Privacy protection based customer management data hierarchical classification method

Abstract

This invention discloses a privacy-preserving method for classifying and grading customer management data, specifically relating to the fields of data processing and information security. First, it establishes a mapping relationship between sensitive attributes and observable features, and weakens the impact of sensitive attributes on features through path-level decoupling. Then, it performs subspace projection on the decoupled features, extracts distribution perturbation features, and merges them with the original features to generate an implicit risk representation tensor. A tensor decomposition method with path consistency constraints is introduced to extract consistency risk factors across subspaces. Further, a graded decision path tree is constructed based on these risk factors, and the classification process is corrected through the constraint of unidentifiable sensitive attributes. Privacy leakage metrics are generated through counterfactual reconstruction analysis, and the decision path tree is structurally backtracked and optimized to obtain the final graded classification result. This invention effectively reduces the risk of sensitive attribute leakage while ensuring classification accuracy, thus improving the privacy protection capability of the data grading process.

Description

Technical Field

[0001] This invention relates to the fields of data processing and information security technology, specifically to a method for classifying and categorizing customer management data based on privacy protection. Background Technology

[0002] In cross-domain customer management systems (such as scenarios where banks and third-party risk control platforms jointly model data), data needs to flow between multiple entities and be classified and categorized. However, existing technologies typically model directly based on observable features or perform overall perturbation through differential privacy, which cannot solve the problem of "the classification boundary itself leaking privacy structures." Especially in extremely small, highly sensitive subgroups (such as combinations of specific professions and abnormal transactions), even if the original data has been desensitized, potential sensitive attributes can still be inferred through the coupling relationship between classification labels and feature distribution, resulting in "classification back-inference leakage." In addition, existing methods lack constraints on the classification process path, leading to the gradual accumulation of privacy risks in multiple rounds of dynamic updates, which are difficult to detect and suppress. Summary of the Invention

[0003] The purpose of this invention is to provide a privacy-preserving method for classifying and categorizing customer management data to address the shortcomings of the prior art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a privacy-protected customer management data hierarchical classification method, comprising: S100, Obtain customer multi-source data set D, and construct sensitive attribute-observable feature mapping relationship matrix M on D; S200, based on M to D, performs path-level decoupling processing to generate a decoupled feature set D1; S300, perform subspace projection on D1 to construct a set of several feature subspaces, and calculate the local density change rate in each subspace to form the distribution perturbation description vector V; S400, which merges V and D1 to generate an implicit risk representation tensor, and extracts the consistent risk factor R across subspaces through tensor decomposition; S500, based on R, constructs a hierarchical decision path tree W, and introduces unidentifiable constraints of sensitive attributes during the path generation process, suppresses and corrects the discrimination gain of each branch, and obtains the initial hierarchical result L; S600 performs counterfactual reconstruction analysis on L, calculates the stability difference of classification results under different sensitive attribute values, and generates a privacy leakage metric E; S700, according to E, perform structural backtracking and reconstruction on the hierarchical decision path tree W, adjust the division strategy of the corresponding branches and update the hierarchical results to obtain the final data hierarchical classification results.

[0005] Preferably, the process of constructing the sensitive attribute-observable feature mapping relationship matrix M for D includes the following steps: S110, perform field semantic parsing and standardization on the customer's multi-source data set D, label each field with sensitive attributes based on the preset sensitive attribute identification rules, form a sensitive attribute identifier set P, and establish an initial association set between each field and the sensitive attributes; S120, Based on the initial association set, the distribution difference of each observable feature under different sensitive attribute values ​​is calculated using the conditional probability estimation method, and a set of feature conditional response functions is generated; S130, Based on the set of feature conditional response functions, construct a measure of the influence intensity of the sensitive attribute on each observable feature, and form a corresponding set of influence weights; S140, The set of influence weights is expressed in matrix form to obtain the sensitive attribute-observable feature mapping relationship matrix M.

[0006] Preferably, the process of generating the decoupled feature set D1 includes the following steps: S210, Based on the sensitive attribute-observable feature mapping relationship matrix M, extract the set of sensitive influence paths corresponding to each observable feature, and construct a feature-path association index table to identify the sensitive attribute propagation path associated with each observable feature; S220, according to the feature-path association index table, perform path contribution decomposition on each observable feature in the customer multi-source data set D, and calculate the contribution weight value of each sensitive attribute to the observable feature along the corresponding propagation path; S230, based on the contribution weight value, perform path-level suppression transformation on each observable feature, and generate an intermediate decoupled feature set by weakening the feature components corresponding to high contribution paths and retaining the feature components corresponding to low contribution paths. S240, perform consistency reconstruction processing on the intermediate decoupled feature set, so that each observable feature reduces its dependence on sensitive attributes while maintaining the original data distribution structure, thereby obtaining the decoupled feature set D1.

[0007] Preferably, the process of calculating the local density change rate in each subspace to form the distribution perturbation description vector V includes the following steps: S310, perform feature correlation calculation on the decoupled feature set D1, construct a feature association matrix based on the Pearson correlation coefficient between features, and perform feature clustering according to the correlation coefficient threshold to generate several initial feature subspace sets; S320, perform orthogonal projection transformation within each initial feature subspace set, construct the covariance matrix and perform eigenvalue decomposition to extract the principal direction vector, form a low-dimensional subspace representation, and thus obtain a stable feature subspace set; S330: In each stable feature subspace, a neighborhood range is constructed with the sample point as the center. The local sample density is calculated based on the preset neighborhood radius, and the density difference between adjacent regions is further calculated to obtain the local density change rate. S340, the local density change rates in each stable feature subspace are spliced ​​together according to the subspace order to construct a distribution perturbation description vector V with a unified dimension.

[0008] Preferably, the process of extracting the consistency risk factor R across subspaces through tensor decomposition includes the following steps: S410, the decoupled feature set D1 and the distribution perturbation description vector V are aligned and expanded according to the sample dimension, and a third-order implicit risk representation tensor is constructed by outer product operation, wherein each dimension corresponds to the sample, feature and perturbation description respectively; S420: Construct a path association graph based on the obtained set of sensitive impact paths, and calculate the path similarity weights between each node to form a path consistency constraint matrix; S430, When decomposing the implicit risk representation tensor, a path consistency constraint matrix is ​​introduced as a regularization term. The decomposition factor matrix is ​​solved by minimizing the objective function of the sum of tensor reconstruction error and path consistency deviation. S440, Based on the common components in the decomposition factor matrix that maintain stable responses in each feature subspace, extract the consistency risk factor R across subspaces.

[0009] Preferably, the process of constructing a hierarchical decision path tree W based on R includes the following steps: S510: Construct a candidate splitting feature set based on the consistency risk factor R, determine the initial splitting node and corresponding splitting threshold by calculating the information gain value of each risk factor on the sample classification result, and generate the initial decision structure. S520, based on the initial decision structure, introduce the sensitive attribute indistinguishability constraint for each candidate branch, and construct the sensitive leakage measurement function by calculating the difference in the distribution of sensitive attributes after branch division; S530, the information gain value of each branch is suppressed and corrected according to the sensitive leakage metric function to obtain the corrected discrimination gain value, and the optimal splitting path is reselected based on the corrected discrimination gain value; S540: Recursively construct a hierarchical decision path tree W according to the corrected optimal splitting path until the preset stopping condition is met, and output the corresponding initial hierarchical result L.

[0010] Preferably, the process of generating the privacy breach metric E includes the following steps: S610, based on the initial classification result L and the original sensitive attribute identifier set, construct counterfactual sample generation rules, and replace the sensitive attribute values ​​of each sample while keeping the non-sensitive features unchanged to generate the corresponding counterfactual sample set. S620, input the counterfactual sample set into the constructed hierarchical decision path tree W to obtain the classification result corresponding to each counterfactual sample and form a counterfactual classification result set; S630, based on the set of original classification results and counterfactual classification results, calculates the classification change rate of each sample under different sensitive attribute values, and constructs a classification stability measurement function; S640, Statistically summarize the classification stability measurement function to generate an overall privacy leakage measurement index E.

[0011] Preferably, the process of structural backtracking and reconstruction of the hierarchical decision path tree W based on E includes the following steps: S710, based on the privacy leakage metric E, performs risk location on each leaf node in the hierarchical decision path tree W, identifies the path branches whose stability is lower than the preset stability threshold, and marks them as branches to be backtracked. S720, For the branch to be backtracked, backtrack from the leaf node upwards layer by layer to the nearest parent node that satisfies the stability constraint, extract the partitioning features and corresponding partitioning thresholds on the path, and construct a path adjustment candidate set; S730, recalculate the corrected discriminant gain value of each candidate partitioning feature in the path adjustment candidate set, and introduce the privacy leakage metric E as a constraint term to dynamically optimize and adjust the partitioning threshold to generate a reconstructed branch structure; S740, replace the corresponding branch in the original hierarchical decision path tree W with the reconstructed branch structure, perform global consistency verification on the updated hierarchical decision path tree, and output the final data hierarchical classification result.

[0012] The technical effects and advantages provided by the present invention in the above technical solution are as follows: 1. This invention constructs a sensitive attribute-observable feature mapping matrix and introduces path-level decoupling processing, which explicitly characterizes and suppresses potential sensitive propagation paths in the original data, cutting off the implicit influence of sensitive attributes on classification features at the data source level. Furthermore, by combining the distribution perturbation description vector with the tensor fusion representation of the decoupled feature set, the data has a unified expressive ability in multiple subspaces. And by using the tensor decomposition method with path consistency constraints to extract cross-subspace stable consistency risk factors, the invention effectively reduces the identifiability of sensitive attributes in the feature space while ensuring the integrity of feature expression, and significantly improves the privacy and security level in the data classification process.

[0013] 2. This invention introduces unidentifiable constraints on sensitive attributes and a counterfactual reconstruction analysis mechanism into the hierarchical decision-making process. By dynamically suppressing and correcting the classification path and reconstructing the structure through backtracking, it ensures that the classification results remain stable and consistent under different sensitive attribute values. Simultaneously, it utilizes privacy leakage metrics to perform closed-loop optimization of the model, achieving quantitative assessment and continuous correction of privacy risks. Compared to existing technologies, this invention not only achieves more refined data hierarchical classification in complex multi-source data scenarios but also effectively suppresses the dependence of classification results on sensitive attributes, reducing the risk of reverse inference. While ensuring classification accuracy, it significantly improves overall privacy protection capabilities, exhibiting stronger practicality and security. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0015] Figure 1 This is a flowchart of the privacy-protected customer management data hierarchical classification method of the present invention.

[0016] Figure 2 This is a flowchart of the method for calculating the consistency risk factor of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] For examples, please refer to Figure 1 and Figure 2 As shown in this embodiment, the privacy-protected customer management data hierarchical classification method includes: In this embodiment, step S100 is used to obtain a multi-source customer data set D, and construct a sensitive attribute-observable feature mapping relationship matrix M on D, specifically including the following steps: In step S110, when performing field semantic parsing and standardization on the customer multi-source data set D, firstly, unified encoding processing is performed on the fields from different business sources in the data set D. Specifically, a field mapping table is established to map fields with the same semantics but different names from different data sources to unified field identifiers. Then, the data types of each field are standardized, converting numeric fields to floating-point representations, converting categorical fields to numeric vectors through one-hot encoding, and splitting time fields into timestamps and periodic features. Based on this, preset sensitive attribute identification rules are constructed. The rules are generated as follows: a historical dataset with known sensitive attribute labels is selected, the mutual information value between each field and each sensitive attribute is calculated, and the mutual information value is normalized. When the normalization result is greater than 0.6, a mapping relationship is established between the field and the corresponding sensitive attribute. Further, each field in the data set D to be processed is matched with the rule base one by one. If the match is successful, it is marked as the corresponding sensitive attribute, and finally a sensitive attribute identifier set P is formed. The one-to-many association relationship between fields and sensitive attributes is recorded to obtain the initial association set.

[0019] In step S120, based on the initial association set, when calculating the distribution difference of each observable feature under different sensitive attribute values, the data set D is first grouped according to the different values ​​of the sensitive attribute to form several subsets. For each subset, for each observable feature, a kernel density estimation method is used to construct its probability distribution function. The kernel function is selected as a Gaussian kernel function, and the bandwidth parameter is determined by cross-validation, that is, selecting the bandwidth value that minimizes the distribution fitting error from a preset candidate bandwidth set. Subsequently, the probability distribution functions of the same observable feature under different sensitive attribute values ​​are compared pairwise, and the Kullback-Leibler divergence is used for measurement. The calculation method is as follows: the logarithmic ratio of the discretized probability distribution function is calculated point by point, and a weighted sum is performed over all values. To ensure the stability of the calculation, the part with a probability value less than 0.001 is smoothed and replaced with 0.001. Finally, the distribution difference function corresponding to each observable feature is obtained, forming a set of feature conditional response functions.

[0020] In step S130, when constructing the influence intensity measure of sensitive attributes on each observable feature based on the set of feature conditional response functions, the distribution difference values ​​of each observable feature under different sensitive attribute values ​​are first summarized. Specifically, the mean Kullback-Leibler divergence of the feature under all combinations of sensitive attribute values ​​is calculated. Then, a sensitive attribute weight coefficient is introduced, which is obtained by statistically analyzing the frequency of each sensitive attribute in the dataset D and normalizing it. The mean distribution difference value and the corresponding sensitive attribute weight coefficient are weighted and summed to obtain the initial influence intensity value. Further, the initial influence intensity values ​​of all observable features are subjected to maximum-minimum normalization. The calculation method is as follows: subtract the minimum value from each value and divide by the difference between the maximum value and the minimum value, so that the result falls within the interval [0,1], and finally, the influence weight set is formed.

[0021] In step S140, when representing the set of influence weights in a matrix format, the row and column structure of the matrix is ​​first determined, where rows correspond to each sensitive attribute in the sensitive attribute set P, and columns correspond to the standardized set of observable features. Then, according to the influence weights obtained in step S130, the influence strength value between each sensitive attribute and its corresponding observable feature is filled into the corresponding position of the matrix. For sensitive attribute and observable feature combinations that do not appear in the initial association set, their influence strength value is set to 0. After filling, the matrix is ​​tested for sparsity. When the proportion of non-zero elements in the matrix is ​​less than 0.2, the non-zero elements are expanded by assigning the average weight value of adjacent features to that position to enhance the connectivity of the matrix. Finally, the sensitive attribute-observable feature mapping relationship matrix M is obtained, which is used to accurately characterize the potential influence path of sensitive attributes on observable features and provide a quantitative basis for subsequent path-level decoupling processing.

[0022] S200 performs path-level decoupling processing on M to D, generating a decoupled feature set D1.

[0023] In step S210, when extracting the set of sensitive influence paths corresponding to each observable feature based on the sensitive attribute-observable feature mapping relationship matrix M, the sensitive attribute-observable feature mapping relationship matrix M is first subjected to threshold filtering processing. Elements in the matrix that are greater than a preset path activation threshold are identified as valid connections. The path activation threshold is set to 0.3. Then, a directed weighted graph structure is constructed with the sensitive attribute as the starting node and the observable feature as the ending node. The edge weights are determined by the corresponding values ​​in the sensitive attribute-observable feature mapping relationship matrix M. Furthermore, a depth-first traversal method is used to traverse from each sensitive attribute to all observable features, recording all paths that satisfy the cumulative path weight greater than 0.5, forming a set of sensitive influence paths. On this basis, a corresponding path index number is established for each observable feature, and a feature-path association index table is generated to clarify all sensitive propagation paths corresponding to each observable feature.

[0024] In step S220, when calculating the path contribution based on the feature-path association index table, each observable feature value in the customer multi-source data set D is first decomposed and represented, specifically: the feature value is represented as a linear combination of each path contribution component; then, for each sensitive influence path, the contribution weight value of that path is calculated, the calculation method is: multiplying the weight values ​​of each edge on the path together, and multiplying by the probability of the sensitive attribute at the path starting point appearing in the data set D; further, all path contribution weights are normalized so that the sum of the path contribution weights is 1; finally, the normalized path contribution weights are allocated to the corresponding observable features to obtain the contribution weight value of each sensitive attribute to the observable feature along the corresponding propagation path.

[0025] In step S230, when performing path-level suppression transformation based on the contribution weight value, the path suppression threshold is first set to 0.4, and paths with contribution weights greater than the path suppression threshold are marked as high-sensitivity paths. Then, the values ​​of each observable feature are reconstructed, specifically: the contribution components belonging to high-sensitivity paths are multiplied by an attenuation coefficient, the attenuation coefficient being 0.2, while the contribution components belonging to low-sensitivity paths remain unchanged. After adjusting each path component, all path components are summed to obtain the feature values ​​after path-level suppression, thereby generating an intermediate decoupling feature set.

[0026] In step S240, when performing consistency reconstruction processing on the intermediate decoupling feature set, the distribution deviation between the intermediate decoupling feature set and the original customer multi-source data set D is first calculated. The distribution deviation is obtained by calculating the difference between the mean and variance of each feature. Then, a distribution compensation function is constructed to adjust each feature value. Specifically, the offset between the intermediate decoupling feature value and the original feature mean is scaled proportionally. The scaling factor is obtained by solving the objective function that minimizes the overall distribution deviation. The objective function is defined as the sum of the squares of the differences between the means and variances of all features. The scaling factor is solved through iterative optimization. The iteration termination condition is set to the change in the objective function being less than 0.001. Finally, a decoupling feature set D1 is obtained that reduces the dependence on sensitive attributes while maintaining the original data distribution structure.

[0027] S300, perform subspace projection on D1 to construct a set of several feature subspaces, and calculate the local density change rate in each subspace to form the distribution perturbation description vector V.

[0028] In step S310, when calculating the feature correlation of the decoupled feature set D1, each observable feature in the decoupled feature set D1 is first standardized. Specifically, each feature value is subtracted from its mean and then divided by its standard deviation to eliminate dimensional differences. Then, the Pearson correlation coefficient between any two observable features is calculated to form a feature correlation matrix. The Pearson correlation coefficient is calculated as the ratio of the product of the covariance of the two features and their respective standard deviations. Further, a correlation coefficient threshold of 0.5 is set, and feature pairs with an absolute correlation coefficient greater than the threshold are considered as strongly correlated features. Based on the strongly correlated feature relationship, a hierarchical clustering method is used to cluster all observable features, where the clustering distance is defined as the complement of the correlation coefficient. Finally, several initial feature subspace sets are generated, so that the features in each subspace have high correlation.

[0029] In step S320, when performing orthogonal projection transformation within each initial feature subspace set, a corresponding data submatrix is ​​first constructed for each initial feature subspace set; then, the covariance matrix of the data submatrix is ​​calculated, and eigenvalue decomposition is performed on the covariance matrix to obtain eigenvalues ​​and corresponding eigenvectors; the eigenvalues ​​are sorted from largest to smallest, and the top few eigenvectors with a cumulative contribution rate of 0.9 are selected as principal direction vectors; the original subspace data is projected onto the orthogonal basis formed by the principal direction vectors to obtain a low-dimensional representation result; in this way, the original high-dimensional related features are mapped to a stable low-dimensional feature representation, thereby forming a stable feature subspace set.

[0030] In step S330, when calculating the local density change rate in each stable feature subspace, a neighborhood range is first constructed for each sample point in each subspace. The neighborhood range is determined by calculating the Euclidean distance between the sample points, and the neighborhood radius is set to 0.3 times the average distance of all sample points. For each sample point, the number of samples in its neighborhood range is counted and divided by the neighborhood volume to obtain the local density value. Then, the nearest neighbor sample of each sample point is selected, the difference between the local density values ​​of the two is calculated, and divided by the distance between the two to obtain the local density change rate. To avoid the influence of outliers, the local density change rate is truncated, and the result exceeding the preset upper limit value of 3 is limited to 3.

[0031] In step S340, when constructing the local density change rate in each stable feature subspace into a distribution perturbation description vector V, firstly, according to the subspace numbering order, the local density change rate of all samples in each subspace is statistically analyzed, and its mean, variance, and maximum value are extracted as subspace perturbation features; then, the perturbation features of each subspace are concatenated in order to form a vector representation with a unified dimension; further, the vector is normalized so that the values ​​of each dimension fall within the interval [0,1]; finally, the distribution perturbation description vector V is obtained, which is used to comprehensively characterize the local distribution change characteristics of the decoupled feature set D1 in different subspaces.

[0032] S400 merges V and D1 to generate an implicit risk representation tensor, and extracts the consistent risk factor R across subspaces through tensor decomposition.

[0033] In step S410, when fusing the distribution perturbation description vector V with the decoupled feature set D1 to generate the implicit risk representation tensor, the distribution perturbation description vector V is first expanded in dimension to ensure consistency with the decoupled feature set D1 in sample dimension. Specifically, the distribution perturbation description vector corresponding to each sample is copied to the length of the feature dimension corresponding to that sample. Then, an outer product operation is performed on the decoupled feature set D1 and the expanded distribution perturbation description vector. For any sample, its feature vector is combined element-wise with the corresponding distribution perturbation description vector to generate a two-dimensional matrix. This operation is repeated for all samples and stacked according to the sample index to construct a third-order implicit risk representation tensor, where the first dimension represents the sample, the second dimension represents the observable feature, and the third dimension represents the distribution perturbation description dimension, thereby achieving a structured fusion representation.

[0034] In step S420, when constructing a path association graph and forming a path consistency constraint matrix based on the set of sensitive influence paths, each sensitive influence path is first regarded as a connection in the graph, and the observable features involved in the path are regarded as nodes; then, the similarity between any two paths is calculated, and the similarity is determined by the ratio of the number of shared features in the path to the total number of features in the path; further, the path similarity is mapped to the weight values ​​between nodes to construct a weighted undirected graph; on this basis, a path consistency constraint matrix is ​​generated, and its element values ​​are defined as the path similarity weights between corresponding nodes; to enhance the stability of the constraint, the path consistency constraint matrix is ​​normalized so that the sum of the elements in each row is 1.

[0035] In step S430, when introducing path consistency constraints during the decomposition of the implicit risk representation tensor, a decomposition rank parameter is first set. This rank parameter is determined through singular value pre-analysis of the implicit risk representation tensor, specifically by selecting the rank value corresponding to a cumulative contribution rate of 0.85. Then, an optimization objective function is constructed, consisting of two parts: the first part is the tensor reconstruction error, defined as the sum of squared errors between the original implicit risk representation tensor and the reconstructed tensor after decomposition; the second part is the path consistency deviation, defined as the Laplace smoothing term of the decomposed feature factor matrix under the influence of the path consistency constraint matrix. Furthermore, the objective function is iteratively solved using the alternating least squares method. In each iteration, a portion of the factor matrix is ​​fixed while the remaining factor matrices are updated. The iteration termination condition is set to a change in the objective function of less than 0.0001.

[0036] In step S440, when extracting the cross-subspace consistency risk factor R based on the decomposed factor matrix, stability analysis is first performed on each factor matrix obtained from the decomposition, specifically by calculating the response variance of each factor in different feature subspaces; then, factor components with response variances lower than a preset stability threshold (the stability threshold is set to 0.05) are selected as candidate consistency components; further, the candidate consistency components are weighted and fused, with the weights determined based on their contribution to tensor reconstruction; finally, the cross-subspace consistency risk factor R is formed, which is used to characterize the unified risk feature that maintains stable expression in different feature subspaces.

[0037] To verify the effectiveness of the method described in step S400 in practical applications, real de-identified data from a financial customer management scenario is selected as an example for illustration. The selected data contains 10,000 customer samples, each sample containing 20-dimensional decoupled features and an 8-dimensional distribution perturbation description vector.

[0038] In step S410, the 20-dimensional decoupled feature vector of each customer sample is outer-producted with the 8-dimensional distribution perturbation description vector to generate a 20×8 two-dimensional matrix. All 10,000 samples are then stacked to construct a third-order implicit risk representation tensor of size 10,000×20×8. This method fuses the original feature information with local distribution perturbation information, enabling the data to structurally possess cross-subspace expressive capabilities.

[0039] In step S420, based on the aforementioned set of sensitive influence paths, a total of 45 valid paths are extracted, and the similarity between the paths is calculated to construct a 45×45 path consistency constraint matrix. After normalization, the sum of the elements in each row is 1. This matrix is ​​used to constrain the consistency of the expression of features corresponding to different paths in the subsequent decomposition process.

[0040] In step S430, the implicit risk representation tensor is decomposed. First, the decomposition rank is determined to be 6 through singular value analysis. Then, an objective function containing tensor reconstruction error and path consistency deviation is constructed and solved using the alternating least squares method. The objective function converges to a stable state (change amount less than 0.0001) after 120 iterations. After decomposition, three factor matrices are obtained, corresponding to the sample dimension, feature dimension and perturbation dimension, respectively.

[0041] In step S440, stability analysis is performed on the feature factor matrix obtained by decomposition, the response variance of each factor in different feature subspaces is calculated, and a total of 4 factor components with response variances lower than 0.05 are selected. Furthermore, they are weighted and fused according to their contribution to tensor reconstruction to finally form a 4-dimensional consistency risk factor R.

[0042] Comparative experiments revealed that without path consistency constraints, the stability index of the classification results to changes in sensitive attributes was 0.62; however, after adopting the method of this application, the stability index increased to 0.81, while the classification accuracy decreased by only 1.8%. These results demonstrate that this application, while maintaining high classification performance, significantly reduces the dependence of the classification results on sensitive attributes, thereby effectively improving privacy protection capabilities.

[0043] S500 constructs a hierarchical decision path tree W based on R, and introduces unidentifiable constraints of sensitive attributes during the path generation process to suppress and correct the discrimination gain of each branch, thereby obtaining the initial hierarchical result L.

[0044] In step S510, when constructing a candidate splitting feature set based on the consistency risk factor R, each dimension factor in the consistency risk factor R is first used as a candidate splitting feature, and each dimension factor is sorted. Then, for each candidate splitting feature, all possible splitting thresholds within its value range are traversed. The splitting threshold is obtained by taking the midpoint of adjacent samples after arranging the feature values ​​in ascending order. Further, the information gain value corresponding to each splitting method is calculated. The information gain value is obtained by calculating the difference in information entropy between the sample sets before and after the split, where the information entropy is calculated using an entropy function with a logarithmic base of 2. Finally, the splitting feature with the largest information gain value and the corresponding splitting threshold are selected as the initial splitting node to generate the initial decision structure.

[0045] In step S520, when introducing the sensitive attribute unidentifiable constraint based on the initial decision structure, firstly, for each candidate branch, the probability distribution of sensitive attributes in each child node after the branch division is statistically analyzed; then, the difference in the sensitive attribute distribution between different child nodes is calculated, and the difference is measured by Jensen-Shannon divergence, which is calculated by weighted summation of the average distribution of the two probability distributions and the Kullback-Leibler divergence between their respective distributions; further, a sensitive leakage measurement function is constructed, and the difference is used as the function output value to characterize the degree of influence of the current branch division on the identifiability of sensitive attributes.

[0046] In step S530, when suppressing and correcting the information gain value according to the sensitive leakage metric function, firstly, a suppression coefficient function is set. The suppression coefficient function is defined as an exponential decay function of the sensitive leakage metric value, specifically expressed as a negative exponential function with the natural constant as the base. Then, the original information gain value is multiplied by the suppression coefficient to obtain the corrected discriminant gain value. When the sensitive leakage metric value is greater than the preset leakage threshold (the leakage threshold is set to 0.2), the corrected discriminant gain value of the corresponding branch is further multiplied by a penalty factor of 0.5. Finally, based on the corrected discriminant gain values ​​of all candidate branches, the optimal splitting path is reselected.

[0047] In step S540, when recursively constructing the hierarchical decision path tree W according to the corrected optimal splitting path, the optimal splitting node selected in step S530 is first used as the current node to divide the sample set; then, steps S510 to S530 are repeated for each child node to build the tree structure layer by layer; during the construction process, a stopping condition is set, which includes: the number of samples in the current node is less than a preset minimum sample threshold (the threshold is set to 1% of the total number of samples) or the corrected discriminant gain value is less than 0.01; when the stopping condition is met, the current node is marked as a leaf node, and its classification label is determined according to the risk distribution of the samples in the node; finally, the construction of the hierarchical decision path tree W is completed, and the corresponding initial classification result L is output.

[0048] S600 performs counterfactual reconstruction analysis on L, calculates the stability differences of classification results under different sensitive attribute values, and generates a privacy leakage metric E.

[0049] In step S610, when constructing the counterfactual sample generation rules based on the initial classification result L and the original sensitive attribute identifier set, the sensitive attribute values ​​of each sample are first enumerated. Specifically, for each sensitive attribute, all possible values ​​in the dataset are extracted. Then, the original sample is copied, and its sensitive attribute values ​​are replaced one by one, while keeping the values ​​of non-sensitive features in the decoupled feature set D1 unchanged, thereby generating the corresponding counterfactual sample. To ensure the rationality of the counterfactual sample, the replaced sample is constrained and screened. The constraint condition is that the Euclidean distance of the replaced sample in the consistency risk factor R space does not exceed 0.2 times the corresponding value of the original sample. Finally, the counterfactual sample set is obtained.

[0050] In step S620, when the set of counterfactual samples is input into the hierarchical decision path tree W, each counterfactual sample is first judged layer by layer according to the splitting rules of the hierarchical decision path tree W constructed in step S540. At each node, the sample is assigned to the corresponding child node according to the corresponding splitting feature and splitting threshold. After reaching the leaf node, the hierarchical label corresponding to the leaf node is read as the classification result of the counterfactual sample. The above process is repeated for all counterfactual samples to form a set of counterfactual classification results, and the classification result sequence of each sample under different sensitive attribute values ​​is recorded.

[0051] In step S630, when calculating the classification stability metric based on the original classification results and the counterfactual classification results set, the original classification result of each sample is first compared with its counterfactual classification result to count the number of times the classification result changes; then, the classification change rate is calculated, which is defined as the ratio of the number of times the classification result changes to the total number of counterfactual samples; further, a classification stability metric function is constructed to map the classification change rate to a stability index, specifically by subtracting the classification change rate from the original value to obtain the stability value; to enhance the discriminative ability, the stability value is subjected to an exponential transformation so that samples with larger changes receive lower stability scores.

[0052] In step S640, when statistically summarizing the classification stability metric function to generate the privacy leakage metric E, firstly, a weighted average of the stability values ​​of all samples is calculated. The weights are determined based on the risk level of each sample in the initial classification result L, with higher risk levels corresponding to larger weights. Then, the variance of the stability values ​​is calculated to measure the dispersion of overall stability. Further, the privacy leakage metric E is constructed, defined as a combination function of weighted average stability and stability variance, where the variance term is introduced in a negative form to reflect the impact of unstable samples. Finally, the privacy leakage metric E is obtained, which is used to quantify the dependence of the classification result on changes in sensitive attributes, providing a basis for subsequent optimization of classification results.

[0053] S700, according to E, perform structural backtracking and reconstruction on the hierarchical decision path tree W, adjust the division strategy of the corresponding branches and update the hierarchical results to obtain the final data hierarchical classification results.

[0054] In step S710, when performing risk localization on each leaf node in the hierarchical decision path tree W based on the privacy leakage metric E, the privacy leakage metric E is first decomposed into stability scores corresponding to each sample, and the stability scores of the samples contained in each leaf node are summarized to calculate the average stability value of the leaf node; then, a preset stability threshold of 0.7 is set, and the path branches corresponding to leaf nodes with average stability values ​​lower than the preset stability threshold are marked as branches to be backtracked; further, the path node sequence corresponding to each branch to be backtracked, its partitioning features, and partitioning threshold are recorded to provide basic data for subsequent structural adjustments.

[0055] In step S720, when performing path backtracking processing on the branch to be backtracked, the process starts from the leaf node corresponding to the branch to be backtracked and backtracks upwards layer by layer along the hierarchical decision path tree W until the first parent node that satisfies the stability constraint condition is reached. The stability constraint condition is that the average stability value of the sample set corresponding to the node is not lower than the preset stability threshold. During the backtracking process, the partitioning features and corresponding partitioning thresholds on the path are extracted layer by layer, and the corrected discriminant gain value of each node is recorded. Then, the extracted partitioning features and partitioning thresholds are combined to form a path adjustment candidate set for subsequent optimization selection.

[0056] In step S730, when reconstructing the partitioning strategy in the path adjustment candidate set, the modified discriminant gain value of each candidate partitioning feature is first recalculated. The calculation method is as follows: based on the original information gain, a privacy leakage metric E is introduced as a penalty term, and the information gain value is multiplied by a decay function with the privacy leakage metric as the independent variable. The decay function is defined as a linear function of minus the privacy leakage metric. Subsequently, for the value range of each candidate partitioning feature, the optimal partitioning threshold is re-determined using an interval partitioning search method. Specifically, the feature value range is divided into several candidate threshold points at equal intervals, and the corresponding modified discriminant gain value is calculated one by one. The threshold corresponding to the maximum value is selected as the optimization result. Finally, the reconstructed branch structure is generated.

[0057] In step S740, when replacing the corresponding branch in the original hierarchical decision path tree W with the reconstructed branch structure and performing global consistency verification, the optimized branch structure is first embedded into the original hierarchical decision path tree W, replacing the original branch to be backtracked; then, the updated hierarchical decision path tree W is reclassified for all samples to obtain new classification results; further, the consistency index between the classification results before and after the update is calculated, the consistency index is measured by the consistency ratio of sample classification labels, and a consistency lower limit is set to 0.85; when the consistency index is lower than the consistency lower limit, the reconstructed branch is rolled back and a suboptimal candidate partitioning strategy is reselected; finally, under the condition of satisfying the consistency constraint, the final data hierarchical classification result is output.

[0058] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A privacy protection based customer management data hierarchical classification method, characterized in that: include: S100, Obtain customer multi-source data set D, and construct sensitive attribute-observable feature mapping relationship matrix M on D; S200, based on M to D, performs path-level decoupling processing to generate a decoupled feature set D1; S300, perform subspace projection on D1 to construct a set of several feature subspaces, and calculate the local density change rate in each subspace to form the distribution perturbation description vector V; S400, fusing V and D1 to generate an implicit risk representation tensor, and extracting a consistent risk factor R across subspaces through tensor decomposition, specifically includes the following steps: S410, the decoupled feature set D1 and the distribution perturbation description vector V are aligned and expanded according to the sample dimension, and a third-order implicit risk representation tensor is constructed by outer product operation, wherein each dimension corresponds to the sample, feature and perturbation description respectively; S420: Construct a path association graph based on the obtained set of sensitive impact paths, and calculate the path similarity weights between each node to form a path consistency constraint matrix; S430, When decomposing the implicit risk representation tensor, a path consistency constraint matrix is ​​introduced as a regularization term. The decomposition factor matrix is ​​solved by minimizing the objective function of the sum of tensor reconstruction error and path consistency deviation. S440, Based on the common components in the decomposition factor matrix that maintain stable responses in each feature subspace, extract the consistency risk factor R across subspaces; S500, based on R, constructs a hierarchical decision path tree W, and introduces unidentifiable constraints of sensitive attributes during the path generation process, suppresses and corrects the discrimination gain of each branch, and obtains the initial hierarchical result L; S600, perform counterfactual reconstruction analysis on L, calculate the stability differences of classification results under different sensitive attribute values, and generate a privacy leakage metric E, specifically including the following steps: S610, based on the initial classification result L and the original sensitive attribute identifier set, construct counterfactual sample generation rules, and replace the sensitive attribute values ​​of each sample while keeping the non-sensitive features unchanged to generate the corresponding counterfactual sample set. S620, input the counterfactual sample set into the constructed hierarchical decision path tree W to obtain the classification result corresponding to each counterfactual sample and form a counterfactual classification result set; S630, based on the set of original classification results and counterfactual classification results, calculates the classification change rate of each sample under different sensitive attribute values, and constructs a classification stability measurement function; S640, Perform statistical summarization on the classification stability measurement function to generate an overall privacy leakage measurement index E; S700, according to E, perform structural backtracking and reconstruction on the hierarchical decision path tree W, adjust the division strategy of the corresponding branches and update the hierarchical results to obtain the final data hierarchical classification results.

2. The privacy-preservation based customer management data triage method of claim 1, wherein: The process of constructing the sensitive attribute-observable feature mapping relationship matrix M for D includes the following steps: S110, perform field semantic parsing and standardization on the customer's multi-source data set D, label each field with sensitive attributes based on the preset sensitive attribute identification rules, form a sensitive attribute identifier set P, and establish an initial association set between each field and the sensitive attributes; S120, Based on the initial association set, the distribution difference of each observable feature under different sensitive attribute values ​​is calculated using the conditional probability estimation method, and a set of feature conditional response functions is generated; S130, Based on the set of feature conditional response functions, construct a measure of the influence intensity of the sensitive attribute on each observable feature, and form a corresponding set of influence weights; S140, The set of influence weights is expressed in matrix form to obtain the sensitive attribute-observable feature mapping relationship matrix M.

3. The privacy-preservation based customer management data triage method of claim 2, wherein: The process of generating the decoupled feature set D1 includes the following steps: S210, Based on the sensitive attribute-observable feature mapping relationship matrix M, extract the set of sensitive influence paths corresponding to each observable feature, and construct a feature-path association index table to identify the sensitive attribute propagation path associated with each observable feature; S220, according to the feature-path association index table, perform path contribution decomposition on each observable feature in the customer multi-source data set D, and calculate the contribution weight value of each sensitive attribute to the observable feature along the corresponding propagation path; S230, based on the contribution weight value, perform path-level suppression transformation on each observable feature, and generate an intermediate decoupled feature set by weakening the feature components corresponding to high contribution paths and retaining the feature components corresponding to low contribution paths. S240, perform consistency reconstruction processing on the intermediate decoupled feature set, so that each observable feature reduces its dependence on sensitive attributes while maintaining the original data distribution structure, thereby obtaining the decoupled feature set D1.

4. The privacy-protected based customer management data triage method of claim 1, wherein: The process of calculating the local density change rate in each subspace to form the distribution perturbation description vector V includes the following steps: S310, perform feature correlation calculation on the decoupled feature set D1, construct a feature association matrix based on the Pearson correlation coefficient between features, and perform feature clustering according to the correlation coefficient threshold to generate several initial feature subspace sets; S320, perform orthogonal projection transformation within each initial feature subspace set, construct the covariance matrix and perform eigenvalue decomposition to extract the principal direction vector, form a low-dimensional subspace representation, and thus obtain a stable feature subspace set; S330: In each stable feature subspace, a neighborhood range is constructed with the sample point as the center. The local sample density is calculated based on the preset neighborhood radius, and the density difference between adjacent regions is further calculated to obtain the local density change rate. S340, the local density change rates in each stable feature subspace are spliced ​​together according to the subspace order to construct a distribution perturbation description vector V with a unified dimension.

5. The privacy-protected based customer management data triage method of claim 1, wherein: The process of constructing a hierarchical decision path tree W based on R includes the following steps: S510: Construct a candidate splitting feature set based on the consistency risk factor R, determine the initial splitting node and corresponding splitting threshold by calculating the information gain value of each risk factor on the sample classification result, and generate the initial decision structure. S520, based on the initial decision structure, introduce the sensitive attribute indistinguishability constraint for each candidate branch, and construct the sensitive leakage measurement function by calculating the difference in the distribution of sensitive attributes after branch division; S530, the information gain value of each branch is suppressed and corrected according to the sensitive leakage metric function to obtain the corrected discrimination gain value, and the optimal splitting path is reselected based on the corrected discrimination gain value; S540: Recursively construct a hierarchical decision path tree W according to the corrected optimal splitting path until the preset stopping condition is met, and output the corresponding initial hierarchical result L.

6. The privacy-protected based customer management data triage method of claim 1, wherein: The process of structural backtracking and reconstruction of the hierarchical decision path tree W based on E includes the following steps: S710, based on the privacy leakage metric E, performs risk location on each leaf node in the hierarchical decision path tree W, identifies the path branches whose stability is lower than the preset stability threshold, and marks them as branches to be backtracked. S720, For the branch to be backtracked, backtrack from the leaf node upwards layer by layer to the nearest parent node that satisfies the stability constraint, extract the partitioning features and corresponding partitioning thresholds on the path, and construct a path adjustment candidate set; S730, recalculate the corrected discriminant gain value of each candidate partitioning feature in the path adjustment candidate set, and introduce the privacy leakage metric E as a constraint term to dynamically optimize and adjust the partitioning threshold to generate a reconstructed branch structure; S740, replace the corresponding branch in the original hierarchical decision path tree W with the reconstructed branch structure, perform global consistency verification on the updated hierarchical decision path tree, and output the final data hierarchical classification result.

Images