AI intelligent talent decision-making method and system based on employee behavior analysis and application

Abstract

The invention relates to the field of talent decision making, in particular to an AI intelligent talent decision making method and system based on employee behavior analysis and application, and the method comprises the steps: employing a graph neural network to carry out the structural expression of the multidimensional features of an employee based on the multisource behavior data of the employee, and generating an employee capability graph; dynamically adjusting the label weight of each feature in the employee ability atlas by adopting a reinforcement learning method based on the business feedback of the employees; performing decision evaluation on the employees based on the label weights; according to the method, an employee ability graph is constructed through GNN, node / edge representation can be carried out based on GNN learning, and the hidden ability migration and team synergistic effect can be captured; meanwhile, a reinforcement learning method is adopted to dynamically adjust the label weight of each feature in the employee ability map, so that the label weight and the portrait strategy can be optimized by taking the business index as a target, and blind overfitting of historical deviation is avoided.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of talent decision-making, and particularly to an AI intelligent talent decision-making method and system based on employee behavior analysis and application. BACKGROUND

[0002] With the development of technology, rational use of time and talent has become an important lever for enterprises to improve efficiency. The existing human resource management system can establish a static talent file around structured data such as personnel files, contracts, attendance, performance evaluation, etc., generate scores and portraits through rule tables or fixed calculation models, and rely on artificial batch tasks to update the portraits. In the aspect of performance management, KPI / KR / OKR indicators are often taken as the core, statistical reports and visualizations are provided, and they are mostly summarized by department / post, which is difficult to reflect the cross-project collaboration relationship and ability transfer of employees, and the unstructured data (comments, weekly reports, email minutes) is underutilized. In the establishment of talent pool, it is often based on keywords or tags to build, and the resume analysis and keyword matching of the post JD form tags, but the source of the tags is single, lacking of behavior timing, project context and growth trajectory modeling, leading to static and subjective matching results. A small number of systems integrate project management (such as task completion, code submission, etc.) and learning platform (such as course completion) data, but often lack a unified feature warehouse and a unified user identity mapping, making it difficult to support the "real-time, dynamic, cross-scenario" update of the portrait.

[0003] In summary, the existing human resource management system mainly has the following problems: 1) Data silos and inconsistent standards: collaboration systems, performance systems, learning platforms, access control / behavior logs lack unified primary keys and time standards, making it difficult to form a sustainable and iterative dynamic portrait; 2) Static portrait, update lag: portraits are mostly batch offline calculations, unable to adjust weights based on recent performance in projects, training, and collaboration, making it difficult to support "current" talent decision-making; 3) Subjective and non-evolving tag system: tag sources rely on manual or fixed rules, making it difficult to dynamically converge to more optimal tag weights based on business feedback (such as hiring or promotion results, project success rates); 4) Lack of non-result-oriented and time-series behavior utilization: performance comments, weekly reports, code reviews, meeting minutes lack NLP semantic analysis and time-series modeling, resulting in incomplete ability portrayal; 5) Lack of structured ability map: the relationship network of employee-skills-tasks-projects is not explicitly modeled, making it difficult to understand ability transfer, potential path, and team combination synergy; 6) Difficulty in cross-module integration with ERP (Enterprise Resource Planning): The profile lacks standardized interfaces, making it impossible to reuse and link it across multiple modules such as recruitment, job transfer, promotion, and training recommendations; 7) Privacy and compliance gaps: The lack of systematic de-identification, access control, auditing, and compliance modules makes it difficult to implement in production-level enterprise environments; 8) Insufficient explainability and controllability: The algorithm recommendation lacks traceable causes and visual explanations, which affects the adoption and trust of HR and business departments. Summary of the Invention

[0004] To address the above problems, this invention proposes an AI-powered intelligent talent decision-making method based on employee behavior analysis, comprising: A graph neural network is used to structurally represent the multidimensional features of employees based on multi-source behavioral data, generating an employee capability map; The label weights of each feature in the employee competency graph are dynamically adjusted based on employee business feedback using reinforcement learning methods. Employees are evaluated and decisions are made based on the aforementioned label weights.

[0005] Optionally, the step of using a graph neural network to structurally represent the multidimensional features of employees based on multi-source behavioral data to generate an employee competency map includes: Acquire multi-source behavioral data of employees; wherein, the multi-source behavioral data includes structured and / or unstructured data; Feature extraction, semantic parsing, and format standardization are performed on the multi-source behavioral data; A graph neural network is used to construct an employee capability map with multidimensional features.

[0006] Optionally, the acquisition of multi-source behavioral data of employees specifically includes: Obtain structured and unstructured data of employees from the enterprise resource planning system or human resource management system within the enterprise.

[0007] Optionally, the multidimensional features include several of the following: Skills, potential, behavioral preferences, and collaboration patterns.

[0008] Optionally, the step of dynamically adjusting the label weights of each feature in the employee competency map based on employee business feedback using reinforcement learning methods includes: The Bandit or RL agent is used to calculate different weight combination strategies in multidimensional features; Develop an optimal weighting strategy with employee performance as the primary objective. The optimal weight combination strategy is verified using A / B Testing or PSM algorithm, and the verification results are fed back to the Bandit or RL agent so that the Bandit or RL agent can be updated.

[0009] Optionally, before employing a graph neural network to structurally represent the multidimensional features of employees based on multi-source behavioral data, the method further includes: The multi-source behavioral data is anonymized, encrypted, audited for access, or subjected to differential privacy processing.

[0010] A second aspect of the present invention provides an AI-powered intelligent human decision-making system based on employee behavior analysis, comprising: The graph generation module is used to use graph neural networks to structurally represent the multidimensional features of employees based on multi-source behavioral data of employees, and generate an employee capability graph. The dynamic adjustment module is used to dynamically adjust the label weights of each feature in the employee competency graph based on the employee's business feedback using reinforcement learning methods. The decision evaluation module is used to evaluate employees based on the label weights.

[0011] Optionally, the graph generation module uses a graph neural network to structurally represent the multidimensional features of employees based on multi-source behavioral data, generating an employee capability graph. The steps include: Acquire multi-source behavioral data of employees; wherein, the multi-source behavioral data includes structured and / or unstructured data; Feature extraction, semantic parsing, and format standardization are performed on the multi-source behavioral data; A graph neural network is used to construct an employee capability map with multidimensional features.

[0012] Optionally, the dynamic adjustment module dynamically adjusts the label weights of each feature in the employee competency map based on employee business feedback using reinforcement learning methods. The steps include: The Bandit or RL agent is used to calculate different weight combination strategies in multidimensional features; Develop an optimal weighting strategy with employee performance as the primary objective. The optimal weight combination strategy is verified using A / B Testing or PSM algorithm, and the verification results are fed back to the Bandit or RL agent so that the Bandit or RL agent can be updated.

[0013] Furthermore, this invention also provides an application of an AI-powered intelligent human decision-making system based on employee behavior analysis. This AI-powered intelligent human decision-making system connects to the enterprise resource planning system and / or human resource management system within the enterprise through a standardized interface. Compared with existing technologies, the beneficial effects of this invention are: This invention provides an AI-powered intelligent talent decision-making method, system, and application based on employee behavior analysis. The method includes using a graph neural network to structure and represent the multidimensional features of employees based on multi-source behavioral data, generating an employee competency map; dynamically adjusting the label weights of each feature in the employee competency map based on employee business feedback using reinforcement learning; and evaluating employees based on these label weights. This invention has the following beneficial effects: 1) Construct an employee competency graph through GNN. Based on GNN learning, node / edge representation can be performed, which can capture implicit competency transfer and team synergy effects; 2) Employ reinforcement learning methods to dynamically adjust the label weights of each feature in the employee competency profile, so that the label weights and profiling strategies can be optimized with business indicators as the goal, avoiding blind overfitting of historical biases.

[0014] 3) Combining natural language semantic vectors obtained through NLP technology with behavioral temporal embeddings obtained through GNN can simultaneously characterize the difference between "can do" and "does often", improving the precision of ability judgment. Attached Figure Description

[0015] Figure 1 This is a flowchart of the AI-powered intelligent talent decision-making method based on employee behavior analysis proposed in this invention. Figure 2 The present invention proposes Figure 1 A detailed step diagram of step S1 is shown below; Figure 3 The present invention proposes Figure 1 A detailed step diagram of step S2 is shown below; Figure 4 This is a detailed application diagram of the AI-powered intelligent talent decision-making method based on employee behavior analysis proposed in this invention; Figure 5 This is a schematic diagram of the structure of the AI-powered intelligent talent decision-making system based on employee behavior analysis proposed in this invention. Detailed Implementation

[0016] This invention proposes an AI-powered intelligent talent decision-making method, system, and application based on employee behavior analysis. It constructs an employee capability map that integrates multi-source or heterogeneous data, including behavioral data and performance indicators, and introduces an intelligent tagging system, multi-dimensional capability mapping, and a self-learning feedback mechanism to achieve real-time construction and accurate updating of employee profiles. While adhering to privacy and security regulations, the profile capabilities are modularly integrated and applied across scenarios in enterprise resource planning (ERP) systems for recruitment, job transfer, promotion, and training, enhancing the intelligence and scientific nature of human resource allocation.

[0017] Example 1: An AI-powered intelligent talent decision-making method based on employee behavior analysis, such as Figure 1 As shown, it includes the following steps S1 to S3.

[0018] S1: A graph neural network is used to structurally represent the multidimensional features of employees based on multi-source behavioral data of employees, and an employee capability map is generated.

[0019] In further optimized solutions, such as Figure 2 As shown, step S1 specifically includes: S11: Obtain multi-source behavioral data of employees; wherein, the multi-source behavioral data includes structured and / or unstructured data; S12: Perform feature extraction, semantic parsing, and format unification on the multi-source behavioral data; S13: Use graph neural networks to construct an employee capability map with multidimensional features.

[0020] In step S11, you can connect to the enterprise's internal enterprise resource planning system (ERP) or human resource management system (HRMS) to obtain structured and unstructured data such as employee performance, behavior logs, training records, and collaboration records.

[0021] For reasons of data privacy and security compliance, data privacy or security compliance modules can be used to de-identify, encrypt, audit access, or perform differential privacy processing on multi-source behavioral data to meet corporate compliance and regulatory requirements.

[0022] In step S12, ETL (Extract, Transform, and Load) technology can be used to extract useful data from different, scattered data sources, clean and standardize the data, and then write it into a unified talent profile database.

[0023] Natural Language Processing (NLP) technology is used to process and format data, making it into human language that computers can understand, interpret, and manipulate.

[0024] In step S13, the employee competency graph includes nodes and edges. Nodes can represent employees, skills, projects, departments, or even certificates, while edges represent the relationships between nodes.

[0025] Graph Neural Networks (GNNs) first construct an initial, rudimentary "employee competency graph" based on the raw data. At this stage, the nodes and edges are still just raw data. Then, GNNs are used to perform deep learning on this graph. GNNs allow information to propagate and aggregate in the graph. After processing by GNNs, each node in the graph is transformed into a dense numerical vector. This vector structurally encodes the node's position in the graph, its connections, and its surrounding environment.

[0026] By employing Generative Neural Networks (GNNs), we can infer the depth and interconnectedness of skills, capture potential ability transfer and team synergy, and achieve tacit skill discovery, ability path assessment, and cross-position potential identification. For example, by analyzing an employee's work and project difficulty, we can determine whether their Python skills are at a "beginner" or "expert" level. Furthermore, we can infer an employee's potential by analyzing their "position" in the network graph. For instance, if an employee's network connects to many closely related high-value skill nodes that they haven't yet mastered, or if an employee has close collaborative relationships with several recognized "high-potential" talents or domain experts, these can all be captured by GNNs as signals of "potential."

[0027] In a further preferred embodiment, the multidimensional features include several of the following: Skills, potential, behavioral preferences, and collaboration patterns can be used to construct skill tags, potential tags, behavioral preference tags, and collaboration pattern tags, and initial weights can be assigned to each tag.

[0028] S2: Based on employee business feedback, reinforcement learning methods are used to dynamically adjust the label weights of each feature in the employee competency graph.

[0029] In further optimized solutions, such as Figure 3 As shown, step S2 includes: S21: Use Bandit or RL agents to calculate different weight combination strategies in the employee competency graph; S22: A strategy to obtain the optimal weight combination based on employee performance; S23: Verify the optimal weight combination strategy using A / B Testing or PSM algorithm, and feed the verification results back to the Bandit or RL agent so that the Bandit or RL agent can update.

[0030] This application employs a "Bandit or RL agent" and an "A / B Testing or PSM algorithm" to achieve an intelligent closed loop of "rapid exploration, rigorous verification, and continuous calibration".

[0031] Bandit or RL agents can maintain or run multiple "policies" simultaneously. These policies can be different combinations of label weights. The Bandit algorithm treats these policies as different "arms," ​​allocating most traffic to the policy that appears "optimal" based on short-term feedback, and a small portion to other policies. It collects business feedback from these policies in real time, using this feedback as short-term reward signals. Based on these short-term reward signals, the Bandit or RL agent quickly updates its estimate of the "value" of each policy. Policies that perform well receive higher valuations, thus gaining more traffic in the next round (being "utilized" more frequently).

[0032] Periodically trigger A / B testing or PSM algorithms for causal verification: Initiate rigorous A / B testing or PSM algorithm analysis for causal verification, and feed the results back to the Bandit or RL agent. Based on this verified and reliable signal, the Bandit or RL agent recalibrates its value estimation model. This allows for the collection of feedback data in real-world business applications to automatically correct labels and capability assessment models. Based on long-term dynamic changes in employee behavior and performance, it continuously optimizes label weights and capability paths, achieving automatic updates to the profile.

[0033] Business results (performance, promotion, training effectiveness, etc.) are fed back into the profile model, and strategies are optimized using Bandit / RL and causal analysis. This ensures that profile adjustments are not only based on historical data but also adapt to future business changes, reducing overfitting and bias.

[0034] During application, the system monitors in real time whether the prediction performance of Bandit or RL agents, A / B Testing or PSM algorithms declines over time. Once the accuracy falls below the threshold, an alarm is immediately triggered, prompting the model to be retrained.

[0035] S3: Decision evaluation of employees based on the aforementioned label weights.

[0036] In a further preferred embodiment, while evaluating employee decisions based on the label weights, the method also includes: using bias detection or fairness constraints to constrain the label weights.

[0037] The employee competency mapping and tagging system can be seamlessly integrated into multiple business modules such as recruitment, job transfer, promotion, and training through standardized interfaces. This improves HR decision-making efficiency, reduces redundant development costs, and ensures the operability of employee profiles across the entire enterprise process. Continuous monitoring of interface response time and success rate ensures compliance with service level agreements agreed upon with business departments, guaranteeing enterprise-level availability.

[0038] The final result is quantifiable performance metrics, such as: 1) Screen freshness delay: P50, P95 (target: P50 < 5 minutes, P95 < 30 minutes, adjustable depending on company size); 2) Matching / recommendation conversion rate improvement: candidate interview pass rate, probation to full-time rate, performance improvement rate after job transfer, etc. (target: 10%+ improvement relative to baseline); 3) Job matching AUC / Precision@K: used for offline evaluation of matching model effectiveness; 4) Model stability and drift alarm rate: frequency of data drift / model performance degradation triggers and mean time to repair (MTTR); 5) Compliance audit coverage: completeness of audit events / access records (target: 100% traceability of key calls).

[0039] In summary, as Figure 4 As shown, the main process of this invention includes a multi-dimensional data acquisition layer that collects employee behavior logs, performance indicators, training trajectories, and other data. Data privacy or security compliance checks can be used to encrypt and audit the data. Then, data cleaning, semantic parsing, and fusion are performed. After that, a capability vector graph is constructed, dynamic labels are generated, the labels are optimized through a self-learning feedback mechanism, the capability paths are evaluated, and the generated talent profile is stored and applied.

[0040] This invention has significant advantages over existing technologies in terms of design objectives, implementation path, and deployment strategy, mainly including: 1) Functional advantages End-to-end profiling capabilities: forming an integrated closed loop from multi-source data collection, preprocessing, feature extraction, profiling modeling to decision invocation; unlike traditional tools that only perform "reports / statistics", this method can output executable matching / recommendation / training paths.

[0041] Integrated batch processing and real-time response: Event-driven streaming updates combined with regular batch calibration enable profiles to reflect short-term behavioral changes (e.g., 7 / 30 / 90 days) while maintaining a stable long-term baseline.

[0042] Explainable recommendations: Provide multi-granular evidence for each recommendation (such as feature contribution, graph neighbor evidence, and real-time timeline) to increase the trust of HR and business departments in the output.

[0043] 2) Advantages at the technological level Heterogeneous Employee Competency Graph and GNN Representation: By constructing an employee competency graph of employees, skills, projects, and positions and learning node / edge representations based on GNN, it is possible to capture implicit competency transfer and team synergy effects, which cannot be achieved by traditional vector retrieval or keyword matching.

[0044] Tag self-learning and strategy optimization: By combining Bandit / RL with causal verification (A / B, PSM), tag weights and profiling strategies can be optimized with business metrics as the target, avoiding blind overfitting of historical biases.

[0045] Text behavior and behavior temporal sequence fusion: By combining natural language semantic vectors obtained through NLP technology with behavior temporal embeddings obtained through GNN, the difference between "can do" and "often do" can be simultaneously depicted, improving the precision of ability judgment.

[0046] 3) Business advantages High cross-scenario reuse value: The same set of profiles / tags / maps can drive recruitment, job transfers, promotions, training and succession planning, reducing the data reconstruction costs between different business modules.

[0047] Improve decision-making efficiency and accuracy: By using graph neural networks and reinforcement learning, we can automate screening, candidate / employee ranking, gap diagnosis and training package recommendation, which can significantly shorten the human resource decision-making cycle and improve the accuracy of job matching.

[0048] 4) Compliance and Governance Advantages Privacy-first design: Supports data minimization, layered anonymization, in-transmission / in-storage encryption, audit links, and configurable retention periods to facilitate compliance with GDPR, personal information protection laws, and other requirements.

[0049] Bias and fairness verification: The system has a built-in bias detection and fairness constraint module, which can introduce constraints in recommendation and scoring (such as masking sensitive attributes and enforcing diversity thresholds) to reduce the social / legal risks brought about by the algorithm.

[0050] 5) Advantages in operation and maintenance and engineering Microservices and Observability: Modular service partitioning, monitoring, and alerting (data quality / model drift / SLA) ensure enterprise-level availability; the stream-batch tiered deployment scheme facilitates horizontal scaling and capacity elasticity. Stream processing refers to handling tasks with high real-time requirements (such as collecting user feedback in real time and updating the reward signal of the Bandit model), while batch processing refers to handling computationally intensive tasks with low real-time requirements (such as running the GNN model at a set time every morning to update the capability map of all employees in the company). This tiered architecture can meet both the needs of rapid response to real-time events and efficient handling of heavy computations.

[0051] Degradation and protection mechanisms: In case of abnormal results or deterioration in data quality, the system can revert to rule-based strategies or manual approval workflows to ensure business continuity.

[0052] Example 2: Based on the same inventive concept, this invention also provides an AI-powered intelligent human decision-making system based on employee behavior analysis, such as... Figure 5 As shown, it includes: The graph generation module is used to use graph neural networks to structurally represent the multidimensional features of employees based on multi-source behavioral data of employees, and generate an employee capability graph. The dynamic adjustment module is used to dynamically adjust the label weights of each feature in the employee competency graph based on the employee's business feedback using reinforcement learning methods. The decision evaluation module is used to evaluate employees based on the label weights.

[0053] In a further preferred embodiment, the graph generation module uses a graph neural network to structurally represent the multidimensional features of employees based on multi-source behavioral data, generating an employee capability graph. The steps include: Acquire multi-source behavioral data of employees; wherein, the multi-source behavioral data includes structured and / or unstructured data; Feature extraction, semantic parsing, and format standardization are performed on the multi-source behavioral data; A graph neural network is used to construct an employee capability map with multidimensional features.

[0054] In a further preferred embodiment, the dynamic adjustment module dynamically adjusts the label weights of each feature in the employee competency map based on employee business feedback using a reinforcement learning method. The steps include: The Bandit or RL agent is used to calculate different weight combination strategies in multidimensional features; Develop an optimal weighting strategy with employee performance as the primary objective. The optimal weight combination strategy is verified using A / B Testing or PSM algorithm, and the verification results are fed back to the Bandit or RL agent so that the Bandit or RL agent can be updated.

[0055] This system can achieve the following: 1) Offline evaluation: Use historical data for cross-validation, with metrics including matching AUC, Precision@K, recall, and Calibration; at the same time, perform counterfactual simulations to evaluate the impact of strategy changes.

[0056] 2) Online evaluation: A / B testing, PSM algorithm, Bandit or RL agent control experiment are used to measure real business KPIs (probation conversion rate, performance improvement, recruitment cycle, turnover rate, etc.) and significance test is used to determine whether the strategy improvement is better than the baseline.

[0057] 3) Safety assessment: Conduct sensitive attribute impact analysis on the model and evaluate fairness indicators (statistical differences, equilibrium errors, etc.).

[0058] 4) Mitigating data quality risks: Establish a data quality rule base (missing rate, duplication rate, outlier threshold), and perform cleaning at the ETL layer or rollback to manual verification.

[0059] 5) Improve algorithm bias: Add sensitive attribute screening, cluster stability detection and fairness constraints (such as fairness regularization or threshold adjustment) to the training and deployment process.

[0060] 6) Avoid the risk of over-automation: Key talent decisions (such as senior promotions) default to the "model suggestion + manual review" process.

[0061] 7) Avoid privacy risks: Employ layered desensitization, minimize raw data access, and support options for differential privacy or homomorphic encryption (enabled according to enterprise compliance requirements).

[0062] Deployment and Implementation Recommendations Preparation phase: Analyzing data sources, establishing a unified identity strategy, and creating a data dictionary and permission matrix; Pilot phase: Select 1-2 business lines (such as recruitment and L&D in the technology department) for a 3-6 month pilot program; monitor key KPIs and conduct regular reviews; Promotion phase: Based on the success of the pilot program, expand to more organizational units, gradually open up automation permissions and improve governance processes; Governance and Operations: Establish a model / data audit committee, a regular drift detection and assessment mechanism, and SLAs and contingency plans.

[0063] The following are some recruitment scenarios where this system can be used: Data input: Job description (JD), historical competency profiles, candidate resumes, online assessment scores, etc.; Processing steps: Job descriptions are used to extract job competency vectors via NLP, and the similarity between candidate profiles and job vectors is calculated; candidates with high similarity are further "verified with behavioral evidence" (project contributions / learning behaviors within the last 90 days). Output: Generate two types of candidate lists (preferred list and alternative list) according to the scenario, with explanations (e.g., 3 core competency gaps, evidence of key events in the last 90 days, etc.).

[0064] Feedback: Performance and retention rate within 30 / 90 / 180 days after hiring are fed back as reward signals to optimize JD→profile mapping and tag weights.

[0065] Training planning scenarios (implementation details): Inputs: Employee profile, job competency map, course library metadata (skill tags, expected learning time, rating mode); Processing: Calculate the shortest skill path from an individual to their target position based on graph path cost, and prioritize matching courses / project exercises that can reduce costs along the path; Outputs: Personalized learning roadmap (short-term 3 months, long-term 12 months) and expected ability improvement assessment (quantitative indicators).

[0066] The main code process is as follows: (1) API / Interface Sample (Example) 1) Triggering a portrait refresh: POST / api / v1 / profile / {person_id} / refresh Authorization: Bearer <token> { "trigger":"event", "event_type":"commit / pull_request / assessment / completion", "payload":{...} } 2) Obtaining portraits and evidence GET / api / v1 / profile / {person_id} Response: { "person_id":"P10001", "profile_vec":"...", "tags":[{"tag":"communication skills","weight":0.68,"evidence":[...]}], "last_updated":"2025-08-20T06:12:00Z" } 3) Send back business feedback (for self-learning) POST / api / v1 / feedback { "person_id":"P10001", "scene":"hiring / rotation / training", "action":"hired / promoted / completed", "outcome":"succeeded / failed", "metrics":{"90d_performance_delta":+0.12} } (2) Example of weight update reward = map_business_outcome_to_reward(feedback) For tags in profile tags: tag.weight = decay(tag.weight, delta_t) tag.weight += learning_rate * attribution(tag, reward) normalize(profile.tags) (3) Data Dictionary (Example) person_id: Employee unique identification (Internal ID of the company).

[0067] event_ts: Event timestamp (UTC).

[0068] source: Data source (ERP, Git, LMS, Survey, Manual).

[0069] feature_vector: Numerical / dense vector after multi-source feature splicing (JSON / BINARY).

[0070] tags: Tag set ([{"tag":"code quality","weight":0.72,"confidence":0.85}]).

[0071] profile_version: Profile version number (semantics: major.minor).

[0072] skill_node_id: Skill graph node ID.

[0073] Embodiment 3 The application also provides an application of an AI intelligent artificial decision system based on employee behavior analysis, wherein the AI intelligent artificial decision system accesses an enterprise resource planning system and / or a human resource management system in an enterprise through a standardized interface.

[0074] The above is only an embodiment of the present application and is not used to limit the present application, and any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application is included in the scope of the claims of the present application.< / token>

Claims

1. An AI-powered intelligent talent decision-making method based on employee behavior analysis, characterized in that, include: A graph neural network is used to structurally represent the multidimensional features of employees based on multi-source behavioral data, generating an employee capability map; The label weights of each feature in the employee competency graph are dynamically adjusted based on employee business feedback using reinforcement learning methods. Employees are evaluated and decisions are made based on the aforementioned label weights.

2. The AI-powered intelligent talent decision-making method based on employee behavior analysis according to claim 1, characterized in that, The step of using a graph neural network to structurally represent the multidimensional features of employees based on multi-source behavioral data to generate an employee competency map includes: Acquire multi-source behavioral data of employees; wherein, the multi-source behavioral data includes structured and / or unstructured data; Feature extraction, semantic parsing, and format standardization are performed on the multi-source behavioral data; A graph neural network is used to construct an employee capability map with multidimensional features.

3. The AI-powered intelligent talent decision-making method based on employee behavior analysis according to claim 1, characterized in that, The specific steps involved in acquiring multi-source behavioral data of employees are as follows: Obtain structured and unstructured data of employees from the enterprise resource planning system or human resource management system within the enterprise.

4. The AI-powered intelligent talent decision-making method based on employee behavior analysis according to claim 1 or 2, characterized in that, The multidimensional features include several of the following: Skills, potential, behavioral preferences, and collaboration patterns.

5. The AI-powered intelligent talent decision-making method based on employee behavior analysis according to claim 1, characterized in that, The method of dynamically adjusting the label weights of each feature in the employee competency graph based on employee business feedback using reinforcement learning includes the following steps: The Bandit or RL agent is used to calculate different weight combination strategies in multidimensional features; Develop an optimal weighting strategy with employee performance as the primary objective. The optimal weight combination strategy is verified using A / B Testing or PSM algorithm, and the verification results are fed back to the Bandit or RL agent so that the Bandit or RL agent can be updated.

6. The AI-powered intelligent talent decision-making method based on employee behavior analysis according to claim 1, characterized in that, Before employing a graph neural network to structurally represent the multidimensional features of employees based on their multi-source behavioral data, the process also includes: The multi-source behavioral data is anonymized, encrypted, audited for access, or subjected to differential privacy processing.

7. An AI-powered intelligent human decision-making system based on employee behavior analysis, characterized in that, include: The graph generation module is used to use graph neural networks to structurally represent the multidimensional features of employees based on multi-source behavioral data of employees, and generate an employee capability graph. The dynamic adjustment module is used to dynamically adjust the label weights of each feature in the employee competency graph based on the employee's business feedback using reinforcement learning methods. The decision evaluation module is used to evaluate employees based on the label weights.

8. The AI-powered intelligent human decision-making system according to claim 7, characterized in that, The graph generation module uses a graph neural network to structurally represent the multidimensional features of employees based on multi-source behavioral data, generating an employee competency graph. The steps include: Acquire multi-source behavioral data of employees; wherein, the multi-source behavioral data includes structured and / or unstructured data; Feature extraction, semantic parsing, and format standardization are performed on the multi-source behavioral data; A graph neural network is used to construct an employee capability map with multidimensional features.

9. The AI-powered intelligent human decision-making system according to claim 7, characterized in that, The dynamic adjustment module dynamically adjusts the label weights of each feature in the employee competency map based on employee business feedback using reinforcement learning methods. The steps include: The Bandit or RL agent is used to calculate different weight combination strategies in multidimensional features; Develop an optimal weighting strategy with employee performance as the primary objective. The optimal weight combination strategy is verified using A / B Testing or PSM algorithm, and the verification results are fed back to the Bandit or RL agent so that the Bandit or RL agent can be updated.

10. An application of the AI-powered intelligent human decision-making system based on employee behavior analysis as described in any one of claims 7-8, characterized in that, The AI-powered intelligent human decision-making system is connected to the enterprise's internal enterprise resource planning system and / or human resource management system through standardized interfaces.

Images