A performance management method, device and medium based on cloud computing and artificial intelligence

Abstract

The present application relates to the technical field of performance management, and discloses a performance management method and device based on cloud computing and artificial intelligence and a medium, comprising: obtaining a collaborative system log to perform multi-source behavior data conversion, combining point-in-time accurate historical data to splice and generate an aggregated representation; then performing multi-view representation learning and block fusion processing, introducing graph network smoothing operation to perform semantic matching to obtain associated semantic representation; then extracting a rating main axis based on sample dispersion and establishing it as a reference distribution benchmark; filtering a business stable anchor set to calculate parameters, constructing a transformation model containing a unit matrix and a one-way projection scaling mapping, performing axial re-labeling on the associated semantic representation, so that scale correction only produces anisotropic scaling in the direction of the rating main axis, and the non-main axis orthogonal space remains unchanged; finally, segmenting and mapping the corrected score to output a rating level label.

Description

Technical Field

[0001] This invention relates to the field of performance management technology, and more specifically, to a performance management method, device, and medium based on cloud computing and artificial intelligence. Background Technology

[0002] As enterprises deepen their digital transformation, performance management has gradually evolved from the traditional one-time year-end evaluation to a cross-cycle version evolution scoring system driven by continuous cloud data collection and periodic retraining. Existing cloud-based machine learning operation and maintenance systems widely adopt closed-loop automated pipelines of training, deployment, and retraining. This continuous training mechanism emphasizes repeatedly running the pipeline with new data, ensuring traceability of component versions for each execution. Simultaneously, the underlying infrastructure connecting training and online services often employs a dual-mode data plane, providing both offline high-throughput batch reads and online low-latency random reads, emphasizing point-in-time accuracy to ensure that data snapshots at specific points in time are obtained, rather than current values. In large group enterprises, the scoring service backend typically shares computing power and inference services with multiple organizational units. In this model, the system outputs quantified scores, while the tier boundaries that determine actual business actions are usually externally managed and maintained separately by the operational constraints within each organization. However, the aforementioned collaborative backend relying on continuous cloud iteration and multi-party sharing has significant shortcomings in practical applications: First, during continuous deployment, the latent vector space extracted by the representation learning model inherently lacks direct interchangeability. Even if the vector space obtained after retraining encodes similar information, it cannot be directly interchanged across versions due to orthogonality indeterminism.

[0003] Secondly, continuous retraining across multiple dimensions inevitably causes an overall shift in the distribution of quantized scores. Within the existing business framework of multi-dimensional dimensionality reduction and labeling output, if the drift along the main scoring axis is not identified and compensated, the semantics of the entire chain hierarchy will become extremely unstable.

[0004] Finally, in scenarios where multiple organizations share a scoring backend, the shift in scale distribution leads to a sharp increase in the maintenance costs of maintaining external echelon boundaries.

[0005] In summary, the actual performance and behavior of the same employee remain approximately constant, but due to the drift of the scoring scale with the model version, the employee is incorrectly pushed into or out of the predetermined tier by the system, resulting in a completely non-ability-driven tier jump phenomenon. Summary of the Invention

[0006] This invention provides a performance management method, device, and medium based on cloud computing and artificial intelligence, which solves the technical problems mentioned in the background art.

[0007] Firstly, a performance management method based on cloud computing and artificial intelligence includes: The system acquires multi-source collaborative system logs, performs multi-source behavioral data transformation, and combines accurate historical data splicing at specific moments with time decay weights to generate an aggregated representation. The aggregated representation is subjected to multi-view representation learning and block fusion processing, and graph network smoothing operation is introduced to perform semantic matching between the competency model and the historical behavior graph to obtain the associated semantic representation. Based on the sample dispersion, the tiered scoring axis is extracted from the associated semantic representation, and the tiered scoring axis is established as the reference distribution benchmark. Filter the set of stable business anchor points, and calculate the mean and variance of the scores of the set of stable business anchor points under different model versions to obtain the scaling coefficient and bias parameter. Based on the scaling factor and the bias parameter, a transformation model containing the identity matrix and unidirectional projection scaling mapping is constructed. The associated semantic representation is axially rescaled so that the model update and the scaling correction under multi-organization division only produce anisotropic scaling in the principal axis direction of the echelon scoring, while keeping the non-principal axis orthogonal space unchanged, thereby obtaining the correction score. A fixed allocation threshold is set on the reference distribution benchmark, and the correction scores are segmented and mapped to output tier rating labels.

[0008] In a second aspect, an electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the steps of the performance management method based on cloud computing and artificial intelligence as described in any one of the claims.

[0009] Thirdly, a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the performance management method based on cloud computing and artificial intelligence as described in any one of the claims.

[0010] The beneficial effects of this invention include: by introducing a set of stable business anchor points to obtain scaling coefficients and bias parameters, and on this basis constructing a transformation model containing an identity matrix and unidirectional projection scaling mapping, the dimensional distribution drift caused by continuous cloud retraining and model version updates can be eliminated. This invention restricts scale correction under cross-version evolution and multi-organizational division to anisotropic scaling along the principal axis of tiered scoring, while perfectly preserving the original business semantics of the non-principal axis orthogonal space. Thus, without compromising the competency connotation, it completely eliminates the problem of non-capability-driven tier jumps caused by system iteration, significantly reduces the management cost when multiple organizations share a scoring backend, and achieves cross-cycle stability, traceability, and high interpretability of rating results. Attached Figure Description

[0011] Figure 1 This is the present invention; Figure 2 This is the invention. Detailed Implementation

[0012] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0013] Example 1: As Figure 1 As shown, a performance management method based on cloud computing and artificial intelligence includes: The system acquires multi-source collaborative system logs, performs multi-source behavioral data transformation, and combines accurate historical data splicing at specific moments with time decay weights to generate an aggregated representation. The aggregated representation is subjected to multi-view representation learning and block fusion processing, and graph network smoothing operation is introduced to perform semantic matching between the competency model and the historical behavior graph to obtain the associated semantic representation. Based on the sample dispersion, the tiered scoring axis is extracted from the associated semantic representation, and the tiered scoring axis is established as the reference distribution benchmark. Filter the set of stable business anchor points, and calculate the mean and variance of the scores of the set of stable business anchor points under different model versions to obtain the scaling coefficient and bias parameter. Based on the scaling factor and the bias parameter, a transformation model containing the identity matrix and unidirectional projection scaling mapping is constructed. The associated semantic representation is axially rescaled so that the model update and the scaling correction under multi-organization division only produce anisotropic scaling in the principal axis direction of the echelon scoring, while keeping the non-principal axis orthogonal space unchanged, thereby obtaining the correction score. A fixed allocation threshold is set on the reference distribution benchmark, and the correction scores are segmented and mapped to output tier rating labels.

[0014] Preferably, the system logs from multiple sources are acquired, multi-source behavioral data are transformed, and then combined with accurate historical data at specific points in time and time decay weights to generate an aggregated representation, including: Extract event records from the multi-source collaborative system logs, and calculate the event weight by combining the event baseline value, completion status, task severity, and rework flag, as shown in the following formula: in, Log events; Event weights; For system source Next event type The benchmark value of the event; The status is now complete. This is the severity amplification factor; Task severity; This is the rework penalty coefficient; Marked as rework.

[0015] The event weights are summed and divided by the task opportunity size penalty to obtain the quantitative value of behavior intensity, thus completing the multi-source behavior data transformation. The formula is as follows: in, For employees In system source Event Type Time window The quantitative value of behavioral intensity; This refers to the collection of event records within the corresponding time window. This is a penalty for the scale of task opportunities.

[0016] Based on the scoring time point To accurately stitch together historical data at the boundary execution point, a historical behavior sequence is obtained. A time-decaying diagonal matrix is ​​constructed based on the half-life mechanism as the time-decay weight. A dimensionality expansion operation is then performed on the historical behavior sequence to generate an aggregated representation, as shown in the following formula: in, It is a time-decaying diagonal matrix; The decay coefficient is determined based on the half-life mechanism; For the first Each time window is close to the scoring time point Time difference; Total number of time windows; For polymer characterization; Operations for dimension expansion; For dimension The unit square matrix, in which Total number of event types; It is a historical behavior sequence formed by stacking the behavior intensity quantification values ​​of each time window.

[0017] Event logs are single records related to various work behaviors of employees in the logs of multi-source collaborative systems. They can be collected by connecting to the log interfaces of collaborative systems such as enterprise instant messaging, work order management, code repository, and customer relationship management.

[0018] The baseline value of an event is the basic value coefficient corresponding to different event types under different system sources. It is preferably a value between 0.1 and 2.0 configured according to job families, because different event types have different importance to employee performance evaluation. Core work events are configured with higher coefficients, while auxiliary work events are configured with lower coefficients.

[0019] The completion status refers to the completion status of the corresponding work tasks performed by the employee as recorded in the event log. It can be collected by parsing the status field of the task in the collaborative system log.

[0020] The severity amplification factor is a factor that amplifies the severity of a task. It is preferably 0.3. This value can distinguish the performance contribution of tasks with different severity levels within a reasonable range, and avoid excessive amplification that leads to a single high-severity task dominating the performance evaluation.

[0021] Task severity is the level of importance of a work task recorded in the event log, which can be collected by parsing the severity field of the task in the collaborative system log.

[0022] The rework penalty coefficient is a coefficient used to penalize tasks that require rework. The preferred value is 0.5. This value can reasonably punish rework without completely negating the performance contribution of the task due to a single instance of rework.

[0023] Rework flags are indicators that identify whether a task corresponding to an event has undergone rework. They can be collected by parsing the rework flag field of the task in the collaborative system log or by judging whether rework has occurred based on the task's processing records.

[0024] An employee ID is a unique identification code assigned to each employee by the company. It can be collected through the employee identity field in the company's human resource management system or various collaborative systems.

[0025] System source refers to the collaborative system type that generates employee work behavior logs, which can be collected by identifying the source system identifier field of the log data.

[0026] Event type is a subcategorization of work behavior under the system source, which can be collected by parsing the event type classification field in the collaborative system log.

[0027] A time window is a unit of time used to divide employee work behavior data into time dimensions. The preferred time is the natural week because the time granularity of the natural week can fit the daily work management rhythm of the enterprise and can realize the periodic collection and updating of performance data.

[0028] The event record set within a time window is the collection of all event records of a certain type for a certain employee under a certain system source within a specified time window.

[0029] The task opportunity scale penalty is a value used to quantify the scale of work tasks assigned to an employee under a certain system source within a specified time window. It is used to avoid the problem of excessively high performance quantification values ​​due to a large number of tasks assigned.

[0030] Event weights are calculated by combining the event baseline value, completion status, task severity, and rework flags, and are used to quantify the performance contribution of a single event record.

[0031] The behavioral intensity quantification value is obtained by summing the event weights within a specified time window and dividing by the square root of the task opportunity scale penalty term plus 1. It is used to quantify the performance contribution of an employee's certain work behavior within a specified time window.

[0032] The scoring point is the time node for conducting employee performance evaluations and extracting corresponding historical behavioral data. The preferred time is the last natural day of each month. This value can fit the company's regular monthly performance evaluation management cycle and can define a clear time boundary for splicing data at specific points in time.

[0033] Historical behavior sequence is a sequence of data formed by stacking the quantitative values ​​of an employee's behavior intensity in different time windows under a certain system source in chronological order.

[0034] The decay coefficient is determined based on the half-life mechanism and is used to decay the quantitative values ​​of behavior intensity over time in different time windows. The preferred value is 0.0248, which corresponds to a half-life of 28 days. This value can fit the phased characteristics of employee performance behavior, with low decay of recent behavior and high decay of long-term behavior.

[0035] The time difference is the time interval between the Nth time window and the scoring point.

[0036] The total number of time windows is the total number of time windows used to extract historical employee behavior data. It is preferably 12, because a time span of 12 natural weeks can cover enough work behavior data of employees, and at the same time, the historical data will not interfere with the current performance evaluation due to the time span being too long.

[0037] The time decay diagonal matrix is ​​a diagonal matrix constructed based on the decay coefficient and the time difference between each time window and the scoring point. It is used to perform time-dimensional decay processing on historical behavior sequences.

[0038] The total number of event types is the total number of work behavior event types contained under a certain system source.

[0039] An identity matrix with dimension equal to the total number of event types is an identity matrix with both the number of rows and columns equal to the total number of event types. Its main diagonal elements are all 1s, and all other elements are 0s.

[0040] Aggregated representation is a feature vector obtained by performing time decay and dimensional expansion operations on historical behavior sequences, which can comprehensively reflect the performance contribution of employees' historical work behaviors.

[0041] The dimension expansion operation is an operator that performs a Kronecker product operation on a time decay diagonal matrix and an identity matrix, which is used to extend the effect of time decay to all event type dimensions.

[0042] In detail, this invention breaks through the traditional single-dimensional method of quantifying employee behavioral performance. It calculates event weights through multi-dimensional feature fusion, while introducing a task opportunity scale penalty to prevent the number of tasks from dominating performance evaluation. Combined with point-in-time data splicing rules, it ensures no future information leakage. Based on a half-life mechanism, it constructs a time-decay diagonal matrix to achieve phased weighting of behavioral data. Then, through dimensional expansion operations, it applies time decay to all event type dimensions. For example, for a R&D employee's fault handling event in the work order system, its baseline value is configured as 1.5, completion status is "on time," task severity is 3, and there is no rework mark. The calculated weight of this event is 1.5 multiplied by 1 multiplied by 1 plus 0.3 multiplied by 3, then multiplied by 1, resulting in 2.85. If this employee has 10 such events in the current week, the task... With an opportunity size penalty of 10, the behavioral intensity quantification value is the sum of the weights of 10 events (28.5) divided by the square root of 10 plus 1, resulting in approximately 8.59. This calculation method considers both the performance contribution of the events themselves and avoids inflated values ​​due to a large number of assigned tasks. Furthermore, when the scoring point is the 30th of each month, only behavioral data prior to that point is extracted to eliminate interference from subsequently added data. The behavioral intensity quantification value for the past 12 weeks is then decayed using a 28-day half-life. The decay coefficient for the most recent week is close to 1, while the decay coefficient for 12 weeks ago is approximately 0.1, making recent behaviors contribute more to performance. Finally, through dimensional expansion calculations, this time decay rule is applied to all event types in the work order system, including fault handling, requirement coordination, and problem answering, forming an aggregated representation that truly reflects the employee's phased work behavior.

[0043] It should be noted that the specific calculation method for the task opportunity scale penalty item is the total number of tasks assigned to employees under a certain system source within a specified time window, such as the number of work orders assigned in the work order system, the number of times an employee is mentioned in the instant messaging system, and the number of code submission tasks in the code repository; the specific value of the half-life is 28 days, and the decay coefficient is obtained by dividing the natural logarithm 2 by the half-life; the specific time granularity of the time window is the natural week, that is, each week from Monday to Sunday is a time window; the specific rule for setting the scoring point is the last natural day of the month, and if the last day of the month is a statutory holiday, it is moved to the previous working day; the specific unit for calculating the time difference is the natural day, that is, the number of natural days from the last day of the time window to the scoring point; The specific configuration rules for the event baseline value are based on enterprise job families and set according to the relevance of the event to the core work of the job. For example, the baseline value for code submission events for R&D positions is 1.2, the baseline value for test case design events for testing positions is 1.0, and the baseline value for process approval events for administrative positions is 0.8. The specific mapping value rules for completion status are as follows: on-time completion is 1, overdue completion is 0.6, overdue non-completion is 0.2, and task cancellation after initiation is 0. The total number of time windows needs to fit the management needs of the enterprise's quarterly performance evaluation, selecting a time span of 12 natural weeks, i.e., 3 natural months. At the same time, it can be flexibly adjusted within the range of 6 to 24 natural weeks according to the actual management needs of the enterprise, and the adjustment needs to be recorded in the system for versioning.

[0044] Preferably, the aggregated representation undergoes multi-view representation learning and block fusion processing, and a graph network smoothing operation is introduced to perform semantic matching between the competency model and the historical behavior graph, resulting in an associated semantic representation, including: Based on the available sample size, a multi-view representation learning model is applied to extract the view output from the aggregated representation, as shown in the following formula: in, For employees In system source The view output below; For multi-view representation learning models; For polymer characterization; For model version parameters.

[0045] A masking identifier is constructed for the missing source and combined with the view output to generate a processed output, which is then concatenated into a block structure vector, as shown in the following formula: in, This is the output after processing; This is a masking flag, with a value of either zero or one. It is a block structure vector.

[0046] The block-based fusion process is performed on the block structure vector using a block diagonal fusion weight matrix to obtain uniformly defined dimensional features, as shown in the following formula: in, To uniformly define dimensional features; This is the block diagonal fusion weight matrix containing the source structure.

[0047] Based on co-activation associations, a behavioral association Laplacian matrix is ​​constructed from the historical behavioral graph. A graph network smoothing operation is then constructed by combining the identity matrix and the smoothing intensity coefficient. Inverse operations are performed on the behavioral association Laplacian matrix and the uniformly defined dimensional features to execute the semantic matching, resulting in an association semantic representation, as shown in the following formula: in, For semantic representation of association; A unit square matrix; The smoothness strength coefficient; The Laplace square for behavior association; Smoothing operations are performed on the constructed graph network.

[0048] The multi-view representation learning model is a dedicated model for feature extraction from aggregated representations of different system sources. It is preferably a two-layer multilayer perceptron or a linear support vector machine. The model is dynamically selected according to the available sample size of each system source. When the sample size is large, the multilayer perceptron is used to improve the feature representation capability, and when the sample size is small, the linear support vector machine is used to ensure the model training efficiency and generalization.

[0049] Model version parameters are the training parameters of the multi-view representation learning model under different model versions. They are automatically generated and versioned during the model training process.

[0050] The view output is a feature vector obtained by the multi-view representation learning model after extracting features from the aggregated representation of the corresponding system source.

[0051] The masking identifier is a binary identifier used to indicate whether there is missing data in the system source. It is preferably 0 or 1. The binary value clearly distinguishes the missing state of the system source. 0 corresponds to missing source and 1 corresponds to present source, avoiding interference from non-binary values ​​to subsequent feature fusion.

[0052] The processed output is a feature vector obtained by multiplying the view output with the corresponding masking identifier.

[0053] The block structure vector is a high-dimensional feature vector obtained by concatenating the processed outputs of all system sources with the masking identifier in a fixed order.

[0054] The block diagonal fusion weight matrix is ​​a diagonal matrix that divides the system into block structures according to the system source. It is used to fuse features of the block structure vectors and map them to a unified dimension.

[0055] The unified definition of dimensional features is a fixed-dimensional feature vector obtained by processing the block structure vector through the block diagonal fusion weight matrix, and the number of dimensions is consistent with the number of dimensions of the enterprise competency model.

[0056] The behavioral association Laplace matrix is ​​a Laplace matrix constructed based on the co-activation associations of employees' historical behavior maps, used to quantify the structural association features of employees' work behavior.

[0057] The smoothing intensity coefficient is used to adjust the filtering intensity of the graph network smoothing operation on the uniformly defined dimensional features. It is preferably 0.1. This value can incorporate the structural correlation information of the behavior graph within a reasonable range, which can improve the semantic fit of the representation without masking the individual behavioral representation features of employees. When the numerical calculation is unstable, it can be adjusted to 0.01.

[0058] The identity matrix is ​​a matrix with the same number of rows and columns, and all elements on the main diagonal are 1s while all other elements are 0s. It is the basic matrix for constructing smoothing operations in graph networks.

[0059] The associated semantic representation is a feature vector obtained by processing the uniformly defined dimensional features through graph network smoothing operations, which integrates the semantic information of the competency model and the historical behavior graph.

[0060] The total number of system sources is the total number of collaborative systems participating in performance data collection, which is determined by the type of collaborative systems that the enterprise actually accesses.

[0061] It should be noted that this invention breaks through the traditional single-model approach to multi-source aggregated representation. It selects a multi-view representation learning model based on the adaptability of the available sample size for each system source. Simultaneously, it constructs masking labels for system sources with missing data and combines them with the view output to avoid representation bias caused by data gaps. The processed outputs are then concatenated into block structure vectors and fused in blocks using a block diagonal fusion weight matrix, preserving the structural features of each system source and avoiding semantic dilution. Subsequently, based on the co-activation associations of the competency dimension, a behavioral association Laplacian matrix of the historical behavior graph is constructed. Combined with the unit matrix and smoothing strength coefficient, a graph network smoothing operation is constructed. Semantic matching between the competency model and the historical behavior graph is achieved through matrix inverse operations, allowing the representation to simultaneously integrate individual employee behavior and structural association features. For example, in a certain enterprise... The system integrates three sources: instant messaging, work orders, and a code repository. The work order system has 60,000 available samples and uses a two-layer multilayer perceptron. The instant messaging system has 30,000 available samples and uses a linear support vector machine. If there is no data in the code repository system for three consecutive time windows, the masking flag is set to 0. The processed outputs of the three sources are concatenated with the masking flag in sequence to form a block structure vector. The block diagonal fusion weight matrix is ​​used to fuse the vectors into a unified definition dimension feature consistent with the number of competency dimensions. Then, a behavioral association Laplace matrix is ​​constructed based on the co-activation associations of the three competency dimensions of communication, delivery, and quality. A graph network smoothing operation is constructed by combining the unit matrix with a smoothing strength coefficient of 0.1. The unified definition dimension feature is filtered through inverse operation to finally obtain an association semantic representation that integrates individual behavior and collaborative structure.

[0062] It should be noted that the specific threshold for the available sample size of the multi-view representation learning model is 50,000. When the sample size is greater than or equal to 50,000, a two-layer multilayer perceptron is used; when the sample size is less than 50,000, a linear support vector machine is used. The specific structure of the two-layer multilayer perceptron is as follows: the input layer dimension is consistent with the aggregate representation dimension of the corresponding system source; the hidden layer has 64 neurons; the output layer dimension is consistent with the competency dimension; and the activation function is the ReLU function. The kernel function of the linear support vector machine is fixed as a linear kernel, and the penalty coefficient is set to 1.0. The specific judgment rule for the masking indicator is that if a system source has no event records for three consecutive time windows, it is judged as a source missing, and the masking indicator is set to 0; otherwise, it is set to 1. The specific training method for the block diagonal fusion weight matrix is ​​stochastic gradient descent, the learning rate is set to 0.001, the training epochs are 100, and the loss function is cross-entropy loss. The specific calculation of co-activation association... The rules are as follows: the activation threshold of the competency dimension is taken as the 80th percentile of the dimension in the reference period; the number of times any two competency dimensions are simultaneously activated within the last 12 time windows is used as the co-activation correlation value; the specific steps for constructing the behavioral correlation Laplace matrix are to first construct the adjacency matrix based on the co-activation correlation value, then calculate the degree matrix of the adjacency matrix, and finally subtract the adjacency matrix from the degree matrix to obtain the Laplace matrix; the specific solution method for matrix inversion in the graph network smoothing operation is the conjugate gradient iteration method, with a maximum of 50 iteration steps to avoid numerical instability caused by explicit inversion; the specific dimension value rules for the unified definition of dimensional features are to be completely consistent with the number of dimensions in the enterprise job family competency model, and different job families can be flexibly adjusted according to the competency model; the rule for adjusting the smoothing intensity coefficient is that when the residual of the conjugate gradient iteration method is greater than 0.001, it is judged as numerically unstable, and the smoothing intensity coefficient is adjusted from 0.1 to 0.01.

[0063] Preferably, extracting the tiered scoring axis from the associated semantic representation based on sample dispersion, and establishing the tiered scoring axis as a reference distribution benchmark, includes: The echelon representation center of each echelon is calculated based on the historical echelon set. The internal discreteness matrix is ​​extracted based on the deviation between the associated semantic representation and the echelon representation center of the corresponding echelon, as shown in the following formula: in, It is an internal discreteness matrix; To assemble the historical echelons; Index for employees; For semantic representation of association; It serves as the echelon representation center for its respective echelon.

[0064] The stability adjustment coefficient is calculated based on the trace of the internal discreteness matrix and the total number of dimensions, using the following formula: in, To stabilize the adjustment coefficient; The trace of the internal discreteness matrix; This represents the total number of dimensions.

[0065] By combining the unit matrix and the stability adjustment coefficient to correct the internal dispersion matrix, the stable dispersion matrix is ​​obtained, as shown in the following formula: in, For a stable discreteness matrix; It is a unit square.

[0066] Combining the inverse operation of the stable discreteness matrix with the extreme value difference vectors of the highest and lowest echelon characterization centers, the echelon scoring principal axis is extracted after normalization and solidified as the reference distribution benchmark, as shown in the following formula: in, The main focus of the team scoring; This is the inverse operation of the stable discreteness matrix; As the center representing the highest echelon; As the lowest echelon representation center; This is the extreme value difference vector; This is for normalization purposes.

[0067] The historical tier set is a collection of semantic representations of employees whose tiers have been defined in the company's past performance evaluations. It can be collected by retrieving historical performance rating results and corresponding semantic representation data of employees from the company's performance management system.

[0068] The echelon representation center is the mean vector of the associated semantic representations of all employees in the same echelon within the historical echelon set, which is calculated by averaging the associated semantic representations of the same echelon.

[0069] The internal dispersion matrix is ​​a matrix that quantifies the dispersion of the semantic representations of employees within each echelon in the historical echelon set. It is calculated from the deviation between the semantic representations of the employees and the representation center of their respective echelons.

[0070] The stability adjustment coefficient is a coefficient used to correct the internal discreteness matrix to solve its numerical inversion instability problem. It is calculated from the trace and the total number of dimensions of the internal discreteness matrix.

[0071] The trace of the internal discreteness matrix is ​​the sum of the elements on the main diagonal of the internal discreteness matrix, and it is an indicator for quantifying the overall numerical characteristics of the internal discreteness matrix.

[0072] The total number of dimensions is the number of dimensions in the semantic representation of the association. It is preferable to use the number of dimensions in the enterprise job family competency model, because keeping the total number of dimensions consistent with the number of dimensions in the competency model can ensure that the main axis of the tiered scoring aligns with the core competency direction of the enterprise's performance evaluation, and avoid the main axis deviating from the business semantics due to redundancy or absence of dimensions.

[0073] A stable discreteness matrix is ​​a matrix obtained by combining the internal discreteness matrix with the identity matrix and the stability adjustment coefficient, and it has the stability of numerical inversion.

[0074] The top echelon representation center is the echelon representation center of the highest performance rating in the historical echelon set, and is calculated by averaging the associated semantic representations of employees in the top echelon.

[0075] The lowest tier representation center is the tier representation center of the lowest performance rating tier in the historical tier set, and is calculated by averaging the associated semantic representations of employees in the lowest tier.

[0076] The extreme value difference vector is the vector obtained by performing a difference operation between the highest-tier representation center and the lowest-tier representation center. It is the core vector for quantifying the difference in performance between the highest and lowest tiers.

[0077] The tiered scoring axis is a feature direction vector that maximizes the differentiation between different performance tiers. It is obtained by combining the inverse operation of the stable discreteness matrix with the extreme value difference vector and then normalizing it. It is the core reference dimension for performance scoring.

[0078] It should be noted that this invention breaks through the black-box operation of multi-dimensional feature dimensionality reduction in traditional performance management. Instead of using a generic dimensionality reduction method without business semantics, it calculates the representation center of each echelon based on the company's actual historical echelon set. By deviating between the semantic representation and the echelon representation center, it extracts an internal dispersion matrix that reflects the discrete features within the echelon. Then, it combines the trace of the internal dispersion matrix with the total number of dimensions to obtain a stabilization adjustment coefficient, achieving regularization correction of the internal dispersion matrix and solving the numerical instability problem of direct inversion. Subsequently, it combines the inverse operation of the stable dispersion matrix with the extreme value difference vector to extract feature directions that fit the core differences of the performance echelons. After normalization, it obtains the echelon scoring main axis with clear business semantics and solidifies it as a reference. The distribution benchmark ensures that the dimensionality-reduced scoring dimensions perfectly match the core needs of enterprise performance evaluation. For example, if an enterprise divides performance into four tiers, it first retrieves the historical tier set and corresponding semantic representations from the past three months to calculate the representation centers of the four tiers. Then, it extracts the internal dispersion matrix within each tier. If the total number of dimensions is 6, i.e., the 6 dimensions of the competency model, it calculates the trace of the internal dispersion matrix and obtains the stability adjustment coefficient. After correcting to obtain the stable dispersion matrix, it is inverted. The inverse matrix is ​​multiplied by the extreme value difference vectors of the highest and lowest tiers. After normalization, the tier scoring axis is obtained. This axis can accurately point to the most critical feature direction for distinguishing performance tiers among the 6 competency dimensions, and after being solidified, it becomes the reference benchmark for all subsequent versions of performance scoring.

[0079] It should be noted that the specific selection rules for the historical tier set are as follows: the performance rating results of the company for the past 12 natural weeks are selected, and the single sample size of each tier is not less than 50, to avoid the bias in the calculation of the representation center caused by too small a sample size; the stability adjustment coefficient is a negative cubic coefficient of 10, which can effectively solve the singularity problem of the internal dispersion matrix and ensure the stability of the inversion while minimizing the interference of regularization on the original dispersion data, so that the corrected stable dispersion matrix fits the original data characteristics as closely as possible; the specific rules for setting the total number of dimensions are that they correspond one-to-one with the number of dimensions of the competency model of each job family in the company. Different job families can set the total number of dimensions separately, such as 6 for R&D positions, 5 for operations positions, and 4 for administrative positions, and the setting is then versioned and fixed; the specific calculation method for normalization is L2 normalization. First, calculate the magnitude of the vector obtained by multiplying the stable discreteness matrix by the extreme value difference vector. Then, divide each element of this vector by the magnitude to obtain the tier scoring axis with a magnitude of 1. The specific verification rule for the tier scoring axis is to calculate the ratio of the average inter-class distance between tiers along the axis to the average intra-class distance within each tier. If the ratio is greater than 1.5, the tier discrimination of the axis is deemed qualified. If it is less than 1.5, the historical tier set is reselected to expand the sample size and then extracted again. The specific definition rule for the extreme value tier is that the highest tier is the first tier of the enterprise performance rating, corresponding to the top 10% of employees in performance rating, and the lowest tier is the fourth tier of the enterprise performance rating, corresponding to the bottom 20% of employees in performance rating. The tiers are determined according to the performance quota ratio of the enterprise's human resource management, and the determination rule is consistent with the enterprise's actual performance management system.

[0080] Preferably, a set of stable business anchor points is selected, and the mean and variance of the scores of the set of stable business anchor points under different model versions are calculated to obtain the scaling coefficient and bias parameter, including: The candidate template set is filtered based on template stability attributes, personnel coverage, and result measurability attributes; employees whose number of completions on the candidate template set falls within a set median range are selected to form the business stability anchor point set. The reference mean and reference standard deviation of the set of stable business anchor points under the reference model version, and the current mean and current standard deviation under the current model version are calculated using the following formulas: in, This is the current mean value under the current model version; This refers to the current model version. To divide the organization into units; A set of anchor points for business stability; The number of samples in the set of stable business anchor points; Index for employees; Original projected scores for employees; This represents the current standard deviation for the current model version.

[0081] The scaling factor is obtained by dividing the reference standard deviation by the sum of the current standard deviation and the zero-prevention minimum; the bias parameter is obtained by subtracting the product of the scaling factor and the current mean from the reference mean, as shown in the following formula: in, This is the scaling factor; The reference standard deviation is the reference model version. To prevent extremely small quantities; These are bias parameters; This is the reference mean under the reference model version.

[0082] The candidate template set is a set of task templates selected from enterprise work task templates based on template stability attributes, personnel coverage, and result measurability attributes. It can be obtained by retrieving the full set of task template data in the enterprise task management system and filtering according to the preset three attribute judgment rules.

[0083] The business stability anchor set is a set of employees selected from the company's employees whose number of tasks completed on the candidate template set falls within a set median range. It is calculated from the task completion data of the candidate template set and the work behavior data of the company's employees.

[0084] The current model version is the training iteration version identifier of the cloud-based performance management model currently used for performance scoring. It is preferably the latest training iteration version number of the cloud-based performance management model. This version is the model version currently used for performance scoring and can match the dimensional calibration requirements in real-time scoring scenarios.

[0085] Organizational units are independent performance evaluation units defined by an enterprise based on its business management structure. They can be obtained by retrieving organizational structure and job family division data from the enterprise's human resource management system, and then extracting them by job family combined with department.

[0086] The current mean is the arithmetic mean of the original projected scores of the business stability anchor set under the current model version, which is calculated from the original projected scores of all employees in the business stability anchor set.

[0087] The current standard deviation is the standard deviation of the original projected scores of the business stability anchor set under the current model version, which is calculated from the original projected scores of all employees in the business stability anchor set.

[0088] The sample size of the anchor set is the total number of employees of the enterprises included in the business stable anchor set, which is obtained by statistical analysis of the actual employee composition of the business stable anchor set.

[0089] The original projected score is the projection value of the employee's associated semantic representation onto the principal axis of the tier score, which is calculated by vector projection of the employee's associated semantic representation and the principal axis of the tier score.

[0090] The reference model version is the version identifier of the cloud-based performance management model that serves as the benchmark for dimensional calibration. It is preferably the benchmark version number of the cloud-based performance management model that was first launched and completed enterprise performance calibration. The scoring distribution of this version has been verified by the enterprise's actual business and can serve as a unified reference standard for cross-version and cross-organizational scoring dimensional calibration.

[0091] The reference mean is the arithmetic mean of the original projected scores of the business stability anchor set under the reference model version, which is calculated from the original projected scores of all employees in the business stability anchor set.

[0092] The reference standard deviation is the standard deviation of the original projected scores of the business stability anchor set under the reference model version, which is calculated from the original projected scores of all employees in the business stability anchor set.

[0093] The zero-prevention minimum is a very small value set to avoid the denominator being zero when calculating the scaling factor. It is preferably 10 to the power of negative 6. This value is extremely small, which can effectively avoid the numerical calculation error of the denominator being zero, and will not substantially interfere with the actual calculation result of the scaling factor.

[0094] The scaling factor is a coefficient used to scale the score of the current model version. It is calculated by a fixed formula from the reference standard deviation, the current standard deviation, and the zero minima.

[0095] The bias parameter is a coefficient used to numerically shift and calibrate the score of the current model version. It is calculated from the reference mean, the scaling factor, and the current mean using a fixed formula.

[0096] It should be noted that this invention overcomes the problem of lacking a unified and stable anchor point for cross-version scoring calibration in traditional performance management. Instead of using a general score adjustment method without business basis, it selects a set of candidate templates with stable business semantics based on three hard conditions: template stability attribute, personnel coverage, and measurable result attribute. Then, it selects employees whose task completion numbers fall within a set median range to form a set of stable business anchor points, avoiding bias in statistical results caused by employees with extreme task completion numbers. Subsequently, it calculates the mean and standard deviation of the anchor point set's scores under both the reference model version and the current model version. Dividing the reference standard deviation by the sum of the current standard deviation and the zero-prevention factor yields the scaling coefficient, achieving cross-version scoring scaling calibration. Finally, the bias parameter is obtained by subtracting the product of the scaling coefficient and the current mean from the reference mean. Cross-version scoring numerical shift calibration provides clear business anchors and mathematical basis for dimensional calibration. For example, a company selects fault handling templates from its work order system that have had no field changes in the past four quarters, cover more than 60% of employees, and include completion and deadline times as a candidate template set. It then selects 200 employees whose number of completed tasks in this template falls between the 30th and 70th percentiles to form an anchor set. The calculation shows that this set has a reference mean of 80 and a reference standard deviation of 10 under the reference model version, and a current mean of 75 and a current standard deviation of 8 under the current model version. Substituting these values ​​into the formula, the scaling factor is approximately 1.25 (10 divided by 8 plus 10 to the power of -6). The bias parameter is 80 minus 1.25 multiplied by 75, which equals -6.25. These two parameters allow for accurate dimensional calibration of the current model version's scoring.

[0097] It should be noted that the specific criteria for determining the template stability attribute are that the hash values ​​of all fields of the task template have not changed in the past four quarters; the specific threshold for personnel coverage is that the proportion of employees covered by the template within the corresponding organizational unit is not less than 60%; the specific criteria for determining the measurable attribute are that the task record corresponding to the template contains both the task deadline and actual completion time fields; the median range for the number of tasks completed on the candidate template set is set to the 30th to 70th percentile, that is, employees whose completion numbers fall within this percentile range constitute the business stability anchor set; the specific selection rule for the reference model version is to prioritize the cloud performance management model that has been launched for the first time and passed the test. The performance calibration and verification version used by the enterprise's human resources department is selected. If no such version is available, the model version with the most stable historical score distribution is chosen. The stability of the score distribution is judged by the minimum standard deviation coefficient of variation. The specific value for the zero-minimum measure is 10 to the power of -6, which is a fixed value and is not adjusted with the scenario. The specific calculation precision of the original projected score is to retain 4 decimal places, and this precision standard is adopted for the score statistics of all versions. The sample size of the business stability anchor set must meet the requirement of not less than 30 people. If the sample size after the initial screening is lower than this value, the screening range of the candidate template set is expanded, and the time period requirement of the template stability attribute is appropriately reduced until the sample size meets the standard.

[0098] Preferably, based on the scaling coefficient and the bias parameter, a transformation model including an identity matrix and a unidirectional projection scaling mapping is constructed. The associated semantic representation is then axially rescaled, ensuring that model updates and scaling corrections under multi-organizational partitioning only produce anisotropic scaling along the principal axis of the echelon scoring, while maintaining the non-principal axis orthogonal space unchanged, thereby obtaining a corrected score. This includes: Obtain the outward product matrix of the principal axis of the tier rating; multiply the difference between the scale coefficient and the numerical value of one by the outward product matrix to obtain the unidirectional projection scaling map; add the identity matrix to the unidirectional projection scaling map to construct the transformation model; multiply the bias parameter by the principal axis of the tier rating to obtain the axial bias vector, as shown in the following formula: in, For transformation model; It is the identity matrix; This is the scaling factor; The main focus of the team scoring; It is a square matrix that integrates outwards; For unidirectional projection scaling mapping; This is the axial offset vector; This is the bias parameter.

[0099] The transformation model is applied to perform mapping processing on the associated semantic representation, and then added to the axial bias vector to obtain the rescaled representation to complete the axial rescale; the inner product of the echelon scoring axis and the rescaled representation is calculated to obtain the corrected score, as shown in the following formula: in, This is a characterization after recalibration; For semantic representation of association; To correct the score.

[0100] The self-outward product matrix is ​​a square matrix obtained by multiplying the echelon scoring principal axis with its own transpose, and is used to construct a one-way projection scaling map.

[0101] The unidirectional projection scaling mapping is a matrix obtained by multiplying the difference between the scale factor and the numerical value by the self-outward product matrix, and it only applies to the principal axis direction of the echelon scoring.

[0102] The transformation model is a matrix model constructed by adding the identity matrix and the unidirectional projection scaling mapping. It is used to perform axial rescaling of the associated semantic representation. The sum of the identity matrix and the unidirectional projection scaling mapping is preferred because this model can achieve anisotropic scaling only in the principal axis direction of the echelon score, while keeping the non-principal axis orthogonal space unchanged, which meets the cross-version dimension calibration requirements.

[0103] The axial bias vector is a vector obtained by multiplying the bias parameter by the principal axis of the echelon score, and is used to numerically shift the associated semantic representation along the principal axis.

[0104] The recalibrated representation is a feature vector obtained by applying a transformation model to perform mapping processing on the associated semantic representation and then adding it to the axial offset vector. It is a representation after dimensional calibration.

[0105] The corrected score is a value obtained by calculating the inner product of the tier score principal axis and the recalibrated representation. It is a standardized score used for subsequent tier segmentation.

[0106] It should be noted that this invention breaks through the traditional performance management approach of requiring complex transformations of all dimensions for cross-version and cross-organizational dimensional calibration. Instead, it constructs a transformation model that only applies to the main axis of the tiered scoring system. By extracting the self-outward product matrix of the tiered scoring main axis and combining it with scaling coefficients and bias parameters, it achieves unidirectional scaling and translation, eliminating dimensional drift without destroying competency semantics in non-main axis directions. In specific implementation, the fixed tiered scoring main axis is first obtained and its self-outward product matrix is ​​calculated. The difference between the scaling coefficient obtained in claim 5 and one is multiplied by this matrix to obtain a projection scaling mapping that only applies to the main axis. This is then added to the identity matrix to construct the transformation model. Simultaneously, the bias parameter is multiplied by the main axis to obtain the axial bias vector. After applying this transformation model to the employee's associated semantic representation and superimposing the axial bias vector, a rescaled representation is obtained. At this point, the features of the non-main axis orthogonal space remain unchanged, and only the main axis direction undergoes dimensional calibration. Finally, the corrected score is obtained through the inner product of the main axis and the rescaled representation. For example, an organization has a scaling factor of 1.25, a bias parameter of -6.25, and a tiered scoring principal axis that is a six-dimensional vector. Its self-outward product matrix is ​​a sixth-order square matrix. The transformation model is the sum of the difference between the sixth-order identity matrix and 1.25 minus one, multiplied by the sum of the square matrix. After applying this model to map the semantic representation of employee associations, the product vector of the bias parameter and the principal axis is superimposed to obtain the rescaled representation. The inner product of the principal axis and this representation is then calculated, which is the corrected score after eliminating dimensional drift.

[0107] It should be noted that the specific calculation precision requirement for the self-outward product matrix is ​​to retain 6 decimal places. All elements are calculated and stored with this precision to ensure the accuracy of matrix operations. The specific storage and retrieval method for the transformation model is to classify and store it according to the current model version and organizational unit. When retrieving, the corresponding transformation model is accurately matched by version number and organizational identifier. The storage format adopts the standard matrix storage format, which supports fast reading. The specific optimization method for matrix operations in the axial rescaling process is to use batch matrix operations for the semantic representation of associations of a large number of employees. The parallel computing capability of cloud computing is used to improve the computing efficiency. The single batch processing volume is set to 1000 employees to balance computing speed and resource consumption. The precision retention rule for the correction score is to retain 4 decimal places and use rounding to ensure that the scores of different versions and different organizations are comparable. The numerical stability judgment rule for the transformation model to map the semantic representation of associations is to calculate the variance change rate of the non-principal axis orthogonal space features before and after the mapping. If the variance change rate is less than 0.01, it is judged as stable. If it is greater than or equal to 0.01, the calculation process of the scaling coefficient and bias parameter is re-examined, and the set of stable business anchor points is re-selected if necessary.

[0108] Preferably, a fixed allocation threshold is set on the reference distribution benchmark, and the correction scores are segmented and mapped to output tier rating labels, including: In the reference model version, the first allocation threshold, the second allocation threshold, and the third allocation threshold are calculated based on the preset quota quantile ratio and used as the fixed allocation threshold, as shown in the following formula: in, The third allocation threshold; The second allocation threshold; Assign the first threshold; This is a function for calculating quantiles; The cumulative value of the preset quota percentile ratio; Reference model version Sub-organizational division unit The full corrected score.

[0109] The corrected score is compared with the first allocation threshold, the second allocation threshold, and the third allocation threshold using interval values ​​to perform the segmented mapping, and a first tier label, a second tier label, a third tier label, or a fourth tier label is assigned to obtain the tier rating label, as shown in the following formula: in, For employees Tier rating labels; To correct the score.

[0110] Calculate the element-wise product of the tiered score axis and the re-scaled representation to obtain the axis contribution explanation vector; extract the index of the highest contribution dimension based on the absolute value of the axis contribution explanation vector, and map it to the competency dimension name to output the axis contribution explanation, as shown in the following formula: in, The contribution explanation vector of the main axis; The main focus of the team scoring; This is an element-wise multiplication operation; This is the characterization after relabeling.

[0111] The third allocation threshold is the highest tier division critical value calculated based on the preset quota quantile ratio under the reference model version, used to distinguish between the first tier and the second tier.

[0112] The second allocation threshold is the critical value for dividing the middle high tiers, calculated based on the preset quota quantile ratio under the reference model version, and is used to distinguish between the second tier and the third tier.

[0113] The first allocation threshold is the critical value for dividing the middle and lower tiers, calculated based on the preset quota quantile ratio under the reference model version, and is used to distinguish between the third and fourth tiers.

[0114] The quantile calculation function is used to calculate the allocation threshold based on the full correction score and the preset quota quantile ratio. The preferred function is the linear interpolation quantile function because it can accurately calculate the critical value of any quantile, adapt to the allocation requirements of different quota ratios, and has strong calculation stability.

[0115] The preset quota percentile cumulative value is a percentile reference value used to determine each allocation threshold. It is preferably 0.10, 0.30, and 0.80. This set of values ​​corresponds to common performance quota ratios, namely 10% for the first tier, 20% for the second tier, 50% for the third tier, and 20% for the fourth tier, which is in line with the conventional configuration of enterprise performance management.

[0116] The full-scale corrected score of an organizational unit under the reference model version is the set of corrected scores of all employees within the corresponding organizational unit in the reference model version. It is calculated by processing the associated semantic representation of employees within the organizational unit through a transformation model.

[0117] Tier rating labels are performance rating identifiers assigned to employees based on the comparison results of the corrected scores with the intervals of each allocation threshold, including first tier labels, second tier labels, third tier labels, and fourth tier labels.

[0118] The principal axis contribution explanation vector is a vector obtained by element-wise multiplication of the principal axis of the echelon rating and the rescaled representation. It is used to quantify the contribution of each competency dimension to the corrected score.

[0119] Element-wise multiplication is an operator that multiplies corresponding elements of two vectors with the same dimension to obtain a new vector. It is used to calculate the principal axis contribution explanation vector.

[0120] The highest contribution dimension index is the dimension number corresponding to the element with the largest absolute value in the main axis contribution explanation vector, used to locate the competency dimension that has the greatest impact on performance rating.

[0121] The competency dimension name is the specific name of each dimension in the enterprise competency model, which can be collected by retrieving the fixed competency model dimension dictionary in the enterprise human resource management system.

[0122] It should be noted that this invention overcomes the problems of frequent threshold adjustments and uninterpretable rating results in traditional performance management. Under the reference model version, a fixed allocation threshold is calculated based on a preset quota quantile ratio. Subsequent versions will not change due to model updates or organizational adjustments. Only segmented mapping is performed based on the correction score, ensuring the consistency of rating standards across versions and organizations. At the same time, through the element-wise product operation of the tiered scoring axis and the re-rating representation, the axis contribution explanation vector is generated, which accurately locates the core competency dimensions that affect performance rating and achieves the interpretability of rating results. For example, in a reference model version, the 90th quantile of a company's full corrected score is 92, the 70th quantile is 80, and the 20th quantile is 65, corresponding to preset quota quantile ratio cumulative values ​​of 0.10, 0.30, and 0.80. The third allocation threshold of 92, the second allocation threshold of 80, and the first allocation threshold of 65 are calculated. In a later version, an employee's corrected score is 85. After interval comparison, the employee is assigned a second-tier label. At the same time, the principal contribution explanation vector is calculated. If the absolute value of the element corresponding to the delivery quality dimension in the vector is the largest, the delivery quality dimension can be output as the principal contribution dimension, allowing employees to clearly understand their performance advantages.

[0123] It should be noted that the specific rules for setting the cumulative value of the preset quota quantile ratio are as follows: the quota ratio for each tier is determined based on the company's human resources strategy. The quota ratio for the first tier corresponds to 1 minus 0.10, the quota ratio for the second tier corresponds to 0.30 minus 0.10, the quota ratio for the third tier corresponds to 0.80 minus 0.30, and the quota ratio for the fourth tier corresponds to 0.80. Companies can adjust the quota ratios within the range of 5% to 15% for the first tier, 15% to 25% for the second tier, 40% to 60% for the third tier, and 15% to 25% for the fourth tier, according to their actual situation. Adjustments must be recorded in the system version. The specific quantile calculation function is a linear interpolation quantile function. During calculation, the full corrected scores are first sorted in ascending order, and then the interpolation position is determined according to the quantile ratio. The allocation threshold is obtained through linear interpolation of adjacent elements. The specific application rules for the tier rating labels are: the first tier label corresponds to excellent, the second tier label corresponds to good, and the third tier label corresponds to qualified. The fourth-tier labeling needs improvement. The rules for linking labels with employee promotions, salary adjustments, training, and other business actions should be clearly defined by the company's human resources management system. The specific mapping relationship table between the main contribution dimension and the competency dimension is the company's competency model dimension dictionary, which contains a one-to-one correspondence between dimension indices and dimension names, such as index 1 corresponding to communication skills, index 2 corresponding to delivery quality, etc. This dictionary cannot be modified arbitrarily once it is fixed. The specific extraction quantity of the highest contribution dimension is 3, that is, extract the top 3 dimension indices with the largest absolute values ​​of the main contribution explanatory vector and map them to the corresponding competency dimension names. The numerical interpretation rules for the main contribution explanatory vector are as follows: a positive value indicates that the dimension has a positive contribution to the performance rating, a negative value indicates a negative impact, the larger the absolute value, the stronger the impact, an absolute value greater than 0.3 indicates that the dimension is a core influencing factor, an absolute value between 0.1 and 0.3 indicates a general influencing factor, and an absolute value less than 0.1 indicates a small impact.

[0124] like Figure 2 As shown, Figure 2 The presentation showcases an enterprise performance management system architecture based on cloud computing and artificial intelligence. The enterprise collaboration system on the left (including IM communication, work order processes, code repository, document collaboration, meeting system, OKR system, CRM, etc.) transmits multi-source behavioral data to the performance management service module of the enterprise cloud platform via a log collection gateway. This module uses feature storage (offline / online) as its data foundation, generates iterative model versions (v0 / v1 / v2) through model training and a version library, processes online features using the online inference service, and then completes dimensional calibration by the scoring and correction module in conjunction with correction parameters. Finally, the threshold and label service outputs the corrected score, tier labels (A / B / C / D), and explanations of the main contribution axis, pushing them to each department / job family (A / B / C). Simultaneously, the audit and metadata repository record metadata throughout the entire process, ensuring system traceability and compliance.

[0125] Example 2: An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the performance management method based on cloud computing and artificial intelligence as described in any one of the examples.

[0126] Example 3: A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the performance management method based on cloud computing and artificial intelligence as described in any one of the examples.

[0127] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. A performance management method based on cloud computing and artificial intelligence, characterized in that, include: The system acquires multi-source collaborative system logs, performs multi-source behavioral data transformation, and combines accurate historical data splicing at specific moments with time decay weights to generate an aggregated representation. The aggregated representation is subjected to multi-view representation learning and block fusion processing, and graph network smoothing operation is introduced to perform semantic matching between the competency model and the historical behavior graph to obtain the associated semantic representation. Based on the sample dispersion, the tiered scoring axis is extracted from the associated semantic representation, and the tiered scoring axis is established as the reference distribution benchmark. Filter the set of stable business anchor points, and calculate the mean and variance of the scores of the set of stable business anchor points under different model versions to obtain the scaling coefficient and bias parameter. Based on the scaling factor and the bias parameter, a transformation model containing the identity matrix and unidirectional projection scaling mapping is constructed. The associated semantic representation is axially rescaled so that the model update and the scaling correction under multi-organization division only produce anisotropic scaling in the principal axis direction of the echelon scoring, while keeping the non-principal axis orthogonal space unchanged, thereby obtaining the correction score. A fixed allocation threshold is set on the reference distribution benchmark, and the correction scores are segmented and mapped to output tier rating labels.

2. The performance management method based on cloud computing and artificial intelligence according to claim 1, characterized in that, The system acquires multi-source collaborative system logs, performs multi-source behavioral data transformation, and combines accurate historical data at specific points in time with time decay weights to generate an aggregated representation, including: Extract event records from the multi-source collaborative system logs, and calculate event weights by combining event baseline value, completion status, task severity, and rework markers; The event weights are summed and divided by the task opportunity scale penalty term to obtain the behavior intensity quantification value, thus completing the multi-source behavior data conversion. The historical data at the specified time point is accurately spliced ​​to obtain the historical behavior sequence; A time decay diagonal matrix is ​​constructed based on the half-life mechanism as the time decay weight, and a dimension expansion operation is performed on the historical behavior sequence to generate the aggregated representation.

3. The performance management method based on cloud computing and artificial intelligence according to claim 2, characterized in that, The aggregated representation is subjected to multi-view representation learning and block fusion processing, and a graph network smoothing operation is introduced to perform semantic matching between the competency model and the historical behavior graph, resulting in an associated semantic representation, including: Based on the available sample size, a multi-view representation learning model is applied to extract the view output from the aggregated representation; A masking identifier is constructed for the missing source and combined with the view output to generate a processed output and concatenate them into a block structure vector. The block structure vector is then subjected to the block fusion process using a block diagonal fusion weight matrix to obtain uniformly defined dimensional features. Construct the behavior association Laplace matrix of the historical behavior graph based on co-activation associations; The graph network smoothing operation is constructed by combining the identity matrix and the smoothing intensity coefficient. The inverse operation is calculated on the behavior association Laplacian matrix and the unified definition dimensional feature to perform the semantic matching and obtain the association semantic representation.

4. The performance management method based on cloud computing and artificial intelligence according to claim 3, characterized in that, Extracting the tiered scoring axis from the associated semantic representation based on sample dispersion, and establishing the tiered scoring axis as a reference distribution benchmark, including: Calculate the echelon representation center of each echelon in the historical echelon set based on the historical echelon set; The internal discreteness matrix is ​​extracted based on the deviation between the associated semantic representation and the echelon representation center of the corresponding echelon. The stability adjustment coefficient is obtained based on the trace and the total number of dimensions of the internal discreteness matrix. The internal dispersion matrix is ​​corrected by combining the identity matrix and the stability adjustment coefficient to obtain a stable dispersion matrix; By combining the inverse operation of the stable discreteness matrix with the extreme value difference vectors of the highest and lowest echelon representation centers within each echelon representation center, the echelon scoring principal axis is extracted after normalization and solidified as the reference distribution benchmark.

5. A performance management method based on cloud computing and artificial intelligence according to claim 4, characterized in that, Filter the set of stable business anchor points, calculate the mean and variance of the scores of the set of stable business anchor points under different model versions, and obtain the scaling coefficient and bias parameters, including: Candidate template sets are filtered based on template stability attributes, personnel coverage, and measurable outcome attributes. The set of business stability anchors is formed by selecting employees whose number of completed tasks on the candidate template set falls within a set median range. Calculate the reference mean and reference standard deviation of the set of stable business anchor points under the reference model version, and the current mean and current standard deviation under the current model version; The scaling factor is obtained by dividing the reference standard deviation by the sum of the current standard deviation and the zero minima. The bias parameter is obtained by subtracting the product of the scaling coefficient and the current mean from the reference mean.

6. The performance management method based on cloud computing and artificial intelligence according to claim 5, characterized in that, Based on the scaling coefficient and the bias parameter, a transformation model including an identity matrix and a unidirectional projection scaling mapping is constructed. The associated semantic representation is axially rescaled, ensuring that model updates and scaling corrections under multi-organizational partitioning only produce anisotropic scaling along the principal axis of the echelon scoring, while maintaining the non-principal axis orthogonal space unchanged. This process yields a corrected score, including: Obtain the outward product matrix of the tier scoring axis; Multiply the difference between the scale coefficient and the value of one by the self-out product matrix to obtain the unidirectional projection scaling map; The transformation model is constructed by adding the identity matrix to the unidirectional projection scaling map; Multiply the bias parameter by the echelon scoring axis to obtain the axial bias vector; The transformation model is applied to perform mapping processing on the associated semantic representation, and then added to the axial bias vector to obtain the re-labeled representation to complete the axial re-labeling; The correction score is obtained by calculating the inner product of the tier score axis and the recalibrated representation.

7. A performance management method based on cloud computing and artificial intelligence according to claim 6, characterized in that, A fixed allocation threshold is set on the reference distribution benchmark, and the correction scores are segmented and mapped to output tier rating labels, including: In the reference model version, a first allocation threshold, a second allocation threshold, and a third allocation threshold are calculated based on a preset quota percentile ratio and used as the fixed allocation threshold. The corrected score is compared with the first allocation threshold, the second allocation threshold and the third allocation threshold in an interval value comparison to perform the segmented mapping, and a first tier label, a second tier label, a third tier label or a fourth tier label is assigned to obtain the tier rating label; Calculate the element-wise product of the tier score main axis and the re-labeled representation to obtain the main axis contribution explanation vector; The index of the highest contribution dimension is extracted based on the absolute value of the principal axis contribution explanation vector and mapped to the competency dimension name to output the principal axis contribution explanation.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the performance management method based on cloud computing and artificial intelligence as described in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the performance management method based on cloud computing and artificial intelligence as described in any one of claims 1 to 7.

Images