summation of reconstruction loss, in-view contrast loss, and out-of-view contrast loss. By minimizing the total loss during the warm-up phase, the parameters of the encoder and decoder are jointly optimized to improve the initial representation quality of multi-view data.
3. The robust incomplete multi-view clustering method for pseudo-label-guided general negative class learning according to claim 2, characterized in that, The method for obtaining the reconstruction loss is as follows: ； in, For reconstruction loss, v represents the view number. The encoder for the v-th view data, For the decoder of the v-th view data, For the i-th sample of the v-th view data, This represents the total number of samples in the v-th view data.
4. The robust incomplete multi-view clustering method for pseudo-label-guided general negative class learning according to claim 3, characterized in that, The methods for performing intra-view comparison learning and inter-view comparison learning are as follows: Using Dropout to represent the latent features of the i-th sample in the v-th view Apply random perturbations to generate enhanced representations. Used to construct positive contrast samples within the view, v=1,2,…,V, i=1,2,…,N; Perform in-view comparison: ； in, For in-view contrast loss; Compare the first view data and the second view data between views: ； in, Loss due to comparison between first view data and second view data; All data in the p-th view and the q-th view are processed using the same method described above; The sum of the inter-view contrast losses for all p-th view data and q-th view data is used as the inter-view contrast loss.
5. The robust incomplete multi-view clustering method for general negative class learning guided by pseudo-labels according to claim 1, characterized in that, The method for obtaining the pseudo-tags is as follows: Multiple view data are merged and concatenated, and k-means clustering is used to obtain pseudo-labels; The method for completing missing view samples using pseudo-labels includes: For samples with missing views, the pseudo-labels generated in other views are used to return to the missing view. Based on the pseudo-labels, a set of samples with the same pseudo-labels and complete features in the missing view is retrieved. The complete feature representation of the missing view is constructed based on the feature mean of the set. Define a set of category labels , c represents the category label, c=1,2,…,C; Let the first The i-th sample is missing in each view of data, based on the corresponding mean feature. For the missing data Fill in: ， ； in, This represents the set of all pseudo-labeled samples within the view data that belong to the missing view samples. These are pseudo-labels filled in for the missing sample using other views. Representing the Latent features of the j-th sample in each view; For all partially missing view samples, the confidence level is estimated by calculating the similarity between their completion features and the class center to which the pseudo-label belongs. Only completed samples with a confidence level higher than a set threshold are retained for subsequent training.
6. The robust incomplete multi-view clustering method for pseudo-label-guided general negative class learning according to claim 1, characterized in that, In the scenario of comparing views, the method for generating negative class samples based on the multi-view data is as follows: Using the first view as the main view, identify all data in the first view that have pseudo-labels. Potential characteristics The negative class of the view, where i is the sample index; Define a set of category labels , c represents the category label, c=1,2,…,C; Using the second view as an auxiliary view, the second view data defines a sample index set with pseudo-labels. : ； in, For the j-th latent feature of the second view data pseudo-tags; Based on sample index set Obtain the negative class centroid of the second view data with category label c. : ； For all category labels, for those that satisfy And c is not equal to Category labels will With the corresponding negative class centroid Generate corresponding initial negative class samples through mixing : ； in, The mixing coefficient, ; Based on all initial negative class samples The average value is obtained from the average value of ... negative class samples , i=1,2,…,N: ； The same method is applied to all p-th and q-th view data to obtain a set of negative class samples between all p-th and q-th view data. .
7. The robust incomplete multi-view clustering method for pseudo-label-guided general negative class learning according to claim 6, characterized in that, In a scenario involving in-view comparison, the method for generating negative class samples based on the multi-view data is as follows: 。