Cross-team evaluation method and system based on adversarial learning and optimal transmission theory

Abstract

The invention provides a cross-team evaluation method and system based on adversarial learning and an optimal transmission theory, and the method is applied to an evaluation system which comprises a data processing module, a capability representation extraction module, a template gap recognition module and an evaluation module. The method specifically comprises the following steps: constructing a corresponding multi-dimensional capability feature vector based on multi-dimensional working data of each team employee in a predetermined assessment period; generating corresponding capability representation based on adversarial learning; in combination with an optimal transmission theory, obtaining a template difference between the template distribution and preset template distribution; and calculating a performance alignment score of each employee based on the template gap, and mapping the performance alignment score to a preset score interval to obtain a fair performance index. According to the method, adversarial learning is utilized to extract team and group irrelevant ability characterization, the unified value scale is established in combination with the optimal transmission theory, and then team and group systematic deviation is avoided through accurate alignment of the ability characterization and the unified value scale, and the accuracy and reliability of a final evaluation result are improved.

Description

Technical Field

[0001] This invention relates to the field of performance evaluation technology, and in particular to a cross-team evaluation method and system based on adversarial learning and optimal transmission theory. Background Technology

[0002] Within the power company's production and operation system, there are multiple work teams with distinct divisions of labor and responsibilities, such as transmission line operation teams, substation operation and maintenance teams, and power dispatching teams. The work content and task attributes of these different teams differ fundamentally. For example, the transmission line operation team primarily focuses on high-frequency, low-tariff line inspections and fault handling, emphasizing breadth and frequency of work; the substation operation and maintenance team focuses on low-frequency, high-tariff equipment maintenance and live-line testing, emphasizing technical depth and responsibility; and the power dispatching team focuses on core businesses such as real-time monitoring and load adjustment. Their evaluation criteria also differ. To facilitate cross-team talent evaluation, resource allocation, and incentive decisions, most companies adopt an evaluation system that maps the performance of employees from different teams to a unified scale, thereby achieving comparable and quantifiable cross-team performance.

[0003] Currently, cross-team performance evaluation in power companies mainly relies on three types of methods. The first type involves setting fixed work point conversion coefficients or weighting rules to unify work point data of different natures to the same scale. However, this method relies on expert subjective experience, the rules are rigid and lack flexibility, making it difficult to dynamically adapt to changes in task complexity and actual value, and it cannot distinguish systematic biases between teams. The second type involves first normalizing and ranking employee performance within the team, and then mapping each team's ranking proportionally to the company's unified ranking system. This method ignores the differences in overall performance levels between different teams, potentially leading to lower evaluation results for ordinary employees in high-performing teams compared to outstanding employees in low-performing teams, making the accuracy of the evaluation difficult to guarantee. The last type uses models such as linear regression and gradient boosting trees to learn the non-linear relationship between features and performance scores from historical data to improve the accuracy of performance prediction within the team. However, this method lacks a mechanism to actively identify and eliminate the confounding factor of the team, indiscriminately learning the statistical correlation between team attributes and performance, rather than the causal correlation between individual ability and performance. This results in evaluation results being influenced by institutional differences within the team, failing to reflect the true abilities of employees.

[0004] In summary, existing cross-team performance evaluation methods lack an effective mechanism for identifying and eliminating systematic biases within teams, resulting in incomparability of cross-team performance evaluations and ultimately lower accuracy and reliability of the evaluation results. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies, which lack an effective mechanism for identifying and removing systematic biases in work teams, leading to incomparability in cross-work team performance evaluations and low accuracy and reliability of evaluation results. This invention provides a cross-work team evaluation method and system based on adversarial learning and optimal transfer theory. It utilizes adversarial learning to extract work team-independent capability representations and combines them with optimal transfer theory to establish a unified value scale. Through the precise alignment of capability representations and the unified value scale, systematic biases in work teams are avoided, thereby improving the accuracy and reliability of the final evaluation results.

[0006] The objective of this invention is achieved through the following technical solution: Cross-team evaluation methods based on adversarial learning and optimal transport theory include: Extract multi-dimensional work data of employees in each work group during the predetermined assessment period, and construct corresponding multi-dimensional capability feature vectors; Based on adversarial learning, capability representations corresponding to multi-dimensional capability feature vectors are generated. Based on the optimal transmission theory, the ability representation of each employee is compared with the preset template distribution to obtain the corresponding template gap; The performance alignment score for each employee is calculated based on the template gap, and then mapped to a preset score range to obtain the corresponding fair performance index.

[0007] Furthermore, the collection of multi-dimensional work data from employees in each work group within the predetermined assessment period, and the construction of corresponding multi-dimensional capability feature vectors, includes: Extract multi-dimensional work data of employees in each work group within the predetermined assessment period, and preprocess the extracted work data; Based on the preprocessed work data, the corresponding features of workload, work complexity, work quality, safety contribution, and supporting contribution are extracted. The extracted features are normalized to form a multidimensional capability feature vector.

[0008] Furthermore, before generating the capability representation corresponding to the multi-dimensional capability feature vector based on adversarial learning, the following is also performed: Performance benchmarks are calculated based on historical performance data, and corresponding expert scores are generated by combining expert evaluations. Using expert ratings as the target, a regression model is trained by combining performance benchmarks and auxiliary information from employees in each work group. Based on the trained regression model, corresponding fair performance target variables are generated. Training samples for corresponding competency representation learning are constructed using the fair performance target variables and multidimensional competency feature vectors of employees in each work group.

[0009] Furthermore, the ability representation corresponding to the multi-dimensional ability feature vector generated based on adversarial learning includes: Construct a collaborative training network consisting of a generator, a discriminator, and a performance predictor; Using the multi-dimensional capability feature vectors in the training samples as input, the generator learns to output the latent capability representation stripped of the class group attributes. The discriminator receives potential capability representations, identifies the corresponding work group origins, combines them with the corresponding real work group labels in the training samples, and optimizes the discriminator's network parameters by calculating adversarial loss. While maintaining the network parameters of the discriminator, the potential capability representation is regenerated through the generator, and the corresponding adversarial loss and reconstruction loss are calculated. At the same time, the performance score of the potential capability representation is predicted based on the performance predictor, and the performance retention loss is calculated in combination with the corresponding fair performance target variable in the training samples. Optimize the generator's network parameters based on adversarial loss, reconstruction loss, and performance preservation loss; Repeatedly train the generator and discriminator alternately until the Nash equilibrium point is reached, and obtain the generator that has converged during training; The generator generates the capability representations of employees in each work group based on the training convergence.

[0010] Furthermore, after generating capability representations corresponding to multi-dimensional capability feature vectors based on adversarial learning, the following steps are also performed: All team members are sorted in descending order based on the fair performance target variable, and the team members with the highest pre-set percentage of the ranking are selected to construct a high-performance group. Extract the competency representations of employees in each work group from the high-performing group and construct a pre-defined template distribution.

[0011] Furthermore, based on optimal transmission theory, the ability representation of each employee is compared with a preset template distribution to obtain the corresponding template gap, including: Each employee's capability is defined as a unit quality, and the preset template distribution is defined as the target warehouse set. The corresponding transportation cost is determined based on the spatial distance between each employee's ability representation and the ability representation of each team of employees in the preset template distribution. The template gap for each employee is obtained by calculating the minimum total transportation cost.

[0012] Furthermore, the step of calculating the performance alignment score for each employee based on the template gap, mapping each employee's performance alignment score to a preset score range, and obtaining the corresponding fair performance index includes: The template gaps for each employee are standardized, and the standardized template gaps are then reversed to obtain the corresponding performance alignment scores. The performance alignment score is linearly mapped to a preset score range to obtain the corresponding fair performance index.

[0013] Furthermore, after obtaining the corresponding fairness performance index, the following steps are also performed: For each employee, calculate the correlation between each dimension feature in the corresponding multidimensional capability feature vector and the corresponding fair performance index, extract the feature influence weight of each dimension feature, and convert it into the dimension contribution ratio; Based on the preset template distribution, set template values ​​and combine each employee's multidimensional capability feature vector to construct a corresponding personal radar chart; An evaluation report for each employee is constructed based on the contribution percentage of each dimension and the individual radar chart.

[0014] A cross-shift evaluation system based on adversarial learning and optimal transport theory, used to perform any of the above-mentioned cross-shift evaluation methods, includes: The data processing module is used to extract multi-dimensional work data of employees in each shift within the predetermined assessment period and construct corresponding multi-dimensional capability feature vectors. The capability representation extraction module is used to generate capability representations corresponding to multi-dimensional capability feature vectors based on adversarial learning. The template gap identification module is used to compare each employee's ability representation with a preset template distribution based on the optimal transmission theory to obtain the corresponding template gap; The evaluation module is used to calculate the performance alignment score for each employee based on the template gap, map each employee's performance alignment score to a preset score range, and obtain the corresponding fair performance index.

[0015] Furthermore, the evaluation module also includes: The evaluation results analysis unit is used to calculate the dimensional contribution ratio and generate an individual radar chart based on the corresponding multidimensional capability feature vector and fair performance index, and to construct an evaluation report for each employee.

[0016] The beneficial effects of this invention are: By introducing an adversarial learning mechanism, the generator extracts competency representations stripped of team information, retaining competency features relevant to employees' actual performance. Simultaneously, adversarial training of the discriminator suppresses the incorporation of team-related features, thus avoiding systematic bias in cross-team evaluation. Furthermore, based on optimal transport theory, a multivariate unified template distribution is constructed using the competency representations of high-performing employees across teams. The average alignment cost between each employee and this template distribution is quantified by calculating the corresponding spatial distance, transforming cross-team performance comparison into a gap metric with a unified high-performance standard, achieving comparability across team performance. Finally, the alignment cost is normalized and inverted to obtain a performance alignment score, which is then linearly mapped to a fair performance index to obtain accurate and reliable intuitive quantitative results. Attached Figure Description

[0017] Figure 1This is a flowchart of the present invention. Detailed Implementation

[0018] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0019] Example: Cross-team evaluation methods based on adversarial learning and optimal transport theory, such as Figure 1 As shown, it includes: Extract multi-dimensional work data of employees in each work group during the predetermined assessment period, and construct corresponding multi-dimensional capability feature vectors; Based on adversarial learning, capability representations corresponding to multi-dimensional capability feature vectors are generated. Based on the optimal transmission theory, the ability representation of each employee is compared with the preset template distribution to obtain the corresponding template gap; The performance alignment score for each employee is calculated based on the template gap, and then mapped to a preset score range to obtain the corresponding fair performance index.

[0020] Performance data of power company employees is scattered across multiple platforms, including the Production Management System (PMS), work order management system, operation ticket system, defect management system, safety incident recording system, and training and examination system. Furthermore, the data formats differ significantly between different work teams, leading to data fragmentation and hindering unified analysis. Therefore, a systematic approach is adopted to collect and initially clean multi-source data, integrating the scattered objective data into a structured dataset. Multi-dimensional capability feature vectors are then extracted from this structured dataset to overcome the incomparability caused by differences in task formats across different work teams, and to address the issue that a single indicator cannot reflect the true capabilities of employees. Specific tasks are transformed into general capability dimensions decoupled from task formats, preventing distortion of subsequent evaluation results due to one-sided indicators and ensuring the reliability of the final evaluation.

[0021] The process of collecting multi-dimensional work data from employees in each work group during the predetermined assessment period and constructing corresponding multi-dimensional capability feature vectors includes: Extract multi-dimensional work data of employees in each work group within the predetermined assessment period, and preprocess the extracted work data; Based on the preprocessed work data, the corresponding features of workload, work complexity, work quality, safety contribution, and supporting contribution are extracted. The extracted features are normalized to form a multidimensional capability feature vector.

[0022] The predetermined assessment period can be selected according to the current evaluation needs, such as monthly, quarterly or annually, and then relevant work data within the corresponding period can be extracted from data sources such as the PMS production management system.

[0023] The multi-dimensional work data specifically includes: baseline work points, actual execution time, number of participants, and task weights for calculating workload; inherent attribute data and external condition data for calculating work complexity, such as voltage level, equipment importance, operation type, weather conditions, and whether operation is live; first-time success rate, rework rate, number and accuracy of defects and hidden dangers found, and post-maintenance health data for calculating work quality; basic safety score, violation operation records, and details of unsafe events for calculating safety contributions; bonus data such as safety hazard elimination records, adoption of safety rationalization suggestions, and emergency rescue performance; and training and lecture hours, participation time in technical innovation projects, mentorship hours, time invested in team building, workload of data processing, and tool maintenance records for calculating supporting contributions.

[0024] The extracted multi-dimensional working data is preprocessed, including removing erroneous data, standardizing the format, and identifying and handling missing values.

[0025] Based on the preprocessed multi-dimensional work data, in order to eliminate the task differences between different work groups, the specific work tasks completed by employees within the predetermined assessment period are transformed into a standardized, decoupled multi-dimensional capability feature vector. The multi-dimensional capability feature vector represents the comprehensive capabilities demonstrated by employees within the assessment period. After calculating all features, each feature is normalized to construct the final multi-dimensional capability feature vector.

[0026] The workload characteristics The formula used to measure the total workload undertaken by an employee during the assessment period is as follows: ; in, This represents the baseline work points required to complete the i-th task, which can be pre-determined based on actual needs. This indicates an adjustment of the weighting coefficient, used to adjust the base work points based on factors such as the actual execution time and number of participants for the i-th task. This represents the total number of tasks.

[0027] The complexity of the work The formula for calculating the technical difficulty, professional expertise, and risk of the work handled by employees is as follows: ; in, This represents the objective complexity score of the i-th task, which is obtained according to preset rules based on the inherent attributes of the task and external conditions. This indicates the weight or proportion of the i-th task in the employee's total workload.

[0028] The quality of work The formula used to measure the quality and effectiveness of employees' work is as follows: ; in, Let be the success rate of completing the i-th task in one attempt. The rework rate of the i-th task, The number and accuracy of defects or hidden dangers found in the i-th task.

[0029] The security contribution The formula used to measure an employee's safety performance and contribution to safe production is as follows: ; in, For the basic score of safety, For the deduction data in the i-th task, This represents the bonus data for the i-th task.

[0030] The supporting contributions This is used to measure an employee's supporting contribution to the team and work group, beyond their primary business performance. The calculation formula is as follows: ; in, This refers to the actual time an employee spends on the i-th task; It is a value multiplier used to differentiate the relative importance, technical content, and contribution to the organization of different types of support work.

[0031] The extracted features are normalized to form a multidimensional capability feature vector. .

[0032] Before generating the capability representation corresponding to the multi-dimensional capability feature vector based on adversarial learning, the following is also performed: Performance benchmarks are calculated based on historical performance data, and corresponding expert scores are generated by combining expert evaluations. Using expert ratings as the target, a regression model is trained by combining performance benchmarks and auxiliary information from employees in each work group. Based on the trained regression model, corresponding fair performance target variables are generated. Training samples for corresponding competency representation learning are constructed using the fair performance target variables and multidimensional competency feature vectors of employees in each work group.

[0033] Because of significant differences in job content, work point allocation mechanisms, and performance evaluation standards among different work teams in power companies—for example, transmission line operation teams primarily perform high-frequency, low-work-point line patrol tasks, while substation operation and maintenance teams focus on low-frequency, high-work-point specialized technical operations—traditional evaluation methods either rely on expert subjective experience or indiscriminately learn the statistical correlation between work team attributes and work points, rather than the causal relationship between individual ability and performance. This leads to the model incorrectly coupling work team identity with individual performance, resulting in a distortion that overestimates the work points of employees in high-performing teams and underestimates the true contributions of other teams. This lack of a unified, fair benchmark across work teams makes it difficult for subsequent model training to learn the mapping relationship between general abilities and performance, and the construction of high-performance benchmarks loses its representativeness due to work team biases.

[0034] Therefore, based on historical data and cross-team expert scoring training, a fair performance target variable independent of team attributes is generated. This variable incorporates objective performance information and reflects unified value judgments, providing a reference anchor for the performance maintenance constraints of the subsequent adversarial learning model, the selection of high-performing groups, and the fairness of the entire evaluation system, thus avoiding team-specific bias in the evaluation process.

[0035] First, based on the existing performance evaluation methods of each work group, such as comprehensive work points and historical assessment grades, a preliminary performance benchmark value is calculated. Then, combined with expert evaluation, a representative sample of employees is evaluated across work groups to generate expert scores. A representative employee sample can be selected based on the silhouette coefficient, the expression of which is: ; in, Indicates cohesion, indicating the number of sample points. The smaller the average difference or distance between itself and all other points in its own group, the more representative it is of the typical characteristics of the group. The resolution indicates the number of sample points. The larger the average difference or distance between it and all points in the nearest other group, the stronger the difference. The farther away from other groups, the clearer it is to distinguish one's own group from the others. Ultimately, the silhouette coefficient... This takes into account the degree of cohesion. and resolution Then, a measure was derived. The final score of the sample point grouping quality can be used to select the silhouette coefficient. The highest employee sample is used as a representative employee sample.

[0036] Using expert ratings as the target, a regression model is trained by combining performance benchmarks and auxiliary information of employees in each work group. The auxiliary information includes years of service, professional titles, etc. The regression model can be a gradient boosting tree or a simple linear regression, used to predict fair performance scores for all employees.

[0037] Then, the data of all employees is input into the trained regression model, and the resulting prediction is the final fair performance target variable. . It is in The score within the interval represents the employee's fair performance level under uniform standards, independent of work group. Each employee is then represented as a training sample containing a corresponding multi-dimensional ability feature vector and a fair performance target variable.

[0038] Although the generated multidimensional capability feature vectors have achieved data scale uniformity and task decoupling, these vectors still implicitly contain interfering information strongly correlated with team attributes. This residual team information can lead to subsequent evaluations still being affected by systematic biases, failing to truly focus on employees' actual capabilities. Therefore, by further utilizing the dynamic game between the generator and discriminator in adversarial learning, while retaining core performance information, we proactively erase the traces of team attributes in the capability feature vectors, severing the link between team identity and capability assessment, and ensuring the accuracy of subsequent employee performance evaluations.

[0039] The ability representation based on adversarial learning, which generates multi-dimensional ability feature vectors, includes: Construct a collaborative training network consisting of a generator, a discriminator, and a performance predictor; Using the multi-dimensional capability feature vectors in the training samples as input, the generator learns to output the latent capability representation stripped of the class group attributes. The discriminator receives potential capability representations, identifies the corresponding work group origins, combines them with the corresponding real work group labels in the training samples, and optimizes the discriminator's network parameters by calculating adversarial loss. While maintaining the network parameters of the discriminator, the potential capability representation is regenerated through the generator, and the corresponding adversarial loss and reconstruction loss are calculated. At the same time, the performance score of the potential capability representation is predicted based on the performance predictor, and the performance retention loss is calculated in combination with the corresponding fair performance target variable in the training samples. Optimize the generator's network parameters based on adversarial loss, reconstruction loss, and performance preservation loss; Repeatedly train the generator and discriminator alternately until the Nash equilibrium point is reached, and obtain the generator that has converged during training; The generator generates the capability representations of employees in each work group based on the training convergence.

[0040] In the collaborative training network, the generator's main function is to learn the group-invariant representation. Its input is the multi-dimensional ability feature vector of the employees. After mapping, a low-dimensional latent capability representation is output. This representation retains performance-related information while removing department-related information as much as possible. The generator uses a multilayer perceptron or encoder network as its basic structure.

[0041] The function of the discriminator is to identify the latent representations produced by the generator. From which department? Discriminator receives. As input, the output is a probability value used to determine the likelihood that the representation belongs to each work group type. The discriminator is implemented using a multilayer perceptron network suitable for classification tasks.

[0042] The function of a performance predictor is to validate latent representations. Has sufficient performance information been retained? It is based on... As input, output a predicted performance score. To ensure that the representation still has strong performance prediction capabilities while being de-grouped, a multilayer perceptron network suitable for regression tasks is used for construction.

[0043] Based on the constructed collaborative training network, training is performed using the constructed training samples. During training, the generator and discriminator will discriminate against each other, competing against each other. The discriminator's goal is to accurately determine the representation as much as possible. The original work group of the employees is determined. The training process minimizes the classification cross-entropy loss. This is achieved by generating representations that make it difficult for the discriminator to distinguish between true and false representations. Learn mapping functions This eliminates department-specific biases, retains performance-related information, and makes the accuracy of the discriminator's judgment approach that of random guessing.

[0044] Therefore, maximizing the adversarial loss of the discriminator is taken as the training objective of the discriminator, and its calculation formula is as follows: ; in, To combat the losses, For mathematical expectation, These are capability representation samples that have not undergone the generator's de-grouping process and possess real-world team attributes. The probability distribution represents the true characteristics of work groups. For discriminator For true representation The recognition results Let this be the multidimensional capability feature vector of one of the employees. The true data distribution of multidimensional capability feature vectors for all employees. For discriminator For generator Output the recognition results of the representation.

[0045] Meanwhile, to ensure that the generator does not lose its original multidimensional capability feature vector during the process of deceiving the discriminator. In addition to valuable information, the generator's training is also constrained by two other loss functions: Reconstruction loss The purpose is to ensure capability representation. Able to be restored to the original multidimensional capability feature vector by a decoder network with as little loss as possible. This ensures capability representation. Multidimensional capability feature vectors are preserved. Most of the original information is expressed as follows: ; in, The input is a multidimensional capability feature vector. For decoder networks.

[0046] Performance retention loss The purpose is to ensure capability representation. Can be used by performance predictors Accurately map to the true fair performance target variable This ensures capability representation The information most relevant to performance is highlighted and retained, and its calculation method is as follows: ; in, For performance predictors.

[0047] Optimization goal of generator It minimizes a weighted total loss consisting of three loss components, and its expression is: ; in, , and These are hyperparameters used to balance the importance of the three objectives; they are preset parameters.

[0048] By repeatedly training the generator and discriminator alternately, the model will eventually reach a Nash equilibrium point. At this point, the generator can produce a capability representation that the discriminator cannot effectively distinguish from the department source. At the same time, this capability representation still contains enough information to reconstruct the original input and predict the final performance. This final capability representation is the capability representation that is independent of the work group.

[0049] After generating capability representations corresponding to multi-dimensional capability feature vectors based on adversarial learning, the following is also performed: All team members are sorted in descending order based on the fair performance target variable, and the team members with the highest pre-set percentage of the ranking are selected to construct a high-performance group. Extract the competency representations of employees in each work group from the high-performing group and construct a pre-defined template distribution.

[0050] To establish a benchmark that is both fair and consistent while also reflecting the diversity of high performance, instead of constructing a single average profile, we will build a distribution of high-performing employees' abilities. First, based on the fair target variable calculated for all employees, we will select employees from the work groups with the highest pre-defined percentage of scores, such as the top 20% of employees, as the high-performing employee group.

[0051] Then, the ability representations corresponding to these high-performing employees are extracted to form a set of ability representations, which constitutes an experience probability distribution. The resulting experience probability distribution is the preset template distribution.

[0052] Compared to a single average vector, a unified preset template distribution can more completely capture the various possibilities and forms of high performance, ensuring the authenticity and reliability of subsequent evaluation results.

[0053] Based on this, and using optimal transmission theory, the gap between each employee and the preset template distribution is calculated, thereby achieving performance alignment.

[0054] The process, based on optimal transmission theory, compares each employee's capability representation with a preset template distribution to obtain the corresponding template gap, including: Each employee's capability is defined as a unit quality, and the preset template distribution is defined as the target warehouse set. The corresponding transportation cost is determined based on the spatial distance between each employee's ability representation and the ability representation of each team of employees in the preset template distribution. The template gap for each employee is obtained by calculating the minimum total transportation cost.

[0055] Each employee's capability is defined as a unit of quality, and the preset template distribution consists of a set of warehouses. Optimal transport aims to calculate the minimum total transportation cost required to allocate and transport this unit of employee quality to the various warehouses in the most economical way. This cost, or Wasserstein distance, quantifies the gap between the employee and the high-performing group; the smaller the cost, the closer the employee is to the high-performing group, and the better their performance.

[0056] Specifically, transportation cost is defined as the square of the Euclidean distance between the ability representation of each employee and the ability representation of each shift employee in the preset template distribution.

[0057] Because the preset template distribution is a collection The minimum cost of a uniform distribution of points equals the average transportation cost from an employee to all points in the pre-defined template distribution, which can be defined as the alignment cost of the employee, expressed as: ; in, For the first Alignment costs per employee For the first Multidimensional capability feature vector of an employee The first among high-performing groups Multidimensional capability feature vector of an employee For the preset template distribution, For Wasserstein distance, The sample size for the high-performing group. For the first The first employee to be placed in the high-performing group Transportation costs per employee. The smaller the value, the closer the employee's ability characteristics are to those of high-performing individuals.

[0058] The resulting alignment cost is the template gap, and the smaller the value, the better. In order to convert it into a more intuitive performance score, after obtaining the template gap, it is also converted into a structured fair performance index.

[0059] The step of calculating each employee's performance alignment score based on template gaps, mapping each employee's performance alignment score to a preset score range, and obtaining the corresponding fair performance index includes: The template gaps for each employee are standardized, and the standardized template gaps are then reversed to obtain the corresponding performance alignment scores. The performance alignment score is linearly mapped to a preset score range to obtain the corresponding fair performance index.

[0060] The formula for calculating the performance alignment score is as follows: ; in, For the first The performance alignment score of each employee, with a value range of... , This is a normalization function. The higher the performance alignment score, the closer the employee's ability pattern is to that of the high-performing group.

[0061] To further improve the interpretability of the output results, the performance alignment score is further linearly mapped to the standard. The score range forms the final fairness performance index, which is expressed as: ; in, For the first Fair performance index for each employee.

[0062] After obtaining the corresponding fairness performance index, the following steps are also performed: For each employee, calculate the correlation between each dimension feature in the corresponding multidimensional capability feature vector and the corresponding fair performance index, extract the feature influence weight of each dimension feature, and convert it into the dimension contribution ratio; Based on the preset template distribution, set template values ​​and combine each employee's multidimensional capability feature vector to construct a corresponding personal radar chart; An evaluation report for each employee is constructed based on the contribution percentage of each dimension and the individual radar chart.

[0063] Because the model is end-to-end resolvable, it is possible to clearly trace how the final evaluation results are calculated and derived from characteristic dimensions such as original workload, complexity, quality, safety contribution, and supporting contribution, thereby clearly identifying employees' strengths and weaknesses and performance drivers.

[0064] Specifically, model interpretability analysis, such as feature gradient contribution and attention weight quantification, can be used to determine the degree of influence of each dimension on the employee's individual fair performance index. Then, after normalizing these degrees of influence, the corresponding dimension contribution percentage can be formed.

[0065] Then, by solving for the mean values ​​of each dimension of the ability characteristics of all high-performing employees in the preset template distribution, template values ​​are set, and then compared with each employee's multidimensional ability characteristic vector in the same coordinate system to form a corresponding personal radar chart. Through the difference in the length of the rays in the radar chart, the gap and advantages of employees and high-performing employees in each ability dimension are intuitively displayed.

[0066] Finally, the composition of performance scores is explained from a causal perspective by the contribution ratio of each dimension, identifying driving factors and shortcomings. Individual radar charts are used to show the specific gaps with high performance standards from a comparative perspective, forming the final evaluation report. This provides accurate and objective data support for resource allocation, incentive formulation, and personalized training program design.

[0067] Another aspect of this embodiment provides a cross-team evaluation system based on adversarial learning and optimal transmission theory, including: The data processing module is used to extract multi-dimensional work data of employees in each shift within the predetermined assessment period and construct corresponding multi-dimensional capability feature vectors. The capability representation extraction module is used to generate capability representations corresponding to multi-dimensional capability feature vectors based on adversarial learning. The template gap identification module is used to compare each employee's ability representation with a preset template distribution based on the optimal transmission theory to obtain the corresponding template gap; The evaluation module is used to calculate the performance alignment score for each employee based on the template gap, map each employee's performance alignment score to a preset score range, and obtain the corresponding fair performance index.

[0068] The evaluation module also includes: The evaluation results analysis unit is used to calculate the dimensional contribution ratio and generate an individual radar chart based on the corresponding multidimensional capability feature vector and fair performance index, and to construct an evaluation report for each employee.

[0069] The data processing module, capability characterization extraction module, template gap identification module, evaluation module, and evaluation result analysis unit are all data processing devices with corresponding data processing capabilities, such as computers, and are all equipped with external information ports to extract the required data from various external systems.

[0070] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Other variations and modifications are possible without departing from the technical solutions described in the claims.

Claims

1. A cross-team evaluation method based on adversarial learning and optimal transport theory, characterized in that, include: Extract multi-dimensional work data of employees in each work group during the predetermined assessment period, and construct corresponding multi-dimensional capability feature vectors; Based on adversarial learning, capability representations corresponding to multi-dimensional capability feature vectors are generated. Based on the optimal transmission theory, the ability representation of each employee is compared with the preset template distribution to obtain the corresponding template gap; The performance alignment score for each employee is calculated based on the template gap, and then mapped to a preset score range to obtain the corresponding fair performance index.

2. The cross-team evaluation method based on adversarial learning and optimal transmission theory according to claim 1, characterized in that, The step involves extracting multi-dimensional work data from employees in each work group within a predetermined assessment period and constructing corresponding multi-dimensional capability feature vectors, including: Extract multi-dimensional work data of employees in each work group within the predetermined assessment period, and preprocess the extracted work data; Based on the preprocessed work data, the corresponding features of workload, work complexity, work quality, safety contribution, and supporting contribution are extracted. The extracted features are normalized to form a multidimensional capability feature vector.

3. The cross-team evaluation method based on adversarial learning and optimal transmission theory according to claim 1, characterized in that, Before generating the capability representation corresponding to the multi-dimensional capability feature vector based on adversarial learning, the following is also performed: Performance benchmarks are calculated based on historical performance data, and corresponding expert scores are generated by combining expert evaluations. Using expert ratings as the target, a regression model is trained by combining performance benchmarks and auxiliary information from employees in each work group. Based on the trained regression model, corresponding fair performance target variables are generated. Training samples for corresponding competency representation learning are constructed using the fair performance target variables and multidimensional competency feature vectors of employees in each work group.

4. The cross-team evaluation method based on adversarial learning and optimal transmission theory according to claim 3, characterized in that, The capability representation based on adversarial learning, which generates multi-dimensional capability feature vectors, includes: Construct a collaborative training network consisting of a generator, a discriminator, and a performance predictor; Using the multi-dimensional capability feature vectors in the training samples as input, the generator learns to output the latent capability representation stripped of the class group attributes. The discriminator receives potential capability representations, identifies the corresponding work group origins, combines them with the corresponding real work group labels in the training samples, and optimizes the discriminator's network parameters by calculating adversarial loss. While maintaining the network parameters of the discriminator, the potential capability representation is regenerated through the generator, and the corresponding adversarial loss and reconstruction loss are calculated. At the same time, the performance score of the potential capability representation is predicted based on the performance predictor, and the performance retention loss is calculated in combination with the corresponding fair performance target variable in the training samples. Optimize the generator's network parameters based on adversarial loss, reconstruction loss, and performance preservation loss; Repeatedly train the generator and discriminator alternately until the Nash equilibrium point is reached, and obtain the generator that has converged during training; The generator generates the capability representations of employees in each work group based on the training convergence.

5. The cross-team evaluation method based on adversarial learning and optimal transmission theory according to claim 1, characterized in that, After generating capability representations corresponding to multi-dimensional capability feature vectors based on adversarial learning, the following is also performed: All team members are sorted in descending order based on the fair performance target variable, and the team members with the highest pre-set percentage of the ranking are selected to construct a high-performance group. Extract the competency representations of employees in each work group from the high-performing group and construct a pre-defined template distribution.

6. The cross-team evaluation method based on adversarial learning and optimal transmission theory according to claim 5, characterized in that, Based on optimal transmission theory, the ability representation of each employee is compared with a preset template distribution to obtain the corresponding template gap, including: Each employee's capability is defined as a unit quality, and the preset template distribution is defined as the target warehouse set. The corresponding transportation cost is determined based on the spatial distance between each employee's ability representation and the ability representation of each team of employees in the preset template distribution. The template gap for each employee is obtained by calculating the minimum total transportation cost.

7. The cross-team evaluation method based on adversarial learning and optimal transmission theory according to claim 1, characterized in that, The process of calculating each employee's performance alignment score based on template gaps, mapping each employee's performance alignment score to a preset score range, and obtaining the corresponding fair performance index includes: The template gaps for each employee are standardized, and the standardized template gaps are then reversed to obtain the corresponding performance alignment scores. The performance alignment score is linearly mapped to a preset score range to obtain the corresponding fair performance index.

8. The cross-team evaluation method based on adversarial learning and optimal transmission theory according to claim 1, characterized in that, After obtaining the corresponding fairness performance index, the following steps are also performed: For each employee, calculate the correlation between each dimension feature in the corresponding multidimensional capability feature vector and the corresponding fair performance index, extract the feature influence weight of each dimension feature, and convert it into the dimension contribution ratio; Based on the preset template distribution, set template values ​​and combine each employee's multidimensional capability feature vector to construct a corresponding personal radar chart; An evaluation report for each employee is constructed based on the contribution percentage of each dimension and the individual radar chart.

9. A cross-shift evaluation system based on adversarial learning and optimal transport theory, used to execute the cross-shift evaluation method according to any one of claims 1 to 8, characterized in that, include: The data processing module is used to extract multi-dimensional work data of employees in each shift within the predetermined assessment period and construct corresponding multi-dimensional capability feature vectors. The capability representation extraction module is used to generate capability representations corresponding to multi-dimensional capability feature vectors based on adversarial learning. The template gap identification module is used to compare each employee's ability representation with a preset template distribution based on the optimal transmission theory to obtain the corresponding template gap; The evaluation module is used to calculate the performance alignment score for each employee based on the template gap, map each employee's performance alignment score to a preset score range, and obtain the corresponding fair performance index.

10. The cross-team evaluation system based on adversarial learning and optimal transmission theory according to claim 9, characterized in that, The evaluation module also includes: The evaluation results analysis unit is used to calculate the dimensional contribution ratio and generate an individual radar chart based on the corresponding multidimensional capability feature vector and fair performance index, and to construct an evaluation report for each employee.

Images