Multi-view clustering method and system based on self-paced learning and view weighting
By employing self-learning and view weighting methods, the multi-view clustering model is optimized, addressing the issues of local minima and view quality differences, and achieving more efficient clustering accuracy and robustness.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- INST OF ELECTRONICS & INFORMATION ENG OF UESTC IN GUANGDONG
- Filing Date
- 2022-09-27
- Publication Date
- 2026-05-19
AI Technical Summary
Multi-view clustering is prone to getting stuck in local minima, and the quality of different views varies greatly, resulting in poor clustering performance.
We employ a self-stepping learning and view weighting approach. We gradually train simple and complex data using a self-stepping learning model, and then combine it with view weighting to optimize the objective function, update the clustering kernel and assignment matrix, and iteratively optimize the clustering results.
It improves the accuracy and robustness of multi-view clustering, reduces the impact of noise and outliers, and provides better clustering performance.
Smart Images
Figure CN115601578B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of multi-view data processing technology, specifically relating to a multi-view clustering method and system based on self-stepping learning and view weighting. Background Technology
[0002] With the development of information technology, the amount of data accumulated in various fields is increasing. Uncovering the valuable information hidden behind this data requires data mining techniques. Clustering, as an important data mining method, typically involves unsupervised learning scenarios. Its purpose is to divide a group of objects into different groups based on their characteristics. In this way, objects with similar characteristics or attributes will be grouped into the same cluster.
[0003] Ordinary clustering, which clusters data based on a single view, is called single-view clustering. Multi-view clustering, on the other hand, uses data with multiple different descriptive methods for clustering. Multi-view clustering is widely used in real-world data clustering, such as web page clustering, natural language clustering, image clustering, and multimodal learning. For specific learning problems, the minimum feature set is unknown. Without prior information, the more information we collect, the better the training results the model can achieve. Furthermore, with the development of information gathering techniques, people can describe objects from more perspectives. Single-view clustering methods are unsuitable for these problems because different features (or views) have different attributes. Therefore, multi-view clustering methods are more suitable for solving these problems.
[0004] However, both multi-view clustering and single-view clustering require a non-convex optimization function. However, during clustering, multi-view clustering often gets stuck in local minima, resulting in poor clustering performance.
[0005] The current popular solution is to perform multiple clustering operations using different initialization parameters, obtain different results, and select the optimal one. However, multi-view clustering can still get stuck in local minima, and this method is also too computationally expensive.
[0006] Self-walking learning (SPL) is a novel machine learning architecture whose core concept is to mimic the mechanisms of human cognition. First, the model applying self-walking learning is trained using simple, high-confidence data. Then, it is trained again using more difficult, complex, and low-confidence data. Simple data can be interpreted as data with low loss or high likelihood during training, while difficult data can be interpreted as data with high loss or low likelihood during training.
[0007] Experience has shown that SPL has the ability to avoid bad local minima, and therefore has better generality. Therefore, SPL is often used to find better solutions to non-convex problems with multiple local minima.
[0008] Furthermore, since the quality of different views in multi-view clustering often varies greatly, simply fusing information from different views in a general way rarely yields satisfactory results. To address this issue in multi-view clustering, assigning appropriate weights to different views is a necessary step to improve the final clustering performance. Summary of the Invention
[0009] This invention aims to solve the technical problems of multi-view clustering getting stuck in local minima and being susceptible to the influence of poor-quality views, so as to improve the clustering accuracy of the target dataset.
[0010] On one hand, the present invention provides a multi-view clustering method based on self-stepping learning and view weighting, the method comprising:
[0011] Step 1: Input the dataset to be clustered and the number of clusters k for the clustering task. Each data point in the dataset includes m views, where m is greater than 1.
[0012] The dataset to be clustered is then normalized to obtain the normalized dataset X = {X}. 1 X 2 , ..., X v , ..., X m}, where X v The view data representing view v, X v ∈R Dv×n Dv represents the feature dimension of view v, n represents the number of data points included in dataset X, and the view number v = 1, 2, ..., m;
[0013] Step 2, for each view data X v Feature concatenation is performed, and the data points of each concatenated view are processed by k-means clustering. The clustering kernel C of each view is initialized based on the clustering results. v And the allocation matrix B = {b1, b2, ..., b} j ..., b n}∈R k×n , where b j Let b represent the clustering representation of the j-th data point. j Let be a k-dimensional vector, where each dimension takes the value 0 or 1, and satisfies the following condition:
[0014] Step 3, calculate the weight η of each view. v :
[0015] Step 4: Calculate the loss for each data point using the objective function, and update the weight matrix W of each view sample based on the loss for each data point. v ;
[0016] The objective function is:
[0017]
[0018] in, This represents the weight of the i-th data point in the v-th view. Let λ represent the regularization function. v This represents the self-synchronous learning hyperparameters;
[0019] Step 5: Based on the current parameters, minimize the objective function and update the clustering kernel C of each view. v Based on the updated clustering kernel C v Minimize the objective function and update the allocation matrix B;
[0020] Step 6: Iterate through steps 3 to 5 until the iteration update termination condition is met, then output the final clustering result of the dataset: the most recently obtained clustering kernel C. v And the allocation matrix B.
[0021] Furthermore, the termination condition for iterative updates is: the objective function value converges or the maximum number of iterations is reached.
[0022] On the other hand, the present invention provides a multi-view clustering system based on self-stepping learning and view weighting, the system comprising: an acquisition module, a preprocessing module, a construction module, an optimization module, and a clustering output module;
[0023] in,
[0024] The acquisition module is used to acquire the clustering task and the dataset to be clustered, and send the dataset to the preprocessing module. The number of clusters k in the clustering task is sent to the construction module. Each data point in the dataset includes m views, where m is greater than 1.
[0025] The preprocessing module is used to normalize the dataset to be clustered before sending it to the construction module;
[0026] The module builds the view data X in the normalized dataset. v Feature concatenation is performed, and the data points of each concatenated view are processed by k-means clustering. The clustering kernel C of each view is initialized based on the clustering results. v And the allocation matrix B = {b1, b2, ..., b} j ..., b n}∈R k×n And according to the formula Calculate the weight η of each view v Where k represents the number of clusters, bj Let b represent the clustering representation of the j-th data point. j Let be a k-dimensional vector, where each dimension takes the value 0 or 1, and satisfies the following condition:
[0027] The building block will cluster kernel C v Assignment matrix B and the weights η of the view v Send to the optimization module;
[0028] The optimization module calculates the loss for each data point using an objective function and updates the weight matrix W of each view sample based on the loss for each data point. v ;
[0029] The objective function is:
[0030]
[0031] in, Let represent the weight of the i-th data point in the v-th view, and Let λ represent the regularization function. v This represents the self-synchronous learning hyperparameters;
[0032] And based on the current parameters, minimize the objective function and update the clustering kernel C of each view. v Based on the updated clustering kernel C v Minimize the objective function and update the allocation matrix B;
[0033] Iteratively update the view sample weight matrix W. v Clustering kernel C of the view v Assign matrix B until the iteration update termination condition is met, and then assign the most recently obtained clustering kernel C. v The assignment matrix B is sent to the clustering output module;
[0034] The clustering output module, based on the received clustering kernel C v The distribution matrix B outputs the final clustering results of the dataset.
[0035] The technical solution provided by this invention brings at least the following beneficial effects:
[0036] In this invention, a self-learning model is used to sequentially learn from the clustering data and ultimately obtain the clustering result through multi-view clustering. By assigning weights to views, the model can selectively learn information from different views. Compared with existing clustering techniques, the accuracy of clustering and various parameters have been further improved, providing a widely applicable architecture for multi-view clustering. Attached Figure Description
[0037] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0038] Figure 1 This is a flowchart of a multi-view clustering method based on self-stepping learning and view weighting provided in an embodiment of the present invention.
[0039] Figure 2 This is a schematic diagram of the structure of a multi-view clustering system based on self-stepping learning and view weighting provided in an embodiment of the present invention. Detailed Implementation
[0040] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.
[0041] This invention addresses the issues of non-convex optimization in existing multi-view clustering and its susceptibility to blurred views within the cluster structure. To improve the accuracy of clustering results, this invention provides a multi-view clustering method based on auto-stepping learning and view weighting, applicable to image data clustering. This method fully leverages the ability of auto-stepping learning to solve non-convex optimization problems while reducing the impact of outliers, noise, and poor-quality views on the clustering results. This invention offers a novel approach to processing clustered data, providing a new and universal architecture to improve clustering performance for most current multi-view clustering methods. This invention can be used for retrieval applications of database data (such as image databases, text databases, etc.). First, the target database is clustered into multiple clusters using this method, and the feature information of each cluster center is obtained. Then, the input information to be retrieved is matched with the feature information of each cluster center. The cluster to which the object to be retrieved belongs is determined based on the most matching cluster center. Finally, the object to be retrieved is retrieved within the matched clusters, and the retrieval results are output. For example, in image retrieval, the target image database is first clustered into multiple clusters based on this invention, and then the image feature vector of the center of each cluster is extracted (e.g., feature extraction by convolutional neural network, descriptive operator, etc.). For the image to be retrieved input by the user, the image feature vector is first extracted, and then the cluster to which the image to be retrieved belongs is matched by the distance between the image feature vectors. Then, the corresponding retrieval result is matched in the matched cluster by the image feature vector and fed back to the user.
[0042] like Figure 1 As shown in the figure, the specific implementation process of a multi-view clustering method based on self-stepping learning and view weighting provided in this embodiment of the invention is as follows:
[0043] Step 1: Construct the dataset used in the experiment. In this embodiment, the dataset consists of handwritten numerals, which contain 2000 data points, 10 data classes, each with the same amount of data, and each data point contains six views.
[0044] Step 2: Normalize the input dataset to obtain the normalized dataset X = {X} 1 X 2 , ..., X m}, where X m This represents the samples in the dataset, m represents the number of views, input the number of clusters K=10, and input other hyperparameters.
[0045] For a view v in the dataset, assuming the original X v ∈R Dv×n Where Dv is the feature dimension of view v, and n is the number of samples; in this embodiment, n is set to 2000. These matrices are normalized to obtain the updated multi-view data set, and the target cluster number k and each hyperparameter are input.
[0046] Step 3: Concatenate the multi-view data and initialize the clustering kernel C for each view using the k-means algorithm. v And the allocation matrix B.
[0047] After obtaining the dataset X from the normalization operation, the data from each view are concatenated to obtain... right The data was analyzed using the k-means algorithm to determine the cluster centers. Restore to each view to obtain the initial clustering kernel C for each view. v With the initialization allocation matrix B = {b1, b2, ..., b} j ..., b n}∈R k×n Among them, b j Let represent the clustering representation of the j-th sample. It is a k-dimensional vector, where each dimension takes the value 0 or 1, and satisfies ... Unlike the clustering kernel C, the assignment matrix B is universal across all views of the dataset, and its constraint is b. ij For elements ∈{0,1}, the assignment matrix B is intuitively represented as a k×n matrix where each column contains only one non-zero element with a value of 1. Furthermore, based on the self-stepping learning method, view weighting, and the objective function form of multi-view clustering, the objective function used in this invention is obtained:
[0048]
[0049] Where, η v Let λ be the weight of the v-th view, m represent the number of views, and λ be the weight of the v-th view. v For the self-learning control parameters of the v-th view, The sample weights are the weights of the i-th sample in the v-th view, represented by the view sample weight matrix. The function f() is a regularizer, and its specific expression can be found in formula (4) below.
[0050] Step 4: Calculate the weight η of each view. v The initial iteration time t is 0, which means t can also be expressed as the number of iterations.
[0051] In multi-view clustering, the quality of different views often varies, and simply fusing view information indiscriminately rarely yields satisfactory results. Therefore, this invention initializes loss information and assigns corresponding weights η to different views. v :
[0052]
[0053] Step 5: Calculate the loss caused by each sample point using the objective function, and update the weight matrix W of each view sample based on this. v .
[0054] After obtaining the allocation matrix B and the cluster kernel C v With view weight η v Then, the weight matrix W = {W} can be calculated. 1 W 2 , ..., W m}, where any W v Given an n×n diagonal matrix, calculate W based on the objective function. v The principle formula is:
[0055]
[0056] In the above formula, λ v This represents the self-synchronous learning hyperparameter, whose value increases with each iteration, thereby allowing more samples to be added to the training. Additionally:
[0057]
[0058] in, Because of l 2，1 The properties of the norm can transform the original problem into one where, for any sample in any view, let Seeking Apply this formula to Taking the derivative and setting the result to 0, we can obtain
[0059]
[0060] During training, the model continuously increases λ during iterations. v By gradually learning more and more samples, the value of is obtained, thus achieving self-paced learning. This involves finding all... Then, the weight matrix W was completed. v Update.
[0061] Step 6: Fix other parameters, minimize the objective function, and update the clustering kernel C for each view. v .
[0062] For each view clustering kernel C v The update, since the objective function is independent on different views, can be transformed into finding the objective function for any view v:
[0063]
[0064] stC v ≥0#(6)
[0065] Since the data has already been normalized in step 1, therefore C v It must satisfy the constraints, and C satisfies the conditions of the above equation. v It needs to be obtained through an iterative method, and auxiliary functions J and Z need to be constructed for derivation.
[0066] The auxiliary function J is defined as follows:
[0067] J(C v )=Tr((X v -C v B)D v (X v -C v B) T )#(7)
[0068] The auxiliary function Z is defined as follows:
[0069]
[0070] -2Tr(C vT X v D v B T )
[0071] stC v′ ≥0#(8)
[0072] Where Tr() represents the trace of the matrix, a diagonal matrix. have:
[0073]
[0074] As mentioned earlier, this step is performed separately for each view, therefore the superscript v is omitted. This rule will be followed in the subsequent derivation. B is the assignment matrix, W is the view weight matrix, and as shown in the formula, D is a diagonal matrix. X is used in the calculation process. i With G i C represents the i-th column vector of matrices X and G, respectively. v Represents the current clustering kernel of the view.
[0075] Z was chosen as the auxiliary function because it has two special properties:
[0076] Z(C v C v )=J(C v )
[0077] Z(C v C v )≤Z(C v C v′ )#(10)
[0078] The first property is easy to prove, while the second property can be proven by referring to the following inequality:
[0079]
[0080] Let A = I, H = C, and B = BD. v B T Substituting this into the equation, we can prove the second property of Z. Let f(C) v(t+1) )=Z(C v(t+1) C v(t) ), where t represents the current iteration time, for C v(t+1) Taking the derivative and setting it to zero, we finally get:
[0081]
[0082] Make f(C) v(t+1) C that achieves the minimum value t+1 for:
[0083]
[0084] in, Let i represent the cluster kernel representation of the v-th view v before and after the update, respectively, where i represents the cluster number and j represents the sample (data point) number.
[0085] Based on the properties of the matrices involved in the operation, the resulting C v(t+1) The constraints can still be met.
[0086] Furthermore, if we take the second derivative of Z(C, C′), we can obtain:
[0087]
[0088] Where, δ jl The value is 1 when j = l, and 0 in all other cases.
[0089] Obviously, Z(C) v(t+1) C v(t) The Hessian matrix of Z(C) is a positive semi-definite matrix, therefore, Z(C) v(t+1) C v(t) ) is a convex function, therefore the result obtained in equation (12) is It is the global minimum value.
[0090] Therefore, we can conclude that:
[0091] J(C v(t+1) )=Z(C v(t+1) C v(t+1) )≤Z(C v(t+1) C v(t) )≤Z(C v(t) C v(t) )=J(C v(t) )#(14)
[0092] Therefore:
[0093] Tr(X v -C v(t+1) B)D v (X v -C v(t+1) B) T ≤Tr(X v -C v(t) B)D v (X v -C v(t) B) T #(15)
[0094] After obtaining equation (15), it can be proved that:
[0095]
[0096]
[0097] Representing the left half of equation (16) as LHS (left hand side) and the right half as RHS (right hand side), we can obtain LHS-RHS:
[0098]
[0099] Where both LHS and RHS are ≤0, LHS-RHS means that the value of the objective function (6) can be reduced, thereby obtaining the optimal clustering result. Let b represent the i-th sample in the v-th view. i This represents the clustering representation of the i-th sample.
[0100] Therefore, we can conclude that:
[0101] ||(XC t+1 B)diag(W v )|| 2，1 ≤||(XC t+1 B)diag(W v )|| 2，1 #(18)
[0102] As can be seen from the above, the method proposed in equation (12) can find the C that makes the objective function value decrease under other determined conditions. v .
[0103] Step 7: Fix other parameters, minimize the objective function, update the allocation matrix B, and determine whether to continue iterating.
[0104] After obtaining the updated clustering kernel C through step 6, the allocation matrix B needs to be updated. According to the form of the objective function (1), the update rule for the allocation matrix is:
[0105]
[0106] As can be seen from equation (17), in this step, the influence of each sample i on the objective function is independent of each other, and since b i Let e be a column vector where each column has only one component that is 1 and the rest are 0. Therefore, for each sample i, we can directly set the column vector e. j Represents the identity matrix I∈R k×k The j-th column. Then, for this sample, the index j of the cluster center closest to it is calculated, and the column vector corresponding to this sample in B is updated to e. j :
[0107]
[0108] b i =e j (twenty one)
[0109] By using equations (20) and (21), the nearest cluster center for each sample is used as the reference for its assignment matrix, thus updating the assignment matrix B. Simultaneously, the iteration count t is incremented by one.
[0110] After updating the allocation matrix B, the objective function is recalculated. If the objective function has converged or the number of iterations t exceeds the upper limit of the number of iterations in the hyperparameters, the algorithm terminates; otherwise, it returns to step 5 to continue iterating.
[0111] Step 8: Output the final clustering core C v With the allocation matrix B.
[0112] Taking different types of multi-view datasets as examples, the multi-view clustering method based on self-stepping learning and view weighting provided in this embodiment of the invention is further illustrated. The specific implementation process of this embodiment includes:
[0113] S1. Normalize the input dataset to obtain the normalized dataset X = {X 1 X 2 , ..., X m Input the number of clusters k, and other hyperparameters;
[0114] S2. Concatenate the multi-view data and initialize the clustering kernel C for each view using the k-means algorithm. v and the allocation matrix B;
[0115] S3. Calculate the weights η of each view. v The initial iteration time t (number of iterations) is set to 0;
[0116] S4. Calculate the loss caused by each sample point using the objective function, and update the weight matrix W of each view sample based on this. v ;
[0117] S5. Fix other parameters, minimize the objective function, and update the clustering kernels C for each view. v ;
[0118] S6. Fix other parameters, minimize the objective function, update the allocation matrix B, and determine whether to continue iterating;
[0119] S7. Output the final clustering core C v With the allocation matrix B.
[0120] This embodiment evaluates the clustering performance of the proposed method on four widely used datasets, the benchmark datasets used in this example are shown in Table 1. The multi-view clustering method based on self-stepping learning and view weighting is compared with three advanced multi-view clustering methods: MVKKM, CAMVC, and MSPL, verifying the effectiveness of the algorithm.
[0121] Clustering performance was evaluated using clustering accuracy (ACC), normalized mutual information (NMI), and purity. Higher NMI, ACC, and purity values indicated better clustering performance. Experimental results are shown in Table 2.
[0122] Table 1. Benchmark Dataset Information
[0123]
[0124]
[0125] The Handwritten numerals dataset was selected from the UCI Machine Learning Repository. This dataset consists of 2000 points and features of handwritten digits (0-9). For each of the 10 data classes, there is the same amount of data, and these data points are represented by the following 6 features: 76 Fourier coefficients for feature shapes, 216 profile correlations, 64 Karhunen-Lovecoeffificients, a 240-pixel average of a 2×3 window, 47 Zernike moments, and 6 morphological features.
[0126] The BBCsport dataset comes from sports news reported by BBCSport. BBCsport consists of five types of articles published in 2012. Each article is divided into two sections, each representing a view and containing over 200 words logically related to the original article.
[0127] The Movies dataset is collected from IMDB and contains 617 movies with over 17 tags. Two views of the data are 1878 keywords used in three or more movies and 1398 actors.
[0128] The Reuters dataset selected 1200 articles from six categories (C15, CCAT, E21, ECAT, GCAT, and M11), with 200 articles provided for each category. Each document was written in five different languages (English, French, German, Italian, and Spanish), corresponding to five different views in the experiment.
[0129] Table 2
[0130]
[0131] As one possible implementation, a processing system employing the multi-view clustering method based on self-stepping learning and view weighting of this invention is as follows: Figure 2 As shown, its functional modules include:
[0132] The acquisition module is used to acquire clustering tasks and data samples;
[0133] The preprocessing module is used to normalize the input data;
[0134] The module constructs the data from multiple views, concatenates them, and uses the k-means algorithm to initialize the clustering kernel C for each view. v And the allocation matrix B; calculate the weight η of each view. v The initial iteration time t (number of iterations) is set to 0;
[0135] The optimization module calculates the loss caused by each sample point using an objective function, and updates the weight matrix W of each view sample based on this loss. v With other parameters fixed, minimize the objective function and update the clustering kernel C for each view. v With other parameters fixed, minimize the objective function, update the allocation matrix B, and determine whether to continue iterative optimization.
[0136] The clustering output module outputs the final clustering core, assignment matrix B, and the final clustering results.
[0137] Furthermore, in the optimization module, the specific optimization method is to calculate the loss caused by each sample point through the objective function, and update the weight matrix W of each view sample based on this. v .
[0138] After obtaining the allocation matrix B and the cluster kernel C v With view weight η v Then, the weight matrix W = {W} can be calculated. 1 W 2 , ..., W m}, where any W v Given an n×n diagonal matrix, calculate W based on the objective function. v The principle formula is:
[0139]
[0140] In the above formula, λ v This represents the self-synchronous learning hyperparameter, whose value increases with each iteration, thereby allowing more samples to be added to the training. Additionally:
[0141]
[0142] in, Because of l 2，1 The properties of the norm can transform the original problem into one where, for any sample in any view, let Seeking Apply this formula to Taking the derivative and setting the result to 0, we can obtain
[0143]
[0144] During training, the model continuously increases λ during iterations. v By gradually learning more and more samples, the value of is obtained, thus achieving self-paced learning. This involves finding all... Then, the update of the weight matrix Wv was completed.
[0145] It should be noted that the apparatus provided in the above embodiments is only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the apparatus and method embodiments provided in the above embodiments belong to the same concept, and their specific implementation process can be found in the method embodiments, which will not be repeated here.
[0146] This example integrates auto-learning into a multi-view clustering model, addressing the non-convexity problem by progressively selecting training samples from simple to complex, while utilizing a soft-weighted regularizer to significantly reduce the negative impact of noise and outliers.
[0147] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
[0148] The above descriptions are merely some embodiments of the present invention. Those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the scope of protection of the present invention.
Claims
1. A multi-view clustering method based on self-stepping learning and view weighting, characterized in that, This process is used for retrieving data from a database. It involves clustering the target database into multiple clusters and obtaining the feature information of the center of each cluster. The target database can be an image database or a text database. The clustering process includes: Step 1: Input the dataset to be clustered and the number of clusters k for the clustering task. The dataset is an image database or a text database. Each data point in the dataset includes m views, where m is greater than 1. The dataset to be clustered is then normalized to obtain the normalized dataset. ,in, This represents the view data of view v. Data dimensions are , n represents the feature dimension of view v, and n represents the dataset. Number of data points included, view number ; Step 2, process the data for each view. Feature concatenation is performed, and the data points of each concatenated view are processed by k-means clustering. The clustering kernels of each view are then initialized based on the clustering results. and allocation matrix ,in, This represents the clustering representation of the j-th data point. Let be a k-dimensional vector, where each dimension takes the value 0 or 1, and satisfies the following condition: ; Step 3: Calculate the weights of each view. : ; Step 4: Calculate the loss for each data point using the objective function, and update the weight matrix of each view sample based on the loss for each data point. ; The objective function is: ； in, , Let represent the weight of the i-th data point in the v-th view, and ; Represents the regularization function. This represents the self-synchronous learning hyperparameters; Step 5: Based on the current parameters, minimize the objective function and update the clustering kernels of each view. Based on the updated clustering kernel Minimize the objective function and update the allocation matrix B; Step 6: Iterate through steps 3 to 5 until the iteration update termination condition is met, then output the final clustering result of the dataset: the most recently obtained clustering kernel. And the allocation matrix B; For the input information to be retrieved, it is matched with the feature information of each cluster center. The cluster to which the object to be retrieved belongs is determined based on the most matching cluster center. Then, the object to be retrieved is retrieved in the matching cluster and the retrieval results are output. The current information to be retrieved is image data or text data.
2. The method as described in claim 1, characterized in that, Update the weight matrix of each view sample Specifically: ； Among them, the weight matrix .
3. The method as described in claim 1, characterized in that, Update the clustering kernel of each view Specifically: when satisfying Under the premise, based on Complete clustering kernel Update.
4. The method as described in claim 1, characterized in that, The specific steps for updating the allocation matrix B are as follows: .
5. The method according to any one of claims 1 to 4, characterized in that, Regularization function for: ； in, , Represents the natural base.
6. The method as described in claim 1, characterized in that, The iteration update ends when the objective function value converges or the maximum number of iterations is reached.
7. A multi-view clustering system based on self-stepping learning and view weighting, characterized in that, include: The module includes an acquisition module, a preprocessing module, a construction module, an optimization module, and a clustering output module. The acquisition module is used to acquire the clustering task and the dataset to be clustered, wherein the dataset is an image database or a text database; and sends the dataset to the preprocessing module, and sends the cluster number k in the clustering task to the construction module. Each data point in the dataset includes m views, where m is greater than 1; wherein the data points are image data or text data. The preprocessing module is used to normalize the dataset to be clustered before sending it to the construction module; The module builds the view data in the normalized dataset. Feature concatenation is performed, and the data points of each concatenated view are processed by k-means clustering. The clustering kernels of each view are then initialized based on the clustering results. and allocation matrix ∈ And according to the formula Calculate the weights of each view. ,in, This represents the clustering representation of the j-th data point. Let be a k-dimensional vector, where each dimension takes the value 0 or 1, and satisfies the following condition: ; The building block will cluster the kernel Allocation matrix and view weight Send to the optimization module; The optimization module calculates the loss for each data point using an objective function and updates the weight matrix of each view sample based on the loss for each data point. ; The objective function is: ； in, , Let represent the weight of the i-th data point in the v-th view, and ; Represents the regularization function. This represents the self-synchronous learning hyperparameters; And based on the current parameters, minimize the objective function and update the clustering kernels of each view. Based on the updated clustering kernel Minimize the objective function and update the allocation matrix B; Iteratively update the view sample weight matrix one by one. Clustering kernel of views Assign matrix B until the iteration update termination condition is met, and then assign the most recently obtained clustering kernel. The assignment matrix B is sent to the clustering output module; The clustering output module, based on the received clustering kernels The distribution matrix B outputs the final clustering results of the dataset.