Performance correlation analysis and visualization system for differential privacy protection

Abstract

The invention relates to the technical field of data privacy protection, and discloses a differential privacy protection performance correlation analysis and visualization system, which is used for comprehensively representing multi-dimensional correlation of performance data by collecting employee performance data and carrying out structured preprocessing to construct a four-order tensor. And a differential privacy mechanism is utilized to perform decomposition and noise injection on the tensor, so that the security of sensitive information is guaranteed. In addition, the system evaluates the privacy sensitivity of each node by constructing a factor map, generates an adaptive visualization result in combination with the remaining privacy budget, and dynamically adjusts a display strategy. By recording user interaction behaviors and budget changes in real time, a game strategy optimizes a display mode, and balance between privacy protection and data availability is achieved. The requirements of privacy protection and data display are effectively balanced, a safe, intelligent and dynamically-adjusted performance analysis and visual display tool can be provided for an enterprise, and the capacity of the enterprise in the aspects of data safety and decision support is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of data privacy protection, and in particular to a performance correlation analysis and visualization system for differential privacy protection. Background Art

[0002] In recent years, with the advancement of digital management in enterprises, multi-dimensional analysis and correlation mining based on employee performance data have become important tools for organizational management. In particular, in the big data environment, companies have accumulated vast amounts of structured and unstructured human resources data. This data contains a wealth of implicit information about employee performance, collaborative relationships, and performance trends, providing valuable insights into organizational optimization, talent assessment, and management decision-making.

[0003] Existing performance data analysis and visualization applications generally use tables, statistical reports, or relationship graphs to present employee metrics and collaborative networks. Some advanced systems are beginning to incorporate graph computing and network relationship analysis to model the intensity or influence of collaboration among employees. However, traditional visualization solutions mostly remain at the level of static data display, lacking the ability to integrate multidimensional data relationships and failing to consider privacy protection mechanisms in a data security environment. While these applications may be adequate for general data analysis scenarios, they lack security assurance when processing sensitive personal information.

[0004] However, existing performance analysis systems generally fail to establish a unified multi-dimensional data expression framework and lack structured modeling of complex relationships such as employee-indicator-time-collaboration, making it difficult to deeply explore the potential value between data. Secondly, current common privacy protection methods mainly rely on data desensitization or pseudo-anonymity processing, which cannot adapt to dynamic interactive display scenarios and easily leak sensitive information during high-dimensional data visualization. In addition, traditional visualization systems use fixed display strategies and cannot flexibly adjust the display granularity or security level according to the user's actual operation behavior, posing a security risk of privacy leakage. Especially in the case of continuous and multiple user interactions, there is a lack of a complete budget control mechanism, which cannot dynamically evaluate and limit privacy consumption. These technical shortcomings have made it difficult to meet the high-standard data analysis needs of modern enterprises for security, intelligence, and adaptability. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the present invention provides a performance correlation analysis and visualization system with differential privacy protection, which solves the problems in existing performance data analysis such as the lack of multi-dimensional correlation expression, weak privacy protection mechanism, and rigid and non-adaptive visualization strategy.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: a performance correlation analysis and visualization system with differential privacy protection, comprising:

[0007] The raw data collection and pre-processing module is used to collect data including employee identification, performance indicators, time periods and collaboration relationships, and perform structured processing;

[0008] Tensor building module, used to encode structured processed data into multi-order tensor structures;

[0009] The differential privacy tensor decomposition module is used to decompose multi-order tensors and introduce structural perturbation differential privacy noise into the decomposed factor matrix;

[0010] The factor graph construction and sensitivity analysis module is used to construct the factor graph based on the perturbed factor matrix and calculate the node sensitivity;

[0011] A visualization rendering module, which dynamically adjusts the display granularity and method of the factor graph based on node sensitivity and the system's remaining privacy budget;

[0012] The interaction and budget control module is used to track user interaction behavior, evaluate privacy budget consumption trends, and guide the selection of display strategies through a game mechanism.

[0013] Preferably, the raw data acquisition and preprocessing module includes:

[0014] Employee information collection unit, used to collect employees' unique identification, job category, and organizational hierarchy;

[0015] Performance indicator processing unit, used to clean and standardize employee performance ratings over multiple time periods;

[0016] Collaboration data coding unit, used to construct a collaboration matrix based on project assignments or communication records.

[0017] Preferably, the tensor building module includes:

[0018] Tensor structure generation unit, used to construct fourth-order tensors

[0019] Among them, n represents the number of employees, m represents the number of performance indicators, p represents the number of time periods, and d represents the order of collaborative influence propagation.

[0020] The score normalization unit is used to normalize the employee's performance score to ensure that the score value conforms to the standard range of tensor element calculation;

[0021] Collaboration impact calculation unit, used to calculate employee u i In a collaborative network, the l-th order neighbor propagation influence φ(u i ,l);

[0022] Each element of the tensor is calculated using the following formula:

[0023]

[0024] Among them, S i,j,k is the normalized performance index j score of employee i at time k, and ∈ is a smoothing constant.

[0025] Preferably, the differential privacy tensor decomposition module includes:

[0026] CP decomposition unit, used to decompose a tensor into multiple factor vector groups;

[0027] The factor matrix perturbation unit is used to apply differential privacy noise to the decomposed vector group;

[0028] The budget control unit adjusts the perturbation intensity according to the privacy budget value set by the user.

[0029] Preferably, the factor matrix perturbation unit is based on the Laplace mechanism for the vector a r Add disturbance, the disturbance mode is:

[0030]

[0031] Among them: a r is the original factor vector, is the disturbance vector after adding noise, Δa r is the sensitivity, ε is the privacy budget;

[0032] The Laplace function Lap is Sampling the scale parameter.

[0033] Preferably, the factor graph construction and sensitivity analysis module includes:

[0034] Graph node generation unit, used to generate employee, indicator, time and collaboration type nodes v from the disturbance factor vector i ;

[0035] Edge weight calculation unit, used to generate edge weight e based on cosine similarity between vectors i,j , defined as:

[0036]

[0037] The node sensitivity scoring unit is used to calculate the node sensitivity according to the following formula:

[0038] S(v)=α·f access (v)+β·f impact (v)+γ·f degree (v);

[0039] in:

[0040] f access (v): node access frequency;

[0041] f impact (v): the influence weight of the node on the analysis results;

[0042] f degree (v): degree of nodes in the graph;

[0043] α, β, γ: Adjust the weight to satisfy α+β+γ=1.

[0044] Preferably, the visual rendering module includes:

[0045] Graph layout unit, used to perform layered and force-directed layout of factor graphs;

[0046] A node display control unit, configured to filter display content based on node sensitivity;

[0047] Layer granularity adjustment unit switches display accuracy based on remaining budget and map complexity.

[0048] Preferably, the display control strategy follows the following mapping relationship:

[0049]

[0050] in:

[0051] S(v) is the node sensitivity, ε r is the current remaining privacy budget, θ1, θ2 are the sensitivity thresholds, and η1 is the system preset threshold.

[0052] Preferably, the interaction and budget control module includes:

[0053] Interaction recording unit, used to collect user operation behaviors on the graph interface;

[0054] Budget update unit, used to consume the budget and update the system status according to the interaction behavior;

[0055] The display strategy decision unit is used to select the display strategy based on the comprehensive budget margin and sensitivity score.

[0056] The present invention also provides a performance correlation analysis and visualization method for differential privacy protection, comprising the following steps:

[0057] S1. Collect employee performance data and perform structured preprocessing;

[0058] S2, construct a fourth-order tensor structure to represent employee, indicator, time and collaboration information;

[0059] S3, calculate the propagation impact and generate tensor elements based on the normalized score;

[0060] S4, perform tensor CP decomposition and inject differential privacy noise into the factor matrix;

[0061] S5. Construct factor graph and evaluate node sensitivity;

[0062] S6, generate adaptive visualization results based on sensitivity and remaining budget;

[0063] S7. Record user interaction behavior and estimate budget changes;

[0064] S8. Adjust the display method through game strategy and update the graph view.

[0065] The present invention provides a performance correlation analysis and visualization system with differential privacy protection.

[0066] Beneficial effects:

[0067] 1. This invention utilizes a technical solution that integrates multidimensional tensor modeling with collaborative influence to achieve a structured representation of employee performance data. Compared to existing analysis methods based solely on tables or simple graphs, this overcomes the issues of information isolation and feature loss, achieving the beneficial effect of comprehensively and relatively expressing performance data.

[0068] 2. This invention utilizes a differentially private tensor decomposition and factor matrix perturbation mechanism to effectively safeguard sensitive data. Unlike traditional anonymization or data desensitization techniques, this approach addresses the privacy limitations of existing algorithms in high-dimensional data processing, achieving a balanced approach to privacy protection and feature extraction in complex data scenarios.

[0069] 3. This invention achieves a dynamic assessment of node privacy risk by establishing a factor graph and sensitivity scoring model. Unlike existing static graph display solutions, this overcomes the mismatch between displayed content and security levels, achieving the beneficial effect of a controllable, visual display of risk while ensuring the value of the display.

[0070] 4. This invention utilizes user-interaction-driven budget control and game optimization to dynamically adjust display strategies. Unlike existing fixed display strategy systems, this system avoids the risk of privacy leaks caused by uncontrollable user operations and achieves interactive, intelligent, and visual display management. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 It is a system structure diagram of the present invention;

[0072] Figure 2Schematic diagram of the method of the present invention. DETAILED DESCRIPTION

[0073] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the present specification. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0074] Please see the attached Figure 1 , an embodiment of the present invention provides a performance correlation analysis and visualization system with differential privacy protection, including:

[0075] The raw data collection and pre-processing module is used to collect data including employee identification, performance indicators, time periods and collaboration relationships, and perform structured processing;

[0076] As the basic module of the performance correlation analysis and visualization system with differential privacy protection, the raw data collection and preprocessing module is mainly used to collect and standardize employees' basic information, performance indicator data, and collaborative relationship data, providing standardized input data for the subsequent construction of multi-dimensional tensors.

[0077] The raw data collection and preprocessing module includes an employee information collection unit, a performance indicator processing unit, and a collaborative data encoding unit. The three are connected in sequence according to the timing logic of data processing. The functions of each sub-module are described in detail below.

[0078] The employee information collection unit is used to collect the employee's unique identification, job category, and organizational level, including the employee identification collection unit, job category collection unit, and organizational level collection unit. The employee identification collection unit is used to collect the employee's unique identification number u i , where i is the employee index, satisfying 1≤i≤n, and n is the total number of employees. The job category collection unit is used to record employee job categories p i , the organizational level collection department is used to record the organizational level to which the employee belongs. i .

[0079] The performance index processing unit is used to clean and standardize the performance scores of employees in multiple time periods, including the original score collection unit, the index standardization processing unit and the time period management unit. The original score collection unit is used to obtain the performance scores of employees in different time periods. k Raw performance ratings within i,j,k , where j is the performance indicator index, satisfying 1≤j≤m, m is the number of performance indicators, k is the time period index, satisfying 1≤k≤p, p is the number of time periods. The time period management unit is used to record each time period t kAfter the original score collection is completed, the original score is normalized by the indicator standardization processing unit to eliminate the differences in the dimensions of different indicators and unify them to the [0,1] interval. The calculation formula is:

[0080]

[0081] in:

[0082] S i,j,k :Employeeu i Normalized performance score for indicator j, time period k;

[0083] The minimum value of indicator j under all employees and time periods;

[0084] The maximum value of indicator j across all employees and time periods;

[0085] ∈: value is 10 -6 , a smoothing constant used to prevent the denominator from being zero.

[0086] The collaboration data encoding unit is used to construct a collaboration matrix based on project assignments or communication records. It includes a collaboration relationship collection unit, a collaboration matrix generation unit, and a collaboration influence calculation unit. The collaboration relationship collection unit is used to capture records of collaboration relationships between employees. These relationships can be based on email exchanges, project collaboration records, or organizational charts.

[0087] The collaboration matrix generation unit is used to construct the employee collaboration matrix C = [c i,q ],in:

[0088]

[0089] The collaborative impact calculation unit is used to calculate employee u i The influence degree φ(u i ,l), the propagation order l satisfies 1≤l≤d; where d is the maximum propagation order set by the system.

[0090] The influence degree is calculated using the adjacency matrix multiplication recursive model, and the initialization conditions are:

[0091]

[0092] The recursive formula is:

[0093]

[0094] in:

[0095] φ(u i ,l): employee u iThe degree of collaborative influence at the diffusion order l;

[0096] c i,q :Employeeu i with u q collaborative relationship indicators;

[0097] φ(u q ,l-1): employee u q The degree of influence at propagation order l-1.

[0098] After obtaining the basic information, performance indicators and collaboration relationship data of employees, the system stores all data in a structured manner to form a standardized data table. The record format of this data table is:

[0099] D={(u i ,p i ,o i ,S i,j,k ,φ(u i ,l))};

[0100] The above data table serves as input data for the subsequent tensor construction module to ensure that the data meets the system's processing requirements in terms of dimensional consistency, indicator standardization, and collaborative impact supplementation.

[0101] The output of the employee information collection unit is connected to the input of the performance indicator processing unit for synchronizing employee identification data. The output of the performance indicator processing unit is connected to the input of the collaboration data encoding unit for providing integrated processing of standardized performance scores and collaboration data. The output of the collaboration data encoding unit is connected to the data table generation unit for forming the final standardized data table D. This embodiment enables the comprehensive collection of basic employee data, the standardized processing of multi-indicator performance data, and the calculation of multi-level influence of collaboration relationships, thereby ensuring data consistency and improving the efficiency of subsequent modeling.

[0102] Tensor building module, used to encode structured processed data into multi-order tensor structures;

[0103] The tensor construction module serves as the data modeling core in the differential privacy-protected performance correlation analysis and visualization system. It is mainly used to perform unified tensor expression modeling on the structured data output by the original data collection and preprocessing modules, realize the standardized expression of employees' multi-dimensional performance data and the fusion calculation of collaborative influence, and ensure the input data quality of subsequent privacy calculations.

[0104] The tensor construction module includes a tensor structure generation unit, a score normalization unit, and a collaborative influence calculation unit. The three are logically connected in sequence according to the data processing flow. The functions of each sub-module are described in detail below.

[0105] Tensor structure generation unit, used to construct fourth-order tensors Its input is connected to the output of the score normalization unit and the collaborative influence calculation unit, and its output is connected to the input of the differential privacy tensor decomposition module.

[0106] Where n represents the number of employees, m represents the number of performance indicators, p represents the number of time periods, and d represents the order of collaborative influence diffusion;

[0107] d refers to the maximum number of hops (or adjacency levels) that information or influence takes to propagate from one employee to another in a collaborative network, essentially the propagation depth of the collaborative path between employees. This value measures the diffusion of collaborative relationships within the network and is a key parameter in constructing the collaborative propagation axis in the fourth-order tensor dimension. The setting of d is based on two factors: the actual depth of the collaborative network, which is determined by measuring the maximum connection level of project collaboration or organizational communication networks to determine a reasonable propagation limit; and empirical tuning, which is based on the balance between privacy protection and performance evaluation in real-world applications.

[0108] The score normalization unit is used to receive the original performance score s output by the original data collection and preprocessing module i,j,k ,in:

[0109] i is the employee index, satisfying 1≤i≤n, and n is the total number of employees;

[0110] j is the performance indicator index, satisfying 1≤j≤m, and m is the total number of performance indicators;

[0111] k is the time period index, satisfying 1≤k≤p, and p is the number of time periods;

[0112] t k is the specific time interval of time period k.

[0113] In order to eliminate the differences in measurement units and scoring intervals of different performance indicators, the scoring normalization unit uses the interval normalization algorithm to process the original score to obtain the standardized performance score S i,j,k , the calculation formula is:

[0114]

[0115] in:

[0116] is the minimum score of indicator j across all employees and time periods;

[0117] is the maximum score of indicator j across all employees and time periods;

[0118] ∈ is a positive number used for smoothing calculations to prevent division by zero errors, and its value is 10 -6 .

[0119] The collaborative influence calculation unit is used to receive the collaborative relationship matrix C constructed by the original data acquisition and pre-processing modules. i,q ],in:

[0120]

[0121] Based on the collaboration matrix C, the collaboration impact calculation unit calculates employee u i The influence degree φ(u i ,l), the propagation order l satisfies 1≤l≤d, where d is the maximum propagation order set by the system.

[0122] The calculation of influence degree adopts recursive algorithm, and the initial conditions are:

[0123]

[0124] When l>1, the recursive calculation formula is:

[0125]

[0126] in:

[0127] φ(u i ,l) represents employee u i The degree of collaborative influence at the diffusion order l;

[0128] φ(u q ,l-1) represents employee u q The degree of influence at propagation order l-1.

[0129] The tensor structure generation unit receives the S output by the score normalization unit. i,j,k and collaboratively influence the output of the computing unit φ(u i ,l), construct a fourth-order tensor This tensor is used to uniformly encode employees, performance indicators, time periods, and collaborative influence relationships.

[0130] Tensor Chinese elements The specific calculation formula is:

[0131]

[0132] in:

[0133] Indicates that employees i , the tensor value of the position of indicator j, time period k, and collaboration order l;

[0134] Indicates that for fixed employees u i, the maximum product value under all indicators j, time periods k and collaboration orders l;

[0135] ∈ is a smoothing constant, which is set to 10 -6 .

[0136] The design of the tensor structure generation unit ensures the integrated expression of standardized performance scores of different indicators and employee collaboration influence within a unified framework. Its core advantage lies in preserving the temporal characteristics, indicator characteristics, and collaborative propagation characteristics of multi-dimensional performance data, providing structured input data for subsequent differential privacy protection and correlation analysis. Through the above technical solution, this implementation method can achieve a unified expression of employee performance and collaboration data of different dimensions, multiple levels, and multiple time series, providing strong data support for the subsequent implementation of differential privacy calculations, and has the technical effect of ensuring data integrity and expression standardization.

[0137] The physical connection relationship and logical flow of each unit are as follows:

[0138] First, the score normalization unit obtains the original performance score and calculates the standardized score S i,j,k ;

[0139] Then, the collaborative influence calculation unit recursively calculates the collaborative influence degree φ(u i ,l);

[0140] Finally, the tensor structure generation unit generates a fourth-order tensor And output it to the differential privacy tensor decomposition module.

[0141] The differential privacy tensor decomposition module is used to decompose multi-order tensors and introduce structural perturbation differential privacy noise into the decomposed factor matrix;

[0142] The differential privacy tensor decomposition module is the core privacy computing module of the performance correlation analysis and visualization system with differential privacy protection. It is mainly used to perform low-rank decomposition on the multi-dimensional tensor data output by the tensor construction module while ensuring data privacy, extract the main feature information, and prevent the leakage of sensitive information through the differential privacy perturbation mechanism.

[0143] The differential privacy tensor decomposition module includes a CP decomposition unit, a factor matrix perturbation unit, and a budget control unit. Each unit is connected in sequence according to the logical flow and physically connected in sequence to form a complete data processing path. The function of each sub-module is explained in detail below.

[0144] The CP decomposition unit is used to decompose the input fourth-order tensor Perform low-rank decomposition and finally decompose into multiple factor matrices, tensors The dimensions are: n is the number of employees; m is the number of performance indicators; p is the number of time periods; and d is the maximum order of collaborative influence transmission.

[0145] When constructing a factor graph, it is necessary to ensure that the elements of each factor matrix are correctly mapped to the nodes of the graph. If only the employee factor matrix A is used, while other factor matrices (such as B, C, and D) are ignored, the structure of the graph will be distorted and will not fully reflect the interactive nature of the task and the high-dimensional nature of the data.

[0146] Tensor The decomposition uses the CANDECOMP / PARAFAC (CP) decomposition method to represent the tensor as the sum of the outer products of multiple vectors of rank R. The calculation expression is:

[0147]

[0148] in:

[0149] R is the rank of tensor decomposition, which depends on the system's requirement for feature preservation; is the rth employee feature vector; is the rth performance indicator feature vector; is the feature vector of the rth time period; is the rth collaborative propagation order eigenvector; It is the outer product operation of vectors.

[0150] The factor matrices obtained by decomposition are:

[0151]

[0152] Among them, matrices A, B, C, and D represent the employee factor matrix, performance indicator factor matrix, time period factor matrix, and collaborative communication factor matrix, respectively.

[0153] The factor matrix perturbation unit is used to apply differential privacy noise to the factor matrix obtained by decomposition to prevent feature leakage. The perturbation mechanism uses the Laplace mechanism, which achieves differential privacy protection by independently adding Laplace noise to each column vector in the factor matrix.

[0154] For the rth column vector a of the factor matrix A r , the perturbation formula is:

[0155]

[0156] in:

[0157] a r is the original factor vector; is the disturbance vector after adding noise; Δa ris sensitivity, which indicates the maximum impact of a single change in input data on the factor vector; ε r is the privacy budget for the rth factor; Lap(·) is the privacy budget for the rth factor; is a Laplace distribution with a scale parameter.

[0158] Sensitivity Δa r The calculation of is determined based on the normalization characteristics of the input data. If the input data has been normalized to [0,1], then Δa r =1.

[0159] To prevent the privacy budget from accumulating too quickly and exceeding the limit, this embodiment proposes the following control strategy:

[0160] Privacy budget consumption records: Each tensor decomposition operation is independently recorded, the ε value of each operation is calculated, and the consumption of the privacy budget is tracked during each decomposition process. By dynamically adjusting the budget consumption, the total privacy budget is ensured not to exceed the limit.

[0161] Hierarchical Differential Privacy: Using a Hierarchical Differential Privacy strategy, each decomposition step requires calculating an independent privacy budget and ensuring that the budget consumption of each layer does not exceed the preset upper limit. This ensures that the budget constraint will not be violated even if multiple decomposition steps are accumulated.

[0162] Application of the Serial Combination Theorem: In this embodiment, for multiple decomposition operations (such as multiple iterations of CP decomposition), the privacy budget consumption of each operation will be accumulated through the Combination Theorem. The specific formula is:

[0163] ε total =ε1+ε2+…+ε n ;

[0164] Among them, ε n Represents the privacy budget consumption of each operation. To avoid exceeding the limit, we will strictly calculate and plan the budget consumption of each operation during the design phase.

[0165] Privacy budget allocation and adjustment: The system dynamically adjusts the budget allocation based on the privacy budget consumption of each operation. If the system detects that the budget is about to be exceeded, it will reduce the budget consumption of subsequent operations, for example by reducing noise intensity or simplifying certain calculation steps.

[0166] The perturbation processing method of factor matrices B, C, and D is the same as that of factor matrix A, and the perturbation vectors are generated respectively.

[0167] Perturbation operations (such as differential privacy mechanisms) may affect the content of the factor matrix. Therefore, the perturbed factor matrix needs to be graphed based on the perturbed data.

[0168] When mapping graph nodes, it's important to ensure that the data in each factor matrix can be associated with the corresponding node in the graph, and that there are no dimensional inconsistencies. Specifically, employee factor matrix A might map to the employee node, while other factor matrices B, C, and D should be mapped to other related nodes or relationships.

[0169] In this example, to avoid data dimension mismatch, it is recommended to use appropriate methods to process these factor matrices when constructing the graph. The following methods can be considered:

[0170] Mapping rules: Clarify the mapping rules to ensure that the data in each factor matrix can be correctly connected to the nodes or edges in the graph.

[0171] Data fusion: The data of multiple factor matrices are fused before constructing the graph to ensure that the data of each dimension is reflected.

[0172] Post-perturbation correction: For the perturbed factor matrix, it may be necessary to correct or rebuild part of the graph structure to ensure the integrity and consistency of the information.

[0173] The budget control unit is used to manage the privacy budget of the system as a whole total , and dynamically allocate the disturbance budget of each factor matrix. The budget allocation formula for a single factor vector is:

[0174]

[0175] in:

[0176] N f is the number of factor matrices to which noise is added. In this embodiment, N f =4; R is the rank of tensor decomposition.

[0177] To prevent the privacy budget ε total When fixed, if the rank R is large, the noise may be overly diluted. Therefore, the following experimental verification and theoretical analysis are performed:

[0178] First, from a theoretical perspective, we can see that when the number of factor matrices N f Or when the rank R of the tensor decomposition is large, the privacy budget ε allocated to a single factor vector is r The reduction will be significant, resulting in an increase in the Laplace perturbation amplitude, which may introduce a large reconstruction error, affecting the interpretability of the data and the effectiveness of subsequent graph display. rUnder the premise of being fixed, the rank R of tensor decomposition should not be too large and should be reasonably set according to the data sparsity and reconstruction accuracy requirements to achieve an effective compromise between privacy and availability.

[0179] Secondly, in the system prototype verification of this implementation, we designed experimental schemes under different tensor ranks R = 10, 20, and 40 for multiple actual enterprise performance data sets, and total =1.0, number of factor matrices N f =4, we evaluate the factor matrix reconstruction error (Root Mean Square Error, RMSE), the number of nodes that can be identified in the graph, and the sensitivity masking rate. The experimental results show that:

[0180] When R≤20, the system can maintain high reconstruction accuracy and graph connectivity while ensuring controllable privacy perturbation intensity, and the node information retention rate is good;

[0181] When R≥40, the disturbance intensity increases significantly, the reconstruction error increases, and the number of high-sensitivity nodes in the atlas increases significantly, causing the display masking ratio to increase to about 37%, affecting the integrity and expression effect of the atlas.

[0182] To further optimize the allocation strategy, this implementation supports extending budget allocation to an adaptive weighted allocation mechanism, that is, setting different budget weights based on the contribution or sensitivity of the factor dimension to the analysis, thereby improving the retention capacity of key factors under budget constraints.

[0183] In summary, to ensure a balance between system performance and privacy strength, this implementation recommends controlling the tensor decomposition rank R to 10 to 30 by default, and supports developers to manually set the budget ratio distribution in advanced configuration to achieve flexible differential privacy policy adjustment.

[0184] The factor matrices after perturbation are:

[0185]

[0186] In this embodiment, the input end of the CP decomposition unit is connected to the output end of the tensor construction module, receiving multi-order tensors As input data; the input end of the factor matrix perturbation unit is connected to the output end of the CP decomposition unit, which is responsible for performing privacy perturbation processing on the factor matrix; the budget control unit controls the noise addition budget of the factor matrix perturbation unit. Finally, the perturbed factor matrix The output of the differential privacy tensor decomposition module is transmitted to the factor graph construction and sensitivity analysis module for subsequent factor graph construction and privacy sensitivity control.

[0187] First, a low-rank decomposition model of a multidimensional tensor is established, and then the Laplace mechanism is used to introduce differential privacy noise into the decomposed factor matrix. Finally, a factor matrix that meets the privacy protection requirements is output. This achieves the goal of protecting data privacy while retaining the feature expression ability of the data, and has the technical effect of ensuring privacy security and improving data utilization efficiency.

[0188] The factor graph construction and sensitivity analysis module is used to construct the factor graph based on the perturbed factor matrix and calculate the node sensitivity;

[0189] As an important component of the performance correlation analysis and visualization system for differential privacy protection, the factor graph construction and sensitivity analysis module is mainly used to construct a factor relationship graph based on the factor matrix output by the differential privacy tensor decomposition module, and perform sensitivity scoring and privacy security assessment on each node in the graph, providing a data basis and display control basis for the subsequent visualization rendering module.

[0190] The factor graph construction and sensitivity analysis module includes a graph node generation unit, an edge weight calculation unit, a node sensitivity scoring unit, and a graph data management unit. Each unit is logically connected in sequence and physically connected to each other to form a complete graph construction and sensitivity analysis process. The following is a further description of each sub-module.

[0191] First, the input of the graph node generation unit is connected to the output of the differential privacy tensor decomposition module to receive the perturbed employee factor matrix

[0192] Where: n is the number of employees; R is the rank of tensor decomposition, that is, the number of feature dimensions;

[0193] For employees i The perturbed eigenvector of .

[0194] The graph node generation unit maps each employee feature vector in the employee factor matrix to a node in the graph, and sets the factor graph G = (V, E).

[0195] in:

[0196] V={v1,v2,…,v n} is a node set, corresponding to n employees;

[0197] Node v i The attribute vector is the perturbed eigenvector

[0198] The edge weight calculation unit is used to calculate the edge weight between any two nodes v in the graph. i With v j The edge weight e between i,j, the cosine similarity calculation formula is:

[0199]

[0200] in: represents the inner product of eigenvectors; Represents a vector The Euclidean norm of is calculated as:

[0201]

[0202] After the edge weight calculation is completed, all the calculated edge weights e i,j The edge set E is formed to form a complete factor graph G = (V, E).

[0203] The node sensitivity scoring unit is used to calculate the sensitivity of each node v in the graph. i Sensitivity score S(v i ), which is used to evaluate the privacy risk of the node in the visualization display. The sensitivity scoring formula is:

[0204] S(v i )=α×f access (v i )+β×f impact (v i )+γ×f degree (v i );

[0205] in:

[0206] α, β, γ are weight coefficients, satisfying α+β+γ=1;

[0207] f access (v i ) is the node v i The access frequency indicates the number of times the user visits the node during the system interaction process;

[0208] f impact (v i ) is the node v i The feature importance index of

[0209] f degree (v i ) is the node v i The degree of connectivity in the graph.

[0210] In the calculation formula for the node sensitivity score above, the weight parameters α, β, and γ are used to adjust the different types of sensitivity effects. Specifically, α, β, and γ control the weights of the following aspects respectively:

[0211] α: affects the node connectivity (i.e. the importance of the node in the graph);

[0212] β: the attribute value that affects the node (i.e., the privacy sensitivity of the node);

[0213] γ: affects the interaction frequency of the node (i.e., the frequency of collaboration between the node and other nodes);

[0214] Due to the characteristics of different application scenarios and data sets, the weight parameters need to be set according to actual needs and can be adjusted in the following ways:

[0215] Expert experience setting: Based on the experience of domain experts and their understanding of the data, the initial setting of weight parameters can be:

[0216] α=0.4,β=0.3,γ=0.3;

[0217] This setting is based on general privacy protection requirements, where connectivity has the greatest impact on node sensitivity and is therefore given a higher weight.

[0218] Data-driven optimization: When data is abundant and model training is possible, weight parameters can be dynamically adjusted using data-driven optimization algorithms (such as gradient descent or Bayesian optimization). Through backpropagation of errors during training, each weight value is automatically optimized to maximize the accuracy of node sensitivity scores.

[0219] Scenario-based adjustment: For different application scenarios, weight parameters can be dynamically adjusted according to specific goals:

[0220] In highly privacy-sensitive scenarios (such as medical data analysis), the weights of β and γ can be increased to enhance the emphasis on node privacy attributes and interaction frequency, and the weight of α can be reduced to reduce the impact of connectivity on sensitivity.

[0221] In collaboration-intensive scenarios (such as social network analysis), the weight of α can be appropriately increased to make the impact of connectivity on the sensitivity score more significant.

[0222] User-defined adjustments: In the case of user-defined weighting, the system provides a parameter adjustment interface, allowing users to manually set weighting parameters according to actual needs to adapt to the complexity of different data analyses. For example, users can flexibly select and adjust weighting parameters based on their privacy protection policies and business needs.

[0223] Dynamic learning and adjustment: If the system has online learning capabilities, weight parameters can be dynamically adjusted during model execution. By continuously collecting real-time data feedback, the weight parameters will be gradually optimized as new data is input, allowing the model to automatically adjust to changes in data over time.

[0224] Through the above method, it can be ensured that the calculation of node sensitivity scores has high flexibility and adaptability in different application scenarios, thereby optimizing the balance between privacy protection and data availability.

[0225] Access frequency f access (v i ) Through system log recording unit statistics, the number of user clicks or query behaviors is accumulated per unit time to ensure objectivity.

[0226] Feature importance index f impact (v i ) According to node v i The norm calculation of the corresponding eigenvector is calculated as follows:

[0227]

[0228] in:

[0229] is the Euclidean norm of the eigenvector.

[0230] Connectivity f degree (v i ) According to node v i The edge weights and calculations between all other nodes are calculated as follows:

[0231]

[0232] in:

[0233] e i,j For node v i With v j The edge weights between .

[0234] The graph data management unit is used to manage the factor graph data G = (V, E) and the sensitivity score S (v i ) for unified management and transmits the results to the visualization rendering module for dynamic display control.

[0235] First, the nodes and edge weights of the factor graph are generated based on the perturbed factor matrix. The sensitivity score of each node in the graph is then calculated. Finally, the graph data is managed and output, providing the foundational data for privacy-aware visualization. This implementation enables the graphical representation of multidimensional performance characteristics and collaborative relationships, while also controlling the sensitivity of key nodes. This approach has the technical effect of ensuring data privacy and security while enhancing the flexibility of graph presentation.

[0236] A visualization rendering module, which dynamically adjusts the display granularity and method of the factor graph based on node sensitivity and the system's remaining privacy budget;

[0237] The visualization rendering module is mainly used to visualize the factor graph data output by the factor graph construction and sensitivity analysis modules under controlled privacy budget conditions, and dynamically adjust the graph display strategy and granularity to achieve safe and controllable graph interaction and information display.

[0238] The visualization rendering module includes a graph layout unit, a node display control unit, a layer granularity adjustment unit, and a graph rendering execution unit. Each unit is logically connected in sequence and physically interconnected to form a complete privacy-controllable visualization rendering process. The following is a detailed description of each sub-module.

[0239] The input end of the graph layout unit is connected to the output end of the factor graph construction and sensitivity analysis module, and is used to receive the factor graph data G = (V, E);

[0240] in:

[0241] V={v1,v2,…,v n} represents the node set in the factor graph;

[0242] E={e i,j} represents the set of edges between nodes.

[0243] Edge weight e i,j The calculation uses cosine similarity, which satisfies:

[0244]

[0245] in:

[0246] For employees i The eigenvector of comes from the perturbation factor matrix of the differential privacy tensor decomposition module is the Euclidean norm, which is calculated as:

[0247]

[0248] The graph layout unit adopts a node layout method based on a force-directed algorithm to simulate the attraction and repulsion between nodes. By iteratively optimizing the node coordinates, the graph has a clear structure in the two-dimensional or three-dimensional display space, avoiding node overlap and improving visual readability.

[0249] The input of the node display control unit is connected to the output of the graph layout unit to control the display status of each node. The node display strategy is based on the node sensitivity score S(v i ) and the current remaining privacy budget ε r Dynamic adjustment, specifically showing the strategy function P(v i ) is defined as:

[0250]

[0251] in:

[0252] S(v i ) is the node v i The sensitivity score is calculated as:

[0253] S(v i )=α×f access (v i )+β×f impact (v i )+γ×f degree (v i );

[0254] The parameters have the following meanings:

[0255] f access (v i ) is the access frequency, which is the number of node accesses per unit time counted by the system log;

[0256] f impact (v i ) is the node feature influence, calculated as the Euclidean norm of the node feature vector;

[0257] f degree (v i ) is the connectivity of the node in the graph, calculated as the sum of all its edge weights:

[0258]

[0259] in:

[0260] α, β, γ are sensitivity score weight coefficients, satisfying: α+β+γ=1;

[0261] The threshold parameter is defined as:

[0262] θ1 and θ2 are different sensitivity thresholds; θ1>θ2;

[0263] η1 is the privacy budget threshold, which is used to determine whether to enter the blurred or hidden display state.

[0264] In this embodiment, the policy function P(v i ) is used to determine whether to display a node v in the visualization graph i The key parameters involved in its decision logic include the display threshold θ1, the masking threshold θ2, and the display priority filtering threshold η1. These parameters have a significant impact on the accuracy of the visualization results and the strength of privacy protection.

[0265] To enhance the practicality and adaptability of the system, the setting basis and adjustment mechanism of each threshold parameter are explained as follows:

[0266] Threshold setting logic:

[0267] The initial value of each threshold parameter is set based on historical data statistics and preset rules:

[0268] θ1 represents the upper limit of the sensitivity of the display node, which is usually set by 15% to 25% higher than the median of the previous sensitivity score distribution;

[0269] θ2 represents the strong shielding threshold. When the node sensitivity is higher than this value, it will be shielded unconditionally. The setting principle is the upper quartile point in the sensitivity distribution.

[0270] η1 represents the display priority of the node in the entire graph. Usually, the top 30% to 50% of the nodes are set as candidate nodes for display after ranking based on indicators such as node connectivity and weight score.

[0271] Adaptive adjustment mechanism:

[0272] This system supports dynamic parameter adjustment based on running state, including:

[0273] Dynamic threshold sliding mechanism: When the remaining visualization budget decreases, the system automatically increases the θ1 value to tighten the display standard; when the budget is sufficient, it is appropriately lowered to increase the number of display nodes;

[0274] Real-time sensitivity statistics update: The system can automatically adjust θ2 based on the latest sensitivity distribution to ensure that highly sensitive nodes are continuously shielded;

[0275] User-defined strategy interface: allows advanced users to customize the specific value ranges of θ1, θ2, and η1 in the settings panel to adapt to different business needs.

[0276] Recommended default values:

[0277] To facilitate initial deployment and quick startup, the system uses the following parameter configuration by default:

[0278] θ1=0.6, θ2=0.85, η1=0.35;

[0279] The sensitivity range is normalized to [0, 1]. The default value is suitable for most general privacy visualization scenarios and has a good balance.

[0280] Through the above-mentioned mechanism combining static setting with dynamic adjustment, it is possible to ensure that the strategy function P(v i ) has good flexibility and robustness under different sensitivity distributions, budget constraints and display requirements, thereby improving the actual application performance of the system.

[0281] The input of the layer granularity adjustment unit is connected to the output of the node display control unit, which is used to adjust the overall sensitivity score of the current system according to the average value.

[0282] With the remaining privacy budget ε r , dynamically adjust the granularity of the displayed graph.

[0283] The formula for calculating the mean overall sensitivity score is:

[0284]

[0285] The granularity adjustment strategy is:

[0286]

[0287] in:

[0288] G overview To automatically aggregate the simplified overview graph, only nodes with low sensitivity scores are retained;

[0289] θ m is the overall sensitivity mean threshold; η2 is the remaining budget granularity control threshold; ε r Budget for remaining privacy.

[0290] The input end of the atlas rendering execution unit is connected to the output end of the layer granularity adjustment unit, which is used to display the atlas G according to the final determination. display , generate the visual rendering results of the graph and provide it to the system front-end for display.

[0291] First, a spatial layout model of the factor graph is established. The node display control algorithm then dynamically adjusts the node display mode. A granularity adjustment strategy is then used to control the overall display hierarchy. Finally, the rendering execution unit outputs the graph display results, providing the system with privacy-controllable graph visualization capabilities. This implementation enables secure visualization and dynamic privacy control of factor graphs in multi-dimensional sensitive data scenarios, ensuring data privacy and improving system usability and visualization.

[0292] The interaction and budget control module is used to track user interaction behavior, evaluate privacy budget consumption trends, and guide the selection of display strategies through a game mechanism.

[0293] The interaction and budget control module is a key control module in the performance correlation analysis and visualization system with differential privacy protection. It is mainly used to track users' interactive behavior in real time, calculate privacy budget consumption based on behavior logs, dynamically adjust visualization display strategies, and achieve a dynamic game balance between system privacy protection strength and data availability.

[0294] The interaction and budget control module includes an interaction recording unit, a budget update unit, a display strategy decision unit, and a feedback transmission unit. These four units are logically connected and physically interconnected in sequence to form a complete privacy budget control and display strategy adaptive adjustment process. The following is a detailed description of each sub-module.

[0295] The input end of the interaction recording unit is connected to the front-end interaction interface of the visualization rendering module to obtain the user interaction behavior set U = {u1,u2,…,u k},in:

[0296] Each interaction behavior u k The record structure is a tuple u k =(type(k),t k ,v k ),

[0297] type(k) is the operation type, including click, zoom, search, view details, etc.; k timestamp for the behavior; v k The target node number.

[0298] The input of the budget update unit is connected to the output of the interaction recording unit, which is used to calculate the corresponding privacy budget consumption Δε according to the behavior set U k ,The calculation model of budget consumption considers the weight of the operation type and the ,sensitivity of the target node, and the calculation formula is:

[0299] Δε k =w type(k) ×S(v k );

[0300] in:

[0301] w type(k) is the weight of the interactive operation type. The specific value is set according to the impact of the operation on the risk of privacy leakage. The value range is 0 <w type(k) ≤1;

[0302] S(v k ) is the target node v k The sensitivity score of is calculated to satisfy:

[0303] S(v k )=α×f access (v k )+β×f impact (v k )+γ×f degree (v k );

[0304] in:

[0305] faccess (v k ) is the access frequency of the node, which is counted by the interaction records;

[0306] f impact (v k ) is the characteristic influence of the node, is the Euclidean norm of the characteristic vector, and satisfies:

[0307]

[0308] in:

[0309] The weight parameter α+β+γ=1.

[0310] Remaining privacy budget ε r The real-time update formula is:

[0311]

[0312] in:

[0313] K is the cumulative number of interactive behaviors in the current stage; the initial budget ε0 is set uniformly by the system.

[0314] In order to improve the adaptability and protection capability of the budget update mechanism to complex interactive behaviors, this implementation method introduces an interaction risk factor model (IRM) based on the basic budget consumption calculation formula to quantitatively model the potential privacy leakage caused by high-risk behaviors such as continuous queries, combined operations, and association analysis, and dynamically adjust the budget consumption accordingly.

[0315] The original budget consumption is calculated as follows:

[0316] Δ∈=w op ·S(v i );

[0317] where w op is the operation type weight, S(v i ) is the sensitivity of the target node.

[0318] Considering that this approach may underestimate the real risk when facing complex user behaviors, this implementation introduces the following extension mechanism:

[0319] Add interaction context parameters:

[0320] Based on the above, the interaction context weight w is introduced ctx ,Taking into account factors such as operation frequency, time continuity and target node set overlap, the update budget consumption formula is:

[0321] Δ∈=w op ·S(v i)·w ctx ;

[0322] where w ctx ∈[1.0,3.0], the higher the value, the higher the privacy leakage risk of the current operation and the greater the budget consumption.

[0323] The calculation logic of context weight is as follows:

[0324] Continuous query penalty factor: If a user initiates continuous query operations n≥3 on the same type of node within a short period of time (e.g., within 10 seconds), an increasing factor is introduced:

[0325]

[0326] Overlapping target gain factor: If the intersection ratio of the current query and the historical query target nodes is higher than 60%, the weight is increased by an additional 0.5.

[0327] Combined behavior correlation factor: For combined operations that are known to cause privacy linkage (such as "view node attributes" + "export graph relationships"), the system defines a blacklist operation pair and directly sets w ctx =2.5~3.0.

[0328] Adaptive protection mechanisms:

[0329] When the accumulated budget is consumed quickly or the system detects high-frequency aggregation characteristics in user operations, the system will automatically trigger:

[0330] Display delay (to prevent image scanning); fuzzy display (node ​​names are blurred); partial shielding (temporarily hiding sensitive nodes that have been visited frequently); or recommending users to sleep (cool down) for a certain period of time.

[0331] This mechanism can effectively prevent attackers from inferring sensitive information through frequent interactive operations, thereby improving the overall privacy robustness of the system.

[0332] The input of the strategy decision unit is connected to the output of the budget update unit, which is used to update the budget according to the latest remaining budget ε r The sensitivity of the whole spectrum Dynamically decide the node display strategy and display granularity of the visualization rendering module. The overall sensitivity mean calculation formula is:

[0333]

[0334] The goal of display strategy optimization is to maximize the visual display effect while ensuring that the privacy budget is controlled. The optimization objective function is the display distortion function L display Minimization of:

[0335]

[0336] Where: P(v i ) is given in the above-mentioned visualization rendering module.

[0337] The output of the display strategy decision unit is connected to the visualization rendering module, which transmits optimized display strategy control instructions and adjusts the display method in real time. The input of the feedback transmission unit is connected to the output of the display strategy decision unit to record complete budget consumption logs, display strategy adjustment logs, and interaction behavior logs, and synchronize them to the system database to ensure the traceability and security of the interaction and budget control process.

[0338] First, a complete user interaction data model is constructed. The privacy budget is then dynamically updated using a budget consumption calculation model. The display strategy is then optimized through a privacy budget-availability game mechanism. Finally, logging and feedback are performed, completing a complete technical chain of interaction-driven, privacy-controlled, and adaptive display. This implementation enables dynamic privacy protection and interactive control of sensitive data visualization, effectively safeguarding the system's data security, interactive experience, and policy adaptability.

[0339] The performance correlation analysis and visualization method for differential privacy protection described below and the performance correlation analysis and visualization system for differential privacy protection described above can be referenced to each other.

[0340] See also Figure 2 The present invention also provides a performance correlation analysis and visualization method for differential privacy protection, comprising the following steps:

[0341] S1. Collect multi-dimensional employee performance data, including personal information, performance indicators, timestamps, and collaboration information. Pre-processing, such as data cleaning, formatting, and standardization, ensures data quality and consistency, providing a foundation for subsequent analysis. The processed data will be used to construct a multi-dimensional performance analysis model.

[0342] S2. Construct a fourth-order tensor, dividing the data into four dimensions: employees, performance indicators, time, and collaboration. Each dimension represents an employee, an assessment indicator, a time point, and the collaborative relationship between employees, ensuring that all related data is comprehensively considered and providing comprehensive input for subsequent analysis.

[0343] S3. Calculate the collaborative impact between employees using a propagation algorithm from graph theory and generate tensor elements based on the normalized scores. Propagation impact quantifies the degree of collaboration between employees, allowing the tensor to reflect complex interactive relationships and enrich data representation.

[0344] S4. Perform CP decomposition on the fourth-order tensor to extract a low-rank factor matrix. To protect privacy, differential privacy noise is injected into the factor matrix. Laplace or Gaussian noise is used to control the privacy budget to ensure that the data analysis results do not leak sensitive information.

[0345] S5. Based on the factor matrix after CP decomposition, a factor graph is constructed to evaluate the sensitivity of nodes (employees). The sensitivity of nodes is assessed based on the privacy risk of employee information, taking into account factors such as the degree of disclosure and access frequency, providing data support for subsequent visualization and presentation;

[0346] S6. Generate adaptive visualization results based on the node's sensitivity score and remaining privacy budget. Highly sensitive nodes adopt a conservative display strategy, while low-sensitivity nodes provide detailed information, dynamically balancing the needs of data display and privacy protection.

[0347] S7. Record user interactions (such as clicks, viewing details, etc.) and estimate changes in the privacy budget during the interaction. Each user action may increase or decrease the privacy budget. The system updates the budget in real time to ensure that privacy protection is strictly controlled during the display process.

[0348] S8. Based on user interaction records and remaining budget, use game-playing strategies to adjust the display method to ensure optimal data display within budget constraints. When budget is insufficient, blur or hide certain information to protect privacy while improving the user experience. Update the graph view to adapt to new display requirements.

[0349] The device of this embodiment can be used to execute the above method embodiment, and its principles and technical effects are similar, so they will not be repeated here.

[0350] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A performance correlation analysis and visualization system with differential privacy protection, characterized by: include: The raw data collection and pre-processing module is used to collect data including employee identification, performance indicators, time periods and collaboration relationships, and perform structured processing; Tensor building module, used to encode structured processed data into multi-order tensor structures; The differential privacy tensor decomposition module is used to decompose multi-order tensors and introduce structural perturbation differential privacy noise into the decomposed factor matrix; The factor graph construction and sensitivity analysis module is used to construct the factor graph based on the perturbed factor matrix and calculate the node sensitivity; A visualization rendering module, which dynamically adjusts the display granularity and method of the factor graph based on node sensitivity and the system's remaining privacy budget; The interaction and budget control module is used to track user interaction behavior, evaluate privacy budget consumption trends, and guide the selection of display strategies through a game mechanism.

2. The performance correlation analysis and visualization system with differential privacy protection according to claim 1, characterized in that: The raw data acquisition and preprocessing module includes: Employee information collection unit, used to collect employees' unique identification, job category, and organizational hierarchy; Performance indicator processing unit, used to clean and standardize employee performance ratings over multiple time periods; Collaboration data coding unit, used to construct a collaboration matrix based on project assignments or communication records.

3. The performance correlation analysis and visualization system with differential privacy protection according to claim 1, characterized in that: The tensor building block includes: Tensor structure generation unit, used to construct fourth-order tensors Where n represents the number of employees, m represents the number of performance indicators, p represents the number of time periods, and d represents the order of collaborative influence diffusion; The score normalization unit is used to normalize the employee's performance score to ensure that the score value conforms to the standard range of tensor element calculation; Collaboration impact calculation unit, used to calculate employee u i In a collaborative network, the l-th order neighbor propagation influence φ(u i ,l); Each element of the tensor is calculated using the following formula: Among them, S i,j,k is the normalized performance index j score of employee i at time k, and ∈ is a smoothing constant.

4. The performance correlation analysis and visualization system with differential privacy protection according to claim 1, characterized in that: The differential privacy tensor decomposition module includes: CP decomposition unit, used to decompose a tensor into multiple factor vector groups; The factor matrix perturbation unit is used to apply differential privacy noise to the decomposed vector group; The budget control unit adjusts the perturbation intensity according to the privacy budget value set by the user.

5. The performance correlation analysis and visualization system with differential privacy protection according to claim 4, characterized in that: The factor matrix perturbation unit is based on the Laplace mechanism for the vector a r Add disturbance, the disturbance mode is: Among them: a r is the original factor vector, is the disturbance vector after adding noise, Δa r is the sensitivity, ε is the privacy budget; The Laplace function Lap is Sampling the scale parameter.

6. The performance correlation analysis and visualization system with differential privacy protection according to claim 1, characterized in that: The factor graph construction and sensitivity analysis module includes: Graph node generation unit, used to generate employee, indicator, time and collaboration type nodes v from the disturbance factor vector i ; Edge weight calculation unit, used to generate edge weight e based on cosine similarity between vectors i,j , defined as: The node sensitivity scoring unit is used to calculate the node sensitivity according to the following formula: S(v)=α·f access (v)·+β·f impact (v)+γ·f degree (v); in: f access (v): node access frequency; f impact (v): the influence weight of the node on the analysis results; f degree (v): degree of nodes in the graph; α, β, γ: Adjust the weight to satisfy α+β+γ=1.

7. The performance correlation analysis and visualization system with differential privacy protection according to claim 1, characterized in that: The visual rendering module includes: Graph layout unit, used to perform layered and force-directed layout of factor graphs; A node display control unit, configured to filter display content based on node sensitivity; Layer granularity adjustment unit switches display accuracy based on remaining budget and map complexity.

8. The performance correlation analysis and visualization system with differential privacy protection according to claim 7, characterized in that: The display control strategy follows the following mapping relationship: in: S(v) is the node sensitivity, ε r is the current remaining privacy budget, θ1, θ2 are sensitivity thresholds, and η1 is the system preset threshold.

9. The performance correlation analysis and visualization system with differential privacy protection according to claim 1, characterized in that: The interaction and budget control module includes: Interaction recording unit, used to collect user operation behaviors on the graph interface; Budget update unit, used to consume the budget and update the system status according to the interaction behavior; The display strategy decision unit is used to select the display strategy based on the comprehensive budget margin and sensitivity score.

10. A differential privacy-preserving performance correlation analysis and visualization method, applied to the differential privacy-preserving performance correlation analysis and visualization system according to any one of claims 1 to 9, characterized in that: The steps include: S1. Collect employee performance data and perform structured preprocessing; S2, construct a fourth-order tensor structure to represent employee, indicator, time and collaboration information; S3, calculate the propagation impact and generate tensor elements based on the normalized score; S4, perform tensor CP decomposition and inject differential privacy noise into the factor matrix; S5. Construct factor graph and evaluate node sensitivity; S6, generate adaptive visualization results based on sensitivity and remaining budget; S7. Record user interaction behavior and estimate budget changes; S8. Adjust the display method through game strategy and update the graph view.

Images