Employee ability assessment method and system based on structure attribution neural network, and medium

Abstract

The invention discloses an employee ability assessment method and system based on a structure attribution neural network, and a medium, and belongs to the field of power systems, and the method comprises the steps: extracting the partial derivative information of an ability assessment model based on a gradient back propagation mechanism, determining the gradient sensitivity based on each partial derivative information, and determining a weight parameter; processing the real-time evaluation event according to the weight parameter to obtain a target evaluation event, and performing matrix processing and division according to the index service dimension to obtain a plurality of sub-matrixes, so that the capability evaluation model splices feature extraction results output after extraction of the plurality of structure sub-networks for each sub-matrix into an intermediate vector, the intermediate vector is predicted through the summary regression network, a target evaluation result is output, the structure dimension of each structure sub-network is in one-to-one correspondence with the service dimension, and the model adopts a loss function based on attribution guidance, so that objective and accurate evaluation of the ability of the to-be-evaluated employee can be realized through implementation of the method and the system.

Description

Technical Field

[0001] This invention relates to the field of power systems, and in particular to a method, system, and medium for employee competence assessment based on structural attribution neural networks. Background Technology

[0002] With the continuous advancement of the construction of new power systems, the power grid system is placing higher demands on the comprehensive capabilities of its employees. For example, employees need to possess interdisciplinary innovation capabilities, encompassing multiple composite technical fields such as smart grids, energy internet, and big data analytics. Correspondingly, the indicator system for employee capability assessment is becoming increasingly complex. Therefore, how to scientifically and accurately construct an employee indicator assessment system to achieve objective and accurate evaluation of employee capabilities is a crucial issue in employee capability assessment within the current power grid system.

[0003] Existing employee competency assessment methods mostly rely on traditional indicator systems and expert subjective scoring. They generally lack data-driven dynamic optimization mechanisms, making it difficult to adapt to the variability of indicators in different business scenarios. Furthermore, the allocation of indicator weights depends on human experience, which introduces a certain degree of subjectivity. This makes it difficult to unify assessment standards and results in weak interpretability. While existing deep learning methods have non-linear modeling capabilities, they often treat input indicators as equal, ignoring the structural attribution and interdependence between indicators, making it difficult to balance prediction accuracy and interpretability. Summary of the Invention

[0004] This invention provides an employee competency assessment method, system, and medium based on structural attribution neural networks. It can solve the problems of inaccurate competency assessment results caused by the lack of consideration of the structural attribution relationship between indicators and the subjective allocation of indicator weights in existing technologies, thereby achieving an objective and accurate assessment of the competency of the employee to be assessed.

[0005] This invention provides an employee competence assessment method based on a structural attribution neural network, comprising:

[0006] The partial derivative information of each evaluation index with respect to historical capability evaluation results in the capability evaluation model is extracted based on the gradient backpropagation mechanism. The gradient sensitivity of each evaluation index is determined based on the partial derivative information, and the weight parameter of each evaluation index is determined based on the gradient sensitivity.

[0007] The real-time evaluation events of the employees to be evaluated are obtained. The real-time evaluation events are processed according to the weight parameters corresponding to each evaluation indicator to obtain the target evaluation event. The target evaluation event is then matrix-processed and divided according to the business dimension of the indicator to obtain multiple sub-matrices.

[0008] Each of the sub-matrices is input into the capability assessment model, so that the capability assessment model extracts features from each of the sub-matrices based on multiple structural sub-networks, concatenates the feature extraction results of all structural sub-networks into an intermediate representation vector, and predicts the intermediate representation vector through a multilayer perceptron aggregation regression network, outputting the target assessment result of the employee to be assessed. The structural dimension of each of the structural sub-networks corresponds one-to-one with the business dimension, and the capability assessment model adopts an attribution-guided loss function.

[0009] This invention utilizes the gradient backpropagation mechanism to extract partial derivative information and obtain gradient sensitivity, thereby determining weight parameters. This quantifies the impact of each evaluation indicator on the capability assessment results, providing an objective basis for subsequent updates to real-time evaluation events and making the assessment results more scientific and adaptable. By dividing the evaluation matrix into sub-matrices according to business dimensions, indicators from different business dimensions can be input into the corresponding structural sub-networks for targeted feature extraction. This allows the model to better capture the unique contribution of each business dimension indicator to the capability assessment, improving the accuracy of the assessment. After extracting features from multiple structural sub-networks, they are concatenated into an intermediate representation vector for summarizing and regression prediction. This integrates the feature information from each business dimension, comprehensively considering the synergistic impact of each dimension on the capability assessment, making the assessment results more comprehensive and accurate in reflecting employee capabilities.

[0010] Furthermore, the attribution-guided loss function is specifically as follows:

[0011] Weight parameters are determined from the candidate value set through cross-validation and grid search;

[0012] The mean square error between the predicted evaluation results and the actual evaluation results of the capability assessment model is calculated, and the structural attribution loss is calculated, wherein the structural attribution loss is used to guide the capability assessment model to learn the sensitivity structure of the input index vector.

[0013] The loss value of the capability assessment model is determined by adjusting the mean square error and the structural attribution loss using the weight parameters.

[0014] This approach, employing cross-validation and grid search, effectively determines the optimal values ​​of the weight parameters. This allows the model to better balance prediction error and structural attribution loss during training, improving its generalization ability and stability, and ensuring good evaluation results across different datasets. The mean squared error (MSE) measures the difference between the model's predicted evaluation results and the actual evaluation results. As a direct indicator of model prediction accuracy, it guides the model to optimize its prediction capabilities, making the evaluation results closer to reality. The introduction of structural attribution loss allows the model to focus on the sensitivity structure of the input indicator vector during training. This helps the model learn the degree of influence and intrinsic relationships of each indicator on the evaluation results, improving the model's interpretability. The evaluation results are not only accurate but also clearly reflect the role of each indicator, providing strong support for the optimization and adjustment of the evaluation system. Adjusting the weight parameters for both losses allows the model to balance prediction accuracy and structural interpretability during training, finding the optimal loss value and achieving comprehensive optimization of model performance, thus improving the model's effectiveness and reliability in practical applications.

[0015] Furthermore, the structural sub-network specifically comprises:

[0016] Each of the structural subnetworks includes multiple hidden layers. Each hidden layer uses the Mish activation function, and the input index vector of each hidden layer is processed according to the weight matrix and bias vector to obtain the output of each hidden layer.

[0017] By introducing the Mish activation function, the model gains better nonlinear fitting ability and gradient propagation characteristics. This enables the hidden layer to extract complex feature information more effectively when processing input indicator vectors, enhancing the model's ability to learn complex relationships between different indicators. Consequently, the accuracy of the competency assessment model in evaluating employee competencies is improved, and subtle differences in employee competencies are better captured.

[0018] Furthermore, the output of the hidden layer is specifically as follows:

[0019] h (l) =σ(W (l) h (l-1) +b (l) ), l=1,2;

[0020] in, Let σ be the input index vector, and σ be the Mish activation function σ(x) = x·

[0021] tanh(ln(1+e x W (l) With b (l) These are the weight matrix and bias vector of the l-th layer, respectively.

[0022] By explicitly defining the calculation method of the hidden layer, linearly transforming the input indicator vector through the weight matrix and bias vector, and then processing it with the Mish activation function, non-linear feature extraction of the input indicators is achieved. This calculation method enables the model to automatically adjust the weights and bias parameters during the learning process to better fit the data, thereby more accurately extracting the features of indicators in each business dimension. This provides a more reliable feature representation for subsequent capability assessment, thus improving the accuracy of the assessment results.

[0023] Furthermore, the determination of the weight parameters corresponding to each of the initial evaluation indicators based on each of the gradient sensitivities specifically involves:

[0024] The gradient sensitivities are normalized to obtain the weight parameters corresponding to each initial evaluation index.

[0025] By normalizing the data, gradient sensitivity values ​​at different scales can be transformed into comparable and interpretable weight parameters. This gives the weights of each initial evaluation indicator a unified dimension and range, facilitating weight comparison and analysis. This allows for a more intuitive understanding of the importance of each indicator in capability assessment, ensuring the rationality and interpretability of the weight parameters, providing a clear basis for optimizing the evaluation indicator system, and further enhancing the scientific rigor and accuracy of the evaluation system.

[0026] Furthermore, the attribution-guided loss function is specifically as follows:

[0027]

[0028] Among them, L MSE For predicting scores The first term is the mean squared error function of the actual score y; the second term is the structural attribution loss; and λ is the weighting parameter that adjusts the two losses.

[0029] This loss function explicitly combines prediction error and structural attribution loss, adjustable through the weight parameter λ. This allows the model to simultaneously optimize prediction accuracy and structural interpretability during training. This comprehensive loss function design enables the model to better explain the contribution of each evaluation metric to the results while ensuring accuracy. It improves the transparency and credibility of the evaluation system, providing clear direction and basis for subsequent adjustments and optimizations of evaluation metrics, and enhancing the model's practicality and operability in real-world applications.

[0030] Furthermore, the step of calculating the partial derivative information of each initial evaluation index with respect to the capability evaluation result based on the gradient backpropagation mechanism to obtain the gradient sensitivity corresponding to each initial evaluation index is specifically as follows:

[0031]

[0032] in, x represents the predicted score output by the model. i Let w represent the i-th input metric. i The gradient sensitivity of this indicator, wherein the gradient sensitivity value is used to reflect the input indicator x. i The extent to which small perturbations affect the evaluation results.

[0033] By calculating gradient sensitivity, the impact of each initial assessment indicator on the competency assessment results can be quantified, clarifying the importance of each indicator in the assessment process. This quantitative approach provides a direct basis for optimizing assessment indicators. Indicators can be adjusted and filtered based on the magnitude of gradient sensitivity, removing unimportant indicators and strengthening important ones. This allows the assessment system to more accurately reflect employee capabilities, improve the reliability and effectiveness of assessment results, and better serve the company's talent management decisions.

[0034] Another embodiment of the present invention provides an employee competency assessment system based on a structural attribution neural network, comprising: a matrix partitioning module, a model calculation module, a sensitivity module, and an indicator optimization module;

[0035] The matrix partitioning module is used to obtain the initial matrix corresponding to the initial evaluation index set and the evaluation score set, and to partition the initial matrix according to the business dimension corresponding to each index in the initial evaluation index set to obtain multiple sub-matrices.

[0036] The model calculation module is used to input each of the sub-matrices into the capability assessment model, so that the capability assessment model extracts features from each of the sub-matrices based on multiple structural sub-networks, concatenates the feature extraction results of all structural sub-networks into an intermediate representation vector, and predicts the intermediate representation vector through a multilayer perceptron aggregation regression network to obtain the capability assessment result output by the capability assessment model. The number of each of the sub-matrices and each of the structural sub-networks is equal. The initial assessment index set includes multiple initial assessment indices. The capability assessment model adopts an attribution-guided loss function.

[0037] The sensitivity determination module is used to calculate the partial derivative information of each initial evaluation index with respect to the capability evaluation result based on the gradient backpropagation mechanism, obtain the gradient sensitivity corresponding to each initial evaluation index, and determine the weight parameter corresponding to each initial evaluation index based on the gradient sensitivity.

[0038] The indicator optimization module is used to update the initial evaluation indicator set based on the weight parameters to obtain the target evaluation indicator set, and to perform capability evaluation based on the target evaluation indicators and the capability evaluation model.

[0039] This invention, through dividing the initial matrix into sub-matrices according to business dimensions, allows indicators from different business dimensions to be input into corresponding structural sub-networks for targeted feature extraction. This enables the model to better capture the unique contribution of each business dimension indicator to capability assessment, improving the accuracy of the assessment. Multiple structural sub-networks extract features and then concatenate them into an intermediate representation vector before summarizing and regressing for prediction. This integrates feature information from various business dimensions, comprehensively considering the synergistic impact of each dimension on capability assessment, resulting in a more comprehensive and accurate reflection of employee capabilities. Gradient sensitivity is obtained by calculating partial derivative information using the gradient backpropagation mechanism, which then determines the weight parameters. This quantifies the influence of each initial assessment indicator on the capability assessment results, providing an objective basis for subsequent updates to the assessment indicator set. This makes the assessment indicator system more scientific and adaptable, allowing for dynamic adjustment of indicator weights based on actual assessment conditions, thus optimizing the assessment system. The target assessment indicator set is obtained by updating the initial assessment indicator set based on the weight parameters, and capability assessment is then performed. This achieves dynamic optimization and self-improvement of the assessment indicator system, enabling it to better adapt to changes in the company's needs for employee capability assessment, improving the accuracy and effectiveness of the assessment.

[0040] Another embodiment of the present invention provides a terminal device, including: a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, it implements the steps of the employee competence assessment method based on structural attribution neural network of the present invention.

[0041] Another embodiment of the present invention provides a computer-readable storage medium item, including: a stored computer program, which, when the computer program is running, controls the device where the computer-readable storage medium is located to perform the steps of the employee competence assessment method based on structural attribution neural network of the present invention. Attached Figure Description

[0042] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0043] Figure 1 This is a schematic diagram illustrating the steps of an employee competence assessment method based on a structural attribution neural network provided in an embodiment of the present invention;

[0044] Figure 2 This is a schematic diagram of the structure of an employee competence assessment system based on a structural attribution neural network provided in an embodiment of the present invention. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0046] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms “comprising” and “having”, and any variations thereof, in the specification, claims, and foregoing description of the drawings are intended to cover non-exclusive inclusion.

[0047] In the description of the embodiments of this application, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly defined.

[0048] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0049] In the description of the embodiments in this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0050] In the description of the embodiments of this application, the term "multiple" refers to two or more (including two), similarly, "multiple sets" refers to two or more (including two sets), and "multiple pieces" refers to two or more (including two pieces).

[0051] In the description of the embodiments of this application, unless otherwise expressly specified and limited, technical terms such as "installation," "connection," "joining," and "fixing" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. For those skilled in the art, the specific meaning of the above terms in the embodiments of this application can be understood according to the specific circumstances.

[0052] See Figure 1 To address the problem of inaccurate competency assessment results in existing technologies, an embodiment of the present invention provides an employee competency assessment method based on a structural attribution neural network, comprising:

[0053] Step S101: Extract the partial derivative information of each evaluation index in the capability assessment model with respect to historical capability assessment results based on the gradient backpropagation mechanism, determine the gradient sensitivity of each evaluation index based on the partial derivative information, and determine the weight parameter of each evaluation index based on the gradient sensitivity.

[0054] In this embodiment, historical assessment data of employee competence evaluations from past power grid companies are collected. This data should include multiple assessment indicators, such as a core indicator set of eight indicators including education level, work experience, industry communication, project planning ability, project implementation level, innovation achievement ability, platform construction ability, and social organization construction and operation ability, forming an indicator input vector x = {x1,…,x8} to support the training of the competence evaluation model. The historical assessment data also includes historical assessment events for 200 employees collected from relevant systems of the power grid company, such as multi-dimensional data on project results, project implementation level, and intellectual property. The collected historical assessment data is then processed by missing data completion, noise reduction, and normalization to form a 200-row, 8-column input matrix. This input matrix is ​​then input into the competence evaluation model for forward propagation calculation, yielding the model's predicted competence evaluation result for each employee. Based on the model's prediction results and the actual historical competence evaluation results, the loss function is calculated. The loss function value reflects the accuracy of the model's prediction. Using a gradient backpropagation mechanism, the gradient of each parameter in the model is calculated layer by layer, starting from the loss function value. In this process, the partial derivative information of each evaluation indicator with respect to historical capability assessment results is calculated simultaneously. The partial derivative information represents the marginal influence of each evaluation indicator on the final result during model prediction; that is, the change in the model output when the value of the evaluation indicator changes slightly. For each evaluation indicator, its corresponding partial derivative information is analyzed. The larger the absolute value of the partial derivative information, the greater the influence of the evaluation indicator on the historical capability assessment results, meaning the more sensitive the indicator is in the model. The absolute value of the partial derivative information of each evaluation indicator is used as its gradient sensitivity. Gradient sensitivity quantifies the importance of each evaluation indicator in the model prediction process and reflects the degree of contribution of each evaluation indicator to the capability assessment results. To make the weight parameters of each evaluation indicator comparable, the gradient sensitivity of all evaluation indicators can be normalized to obtain the corresponding weight parameters for each evaluation indicator.

[0055] As an example of an embodiment of the present invention, the step of extracting the partial derivative information of each evaluation index in the capability assessment model with respect to historical capability assessment results based on the gradient backpropagation mechanism, and determining the gradient sensitivity corresponding to each evaluation index based on the partial derivative information, specifically involves:

[0056]

[0057] in, x represents the predicted score output by the model. i Let w represent the i-th input metric. i The gradient sensitivity of this indicator, wherein the gradient sensitivity value is used to reflect the input indicator x. i The extent to which small perturbations affect the evaluation results.

[0058] For example, the weight parameters of each indicator after normalization are shown in Table 1 below.

[0059] Table 1

[0060]

[0061] As an example of an embodiment of the present invention, the step of determining the weight parameters corresponding to each evaluation index based on each gradient sensitivity specifically involves:

[0062] The gradient sensitivities are normalized to obtain the weight parameters corresponding to each initial evaluation index.

[0063] In this embodiment, to ensure the comparability of the weight parameters of each evaluation indicator, the gradient sensitivity of all evaluation indicators is normalized. The normalization method can be to divide each gradient sensitivity by the sum of all gradient sensitivities, ensuring it ranges between 0 and 1, and that the sum of all normalized gradient sensitivities equals 1. The normalized gradient sensitivity is then used as the weight parameter for each evaluation indicator. A larger weight parameter indicates a greater proportion of that evaluation indicator in the competency assessment model, and a greater impact on the final employee competency assessment result. The normalization calculation formula is as follows:

[0064]

[0065] Where n is the total number of all input metrics, w is the weight parameter. i For gradient sensitivity.

[0066] Step S102: Obtain the real-time evaluation events of the employees to be evaluated, process the real-time evaluation events according to the weight parameters corresponding to each evaluation indicator to obtain the target evaluation events, and perform matrix processing and division on the target evaluation events according to the business dimension of the indicators to obtain multiple sub-matrices.

[0067] In this embodiment, m real-time evaluation events of employees to be evaluated are collected from the relevant systems of the power grid enterprise. Based on the weight parameters of each evaluation indicator previously determined through gradient sensitivity, the real-time evaluation events of the employees to be evaluated are weighted. Specifically, the real-time evaluation event of each evaluation indicator is multiplied by its corresponding weight parameter to obtain the target evaluation event. For example, if the weight of a certain evaluation indicator is 0.3 and the real-time score is 8 points, then the weighted value is 0.3 × 8 = 2.4. All weighted evaluation indicator values ​​are comprehensively calculated to obtain a comprehensive evaluation result, i.e., the target evaluation event. This target evaluation event reflects the comprehensive performance of the employee to be evaluated on each evaluation indicator, while also considering the importance (weight) of each indicator. All weighted evaluation indicator values ​​are arranged in a certain order to construct an original matrix. Assuming there are n evaluation indicators, the original matrix can be an n×1 column vector, where each row corresponds to a weighted evaluation indicator value. In the predefined evaluation indicator system, three different indicator business dimensions are designed according to the results, processes, and background to which the indicators belong. Based on the division of business dimensions, the original matrix is ​​divided into multiple sub-matrices. Each submatrix contains the weighted values ​​of all evaluation metrics for the corresponding business dimension. For example, if a company has three business dimensions (technical capability, communication capability, and teamwork capability), then three submatrices will be generated, each corresponding to one business dimension. Ensure that the format of each submatrix conforms to the input requirements of the capability assessment model. For example, the number of rows and columns of the submatrix should be consistent with the input dimensions of the structural subnetworks in the model. If the input dimension of a structural subnetwork in the model is 3×1, then the submatrix for the corresponding business dimension should also be in 3×1 format.

[0068] Step S103: Input each of the sub-matrices into the capability assessment model, so that the capability assessment model extracts features from each of the sub-matrices based on multiple structural sub-networks, concatenates the feature extraction results of all structural sub-networks into an intermediate representation vector, and predicts the intermediate representation vector through a multilayer perceptron aggregation regression network, and outputs the target assessment result of the employee to be assessed. The structural dimension of each of the structural sub-networks corresponds one-to-one with the business dimension, and the capability assessment model adopts an attribution-guided loss function.

[0069] In this embodiment, each sub-matrix obtained according to the business dimension is input into the corresponding structural sub-network in the capability assessment model. The input dimension of each structural sub-network is consistent with the sub-matrix dimension of the corresponding business dimension. Each structural sub-network extracts features from the input sub-matrix through its internal neural network layers (such as fully connected layers, convolutional layers, etc.). These feature extraction layers can learn the complex relationships between the various assessment indicators in the sub-matrix and extract feature vectors reflecting the comprehensive capabilities of employees under that business dimension. After completing feature extraction, each structural sub-network outputs a feature vector. These feature vectors represent the capability characteristics of employees under different business dimensions. The feature vectors output by all structural sub-networks are concatenated in a certain order to form an intermediate representation vector. This intermediate representation vector integrates the capability characteristics of employees under all business dimensions, providing comprehensive input information for subsequent aggregation regression. The concatenated intermediate representation vector is input into the aggregation regression network of the multilayer perceptron. The aggregation regression network consists of multiple fully connected layers, which can further process and analyze the intermediate representation vector. The summation regression network, through its internal weights and bias parameters, performs a weighted summation and nonlinear transformation on the intermediate representation vector, ultimately outputting a predicted value. This predicted value is the target assessment result for the employee being evaluated, reflecting the employee's comprehensive ability level. The ability assessment model employs an attribution-guided loss function. This loss function considers not only the difference between the model's predicted results and the actual historical ability assessment results, but also the contribution of each assessment indicator to the prediction result. Through attribution analysis, the importance of each assessment indicator in the model's prediction process can be determined, i.e., the contribution of each indicator to the final prediction result. This helps the model focus more on important assessment indicators during training, improving the model's prediction accuracy and interpretability. The intermediate representation vector is represented as follows:

[0070]

[0071] in, This is the output vector of the k-th structural subnetwork.

[0072] As an example of an embodiment of the present invention, the attribution-guided loss function is specifically as follows:

[0073] Weight parameters are determined from the candidate value set through cross-validation and grid search; the mean squared error between the predicted evaluation result and the actual evaluation result of the capability assessment model is calculated, and the structural attribution loss is calculated, wherein the structural attribution loss is used to guide the capability assessment model to learn the sensitivity structure of the input index vector; the mean squared error and the structural attribution loss are adjusted by the weight parameters to determine the loss value of the capability assessment model.

[0074] In this embodiment, firstly, a set of candidate values ​​for the weight parameters is determined. These candidate values ​​can be several sets of possible weight parameter values ​​set based on experience or pre-experimentation. For example, multiple candidate values ​​(such as 0.1, 0.2, 0.3, etc.) can be set for the weight parameters of each evaluation metric. Cross-validation is then used to evaluate the model's performance under different combinations of weight parameters. The historical ability assessment dataset is divided into several subsets, typically using k-fold cross-validation. For example, the dataset is divided into 10 subsets, with 9 subsets used as the training set and 1 subset as the validation set each time, repeated 10 times, selecting a different subset as the validation set each time. During each cross-validation process, a grid search is performed on the set of candidate weight parameter values. That is, all possible combinations of weight parameters are traversed, the ability assessment model is trained, and the model's performance is evaluated on the validation set. The performance evaluation metric can be mean squared error (MSE) or other relevant metrics. By comparing the model's performance on the validation set under different combinations of weight parameters, the weight parameter combination that optimizes the model's performance is selected. For example, the weight parameter combination that minimizes the mean squared error is selected as the final weight parameters. After training the capability assessment model using defined weight parameters, predictions are made on the training and validation sets to obtain the model's predicted assessment results. For each sample, the squared error between the model's predicted assessment result and the actual assessment result is calculated. Then, the average of the squared errors for all samples is taken to obtain the mean squared error (MSE). The MSE reflects the degree of difference between the model's predicted result and the actual result; the smaller the MSE value, the more accurate the model's prediction. The structural attribution loss is used to guide the capability assessment model to learn the sensitivity structure of the input indicator vector. Specifically, it measures whether the model's predicted sensitivity to the input indicator vector conforms to the expected gradient sensitivity structure. For example, if the gradient sensitivity of a certain assessment indicator is high, the model's predicted result should be more sensitive to changes in that indicator. For each sample, the gradient (i.e., partial derivative) of the model's predicted result with respect to the input indicator vector is calculated and compared with the pre-calculated gradient sensitivity. The difference between the two can be calculated using some distance metric (such as Euclidean distance). The sensitivity differences of all samples are summarized to obtain the structural attribution loss. The structural attribution loss reflects the overall difference between the model's predicted sensitivity structure to the input indicator vector and the expected structure. The mean squared error and structural attribution loss are weighted and summed to obtain the total loss value of the capability assessment model. Weight parameters are used to adjust the relative importance of mean squared error and structural attribution loss in the total loss. For example, if more emphasis is placed on the model's prediction accuracy, a larger weight can be assigned to the mean squared error; if more emphasis is placed on the model's sensitivity structure to the input index vector, a larger weight can be assigned to the structural attribution loss. The total loss value calculated through weighted summation is used to evaluate the overall performance of the model.During model training, the total loss value is minimized through optimization algorithms (such as gradient descent), thereby adjusting the model's parameters to achieve an optimal balance between prediction accuracy and sensitivity structure. The loss function is expressed as follows:

[0075]

[0076] Among them, L MSE The first term is the mean squared error between the predicted score and the actual score y; the second term is the structural attribution loss, which guides the model to focus on the sensitivity structure of the input index; λ is the weighting parameter that adjusts the two losses.

[0077] As an example of an embodiment of the present invention, the attribution-guided loss function is specifically as follows:

[0078]

[0079] Among them, L MSE For predicting scores The first term is the mean squared error function of the actual score y; the second term is the structural attribution loss; and λ is the weighting parameter that adjusts the two losses.

[0080] As an example of an embodiment of the present invention, the structural sub-network is specifically as follows:

[0081] Each of the structural subnetworks includes multiple hidden layers. Each hidden layer uses the Mish activation function, and the input index vector of each hidden layer is processed according to the weight matrix and bias vector to obtain the output of each hidden layer.

[0082] In this embodiment, each subnetwork contains multiple hidden layers used to extract features from the input index vector. The number of hidden layers and the number of neurons per layer can be designed according to the specific task and data complexity. For each hidden layer, a weight matrix and a bias vector are initialized. The weight matrix maps the input index vector to a new feature space, and the bias vector adjusts the activation threshold of the neurons. The initial values ​​of the weight matrix and bias vector are typically set using random initialization methods (such as Xavier initialization or He initialization) to ensure the stability and convergence speed of network training. For each hidden layer, the output of the previous layer is received as the input index vector. If it is the first hidden layer, the input index vector is a weighted submatrix. In each hidden layer, the input index vector is first linearly transformed. Specifically, the input index vector is multiplied by the weight matrix of the hidden layer, and then the bias vector is added. The Mish activation function is a self-gated activation function that combines a nonlinear activation function (such as ReLU) and a gating mechanism (such as Tanh). The result Z after the linear transformation is nonlinearly processed by the Mish activation function. The Mish activation function introduces non-linearity while avoiding the vanishing gradient problem, improving the model's expressive power and training stability. The output of each hidden layer serves as the input index vector for the next layer. This process is repeated sequentially through all hidden layers until the last one. The output of the last hidden layer is the feature vector extracted by this sub-network. This feature vector is then used in subsequent concatenation operations, combining with feature vectors from other sub-networks to form an intermediate representation vector, which is then input into the aggregation regression network for final prediction.

[0083] As an example of an embodiment of the present invention, the output of the hidden layer is specifically as follows:

[0084] h (l) =σ(W (l) h (l-1) +b (l) ), l=1,2;

[0085] in, Let be the input index vector, σ be the Mish activation function, and σ(x) = x·tanh(ln(1+e)). x W (l) With b (l) These are the weight matrix and bias vector of the l-th layer, respectively.

[0086] like Figure 2 As shown, based on the above method embodiments, corresponding system embodiments are provided;

[0087] An embodiment of the present invention provides an employee competency assessment system 200 based on a structural attribution neural network, comprising: a weight determination module 201, a matrix partitioning module 202, and an assessment module 203;

[0088] The weight determination module 201 is used to extract the partial derivative information of each evaluation index in the capability evaluation model with respect to the historical capability evaluation results based on the gradient backpropagation mechanism, determine the gradient sensitivity corresponding to each evaluation index based on the partial derivative information, and determine the weight parameter corresponding to each evaluation index according to the gradient sensitivity.

[0089] The matrix partitioning module 202 is used to acquire real-time evaluation events of employees to be evaluated, process the real-time evaluation events according to the weight parameters corresponding to each evaluation indicator to obtain target evaluation events, and perform matrix processing and partitioning on the target evaluation events according to the business dimensions of the indicators to obtain multiple sub-matrices.

[0090] The evaluation module 203 is used to input each of the sub-matrices into the capability evaluation model, so that the capability evaluation model extracts features from each of the sub-matrices based on multiple structural sub-networks, concatenates the feature extraction results of all structural sub-networks into an intermediate representation vector, and predicts the intermediate representation vector through a multilayer perceptron aggregation regression network, and outputs the target evaluation result of the employee to be evaluated. The structural dimension of each of the structural sub-networks corresponds one-to-one with the business dimension, and the capability evaluation model adopts an attribution-guided loss function.

[0091] As an example of an embodiment of the present invention, the evaluation module 203 is specifically as follows:

[0092] Used to determine weight parameters from a set of candidate values ​​through cross-validation and grid search;

[0093] The mean square error between the predicted evaluation results and the actual evaluation results of the capability assessment model is calculated, and the structural attribution loss is calculated, wherein the structural attribution loss is used to guide the capability assessment model to learn the sensitivity structure of the input index vector.

[0094] The loss value of the capability assessment model is determined by adjusting the mean square error and the structural attribution loss using the weight parameters.

[0095] As an example of an embodiment of the present invention, the evaluation module 203 is specifically as follows:

[0096] Each of the structural subnetworks includes multiple hidden layers. Each hidden layer uses the Mish activation function, and the input index vector of each hidden layer is processed according to the weight matrix and bias vector to obtain the output of each hidden layer.

[0097] As an example of an embodiment of the present invention, the evaluation module 203 is specifically as follows:

[0098] h (l) =σ(W (l) h (l-1) +b (l) ), l=1,2;

[0099] in, Let be the input index vector, σ be the Mish activation function, and σ(x) = x·tanh(ln(1+e)). x W (l) With b (l) These are the weight matrix and bias vector of the l-th layer, respectively.

[0100] As an example of an embodiment of the present invention, the weight determination module 201 specifically comprises:

[0101] This is used to normalize the gradient sensitivity of each of the above, so as to obtain the weight parameters corresponding to each of the initial evaluation indicators.

[0102] As an example of an embodiment of the present invention, the evaluation module 203 is specifically as follows:

[0103]

[0104] Among them, L MSE For predicting scores The first term is the mean squared error function of the actual score y; the second term is the structural attribution loss; and λ is the weighting parameter that adjusts the two losses.

[0105] As an example of an embodiment of the present invention, the weight determination module 201 specifically comprises:

[0106]

[0107] in, x represents the predicted score output by the model. i Let w represent the i-th input metric. i The gradient sensitivity of this indicator, wherein the gradient sensitivity value is used to reflect the input indicator x. i The extent to which small perturbations affect the evaluation results.

[0108] It is understood that the above system item embodiments correspond to the method item embodiments of the present invention, and can implement the employee competence assessment method based on structural attribution neural network provided by any of the above method item embodiments of the present invention.

[0109] It should be noted that the system embodiments described above are merely illustrative, and some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0110] Based on the above embodiments of the employee competency assessment method based on structural attribution neural networks, another embodiment of the present invention provides a terminal device, which includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the employee competency assessment method based on structural attribution neural networks of any embodiment of the present invention.

[0111] For example, in this embodiment, the computer program can be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the terminal device.

[0112] The terminal device may be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.

[0113] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the terminal device, connecting all parts of the terminal device via various interfaces and lines.

[0114] Based on the above-described method embodiments, another embodiment of the present invention provides a computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to execute the employee competency assessment method based on structural attribution neural network described in any of the above-described method embodiments of the present invention.

[0115] The modules / units integrated in the device / terminal equipment, if implemented as software functional units and sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0116] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for employee competency assessment based on structural attribution neural networks, characterized in that, include: The partial derivative information of each evaluation index with respect to historical capability evaluation results in the capability evaluation model is extracted based on the gradient backpropagation mechanism. The gradient sensitivity of each evaluation index is determined based on the partial derivative information, and the weight parameter of each evaluation index is determined based on the gradient sensitivity. The real-time evaluation events of the employees to be evaluated are obtained. The real-time evaluation events are processed according to the weight parameters corresponding to each evaluation indicator to obtain the target evaluation event. The target evaluation event is then matrix-processed and divided according to the business dimension of the indicator to obtain multiple sub-matrices. Each of the sub-matrices is input into the capability assessment model, so that the capability assessment model extracts features from each of the sub-matrices based on multiple structural sub-networks, concatenates the feature extraction results of all structural sub-networks into an intermediate representation vector, and predicts the intermediate representation vector through a multilayer perceptron aggregation regression network, outputting the target assessment result of the employee to be assessed. The structural dimension of each of the structural sub-networks corresponds one-to-one with the business dimension, and the capability assessment model adopts an attribution-guided loss function.

2. The employee competence assessment method based on structural attribution neural networks as described in claim 1, characterized in that, The attribution-guided loss function is specifically as follows: Weight parameters are determined from the candidate value set through cross-validation and grid search; The mean square error between the predicted evaluation results and the actual evaluation results of the capability assessment model is calculated, and the structural attribution loss is calculated, wherein the structural attribution loss is used to guide the capability assessment model to learn the sensitivity structure of the input index vector. The loss value of the capability assessment model is determined by adjusting the mean square error and the structural attribution loss using the weight parameters.

3. The employee competency assessment method based on structural attribution neural networks as described in claim 1, characterized in that, The structural sub-network is specifically as follows: Each of the structural subnetworks includes multiple hidden layers. Each hidden layer uses the Mish activation function, and the input index vector of each hidden layer is processed according to the weight matrix and bias vector to obtain the output of each hidden layer.

4. The employee competency assessment method based on structural attribution neural networks as described in claim 3, characterized in that, The output of the hidden layer is specifically as follows: h (l) =σ(W (l) h (l-1) +b (l) ),l=1,2; in, Let be the input index vector, σ be the Mish activation function, and σ(x) = x· tanh(ln(1+e x W (l) With b (l) These are the weight matrix and bias vector of the l-th layer, respectively.

5. The employee competence assessment method based on structural attribution neural networks as described in claim 1, characterized in that, The determination of the weight parameters corresponding to each evaluation index based on each gradient sensitivity is specifically as follows: The gradient sensitivities are normalized to obtain the weight parameters corresponding to each initial evaluation index.

6. The employee competence assessment method based on structural attribution neural networks as described in claim 2, characterized in that, The attribution-guided loss function is specifically as follows: Among them, L MSE For predicting scores The first term is the mean squared error function of the actual score y; the second term is the structural attribution loss; and λ is the weighting parameter that adjusts the two losses.

7. The employee competence assessment method based on structural attribution neural networks as described in claim 1, characterized in that, The method of extracting the partial derivative information of each evaluation index with respect to historical capability evaluation results in the capability evaluation model based on the gradient backpropagation mechanism, and determining the gradient sensitivity corresponding to each evaluation index based on the partial derivative information, specifically involves: in, x represents the predicted score output by the model. i Let w represent the i-th input metric. i The gradient sensitivity of this indicator, wherein the gradient sensitivity value is used to reflect the input indicator x. i The extent to which small perturbations affect the evaluation results.

8. An employee competency assessment system based on structural attribution neural networks, characterized in that, include: The module includes a weight determination module, a matrix partitioning module, and an evaluation module. The weight determination module is used to extract the partial derivative information of each evaluation index in the capability assessment model with respect to historical capability assessment results based on the gradient backpropagation mechanism, determine the gradient sensitivity corresponding to each evaluation index based on the partial derivative information, and determine the weight parameter corresponding to each evaluation index according to the gradient sensitivity. The matrix partitioning module is used to acquire real-time assessment events of employees to be assessed, process the real-time assessment events according to the weight parameters corresponding to each assessment indicator to obtain target assessment events, and perform matrix processing and partitioning of the target assessment events according to the business dimensions of the indicators to obtain multiple sub-matrices. The evaluation module is used to input each of the sub-matrices into the capability evaluation model, so that the capability evaluation model extracts features from each of the sub-matrices based on multiple structural sub-networks, concatenates the feature extraction results of all structural sub-networks into an intermediate representation vector, and predicts the intermediate representation vector through a multilayer perceptron aggregation regression network, and outputs the target evaluation result of the employee to be evaluated. The structural dimension of each of the structural sub-networks corresponds one-to-one with the business dimension, and the capability evaluation model adopts an attribution-guided loss function.

9. A terminal device, characterized in that, The device includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, it implements the employee competency assessment method based on a structural attribution neural network as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, include: A stored computer program, wherein, when the computer program is executed, it controls the device containing the computer-readable storage medium to perform the employee competency assessment method based on a structural attribution neural network as described in any one of claims 1-7.

Images