Enterprise human resource performance evaluation method based on knowledge graph

Abstract

The invention discloses an enterprise human resource performance evaluation method based on a knowledge graph, and the method comprises the following steps: S1, collecting employee daily work data, task cooperation data, process tracking data, enterprise knowledge base data and external market dynamic data through a multi-source heterogeneous data interface, and obtaining initial multi-modal data; s2, carrying out cleaning and structured processing on the initial multi-modal data; s3, automatically extracting performance semantic features of the employees on the basis of an improved Mamba framework; s4, constructing a multi-granularity logical reasoning knowledge graph library, and learning and updating the weight of logical reasoning rules in real time; s5, performing neural logic reasoning by using the dynamic logic rule set to obtain employee performance grade scores; s6, realizing cross-department employee performance data collaborative analysis through a federal knowledge distillation method; and S7, optimizing the performance evaluation result according to the deviation between the evaluation result and the actual performance feedback data. According to the invention, objective, dynamic and comprehensive performance evaluation is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of human resource management, and in particular to a method for evaluating enterprise human resource performance based on a knowledge graph. Background Art

[0002] With the rapid development of big data, artificial intelligence, and knowledge graph technologies, the field of corporate human resources management is gradually showing a trend towards intelligent and data-driven approaches. Existing technologies generally employ performance management methods based on key performance indicators (KPIs). These methods periodically evaluate employee performance through manually set quantitative indicators and scoring rules. This typically relies on the company's human resources department or management staff to determine indicators based on experience. Internal management systems then regularly collect data, manually assign scores, and perform summary analysis to generate employee performance evaluation results.

[0003] To address the subjectivity of manual scoring, some companies are attempting to automate the analysis of employee behavioral or performance-related data using data mining, machine learning, or traditional statistical analysis. For example, existing technologies offer automated scoring models based on employee attendance, task completion, or process compliance. These methods typically employ supervised or unsupervised learning methods, analyzing employee behavioral characteristics through mathematical models to generate performance scores or rankings.

[0004] However, the above technical solutions still have some obvious limitations in practical application:

[0005] (1) Performance evaluation methods based on traditional KPI indicators or a single data source have relatively single data dimensions and cannot fully and objectively reflect employees' actual work performance and potential contributions;

[0006] (2) Although data mining and machine learning methods have improved the automation level of performance evaluation, the feature extraction capabilities of such methods are still limited, and they lack in-depth analysis and modeling of the semantic characteristics of employee behavior in complex scenarios;

[0007] (3) Existing performance evaluation methods often ignore cross-departmental data collaborative analysis. Due to the existence of data silos, data from various departments cannot be effectively shared, making it difficult to conduct enterprise-level, global, and dynamic evaluations of employee performance.

[0008] (4) Existing methods usually lack an effective feedback mechanism and fail to dynamically respond to the deviation between actual performance feedback data and evaluation results, making it impossible to achieve continuous optimization and evolution of the performance evaluation method itself.

[0009] Therefore, how to provide an enterprise human resource performance evaluation method based on knowledge graph is an urgent problem that technicians in this field need to solve. Summary of the Invention

[0010] One purpose of the present invention is to propose a method for enterprise human resource performance evaluation based on knowledge graph. The present invention adopts technical means based on an improved Mamba architecture and a neural logic reasoning network, which effectively solves the problems in the existing technology such as strong subjectivity of performance evaluation, serious data silos, insufficient semantic mining capabilities, and lack of dynamic optimization feedback mechanism, and realizes objective, intelligent and real-time optimized enterprise human resource performance evaluation.

[0011] The enterprise human resources performance evaluation method based on knowledge graph according to an embodiment of the present invention includes the following steps:

[0012] S1. Collect employee daily work data, task collaboration data, process tracking data, enterprise knowledge base data, and external market dynamics data through multi-source heterogeneous data interfaces to obtain initial multimodal data;

[0013] S2. Clean and structure the initial multimodal data to obtain a structured data set;

[0014] S3. Based on the improved Mamba architecture, deep semantic features corresponding to employee performance are automatically extracted from structured datasets to obtain performance semantic feature representation;

[0015] S4. Build a multi-granularity logical reasoning knowledge graph based on enterprise performance evaluation rules, labor law compliance requirements, and industry standards. Learn and update the weights of logical reasoning rules in real time through a neural symbolic reasoning framework to obtain a dynamic logical rule set.

[0016] S5. Use a dynamic logic rule set to perform neural logic reasoning on the performance semantic feature representation, and verify the reasoning results based on a probability calibration mechanism to obtain the employee performance grade score;

[0017] S6. Through the federated knowledge distillation method, without directly transmitting the original departmental data, we conduct collaborative analysis of employee performance ratings across departments to obtain enterprise-level global dynamic performance evaluation results.

[0018] S7. Based on the performance evaluation results, build a performance evaluation feedback indicator system, automatically collect actual performance feedback data, calculate the deviation between the evaluation results and the actual performance feedback data, and optimize the enterprise human resources performance evaluation results based on the deviation.

[0019] Optionally, the S2 specifically includes:

[0020] S21, automatically classifying text data, numerical data, image data, audio data, and video data in the initial multimodal data, and marking modality categories;

[0021] S22. Based on the classified text data, automatically identify and delete redundant text information irrelevant to the performance evaluation based on a preset enterprise performance text feature keyword library. Then, based on preset text rules, delete incorrect characters, sensitive words, and redundant symbols in the text to obtain cleaned text data.

[0022] S23. For the classified numerical data, determine the estimated value of the missing value based on the data correlation, complete the missing data filling, and then automatically detect and eliminate abnormal numerical data using a preset numerical threshold range to obtain cleaned numerical data;

[0023] S24, for the classified image data, after removing image noise using a non-local mean filtering method, unifying the image size using an adaptive resizing method based on bilinear interpolation to obtain processed image data;

[0024] S25. For the classified audio data and video data, remove background noise and interference information from the audio data and video data using a spectral subtraction noise reduction method and a wavelet threshold-based noise reduction method, respectively, and cut the length of the audio and video data according to a preset unified time length standard to obtain processed audio data and video data;

[0025] S26, performing modal feature extraction on the processed text data, numerical data, image data, audio data, and video data respectively to obtain corresponding modal feature vectors;

[0026] S27. According to a preset feature vector mapping rule, the modal feature vector is mapped into a unified multimodal feature space to form a structured data set.

[0027] Optionally, the improved Mamba architecture specifically includes:

[0028] The text data, numerical data, image data, audio data and video data in the structured data set are respectively embedded with features to obtain a fixed-dimensional modal feature vector sequence;

[0029] Based on the modal eigenvector sequence, the cosine similarity between any two modal eigenvectors is calculated to construct the modal adaptive graph association matrix;

[0030] Update the state vector through the state space model dynamically adjusted by the modal adaptation graph incidence matrix;

[0031] Based on the state vector at time t, calculate the dynamic modal attention weight matrix;

[0032] Calculate the cross-modal fusion feature vector based on the dynamic attention weight matrix;

[0033] Based on the cross-modal fusion feature vector, long-term, medium-term and short-term state space models are constructed, and the output features of the state space models are calculated respectively;

[0034] After concatenating the feature vectors of the long-term, medium-term, and short-term scale output features, the layer normalization method is used to normalize the features, and a stable performance semantic feature representation is obtained through the residual connection method.

[0035] Optionally, the long-term, medium-term and short-term state space models specifically include:

[0036] Based on the cross-modal fusion feature vector and the fixed-scale state vector, nonlinear gating units at long-term, medium-term, and short-term scales are calculated;

[0037] Based on the modal adaptive graph correlation matrix, the dynamic adjustment parameter matrix of the state space model is defined for the long-term, medium-term and short-term scales respectively;

[0038] Based on the gated unit and the dynamic adjustment parameter matrix, the state vector is updated for the long-term, medium-term and short-term scales respectively;

[0039] The output features of the updated long-term, medium-term and short-term scale state vectors are calculated at fixed scales respectively.

[0040] Optionally, the S4 specifically includes:

[0041] S41. Based on the enterprise performance evaluation rules, labor law compliance requirements, and industry standard requirements, construct a multi-level knowledge node set centered on job responsibility nodes, performance task nodes, compliance standard nodes, and industry indicator nodes, and assign each node a unique identifier;

[0042] S42. Based on the knowledge node set, dynamically calculate the logical association strength between nodes as a dynamic relationship weight through the semantic vector similarity and co-occurrence frequency of node features;

[0043] S43. Based on the dynamic relationship weights, a dynamic logical path expression including a node relationship path and a weight sequence is generated, and based on the performance semantic feature representation, a training sample set is constructed, where each training sample is represented by a performance semantic feature representation and a corresponding dynamic logical path;

[0044] S44, using the neural symbolic reasoning network training sample set based on the dynamic path attention mechanism for initial training;

[0045] S45. Calculating a dynamic attention cross entropy loss function between the real-time predicted performance level probability and the actual performance level based on the real-time obtained performance semantic feature representation and the dynamic logic path representation;

[0046] S46. Use a gradient descent optimization algorithm with adaptive attention adjustment to update the logic rule weight vector, and use the updated logic rule weight vector as the dynamic logic rule set.

[0047] Optionally, the S5 specifically includes:

[0048] S51. Based on the performance semantic feature representation, the job adaptability features, task execution integrity features, and performance indicator correlation features are extracted respectively to construct a structured performance reasoning input feature vector;

[0049] S52. Using a dynamic logic rule set, based on the dynamic logic path and the real-time attention weight of the path, establish a real-time performance reasoning mapping relationship, and use the structured performance reasoning input feature vector as input to obtain an initial performance level probability distribution in real time;

[0050] S53, introducing a probability calibration temperature parameter, and using a nonlinear temperature transformation method to perform dynamic probability calibration on the initial performance level probability distribution, thereby obtaining a dynamically calibrated performance level probability distribution;

[0051] S54. Select the performance level with the largest probability value in the dynamically calibrated performance level probability distribution as the real-time preliminary performance level score, and record the probability value corresponding to this score in real time;

[0052] S55. Calculate in real time the entropy value of the performance grade probability distribution after dynamic calibration, and perform normalization on the entropy value as a performance grade scoring credibility indicator;

[0053] S56. Setting a performance rating credibility threshold. When the performance rating credibility indicator does not reach the threshold, adaptively adjusting the weight of the dynamic logic rule set according to the real-time calculated performance rating credibility indicator, and repeating steps S52 to S55 until the performance rating credibility indicator reaches or exceeds the credibility threshold.

[0054] S57. When the performance rating credibility index reaches or exceeds the credibility threshold, the corresponding performance rating is used as the employee's performance rating.

[0055] Optionally, the S6 specifically includes:

[0056] S61. Build an independent local performance reasoning model for each department within the enterprise. The local performance reasoning model generates employee performance rating scores within the department based on a dynamic logic rule set and performance semantic feature representation.

[0057] S62. Employee performance ratings generated by each department are encrypted locally using a homomorphic encryption algorithm. The encrypted data is not decrypted and does not leave the department's local location. Only the encrypted department performance ratings are uploaded to the enterprise-level federated analysis center.

[0058] S63. At the enterprise-level federated analysis center, based on a secure multi-party computing protocol, without decrypting the original department data, securely aggregate the received encrypted performance rating scores of each department to generate an aggregated result of the enterprise-level encrypted performance rating score.

[0059] S64. Build an enterprise-level global teacher model in the enterprise-level federated analysis center. Using the aggregated enterprise-level encrypted performance ratings as the guiding signal for knowledge distillation, dynamically calculate the difference between each department's local performance inference model and the enterprise-level global teacher model, and generate distillation adjustment parameters for each department.

[0060] S65: Adjust the parameters based on departmental distillation and distribute them securely to the corresponding departments through the secure multi-party computing protocol. Each department then autonomously updates the parameters of the local performance inference model using the distillation adjustment parameters.

[0061] S66. Generate a new encrypted performance rating score for the optimized departmental local performance reasoning model, and execute the secure aggregation and federated knowledge distillation process from steps S62 to S65. Through iterative cycles, continuously narrow the performance rating differences between departments.

[0062] S67. When the change in the performance rating difference between departments is less than the preset convergence threshold for two consecutive times, the iterative cycle is stopped, and the enterprise-level federal analysis center securely decrypts the aggregated enterprise-level encrypted performance rating to obtain the enterprise-level global dynamic performance evaluation results.

[0063] Optionally, the S7 specifically includes:

[0064] S71. Based on the results of enterprise-level global dynamic performance evaluation, construct a set of performance feedback indicators, including job competency matching, task achievement accuracy, execution process standardization, and performance continuity and stability, and set weights for each feedback indicator;

[0065] S72. Based on the enterprise's internal real-time data interface, automatically collect employees' actual work task completion data, job goal achievement data, process compliance review records, and periodic performance stability records to form a standardized actual performance feedback data set;

[0066] S73. Mapping the actual performance feedback data sets into quantified actual performance feedback scores one by one according to the performance feedback indicator set and the weight values, and weighting and summarizing the actual performance feedback scores according to the weight values ​​to obtain a comprehensive feedback score of the actual performance;

[0067] S74. Based on the performance feedback indicator set and weight values, weight the enterprise-level global dynamic performance evaluation results to obtain a comprehensive score of the evaluation performance, and calculate the absolute difference between the comprehensive score of the evaluation performance and the actual comprehensive performance feedback score to obtain the real-time deviation of the performance evaluation;

[0068] S75. Dynamically establish a deviation sensitivity parameter for the real-time deviation of the performance evaluation, and adjust the dynamic adjustment parameter matrix of the modal adaptive graph association matrix in real time according to the deviation sensitivity parameter;

[0069] S76. Based on the dynamically adjusted parameter matrix of the optimized modal adaptive graph association matrix, repeatedly execute the enterprise-level performance evaluation reasoning process, continuously update the enterprise-level global dynamic performance evaluation results, and recalculate the real-time deviation;

[0070] S77 , repeat the iterative optimization process from step S72 to step S76 until the change in the real-time deviation of the performance evaluation obtained by two consecutive iterative calculations is less than a preset convergence threshold.

[0071] The beneficial effects of the present invention are:

[0072] (1) The present invention constructs a data collection method based on a multi-source heterogeneous data interface and adopts an improved Mamba architecture to automatically extract deep performance semantic features, thereby achieving multi-dimensional and semantically accurate modeling of employee performance data, effectively improving the representation ability of performance data features and the accuracy of semantic mining, and enhancing the objectivity and reliability of performance evaluation results.

[0073] (2) By constructing a multi-granularity logical reasoning knowledge graph library and adopting a dynamic learning neural symbolic reasoning framework, the present invention can realize real-time logical reasoning of enterprise performance evaluation rules, labor law compliance requirements and industry standard requirements, significantly improving the accuracy and interpretability of performance rating scores, and showing better adaptability and effectiveness in the complex performance evaluation environment of enterprises.

[0074] (3) In terms of collaborative analysis of cross-departmental performance data, the present invention effectively solves the problem of the inability to securely collaborate on cross-departmental data analysis in the existing technology by combining the federated knowledge distillation method with secure multi-party computing and homomorphic encryption mechanism, breaks through the bottlenecks of internal data silos and privacy protection in the enterprise, and realizes the secure sharing and collaboration of enterprise-level global dynamic performance evaluation results, thereby effectively improving the comprehensiveness and dynamism of enterprise human resource performance evaluation.

[0075] (4) The present invention constructs a performance feedback indicator system and a real-time feedback data collection mechanism, dynamically calculates the deviation between the evaluation results and the actual performance feedback data, and dynamically optimizes the model parameters based on the real-time deviation, thereby realizing adaptive real-time optimization of the performance evaluation model, effectively reducing the performance evaluation error and significantly improving the accuracy and stability of the performance evaluation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0077] Figure 1 This is a schematic diagram of the overall process of the enterprise human resources performance evaluation method based on knowledge graph proposed in the present invention;

[0078] Figure 2 This is a schematic diagram of the structure of the improved Mamba architecture of the enterprise human resources performance evaluation method based on knowledge graph proposed in the present invention;

[0079] Figure 3 Schematic diagram of collaborative analysis of federated knowledge distillation data for the knowledge graph-based enterprise human resource performance evaluation method proposed in this invention. DETAILED DESCRIPTION

[0080] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0081] refer to Figure 1-Figure 3 ,The enterprise human resource performance evaluation method based on knowledge graph includes the following steps:

[0082] S1. Collect employee daily work data, task collaboration data, process tracking data, enterprise knowledge base data, and external market dynamics data through multi-source heterogeneous data interfaces to obtain initial multimodal data;

[0083] The daily work data of employees include daily attendance records, daily work logs, office system operation data, project completion progress and contribution ratio, and work quality inspection records; task collaboration data include the roles and task allocation of employees in team tasks, team task goals and completion status, team communication and collaboration process records, task progress tracking data and key node progress, and employee and team member interaction data; process tracking data includes the execution status of corporate business or project processes, milestone completion status, business approval efficiency, quality control data in process execution, and abnormal event records; corporate knowledge base data includes knowledge documents created or used by employees, internal training and learning outcomes data, employee knowledge contribution and sharing, knowledge management platform usage behavior, and knowledge usage evaluation data; external market dynamics data include performance benchmark data of competitors in the same industry, market talent supply and demand and salary level changes, industry standards and labor laws and policies changes, external training and skill requirements data, customer feedback and service quality evaluation information.

[0084] S2. Clean and structure the initial multimodal data to obtain a structured data set;

[0085] S3. Based on the improved Mamba architecture, deep semantic features corresponding to employee performance are automatically extracted from structured datasets to obtain performance semantic feature representation;

[0086] S4. Build a multi-granularity logical reasoning knowledge graph based on enterprise performance evaluation rules, labor law compliance requirements, and industry standards. Learn and update the weights of logical reasoning rules in real time through a neural symbolic reasoning framework to obtain a dynamic logical rule set.

[0087] S5. Use a dynamic logic rule set to perform neural logic reasoning on the performance semantic feature representation, and verify the reasoning results based on a probability calibration mechanism to obtain the employee performance grade score;

[0088] S6. Through the federated knowledge distillation method, without directly transmitting the original departmental data, we conduct collaborative analysis of employee performance ratings across departments to obtain enterprise-level global dynamic performance evaluation results.

[0089] S7. Based on the performance evaluation results, build a performance evaluation feedback indicator system, automatically collect actual performance feedback data, calculate the deviation between the evaluation results and the actual performance feedback data, and optimize the enterprise human resources performance evaluation results based on the deviation.

[0090] Through multi-source heterogeneous data interfaces, the system automatically collects multi-dimensional employee data. Through data cleaning and structuring, it generates a structured data set. The improved Mamba architecture is then used to automatically extract the semantic features of employee performance. Furthermore, it integrates enterprise performance rules, labor law compliance requirements, and industry standards to construct a multi-granularity logical reasoning knowledge graph. Neural symbolic reasoning is used to dynamically update logical rule weights and implement performance ratings. Federated knowledge distillation methods are employed to effectively resolve cross-departmental data silos and privacy protection issues, resulting in global dynamic performance evaluation results. Furthermore, the system dynamically optimizes the evaluation model based on a feedback indicator system to reduce evaluation bias. This effectively improves the accuracy, objectivity, and real-time nature of performance evaluations, enhancing the intelligence of enterprise performance management.

[0091] In this embodiment, S2 specifically includes:

[0092] S21, automatically classifying text data, numerical data, image data, audio data, and video data in the initial multimodal data, and marking modality categories;

[0093] S22. Based on the classified text data, automatically identify and delete redundant text information irrelevant to the performance evaluation based on a preset enterprise performance text feature keyword library. Then, based on preset text rules, delete incorrect characters, sensitive words, and redundant symbols in the text to obtain cleaned text data.

[0094] S23. For the classified numerical data, determine the estimated value of the missing value based on the data correlation, complete the missing data filling, and then automatically detect and eliminate abnormal numerical data using a preset numerical threshold range to obtain cleaned numerical data;

[0095] S24, for the classified image data, after removing image noise using a non-local mean filtering method, unifying the image size using an adaptive resizing method based on bilinear interpolation to obtain processed image data;

[0096] S25. For the classified audio data and video data, remove background noise and interference information from the audio data and video data using a spectral subtraction noise reduction method and a wavelet threshold-based noise reduction method, respectively, and cut the length of the audio and video data according to a preset unified time length standard to obtain processed audio data and video data;

[0097] S26, performing modal feature extraction on the processed text data, numerical data, image data, audio data, and video data respectively to obtain corresponding modal feature vectors;

[0098] S27. According to a preset feature vector mapping rule, the modal feature vector is mapped into a unified multimodal feature space to form a structured data set.

[0099] By automatically classifying and labeling the initial multimodal data, and adopting targeted and refined cleaning and structured processing methods for text, numerical, image, audio and video data, the quality of the initial dataset was significantly improved. By automatically screening and cleaning text data based on the enterprise performance text feature keyword library and rule library, the data content related to performance evaluation was retained. At the same time, the non-local mean filtering method and bilinear interpolation image processing technology as well as spectral subtraction and wavelet threshold denoising methods were used to effectively remove interference information in audio and video data, ensuring high fidelity after data processing. Finally, through feature extraction and feature vector mapping, a unified structured dataset was formed.

[0100] In this embodiment, the improved Mamba architecture specifically includes:

[0101] The text data, numerical data, image data, audio data and video data in the structured data set are respectively embedded with features to obtain a fixed-dimensional modal feature vector sequence;

[0102] Based on the modal eigenvector sequence, calculate any two modal eigenvectors and The cosine similarity between them is used to construct the modal adaptive graph association matrix G t :

[0103]

[0104] in, are the feature vectors of the mth mode and the nth mode at the tth moment respectively; N is the total number of modes, specifically 5, representing the five modes of text, numerical, image, audio and video data; the numerator uses the cosine similarity of the two modal feature vectors, specifically indicating the degree of feature similarity between the two modalities at the current moment, and the denominator represents the sum of the cosine similarities between the current mode and all other modal feature vectors, which is used for normalization processing to obtain the relative weight value, modal adaptive graph association matrix The dynamic correlation strength of the mth modality relative to the nth modality at time t is expressed. The modality-adaptive graph association matrix dynamically quantifies and expresses the semantic associations between modalities in real time, effectively enabling adaptive dynamic fusion of multimodal data and improving the model's ability to accurately capture the semantic associations between features from different modalities.

[0105] By the modal adaptive graph incidence matrix G t Dynamically adjusted state space model, update state vector X t+1 :

[0106] X t+1 =(A+α·G t )Xt +(B+β·G t )U t ;

[0107] Among them, X t+1 is the state vector at time t+1; A and B are the initial parameter matrices of the state space model, describing the basic relationship between state transition and input features acting on state change; parameters α and β are learnable scalar parameters used to dynamically adjust the modal adaptive graph association matrix G t The degree of influence on the state space model is set to 0.1 initially and can be dynamically optimized according to real-time data; G t is the modal adaptive graph association matrix at the tth moment, which represents the dynamic association strength between the modes in real time; X t is the state vector at time t, representing the state characteristics of the current system; U t The input feature vector at time t represents the impact of external input on state changes at that moment. By introducing a modal adaptive graph association matrix to dynamically adjust the state space model, real-time adaptive adjustment of state transition parameters is achieved. This effectively enhances the model's sensitivity and accuracy to the dynamic associations between modal features, improving the state model's dynamic adaptability and its ability to capture semantic associations in multimodal data.

[0108] Based on the state vector X at time t t , calculate the dynamic modal attention weight matrix W t :

[0109]

[0110] in, The dynamic attention weight of the mth modal relative to the nth modal at time t represents the degree of dynamic attention between modalities. d is the dimension of the modal feature vector, which is used to normalize the attention weight calculation process, prevent numerical extremes, and improve computational stability. This dynamic attention mechanism captures the interdependencies between modal features in real time and distributes attention to each modal feature, effectively improving the accuracy and performance of multimodal feature fusion.

[0111] Based on the dynamic attention weight matrix W t , calculate the cross-modal fusion feature vector

[0112]

[0113] in, The fused cross-modal feature vector at time t reflects the comprehensive semantic feature information of multiple modalities at the current moment. This approach achieves dynamic fusion of cross-modal features, accurately capturing the real-time correlation between modalities using a dynamic attention weight matrix, and effectively extracting the collaborative semantic information of each modality at a specific moment, improving the accuracy and expressiveness of multimodal data fusion.

[0114] Based on the cross-modal fusion feature vector, long-term, medium-term and short-term state space models are constructed, and the output features of the state space models are calculated respectively;

[0115] After concatenating the feature vectors of the long-term, medium-term, and short-term scale output features, the layer normalization method is used to normalize the features, and a stable performance semantic feature representation is obtained through the residual connection method.

[0116] By constructing a modal adaptive graph association matrix for multimodal feature vector sequences, dynamic modeling of semantic associations between modalities is achieved, and the association matrix is ​​used to adjust the state update process of the state-space model in real time. At the same time, a dynamic modal attention weight matrix is ​​introduced to achieve efficient aggregation of cross-modal fusion features, so that the improved Mamba architecture can effectively capture and fully integrate the deep semantic features of data from different modalities. In addition, by combining long-term, medium-term and short-term state-space models, the dynamic characteristics of employee performance at different time scales are effectively captured, ensuring the comprehensiveness and reliability of feature extraction, and significantly improving the accuracy and robustness of corporate human resource performance evaluation.

[0117] In this embodiment, the long-term, medium-term and short-term state space models specifically include:

[0118] Based on cross-modal fusion feature vector With a fixed-scale state vector Computing nonlinear gated units at long, medium, and short time scales

[0119]

[0120] in, is a nonlinear gating unit at the long-term, medium-term, or short-term scale at time t, which is used to dynamically adjust the contribution of cross-modal fusion features and the state vector; the function σ is a Sigmoid nonlinear activation function, which limits the gated output value between 0 and 1 and clearly distinguishes the retention and update ratio of state information; is the weight matrix of the gate unit, is the bias vector of the gating unit. Both are learnable parameters that determine the impact of input features on the gating mechanism. Through the gating mechanism, the contribution ratio of cross-modal fusion features and fixed-scale state vectors at a specific moment is adaptively determined, achieving precise control and dynamic balance of state information and fusion features at different time scales, effectively improving the selectivity and stability of the model's state information at long-term, medium-term and short-term scales.

[0121] Based on the modal adaptive graph incidence matrix G t , define the dynamic adjustment parameter matrix of the state space model for the long-term, medium-term and short-term scales respectively:

[0122]

[0123] Among them, the matrix and are respectively the dynamic adjustment parameter matrices of the long-term, medium-term or short-term scale state space model at time t, which are used to adjust the state transition process and the degree of effect of input features in real time; A (scale) and B (scale) The initial parameter matrices for the long-term, medium-term, and short-term state-space models, respectively, define the initial transfer characteristics of the state-space model. Learnable scalar parameters α and β control the influence of the modal adaptive graph association matrix on the adjustment of state-space model parameters. Both are initially set to 0.1 and can be optimized as the model trains. By dynamically adjusting the state-space model parameters at the long-term, medium-term, and short-term scales in real time, the model becomes more sensitive to the semantic associations between modal data and offers real-time adaptability, thereby enhancing the state-space model's ability to capture dynamic performance characteristics at different time scales.

[0124] Based on gate control unit and the dynamic adjustment parameter matrix The state vector is calculated for the long-term, medium-term and short-term scales respectively. Update:

[0125]

[0126] Among them, the operator ⊙ represents the element-by-element multiplication operation, ensuring that the gating mechanism acts on each state feature component; the function tanh is a nonlinear activation function, which converts the state information to the [-1,1] interval, effectively enhancing the nonlinear expression ability of the model; U t is the input feature vector at time t, representing the real-time effect of external input on state updates. By adjusting the state vector update process in real time through a gating mechanism and a dynamic parameter matrix, a precise dynamic balance is achieved between state information and input feature information at different scales, effectively improving the ability to capture and update state features at long-term, medium-term, and short-term scales.

[0127] The updated long-term, medium-term and short-term scale state vectors Fixed-scale output features calculate:

[0128]

[0129] Among them, the matrix C (scale) The state output parameter matrix of the long-term, medium-term and short-term scale state space models is used to map the state vector to the output feature; the matrix D (scale) The input feature output parameter matrix of the state-space model at the corresponding scale reflects the contribution of external input to the output feature vector. Through the state-output parameter matrix and the input feature output parameter matrix, the state vector and input features of the state-space model are explicitly mapped to the output feature space, achieving the effective fusion and output of state information and external input information at different time scales, improving the accuracy and stability of the model output features at long-term, medium-term, and short-term scales.

[0130] By introducing nonlinear gating units at long-term, medium-term and short-term scales, effective modeling of state characteristics at different time scales is achieved; at the same time, the modal adaptive graph association matrix is ​​used to perform real-time optimization of the dynamic adjustment parameter matrix of the state space model, significantly improving the sensitivity and dynamic adaptability of the state model to the correlation relationship of multimodal data; further, through the nonlinear fusion of long-term, medium-term and short-term scale state vectors, the dynamic characteristics of employee performance data at different scales are effectively captured and integrated, enhancing the overall stability and accuracy of the model; in addition, the fixed-scale output feature calculation method ensures the interpretability and repeatability of the performance evaluation output results, significantly improving the accuracy and reliability of enterprise human resource performance evaluation.

[0131] In this embodiment, the S4 specifically includes:

[0132] S41. Based on the enterprise performance evaluation rules, labor law compliance requirements, and industry standard requirements, construct a multi-level knowledge node set centered on job responsibility nodes, performance task nodes, compliance standard nodes, and industry indicator nodes, and assign each node a unique identifier;

[0133] S42. Based on the knowledge node set, dynamically calculate the logical association strength between nodes as a dynamic relationship weight through the semantic vector similarity and co-occurrence frequency of node features;

[0134] S43. Based on the dynamic relationship weights, a dynamic logical path expression including a node relationship path and a weight sequence is generated, and based on the performance semantic feature representation, a training sample set is constructed, where each training sample is represented by a performance semantic feature representation and a corresponding dynamic logical path;

[0135] S44, using the neural symbolic reasoning network training sample set based on the dynamic path attention mechanism for initial training;

[0136] S45. Calculating a dynamic attention cross entropy loss function between the real-time predicted performance level probability and the actual performance level based on the real-time obtained performance semantic feature representation and the dynamic logic path representation;

[0137] S46. Use a gradient descent optimization algorithm with adaptive attention adjustment to update the logic rule weight vector, and use the updated logic rule weight vector as the dynamic logic rule set.

[0138] By establishing a multi-level knowledge node set based on corporate performance evaluation rules, labor law compliance requirements and industry standard requirements, and using semantic vector similarity and co-occurrence frequency to calculate the dynamic relationship weights between nodes, accurate dynamic modeling of the semantic logical relationship between knowledge nodes is achieved; at the same time, the neural symbolic reasoning network with dynamic path attention mechanism is used to dynamically update the logical rule weights in real time, thereby improving the adaptability and accuracy of performance reasoning rules; in addition, through the dynamic attention cross-entropy loss function and adaptive attention adjustment gradient descent optimization method, the real-time and reliability of performance reasoning are effectively improved, and the accuracy, robustness and dynamic adaptability of corporate human resource performance evaluation are significantly enhanced.

[0139] In this embodiment, the S5 specifically includes:

[0140] S51. Based on the performance semantic feature representation, the job adaptability features, task execution integrity features, and performance indicator correlation features are extracted respectively to construct a structured performance reasoning input feature vector;

[0141] S52. Using a dynamic logic rule set, based on the dynamic logic path and the real-time attention weight of the path, establish a real-time performance reasoning mapping relationship, and use the structured performance reasoning input feature vector as input to obtain an initial performance level probability distribution in real time;

[0142] S53, introducing a probability calibration temperature parameter, and using a nonlinear temperature transformation method to perform dynamic probability calibration on the initial performance level probability distribution, thereby obtaining a dynamically calibrated performance level probability distribution;

[0143] S54. Select the performance level with the largest probability value in the dynamically calibrated performance level probability distribution as the real-time preliminary performance level score, and record the probability value corresponding to this score in real time;

[0144] S55. Calculate in real time the entropy value of the performance grade probability distribution after dynamic calibration, and perform normalization on the entropy value as a performance grade scoring credibility indicator;

[0145] S56. Setting a performance rating credibility threshold. When the performance rating credibility indicator does not reach the threshold, adaptively adjusting the weight of the dynamic logic rule set according to the real-time calculated performance rating credibility indicator, and repeating steps S52 to S55 until the performance rating credibility indicator reaches or exceeds the credibility threshold.

[0146] S57. When the performance rating credibility index reaches or exceeds the credibility threshold, the corresponding performance rating is used as the employee's performance rating.

[0147] By establishing a structured performance reasoning input feature vector and utilizing a dynamic logic rule set and a real-time attention weight mechanism to obtain the initial probability distribution of the performance grade, the nonlinear probability calibration temperature transformation method is used to achieve dynamic calibration of the probability distribution, thereby improving the accuracy of the performance score. At the same time, based on the dynamically calibrated probability distribution, the entropy value is calculated in real time as a credibility indicator, and the logic rule weight is adaptively adjusted according to the credibility indicator, thereby enhancing the robustness and reliability of the performance grade score. In addition, by setting an iterative optimization strategy for the credibility threshold, the high credibility and stability of the final performance grade score are dynamically guaranteed, effectively solving the deficiency in the existing technology that the performance score is easily disturbed by uncertainty, thereby improving the accuracy and credibility of the enterprise performance evaluation.

[0148] In this embodiment, S6 specifically includes:

[0149] S61. Build an independent local performance reasoning model for each department within the enterprise. The local performance reasoning model generates employee performance rating scores within the department based on a dynamic logic rule set and performance semantic feature representation.

[0150] S62. Employee performance ratings generated by each department are encrypted locally using a homomorphic encryption algorithm. The encrypted data is not decrypted and does not leave the department's local location. Only the encrypted department performance ratings are uploaded to the enterprise-level federated analysis center.

[0151] S63. At the enterprise-level federated analysis center, based on a secure multi-party computing protocol, without decrypting the original department data, securely aggregate the received encrypted performance rating scores of each department to generate an aggregated result of the enterprise-level encrypted performance rating score.

[0152] S64. Build an enterprise-level global teacher model in the enterprise-level federated analysis center. Using the aggregated enterprise-level encrypted performance ratings as the guiding signal for knowledge distillation, dynamically calculate the difference between each department's local performance inference model and the enterprise-level global teacher model, and generate distillation adjustment parameters for each department.

[0153] S65: Adjust the parameters based on departmental distillation and distribute them securely to the corresponding departments through the secure multi-party computing protocol. Each department then autonomously updates the parameters of the local performance inference model using the distillation adjustment parameters.

[0154] S66. Generate a new encrypted performance rating score for the optimized departmental local performance reasoning model, and execute the secure aggregation and federated knowledge distillation process from steps S62 to S65. Through iterative cycles, continuously narrow the performance rating differences between departments.

[0155] S67. When the change in the performance rating difference between departments is less than the preset convergence threshold for two consecutive times, the iterative cycle is stopped, and the enterprise-level federal analysis center securely decrypts the aggregated enterprise-level encrypted performance rating to obtain the enterprise-level global dynamic performance evaluation results.

[0156] By building local performance inference models in each department within the enterprise, using homomorphic encryption algorithms to locally encrypt department data and securely upload it to the enterprise-level federated analysis center, the security and privacy of department performance data are ensured; at the same time, based on the secure multi-party computing protocol, the secure aggregation of enterprise-level performance data is achieved, and the model parameters of each department are dynamically adjusted through the federated knowledge distillation method, which effectively solves the problem of difficulties in cross-departmental data collaborative analysis within the enterprise; further, based on the dynamically adjusted model parameters, the performance evaluation models of each department are continuously optimized, realizing global dynamic collaborative analysis and unified evaluation of enterprise performance data, and significantly improving the objectivity, comprehensiveness and security of enterprise human resource performance evaluation.

[0157] In this embodiment, the S7 specifically includes:

[0158] S71. Based on the results of enterprise-level global dynamic performance evaluation, construct a set of performance feedback indicators, including job competency matching, task achievement accuracy, execution process standardization, and performance continuity and stability, and set weights for each feedback indicator;

[0159] S72. Based on the enterprise's internal real-time data interface, automatically collect employees' actual work task completion data, job goal achievement data, process compliance review records, and periodic performance stability records to form a standardized actual performance feedback data set;

[0160] S73. Mapping the actual performance feedback data sets into quantified actual performance feedback scores one by one according to the performance feedback indicator set and the weight values, and weighting and summarizing the actual performance feedback scores according to the weight values ​​to obtain a comprehensive feedback score of the actual performance;

[0161] S74. Based on the performance feedback indicator set and weight values, weight the enterprise-level global dynamic performance evaluation results to obtain a comprehensive score of the evaluation performance, and calculate the absolute difference between the comprehensive score of the evaluation performance and the actual comprehensive performance feedback score to obtain the real-time deviation of the performance evaluation;

[0162] S75. Dynamically establish a deviation sensitivity parameter for the real-time deviation of the performance evaluation, and adjust the dynamic adjustment parameter matrix of the modal adaptive graph association matrix in real time according to the deviation sensitivity parameter;

[0163] S76. Based on the dynamically adjusted parameter matrix of the optimized modal adaptive graph association matrix, repeatedly execute the enterprise-level performance evaluation reasoning process, continuously update the enterprise-level global dynamic performance evaluation results, and recalculate the real-time deviation;

[0164] S77 , repeat the iterative optimization process from step S72 to step S76 until the change in the real-time deviation of the performance evaluation obtained by two consecutive iterative calculations is less than a preset convergence threshold.

[0165] By constructing a set of feedback indicators for enterprise-level global dynamic performance evaluation results, and automatically collecting, mapping and summarizing actual performance feedback data based on indicator weights, a comprehensive feedback score is formed; by calculating the deviation between the performance evaluation results and the actual feedback score in real time, and dynamically establishing deviation sensitivity parameters, real-time optimization and adjustment of the dynamic adjustment parameter matrix of the modal adaptive graph association matrix is ​​achieved; the optimized dynamic adjustment parameter matrix is ​​used to continuously update the enterprise performance evaluation reasoning process, forming a feedback closed loop until the deviation change amplitude converges to the preset threshold, effectively improving the accuracy and stability of the performance evaluation method, and significantly improving the real-time dynamic adaptability of the enterprise human resources performance evaluation system.

[0166] Example 1:

[0167] In order to verify the feasibility of the present invention in implementation, the present invention is applied to the human resources performance management scenario of a large technology enterprise group. The enterprise group is large in scale, with tens of thousands of employees, multiple departments, and multiple business segments. It faces heavy and complex performance appraisal tasks every year. The traditional performance evaluation method adopts a model of manually setting key performance indicators and regularly manually calculating scores. There is obvious subjectivity, and data sharing between departments is difficult, resulting in a serious data island phenomenon, which makes the evaluation results unable to accurately and objectively reflect the actual performance of employees and the performance level differences between departments. In response to the above problems, the company hopes to explore an objective, fair and dynamically optimized performance evaluation solution to improve the efficiency and accuracy of performance management.

[0168] During implementation, the multi-source heterogeneous data interface solution proposed in this invention was first used to build an internal data collection platform. This platform automatically collected multiple data sources, including employee daily work records, task collaboration information, process tracking logs, enterprise knowledge base interaction data, and external market dynamics data. This platform quickly formed a structured multimodal dataset and automatically classified and fine-tuned text, numerical, image, audio, and video data. For example, image data processed using non-local mean filtering and bilinear interpolation achieved an average denoising rate of over 95%, effectively preserving key details. Audio and video data processed using spectral subtraction and wavelet threshold denoising achieved an average noise reduction rate of 92%.

[0169] Subsequently, by implementing an improved Mamba architecture and a modal adaptive graph association matrix, the system automatically modeled complex semantic relationships between data from different modalities, capable of automatically extracting deep semantic features related to performance. Subsequently, based on corporate performance evaluation rules, labor law compliance requirements, and industry standards, a multi-granularity logical reasoning knowledge graph was constructed. The knowledge graph's logical rules were dynamically updated in real time using a dynamic path attention mechanism.

[0170] In the specific performance reasoning process, the system uses a dynamic logic rule set to perform neural logic reasoning on the extracted performance semantic features, and adopts a probability calibration mechanism for dynamic probability calibration to obtain real-time performance ratings. To achieve cross-departmental data collaborative analysis, the enterprise group implemented federated knowledge distillation technology. Each department established a local performance reasoning model and securely shared departmental performance data through homomorphic encryption and secure multi-party computing protocols. The enterprise federated analysis center established a global teacher model to perform knowledge distillation and parameter optimization on each department's local model. In actual application, the secure transmission and aggregate calculation of each department's performance data takes an average of less than 120 milliseconds, significantly lower than the data interaction delay of more than 300 milliseconds of traditional methods.

[0171] To further improve the real-time and accuracy of performance evaluation, the system has designed a set of performance feedback indicators, including job competency matching, task achievement accuracy, execution process standardization and performance continuity and stability, and set corresponding indicator weights. It automatically collects actual performance feedback data, calculates the real-time deviation between the evaluation results and the actual performance data, and dynamically optimizes the parameters of the modal adaptive graph association matrix based on the deviation sensitivity parameters to continuously improve the evaluation accuracy.

[0172] Through three months of actual operation, the performance evaluation results of the traditional performance evaluation method and the method proposed in this invention were compared, and the significant advantages of the present invention were verified. The specific performance evaluation comparison data are shown in Table 1:

[0173] Table 1 Comparison of performance evaluation results between traditional method and the method of the present invention

[0174]

[0175] As can be seen from the data in Table 1, the proposed method significantly reduces scoring errors compared to traditional performance evaluation methods. The scoring error rate for job competency match decreased from 16.7% to 4.5%, a 73.1% reduction. The scoring error rate for task achievement accuracy decreased from 14.3% to 3.8%, a 73.4% reduction. The scoring error rate for execution process standardization decreased from 18.9% to 5.2%, a 72.5% reduction. The scoring error rate for performance continuity and stability decreased from 20.5% to 6.0%, a 70.7% reduction. The overall comprehensive performance evaluation scoring error decreased from 17.6% to 4.9%, an overall improvement of 72.2%, significantly improving the accuracy and objectivity of performance evaluation results.

[0176] In addition, in terms of cross-departmental collaborative analysis, the federated knowledge distillation method proposed in this invention effectively solves the problem of data silos within the enterprise. The coverage rate of cross-departmental data collaborative analysis has been increased from 45% of traditional methods to more than 95%. At the same time, the level of data security protection has been significantly improved, and no data leakage or privacy issues have occurred.

[0177] Furthermore, the method of this invention enables real-time optimization feedback for performance evaluation, significantly improving the model's adaptability. Traditionally, performance evaluation rules are manually revised quarterly, taking over 10 working days. However, this invention utilizes an automatic feedback optimization mechanism, which automatically adjusts model parameters in real time. Each optimization adjustment is reduced to under 30 minutes, keeping the real-time evaluation error below 5%, and improving the efficiency of enterprise performance evaluation by over 90%.

[0178] It can be clearly seen from the above implementation cases that the present invention effectively overcomes the problems of subjectivity, data silos and insufficient dynamic adjustment in traditional performance evaluation methods by introducing multi-source heterogeneous data collection, improved Mamba architecture, dynamic logical reasoning knowledge graph library, federated knowledge distillation technology and real-time feedback optimization mechanism, and realizes the objectivity, accuracy and real-time dynamism of performance evaluation, providing strong technical support for enterprise performance management.

[0179] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. The enterprise human resource performance evaluation method based on knowledge graph is characterized by: The steps include: S1. Collect employee daily work data, task collaboration data, process tracking data, enterprise knowledge base data, and external market dynamics data through multi-source heterogeneous data interfaces to obtain initial multimodal data; S2. Clean and structure the initial multimodal data to obtain a structured data set; S3. Based on the improved Mamba architecture, deep semantic features corresponding to employee performance are automatically extracted from structured datasets to obtain performance semantic feature representation; S4. Build a multi-granularity logical reasoning knowledge graph based on enterprise performance evaluation rules, labor law compliance requirements, and industry standards. Learn and update the weights of logical reasoning rules in real time through a neural symbolic reasoning framework to obtain a dynamic logical rule set. S5. Use a dynamic logic rule set to perform neural logic reasoning on the performance semantic feature representation, and verify the reasoning results based on a probability calibration mechanism to obtain the employee performance grade score; S6. Through the federated knowledge distillation method, without directly transmitting the original departmental data, we conduct collaborative analysis of employee performance ratings across departments to obtain enterprise-level global dynamic performance evaluation results. S7. Based on the performance evaluation results, build a performance evaluation feedback indicator system, automatically collect actual performance feedback data, calculate the deviation between the evaluation results and the actual performance feedback data, and optimize the enterprise human resources performance evaluation results based on the deviation.

2. The enterprise human resource performance evaluation method based on knowledge graph according to claim 1 is characterized in that: The S2 specifically includes: S21, automatically classifying text data, numerical data, image data, audio data, and video data in the initial multimodal data, and marking modality categories; S22. Based on the classified text data, automatically identify and delete redundant text information irrelevant to the performance evaluation based on a preset enterprise performance text feature keyword library. Then, based on preset text rules, delete incorrect characters, sensitive words, and redundant symbols in the text to obtain cleaned text data. S23. For the classified numerical data, determine the estimated value of the missing value based on the data correlation, complete the missing data filling, and then automatically detect and eliminate abnormal numerical data using a preset numerical threshold range to obtain cleaned numerical data; S24, for the classified image data, after removing image noise using a non-local mean filtering method, unifying the image size using an adaptive resizing method based on bilinear interpolation to obtain processed image data; S25. For the classified audio data and video data, remove background noise and interference information from the audio data and video data using a spectral subtraction noise reduction method and a wavelet threshold-based noise reduction method, respectively, and cut the length of the audio and video data according to a preset unified time length standard to obtain processed audio data and video data; S26, performing modal feature extraction on the processed text data, numerical data, image data, audio data, and video data respectively to obtain corresponding modal feature vectors; S27. According to a preset feature vector mapping rule, the modal feature vector is mapped into a unified multimodal feature space to form a structured data set.

3. The enterprise human resource performance evaluation method based on knowledge graph according to claim 1 is characterized in that: The improved Mamba architecture specifically includes: The text data, numerical data, image data, audio data and video data in the structured data set are respectively embedded with features to obtain a fixed-dimensional modal feature vector sequence; Based on the modal eigenvector sequence, the cosine similarity between any two modal eigenvectors is calculated to construct the modal adaptive graph association matrix; Update the state vector through the state space model dynamically adjusted by the modal adaptation graph incidence matrix; Based on the state vector at time t, calculate the dynamic modal attention weight matrix; Calculate the cross-modal fusion feature vector based on the dynamic attention weight matrix; Based on the cross-modal fusion feature vector, long-term, medium-term and short-term state space models are constructed, and the output features of the state space models are calculated respectively; After concatenating the feature vectors of the long-term, medium-term, and short-term scale output features, the layer normalization method is used to normalize the features, and a stable performance semantic feature representation is obtained through the residual connection method.

4. The enterprise human resource performance evaluation method based on knowledge graph according to claim 3 is characterized in that: The long-term, medium-term and short-term state space models specifically include: Based on the cross-modal fusion feature vector and the fixed-scale state vector, nonlinear gating units at long-term, medium-term, and short-term scales are calculated; Based on the modal adaptive graph correlation matrix, the dynamic adjustment parameter matrix of the state space model is defined for the long-term, medium-term and short-term scales respectively; Based on the gated unit and the dynamic adjustment parameter matrix, the state vector is updated for the long-term, medium-term and short-term scales respectively; The output features of the updated long-term, medium-term and short-term scale state vectors are calculated at fixed scales respectively.

5. The enterprise human resource performance evaluation method based on knowledge graph according to claim 1 is characterized in that: The S4 specifically includes: S41. Based on the enterprise performance evaluation rules, labor law compliance requirements, and industry standard requirements, construct a multi-level knowledge node set centered on job responsibility nodes, performance task nodes, compliance standard nodes, and industry indicator nodes, and assign each node a unique identifier; S42. Based on the knowledge node set, dynamically calculate the logical association strength between nodes as a dynamic relationship weight through the semantic vector similarity and co-occurrence frequency of node features; S43. Based on the dynamic relationship weights, a dynamic logical path expression including a node relationship path and a weight sequence is generated, and based on the performance semantic feature representation, a training sample set is constructed, where each training sample is represented by a performance semantic feature representation and a corresponding dynamic logical path; S44, using the neural symbolic reasoning network training sample set based on the dynamic path attention mechanism for initial training; S45. Calculating a dynamic attention cross entropy loss function between the real-time predicted performance level probability and the actual performance level based on the real-time obtained performance semantic feature representation and the dynamic logic path representation; S46. Use a gradient descent optimization algorithm with adaptive attention adjustment to update the logic rule weight vector, and use the updated logic rule weight vector as the dynamic logic rule set.

6. The enterprise human resource performance evaluation method based on knowledge graph according to claim 1 is characterized in that: The S5 specifically includes: S51. Based on the performance semantic feature representation, the job adaptability features, task execution integrity features, and performance indicator correlation features are extracted respectively to construct a structured performance reasoning input feature vector; S52. Using a dynamic logic rule set, based on the dynamic logic path and the real-time attention weight of the path, establish a real-time performance reasoning mapping relationship, and use the structured performance reasoning input feature vector as input to obtain an initial performance level probability distribution in real time; S53, introducing a probability calibration temperature parameter, and using a nonlinear temperature transformation method to perform dynamic probability calibration on the initial performance level probability distribution, thereby obtaining a dynamically calibrated performance level probability distribution; S54. Select the performance level with the largest probability value in the dynamically calibrated performance level probability distribution as the real-time preliminary performance level score, and record the probability value corresponding to this score in real time; S55. Calculate in real time the entropy value of the performance grade probability distribution after dynamic calibration, and perform normalization on the entropy value as a performance grade scoring credibility indicator; S56. Setting a performance rating credibility threshold. When the performance rating credibility indicator does not reach the threshold, adaptively adjusting the weight of the dynamic logic rule set according to the real-time calculated performance rating credibility indicator, and repeating steps S52 to S55 until the performance rating credibility indicator reaches or exceeds the credibility threshold. S57. When the performance rating credibility index reaches or exceeds the credibility threshold, the corresponding performance rating is used as the employee's performance rating.

7. The enterprise human resource performance evaluation method based on knowledge graph according to claim 1 is characterized in that: The S6 specifically includes: S61. Build an independent local performance reasoning model for each department within the enterprise. The local performance reasoning model generates employee performance rating scores within the department based on a dynamic logic rule set and performance semantic feature representation. S62. Employee performance ratings generated by each department are encrypted locally using a homomorphic encryption algorithm. The encrypted data is not decrypted and does not leave the department's local location. Only the encrypted department performance ratings are uploaded to the enterprise-level federated analysis center. S63. At the enterprise-level federated analysis center, based on a secure multi-party computing protocol, without decrypting the original department data, securely aggregate the received encrypted performance rating scores of each department to generate an aggregated result of the enterprise-level encrypted performance rating score. S64. Build an enterprise-level global teacher model in the enterprise-level federated analysis center. Using the aggregated enterprise-level encrypted performance ratings as the guiding signal for knowledge distillation, dynamically calculate the difference between each department's local performance inference model and the enterprise-level global teacher model, and generate distillation adjustment parameters for each department. S65: Adjust the parameters based on departmental distillation and distribute them securely to the corresponding departments through the secure multi-party computing protocol. Each department then autonomously updates the parameters of the local performance inference model using the distillation adjustment parameters. S66. Generate a new encrypted performance rating score for the optimized departmental local performance reasoning model, and execute the secure aggregation and federated knowledge distillation process from steps S62 to S65. Through iterative cycles, continuously narrow the performance rating differences between departments. S67. When the change in the performance rating difference between departments is less than the preset convergence threshold for two consecutive times, the iterative cycle is stopped, and the enterprise-level federal analysis center securely decrypts the aggregated enterprise-level encrypted performance rating to obtain the enterprise-level global dynamic performance evaluation results.

8. The enterprise human resource performance evaluation method based on knowledge graph according to claim 1 is characterized in that: The S7 specifically includes: S71. Based on the results of enterprise-level global dynamic performance evaluation, construct a set of performance feedback indicators, including job competency matching, task achievement accuracy, execution process standardization, and performance continuity and stability, and set weights for each feedback indicator; S72. Based on the enterprise's internal real-time data interface, automatically collect employees' actual work task completion data, job goal achievement data, process compliance review records, and periodic performance stability records to form a standardized actual performance feedback data set; S73. Mapping the actual performance feedback data sets into quantified actual performance feedback scores one by one according to the performance feedback indicator set and the weight values, and weighting and summarizing the actual performance feedback scores according to the weight values ​​to obtain a comprehensive feedback score of the actual performance; S74. Based on the performance feedback indicator set and weight values, weight the enterprise-level global dynamic performance evaluation results to obtain a comprehensive score of the evaluation performance, and calculate the absolute difference between the comprehensive score of the evaluation performance and the actual comprehensive performance feedback score to obtain the real-time deviation of the performance evaluation; S75. Dynamically establish a deviation sensitivity parameter for the real-time deviation of the performance evaluation, and adjust the dynamic adjustment parameter matrix of the modal adaptive graph association matrix in real time according to the deviation sensitivity parameter; S76. Based on the dynamically adjusted parameter matrix of the optimized modal adaptive graph association matrix, repeatedly execute the enterprise-level performance evaluation reasoning process, continuously update the enterprise-level global dynamic performance evaluation results, and recalculate the real-time deviation; S77 , repeat the iterative optimization process from step S72 to step S76 until the change in the real-time deviation of the performance evaluation obtained by two consecutive iterative calculations is less than a preset convergence threshold.

Images