Unsupervised multi-view clustering method for problems of clustering information imbalance and information loss
Through the dual network structure and contrast learning method of the delay activation strategy, the consistency and complementary information in multi-view data are balanced, and the problems of view missing and information imbalance are solved, which significantly improves clustering performance and is suitable for incomplete multi-views. Analysis of data.
Patent Information
- Application Number
- CN202410983333.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-22
- Publication Date
- 2024-11-26
AI Technical Summary
When existing multi-view clustering methods process incomplete multi-view data, it is difficult to balance consistency and complementary information, resulting in unsatisfactory clustering effect. Especially in the absence of views and unbalanced information, complementary information is ignored, which affects the impact of performance.
Using a dual network structure with a delay activation strategy, it explicitly predicts the view-specific representation by minimizing conditional entropy, and maximizes the mutual information between different views through comparative learning, thereby achieving enhanced consistency learning, balanced complementarity and consistency information. The final clustering is performed by combining the K-means clustering algorithm.
In the case of high missing rates, clustering performance is significantly improved, complementary information in multi-view data can be effectively utilized, the quality of clustering results is improved, and the risk of model crash is reduced. It is suitable for large-scale data sets.
Smart Images

Figure CN119027701A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of artificial intelligence and multi-view data clustering analysis technology, and in particular relates to an unsupervised multi-view clustering method for clustering information imbalance and information missing problems Background Art
[0002] With the development of information technology and the diversification of data collection methods, multi-view data are generated in more and more application scenarios. These data come from different sources or perspectives and contain rich and complementary information. For example, in social network analysis, the behavior data of the same user can be obtained from multiple perspectives such as different social platforms and activity records; in medical image analysis, different types of patient images (such as X-ray, CT, MRI) also provide multi-view data. These multi-view data provide a more comprehensive information basis for data mining and analysis. However, existing multi-view clustering methods face challenges in processing incomplete multi-view data. Specifically, there are missing views and incomplete data in incomplete multi-view data, which makes it difficult to fully utilize the complementary information between different views, affecting the clustering effect. Traditional multi-view clustering methods usually focus on the consistency between views, but often ignore the complementary information between different views. Due to the incompleteness of views and the neglect of complementary information, existing methods are not ideal when processing incomplete multi-view data.
[0003] In order to solve the above problems, multi-view clustering algorithms have attracted widespread attention, but their performance is limited due to the following shortcomings: 1) Most multi-view clustering methods only focus on the consistency information or complementary information in multiple views, and usually cannot effectively balance the use of consistency and complementary information at the same time. 2) Based on consistency learning, only complete data can be used, but complete data in real scenarios is very scarce and insufficient to learn consistent information. In addition, the complementary information hidden in incomplete data is particularly important for multi-view clustering. Summary of the invention
[0004] In order to overcome the shortcomings of the above-mentioned prior art and the problems of complementary information being ignored and consistency learning being insufficient due to information imbalance and view missing, the purpose of the present invention is to provide an unsupervised multi-view clustering method for clustering information imbalance and information missing problems, which can map the original multi-view data into a latent space of unified dimensionality through an automatic encoder.
[0005] In order to achieve the above objectives, the scheme of the present invention can be described as follows: First, an effective balance of complementarity and consistency information in multi-view data is achieved through a dual network structure with a delayed activation strategy. Secondly, for the view missing problem, its specific representation is explicitly predicted by minimizing the conditional entropy. Then, the mutual information between different views is maximized through contrastive learning to enhance consistency learning. Finally, the above schemes are merged into a unified reconstruction process to avoid model collapse and produce trivial solutions. The present invention further improves efficiency based on existing data and achieves fundamentally better clustering results.
[0006] The technical solution adopted by the present invention can be further described in detail as follows:
[0007] An unsupervised multi-view clustering method for clustering information imbalance and information missing problems includes the following steps:
[0008] Step 1: preprocess the multi-view data to make the scales of each view consistent, and divide it into complete data and incomplete data; the complete data means that all views of the sample are not missing, and the incomplete data means that at least one view of the sample is missing;
[0009] Step 2: construct M groups of dual network structures and M decoders, where M is the number of view types of a single multi-view data, and each view corresponds to a group of dual network structures and a decoder, wherein the dual network structure is used for feature mining and information balance; the decoder is used for feature reconstruction; wherein each group of dual network structures is composed of an online network and a target network, wherein the online network is used to learn the differences between different views, and the target network is used to maintain the differences to the greatest extent;
[0010] Step 3: Based on the data from step 1 and the network from step 2, consistency learning is performed to balance complementary information and consistency information.
[0011] Step 4: Under the action of the dual network structure, joint optimization is performed, and iterations are repeated until the network converges to achieve the best balance between complementary information and consistency information. After the network converges, multi-view data containing complete data and incomplete data is input and transmitted to the target network of the dual network structure. The K-means clustering algorithm is used to perform clustering on the obtained public representation to obtain the final clustering result.
[0012] In one embodiment, in step 1, the multi-view data of the sample are represented as views of the same sample at different angles or views of the same sample at different modalities at the same angle.
[0013] In one embodiment, the online network consists of an encoder E θ , a projector P θ and a predictor Q θ The target network consists of an encoder Eδ and a projector P δ In each group of the dual network structures, each encoder is used to project the view into a potential representation, each projector is used to project the potential representation into a specific dimension, and the predictor Q θ Used to linearly transform the features passed through the encoder and projector, with the dimension consistent with the projector.
[0014] In one embodiment, the online network and the target network perform parameter optimization through contrastive learning to achieve a dynamic balance between complementary information and consistency information; the target network in the M groups of dual network structures projects the M views of the sample into M potential representations, respectively, and restores the potential representation of the missing views in the sample by minimizing conditional entropy, while enhancing consistency learning by maximizing the mutual information between views.
[0015] In one embodiment, in each group of the dual network structures, the output results of the online network and the target network are compared and learned to achieve complementary learning; the parameter update method of the online network is gradient descent, and the parameter update method of the target network is the exponential moving average of the online network, so as to ensure that the M groups of dual network structures and the M decoders are forced to pay attention to the characteristics of the data at different time points during the training process, so as to understand the data from multiple angles; through the different parameter update methods of the online network and the target network, that is, the delayed activation mechanism, the balance of complementary information and consistency information in the learning process is achieved.
[0016] In one embodiment, the process of performing consistency learning in step 3 is as follows:
[0017] In the initial stage, the dual network structure and decoder are trained with complete data, and then the incomplete data is input into the dual network structure and decoder. Finally, the potential representation of the missing view in the sample is restored by minimizing the conditional entropy, while maximizing the mutual information between views to learn the consistency information between views.
[0018] The process of balancing complementary information and consistent information is as follows:
[0019] In the initial stage, only complete data is used to train the dual network structure and decoder, and delayed activation parameters are set to dynamically retain the complementary information of each view to achieve a dynamic balance between complementary information and consistent information.
[0020] In one embodiment, in step 4, the original multi-view data is passed to the dual network structure in which the objective function L has converged and the corresponding decoder, the missing data will be predicted and restored, and consistency learning will be performed; then, complementary information and consistent information are optimally balanced; finally, a high-dimensional public representation of the multi-view is obtained, based on which, clustering is performed using the K-means clustering algorithm to obtain the final clustering result.
[0021] Compared with the prior art, the present invention has the following beneficial effects:
[0022] When faced with the problem of missing views, existing multi-view clustering methods can only explore consistent information based on existing complete data. On the one hand, the features contained in the missing views are ignored. On the other hand, complementary information is not fully utilized. In fact, the rational use of the above information can greatly improve the clustering performance. In addition, for the study of public semantics, most of the existing methods are based on traditional machine learning, and the high complexity limits its scalability on large-scale datasets.
[0023] In order to solve the problem of the difficulty in balancing complementary information and consistent information and the problem of missing information in the multi-view clustering process, the present invention restores the representation of the missing view by minimizing the conditional entropy, and enhances the consistency learning by maximizing the mutual information through contrastive learning. In addition, a dual network structure with a delayed activation strategy is designed and implemented to balance complementarity and consistency. This solves the problem of imbalance between complementary information and consistent information and missing information. Even when the data is missing at a high rate, the best clustering performance can be achieved. All network models are based on deep learning, and the introduced deep neural network has a high generalization ability and has significant advantages in processing incomplete multi-view data, providing a new solution for the analysis and application of multi-view data. BRIEF DESCRIPTION OF THE DRAWINGS
[0024] Figure 1 It is a schematic diagram of the process of the present invention.
[0025] Figure 2 It is a dual network structure diagram of the present invention.
[0026] Figure 3 It is a framework diagram of the method of the present invention. DETAILED DESCRIPTION
[0027] In order to make the objectives, technical solutions and advantages of the present invention clearer, the embodiments of the present invention are further described below in conjunction with the accompanying drawings.
[0028] The present invention is an unsupervised multi-view clustering method for the problems of unbalanced and missing clustering information. It is used in the case of unbalanced and missing multi-view data information, where the complementary information generated is ignored and consistency learning is insufficient. An incomplete multi-view clustering method that balances consistency and complementarity is proposed.
[0029] The concepts and parameters that may be involved in the present invention are introduced as follows:
[0030] There may be missing views in the sample. In a given data set, multi-view data of multiple samples are included. Multi-view data refers to views of the same sample from different angles or views of the same sample from different modes at the same angle. Multi-view data contains rich and complementary information. In social network analysis, the behavior data of the same user can be obtained from multiple perspectives such as different social platforms and activity records; in medical image analysis, different imaging types of patients (such as X-ray, CT, MRI) also provide multi-view data. In multi-view data, the incompleteness of views and the neglect of complementary information make it difficult to fully utilize the complementary information between different views, affecting the clustering effect. In view of this phenomenon, the unsupervised multi-view clustering method for information imbalance and information missing problems of the present invention can still maintain good clustering performance when the above situation exists in actual situations.
[0031] refer to Figure 1 , the complete process of the present invention is as follows:
[0032] S1, multi-view data preprocessing.
[0033] The main purpose of preprocessing is to make the scales of each view consistent and to divide it into complete data and incomplete data. Complete data means that all views of the sample are not missing, and incomplete data means that at least one view of the sample is missing.
[0034] For the sake of convenience, taking dual views as an example, this embodiment first collects multi-view data of several samples, and the number of view data of each sample is 2. Secondly, the present invention performs standardization processing on the data to ensure that the scale of the view data is consistent. Finally, the present invention divides the samples into complete data and incomplete data; complete data means that both views of the sample are not missing, and incomplete data means that one of the two views of the sample is missing; obviously, based on this principle, the expansion of the number of view data of the present invention to other situations has the same processing method.
[0035] The data set of this embodiment has two view data, with a total of N samples, where each of the N samples has complete two view data. The data set is divided into two parts: complete data X N , incomplete data W N . N X The view data of samples constitutes the complete data set, N W The view data of samples constitute an incomplete data set, where N = N X +N W In the complete data set, use and They respectively represent the feature vectors of the Vth view of the Nth sample.
[0036] S2, initialize the dual network structure.
[0037] The model of the present invention is mainly composed of M groups of dual network structures and M decoders, where M is the number of view types of a single multi-view data, and each view corresponds to a group of dual network structures and a decoder.
[0038] The dual network structure is used for feature mining and information balance, such as Figure 2 As shown, the dual network A consists of an online network A1 and a target network A2. Specifically, the online network is used to learn the differences between different views, and consists of an encoder A projector and a predictor The target network is used to maximize the difference and consists of an encoder and a projector Each set of dual network structures corresponds to a decoder, which is used for feature reconstruction. Figure 3 As shown, the decoder corresponding to the dual network A1 is D1, and the decoder corresponding to the dual network A2 is D2. The target network in the dual network A1 will be shown in Figure X (1) Projected into the latent representation Z (1) , the target network in the dual network A2 will view X (2) Projected into the latent representation Z (2) Obviously, when the view data volume is other cases, the principle is similar, the projector For Z (1) Projection to a specific dimension, projector For Z (2) Projected to a specific dimension, predictor It is used to linearly transform the features passed through the encoder and projector, and the dimension is consistent with the projector. (1) and Z (2) Prediction recovery is performed by minimizing conditional entropy, and consistency learning is enhanced by maximizing mutual information through contrastive learning.
[0039] Specifically, the online network and the target network in the dual network perform parameter optimization through contrastive learning to dynamically balance the complementary information and consistent information in the learning process. Among them, the target network in the M sets of dual network structures projects the M views of the sample into M potential representations, and restores the potential representation of the missing views in the sample by minimizing the conditional entropy, while enhancing the consistency learning by maximizing the mutual information between views.
[0040] In this embodiment, complementary learning is achieved by comparing the output results of the online network and the target network of each group of dual network structures. The parameter update method of the online network is gradient descent, and the parameter update method of the target network is the exponential moving average of the online network, so as to ensure that the M groups of dual network structures and M decoders are forced to pay attention to the characteristics of the data at different time points during the training process, so as to understand the data from multiple angles; through the different parameter update methods of the online network and the target network, that is, the delayed activation mechanism, the balance of complementary information and consistency information in the learning process is achieved.
[0041] The superiority of the present invention is evaluated on four public datasets below.
[0042] 1) Caltech101-20 consists of 2,386 images of 20 subjects, with views of Histogram of Oriented Gradients (HOG) and Global Signature Information (GIST) features, with 1,984 and 512 feature dimensions, respectively. Scene-15 consists of 4,485 images distributed in 15 scene categories, with views of GIST and Histogram of Oriented Gradients Pyramid (PHOG) features, 59D and 20D feature vectors, respectively. LandUse-21 consists of 2,100 satellite images of 21 categories, with PHOG and Local Binary Pattern (LBP) features, 59D and 40D feature vectors, respectively. Noisy MNIST consists of 70k instances of 10 categories. 15k original instances are randomly selected as views. Figure 1 , and randomly select 15k Gaussian noise instances as the view Figure 2 The detailed statistics of the dataset are summarized in Table 1.
[0043] Table 1 Dataset summary
[0044] Dataset Sample size Number of categories Dimensions Caltech101-20 2,386 20 1,984 / 512 Scene-15 4,485 15 59 / 20 LandUse-21 2,100 21 59 / 40 Noisy MNIST 70k 10 784 / 784
[0045] S3, utilizes complete data training, consistency learning, and balance of complementarity and consistency information.
[0046] In this step, at the initial stage, the dual network structure and decoder are trained with complete data, and then the incomplete data is input into the dual network structure and decoder. Finally, the potential representation of the missing views in the sample is restored by minimizing the conditional entropy, while maximizing the mutual information between views to learn the consistency information between views.
[0047] In the above contrastive learning, three objective functions are constructed based on M potential representations:
[0048] i) The reconstruction loss L obtained by reconstructing different views through the encoder and decoder of the target network rec for:
[0049]
[0050] Among them, X n (v) Represents view X (v) The vth view of the nth sample of E, where N is the total number of samples, (v) represents the target network encoder of the vth view, D (v) Denote the decoder of the v-th view, the representation of the v-th view of the n-th sample is:
[0051] Z n (v) =E (v) (X n (v) )
[0052] Among them, Z (v) Represents view X (v) potential representation.
[0053] ii) The dual network structure mines feature information through contrastive learning without negative samples; the goal of the online network is to learn more about the differences between different views, and the goal of the target network is to maintain this difference to the greatest extent; the loss of the dual network structure consists of two parts, namely the loss of the online network and the loss of the target network; the loss of the online network is expressed as:
[0054]
[0055] Among them, Z (i) =Q(P(f(X (i) ))) represents view X (i) Through the output of the online network, the view X (1) Input into the target network to obtain a symmetric loss function That is, the loss of the target network; then the complementary loss is expressed as:
[0056]
[0057] The balance between complementary information and consistent information in multi-view clustering is achieved by delaying the activation of the dual network; the momentum parameter m of the delay activation mechanism is updated as follows:
[0058] δ←mδ+(1-m)θ
[0059] Among them, θ is the parameter of the online network, δ is the parameter of the target network;
[0060] iii) Consistency learning and missing view recovery are integrated in the same optimization framework. First, the view-specific representation is explicitly predicted by minimizing the conditional entropy. Second, the mutual information between different views is maximized through contrastive learning to ensure consistency learning. Finally, the consistency learning loss is expressed as:
[0061] L cnl =L pre +L ccl
[0062] The loss for minimizing the conditional entropy to predict a view-specific representation is:
[0063]
[0064] Among them, P (i) and P (j) They represent the network that approximately minimizes the conditional entropy, Z (i) and Z (j) They are the representations of mutual predictions between samples;
[0065] The loss for maximizing the mutual information between different views is:
[0066]
[0067] Among them, μ is the balanced regularization hyperparameter, H is the information entropy, the larger H is, the more information it contains, and I represents the mutual information;
[0068] The overall objective function obtained is as follows:
[0069] L=αL cml +βL cnl +γL rec
[0070] Wherein α, β, and γ are balancing factors of complementarity loss, consistency loss, and reconstruction loss, respectively, and are set to 0.001, 0.01, and 1, respectively, in this embodiment.
[0071] For the convenience of description, we still take the dual view as an example, and the complete data is represented as {X (1) ,X (2)}, the complete data X (1) Passed to the target network of dual network A1 Get Z (1) , X (2) Passed to the target network of dual network A2 Get Z (2) , Z (1) and Z (2) are the latent representations of the first and second views, respectively.
[0072] Based on this objective function back propagation, the dual network A and decoder D are updated by gradient descent. Training with complete data provides a good foundation for subsequent clustering.
[0073] Step 4: In the initial stage, only complete data is used to train the dual network structure and decoder, and delayed activation parameters are set to dynamically retain the complementary information of each view to achieve a dynamic balance between complementary information and consistent information.
[0074] In the above steps 3 and 4, complete data is first used for training to converge the overall objective function L, and then incomplete data is used for testing.
[0075] Step 5: Under the dual network structure, joint optimization is performed, and iterations are repeated until the network converges, so that the complementary information and consistency information reach the best balance. Finally, after the network converges, multi-view data containing complete data and incomplete data is input and transmitted to the target network of the dual network structure. The K-means clustering algorithm is used to perform clustering on the obtained public representation to obtain the final clustering result.
[0076] Specifically, the original multi-view data is passed to the dual network structure whose objective function L has converged and the corresponding decoder. The missing data will be predicted and restored, and consistency learning will be performed. When the network converges, complementary information and consistent information reach the best balance. Finally, after the network converges, multi-view data containing complete data and incomplete data is input to obtain a high-dimensional public representation of the multi-view. Based on this, the K-means clustering algorithm is used to perform clustering to obtain the final clustering result.
[0077] The evaluation index selection of the present invention adopts three indicators, namely normalized mutual information (NMI), accuracy (ACC) and adjusted Rand index (ARI), to evaluate clustering performance. The higher the value of these evaluation indicators, the better the clustering performance.
[0078] Table 2 shows the loss components and experimental results corresponding to each component in the method. In (1), it can be seen that due to L pre It is not optimized, so the low-dimensional potential representation will lose more complementary information. Therefore, the contrast prediction loss L is used alone. pre It may lead to trivial solutions or model collapse. By comparing (4)(5)(6) with (1)(2)(3), it can be seen that the double loss is more effective than the single consistency learning module. By comparing (7) with (2), it can be asserted that the two modules of contrastive learning and mutual prediction have greatly improved the clustering effect. It is worth noting that each module of the present invention improves the clustering performance, which further proves the effectiveness of the method.
[0079] Table 2 Ablation experiments on the Caltech101-20 dataset
[0080] <![CDATA[L rec ]]> <![CDATA[L cml ]]> <![CDATA[L ccl ]]> NMI ACC ARI (1) √ 31.22 45.20 21.84 (2) √ 21.03 12.73 3.22 (3) √ 44.14 58.57 41.03 (4) √ √ 30.72 41.97 20.78 (5) √ √ 69.21 70.24 82.31 (6) √ √ 58.78 61.12 66.87 (7) √ √ √ 76.78 70.51 87.01
[0081] Table 3 describes the clustering performance of the present invention under data sets of different scales. It can be seen that the present invention can effectively deal with information imbalance and view loss, and significantly improve the clustering performance.
[0082] Table 3 Clustering performance on different data sets
[0083]
[0084]
[0085] In summary, the present invention first projects the original dataset into a unified latent space, and solves the problem of the difficult balance between complementarity and consistency in multi-view data through a delayed activation mechanism. For the problem of insufficient consistency learning, this method restores incomplete information by minimizing conditional entropy and maximizing mutual information between different views, thereby achieving multi-view Figure 1 Enhanced consistency learning. The delayed activation mechanism allows the network to gradually introduce complementary information from each view during the clustering process, thereby avoiding information loss. The present invention integrates the balance problem and the consistency learning problem into a unified framework, which can better utilize the inherent correlation of multi-view data and significantly improve clustering performance.
Claims
1. An unsupervised multi-view clustering method for clustering information imbalance and information missing problems, characterized in that: The steps include: Step 1: preprocess the multi-view data to make the scales of each view consistent, and divide it into complete data and incomplete data; the complete data means that all views of the sample are not missing, and the incomplete data means that at least one view of the sample is missing; Step 2: construct M groups of dual network structures and M decoders, where M is the number of view types of a single multi-view data, and each view corresponds to a group of dual network structures and a decoder, wherein the dual network structure is used for feature mining and information balance; the decoder is used for feature reconstruction; wherein each group of dual network structures is composed of an online network and a target network, wherein the online network is used to learn the differences between different views, and the target network is used to maintain the differences to the greatest extent; Step 3: Based on the data from step 1 and the network from step 2, consistency learning is performed to balance complementary information and consistency information. Step 4: Under the action of the dual network structure, joint optimization is performed, and iterations are repeated until the network converges to achieve the best balance between complementary information and consistency information. After the network converges, multi-view data containing complete data and incomplete data is input and transmitted to the target network of the dual network structure. The K-means clustering algorithm is used to perform clustering on the obtained public representation to obtain the final clustering result.
2. The unsupervised multi-view clustering method for clustering information imbalance and information missing problems according to claim 1, characterized in that: In step 1, the multi-view data of the sample are represented as views of the same sample at different angles or views of the same sample at different modalities at the same angle.
3. The unsupervised multi-view clustering method for clustering information imbalance and information missing problems according to claim 1, characterized in that: The online network consists of an encoder E θ , a projector P θ and a predictor Q θ The target network consists of an encoder E δ and a projector P δ In each group of the dual network structures, each encoder is used to project the view into a potential representation, each projector is used to project the potential representation into a specific dimension, and the predictor Q θ Used to linearly transform the features passed through the encoder and projector, with the dimension consistent with the projector.
4. The unsupervised multi-view clustering method for clustering information imbalance and information missing problems according to claim 1, characterized in that: The online network and the target network perform parameter optimization through contrastive learning to achieve a dynamic balance between complementary information and consistent information; the target network in the M groups of dual network structures projects the M views of the sample into M potential representations, and restores the potential representation of the missing views in the sample by minimizing the conditional entropy, while enhancing the consistency learning by maximizing the mutual information between views.
5. The unsupervised multi-view clustering method for clustering information imbalance and information missing problems according to claim 4, characterized in that: In each group of the dual network structures, the output results of the online network and the target network are compared and learned to achieve complementary learning; the parameter update method of the online network is gradient descent, and the parameter update method of the target network is the exponential moving average of the online network, so as to ensure that the M groups of dual network structures and the M decoders are forced to pay attention to the characteristics of the data at different time points during the training process, so as to achieve understanding of the data from multiple angles; through the different parameter update methods of the online network and the target network, that is, the delayed activation mechanism, the balance between complementary information and consistent information in the learning process is achieved.
6. The unsupervised multi-view clustering method for clustering information imbalance and information missing problems according to claim 4 or 5, characterized in that: Three objective functions are constructed based on the M potential representations: i) The reconstruction loss L obtained by reconstructing different views through the encoder and decoder of the target network rec for: Among them, X n (v) Represents view X (v) The vth view of the nth sample of E, where N is the total number of samples, (v) represents the target network encoder of the vth view, D (v) Denote the decoder of the v-th view, the representation of the v-th view of the n-th sample is: Z n (v) =E (v) (X n (v) ) Among them, Z (v) Represents view X (v) potential representation of; ii) The dual network structure mines feature information through contrastive learning without negative samples; the goal of the online network is to learn more about the differences between different views, and the goal of the target network is to maintain this difference to the greatest extent; the loss of the dual network structure consists of two parts, namely the loss of the online network and the loss of the target network; the loss of the online network is expressed as: Among them, Z (i) =Q(P(f(X (i) ))) represents view X (i) Through the output of the online network, the view X (1) Input into the target network to obtain a symmetric loss function That is, the loss of the target network; then the complementary loss is expressed as: The balance between complementary information and consistent information in multi-view clustering is achieved by delaying the activation of the dual network; the momentum parameter m of the delay activation mechanism is updated as follows: δ←mδ+(1-m)θ Among them, θ is the parameter of the online network, δ is the parameter of the target network; iii) Consistency learning and missing view recovery are integrated in the same optimization framework. First, the view-specific representation is explicitly predicted by minimizing the conditional entropy. Second, the mutual information between different views is maximized through contrastive learning to ensure consistency learning. The consistency learning loss is expressed as: L cnl =L pre +L ccl The loss for minimizing the conditional entropy to predict a view-specific representation is: Among them, P (i) and P (j) They represent the network that approximately minimizes the conditional entropy, Z (i) and Z (j) They are the representations of mutual predictions between samples; The loss for maximizing the mutual information between different views is: Among them, μ is the balanced regularization hyperparameter, H is the information entropy, the larger H is, the more information it contains, and I represents the mutual information; The overall objective function obtained is as follows: L=αL cml +βL cnl +γL rec Among them, α, β, and γ are the balancing factors of complementarity loss, consistency loss, and reconstruction loss, respectively.
7. The unsupervised multi-view clustering method for clustering information imbalance and information missing problems according to claim 6, characterized in that: α, β, and γ are set to 0.001, 0.01, and 1, respectively.
8. The unsupervised multi-view clustering method for clustering information imbalance and information missing problems according to claim 6, characterized in that: In step 3, first complete data is used for training to converge the overall objective function L, and then incomplete data is used for testing.
9. The unsupervised multi-view clustering method for clustering information imbalance and information missing problems according to claim 8, characterized in that: In step 3, the process of consistency learning is as follows: In the initial stage, the dual network structure and decoder are trained with complete data, and then the incomplete data is input into the dual network structure and decoder. Finally, the potential representation of the missing view in the sample is restored by minimizing the conditional entropy, while maximizing the mutual information between views to learn the consistency information between views. The process of balancing complementary information and consistent information is as follows: In the initial stage, only complete data is used to train the dual network structure and decoder, and delayed activation parameters are set to dynamically retain the complementary information of each view to achieve a dynamic balance between complementary information and consistent information.
10. The unsupervised multi-view clustering method for clustering information imbalance and information missing problems according to claim 6, characterized in that: In step 4, the original multi-view data is passed to the dual network structure in which the objective function L has converged and the corresponding decoder, the missing data is predicted and restored, and consistency learning is performed; then, complementary information and consistent information are optimally balanced; finally, a high-dimensional public representation of the multi-view is obtained, based on which, clustering is performed using the K-means clustering algorithm to obtain the final clustering result.
Citation Information
Cited By
Incomplete multi-view clustering method and system based on prototype matching learning
CN119963862A
