Explanatable general multi-scene clustering method and system based on optimized heuristic network

Abstract

The invention provides an interpretable general multi-scene clustering method and system based on an optimized heuristic network, and aims to provide a general clustering framework suitable for homogeneous and heterogeneous multi-view scenes. According to the method, optimization design and a deep learning network are combined, and a network module capable of learning sharing and self-expression representation is constructed, so that the relationship and similarity between views are extracted from multi-view data, and efficient clustering is realized. Meanwhile, each network component in the method has practical significance, the interpretability of the model is ensured, and the problem of insufficient interpretability of the model caused by a conventional deep method is effectively solved. In addition, the training loss function proposed by the method can effectively enhance the affinity between similar attribute objects. The clustering method has good universality and adaptability, clustering tasks can be effectively executed in different multi-view scenes, and certain practical value is achieved.

Description

Technical Field

[0001] The present invention proposes an interpretable general multi-scenario clustering method and system based on an optimization-inspired network, which relates to the fields of deep learning, optimization-inspired networks and multi-view clustering. Background Art

[0002] With the rapid development of modern data-driven environments, data sources and forms of expression are becoming more and more diverse, making multi-view data analysis one of the important research directions. Multi-view data refers to data observed from multiple different feature spaces or dimensions for the same research object, such as user interaction information on social platforms, image feature representation, and text data. Different views can often provide complementary information, and integrating this information can achieve a more comprehensive and accurate understanding of the data. Multi-view clustering, as a typical data analysis method, aims to explore the inherent consistency and complementarity between data from different views to obtain accurate sample partitioning results. It has been widely used in fields such as recommendation systems, social network analysis, and disease diagnosis.

[0003] The above data forms can be regarded as homogeneous multi-views. In the real world, heterogeneous multi-view data, as a common form of data representation, also reflects the different semantic relationships between objects. For example, in a social network, each user establishes different social relationships based on different attributes, and these different relationships can be regarded as different views. However, existing multi-view clustering methods often mainly deal with homogeneous view data, and their ability to simultaneously process heterogeneous and homogeneous view data is relatively weak. Therefore, how to design an interpretable, versatile and efficient multi-view clustering method that can simultaneously process homogeneous and heterogeneous multi-view data is still an important challenge currently faced.

[0004] At the same time, existing multi-view clustering methods are usually based on deep neural networks. Although they have powerful feature extraction capabilities, they lack interpretability due to their black-box characteristics and are difficult to meet the high requirements for model transparency and credibility. To solve this problem, explainable artificial intelligence has become a hot topic in current research, including post-explanation methods and strategies that incorporate interpretable mechanisms in the model design stage. Optimization-inspired network architecture is an effective way to provide interpretability in the model design stage. It transforms the iterative process of traditional optimization problems into neural network modules with clear explanatory meanings to achieve higher performance and model transparency.

[0005] Although optimized networks have been attempted several times in the fields of signal and image processing and some progress has been made, the research on multi-view data is still relatively limited. Moreover, due to the limitations of the original data structure, these methods are only applicable to homogeneous or heterogeneous multi-view data. Therefore, more multi-view clustering methods based on optimization-inspired networks that are applicable to multiple scenarios need to be developed, taking into account both scenario applicability and interpretability. Summary of the Invention

[0006] To solve the above problems, the present invention proposes an interpretable general multi-scenario clustering method and system based on an optimization-inspired network. By combining optimization design with a deep learning network, the method constructs a network module that can learn shared and self-expressive representations, thereby extracting the relationships and similarities between views in multi-view data to achieve efficient clustering.

[0007] An interpretable general multi-scenario clustering method and system based on an optimization-inspired network, wherein an interpretable general multi-scenario clustering method based on an optimization-inspired network includes the following steps:

[0008] Step S1: Input multi-view data X v and A v , reduce the dimension of the features of each view data through the encoding matrix E v to map the data from each perspective to a unified low-dimensional space, thereby obtaining a consistent shared representation H; and introduce a regularization term to reduce the interference of noise and redundant information and optimize the quality of the shared representation;

[0009] Step S2: By introducing a manifold regularization term, further optimize the shared representation H based on the similarity between samples, learn the local geometric structure and similarity of each sample, so that samples with similar features remain close in the shared representation space;

[0010] Step S3: Design a self-representation matrix S to represent the weighted combination of each sample on other samples; and by introducing a sparsity regularization term, make the representation of each sample more accurate while reducing the influence of redundancy and noise;

[0011] Step S4: Use an alternating optimization strategy to convert the original optimization problem into optimization sub-problems of multiple sub-matrices, and in this process, construct an interpretable neural network module, with each module responsible for solving a specific sub-problem;

[0012] Step S5: Calculate the loss function of the neural network based on the shared representation obtained in Step S2, and update the parameters in each network module and the network according to this loss function using backpropagation and gradient descent algorithms;

[0013] Step S6: Output the optimized self-representation matrix S.

[0014] Further, step S1 specifically includes the following steps:

[0015] Step S11: Dimensionally reduce the data and encode it into a unified representation space H. The optimization objective of this method is:

[0016]

[0017] where is the encoding matrix of each view feature; β is a balancing parameter.

[0018] where represents the l1 norm of the matrix, v represents the subscript of the current view, V is the total number of views; X v represents the feature matrix of the current view, and A v represents the topological matrix of the current view.

[0019] Further, step S2 specifically includes the following steps:

[0020] Step S21: Calculate the pairwise similarity matrix P of the multi-view data through a Gaussian kernel:

[0021]

[0022] where σ represents the weight control parameter. A larger σ value makes points at a distance still have a relatively large similarity weight, and a smaller σ value makes only very close points be considered similar.

[0023] Step S22: Further introduce a manifold regularization term:

[0024]

[0025] where Tr(·) represents the trace of the matrix, L = D - P represents the graph Laplacian matrix, and D represents the degree diagonal matrix; h i , h j respectively represent the i-th and j-th rows of the shared matrix H;

[0026] Step S23: According to the manifold regularization term, make samples with similar features close in the shared representation space, and the optimization objective is further expressed as:

[0027]

[0028] where the parameters α and β are balancing parameters.

[0029] Further, step S3 specifically includes the following steps:

[0030] Step S31: Represent each sample of H as a weighted combination on other samples to obtain the self - representation matrix S; and introduce a sparsity regularization term to enhance the accuracy of the feature representation of each sample, which is specifically expressed as:

[0031]

[0032] where the parameters γ and δ are balance parameters.

[0033] Furthermore, the specific steps of step S4 are as follows:

[0034] Step S41: According to the alternating optimization strategy and the variant of the proximal gradient descent algorithm, decompose the original optimization problem into several sub - problems, and achieve alternating iterative optimization by separately optimizing the shared representation matrix H and the self - representation matrix S. The optimization steps for each matrix are as follows:

[0035] Update H:

[0036]

[0037] Update S:

[0038]

[0039] where J represents the Lipschitz constant of the quadratic term, ξ θ and ζ θ are sparsity - parameterized regularizers;

[0040] Step S42: Replace the iterative variables through learnable network layers to achieve the integration of the deep neural network and interpretability. The optimization process is further expressed as:

[0041]

[0042] where W is the learnable layer, V is the number of views, and I is the identity matrix.

[0043] Furthermore, the specific steps of step S5 are as follows:

[0044] Step S51: Calculate the loss function of the current network according to the formula For homogeneous multi - view data, M v = P v ; while for heterogeneous data, M v = A v ;

[0045] Step S52: According to the loss function calculated in step S51, use the gradient descent and back - propagation algorithms to update the learnable parameters Θ = {W1, W2, W3, W4, θ1, θ2} in the neural network; where θ is a learnable parameter used to control the sparsity intensity.

[0046] Step S53: If the loss function of the neural network converges to a stable value, proceed to step S6; otherwise, return and continue to execute step S4.

[0047] Further, step S6 specifically includes the following steps:

[0048] Step S61: Output the optimal self - representation matrix S.

[0049] According to the second aspect of the present invention, an interpretable general multi - scenario clustering system based on an optimized heuristic network is proposed. An interpretable general multi - scenario clustering based on an optimized heuristic network includes an electronic device. An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. It is characterized in that when the processor executes the computer program, it implements an interpretable general multi - scenario clustering method proposed by the present invention.

[0050] According to the third aspect of the present invention, an interpretable general multi - scenario clustering system based on an optimized heuristic network is proposed. An interpretable general multi - scenario clustering based on an optimized heuristic network includes a computer - readable storage medium. The computer - readable storage medium stores a computer program. It is characterized in that when the computer program is executed by a processor, it implements an interpretable general multi - scenario clustering method proposed by the present invention.

[0051] The present invention has the following advantages:

[0052] Compared with the prior art, the present invention has the following optimizations:

[0053] 1. The present invention can simultaneously process homogeneous and heterogeneous multi - view data, expanding the limitations of traditional methods that only process homogeneous view data. This makes the present invention more versatile in practical applications and can better handle complex data scenarios, such as social network analysis, etc.

[0054] 2. Based on an optimized heuristic network architecture, the present invention transforms traditional optimization problems into interpretable neural network modules. Through a transparent calculation process, it provides higher interpretability. Different from the black - box characteristics of traditional deep - learning methods, the present invention can provide clear explanations for each decision, improving the transparency and credibility of the model.

[0055] 3. By designing an effective shared representation and self - representation matrix optimization mechanism, the present invention can improve the clustering accuracy of multi - view data. When processing multi - view data, by fusing information from different views, the present invention can more accurately capture the similarity between samples, thus achieving a better clustering effect. Description of the Drawings

[0056] Figure 1 This is the process diagram of the present invention.

[0057] Figure 2 This is the neural network architecture diagram of the present invention.

[0058] Figure 3 This is the clustering result of different methods of the present invention on the DBLP dataset.

[0059] Figure 4 This is the quality of the self - representation matrix of the present invention on the Notting - Hill dataset. Detailed implementation manners

[0060] The technical solution of the present invention will be specifically described below with reference to the accompanying drawings.

[0061] It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present application belongs.

[0062] It should be noted that the terms used herein are only for describing specific implementation manners and are not intended to limit the exemplary embodiments according to the present application; as used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0063] As Figure 1 、 Figure 2 、 Figure 3 and Figure 4 shown, the embodiments of the present invention provide an interpretable general multi - scenario clustering method and system based on an optimized heuristic network, including the following:

[0064] An interpretable general multi - scenario clustering method and system based on an optimized heuristic network, wherein an interpretable general multi - scenario clustering method based on an optimized heuristic network includes the following steps:

[0065] Step S1: Input multi - view data X v and A v , reduce the dimension of the features of each view data through the encoding matrix E v , map the data of each perspective to a unified low - dimensional space, so as to obtain a consistent shared representation H; and introduce a regularization term to reduce the interference of noise and redundant information and optimize the quality of the shared representation;

[0066] Step S2: By introducing a manifold regularization term, based on the similarity between samples, further optimize the shared representation H, learn the local geometric structure and similarity of each sample, so that samples with similar features remain close in the shared representation space;

[0067] Step S3: Design a self-representation matrix S to represent the weighted combination of each sample on other samples; and by introducing a sparsity regularization term, make the representation of each sample more accurate, while reducing the influence of redundancy and noise;

[0068] Step S4: Use the alternating optimization strategy to transform the original optimization problem into optimization sub-problems of multiple sub-matrices, and in this process, construct an interpretable neural network module, and each module is responsible for solving a specific sub-problem;

[0069] Step S5: According to the shared representation obtained in Step S2, calculate the loss function of the neural network, and update the parameters in each network module and the network according to this loss function using the backpropagation and gradient descent algorithms;

[0070] Step S6: Output the optimized self-representation matrix S.

[0071] Furthermore, Step S1 specifically includes the following steps:

[0072] Step S11: Reduce the dimensionality of the data and encode it into a unified representation space H. The optimization objective of this method is:

[0073]

[0074] where is the encoding matrix of each view feature; β is a balance parameter,

[0075] where represents the l1 norm of the matrix, v represents the subscript of the current view, V is the total number of views; X v represents the feature matrix of the current view, A v represents the topological matrix of the current view.

[0076] Furthermore, Step S2 specifically includes the following steps:

[0077] Step S21: Calculate the pairwise similarity matrix P of multi-view data through a Gaussian kernel:

[0078]

[0079] where σ represents the weight control parameter. A larger σ value makes points at a distance still have a relatively large similarity weight, and a smaller σ value makes only very close points be considered similar.

[0080] Step S22: Further introduce the manifold regularization term:

[0081]

[0082] where Tr(·) represents the trace of a matrix, L = D - P represents the graph Laplacian matrix, and D represents the degree diagonal matrix; h i , h j represent the i-th and j-th rows of the shared matrix H respectively;

[0083] Step S23: According to the manifold regularization term, make samples with similar features close in the shared representation space, and the optimization objective is further expressed as:

[0084]

[0085] where the parameters α and β are balance parameters.

[0086] Furthermore, step S3 specifically includes the following steps:

[0087] Step S31: Represent each sample of H as a weighted combination of other samples to obtain the self-representation matrix S; and introduce a sparsity regularization term to enhance the accuracy of the feature expression of each sample, which is specifically expressed as:

[0088]

[0089] where the parameters γ and δ are balance parameters.

[0090] Furthermore, step S4 specifically includes the following steps:

[0091] Step S41: According to the alternating optimization strategy and a variant of the proximal gradient descent algorithm, decompose the original optimization problem into several sub-problems, and achieve alternating iterative optimization by separately optimizing the shared representation matrix H and the self-representation matrix S. The optimization steps for each matrix are as follows:

[0092] Update H:

[0093]

[0094] Update S:

[0095]

[0096] where J represents the Lipschitz constant of the quadratic term, ξ θ and ζ θ are sparsity parameterized regularizers;

[0097] Step S42: Replace the iterative variables through learnable network layers to achieve the integration of deep neural networks and interpretability. The optimization process is further expressed as:

[0098]

[0099] Among them, W is a learnable layer, V is the number of views, and I is the identity matrix.

[0100] Further, the step S5 specifically includes the following steps:

[0101] Step S51: According to the formula calculate the loss function of the current network. For homogeneous multi-view data, M v = P v ; while for heterogeneous data, M v = A v ;

[0102] Step S52: According to the loss function calculated in step S51, use the gradient descent and backpropagation algorithms to update the learnable parameters Θ = {W1, W2, W3, W4, θ1, θ2} in the neural network; where θ is a learnable parameter for controlling the sparsity intensity.

[0103] Step S53: If the loss function of the neural network converges to a stable value, go to step S6; otherwise, return to continue executing step S4.

[0104] Further, the step S6 specifically includes the following steps:

[0105] Step S61: Output the optimal self-representation matrix S.

[0106] According to the second aspect of the present invention, an interpretable general multi-scenario clustering system based on an optimized heuristic network is proposed. An interpretable general multi-scenario clustering based on an optimized heuristic network includes an electronic device. An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. It is characterized in that when the processor executes the computer program, it implements an interpretable general multi-scenario clustering method proposed by the present invention.

[0107] According to the third aspect of the present invention, an interpretable general multi-scenario clustering system based on an optimized heuristic network is proposed. An interpretable general multi-scenario clustering based on an optimized heuristic network includes a computer-readable storage medium. The computer-readable storage medium stores a computer program. It is characterized in that when the computer program is executed by a processor, it implements an interpretable general multi-scenario clustering method proposed by the present invention.

[0108] This example starts from practical applications and aims to effectively process homogeneous and heterogeneous multi-view data. At the same time, to ensure the interpretability of the model, the present invention combines an optimization-inspired network architecture to transform traditional optimization problems into learnable neural network modules. The calculation process of each module can provide clear explanations, enhancing the transparency of clustering decisions. In addition, according to the obtained shared representation matrix, the loss function of the neural network is calculated, and the learnable parameters in each module are updated using backpropagation and gradient descent algorithms according to the loss function until the loss function converges to a stable value; finally, the optimized shared representation matrix is output. This embodiment is based on an optimization-inspired network architecture, which can effectively fuse multi-view data, improve the clustering effect, has strong generality and interpretability, and has broad application prospects. This embodiment provides a performance comparison with other existing popular clustering schemes on 2 homogeneous (MIT Indoor and Notting-Hill) and 2 heterogeneous multi-view datasets (DBLP and YELP), as shown in Table 1.

[0109]

[0110] Table 1 Performance comparison between the embodiment of the present invention and the prior art

[0111] As shown in Table 1, based on the above design, this embodiment conducts a performance comparison with other existing popular clustering schemes on 2 homogeneous (MIT Indoor and Notting-Hill) and 2 heterogeneous multi-view datasets (DBLP and YELP) in the real world. MIT Indoor is a scene dataset containing 5,360 images covering 67 categories. In this dataset, four types of features are extracted in this embodiment of the present invention, including 4096-dimensional PHOW features, 3600-dimensional LBP features, 1770-dimensional CENTRIST features, and 1240-dimensional depth features. Notting-Hill is a video-based facial image dataset collected from an existing movie. The dataset contains 4660 facial images from five main actors, distributed in 76 segments. For characterization, 2000-dimensional intensity features, 3304-dimensional local binary pattern (LBP) features, and 6750-dimensional Gabor features are extracted. The DBLP dataset identifies three views, including co-authorship, co-conference, and co-terms. YELP is sourced from a database Sumer review website and contains four types of nodes: business, user, service, and rating.

[0112] For homogeneous multi-view data, the embodiments of the present invention compared MCGC, LMVSC, RCAGL, FMVSC, SDSNE, CVCL, and SCMVC. For heterogeneous multi-view data, the embodiments of the present invention compared MvCGC, CGL, GMC, MVAGC, O2MAC, DuaLGR, and VGMGC, and obtained the results of four performance indicators: clustering accuracy (ACC), normalized mutual information (NMI), adjusted rand index (ARI), and F-score. The higher the value, the better the performance. The embodiments of the present invention conducted ten trials on all the experiments and recorded the mean and standard deviation as the decisive results.

[0113] Under different scenarios, the perspectives and structures of the data may vary. The embodiments of the present invention can adapt to various data forms by designing a shared representation learning and self-representation matrix optimization mechanism. This enables the method to handle the complexities in multi-view clustering simultaneously, such as heterogeneous relationships in social networks and multi-dimensional feature data in recommendation systems, enhancing its wide application ability.

[0114] Under different scenarios, the perspectives and structures of the data may vary. The present invention can adapt to various data forms by designing a shared representation learning and self-representation matrix optimization mechanism. This enables the method to handle the complexities in multi-view clustering simultaneously, such as heterogeneous relationships in social networks and multi-dimensional feature data in recommendation systems, enhancing its wide application ability.

[0115] Specifically, the DBLP dataset demonstrates the effectiveness of the method of the embodiments of the present invention in academic collaboration networks, while the MIT Indoor dataset demonstrates its application in computer vision. It successfully extracts meaningful information from different image features for accurate clustering.

[0116] As can be seen from Table 1, the proposed method exhibits superior performance on each dataset, demonstrating its effective ability to learn comprehensive multi-view information and capture the consistency between multi-views, thereby improving the clustering performance. Compared with traditional optimization methods, the method has significant advantages. It is interpretable at the model design level and combines the advantages of deep learning in nonlinear processing. This combination achieves robust feature extraction and topological structure learning, which helps to more deeply explore the correlation and consistency between views. Compared with existing deep optimization methods, the proposed method also shows satisfactory performance, verifying the rationality of the optimization objective and the effectiveness of the network module in guiding representation learning.

[0117] For homogeneous multi-view data, the framework not only focuses on multi-view features but also considers the local geometric structure within each view. This method can effectively integrate multiple sample attributes and mine the relationships between samples. For heterogeneous multi-view data, the framework can flexibly capture the consistent information in different topological structures, while fusing node attributes and their relationships, enabling nodes with similar features and relationships to be better clustered. By learning the shared representation and self-representation matrices, more accurate and comprehensive multi-view information can be obtained.

[0118] In addition, the embodiments of the present invention introduce a user-friendly method to demonstrate the discriminative ability of various methods through t-SNE visualization on the DBLP dataset shown in Figure 3 . It is worth noting that the method of the embodiments of the present invention shows a more consistent structure and clearly separated clusters, further verifying its superior clustering ability.

[0119] Figure 4 Shows the t-SNE visualization of the learned self-representation matrix at training iterations 1, 30, 60, and 90 when using the Notting-Hill dataset. It can be seen that during the iteration process, the cluster assignment becomes more reasonable, indicating that the self-representation matrix extracts the relationships and similarities between views and is thus continuously optimized during the iteration process to achieve efficient clustering.

[0120] Generally speaking, the proposed method exhibits remarkable superiority and wide applicability, can effectively process homogeneous and heterogeneous multi-view data, verifies its effectiveness in different multi-view scenarios, and demonstrates satisfactory performance in the clustering task. At the same time, it has the interpretability of traditional methods while using deep networks for representation learning, which is conducive to further expansion.

[0121] The above are only the preferred embodiments of the present invention, and all equivalent changes and modifications made according to the scope of the patent application of the present invention shall fall within the scope covered by the present invention.

Claims

1. An interpretable general multi-scenario clustering method based on an optimized heuristic network, characterized in that: Including the following steps: Step S1: Input multi-view data X v and A v , reduce the dimension of the features of each view data through the encoding matrix E v , map the data of each perspective to a unified low-dimensional space, so as to obtain a consistent shared representation H; and introduce a regularization term to reduce the interference of noise and redundant information and optimize the quality of the shared representation; Step S2: By introducing a manifold regularization term, based on the similarity between samples, further optimize the shared representation H, learn the local geometric structure and similarity of each sample, so that samples with similar features remain close in the shared representation space; Step S3: Design a self-representation matrix S to represent the weighted combination of each sample on other samples; and by introducing a sparsity regularization term, make the representation of each sample more accurate while reducing the influence of redundancy and noise; Step S4: Use the alternating optimization strategy to transform the original optimization problem into optimization sub-problems of multiple sub-matrices, and during this process, construct an interpretable neural network module, where each module is responsible for solving a specific sub-problem; Step S5: According to the shared representation obtained in Step S2, calculate the loss function of the neural network, and update the parameters in each network module and the network according to this loss function using the backpropagation and gradient descent algorithms; Step S6: Output the optimized self-representation matrix S.

2. The interpretable general multi-scenario clustering method based on an optimized heuristic network according to claim 1, wherein: The specific steps of Step S1 include the following steps: Step S11: Reduce the dimensionality of the data and encode it into a unified representation space H. The optimization objective of this method is: Among them is the encoding matrix of each view feature, and β is the balance parameter; represents the l1 norm of the matrix, v represents the subscript of the current view, and V is the total number of views; X v represents the feature matrix of the current view, and A v represents the topological matrix of the current view.

3. An interpretable general multi-scenario clustering method based on an optimized heuristic network according to claim 2, characterized in that: The specific steps of Step S2 include the following steps: Step S21: Calculate the pairwise similarity matrix P of multi-view data through a Gaussian kernel: where σ represents the weight control parameter. A larger σ value makes points at a distance still have a relatively large similarity weight, and a smaller σ value makes only very close points be considered similar; Step S22: Further introduce a manifold regularization term: where Tr(·) represents the trace of a matrix, L = D - P represents the graph Laplacian matrix, D represents the degree diagonal matrix; h i , h j represent the i-th and j-th rows of the shared matrix H, respectively; Step S23: According to the manifold regularization term, make samples with similar features remain close in the shared representation space, and the optimization objective is further expressed as: where the parameters α and β are balance parameters.

4. An interpretable general multi-scenario clustering method based on an optimized heuristic network according to claim 3, characterized in that: The specific steps of Step S3 include the following steps: Step S31: Represent each sample of H as a weighted combination on other samples to obtain the self-representation matrix S; and introduce a sparsity regularization term to enhance the accuracy of the feature expression of each sample, specifically expressed as: where the parameters γ and δ are balance parameters.

5. An interpretable general multi-scenario clustering method based on an optimized heuristic network according to claim 4, characterized in that: The specific steps of Step S4 include the following steps: Step S41: According to the alternating optimization strategy and the deformation of the proximal gradient descent algorithm, disassemble the original optimization problem into several sub-problems, and achieve alternating iterative optimization by separately optimizing the shared representation matrix H and the self-representation matrix S. The optimization steps for each matrix are as follows: Update H: Update S: where J represents the Lipschitz constant of the quadratic term, ξ θ and ζ θ are sparsity parameterized regularizers; Step S42: Replace the iterative variables through learnable network layers to achieve the integration of the deep neural network and interpretability; the optimization process is further expressed as: where W is the learnable layer, W1 represents the learnable layer 1, W2 represents the learnable layer 2, W3 represents the learnable layer 3, W4 represents the learnable layer 4, V is the total number of views, and I is the identity matrix.

6. An interpretable general multi-scenario clustering method based on an optimized heuristic network according to claim 5, characterized in that: The specific steps of Step S5 include the following steps: Step S51: According to the formula calculate the loss function of the current network. For homogeneous multi-view data, M v = P v ; while for heterogeneous data, M v = A v ; Step S52: According to the loss function calculated in Step S51, use the gradient descent and backpropagation algorithms to update the learnable parameters Θ = {W1, W2, W3, W4, θ1, θ2} in the neural network; where θ is a learnable parameter used to control the sparsity intensity; Step S53: If the loss function of the neural network converges to a stable value, proceed to step S6; otherwise, return to step S4 and continue execution.

7. An interpretable general multi-scenario clustering method based on an optimized heuristic network according to claim 6, characterized in that: The specific steps of step S6 are as follows: Step S61: Output the optimal self-representation matrix S.

8. An interpretable general multi-scenario clustering system based on an optimized heuristic network, including an electronic device; wherein the electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements an interpretable general multi-scenario clustering method based on an optimized heuristic network as described in any one of claims 1 to 7.

9. An interpretable general multi-scenario clustering system based on an optimized heuristic network, including a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, characterized in that, When the computer program is executed by the processor, it implements an interpretable general multi-scenario clustering method based on an optimized heuristic network as described in any one of claims 1 to 7.

Images