A Distributed Data Privacy-Preserving Grouped Knowledge Mining Method Based on Heterogeneous Anchor Tensor Graph

Abstract

This invention discloses a distributed data privacy-preserving grouped knowledge mining method based on heterogeneous anchor tensor graphs, belonging to the field of data privacy protection technology. The method includes: each client constructing a local hybrid anchor graph based on raw data lacking local sample-level views, generating a sparse hybrid anchor graph for each client; on each client side, using the sparse hybrid anchor graph as input, performing grouped knowledge mining to cluster and measure client heterogeneity, outputting the anchor-cluster mapping matrix and heterogeneity coefficients for each client; on the server side, based on the anchor-cluster mapping matrix uploaded by all clients, constructing a global anchor tensor and performing consistency-diversity joint modeling, outputting a consistency sub-tensor and a diversity sub-tensor optimized by joint modeling. This invention improves the accuracy of high-quality grouped knowledge mining of distributed heterogeneous data while strictly protecting data privacy.

Description

Technical Field

[0001] This invention relates to the field of data privacy protection technology, and in particular to a distributed data privacy protection group knowledge mining method based on heterogeneous anchor tensor graphs. Background Technology

[0002] Multi-view clustering technology under the federated learning framework provides a feasible path for collaborative grouping and knowledge mining of multi-source data stored in a decentralized manner without sharing the original data. Existing technologies are usually based on the model averaging paradigm, which requires each client to complete the learning of view features locally and then update the global model through parameter aggregation and distribution. This method lays the foundation for distributed learning under privacy protection.

[0003] Existing methods are generally based on the ideal assumption that each client view is complete and aligned. In practical applications, due to limitations in device capabilities and data permissions, view missing phenomena are extremely common, resulting in view heterogeneity among clients. This real challenge makes it difficult for traditional methods to build a unified joint learning space. They are either forced to discard incomplete data, resulting in information loss, or they cannot effectively integrate complementary information from multiple views, ultimately limiting their clustering performance and practicality in real-world scenarios. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a distributed data privacy-preserving grouped knowledge mining method based on heterogeneous anchor tensor graphs to solve the problem that existing technologies struggle to effectively construct a unified learning space and fully utilize multi-view information in a federated environment due to view missingness and heterogeneity.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a distributed data privacy-preserving grouped knowledge mining method based on heterogeneous anchor tensor graphs, comprising: application in a distributed system including a server and multiple clients, wherein each client holds original multi-view data with missing sample-level views; the method comprising: Based on the original data missing from the local sample-level view, each client constructs its own local hybrid anchor point map, generating a sparse hybrid anchor point map for each client. On each client side, using the sparse hybrid anchor point graph as input, grouped knowledge mining is performed to cluster and measure client heterogeneity, outputting the anchor point-cluster mapping matrix and heterogeneity coefficient for each client; On the server side, based on the anchor-cluster mapping matrix uploaded by all clients, a global anchor tensor is constructed and consistency-diversity joint modeling is performed, outputting the consistency sub-tensor and diversity sub-tensor after joint modeling optimization; On the server side, heterogeneity coefficients of each client and the consistency subtensor and diversity subtensor are used to perform heterogeneity-aware federated aggregation, generate and output a global anchor tensor. Based on the global anchor tensor, global guidance information is generated and sent to each client for iterative optimization. After the preset convergence condition is met, the final global sample clustering label is output.

[0007] As a preferred embodiment of the distributed data privacy-preserving grouped knowledge mining method based on heterogeneous anchor tensor graphs described in this invention, wherein: each client holds original multi-view data with missing sample-level views, including the following steps: For each existing view of each client, a clustering algorithm is used to select M of the most representative sample centers as physical anchor points; The cross-view mapping function generates virtual anchor points for the corresponding missing views. The training objective of the cross-view mapping function is to minimize the sum of the reconstruction error of the view alignment samples and the regularization term to prevent overfitting. By combining physical anchors and virtual anchors, a unified set of hybrid anchors is formed across all client dimensions.

[0008] As a preferred embodiment of the distributed data privacy-preserving grouped knowledge mining method based on heterogeneous anchor tensor graphs described in this invention, the method involves: each client constructing a local hybrid anchor graph based on the original data lacking local sample-level views, thereby generating a sparse hybrid anchor graph for each client, including the following steps: Based on the hybrid anchor set, the Gaussian kernel similarity between local samples and anchors is obtained, and a sample-anchor bipartite graph is constructed. Perform a truncation operation on the sample-anchor bipartite graph, retaining only the connection weights between each sample and its top s nearest neighbor anchors, and output a sparse mixed anchor graph for each client.

[0009] As a preferred embodiment of the distributed data privacy-preserving grouped knowledge mining method based on heterogeneous anchor tensor graphs described in this invention, the method includes the following steps: On each client side, using the sparse mixed anchor graph as input, grouped knowledge mining is performed to cluster and measure client heterogeneity, outputting the anchor-cluster mapping matrix and heterogeneity coefficients for each client. Using a sparse hybrid anchor graph as input, we construct an objective function that minimizes the difference between the product of the sparse hybrid anchor graph and the cluster indicator matrix and the anchor-cluster mapping matrix. In the objective function, the row sparsity of the anchor-cluster mapping matrix is ​​constrained by the Lp,q grouped sparse regularization term. The objective function with the Lp,q grouped sparse regularization term is solved to obtain the anchor-cluster mapping matrix and the cluster indicator matrix. The cluster index matrix is ​​used to extract the intra-cluster sum of squares, which reflects the intra-cluster compactness, and the cluster average distance, which reflects the inter-cluster separation, is obtained based on the cluster index matrix. The Sigmoid function is used to normalize the ratio of intra-cluster sum of squares to inter-cluster average distance, and the heterogeneity coefficient is output.

[0010] As a preferred embodiment of the distributed data privacy-preserving grouped knowledge mining method based on heterogeneous anchor tensor graphs described in this invention, the method comprises the following steps: On the server side, based on the anchor-cluster mapping matrix uploaded by all clients, a global anchor tensor is constructed and consistency-diversity joint modeling is performed, outputting the consistent sub-tensor and diversity sub-tensor after joint modeling optimization. Based on the anchor-cluster mapping matrix of all clients, a global anchor tensor is constructed by stacking along the third dimension; The global anchor tensor is decomposed into a consistency subtensor and a diversity subtensor. A contrastive loss function of the form of InfoNCE is applied to the consistency subtensor to enhance semantic consistency. Apply a correlation penalty term based on orthogonality constraints to the diversity subtensor to ensure diversity, and output the optimized consistency subtensor and diversity subtensor.

[0011] As a preferred embodiment of the distributed data privacy-preserving grouped knowledge mining method based on heterogeneous anchor tensor graphs described in this invention, the method includes the following steps: On the server side, heterogeneity-aware federated aggregation is performed using the heterogeneity coefficients of each client and the consistency and diversity subtensors to generate and output a global anchor tensor. Using the heterogeneity coefficient, the aggregation weight corresponding to each client is calculated using the Softmax negative correlation form. The consistency subtensor and the diversity subtensor are added together to reconstruct a global anchor tensor containing information about all clients. Extract the anchor-cluster mapping matrix corresponding to each client from the reconstructed global anchor tensor, and use the aggregation weight to perform a weighted summation of the anchor-cluster mapping matrices of each client extracted from the global anchor tensor; The result of the weighted summation is defined as heterogeneity-aware federated aggregation, and the output is a global anchor tensor.

[0012] As a preferred embodiment of the distributed data privacy protection grouped knowledge mining method based on heterogeneous anchor tensor graphs described in this invention, the method includes the following steps: generating global guidance information based on the global anchor tensor and distributing it to each client for iterative optimization. Based on the global anchor tensor, a global anchor prototype matrix is ​​generated through singular value decomposition or weighted average operation. Based on the global anchor prototype matrix, a masking operation or dimension adaptation is performed for the view configuration of different clients to generate a customized global guidance matrix. A customized global guidance matrix is ​​sent to the corresponding client to update the model.

[0013] As a preferred embodiment of the distributed data privacy-preserving grouped knowledge mining method based on heterogeneous anchor tensor graphs described in this invention, the final global sample clustering labels are output after satisfying the preset convergence conditions, including the following steps: After satisfying the preset convergence condition of the loss function, the spectral clustering algorithm is executed based on the global graph structure corresponding to the consistency subtensor. The spectral clustering algorithm processes the global graph structure corresponding to the consistency subtensor to generate the clustering assignment results of the samples. The clustering assignment results of the samples are mapped to the final global sample clustering labels, and the final global sample clustering labels are output as the final result of grouped knowledge mining.

[0014] In a second aspect, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the distributed data privacy-preserving grouped knowledge mining method based on heterogeneous anchor tensor graphs as described in the first aspect of the present invention.

[0015] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the distributed data privacy-preserving grouped knowledge mining method based on heterogeneous anchor tensor graphs as described in the first aspect of the present invention.

[0016] The beneficial effects of this invention are as follows: by combining physical and virtual anchor points, the problems of missing views and heterogeneity are solved; server-side tensor construction and consistency-diversity joint modeling are achieved; and through tensor decomposition and joint optimization, deep decoupling and synergistic enhancement of global consensus and local specific information are realized. Together, these factors enable the method to improve the accuracy of high-quality grouping knowledge mining of distributed heterogeneous data under the premise of strictly protecting data privacy. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the 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.

[0018] Figure 1This is a flowchart of a distributed data privacy-preserving grouped knowledge mining method based on heterogeneous anchor tensor graphs.

[0019] Figure 2 A schematic diagram illustrating the principles of the client-side hybrid anchor point construction, incomplete view completion, and one-step grouping (clustering) modules.

[0020] Figure 3 A schematic diagram of the structure for constructing, decomposing, and co-modeling consistency and diversity of server-side anchor tensors. Detailed Implementation

[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0022] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0023] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0024] Reference Figures 1-3 This is one embodiment of the present invention, which provides a distributed data privacy-preserving grouped knowledge mining method based on heterogeneous anchor tensor graphs, including the following steps: S1. Each client holds the original multi-view data that is missing from the sample-level view.

[0025] S1.1 For each existing view of each client, a clustering algorithm is used to select M of the most representative sample centers as physical anchor points.

[0026] Furthermore, the specific implementation of using clustering algorithms to select M most representative sample centers as physical anchors is as follows: On each client's local machine, for each view it owns, firstly, all sample data under that view are standardized so that the mean of each feature is zero and the variance is one; then, the K-means clustering algorithm is applied, by randomly initializing M cluster centers, and then iteratively executing the two steps of assigning each sample to the nearest cluster center and re-acquiring each cluster center according to the assignment results, until the movement distance of the cluster centers is less than a very small threshold or the maximum number of iterations is reached; finally, the M cluster center vectors obtained after convergence are the physical anchors of that view. The number M of these physical anchors is a preset hyperparameter, the value of which is much smaller than the total number of samples under that view.

[0027] S1.2. Use the cross-view mapping function to generate virtual anchor points for the corresponding missing views. The training objective of the cross-view mapping function is to minimize the sum of the reconstruction error of the view alignment samples and the regularization term to prevent overfitting.

[0028] Furthermore, a cross-view mapping function needs to be trained. The training process uses a set of samples that simultaneously possess both views u and v as view-aligned samples. The cross-view mapping function can be a simple linear transformation matrix or a multilayer perceptron. The training objective is to minimize a loss function consisting of two parts: the first part is the sum of squares of the Euclidean distances between the true feature vectors of the view-aligned samples on view v and the vectors predicted from the features of view u through the cross-view mapping function; the second part is the square of the Frobenius norm of the cross-view mapping function parameter matrix multiplied by a small positive coefficient as a regularization term to prevent overfitting. This total loss function is optimized using the gradient descent algorithm. After training, the physical anchor point of view u is input into the trained cross-view mapping function, and its output is the virtual anchor point of view v.

[0029] S1.3 By splicing physical anchor points and virtual anchor points, a unified set of hybrid anchor points is formed for each client dimension.

[0030] Furthermore, the physical anchors corresponding to all existing views are collected. Then, for each view missing from the client, a corresponding virtual anchor is generated using the cross-view mapping method described above. All these physical and virtual anchors are then concatenated according to a predefined global view order. Specifically, assuming the global view space contains V views, and the anchor representation of each view is a matrix M multiplied by its feature dimensions, the concatenated set of mixed anchors will be a matrix of size M multiplied by the sum of the feature dimensions of all views. This operation ensures that clients with different view configurations have a logically complete and dimensionally aligned set of mixed anchors.

[0031] S2. Based on the original data missing from the local sample-level view, each client constructs its own local hybrid anchor point graph, generating a sparse hybrid anchor point graph for each client.

[0032] S2.1. Based on the mixed anchor point set, obtain the Gaussian kernel similarity between local samples and anchor points, and construct a sample-anchor point bipartite graph.

[0033] Furthermore, the feature vectors of all local samples and the feature vector of each anchor in the mixed anchor set are obtained. Then, the Euclidean distance between each local sample and each mixed anchor is calculated. Then, a Gaussian kernel function is applied to convert the distance into similarity. The Gaussian kernel function is in the form of a negative quadratic exponent, where the bandwidth parameter controls the rate at which the similarity decays with distance. For each sample, the similarity calculated with all anchors is summed and normalized so that the sum of the connection weights between each sample and all anchors is one. Finally, a matrix is ​​constructed, where the rows correspond to all local samples and the columns correspond to all mixed anchors. Each element in the matrix is ​​the normalized Gaussian kernel similarity between the corresponding sample and the anchor. This matrix is ​​the sample-anchor bipartite graph.

[0034] S2.2 Perform a truncation operation on the sample-anchor bipartite graph, retaining only the connection weights between each sample and its first s nearest neighbor anchors, and output the sparse mixed anchor graph for each client.

[0035] Furthermore, for each row in the sample-anchor bipartite graph, i.e., the row vector corresponding to each sample, find the top s elements with the highest similarity values ​​in the row vector, where s is a preset positive integer much smaller than the total number of anchors M; then set the values ​​of all other elements in the row vector except for these top s maximum values ​​to zero; after repeating this operation for all sample rows, the originally dense sample-anchor bipartite graph is transformed into a sparse matrix with most elements being zero; this sparse matrix obtained after the truncation operation is the sparse mixed anchor graph for each client, reducing storage requirements and the complexity of subsequent calculations.

[0036] S3. On each client side, using the sparse hybrid anchor point graph as input, perform grouped knowledge mining to cluster and measure client heterogeneity, and output the anchor point-cluster mapping matrix and heterogeneity coefficient of each client.

[0037] S3.1. Using the sparse mixed anchor graph as input, construct an objective function that minimizes the difference between the product of the sparse mixed anchor graph and the cluster indicator matrix and the anchor-cluster mapping matrix.

[0038] Furthermore, the cluster indicator matrix is ​​defined as a non-negative matrix with the number of rows equal to the number of local samples and the number of columns equal to the preset number of clusters. Ideally, each row has only one element and the rest are zero to represent the hard clustering assignment of the samples. At the same time, the anchor-cluster mapping matrix is ​​defined as a real matrix with the number of rows equal to the number of anchors and the number of columns equal to the number of clusters. The core of the objective function is to calculate the square of the Frobenius norm of the difference between the sparse hybrid anchor graph and the product of the cluster indicator matrix and the anchor-cluster mapping matrix. The norm is calculated as the square root of the sum of the squares of all elements of the matrix. Minimizing this difference aims to find a set of anchor-cluster mapping relationships that best approximates the true cluster indicator matrix after a linear transformation of the sparse hybrid anchor graph.

[0039] S3.2. Constrain the row sparsity of the anchor-cluster mapping matrix by using the Lp,q grouping sparse regularization term in the objective function, solve the objective function with the Lp,q grouping sparse regularization term, and obtain the anchor-cluster mapping matrix and the cluster indicator matrix.

[0040] Furthermore, an Lp,q grouping sparsity regularization term is added to the objective function. This term first calculates the Lp norm of each row vector of the anchor-cluster mapping matrix, then combines the Lp norms of all rows into a vector and calculates the Lq norm of this vector, where p and q are preset norm parameters. The purpose of adding this regularization term is to make all elements of the entire row of the anchor-cluster mapping matrix approach zero simultaneously, thereby automatically selecting anchors that make significant contributions to clustering. The solution of the entire objective function can adopt an alternating optimization strategy, that is, fixing the anchor-cluster mapping matrix and updating the cluster indicator matrix, then fixing the cluster indicator matrix and updating the anchor-cluster mapping matrix, iterating until convergence, and finally obtaining the optimized anchor-cluster mapping matrix and cluster indicator matrix.

[0041] S3.3 Extract the intra-cluster sum of squares reflecting intra-cluster compactness based on the cluster indicator matrix, and obtain the inter-cluster average distance reflecting inter-cluster separation based on the cluster indicator matrix.

[0042] Furthermore, the cluster label assigned to each sample is determined based on the cluster indicator matrix, and the mean of the feature vectors of all samples in each cluster is used as the cluster center; the intra-cluster sum of squares is calculated for each cluster by summing the squares of the Euclidean distances of all samples in that cluster to its cluster center, and the intra-cluster sums of squares of all clusters are added together to obtain the total value; the Euclidean distances between the cluster centers of all different cluster pairs are calculated, and then the arithmetic mean of these distances is used as the inter-cluster average distance.

[0043] S3.4. Use the Sigmoid function to normalize the ratio of the intra-cluster sum of squares to the inter-cluster average distance and output the heterogeneity coefficient.

[0044] Furthermore, the ratio of the sum of squares within clusters to the average distance between clusters is obtained. This ratio reflects the clustering quality; the smaller the ratio, the better the clustering effect. Then, this ratio is multiplied by a scaling factor and input into the Sigmoid function. The Sigmoid function maps the input value to an output between zero and one. Finally, the output value of the Sigmoid function is the normalized heterogeneity coefficient. The closer the coefficient is to one, the stronger the heterogeneity of the client data and the worse the clustering quality. The closer the coefficient is to zero, the weaker the heterogeneity and the better the clustering quality.

[0045] S4. On the server side, based on the anchor-cluster mapping matrix uploaded by all clients, construct a global anchor tensor and perform consistency-diversity joint modeling, outputting the consistency sub-tensor and diversity sub-tensor after joint modeling optimization.

[0046] S4.1 Construct a global anchor tensor by stacking along the third dimension based on the anchor-cluster mapping matrix of all clients.

[0047] Furthermore, the anchor-cluster mapping matrices uploaded from all clients are collected on the federated server side. Each anchor-cluster mapping matrix is ​​a two-dimensional matrix, with its row dimension corresponding to the number of anchors M and its column dimension corresponding to the number of clusters C. These two-dimensional matrices are then treated as slices of a tensor and stacked along the newly added third dimension, i.e., the client dimension. Assuming that K clients participate, the global anchor tensor formed after stacking becomes a third-order tensor with its three dimensions of size M, C, and K, respectively. This global anchor tensor integrates the local knowledge of all clients regarding the relationship between anchors and clusters, providing a unified high-dimensional data representation for subsequent joint modeling.

[0048] S4.2 Decompose the global anchor tensor into a consistency subtensor and a diversity subtensor, and apply a contrastive loss function of the form InfoNCE to the consistency subtensor to enhance semantic consistency.

[0049] Furthermore, tensor decomposition is employed to decompose the global anchor tensor into two parts: a consistency sub-tensor capturing the common clustering structure shared by all clients, and a diversity sub-tensor capturing the unique information of each client. On the consistency sub-tensor, an InfoNCE-style contrastive loss function is applied to the feature vectors of the same anchor point across different client slices. The cosine similarity between these feature vectors and the feature vectors of other anchor points across different clients is calculated as a positive sample pair. Simultaneously, the cosine similarity between the feature vector of the same anchor point and the feature vectors of other anchor points across different clients is calculated as a negative sample pair. The goal of the loss function is to maximize the logarithm of the ratio of the similarity of positive sample pairs to the sum of the similarities of negative sample pairs, thereby bringing the representations of the same anchor point closer across different clients and pushing the representations of different anchor points further apart, thus enhancing global semantic consistency.

[0050] S4.3 Apply a correlation penalty term based on orthogonality constraints to the diversity subtensor to ensure diversity, and output the optimized consistency subtensor and diversity subtensor.

[0051] Furthermore, an orthogonality constraint is imposed on the diversity subtensor. This constraint is achieved by calculating the product of the matrix expansion of the diversity subtensor under its specific pattern and its transpose, and making it as close as possible to the identity matrix. This correlation penalty term based on the orthogonality constraint is specifically calculated as the square of the Frobenius norm of the difference between the above product matrix and the identity matrix. Minimizing this penalty term can force the diversity information vectors from different client sources to be as orthogonal as possible, thereby reducing redundancy and ensuring that the diversity subtensor truly captures complementary and non-repetitive unique information. Finally, the consistent subtensor and the diversity subtensor, after joint modeling by consistency optimization and diversity constraints, are output.

[0052] S5. On the server side, using the heterogeneity coefficients of each client and the consistency subtensor and diversity subtensor, heterogeneity-aware federated aggregation is performed to generate and output the global anchor tensor.

[0053] S5.1. Using the heterogeneity coefficient, the aggregation weight corresponding to each client is calculated in the form of Softmax negative correlation. The consistency subtensor and the diversity subtensor are added together to reconstruct a global anchor tensor containing information about all clients.

[0054] Furthermore, for each client, its heterogeneity coefficient is taken and the exponential function value of the negative value of the coefficient is calculated. Then, the exponential function values ​​of the negative values ​​of the heterogeneity coefficients of all clients are summed. Finally, the exponential function value of the negative value of the heterogeneity coefficient of a single client is divided by the above sum to obtain the aggregate weight corresponding to that client. This calculation process ensures that the sum of all weights is one and that the weights are negatively correlated with the heterogeneity coefficient. Next, the consistency subtensor and the diversity subtensor after joint modeling optimization are added element-wise. Since the consistency subtensor and the diversity subtensor are decomposed from the same global anchor tensor and have the same dimension, the addition operation reconstructs a global anchor tensor that integrates consistency information and diversity information and contains information from all clients.

[0055] The aggregate weight expression is: ; in, For the first The weight of each client in the federated aggregation process For the first Heterogeneity coefficient of each client, Total number of clients For the first Heterogeneity coefficients for each client For client indexing.

[0056] S5.2 Extract the anchor-cluster mapping matrix corresponding to each client from the reconstructed global anchor tensor, and use the aggregation weight to perform a weighted summation of the anchor-cluster mapping matrices of each client extracted from the global anchor tensor.

[0057] Furthermore, the reconstructed global anchor tensor is sliced ​​along the third dimension, i.e., the client dimension, with each slice corresponding to a client's anchor-cluster mapping matrix. Then, for each client, its corresponding aggregation weight is multiplied by the anchor-cluster mapping matrix extracted from the reconstructed global anchor tensor to obtain a weighted anchor-cluster mapping matrix. Finally, these weighted anchor-cluster mapping matrices for all clients are summed element-wise to obtain a comprehensive weighted sum matrix.

[0058] S5.3 Define the weighted summation result as heterogeneity-aware federated aggregation and output it as a global anchor tensor.

[0059] Furthermore, the weighted summation matrix is ​​defined as a new round of global anchor tensor. This global anchor tensor is the final output of the heterogeneity-aware federated aggregation process. It integrates the knowledge of all clients and reflects the emphasis on high-quality client knowledge through aggregation weights.

[0060] S6. Based on the global anchor tensor, generate global guidance information and send it to each client for iterative optimization.

[0061] S6.1 Based on the global anchor tensor, a global anchor prototype matrix is ​​generated through singular value decomposition or weighted average operation. Based on the global anchor prototype matrix, a masking operation or dimension adaptation is performed for the view configuration of different clients to generate a customized global guidance matrix.

[0062] Furthermore, singular value decomposition is performed on the global anchor tensor to extract its principal component eigenvectors, or a weighted average is directly calculated on the client dimension to obtain a global anchor prototype matrix. Then, the actual view configuration of each client is analyzed. For single-view clients, a mask matrix is ​​applied to the global anchor prototype matrix. This mask matrix only retains the columns corresponding to the client's existing views and sets the columns corresponding to missing views to zero. For multi-view clients, column dimensional clipping or splicing may be required to match the input dimensions of their local model. Through this masking operation or dimension adaptation, a customized global guidance matrix that perfectly matches the local view configuration is generated for each client.

[0063] S6.2. Send the customized global guidance matrix to the corresponding client to update the model.

[0064] Furthermore, a customized global guidance matrix generated specifically for each client is sent to the corresponding client via a secure communication channel. Upon receiving the customized global guidance matrix, the client integrates it as a supervision signal into its local objective function. This typically involves adding a regularization term to the local clustering loss function, which measures the difference between the local anchor-cluster mapping matrix and the received customized global guidance matrix. The client then uses optimization algorithms such as gradient descent to minimize this local objective function that incorporates global guidance, thereby updating the local anchor-cluster mapping matrix and achieving the distillation of global knowledge into the local model.

[0065] S7. After the preset convergence condition is met, output the final global sample cluster label.

[0066] S7.1 After satisfying the preset convergence condition of the loss function, the spectral clustering algorithm is executed based on the global graph structure corresponding to the consistency sub-tensor. The spectral clustering algorithm processes the global graph structure corresponding to the consistency sub-tensor to generate the clustering assignment results of the samples.

[0067] Furthermore, it checks whether the change in the global total loss function between two consecutive iterations is less than a preset minimum positive threshold. If it is satisfied, it is considered converged. Then, a global graph structure is constructed based on the consistency sub-tensor output in step S4. The nodes of this graph structure represent all samples, and the edge weights can be determined by the inter-sample correlation derived from the anchor-cluster relationship implied by the consistency sub-tensor. Next, a spectral clustering algorithm is performed on this global graph structure. The spectral clustering algorithm first calculates the Laplacian matrix of the graph, then performs eigenvalue decomposition on the Laplacian matrix and selects the eigenvectors corresponding to the top C smallest eigenvalues ​​to form a feature matrix. Finally, K-means clustering is performed on the row vectors of this feature matrix. The result of K-means clustering is the cluster label assigned to each sample, generating the clustering assignment results of the samples.

[0068] S7.2 Map the clustering assignment results of the samples to the final global sample clustering labels, and output the final global sample clustering labels as the final result of grouped knowledge mining.

[0069] Furthermore, the clustering assignment results of the samples output by the spectral clustering algorithm, i.e., the cluster index number corresponding to each sample, are directly used as the final global sample clustering label for that sample. These global sample clustering labels are a vector or list containing all samples, with each sample having a cluster number. Outputting the final global sample clustering labels marks the completion of the entire distributed data privacy-preserving grouped knowledge mining process based on heterogeneous anchor tensor graphs, and this result is the final grouped knowledge mining result.

[0070] This embodiment also provides a computer device applicable to the distributed data privacy-preserving grouped knowledge mining method based on heterogeneous anchor tensor graphs, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the distributed data privacy-preserving grouped knowledge mining method based on heterogeneous anchor tensor graphs as proposed in the above embodiment.

[0071] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0072] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the distributed data privacy-preserving grouped knowledge mining method based on heterogeneous anchor tensor graphs as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0073] In summary, this invention solves the problems of missing views and heterogeneity by combining physical and virtual anchor points. Server-side tensor construction and consistency-diversity joint modeling, along with tensor decomposition and joint optimization, achieve deep decoupling and synergistic enhancement of global consensus and local specific information. Together, these methods improve the accuracy of high-quality grouping knowledge mining of distributed heterogeneous data while strictly protecting data privacy.

[0074] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A distributed data privacy-preserving grouped knowledge mining method based on heterogeneous anchor tensor graphs, characterized in that: include, The method, applicable to a distributed system comprising a server and multiple clients, wherein each client holds original multi-view data with missing sample-level views, includes: Based on the original data missing from the local sample-level view, each client constructs its own local hybrid anchor point map, generating a sparse hybrid anchor point map for each client. On each client side, using the sparse hybrid anchor point graph as input, grouped knowledge mining is performed to cluster and measure client heterogeneity, outputting the anchor point-cluster mapping matrix and heterogeneity coefficient for each client; On the server side, based on the anchor-cluster mapping matrix uploaded by all clients, a global anchor tensor is constructed and consistency-diversity joint modeling is performed, outputting the consistency sub-tensor and diversity sub-tensor after joint modeling optimization; On the server side, heterogeneity coefficients of each client and the consistency subtensor and diversity subtensor are used to perform heterogeneity-aware federated aggregation, generate and output a global anchor tensor. Based on the global anchor tensor, global guidance information is generated and sent to each client for iterative optimization. After the preset convergence condition is met, the final global sample clustering label is output.

2. The distributed data privacy-preserving grouped knowledge mining method based on heterogeneous anchor tensor graphs as described in claim 1, characterized in that: Each client holds the original multi-view data that is missing at the sample level, including the following steps: For each existing view of each client, a clustering algorithm is used to select M of the most representative sample centers as physical anchor points; The cross-view mapping function generates virtual anchor points for the corresponding missing views. The training objective of the cross-view mapping function is to minimize the sum of the reconstruction error of the view alignment samples and the regularization term to prevent overfitting. By combining physical anchors and virtual anchors, a unified set of hybrid anchors is formed across all client dimensions.

3. The distributed data privacy-preserving grouped knowledge mining method based on heterogeneous anchor tensor graphs as described in claim 2, characterized in that: Based on the raw data lacking local sample-level views, each client constructs its own local hybrid anchor graph, generating a sparse hybrid anchor graph for each client. This process includes the following steps: Based on the hybrid anchor set, the Gaussian kernel similarity between local samples and anchors is obtained, and a sample-anchor bipartite graph is constructed. Perform a truncation operation on the sample-anchor bipartite graph, retaining only the connection weights between each sample and its top s nearest neighbor anchors, and output a sparse mixed anchor graph for each client.

4. The distributed data privacy-preserving grouped knowledge mining method based on heterogeneous anchor tensor graphs as described in claim 3, characterized in that: On each client side, using the sparse hybrid anchor graph as input, grouped knowledge mining is performed to cluster and measure client heterogeneity, outputting the anchor-cluster mapping matrix and heterogeneity coefficient for each client, including the following steps: Using a sparse hybrid anchor graph as input, we construct an objective function that minimizes the difference between the product of the sparse hybrid anchor graph and the cluster indicator matrix and the anchor-cluster mapping matrix. In the objective function, the row sparsity of the anchor-cluster mapping matrix is ​​constrained by the Lp,q grouped sparse regularization term. The objective function with the Lp,q grouped sparse regularization term is solved to obtain the anchor-cluster mapping matrix and the cluster indicator matrix. The cluster index matrix is ​​used to extract the intra-cluster sum of squares, which reflects the intra-cluster compactness, and the cluster average distance, which reflects the inter-cluster separation, is obtained based on the cluster index matrix. The Sigmoid function is used to normalize the ratio of intra-cluster sum of squares to inter-cluster average distance, and the heterogeneity coefficient is output.

5. The distributed data privacy-preserving grouped knowledge mining method based on heterogeneous anchor tensor graphs as described in claim 4, characterized in that: On the server side, based on the anchor-cluster mapping matrix uploaded by all clients, a global anchor tensor is constructed and consistency-diversity joint modeling is performed. The output is a consistent sub-tensor and a diversity sub-tensor optimized by joint modeling, including the following steps: Based on the anchor-cluster mapping matrix of all clients, a global anchor tensor is constructed by stacking along the third dimension; The global anchor tensor is decomposed into a consistency subtensor and a diversity subtensor. A contrastive loss function of the form of InfoNCE is applied to the consistency subtensor to enhance semantic consistency. Apply a correlation penalty term based on orthogonality constraints to the diversity subtensor to ensure diversity, and output the optimized consistency subtensor and diversity subtensor.

6. The distributed data privacy-preserving grouped knowledge mining method based on heterogeneous anchor tensor graphs as described in claim 5, characterized in that: On the server side, using the heterogeneity coefficients of each client and the consistency and diversity subtensors, heterogeneity-aware federated aggregation is performed to generate and output a global anchor tensor, including the following steps: Using the heterogeneity coefficient, the aggregation weight corresponding to each client is calculated using the Softmax negative correlation form. The consistency subtensor and the diversity subtensor are added together to reconstruct a global anchor tensor containing information about all clients. Extract the anchor-cluster mapping matrix corresponding to each client from the reconstructed global anchor tensor, and use the aggregation weight to perform a weighted summation of the anchor-cluster mapping matrices of each client extracted from the global anchor tensor; The result of the weighted summation is defined as heterogeneity-aware federated aggregation, and the output is a global anchor tensor.

7. The distributed data privacy-preserving grouped knowledge mining method based on heterogeneous anchor tensor graphs as described in claim 6, characterized in that: Based on the global anchor tensor, global guidance information is generated and distributed to each client for iterative optimization, including the following steps: Based on the global anchor tensor, a global anchor prototype matrix is ​​generated through singular value decomposition or weighted average operation. Based on the global anchor prototype matrix, a masking operation or dimension adaptation is performed for the view configuration of different clients to generate a customized global guidance matrix. A customized global guidance matrix is ​​sent to the corresponding client to update the model.

8. The distributed data privacy-preserving grouped knowledge mining method based on heterogeneous anchor tensor graphs as described in claim 7, characterized in that, After satisfying the preset convergence conditions, the final global sample cluster labels are output, including the following steps: After satisfying the preset convergence condition of the loss function, the spectral clustering algorithm is executed based on the global graph structure corresponding to the consistency subtensor. The spectral clustering algorithm processes the global graph structure corresponding to the consistency subtensor to generate the clustering assignment results of the samples. The clustering assignment results of the samples are mapped to the final global sample clustering labels, and the final global sample clustering labels are output as the final result of grouped knowledge mining.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the distributed data privacy-preserving grouped knowledge mining method based on heterogeneous anchor tensor graphs as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the distributed data privacy-preserving grouped knowledge mining method based on heterogeneous anchor tensor graphs as described in any one of claims 1 to 8.

Images