An ai agent personnel optimization configuration method and system based on multi-dimensional performance evaluation

Abstract

The application provides an AI Agent personnel optimization configuration method and system based on multi-dimensional performance evaluation, relates to the technical field of artificial intelligence, and the method comprises the following steps: analyzing target text of an enterprise through an AI Agent to construct a knowledge graph; generating a post value vector based on a post capacity index system; generating an employee capacity vector based on time-series multi-source behavior data of the employee; positioning a target post in response to a received task instruction; generating a task risk portrait through a risk assessment model according to requirement description and execution environment information in the task instruction; determining a matching degree threshold and risk response capacity requirement of personnel configuration based on the grade and combination characteristics of each capacity dimension requirement in the task risk portrait; calculating the multi-dimensional matching degree between the employee capacity vector and the post value vector corresponding to the target post; and generating a personnel configuration scheme by using a multi-objective optimization algorithm, with the goal of maximizing team collaboration performance.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method and system for optimizing the allocation of AI Agent personnel based on multi-dimensional performance evaluation. Background Technology

[0002] As businesses diversify, become more complex, and face dynamic changes in the market environment, they are placing higher demands on staffing. This requires not only matching job positions with employees' basic skills, but also dynamically optimizing staffing based on objectives, task risks, employee collaboration adaptability, and historical collaboration results, in order to maximize team collaboration efficiency and support implementation.

[0003] Existing configuration schemes rely on fixed-rule comparisons and simple similarity calculations based on job and employee basic competency data. They only collect basic job responsibilities and employee basic competency data, failing to collect and analyze information related to objectives and task risks, establishing a link between objectives and job / employee capabilities, or incorporating historical team collaboration data. This leads to a disconnect between the configuration scheme and the overall direction, failing to consider the differentiated requirements of personnel capabilities due to task risks, and struggling to adapt to team collaboration efficiency needs. While the selected personnel may meet basic job requirements, they may not be suited for implementation, risk control, and maximizing collaboration efficiency, easily resulting in inefficient collaboration and failing to support enterprises in efficiently completing complex tasks and achieving objectives.

[0004] Therefore, there is an urgent need for a method and system for optimizing the allocation of AI Agent personnel based on multi-dimensional performance evaluation. Summary of the Invention

[0005] To address the aforementioned technical issues, this application provides a method and system for optimizing the allocation of AI Agent personnel based on multi-dimensional performance evaluation.

[0006] A first aspect of this application provides a method for optimizing the configuration of AI Agent personnel based on multi-dimensional performance evaluation, including: AI Agents analyze target text from enterprises to construct knowledge graphs; Based on the job competency index system, the value of the jobs in the knowledge graph is quantified to generate a job value vector. Based on the time-series multi-source behavioral data of employees in the enterprise, an employee capability vector consistent with the dimensions of the job capability index system is generated through a time-series capability assessment model. In response to the received task instructions, locate the target position associated with the task in the knowledge graph; Based on the requirements description and execution environment information in the task instruction, a task risk profile is generated through a risk assessment model; the task risk profile maps to the required level of the target position's competency dimensions. Based on the level and combination characteristics of the capability dimensions required in the task risk profile, the matching threshold of personnel configuration and the risk response capability requirements are determined through predefined mapping rules. A weighted similarity algorithm is used to calculate the multi-dimensional matching degree between the employee's ability vector and the job value vector corresponding to the target job. Based on the target job position, multi-dimensional matching degree, matching degree threshold, risk response capability requirements, and historical team collaboration data, a multi-objective optimization algorithm is adopted to generate a personnel configuration plan with the goal of maximizing team collaboration efficiency.

[0007] A second aspect of this application provides an AI Agent personnel optimization configuration system based on multi-dimensional performance evaluation, comprising: The knowledge graph construction module is used to parse the target text of an enterprise through an AI Agent and build a knowledge graph. The first vector module is used to quantify the value of the positions in the knowledge graph based on the job competency index system and generate a job value vector. The second vector module is used to generate employee capability vectors that are consistent with the dimensions of the job capability index system based on the time-series multi-source behavioral data of employees in the enterprise and through the time-series capability assessment model. The job location module is used to locate the target job associated with the task in the knowledge graph in response to the received task instructions; The risk determination module is used to generate a task risk profile based on the requirement description and execution environment information in the task instruction, through a risk assessment model; the task risk profile maps to the requirement level of the target position's capability dimension. The threshold determination module is used to determine the matching degree threshold and risk response capability requirements of personnel configuration based on the level and combination characteristics of the capability dimensions required in the task risk profile, through predefined mapping rules. The matching calculation module is used to calculate the multi-dimensional matching degree between the employee's ability vector and the job value vector corresponding to the target job using a weighted similarity algorithm; The scheme generation module is used to generate a personnel configuration scheme based on the target position, multi-dimensional matching degree, the matching degree threshold, the risk response capability requirements and historical team collaboration data, using a multi-objective optimization algorithm with the goal of maximizing team collaboration efficiency.

[0008] A third aspect of this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of the above-described AI Agent personnel optimization configuration method based on multi-dimensional performance evaluation.

[0009] A fourth aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described AI Agent personnel optimization configuration method based on multi-dimensional performance evaluation.

[0010] The beneficial effects of the AI ​​Agent personnel optimization configuration method and system based on multi-dimensional performance evaluation provided in this application are as follows: This application constructs a knowledge graph by parsing the target through AI Agent, achieving precise association with the job; the job value vector quantifies the core requirements of the job, and the employee ability vector accurately depicts the employee's strength, both providing a data foundation for matching. Combining task risk profiles to locate the capability requirement level, the matching threshold and risk response standards are dynamically determined to avoid blind matching. A weighted similarity algorithm achieves multi-dimensional precise matching, and a multi-objective optimization algorithm integrates availability and collaboration data. Each technical feature is progressive and mutually supportive, ultimately achieving deep adaptation between personnel configuration and target / task requirements, improving the scientificity and rationality of the configuration, effectively avoiding human resource waste and configuration imbalance, and enhancing the feasibility and effectiveness of the personnel configuration plan. Attached Figure Description

[0011] Figure 1 A flowchart illustrating an AI Agent personnel optimization configuration method based on multi-dimensional performance evaluation, provided in an embodiment of this application; Figure 2 A structural block diagram of an AI Agent personnel optimization configuration system based on multi-dimensional performance evaluation provided in an embodiment of this application; Figure 3 This is a schematic block diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0012] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0013] To make the purpose, technical solution, and advantages of this application clearer, the following will be described in conjunction with the appendix.Figures 1-3 The following is an explanation using specific examples.

[0014] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of an AI Agent personnel optimization configuration method based on multi-dimensional performance evaluation provided in this application. The method includes: S101: Use an AI Agent to analyze the target text of an enterprise and build a knowledge graph.

[0015] In this embodiment, an AI Agent refers to an intelligent entity with autonomous perception, decision-making, and action capabilities, which can perform specific tasks and optimize its behavior by interacting with the environment. In this application, the AI ​​Agent is configured to parse text, collect data, and perform evaluation and optimization operations to achieve intelligent and automated personnel allocation. A knowledge graph refers to a knowledge base that uses a graph structure to represent elements such as enterprise goals, business processes, job responsibilities, required capabilities, resource allocation, and their interrelationships. It uses nodes to represent entities and edges to represent relationships between entities, providing a global view and structured information support for personnel allocation.

[0016] Specifically, the AI ​​Agent can be configured to receive textual information such as planning documents, annual target reports, and departmental responsibility statements issued by enterprises. Through the Natural Language Processing (NLP) module and knowledge extraction model integrated into the AI ​​Agent, the target text of the enterprise is analyzed in layers. First, core indicators (such as market expansion, cost control, and technological innovation), phased goals, and implementation requirements are extracted. Then, the relationships between the goals and various business processes and job systems are explored, clarifying the responsibilities and collaborative logic of different positions in implementation. Based on the analysis results, a knowledge graph is constructed using a Graph Neural Network (GNN). Graph nodes include core elements such as goals, business modules, job categories, job responsibilities, and competency requirements. Edges are used to represent the strength of the relationship between nodes (such as the contribution of a position to the goal and the synergy coefficient between positions). A dynamic update mechanism is used to synchronize enterprise adjustments in real time, ensuring the timeliness and accuracy of the knowledge graph.

[0017] Specifically, the graph neural network (GNN) adopts a two-layer architecture based on a graph attention network (GAT). The first GAT layer is used to learn the embedding representation of entity nodes (such as targets, positions), with 4 attention heads and an output dimension of 128. The second GAT layer is used to learn the embedding representation of relation edges (such as contribution and collaboration coefficients), with an output dimension of 64. The training process for graph construction adopts a self-supervised learning approach, randomly masking some nodes or edges and using a graph autoencoder (GAE) for reconstruction to learn a robust graph structure representation. During training, the Adam optimizer is used with a learning rate of 0.001, a batch size of 32, and a training cycle of 100 rounds. The initial embedding of the knowledge graph is initialized using pre-trained word vectors (such as Word2Vec or BERT) from enterprise historical documents to incorporate semantic information.

[0018] S102: Based on the job competency index system, the value of jobs in the knowledge graph is quantified to generate job value vectors.

[0019] In this embodiment, the job competency indicator system refers to a set of standards and dimensions used to measure and evaluate the various competencies required for a job. This system typically includes multiple competency dimensions, each with specific indicators to quantify the value and requirements of the job. The job value vector, after quantifying the value of jobs in the knowledge graph, is a vector representing the importance or required level of a job across each competency dimension. This vector characterizes the job's potential contribution to the enterprise and the competency requirements for employees.

[0020] First, a standardized job competency indicator system is constructed. This system includes four dimensions: core competencies (professional competencies directly related to the core responsibilities of the job), general competencies (communication and collaboration, problem-solving, learning ability, etc.), adaptability (adaptability to enterprise implementation), and risk response capability (ability to handle risks related to job-related tasks). Each dimension is further divided into several third-level indicators, and weights are assigned to each indicator (dynamically calibrated using the Analytic Hierarchy Process (AHP) combined with enterprise needs). Based on this indicator system, the value of each job in the knowledge graph is quantified in multiple dimensions. Combining factors such as the job's contribution to implementation, the complexity of job responsibilities, job scarcity, and performance impact weight, the indicators of each dimension are normalized using the entropy weight method. The quantification results are mapped to a fixed-dimensional job value vector, where each component of the vector corresponds to a specific indicator in the job competency indicator system, ensuring that the job value vector can comprehensively and accurately represent the job's competency requirements and value positioning.

[0021] S103: Based on the time-series multi-source behavioral data of employees in the enterprise, an employee capability vector with dimensions consistent with the job capability indicator system is generated through a time-series capability assessment model.

[0022] In this embodiment, time-series multi-source behavioral data refers to employee behavioral data with time-series characteristics collected from various sources, including internal and external systems of the enterprise. This data may include, but is not limited to, employee project participation records, performance appraisal results, training and learning records, communication and collaboration logs, and skills certification information, used to dynamically represent changes and developments in employee capabilities. The time-series capability assessment model is a model that dynamically assesses and predicts employee capabilities across various dimensions of the job capability indicator system based on employee time-series multi-source behavioral data and through machine learning or deep learning methods. This model can capture the evolution trend of employee capabilities and generate a vector representing their current capability level. The employee capability vector is a vector representing the employee's capability level across various dimensions of the job capability indicator system in numerical form after the employee's capabilities have been assessed by the time-series capability assessment model. This vector is used to match with the job value vector to assess the fit between the employee and the job.

[0023] Specifically, the process involves collecting time-series multi-source behavioral data from employees within the enterprise. Data sources include employee performance data (monthly / quarterly / annual performance ratings, performance improvement records), work behavior data (task completion quality, task response speed, frequency of cross-departmental collaboration), skills enhancement data (training participation, assessment pass rate, skills certification results), and job suitability data (current job tenure, job adjustment records, performance evaluations). The collected data is cleaned, deduplicated, and normalized, and outliers are removed to construct an employee time-series behavioral dataset. A time-series competency assessment model (based on a fusion architecture of LSTM and Transformer) is then built. The employee time-series behavioral dataset is input into the model, which captures the changing trends of employee competency over time and the correlation characteristics between different behavioral data. The model dynamically evaluates employee performance across various dimensions of the job competency indicator system, generating an employee competency vector consistent with the job value vector dimension. Each component of the vector corresponds to the employee's score on that competency indicator.

[0024] The specific architecture of the temporal competence assessment model in this embodiment is as follows: First, two bidirectional LSTM layers are used to process the temporal behavior sequences of employees, with 256 hidden units in each layer, to capture long-term dependencies in competence. Then, the final hidden state sequence of the LSTM is input into a Transformer encoder layer, which contains four attention heads and a feedforward network dimension of 512, used to capture global dependencies and interaction features between different behavioral data. The model's output layer is a fully connected layer using the Sigmoid activation function, and the output dimension is consistent with the dimension of the job competence indicator system. The model is trained using the mean squared error (MSE) loss function, with the AdamW optimizer and an initial learning rate of 1. e-4The model employs cosine annealing scheduling. Training data consists of employee behavior sequences from those who participated in at least three projects within the past 24 months, divided into training, validation, and test sets in a 7:2:1 ratio. The model uses early stopping (patience=10) to prevent overfitting.

[0025] S104: In response to the received task instructions, locate the target position associated with the task in the knowledge graph.

[0026] In this embodiment, a task instruction refers to an instruction received by the system that includes information such as specific task requirements, objectives, time limits, and resource requirements. This instruction is the initial input that triggers the personnel configuration process.

[0027] Specifically, the AI ​​Agent receives task instructions from enterprises and uses the Natural Language Understanding (NLU) module to parse the core information in the task instructions, including the task objectives, task type, execution time limit, resource requirements, quality standards, and other requirement descriptions, as well as internal and external environmental information during task execution (existing team configuration, market environment, technical constraints, etc.). Based on the parsed core information, it performs association retrieval and matching in the constructed knowledge graph, and uses graph retrieval algorithms to locate target positions that are directly related to or indirectly support the task, clarifying the roles, responsibilities, and collaborative relationships of each target position in task execution. At the same time, it filters out positions that are not related to the task, reducing the computational workload of subsequent personnel allocation and ensuring the accuracy and relevance of target position positioning.

[0028] S105: Based on the requirements description and execution environment information in the task instruction, generate a task risk profile through a risk assessment model; the task risk profile maps to the required level of the target position's competency dimensions.

[0029] In this embodiment, the risk assessment model refers to a model used to analyze the requirement description and execution environment information in task instructions, and to predict the types and levels of risks that the task may face, as well as the special requirements for personnel capabilities. This model outputs a task risk profile. The task risk profile is a portrait generated by the risk assessment model, describing the potential risk characteristics of the task in structured data form. It maps to the required levels of each capability dimension of the target position, guiding the consideration of risk response capabilities when allocating personnel.

[0030] Based on the requirements description and execution environment information in the task instructions, a multi-dimensional task risk assessment model is built. The model covers four core dimensions: task difficulty risk (task complexity, technical threshold, execution difficulty), time risk (the tightness of the execution deadline, the impact of task delay), quality risk (quality standards, consequences of substandard quality), and collaboration risk (difficulty of cross-position / cross-department collaboration, communication costs). Task-related information is input into the risk assessment model, and each dimension of risk is scored using a risk quantification algorithm. The model is calibrated by combining historical task risk data to generate a task risk profile. The task risk profile clearly presents the risk level (high, medium, low) of each dimension, and through predefined risk-capability mapping rules, each dimension of risk level is mapped to the required level of the target position's capability dimension (high difficulty tasks correspond to high core capability requirements, and high collaboration risk corresponds to high communication and collaboration requirements in general capabilities).

[0031] The risk assessment model in this embodiment is a multi-task deep learning model. Its backbone network is a BERT-based text encoder used to extract semantic features from the task requirement description. It also has four parallel sub-network branches, corresponding to four risk dimensions: task difficulty, time, quality, and collaboration. Each branch is a two-layer fully connected network, using Softmax to output the risk level (low, medium, high) for that dimension. The model's training data comes from historical task archives, including task description text, environmental information text, and ground truth level labels for the four risk dimensions annotated post-processed by the project expert group. Training employs a multi-task cross-entropy loss function, with the total loss being the weighted sum of the losses from each branch (the initial weights are all 1, and can be dynamically adjusted according to the classification difficulty of each dimension). The model uses the Adam optimizer, trained for 50 epochs on the training set, and the optimal model is selected on the validation set.

[0032] S106: Based on the level and combination characteristics of the capability dimensions required in the task risk profile, determine the matching threshold of personnel configuration and the risk response capability requirements through predefined mapping rules.

[0033] In this embodiment, based on the task risk profile generated in step 5, the required level and combination features of each capability dimension are extracted (some high-risk tasks require both core capabilities and risk response capabilities to reach a high level, while some collaborative tasks require both general capabilities and adaptability capabilities to meet the standards); combined with the company's human resource management rules and historical personnel allocation experience, a risk-configuration mapping rule is predefined, which clarifies the core constraints of personnel allocation under different risk level combinations; according to this mapping rule, the matching threshold for this personnel allocation is determined (the matching threshold for high-risk tasks is greater than or equal to 85%, for medium-risk tasks it is greater than or equal to 75%, and for low-risk tasks it is greater than or equal to 65%), while clarifying the risk response capability requirements (high-risk tasks require employees to score greater than or equal to 80 points in the risk response capability dimension), ensuring that the personnel allocation not only meets the task capability requirements but also effectively addresses task risks, avoiding task failure or inefficiency caused by improper allocation.

[0034] S107: Employ a weighted similarity algorithm to calculate the multi-dimensional matching degree between the employee's ability vector and the job value vector corresponding to the target job.

[0035] In this embodiment, the multi-dimensional matching degree refers to the comprehensive matching degree between the employee's ability vector and the job value vector corresponding to the target job across multiple ability dimensions, calculated using a weighted similarity algorithm. This matching degree characterizes the employee's comprehensive suitability in terms of skills, experience, and knowledge.

[0036] A weighted similarity algorithm (combining cosine similarity and Manhattan distance) is used to calculate the multi-dimensional matching degree between the employee's ability vector and the corresponding job value vector of the target job. First, based on the weights of each indicator in the job competency indicator system, the corresponding components of the employee's ability vector and the job value vector are weighted to highlight the influence of core competency indicators. Then, the directional consistency between vectors is calculated using cosine similarity (representing the direction of the fit between the employee's ability and the job requirements), and the numerical difference between vectors is calculated using Manhattan distance (representing the degree of fit between the employee's ability and the job requirements). The two calculation results are then normalized and fused to obtain the multi-dimensional matching degree between the employee and the target job. The matching degree ranges from 0 to 1, with a higher value indicating a higher degree of fit between the employee's ability and the job requirements.

[0037] S108: Based on target positions, multi-dimensional matching degree, matching degree threshold, risk response capability requirements, and historical team collaboration data, a multi-objective optimization algorithm is used to generate a personnel allocation plan with the goal of maximizing team collaboration efficiency.

[0038] In this embodiment, the multi-objective optimization algorithm refers to an algorithm that can simultaneously optimize multiple conflicting or interrelated objectives (such as maximizing team collaboration effectiveness, minimizing costs, maximizing risk response capabilities, etc.). This algorithm is used in personnel allocation to balance various indicators and generate the optimal allocation scheme.

[0039] Multi-objective optimization algorithms can be configured to consider multiple objectives, such as maximizing the overall match between team members and target positions, ensuring the team has sufficient risk response capabilities, and to some extent considering the smoothness of historical team collaboration. Historical team collaboration data can simply record which employees have had successful collaboration experiences. The algorithm can use heuristic search or greedy algorithms to select a group of employees from a pool of candidates to form a team. During the selection process, the algorithm iteratively evaluates the performance of different team combinations on the above multiple objectives and selects a configuration scheme that maximizes team collaboration effectiveness based on preset priorities or weights.

[0040] As can be seen from the above, this application, by adopting the aforementioned technical solution, uses an AI Agent to analyze the target and construct a knowledge graph, achieving precise association with the job position; the job value vector quantifies the core requirements of the job, and the employee capability vector accurately depicts the employee's strength, both providing a data foundation for matching. Combining task risk profiles to locate capability requirement levels, the matching threshold and risk response standards are dynamically determined to avoid blind matching. A weighted similarity algorithm achieves multi-dimensional precise matching, and a multi-objective optimization algorithm integrates usability and collaboration data. These technical features are progressively enhanced and mutually supportive, ultimately achieving deep adaptation between personnel allocation and target / task requirements, improving the scientific and rational nature of the allocation, effectively avoiding human resource waste and allocation imbalance, and enhancing the feasibility and effectiveness of the personnel allocation plan.

[0041] In one embodiment of this application, based on a job competency index system, the value of jobs in a knowledge graph is quantified to generate a job value vector, including: From the knowledge graph, target nodes, business process nodes, and skill nodes associated with the job position are extracted to obtain the set of job capability dimensions; The Analytic Hierarchy Process (AHP) is used to compare the relative importance of each capability in the capability dimension set and determine the initial weight vector of each capability dimension. Based on the correlation analysis between historical project success data and resource consumption of each capability dimension, the initial weight vector is calibrated through regression model to obtain the final weight vector; Based on the knowledge graph, the graph centrality index of the target node and business process node associated with the position is calculated. Combined with the contribution coefficient of the corresponding ability dimension of the position to the success of the project in the historical project data, the factor point method is used to generate the initial score vector of the position's ability. The initial job competency score vector and the final weight vector are weighted and calculated to generate the job value vector.

[0042] In this embodiment, the knowledge graph is a structured knowledge base that includes information such as the enterprise's goals, business processes, required skills, and job responsibilities, and represents the complex relationships between them in the form of nodes and edges. By extracting these related nodes, it can be ensured that the defined set of capability dimensions is based on the enterprise's actual and operational needs, rather than generalized capabilities. For example, graph traversal algorithms, such as breadth-first search (BFS) or depth-first search (DFS), can be used to search and collect all relevant target nodes, business process nodes, and skill nodes from the knowledge graph, starting from the target job node, along predefined relationships. The names or attributes of these nodes will constitute the set of capability dimensions for the job. By summarizing and deduplicating these query results, the set of capability dimensions for the job can be obtained.

[0043] The Analytic Hierarchy Process (AHP) is a multi-criteria decision-making method that systematically determines the weights of each factor by decomposing complex problems into multiple levels and making pairwise comparisons. This helps transform the experience-based judgments of experts or management into quantifiable weights, reducing subjectivity and arbitrariness, and improving the objectivity and consistency of weights. For example, domain experts or senior managers within an organization can be organized to make pairwise comparisons of various capabilities within a set of capability dimensions. Experts assess the importance of one capability dimension relative to another in job success using a pre-defined scale (e.g., a 1-9 scale, representing importance from equally important to extremely important). These comparison results are collected, and the mathematical model of the AHP is used to calculate the eigenvectors of each capability dimension. After normalization, the initial weight vectors are obtained.

[0044] In the step of calibrating the initial weight vector using a regression model based on the correlation analysis between historical project success data and resource consumption across various capability dimensions, the aim is to empirically calibrate the initially determined capability dimension weights by incorporating historical project data, making them more closely reflect actual business performance. The correlation analysis between historical project success data and resource consumption across various capability dimensions reveals which capability dimensions play a key role in the success of actual projects, and their return on investment. The regression model quantifies this correlation, thereby objectively adjusting the initial weights to make them more predictive and practical. For example, a large amount of historical project success data can be collected (e.g., whether projects were completed on time, whether expected returns were achieved, customer satisfaction, etc.) and resource data for each project across different capability dimensions (e.g., manpower, time, budget). A regression model (e.g., multiple linear regression or logistic regression) is constructed, using resource consumption across capability dimensions as independent variables and project success as the dependent variable. After model training, the initial weight vector is adjusted based on the regression coefficients or feature importance of each capability dimension; for example, the weights of capability dimensions with larger regression coefficients are appropriately increased to obtain the final weight vector.

[0045] This embodiment uses a ridge regression model. The input features are standardized resource consumption data (such as manpower hours and financial input) for each capability dimension in historical projects, and the output label is the project success coefficient (range 0-1) rated by experts. Before training, the features are Z-score standardized. The model's regularization parameters... The optimal value is typically determined by grid search combined with 5-fold cross-validation, and usually lies between 0.1 and 1.0. Training uses the least squares method to ensure that the weight calibration both fits historical performance and avoids overfitting. After training, the regression coefficients are normalized and then weighted with the initial AHP weights (weight ratio: regression model confidence: expert weight = 7:3) to obtain the final weight vector.

[0046] The graph centrality index in this embodiment can characterize the influence or connectivity of the targets and business processes associated with a job within the entire knowledge graph, while the historical project contribution coefficient directly quantifies the actual driving effect of job capabilities on project outcomes. The factor-point method provides a structured way to integrate this multi-source information. For example, firstly, in the knowledge graph, the graph centrality index of target nodes and business process nodes associated with the target job is calculated. For instance, the degree centrality, betweenness centrality, or closeness centrality of these nodes can be calculated to measure their importance in information flow or control flow. Secondly, from historical project data, the contribution coefficient of the job to project success in different capability dimensions is statistically analyzed or calculated using a causal inference model. Finally, using the factor-point method, a scoring rule is set for each capability dimension. For example, the graph centrality index and contribution coefficient are mapped to certain score intervals, and then these scores are weighted and summed to obtain the initial score of the job in that capability dimension. The initial scores of all capability dimensions together constitute the initial job capability score vector.

[0047] In the step of generating a job value vector by weighting the initial job competency score vector with the final weight vector, this is the final step in quantifying job value. It integrates the empirically calibrated weights of each competency dimension with the initial job competency score that comprehensively considers importance and historical contribution, thereby generating a quantitative representation that can fully characterize job value. This job value vector will serve as the core basis for subsequent personnel allocation matching, ensuring that personnel allocation considers not only basic competencies but also guidance and actual effectiveness. For example, the initial job competency score vector and the final weight vector can be weighted and summed. Each element of the job value vector (corresponding to a competency dimension) can be obtained by multiplying the initial score of that competency dimension by its corresponding final weight. For example, if the initial job competency score vector is [S1,S2,...,Sn] and the final weight vector is [W1,W2,...,Wn], then the job value vector can be represented as [S1×W1,S2×W2,...,Sn×Wn].

[0048] This embodiment can also train a prediction model based on historical project data, job-related node features, and expert rating samples, and use the model to calibrate the initial job competency rating vector to obtain the final job competency rating vector, which is then used to generate the job value vector.

[0049] As can be seen from the above, this application, by extracting target nodes, business process nodes, and skill nodes associated with the job from the knowledge graph, ensures that the constructed set of job competency dimensions closely aligns with the actual situation of the enterprise and business, avoiding a disconnect between competency definitions and reality. The analytic hierarchy process (AHP) is used to determine the initial weight vector, giving each competency dimension a systematic and relatively objective importance assessment. Based on this, a regression model is used to calibrate the initial weight vector using a correlation analysis between historical project success data and the resources consumed by each competency dimension. This ensures that the final weight vector accurately represents the effectiveness contribution of each competency dimension in actual projects, strengthening the empirical basis of the weights. Simultaneously, by combining the graph centrality index in the knowledge graph and the contribution coefficients in historical project data, a factor-point method is used to generate the initial job competency scoring vector. This ensures that the evaluation of job competencies considers not only their structural importance in the network but also their actual driving force for project success, avoiding the one-sidedness of a single-dimensional evaluation. Finally, the initial job competency scoring vector and the final weight vector are weighted and calculated to generate a comprehensive, accurate, and directional job value vector. This job value vector can more accurately represent the importance of a job and its actual performance contribution, thereby improving the matching accuracy and optimization effect of personnel allocation.

[0050] In one embodiment of this application, a weighted similarity algorithm is used to calculate the multi-dimensional matching degree between the employee's ability vector and the job value vector corresponding to the target job, including: Based on the required skill levels of each position mapped in the task risk profile, the weights of the corresponding skill dimensions in the weighted similarity algorithm are adjusted to obtain a dynamic weight vector. The weighted cosine similarity calculation is adopted. After weighting the employee's ability vector and the target position's job value vector with dynamic weight vector, the similarity between the two is calculated as the comprehensive matching degree. The overall matching degree and the individual matching degrees of each capability dimension are combined to form the multi-dimensional matching degree.

[0051] In this embodiment, the importance of each capability dimension in calculating the employee-job fit is dynamically adjusted based on the risk characteristics of a specific task. The task risk profile is generated by a risk assessment model based on the requirements description and execution environment information in the task instructions, mapping the required levels of the target job's capability dimensions. For example, for a high-risk, high-tech task, the required levels for professional skills and stress resistance may be higher; while for a task emphasizing teamwork, the required levels for communication skills and collaborative spirit may be higher. By converting these required levels into weights, the fit calculation can more accurately represent the actual needs of the task. The required levels of each capability dimension in the task risk profile can be directly linearly mapped to weight values; for example, low, medium, and high required levels correspond to weights of 0.5, 1.0, and 1.5, respectively. Alternatively, a preset weight adjustment function can be used, taking the required levels as input and combining them with the overall risk level of the task to non-linearly generate weights; for example, for high-risk tasks, the weights of capability dimensions with high required levels will be exponentially increased.

[0052] This embodiment employs weighted cosine similarity calculation. A dynamic weight vector is used to weight the employee's ability vector and the target position's value vector, and then the similarity between the two is calculated as the overall matching degree. This step aims to quantify the overall matching degree between employee abilities and target position value, while considering the varying importance of different ability dimensions in the current task. Cosine similarity is a commonly used method to measure the similarity between two non-zero vectors. It determines their similarity by calculating the cosine of the angle between the two vectors; the closer the value is to 1, the more similar they are. Introducing dynamic weights means that before calculating the similarity, the employee's ability vector and the position value vector are weighted according to the dynamic weight vector, thus highlighting the influence of key ability dimensions required for the task. The specific formula for weighted cosine similarity calculation can be expressed as: Similarity(A,B)=(W×A)·(W×B) / (||W×A||×||W×B||), Where A is the employee capability vector, B is the job value vector, W is the dynamic weight vector (element-wise multiplication), · represents the dot product, and || represents the L2 norm of the vector.

[0053] Finally, the overall matching score and the individual matching scores for each capability dimension are combined to form a multi-dimensional matching score. This step aims to provide a comprehensive and detailed matching score assessment. The overall matching score characterizes the degree of fit between the employee and the position, while the individual matching scores for each capability dimension reveal the employee's match with the job requirements in each specific capability dimension. Combining the two not only provides a holistic view but also insight into the details, which is of great significance for subsequent personnel allocation decisions and capability development planning. For example, the multi-dimensional matching score can be represented as a vector or structure that includes the overall matching score and all individual matching scores, such as [overall matching score, individual matching score_capability 1, individual matching score_capability 2, ..., individual matching score_capability N].

[0054] As can be seen from the above, this application can flexibly adjust the importance of different capability dimensions in the matching degree calculation according to the actual risk characteristics of the task. This makes the matching degree assessment no longer static and universal, but can accurately represent the differentiated requirements of the current task for specific capability dimensions. For example, for a task that requires high innovation capabilities, the weight of the innovation capability dimension will be significantly increased, thus prioritizing the matching of employees with strong innovation capabilities. This calculation method makes full use of dynamic weight vectors, ensuring that the comprehensive matching degree can accurately reflect the employee's strengths or weaknesses in the key capability dimensions of the task, thereby solving the problem of not being able to accurately capture the overall matching due to the non-dynamic change of weights. At the same time, by combining the comprehensive matching degree with the sub-item matching degree of each capability dimension to form a multi-dimensional matching degree, it not only improves the overall fit between the employee and the target position, but also shows in detail the matching situation of the employee in each specific capability dimension. This multi-dimensional assessment perspective enables personnel allocation decision-makers to have a comprehensive understanding of employees' competency structure, grasp the overall fit, and identify potential competency gaps or strengths. This avoids the shortcomings of a single fit index that cannot represent detailed differences, greatly improving the accuracy and effectiveness of personnel allocation, ensuring that the allocated personnel can better cope with task risks and maximize team collaboration efficiency.

[0055] In one embodiment of this application, an AI Agent personnel optimization configuration method based on multi-dimensional performance evaluation further includes: By analyzing employees’ historical task decision-making data and behavioral feedback, an employee risk preference vector is constructed. The risk preference vector represents the employee’s decision-making tendency and tolerance threshold under different types of risks. The risk assessment model maps the task risk profile into a risk feature vector. Calculate the similarity between the risk preference vector and the task risk feature vector as the risk fit. The overall matching degree, the sub-matching degree of each capability dimension, and the risk suitability are weighted and fused together using preset fusion rules to form the final multi-dimensional matching degree; among them, the weights in the fusion rules are adjusted according to the overall risk level of the task risk profile.

[0056] In this embodiment, constructing an employee risk preference vector aims to quantify the decision-making tendencies and risk tolerance thresholds exhibited by employees when facing different risk situations, thereby characterizing their risk attitudes and behavioral patterns in actual work. For example, data can be collected on employees' handling of risk events in historical tasks, their chosen solutions (e.g., whether to choose a conservative or aggressive solution), and their coping strategies after a risk occurs. This data is then trained using machine learning models (decision trees, support vector machines, or neural networks) to identify employees' decision-making patterns under different risk types (technological risk, market risk, time risk, etc.) and quantify them into vector form.

[0057] Subsequently, the task risk profile is mapped into a risk feature vector using a risk assessment model. This step aims to transform the high-level information, including qualitative descriptions, of the task risk profile into a standardized, computable numerical vector form, facilitating subsequent comparison and matching with employee risk preference vectors. For example, deep learning models (convolutional neural networks or recurrent neural networks) can be used to learn from the raw data of the task risk profile (e.g., task description text, historical risk case data, etc.) to extract and generate vector representations that represent the essential characteristics of task risk.

[0058] Based on this, the similarity between the risk preference vector and the task risk feature vector is calculated as the risk fit. This step quantifies the degree of fit between the employee's risk preference and the task's risk characteristics, that is, the consistency between the employee's risk attitude and tolerance and the task's risk requirements when facing the task's risk. For example, the similarity can be represented by calculating the Euclidean distance between the two vectors and then taking its reciprocal, or by transforming it using a Gaussian kernel function. The smaller the distance, the higher the similarity, indicating that the employee's risk preference is closer to the task's risk characteristics.

[0059] Finally, the overall matching degree, the individual matching degrees of each capability dimension, and the risk suitability are weighted and fused using pre-defined fusion rules to form the final multi-dimensional matching degree. The weights in the fusion rules are adjusted according to the overall risk level of the task risk profile. This step aims to integrate matching information from multiple dimensions (including capability matching and risk matching) into a unified and more comprehensive matching degree indicator, and to dynamically adjust the importance of each dimension based on the risk characteristics of the task. For example, the fusion rules can be implemented using a fuzzy logic system or an expert system. Based on the overall risk level of the task risk profile, a series of fuzzy rules are defined, taking the overall matching degree, individual matching degree, and risk suitability as inputs, and outputting the final multi-dimensional matching degree through a fuzzy inference mechanism.

[0060] As can be seen from the above, this application, by adopting the aforementioned technical solution, accurately characterizes employees' risk preference vectors to depict their risk decision-making tendencies, quantifies task risk attributes by using task risk feature vectors, and achieves accurate assessment of risk suitability through similarity calculation between the two. The comprehensive matching degree, sub-item matching degree, and risk suitability degree are weighted and integrated, with the integration weight dynamically adjusted according to the task risk level, achieving a scientific integration of multi-dimensional suitability indicators. The synergistic effect of these technical features overcomes the shortcomings of traditional matching methods that only focus on ability and ignore risk suitability, making the matching degree assessment more comprehensive. This ensures that the assigned employees not only meet the skill requirements but also that their risk tolerance and decision-making tendencies are adapted to the task requirements, reducing potential risks during task execution and improving the stability of the matching scheme.

[0061] In one embodiment of this application, an AI Agent personnel optimization configuration method based on multi-dimensional performance evaluation further includes: After scheduling the target personnel to perform tasks according to the personnel allocation plan, the AI ​​Agent collects process data and result data during the task execution process. Based on preset process rules and performance baselines, correlation analysis is performed on process data to identify individual behavioral anomalies and inefficient links in team collaboration, and verified collaborative behavior data is recorded. Input the results data, process efficiency indicators, and collaborative behavior data into the task decision evaluation model to generate a comprehensive evaluation report and quantify the contribution of each member to the task objectives. The parameters of the time-series capability assessment model are updated based on the comprehensive assessment report, contribution level, and identified anomalies.

[0062] In this embodiment, process data represents the dynamic trajectory of task execution and employee behavior patterns, while outcome data reflects the quality and effectiveness of task completion. The AI ​​Agent can integrate with the API interfaces of internal enterprise collaboration platforms (project management systems, instant messaging tools, code repositories, etc.) to capture, in real time, employee operation logs, communication records, document modification history, code commit records, etc., as process data during task execution. Simultaneously, by connecting with business systems, it obtains task completion status, output quality scores, customer feedback, etc., as outcome data. Outcome data can be obtained through manual evaluation input, automated test results, or in conjunction with KPI systems.

[0063] Building upon this foundation, and based on pre-defined process rules and performance baselines, correlation analysis is performed on process data to identify individual behavioral anomalies and inefficient team collaboration processes, and to record validated collaborative behavior data. This step, by comparing and analyzing the collected process data against established standards, aims to uncover efficiency bottlenecks, non-compliant operations, or inefficient collaboration patterns during task execution, while simultaneously identifying collaborative behaviors that positively impact task success, providing specific improvement directions and experience accumulation for subsequent optimization. Process rules can be represented as decision trees or rule engines built based on expert experience, while performance baselines can be performance indicators calculated from a large amount of historical data using statistical methods (mean, standard deviation). Correlation analysis can utilize graph neural networks (GNNs) to model the team collaboration network, analyzing the characteristics of nodes (employees) and edges (collaboration relationships), combined with pre-defined rules and baselines, to detect abnormal interaction patterns or inefficient collaboration paths. Valid collaborative behavior data can be used to identify common collaboration patterns among high-performing teams through cluster analysis, and these are recorded as positive samples.

[0064] Subsequently, the results data, process performance indicators, and collaborative behavior data are input into the task decision evaluation model to generate a comprehensive evaluation report and quantify each member's contribution to the task objective. This step integrates multi-source feedback information and performs comprehensive analysis through a specialized evaluation model to comprehensively evaluate the overall task execution and accurately measure each team member's specific contribution to task completion. The task decision evaluation model can be a multi-input, multi-output machine learning model, such as a deep learning-based sequence model or ensemble learning model. Results data (such as task completion rate and quality score), process performance indicators (such as efficiency and resource utilization), and collaborative behavior data (such as the number of effective communications and participation in key problem-solving) serve as input features to the model. Through pre-training or online learning, the model outputs a comprehensive evaluation report (including overall task performance, risk exposure, etc.) and the contribution of each member (e.g., calculating each member's marginal contribution to task success using interpretable AI methods such as Shapley scores and LIME).

[0065] Finally, the parameters of the time-series competency assessment model are updated based on the comprehensive evaluation report, contribution levels, and identified anomalies. This is the core of the entire closed-loop feedback mechanism, aiming to use real feedback after task execution to correct and optimize the employee competency assessment model, ensuring that the model can dynamically and accurately represent the latest state and development trend of employee competencies, thereby improving the accuracy and adaptability of subsequent personnel allocation. The time-series competency assessment model typically includes a series of parameters to capture the patterns of employee competency changes over time. The update process can employ online learning or incremental learning methods. For example, task success / failure labels and member contributions from the comprehensive evaluation report can be used as supervisory signals, and identified anomalies can be used as negative feedback. Gradient descent or other optimization algorithms can be used to adjust the weights, biases, or hidden state parameters related to the employee competency dimension in the model. A forgetting factor can be introduced during the update to make recent data have a greater impact on the model parameters, thus representing rapid changes in competency. Alternatively, updates can also employ Bayesian inference-based methods. The comprehensive evaluation report and contribution levels can be used as observational evidence, and anomalies as negative observations. A Bayesian network or Kalman filter can be used to update the posterior probability distribution of each competency dimension parameter in the time-series competency assessment model. For example, if an employee demonstrates high contribution and no anomalies in a task, the confidence level of their ability parameters in that task-related ability dimension will increase, or the mean will be adjusted upwards. Conversely, if anomalies occur or the contribution is low, the corresponding parameters will be adjusted downwards.

[0066] As can be seen from the above, this application, by adopting the aforementioned technical solution, enables the AI ​​Agent to collect task execution process and result data, achieving closed-loop data collection and providing a reliable basis for subsequent evaluation and optimization. Correlation analysis of process data accurately identifies abnormal behaviors and inefficient collaboration links, while recording effective collaborative behaviors to clarify optimization directions. The task decision evaluation model generates a comprehensive report and quantifies contribution, providing precise support for model updates, and the time-series capability evaluation model parameters are updated to align with actual execution conditions. These technical features form a closed loop of execution-collection-analysis-evaluation-optimization, continuously improving the accuracy of employee capability assessment, resolving evaluation biases caused by fixed model parameters, and enhancing the dynamic optimization capability of personnel allocation plans.

[0067] In one embodiment of this application, based on preset process rules and performance baselines, correlation analysis is performed on process data to identify individual behavioral anomalies and inefficient aspects of team collaboration, and verified collaborative behavior data is recorded, including: The collaborative events and communication records generated during task execution are mapped to the knowledge graph to form a dynamic collaborative subgraph; Calculate the graph structure similarity and path efficiency metrics between dynamic collaborative subgraphs and collaborative subgraphs formed by high-performing teams under similar task backgrounds in the historical success case library; When the graph structure similarity is less than a preset similarity threshold or the path efficiency index is less than a preset efficiency threshold, locate the subgraph structure unit or efficiency bottleneck node with the greatest difference and mark it as an inefficient link in the collaboration.

[0068] In this embodiment, collaborative events and communication records generated during task execution are mapped to a knowledge graph to form a dynamic collaborative subgraph. This aims to transform unstructured collaborative behavior data into a structured graph representation and associate it with the enterprise's context. Natural Language Processing (NLP) technology can be used to parse communication records (emails, chat logs, meeting minutes), extracting entity information such as participants, tasks, time, and key topics, and identifying relationships between entities (A requests assistance from B, C completes task D). Simultaneously, log data of collaborative events (document modifications, code commits, task status updates) can also be structured. These extracted entities and relationships are then mapped to the knowledge graph. For example, employee entities are linked to their corresponding skill nodes in the knowledge graph, and task events are linked to their respective business process nodes or target nodes. Through this mapping, a subgraph representing the interaction patterns between personnel, tasks, and resources during a specific task execution can be dynamically constructed.

[0069] This study calculates graph structure similarity and path efficiency metrics between dynamic collaborative subgraphs and collaborative subgraphs formed by high-performing teams under similar task backgrounds in a historical success case library. The aim is to quantify the gap between current team collaboration patterns and historical best practices. For graph structure similarity, various graph similarity algorithms can be used, such as methods based on graph edit distance, which measure similarity by calculating the minimum number of editing operations (such as adding, deleting, or modifying nodes or edges) required to transform one graph into another. For path efficiency metrics, the average path length for transmitting key information among team members and the number of hops in information transmission can be evaluated. For example, the average length of the shortest path from the task initiator to the task completer can be calculated, or the average number of communications among team members when completing a subtask can be assessed.

[0070] When the graph structure similarity is less than a preset similarity threshold or the path efficiency index is less than a preset efficiency threshold, the subgraph structure unit with the greatest difference or the efficiency bottleneck node is located and marked as an inefficient link in collaboration, aiming to accurately identify the specific reasons for the inefficiency. When the overall similarity or efficiency is found to be substandard, further analysis is conducted on the local differences between the dynamic collaborative subgraph and the efficient historical subgraph. For the subgraph structure unit with the greatest difference, it can be identified by comparing the substructures of the two graphs (such as community structure, interaction sequences of specific patterns). For example, it can be found that the communication pattern within a certain subteam is significantly different from that of an efficient team. For efficiency bottleneck nodes, individuals or specific interaction points that are overloaded in the information flow or task flow, causing delays or blockages, can be identified. For example, an employee may exhibit high betweenness centrality on multiple critical paths but have slow processing speed, or a communication channel may become a bottleneck for information transmission. This can be achieved by performing flow analysis, critical path analysis, or an evaluation based on a combination of node centrality and task completion time on the graph. Another approach is to use anomaly detection algorithms in machine learning to extract features (such as local centrality, interaction frequency, and response time) from each node or edge in the dynamic collaboration subgraph, and then compare them with the baseline of a high-performing team to identify the node or edge that deviates the furthest from the baseline and mark it as an inefficient link in the collaboration.

[0071] As can be seen from the above, this application, by adopting the aforementioned technical solution, maps collaborative events and communication records into dynamic collaborative subgraphs, intuitively presenting the team's collaborative status and achieving a visual representation of the collaborative process. It calculates the similarity and path efficiency indicators between the subgraphs and historically efficient collaborative subgraphs, establishing a benchmark for comparing efficient and current collaborations, and accurately locating differences and bottlenecks. The various technical features work together to solve the problems of traditional collaboration analysis's difficulty in quantification and ambiguous location of inefficient links, achieving accurate identification of inefficient collaborative links, clarifying the specific direction for collaboration optimization, providing precise basis for subsequent adjustments to personnel allocation and optimization of collaboration models, and improving the efficiency and coordination of team collaboration.

[0072] In one embodiment of this application, the parameters of the time-series capability assessment model are updated based on a comprehensive assessment report, contribution level, and identified anomaly data, including: The global performance label of the task execution result, along with the process behavior data and contribution of individual members, will be used as observational evidence. An online Bayesian learning algorithm is used to update the posterior probability distribution of the ability parameters of the corresponding members in each ability dimension in the time-series ability assessment model, taking observational evidence as input. In this update, a time decay factor is introduced, which makes the influence of recent task evidence on model parameters greater than that of long-term evidence. The time decay factor is determined by the interval between the task completion time and the current time, as well as the task priority coefficient.

[0073] In this embodiment, global performance tags can include, but are not limited to, macro-level indicators such as the final delivery quality score of the task, the on-time completion rate of the project, and the customer satisfaction index, representing the overall performance of the task. Individual member process behavior data covers the micro-level activities of employees during task execution, such as system operation logs, code commit records, document modification history, the frequency and content of communication on internal communication platforms, and meeting participation. This data can meticulously depict the specific behavioral patterns of employees in the task. Contribution can be quantified based on the completion status of sub-tasks after task decomposition, the peer evaluation results among team members, or through the analysis of the correlation between individual output and task objectives by an AI agent. By integrating these data of different granularities and from different sources, it ensures that the information used for model updates is rich and representative, thereby avoiding evaluation bias that may be caused by a single data source.

[0074] Building upon this foundation, an online Bayesian learning algorithm is employed, using the aforementioned observational evidence as input, to update the posterior probability distributions of the ability parameters for each member across all ability dimensions in the time-series ability assessment model. Through this method, an employee's ability in a specific ability dimension is no longer treated as a fixed value, but rather exists as a posterior probability distribution. This distribution characterizes the confidence level and uncertainty of the employee's ability value given the latest observational evidence. This probabilistic framework-based update mechanism allows the model to more flexibly integrate new evidence and effectively handle uncertainties in the data, thereby improving the robustness and adaptability of the ability assessment.

[0075] Furthermore, a time decay factor is introduced during updates, giving greater weight to recent task evidence in influencing model parameters than longer-term evidence. The time decay factor is a mechanism used to adjust the influence of data, designed to characterize the importance of information timeliness. For example, an exponential decay function (exp(-λ)) can be used. Δt), where Δt is the time interval and λ is the decay rate. By multiplying this decay factor by the likelihood function of the observed evidence, the information gain, or by directly adjusting the prior weights in the Bayesian update formula, it can be ensured that the model prioritizes the recent performance and ability changes of employees during updates, thereby more accurately representing their current actual ability level and avoiding inappropriate interference from outdated data on the evaluation results.

[0076] Simultaneously, the time decay factor is determined based on the interval between the task completion time and the current time, as well as the task's priority coefficient. The interval between the task completion time and the current time can be directly calculated as days, weeks, or months, representing the physical timeliness of the evidence. The task's priority coefficient can be a preset integer level (e.g., level 1 to 5) with a percentage weight. These two factors can be determined using a multiplicative combination (e.g., decay factor = f(Δt)). g(priority)).

[0077] As can be seen from the above, this application, by adopting the aforementioned technical solution, uses global performance labels, individual behavioral data, and contribution as observational evidence to ensure that model parameter updates have reliable data support; the online Bayesian learning algorithm enables real-time parameter updates, improving update efficiency; the time decay factor highlights the influence of recent task evidence, making the parameters more closely match the current ability status of employees; and the combination of priority coefficients ensures that the update direction meets the requirements. The synergistic effect of these technical features solves the problems of lagging parameter updates and neglecting time factors and priorities in time-series ability assessment models, continuously improving model assessment accuracy and ensuring the timeliness and accuracy of employee ability vectors.

[0078] In one embodiment of this application, an AI Agent personnel optimization configuration method based on multi-dimensional performance evaluation further includes: Based on real-time environmental data and internal collaboration data collected during the task execution process, combined with actual loss or deviation data after the task is completed, a profile of the actual risk of the task is generated. Calculate the differences between the actual risk profile of the task and the predicted risk profile of the task generated by the risk assessment model before the task is executed in each risk dimension, and construct a risk profile difference matrix; Based on the risk profile difference matrix, the risk dimensions with prediction deviations greater than the preset deviation threshold are located, and the key features on which the predicted value of the dimension was generated are traced. The key features include specific keywords in the task requirement description, abnormal indicators in the execution environment information, and correlation patterns in historical project data. Using the actual risk profile of the task and its corresponding process data as new labeled samples, the risk assessment model is incrementally trained. The model weight parameters corresponding to the risk dimension of prediction deviation exceeding the limit are optimized. Based on the tracing results of key features, the sensitivity of the model to relevant signals in the input features and the interpretation path are adjusted.

[0079] In this embodiment, the actual risk profile of the task is an objective and comprehensive quantitative description of the risks faced by the task based on real-world data after the task execution is completed. It characterizes the type and severity of risks encountered during the actual execution process and their impact on the task objectives. Its purpose is to provide a realistic risk benchmark for comparison with the predicted risk profile, thereby assessing the accuracy of the predictive model. For example, real-time environmental data may include system load, network latency, external market fluctuations, and policy and regulatory changes during task execution; internal collaboration data may include communication frequency among team members, usage of collaboration tools, task workflow time, and resource conflict records. This data is continuously collected and analyzed by an AI Agent, combined with a predefined risk indicator system, and mapped to various risk dimensions through a machine learning model, quantifying their risk levels. Furthermore, actual loss or deviation data may include the number of days the task was delayed, the amount of cost exceeding the budget, product defect rate, number of customer complaints, and number of safety incidents. This data is typically obtained after the task is completed through a project management system, financial system, or customer feedback system. By correlating these loss or deviation data with preset risk thresholds and impact factors, the actual exposure and impact of the task on different risk dimensions can be calculated, thereby generating an actual risk profile of the task.

[0080] Based on this, the differences between the actual task risk profile and the predicted task risk profile generated by the risk assessment model before task execution are calculated across each risk dimension, and a risk profile difference matrix is ​​constructed. The risk profile difference matrix is ​​a quantitative tool used to clearly display the specific differences between the actual and predicted task risk profiles across each risk dimension. By calculating the difference values, it is possible to intuitively identify in which aspects the risk assessment model is accurate in its predictions and in which aspects it is biased, thus providing data support for subsequent model optimization. For example, the difference values ​​can be calculated using various mathematical methods. For continuous risk dimensions, the absolute or relative difference between the two can be directly calculated; for discrete risk dimensions, the consistency of their classifications can be calculated. The construction of the difference matrix can present these difference values ​​in tabular or vector form, with each row or column representing a risk dimension, and the elements in the matrix being the difference values ​​for that dimension.

[0081] Furthermore, based on the risk profile difference matrix, the risk dimensions with prediction deviations exceeding a preset deviation threshold are identified, and the key features used to generate the predicted value for that dimension are traced back. This step aims to focus on those significantly deviating risk dimensions from the overall prediction deviation and delve into the root causes of these deviations. By identifying key features, a clear direction and basis can be provided for targeted optimization of the risk assessment model, avoiding blind adjustments. For example, the preset deviation threshold can be set according to business needs and risk tolerance. If the difference value on a certain risk dimension exceeds a certain percentage or absolute value, then that dimension is considered to have a significant deviation. The identification process can be achieved by traversing the risk profile difference matrix and filtering out all risk dimensions that meet the condition of a difference value exceeding the preset deviation threshold. Tracing key features can be achieved through model interpretability techniques, such as using the SHAP (SHapley Additive exPlanations) value method to analyze which input features (such as specific keywords in the task requirement description, abnormal indicators in the execution environment information, and correlation patterns in historical project data) contribute the most to the model's output when generating the predicted value for that risk dimension.

[0082] Finally, using the actual risk profile of the task and its corresponding process data as new labeled samples, the risk assessment model is incrementally trained. The focus is on optimizing the model weight parameters corresponding to the dimension of prediction deviation exceeding the limit risk. Based on the traceability results of key features, the model's sensitivity to relevant signals in the input features and its interpretation path are adjusted. This is the core step of model self-learning and optimization. By using real risk data as new training samples and selectively adjusting model parameters, the prediction accuracy and robustness of the risk assessment model can be continuously improved. For example, incremental training refers to small-batch or online training using new labeled samples based on the existing model. For optimizing the model weight parameters corresponding to the dimension of prediction deviation exceeding the limit risk, the loss function can be adjusted to assign higher weights to these dimensions with larger deviations when calculating the loss, or specific sub-networks for these dimensions can be introduced into the model structure for fine-tuning. Based on the traceability results of key features, the model's sensitivity to relevant signals in the input features can be adjusted. For example, if the model's risk identification ability for a certain keyword is found to be insufficient, feature engineering can be used to enhance the representation of that keyword, or an attention mechanism for that keyword can be added to the model. Adjusting the interpretation path may involve modifying the model's feature interaction methods to enable it to more accurately capture specific patterns during prediction.

[0083] As can be seen from the above, this application, by adopting the aforementioned technical solution, generates an actual risk profile of the task, compares it with the predicted profile to construct a difference matrix, and accurately locates the risk dimensions where the prediction deviation exceeds the limit; traces key features to clarify the root causes of deviations, providing targeted directions for model optimization; incremental training introduces new labeled samples, focusing on optimizing the model weights corresponding to the deviation dimensions, and adjusting feature sensitivity and interpretation paths. Each technical feature progresses layer by layer, forming a closed loop of prediction-verification-tracing-optimization, solving the problems of large prediction deviations and insufficient generalization ability in risk assessment models, continuously improving the accuracy of risk assessment, and providing more reliable support for risk response in personnel allocation.

[0084] In one embodiment of this application, a multi-objective optimization algorithm is used to solve the problem, specifically including: Construct a multi-objective optimization function, which includes at least the following: The first objective function is used to maximize the overall weighted match between the configuration team and the target positions; The second objective function is used to maximize the team's aggregated risk response capability, which is calculated by aggregating the capability values ​​of team members in the dimension of task risk profile requirements. The third objective function is used to minimize the estimated collaboration cost of the configuration team, which is estimated based on the collaboration distance and communication frequency between members in historical team collaboration network data. The constraints include: individual matching degree is greater than or equal to the matching degree threshold, the range of team members, and the real-time availability status of individuals; A multi-objective evolutionary optimization algorithm is used to solve the multi-objective optimization function. From the generated Pareto optimal solution set, the final personnel allocation scheme is selected according to the preset decision preferences.

[0085] In this embodiment, the first objective function aims to maximize the comprehensive weighted matching degree between the configuration team and the target position. This function ensures that the selected personnel's comprehensive qualities in terms of skills, experience, etc., meet the task requirements by quantifying the degree of fit between the team's overall capabilities and the requirements of the target position. For example, it can be represented by the weighted average or geometric mean of the individual matching scores of team members, and the weights can be dynamically adjusted according to the importance of each capability dimension in the task risk profile. The second objective function is used to maximize the configuration team's aggregated risk response capability. This capability is obtained by aggregating the capability values ​​of team members in the dimensions required by the task risk profile, aiming to ensure that the team can effectively cope with various risks that may occur during task execution. For example, the minimum capability value of team members in the high-risk dimension of the risk profile can be calculated to characterize the team's weakest link effect. The third objective function is used to minimize the estimated collaboration cost of the configuration team. This cost is estimated based on the collaboration distance and communication frequency between members in historical team collaboration network data, aiming to reduce the difficulty and resource consumption of internal team communication and coordination. For example, the collaboration cost can be estimated by analyzing the collaboration graph between team members in historical project data and calculating the average length of the shortest path or the density of communication links between selected team members.

[0086] During the solution process, constraints are crucial to ensuring the feasibility and effectiveness of the personnel allocation plan. An individual matching degree of no less than a matching threshold means that each selected employee must meet the minimum standard of matching the target position, avoiding the selection of individuals with insufficient capabilities. For example, a comprehensive matching degree score can be set, requiring all selected employees to have a score above 80. The range of team member numbers limits the team size; for example, a task team might be limited to 3 to 5 people to ensure a moderate team size that can complete the task without becoming overly bloated. Individual real-time availability status considers the employee's current actual workload and schedule, ensuring that selected employees are available during task execution. For example, querying an employee's schedule or project allocation can exclude employees currently occupied by other critical tasks. To effectively solve the above multi-objective optimization function, this application employs a non-dominated sorting genetic algorithm (NSGA-II), which can simultaneously optimize multiple conflicting objectives and generate a Pareto optimal solution set. Each solution in the Pareto optimal solution set is a non-dominated solution, meaning that no objective can be improved without sacrificing at least one other objective. From the generated Pareto optimal solution set, the final personnel allocation plan is selected according to the preset decision preferences.

[0087] Specifically, the multi-objective evolutionary optimization algorithm employs an improved non-dominated sorting genetic algorithm (NSGA-II). The algorithm parameters are set as follows: population size of 100, number of generations of evolution of 200, crossover probability of 0.9 (using simulated binary crossover SBX), and mutation probability of 0.1 (using polynomial mutation). The decision variables are encoded in binary, with each bit representing whether a candidate employee is selected. The fitness function is the aforementioned multi-objective optimization function. When selecting the final solution from the Pareto optimal solution set, a decision-making method based on the ideal point method (TOPSIS) is used. The system administrator sets the preference weights for each objective (matching degree, risk response capability, and collaboration cost) according to the current task priority, calculates the proximity of each solution to the ideal solution, and selects the closest solution as the final personnel allocation plan.

[0088] As can be seen from the above, this application, by adopting the aforementioned technical solutions, constructs a multi-objective optimization function that balances matching degree, risk response capability, and collaboration cost, achieving a balance of multi-dimensional objectives and avoiding configuration imbalance caused by a single objective. It clarifies constraints to ensure that the configuration scheme meets actual usability requirements and avoids infeasible configurations. A multi-objective evolutionary optimization algorithm solves the Pareto optimal solution set, combining decision preferences to select a scheme, balancing scientific rigor and flexibility. The synergistic effect of these technical features addresses the problems of traditional optimization algorithms, such as single objectives, insufficient constraints, and poor scheme flexibility, improving the optimality of personnel allocation schemes, reducing collaboration costs while ensuring efficiency, and enhancing the practicality of the scheme.

[0089] In one embodiment of this application, an AI Agent personnel optimization configuration method based on multi-dimensional performance evaluation further includes: Based on the model confidence when the time-series capability assessment model outputs employee capability vectors, it is determined whether to trigger the adjustment of the weight parameters of the job capability indicator system. In response to the model confidence level being less than the preset confidence threshold, behavioral and outcome data of high-performing employees in key tasks are collected recently. Cluster analysis and correlation tests are used to identify emerging capability characteristics that contribute to task success. Based on the identified emerging competency characteristics and their contribution, the weight parameters of each dimension in the job competency indicator system are redistributed through an optimization algorithm to generate an updated indicator system. Based on the updated indicator system, the time-series capability assessment model is subjected to transfer learning and fine-tuning.

[0090] In this embodiment, model confidence refers to a quantitative measure of the reliability or accuracy of the prediction results of a time-series competency assessment model when generating employee competency vectors. This confidence level characterizes the degree to which the model matches the current input data and internal parameters, as well as its level of confidence in the employee competency assessment results. By judging this confidence level, the system can intelligently identify potential deficiencies in the current assessment system, thereby deciding whether to update and adjust the underlying job competency indicator system. For example, model confidence can be obtained by calculating the entropy value of the model's output probability distribution. The lower the entropy value, the more certain the model is about the prediction results, and the higher the confidence level; conversely, the higher the entropy value, the lower the confidence level. When the entropy value exceeds a preset threshold, adjustment is triggered.

[0091] In response to a model confidence level below a preset threshold, this application collects behavioral and outcome data of recent high-performing employees in key tasks. Cluster analysis and correlation tests are then used to identify emerging competency characteristics contributing to task success. When the model confidence level is below the preset threshold, it indicates that the existing competency assessment system may not have adequately captured the key competencies demonstrated by employees in actual tasks. In this case, behavioral and outcome data of employees who have performed exceptionally well in recent key tasks are proactively collected. This data serves as a crucial basis for identifying emerging competency characteristics. In-depth analysis of this data can uncover new skills, knowledge, or behavioral patterns that are not fully reflected in traditional indicator systems but significantly contribute to task success. For example, behavioral data may include employee operation records in project management tools, code submission frequency, interaction patterns on communication and collaboration platforms, and document output quality; outcome data may include task completion rate, project delivery quality, and customer satisfaction scores. Cluster analysis can employ the K-means algorithm to group the behavioral patterns of high-performing employees and identify common characteristics; correlation tests can use the Pearson correlation coefficient to assess the statistical strength of the association between these characteristics and task success. In addition, behavioral data can also include employee retrieval and contribution records in the internal knowledge base, the types of training courses participated in and their completion status, as well as focus duration and multitasking efficiency recorded by sensors or software. Cluster analysis can use hierarchical clustering models to identify different types of high-performance behavioral patterns; correlation testing can use chi-square tests to determine the impact of specific behavioral characteristics on the probability of task success.

[0092] Based on the identified emerging competency characteristics and their contribution, this application uses an optimization algorithm to redistribute the weight parameters of each dimension in the job competency indicator system, generating an updated indicator system. After identifying emerging competency characteristics and their contribution to task success, these newly discovered competencies need to be integrated into the existing job competency indicator system, and the weights of each competency dimension need to be adjusted to ensure that the indicator system can more accurately represent current business needs and market changes. Through the optimization algorithm, weights can be scientifically redistributed, giving higher weights to competency dimensions that contribute more to task success, thereby improving the effectiveness and foresight of the entire evaluation system. For example, the optimization algorithm can use a genetic algorithm, with the objective function of maximizing the predictive accuracy of the new indicator system in historical high-performance tasks, to iteratively adjust the weights of each competency dimension. Emerging competency characteristics can be added as new dimensions, and their initial weights can be set according to their contribution.

[0093] Finally, based on the updated indicator system, this application performs transfer learning and fine-tuning on the time-series competency assessment model. After the job competency indicator system is updated, the original time-series competency assessment model may not be fully adapted to the new assessment standards. In order for the model to accurately assess employee competency based on the updated indicator system, it needs to be adaptively adjusted. Transfer learning and fine-tuning is an efficient model update strategy that allows the model to quickly learn and adapt to new data distributions and assessment standards while retaining its original knowledge, avoiding the huge overhead of training the model from scratch. For example, transfer learning can use feature extraction to freeze the bottom feature extraction layer of the pre-trained model (the model trained based on the old indicator system) and only train the top-level classifier or regressor to adapt it to the output of the new indicator system. Fine-tuning can perform small-scale parameter updates on the entire model and train it using a small amount of new data (e.g., employee competency data relabeled based on the updated indicator system).

[0094] As can be seen from the above, this application, by adopting the aforementioned technical solutions, determines the need for adjusting the indicator system based on model confidence, ensuring the necessity and relevance of the adjustments; collects data from high-performing employees, and accurately identifies emerging capability characteristics through clustering and correlation tests to align with business development needs; optimizes the algorithm to reallocate weights, generating an updated indicator system; and uses transfer learning to fine-tune the model, achieving synergistic optimization of the indicator system and the model. The synergistic effect of these technical features addresses the problem of outdated job capability indicator systems that cannot adapt to emerging business needs, making the indicator system more forward-looking and improving the adaptability and accuracy of the time-series capability assessment model.

[0095] In one embodiment of this application, an AI Agent personnel optimization configuration method based on multi-dimensional performance evaluation further includes: Based on the identification of inefficient collaborative links, reverse positioning leads to inefficient members' skill deficiencies or team role configuration defects. The inefficient links in collaboration, the corresponding member capability vectors, and the task risk profiles are used as training samples and input into the collaboration effectiveness prediction model to train the model parameters. By using the trained collaboration efficiency prediction model, when generating personnel configuration schemes, the potential collaboration risks and efficiency losses under different configuration combinations are predicted. The prediction results are used as constraints or penalties and fed back into the multi-objective optimization algorithm to dynamically adjust the weight parameters of the team collaboration effectiveness objective in the algorithm, so that the generated personnel allocation plan is more inclined to avoid the inefficient collaboration patterns that have appeared in the past.

[0096] In this embodiment, after identifying inefficient collaboration links, this application further reverse-engineers the shortcomings in member capabilities or deficiencies in team role allocation that lead to inefficiency. This step aims to delve into the root causes of inefficient collaboration, rather than merely remaining at the surface level. For example, by conducting in-depth analysis of member behavior data, communication records, and task allocation records involved in inefficient collaboration links, combined with a pre-defined member capability model and role and responsibility definition, and using causal reasoning or expert system rules, it can infer the member capability deficiencies (e.g., communication skills, technical proficiency, problem-solving abilities) or unreasonable team role allocation (e.g., missing key roles, overlapping roles, role conflicts) directly related to the inefficient links.

[0097] Subsequently, inefficient collaboration links, corresponding member capability vectors, and task risk profiles are used as training samples and input into the collaboration effectiveness prediction model to train the model parameters. This step aims to build a model capable of learning and predicting collaboration effectiveness. For example, the descriptions of inefficient collaboration links (e.g., collaboration type, occurrence time, involved personnel), member capability vectors (representing the various capability values ​​of participating members), and task risk profiles (describing the risk characteristics of the task) can be feature-engineered to form unified training samples. These samples can be input into a machine learning model, and through supervised learning, the model can be trained to identify feature combinations that lead to inefficiency. The collaboration effectiveness prediction model adopts the Gradient Boosting Decision Tree (GBDT) model, specifically implemented using the XGBoost library. Feature engineering of the training samples includes: 1) Collaboration structure features: graph metrics extracted from dynamic collaboration subgraphs, such as team density, maximum connected subgraph size, and critical path length; 2) Member capability features: the values ​​of each dimension of the employee capability vectors of participating members and their variance; 3) Task risk features: risk feature vectors mapped from the task risk profile. The output labels are either binary classification labels (efficient / inefficient) or continuous value labels (such as percentage of efficiency loss). During model training, the maximum tree depth is set to 6, the learning rate is 0.1, 500 decision trees are ensembled, and the optimal number of iterations is determined using early stopping. Feature importance analysis is used to verify the consistency between the model's decisions and business understanding.

[0098] Building upon this foundation, the trained collaboration effectiveness prediction model is used to predict potential collaboration risks and efficiency losses under different configuration combinations when generating personnel allocation schemes. This step applies predictive capabilities to actual personnel allocation decisions. For example, when a multi-objective optimization algorithm generates candidate personnel allocation schemes, for each candidate scheme, the ability vectors of the team members, task risk profiles, and simulated collaboration features are fed into the trained collaboration effectiveness prediction model as input. The model outputs the potential collaboration risk level (e.g., high, medium, low) or quantified efficiency loss values ​​(e.g., expected delay time, increased communication costs) under that configuration scheme.

[0099] Finally, the prediction results are fed back into the multi-objective optimization algorithm as constraints or penalties, dynamically adjusting the weight parameters related to team collaboration effectiveness. This ensures that the generated personnel allocation schemes are more likely to avoid historically observed inefficient collaboration patterns. This step guarantees that the prediction results directly influence the final personnel allocation decision. For example, when the collaboration effectiveness prediction model outputs that a certain configuration combination has a high potential collaboration risk or significant efficiency loss, these predictions can be used as hard constraints on the multi-objective optimization algorithm, disallowing the selection of schemes with risks exceeding a certain threshold; or they can be added as penalties to the optimization objective function, reducing the fitness of high-risk schemes and thus prioritizing their exclusion from the Pareto optimal solution set.

[0100] As can be seen from the above, this application, by adopting the aforementioned technical solution, reversely locates the root causes of inefficient collaboration, clarifies capability shortcomings and role deficiencies, and provides accurate samples for model training. After training, the collaboration efficiency prediction model can predict the collaboration risks and efficiency losses of configuration combinations in advance, achieving risk avoidance in advance. The prediction results are fed back to the optimization algorithm as constraints or penalties, dynamically adjusting the target weights and optimizing the configuration direction. The various technical features form a closed loop of identification-tracing-training-prediction-optimization, solving the problem of traditional configuration ignoring potential collaboration risks, improving the collaboration stability of personnel configuration schemes, and reducing the probability of inefficient collaboration.

[0101] Corresponding to the AI ​​Agent personnel optimization configuration method based on multi-dimensional performance evaluation in the above embodiment, Figure 2 This is a structural block diagram of an AI Agent personnel optimization configuration system based on multi-dimensional performance evaluation, provided as an embodiment of this application. For ease of explanation, only the parts relevant to the embodiment of this application are shown. References Figure 2 The AIAgent personnel optimization configuration system 20 based on multi-dimensional performance evaluation includes: a graph construction module 21, a first vector module 22, a second vector module 23, a job positioning module 24, a risk determination module 25, a threshold determination module 26, a matching calculation module 27, and a solution generation module 28.

[0102] Among them, the graph construction module 21 is used to parse the target text of the enterprise through AI Agent and construct a knowledge graph; The first vector module 22 is used to quantify the value of positions in the knowledge graph based on the job competency index system and generate job value vectors. The second vector module 23 is used to generate employee capability vectors that are consistent with the dimensions of the job capability indicator system based on the time-series multi-source behavioral data of employees in the enterprise through a time-series capability assessment model. The job positioning module 24 is used to locate the target job associated with the task in the knowledge graph in response to the received task instructions. The risk determination module 25 is used to generate a task risk profile based on the requirement description and execution environment information in the task instruction, through a risk assessment model; the task risk profile maps to the required level of the target position's competency dimension. The threshold determination module 26 is used to determine the matching threshold of personnel configuration and risk response capability requirements based on the level and combination characteristics of the capability dimensions required in the task risk profile, through predefined mapping rules. The matching calculation module 27 is used to calculate the multi-dimensional matching degree between the employee's ability vector and the job value vector corresponding to the target job using a weighted similarity algorithm. The scheme generation module 28 is used to generate personnel configuration schemes based on target positions, multi-dimensional matching degree, matching degree threshold, risk response capability requirements and historical team collaboration data, using a multi-objective optimization algorithm with the goal of maximizing team collaboration efficiency.

[0103] See Figure 3 , Figure 3 This is a schematic block diagram of an electronic device provided according to an embodiment of this application. Figure 3 The electronic device 300 in this embodiment may include one or more processors 301, one or more input devices 302, one or more output devices 303, and one or more memories 304. The processors 301, input devices 302, output devices 303, and memories 304 communicate with each other via a communication bus 305. The memories 304 store computer programs, including program instructions. The processors 301 execute the program instructions stored in the memories 304. Specifically, the processors 301 are configured to invoke the program instructions to perform the functions of the modules in the aforementioned device embodiments, for example... Figure 2 The diagram shows the functions of the following modules: map construction module 21, first vector module 22, second vector module 23, job positioning module 24, risk determination module 25, threshold determination module 26, matching calculation module 27, and scheme generation module 28.

[0104] It should be understood that, in the embodiments of this application, the processor 301 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0105] Input device 302 may include a touchpad, a fingerprint sensor (for collecting the user's fingerprint information and fingerprint orientation information), a microphone, etc., and output device 303 may include a display (LCD, etc.), a speaker, etc.

[0106] The memory 304 may include read-only memory and random access memory, and provides instructions and data to the processor 301. A portion of the memory 304 may also include non-volatile random access memory. For example, the memory 304 may also store device type information.

[0107] In specific implementations, the processor 301, input device 302, and output device 303 described in the embodiments of this application can execute the implementation methods described in any embodiment of the AI ​​Agent personnel optimization configuration method based on multi-dimensional performance evaluation provided in the embodiments of this application, or they can execute the implementation methods of the electronic devices described in the embodiments of this application, which will not be repeated here.

[0108] In another embodiment of this application, a computer-readable storage medium is provided. This computer-readable storage medium stores a computer program, which includes program instructions. When executed by a processor, the program instructions implement all or part of the processes in the methods described above. Alternatively, the computer program can instruct related hardware to complete the process. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include any entity or device capable of carrying computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0109] The computer-readable storage medium can be an internal storage unit of the electronic device in any of the foregoing embodiments, such as a hard disk or memory of the electronic device. The computer-readable storage medium can also be an external storage device of the electronic device, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the electronic device. Furthermore, the computer-readable storage medium can include both internal and external storage units of the electronic device. The computer-readable storage medium is used to store computer programs and other programs and data required by the electronic device. The computer-readable storage medium can also be used to temporarily store data that has been output or will be output.

[0110] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.

[0111] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the electronic devices and units described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0112] In the several embodiments provided in this application, it should be understood that the disclosed electronic devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces or units, or it may be an electrical, mechanical, or other form of connection.

[0113] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments of this application, depending on actual needs.

[0114] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0115] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for optimizing the allocation of AI Agent personnel based on multi-dimensional performance evaluation, characterized in that, include: AI Agents analyze target text from enterprises to construct knowledge graphs; Based on the job competency index system, the value of the jobs in the knowledge graph is quantified to generate a job value vector. Based on the time-series multi-source behavioral data of employees in the enterprise, an employee capability vector consistent with the dimensions of the job capability index system is generated through a time-series capability assessment model. In response to the received task instructions, locate the target position associated with the task in the knowledge graph; Based on the requirements description and execution environment information in the task instruction, a task risk profile is generated through a risk assessment model; the task risk profile maps to the required level of the target position's competency dimensions. Based on the level and combination characteristics of the capability dimensions required in the task risk profile, the matching threshold of personnel configuration and the risk response capability requirements are determined through predefined mapping rules. A weighted similarity algorithm is used to calculate the multi-dimensional matching degree between the employee's ability vector and the job value vector corresponding to the target job. Based on the target job position, multi-dimensional matching degree, matching degree threshold, risk response capability requirements, and historical team collaboration data, a multi-objective optimization algorithm is adopted to generate a personnel configuration plan with the goal of maximizing team collaboration efficiency.

2. The method for optimizing the allocation of AI Agent personnel based on multi-dimensional performance evaluation according to claim 1, characterized in that, The method based on the job competency index system quantifies the value of jobs in the knowledge graph and generates job value vectors, including: From the knowledge graph, target nodes, business process nodes, and skill nodes associated with the job position are extracted to obtain the set of capability dimensions for the job position. The Analytic Hierarchy Process (AHP) is used to compare the relative importance of each capability in the capability dimension set and determine the initial weight vector for each capability dimension. Based on the correlation analysis between historical project success data and resource consumption in each capability dimension, the initial weight vector is calibrated using a regression model to obtain the final weight vector; Based on the knowledge graph, the graph centrality index of the target node and business process node associated with the position is calculated, and the contribution coefficient of the corresponding ability dimension of the position to the success of the project is combined with the contribution coefficient of the ability dimension of the position to the success of the project in the historical project data. The factor point method is used to generate the initial score vector of the position ability. The initial job competency score vector and the final weight vector are weighted and calculated to generate the job value vector.

3. The method for optimizing the allocation of AI Agent personnel based on multi-dimensional performance evaluation according to claim 1, characterized in that, The step of employing a weighted similarity algorithm to calculate the multi-dimensional matching degree between the employee's ability vector and the job value vector corresponding to the target job includes: Based on the required skill levels of each job position mapped in the task risk profile, the weights of the corresponding skill dimensions in the weighted similarity algorithm are adjusted to obtain a dynamic weight vector. The weighted cosine similarity calculation is used to calculate the similarity between the employee's ability vector and the target position's job value vector by weighting the dynamic weight vector, and this similarity is used as the comprehensive matching degree. The overall matching degree and the individual matching degrees of each capability dimension together constitute the multi-dimensional matching degree.

4. The method for optimizing the allocation of AI Agent personnel based on multi-dimensional performance evaluation according to claim 3, characterized in that, Also includes: By analyzing employees’ historical task decision-making data and behavioral feedback, an employee risk preference vector is constructed. The risk preference vector represents the employee’s decision-making tendency and tolerance threshold under different types of risks. The risk assessment model maps the task risk profile into a risk feature vector. Calculate the similarity between the risk preference vector and the task risk feature vector, and use it as the risk fit. The overall matching degree, the individual matching degree of each capability dimension, and the risk suitability are weighted and fused together using preset fusion rules to form the final multi-dimensional matching degree; wherein, the weights in the fusion rules are adjusted according to the overall risk level of the task risk profile.

5. The method for optimizing the allocation of AI Agent personnel based on multi-dimensional performance evaluation according to claim 1, characterized in that, Also includes: After scheduling target personnel to perform tasks according to the personnel configuration plan, process data and result data during task execution are collected through AI Agent; Based on preset process rules and performance baselines, correlation analysis is performed on the process data to identify individual behavioral anomalies and inefficient links in team collaboration, and verified collaborative behavior data is recorded. The results data, process efficiency indicators, and collaborative behavior data are input into the task decision evaluation model to generate a comprehensive evaluation report and quantify the contribution of each member to the task objective. Based on the comprehensive evaluation report, the contribution level, and the identified abnormal data, the parameters of the time-series capability evaluation model are updated.

6. The method for optimizing the allocation of AI Agent personnel based on multi-dimensional performance evaluation according to claim 5, characterized in that, Based on preset process rules and performance baselines, the process data is correlated to identify individual behavioral anomalies and inefficient team collaboration, and verified collaborative behavior data is recorded, including: The collaborative events and communication records generated during task execution are mapped to the knowledge graph to form a dynamic collaborative subgraph; Calculate the graph structure similarity and path efficiency index between the dynamic collaborative subgraph and the collaborative subgraph formed by high-performing teams under similar task backgrounds in the historical success case library; When the graph structure similarity is less than a preset similarity threshold or the path efficiency index is less than a preset efficiency threshold, the subgraph structure unit or efficiency bottleneck node with the greatest difference is located and marked as an inefficient link in the collaboration.

7. The method for optimizing the allocation of AI Agent personnel based on multi-dimensional performance evaluation according to claim 5, characterized in that, The step of updating the parameters of the time-series capability assessment model based on the comprehensive assessment report, the contribution level, and the identified abnormal data includes: The global performance label of the task execution result, along with the process behavior data and contribution of individual members, will be used as observational evidence. An online Bayesian learning algorithm is used to update the posterior probability distribution of the ability parameters of the corresponding members in each ability dimension in the time-series ability assessment model, taking the observation evidence as input. In this update, a time decay factor is introduced, which makes the influence weight of recent task evidence on model parameters greater than that of long-term evidence. The time decay factor is determined based on the interval between the task completion time and the current time, as well as the task priority coefficient.

8. An AI Agent personnel optimization allocation system based on multi-dimensional performance evaluation, characterized in that, include: The knowledge graph construction module is used to parse the target text of an enterprise through an AI Agent and build a knowledge graph. The first vector module is used to quantify the value of the positions in the knowledge graph based on the job competency index system and generate a job value vector. The second vector module is used to generate employee capability vectors that are consistent with the dimensions of the job capability index system based on the time-series multi-source behavioral data of employees in the enterprise and through the time-series capability assessment model. The job location module is used to locate the target job associated with the task in the knowledge graph in response to the received task instructions; The risk determination module is used to generate a task risk profile based on the requirement description and execution environment information in the task instruction, through a risk assessment model. The task risk profile mapping includes the required levels of the target position's competency dimensions; The threshold determination module is used to determine the matching degree threshold and risk response capability requirements of personnel configuration based on the level and combination characteristics of the capability dimensions required in the task risk profile, through predefined mapping rules. The matching calculation module is used to calculate the multi-dimensional matching degree between the employee's ability vector and the job value vector corresponding to the target job using a weighted similarity algorithm; The scheme generation module is used to generate a personnel configuration scheme based on the target position, multi-dimensional matching degree, the matching degree threshold, the risk response capability requirements and historical team collaboration data, using a multi-objective optimization algorithm with the goal of maximizing team collaboration efficiency.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 7.

Images