Enterprise core talent loss analysis method based on Markov logic network

Abstract

The invention discloses an enterprise core talent loss analysis method based on a Markov logic network. The method comprises the following steps: S1, collecting a multi-dimensional data set related to enterprise core talent loss; s2, generating a structured multi-dimensional data set; s3, forming an enterprise core talent loss feature set; s4, constructing an initial improved Markov logic network model by using the enterprise core talent loss feature set; s5, forming an optimized causal path structure; and S6, updating the improved Markov logic network model according to the optimized Snake control path, performing online reasoning and risk assessment on the updated improved Markov logic network model by using enterprise core talent loss related data collected in real time, and outputting a loss risk assessment result of each core talent. According to the method, the adaptive capacity of the model to the individual difference of the employees and the modeling depth of a causal chain are remarkably improved, and a more interpretable loss early warning basis is provided for enterprise managers.

Description

Technical Field

[0001] The invention relates to the field of personnel technology, and in particular to a method for analyzing the loss of core talents of an enterprise based on a Markov logic network. Background Art

[0002] With the intensification of corporate competition and the continuous evolution of organizational structure, core talents, as key resources for corporate strategic development, their stability and mobility directly affect the company's operational efficiency and innovation capabilities. However, in actual management, companies often face the problems of frequent loss of core talents, difficulty in accurately identifying the causes, and lack of prediction mechanisms. This is particularly prominent in high-tech, financial, and Internet knowledge-intensive industries. How to effectively identify the risk of core employee loss, analyze its causal relationship, and provide an explainable early warning mechanism has become an important issue that needs to be urgently addressed in the current human resource management field.

[0003] At present, research on the issue of employee turnover is mostly focused on statistical modeling and machine learning methods. Mainstream methods such as logistic regression, decision trees, support vector machines, and random forest models rely on historical data and feature engineering to make binary predictions on whether employees will turnover. Although the prediction of turnover probability has been achieved to a certain extent, there are obvious defects: on the one hand, most methods are only based on the correlation modeling between features and results, which makes it difficult to reveal the causal relationship between employee behavior, organizational environment, and turnover risk, and lack the ability to model complex interactive factors; on the other hand, existing models generally have black box problems, making it difficult to provide management with explainable decision-making basis. They only give a "high risk" judgment but cannot point out the driving path and key influencing factors behind it, limiting the operability of the prediction results in actual management scenarios.

[0004] In addition, current talent loss analysis methods generally lack dynamic adjustment and online reasoning capabilities. Faced with rapid changes in employee status and continuous disturbances in environmental factors, static models are unable to respond to new data in real time, resulting in delayed or even invalid warnings. At the same time, some systems rely on fixed rule bases or expert experience, are unable to adapt to individual differences among employees, and lack the ability to adaptively analyze heterogeneous behaviors.

[0005] Therefore, there is an urgent need for a talent loss analysis method that integrates causal logical reasoning mechanisms, has dynamic optimization capabilities, and can provide transparent and explainable outputs to solve the above problems. Summary of the invention

[0006] One purpose of the present invention is to propose a method for analyzing the loss of core talents in an enterprise based on a Markov logic network. The present invention significantly improves the model's adaptability to individual differences of employees and the modeling depth of the causal chain, providing enterprise managers with a more explainable loss warning basis.

[0007] A method for analyzing the loss of core talents in enterprises based on Markov logic network according to an embodiment of the present invention includes the following steps:

[0008] S1. Collect a multi-dimensional data set related to the loss of core talents in the enterprise;

[0009] S2. Preprocess the multi-dimensional data set to generate a structured multi-dimensional data set;

[0010] S3. Based on the structured multi-dimensional data set, perform feature engineering, extract and vectorize the key feature variables affecting the loss of core talents, and form a feature set for the loss of core talents in the enterprise;

[0011] S4. Use the feature set for the loss of core talents in the enterprise to construct an initial improved Markov logic network model;

[0012] S5. Encode the key causal rule paths in the initial improved Markov logic network model as Snake control paths, where the Snake control paths are used to express the causal interaction relationships between the logical rules in the risk prediction of the loss of core talents in the enterprise. Use the Snake optimization algorithm to iteratively optimize the encoded Snake control paths to form an optimized causal path structure;

[0013] S6. Update the improved Markov logic network model according to the optimized Snake control path, and use the real-time collected data related to the loss of core talents in the enterprise to perform online inference and risk assessment on the updated improved Markov logic network model, and output the loss risk assessment results of each core talent.

[0014] Optionally, the S1 includes the following steps:

[0015] S11. Set a data collection time window, and collect the employee basic information data set from the enterprise internal management system within the time window. The employee basic information data set includes employee number, employment time, position level, department where the employee is located, and current on-the-job status;

[0016] S12. Collect the salary data set related to the employee salary structure. The employee salary data set includes the employee's monthly salary, year-end bonus, total welfare value, and its relative deviation from the average level of the same position;

[0017] S13. Collect the employee career development path data set. The employee career development path data set includes the sequence of position level changes during the employee's tenure;

[0018] S14. Collect the employee satisfaction survey data set. The employee satisfaction survey data set is based on the enterprise satisfaction scale to score the employees in four dimensions: work engagement, leadership relationship, work recognition, and organizational belonging, construct a satisfaction vector, and calculate the comprehensive satisfaction score;

[0019] S15. Collect the employee work environment evaluation data set, which includes the subjective evaluation scores of employees on office resources, team collaboration atmosphere, workload balance, and remote work support, construct a work environment vector, and then calculate the comprehensive environment satisfaction index;

[0020] S16. Uniformly encode and time-align the employee basic information data set, employee salary data set, employee career development path data set, employee satisfaction survey data set, and employee work environment evaluation data set to construct an integrated multi-dimensional data set D across time nodes raw 。

[0021] Optionally, the S2 includes the following steps:

[0022] S21. Check the field integrity of the multi-dimensional data set D raw Remove the records with missing primary key fields and retain the data samples with complete primary key fields;

[0023] S22. Perform outlier detection on the continuous numerical variables in the multi-dimensional data set D raw Adopt the outlier identification rule based on the quartile method, define the upper threshold U k and the lower threshold L k for each variable. For the value x i,k of the k-th variable in the i-th sample, if it satisfies x i,k >U k or x i,k <L k , then mark it as an outlier;

[0024] S23. Correct the detected outlier variables and replace the outliers by filling them with the distribution median;

[0025] S24. Fill in the missing numerical or categorical variables in the multi-dimensional data set D raw . For numerical variables, fill them with the mean value, and for categorical variables, adopt the mode filling strategy;

[0026] S25. Perform normalization processing on all continuous numerical variables, scale the variable values to the interval [0, 1], reorganize the data after outlier correction, missing value filling, and standardization into a structured multi-dimensional data set D struct , and perform alignment processing on the sample time dimension.

[0027] Optionally, the S3 includes the following steps:

[0028] S31. Based on the structured multi-dimensional data set D structExtract the employee behavior characteristics, which include the employee attendance frequency, overtime hours, job transfer frequency, and promotion cycle, and construct a behavior feature vector

[0029]

[0030] S32. Based on the structured multi-dimensional dataset D struct Extract the employee psychological characteristics from the employee satisfaction survey data, and construct a psychological feature vector

[0031] S33. Based on the organizational structure and position trajectory information in the structured multi-dimensional dataset D struct Extract the organizational characteristics, and construct an organizational feature vector The organizational feature vector includes the average turnover rate of employees in the department Department size The current job level l i And the job mobility index

[0032] S34. Based on the work environment evaluation information in the structured multi-dimensional dataset D struct Extract the external environment characteristics, and construct an external environment feature vector

[0033] S35. Cascade and combine the employee behavior feature vector Psychological feature vector Organizational feature vector And the external environment feature vector To form a comprehensive feature vector in the enterprise core talent loss feature set:

[0034]

[0035] Among them, f i Represents the comprehensive feature vector of the i-th employee;

[0036] S36. Perform unified format encoding and storage on the comprehensive feature vectors of all employees, and construct the enterprise core talent loss feature set F = {f1, f2, …, f N}, where N represents the total number of employee samples in the structured multi-dimensional dataset D struct In.

[0037] Optionally, the S4 includes the following steps:

[0038] S41. Based on the enterprise core talent loss feature set F, construct a node set V = {X1, X2, …, X M}, where each node X jIt corresponds to a feature dimension of the comprehensive feature vector of employees concentrated on the loss characteristics of the enterprise's core talents. The feature dimension includes employee behavior characteristics, psychological characteristics, organizational characteristics, and external environment characteristics, and is used to represent multi-dimensional factors affecting employees' turnover tendency;

[0039] S42. Introduce a semantics-driven feature combination operator Γ(·), and according to the causal logic relationship of the loss of core talents, perform nested combination on the potential interaction relationships between nodes to generate a set of first-order logic rules Φ. Each rule φ in the set of first-order logic rules l is composed of the following form:

[0040] φl:

[0041] where X a , X b represent the employee promotion frequency and work load characteristic nodes, and the node of significant turnover risk (X t ) represents the target node of potential turnover tendency. The rule indicates that there is a turnover risk when promotion is restricted and the load is high. It is automatically generated based on the feature combination operator Γ(X a , X b );

[0042] S43. Use a weight assignment method guided by the attention mechanism to weight the set of first-order logic rules Φ and construct a weight set W. The initial value of each weight w in the weight set l is:

[0043]

[0044] where α i,l represents the correlation attention weight between employee i and rule φ l , represents whether the employee feature f i satisfies the rule φ l ;

[0045] S44. Construct the set of first-order logic rules Φ, the node set V, and the weight set W into an initial improved Markov logic network model M = (V, Φ, W). The joint probability distribution of the improved Markov logic network model is defined as the conditional probability of modeling the turnover risk under different employee feature combinations:

[0046]

[0047] where Y is the employee turnover label set, and n l (F, Y) represents the number of matches where rule φ l is consistent with the turnover result under the feature set F, and Z is the normalization factor;

[0048] S45. Construct a dynamic weight update module for the feedback mechanism based on the sensitivity of talent heterogeneity, and adjust the weight set W:

[0049]

[0050] Among them, represents the weight updated in the (t + 1)-th round of iteration of the logical rule, η is the learning rate, and δ l (f i ) represents the historical deviation degree of employee i from the rule φ l , and λ is the adjustment factor;

[0051] S46. After each round of training iteration, execute the logical rule reconstruction process, retain the rule set Φ * with stability higher than the threshold and prediction contribution greater than the threshold, and introduce new rules in combination with the differences in the churn paths between samples to form an incremental rule set ΔΦ. Finally, update the initial improved Markov logic network model to M * =(V, Φ * ∪ΔΦ, W * ).

[0052] Optionally, the S5 includes the following steps:

[0053] S51. Extract the key logical rule paths from the improved Markov logic network model M * and construct a set of key causal paths Each path P k represents a sequence of logical rules connected by an inference path and is used to represent the causal propagation chain between employee characteristics;

[0054] S52. Encode each causal path P k into a Snake control path where s k,t is the feature node or logical state corresponding to the t-th control point on the path, and the Snake control path is used to model the node sequence and rule jump in the talent churn inference process;

[0055] S53. Construct an energy function for the Snake path The energy function consists of a path structure term, a data adaptation term, and a logical consistency term:

[0056]

[0057] Among them, represents the local continuity of the path, represents the matching error between the control point and the feature data f i , represents the matching error between the control point and the original first-order logical rule set Φ *Consistency measure of ∪ΔΦ, where α, β, and γ are weight coefficients;

[0058] S54. Based on the feature set F and the churn label Y, use the gradient descent method to optimize each Snake path so that the energy function is minimized, and the update strategy is:

[0059]

[0060] where η1 is the path step size, is the position of the control point in the n-th round of optimization;

[0061] S55. When three consecutive control points in the Snake control path satisfy the following conditions in three consecutive iterations: the attendance frequency of the behavioral feature sub-vector in the employee feature vector is lower than 70% of the sample mean, and the job mobility index corresponding to the organizational feature sub-vector is higher than 130% of the job average level, and the organizational belonging score of the psychological feature sub-vector is lower than the median, trigger path structure adjustment;

[0062] S56. Map each optimized Snake control path to the optimized causal path to form the optimized set of causal path structures

[0063] Optionally, the path structure adjustment rules are as follows:

[0064] If and then insert the current control point into the behavior-organization rule node

[0065]

[0066] If both are satisfied, then replace the original control point logic rule with a joint rule node that integrates the three factors of behavior, psychology, and organization

[0067]

[0068] The rule node and will be temporarily inserted into the current path control point position and participate in the recalculation of the subsequent energy function and the dynamic evolution of the path structure.

[0069] Optionally, the S6 includes the following steps:

[0070] S61. Map the optimized set of causal path structures to the improved Markov logic network model M * as the inference subgraph structure constraint condition, limiting the inference scope of the improved Markov logic network model to only execute within the path structure;

[0071] S62. Use the newly real-time collected structured multi-dimensional data set to construct a dynamic employee feature input set, where each feature vector in the dynamic employee feature input set represents the behavioral, psychological, organizational, and environmental feature states of the i-th employee in the current time window;

[0072] S63. Input the feature vector F real into the improved Markov logic network model M * and combine it with the causal path structure to perform conditional probability inference, obtaining an employee turnover probability output set Each in the employee turnover probability output set represents the turnover risk value of employee i under the current feature;

[0073] S64. Normalize all the turnover risk values values and generate a turnover risk score vector R = {r1, r2,..., r N}, where r i ∈[0, 1], which is used to represent the employee turnover possibility level, and mark the risk level status of each employee according to the set threshold θ:

[0074] Employee turnover possibility level determination: r i ≥0.8 is a high risk, triggering a red warning;

[0075] Employee turnover possibility level determination: 0.5 ≤ r i <0.8 is a medium risk, triggering a yellow warning;

[0076] Employee turnover possibility level determination: r i <0.5 is a low risk, not triggering a warning;

[0077] S65. Output the corresponding causal path explanation information for each high-risk or medium-risk employee, including the set of triggered logical rules critical path and the sequence of logical rule nodes activated in the current inference, generating a risk explanation report for managers.

[0078] The beneficial effects of the present invention are:

[0079] (1) The present invention uses a Markov logic network model to construct a causal inference structure among employee characteristics. By introducing a semantic composition operator, it automatically generates a set of first-order logic rules from four types of characteristics: behavior, psychology, organization, and external environment, avoiding the subjectivity and limitations of manual rule design. At the same time, it constructs a logical rule weight allocation mechanism in combination with the attention mechanism to guide the model to dynamically learn the importance weights of rules during the training process, forming a set of logical weights, realizing the adaptive enhancement of causal rules and the elimination and update of invalid rules. This mechanism significantly improves the model's adaptability to individual differences among employees and the modeling depth of causal chains, providing a more interpretable basis for employee turnover early warning for enterprise managers.

[0080] (2) The present invention proposes to encode the key rule path in the Markov logic network as a Snake control path and constructs an energy function including three items: path structure continuity, feature adaptability, and logical consistency. In each iteration, the Snake path adjusts its structure according to the historical employee turnover feature feedback and the energy function gradient, dynamically updating the positions and logical states of control points, making the model's inference path closer to the real turnover mechanism. Especially in the scenario of high correlation among multiple variables, the Snake path can enhance the logical consistency of the path by introducing fusion rule nodes, solving the problem that static rules are difficult to express cross-influences.

[0081] (3) The present invention maps the optimized Snake control path to a subgraph inference structure in the Markov logic network, limiting the inference scope to the key causal path, significantly improving the model's computational efficiency and real-time response ability. At the same time, the system generates a risk score vector for each employee during the inference stage and combines the activated path nodes to output its turnover risk explanation report, including key logical rules, triggered paths, and turnover probability values, supporting enterprises to implement differential intervention and retention measures based on traceable causal logic. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation to the present invention. In the drawings:

[0083] Figure 1 is a flowchart of a method for analyzing the turnover of core talents in an enterprise based on a Markov logic network proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0084] Now, the present invention will be further described in detail with reference to the drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0085] Reference Figure 1, A method for analyzing the loss of core talents in enterprises based on Markov logic network, including the following steps:

[0086] S1. Collect a multi-dimensional data set related to the loss of core talents in the enterprise;

[0087] S2. Preprocess the multi-dimensional data set to generate a structured multi-dimensional data set;

[0088] S3. Based on the structured multi-dimensional data set, perform feature engineering, extract and vectorize the key feature variables affecting the loss of core talents, and form a feature set for the loss of core talents in the enterprise;

[0089] S4. Use the feature set for the loss of core talents in the enterprise to construct an initial improved Markov logic network model;

[0090] S5. Encode the key causal rule paths in the initial improved Markov logic network model as Snake control paths. The Snake control paths are used to express the causal interaction relationships between various logical rules in the risk prediction of the loss of core talents in the enterprise. Use the Snake optimization algorithm to iteratively optimize the encoded Snake control paths to form an optimized causal path structure;

[0091] S6. Update the improved Markov logic network model according to the optimized Snake control path, and use the real-time collected data related to the loss of core talents in the enterprise to perform online inference and risk assessment on the updated improved Markov logic network model, and output the risk assessment results of the loss of each core talent.

[0092] In this embodiment, S1 includes the following steps:

[0093] S11. Set a data collection time window, and collect the employee basic information data set from the enterprise internal management system within the time window. The employee basic information data set includes employee number, employment time, position level, department where the employee is located, and current on-the-job status;

[0094] S12. Collect the salary data set related to the employee salary structure. The employee salary data set includes the employee's monthly salary, year-end bonus, total welfare value, and their relative deviation from the average level of the same position;

[0095] S13. Collect the employee career development path data set. The employee career development path data set includes the sequence of rank changes during the employee's tenure;

[0096] S14. Collect the employee satisfaction survey data set. The employee satisfaction survey data set is based on the enterprise satisfaction scale to score the employees in four dimensions: work engagement, leadership relationship, work recognition, and organizational belonging, construct a satisfaction vector, and calculate the comprehensive satisfaction score;

[0097] S15. Collect the employee work environment evaluation data set, which includes the subjective evaluation scores of employees on office resources, team collaboration atmosphere, workload balance, and remote work support, construct a work environment vector, and then calculate the comprehensive environment satisfaction index;

[0098] S16. Uniformly encode and align the time of the employee basic information data set, employee salary data set, employee career development path data set, employee satisfaction survey data set, and employee work environment evaluation data set to construct an integrated multi-dimensional data set D across time nodes raw 。

[0099] In this embodiment, S2 includes the following steps:

[0100] S21. Check the field integrity of the multi-dimensional data set D raw Remove the records with missing primary key fields and retain the data samples with complete primary key fields;

[0101] S22. Perform outlier detection on the continuous numerical variables in the multi-dimensional data set D raw Adopt the outlier identification rule based on the quartile method, and define the upper threshold U k and the lower threshold L k for each variable. For the value x i,k of the k-th variable in the i-th sample, if it satisfies x i,k >U k or x i,k <L k , then mark it as an outlier;

[0102] S23. Perform correction processing on the variables detected as outliers, and replace the outliers by filling them with the distribution median;

[0103] S24. Fill in the missing numerical or categorical variables in the multi-dimensional data set D raw . For numerical variables, fill them with the mean value, and for categorical variables, adopt the mode filling strategy;

[0104] S25. Perform normalization processing on all continuous numerical variables, scale the variable values to the interval [0, 1], reorganize the data after outlier correction, missing value filling, and standardization into a structured multi-dimensional data set D struct , and perform alignment processing on the sample time dimension.

[0105] In this embodiment, S3 includes the following steps:

[0106] S31. Based on the structured multi-dimensional data set D structExtract the behavioral characteristics of employees, where the behavioral characteristics of employees include the employee attendance frequency, overtime duration, job transfer frequency, and promotion cycle, and construct a behavioral feature vector

[0107] S32. Based on the structured multi-dimensional dataset D struct Extract the psychological characteristics of employees from the employee satisfaction survey data, and construct a psychological feature vector

[0108] S33. Based on the organizational structure and position trajectory information in the structured multi-dimensional dataset D struct Extract the organizational characteristics, and construct an organizational feature vector The organizational feature vector includes the average employee turnover rate in the department where the employee is located Department size The current job level l i And the job mobility index

[0109] S34. Based on the work environment evaluation information in the structured multi-dimensional dataset D struct Extract the external environment characteristics, and construct an external environment feature vector

[0110] S35. Concatenate and combine the employee behavioral feature vector Psychological feature vector Organizational feature vector And the external environment feature vector To form a comprehensive feature vector in the enterprise core talent loss feature set:

[0111]

[0112] Among them, f i Represents the comprehensive feature vector of the i-th employee;

[0113] S36. Perform unified format encoding and storage on the comprehensive feature vectors of all employees, and construct the enterprise core talent loss feature set F = {f1, f2,..., f N}, where N represents the total number of employee samples in the structured multi-dimensional dataset D struct In this embodiment, S4 includes the following steps:

[0114] In this embodiment, S4 includes the following steps:

[0115] S41. Based on the enterprise core talent loss feature set F, construct a node set V = {X1, X2,..., X M}, where each node X jIt corresponds to a feature dimension of the comprehensive feature vector of employees concentrated in the loss characteristics of the enterprise's core talents. The feature dimension includes employee behavior characteristics, psychological characteristics, organizational characteristics, and external environment characteristics, and is used to represent multi-dimensional factors affecting employees' loss tendency;

[0116] S42. Introduce the semantic-driven feature combination operator Γ(·), and according to the causal logic relationship of the loss of core talents, perform nested combination on the potential interaction relationships between nodes to generate a set of first-order logic rules Φ. Each rule φ in the set of first-order logic rules l is composed of the following form:

[0117] φl:

[0118] where, X a , X b represent the nodes of the employee promotion frequency and work load characteristics. The significant loss risk (X t ) represents the target node of the potential loss tendency. The rule indicates that there is a loss risk when the promotion is restricted and the load is high. It is automatically generated based on the feature combination operator Γ(X a , X b );

[0119] S43. Use the weight assignment method guided by the attention mechanism to weight the set of first-order logic rules Φ and construct a weight set W. The initial value of each weight w in the weight set l is:

[0120]

[0121] where, α i,l represents the correlation attention weight of employee i and rule φ l , represents whether the employee feature f i satisfies the rule φ l ;

[0122] S44. Construct the initial improved Markov logic network model M=(V,Φ,W) with the set of first-order logic rules Φ, the node set V, and the weight set W. The joint probability distribution of the improved Markov logic network model is defined as the conditional probability of modeling the loss risk under different combinations of employee characteristics:

[0123]

[0124] where, Y is the set of employee loss labels, and n l (F,Y) represents the number of matches where the rule φ l is consistent with the loss result under the feature set F, and Z is the normalization factor;

[0125] S45. Construct a dynamic weight update module for the feedback mechanism based on the sensitivity of talent heterogeneity, and adjust the weight set W:

[0126]

[0127] Among them, represents the weight after update in the (t + 1)-th round of iteration of the logical rule, η is the learning rate, and δ l (f i ) represents the historical deviation degree of employee i from the rule φ l , and λ is the adjustment factor;

[0128] S46. After each round of training iteration, perform a logical rule reconstruction process, retain the rule set Φ * with stability higher than the threshold and prediction contribution greater than the threshold, and introduce new rules in combination with the differences in the churn paths between samples to form an incremental rule set ΔΦ. Finally, update the initial improved Markov logic network model to M * =(V, Φ * ∪ΔΦ, W * ).

[0129] In this embodiment, S5 includes the following steps:

[0130] S51. Extract the key logical rule paths from the improved Markov logic network model M * to construct a set of key causal paths Each path P k represents a sequence of logical rules connected by an inference path, and is used to represent the causal propagation chain between employee characteristics;

[0131] S52. Encode each causal path P k into a Snake control path where s k,t is the feature node or logical state corresponding to the t-th control point on the path, and the Snake control path is used to model the node sequence and rule jump in the talent churn inference process;

[0132] S53. Construct an energy function for the Snake path The energy function consists of a path structure term, a data adaptation term, and a logical consistency term:

[0133]

[0134] Among them, represents the local continuity of the path, represents the matching error between the control point and the feature data f i , represents the matching error between the control point and the original first-order logical rule set Φ *Consistency measure of ∪ΔΦ, where α, β, and γ are weight coefficients;

[0135] S54. Based on the feature set F and the churn label Y, use the gradient descent method to optimize each Snake path, so that the energy function is minimized, and the update strategy is:

[0136]

[0137] where η1 is the path step size, is the position of the control point in the nth round of optimization;

[0138] S55. When three consecutive control points in the Snake control path satisfy the following conditions in three consecutive iterations: the attendance frequency of the behavior feature sub-vector in the employee feature vector is lower than 70% of the sample mean, and the job mobility index corresponding to the organizational feature sub-vector is higher than 130% of the job average level, and the organizational belonging score of the psychological feature sub-vector is lower than the median, trigger path structure adjustment;

[0139] S56. Map each optimized Snake control path to the optimized causal path to form the optimized set of causal path structures

[0140] In this embodiment, the path structure adjustment rules are as follows:

[0141] If and then insert the current control point into the behavior-organization rule node

[0142]

[0143] If are both satisfied at the same time, then replace the original control point logic rule with a joint rule node that integrates the three factors of behavior, psychology, and organization

[0144]

[0145] Rule node and will be temporarily inserted into the current path control point position and participate in the recalculation of the subsequent energy function and the dynamic evolution of the path structure.

[0146] In this embodiment, S6 includes the following steps:

[0147] S61. Map the optimized set of causal path structures to the improved Markov logic network model M * as the inference sub-graph structure constraint condition, restricting the inference scope of the improved Markov logic network model to only execute within the path structure;

[0148] S62. Use the newly real-time collected structured multi-dimensional data set to construct a dynamic employee feature input set, where each feature vector in the dynamic employee feature input set represents the behavioral, psychological, organizational, and environmental feature states of the i-th employee in the current time window;

[0149] S63. Input the feature vector F real into the improved Markov logic network model M * and combine with the causal path structure to perform conditional probability inference, obtaining an employee turnover probability output set Each in the employee turnover probability output set represents the turnover risk value of employee i under the current feature;

[0150] S64. Normalize all the turnover risk values values and generate a turnover risk score vector R = {r1, r2,..., r N}, where r i ∈[0, 1], used to represent the employee turnover possibility level, and mark the risk level status of each employee according to the set threshold θ:

[0151] Employee turnover possibility level determination: r i ≥0.8 is high risk, triggering a red warning;

[0152] Employee turnover possibility level determination: 0.5 ≤ r i <0.8 is medium risk, triggering a yellow warning;

[0153] Employee turnover possibility level determination: r i <0.5 is low risk, not triggering a warning;

[0154] S65. Output the corresponding causal path explanation information for each high-risk or medium-risk employee, including the set of triggered logical rules critical path and the sequence of logical rule nodes activated in the current inference, generating a risk explanation report for managers.

[0155] Example 1:

[0156] On November 15, 2023, a company named "B Electronics Co., Ltd.", an artificial intelligence chip design company located in a certain park in City A, found in its routine quarterly talent stability assessment that from August to October, a total of 5 core algorithm engineers in the R & D department submitted resignation applications without prior notice, accounting for nearly 17% of the core positions in this department, far higher than the company's historical average quarterly turnover rate (about 4.6%). Among them, one departing employee was the main architecture person in charge of the "No. 1" AI chip project. Their departure led to the delay in the delivery of multiple important modules, seriously affecting the delivery progress with upstream customers, and the direct loss exceeded 700,000 yuan.

[0157] On November 16, the company urgently launched the method of the present invention, jointly deployed by the Information Department and the Human Resources Center, to conduct centralized modeling and inference analysis on the data of 1,276 technical employees in the company over the past three years (from November 2020 to November 2023). The data collection scope covers six categories of information:

[0158] Employee basic information, including employee ID, start date, education level, job level, and department;

[0159] Compensation structure, including monthly salary, quarterly performance bonus, and percentile of comparison with peers;

[0160] Career development path, including annual job changes and promotion time intervals;

[0161] Scores of job satisfaction questionnaires, including 8 indicators such as "sense of belonging", "leadership relationship", and "job recognition";

[0162] Work environment evaluation, including "satisfaction with remote work support", "workstation comfort", and "team collaboration score";

[0163] Behavior logs, including attendance records, overtime hours, and job transfer frequencies in the past 18 months.

[0164] On November 18, the data processing was completed, generating a structured feature set. Each employee corresponds to a 32 - dimensional vector composed of behavioral, psychological, organizational, and environmental characteristics. The system inputs these features into the initial Markov logic network and generates causal paths based on the logic rule library. In the embodiment, in the historical trajectory of employee "W0356" (whose position is senior verification engineer), the system automatically identifies and activates the following logic rule sequence:

[0165] Rule φa:

[0166] Rule φb: Rule φc:

[0167] The system deduced that the employee "W0356" had an average monthly overtime of 68 hours in the last three months (163% higher than the company average). The position had not been promoted for 21 consecutive months. In his department, 4 people had job changes within half a year (the average job change frequency was 1.9 people per half year). The comprehensive turnover risk score reached 0.87, marked as "red warning".

[0168] At 9:05 am on November 21st, the system automatically pushed the turnover risk report of "W0356" to the Human Resources Center, including the activation nodes of the key causal path, the feature evolution map, and the behavior fluctuation trend chart for the last 12 months. At 1:30 pm, the HR arranged a "one-on-one interview" and learned that the employee indeed had a sense of career bottleneck and dissatisfaction with "performance appraisal emphasizing overtime hours". Within the following 48 hours, the company coordinated to transfer him to the core development team of the AI chip "Leiguang 2" project and simultaneously adjusted the performance appraisal mechanism. "W0356" finally cancelled his original job-hopping plan and was promoted to the person in charge of the project sub-module in February 2024.

[0169] In this deployment, the system processed the complete records of 1,276 technical employees in total, among which 89 were marked as "high-risk" employees and 242 were marked as "medium-risk" employees. A total of 312 logical rule paths were activated during the reasoning process, and the system updates the risk score dynamically every day.

[0170] At the beginning of December 2023, the company compared the prediction accuracy of the talent turnover of the method of the present invention and the traditional binary classification prediction model used before from mid-November to the beginning of December, and adopted the following comparison samples:

[0171]

[0172]

[0173] The system further counted the recall rate, false alarm rate, and average early warning time of the two methods for the turnover employees within a 30-day window period:

[0174]

[0175] In addition, among the 87 employees marked as high-risk by the system and with intervention strategies taken, 58 did not have turnover behavior within two months, and the intervention success rate was 66.7%, while the success rate of the control group (marked and intervened with the traditional method) was only 41.2%.

[0176] It is particularly worth mentioning that the system's path interpretation and prediction accuracy for "Employee E0219" have been unanimously recognized by senior management. In the satisfaction questionnaire in October 2023, this employee gave 2 points for "sense of belonging" and 3 points for "job recognition". Based on the characteristics that their year-end performance scores have been continuously low, they have not been promoted for 18 consecutive months, and the average transfer frequency in the department is relatively high, the following path was activated:

[0177]

[0178] The risk score reached 0.91, and a warning was triggered on November 5, 2023. The employee originally planned to submit a resignation application in December. After the HR conducted early communication and adjusted their project arrangements, the employee decided to continue to stay.

[0179] From the actual measurement, the method proposed in the present invention can not only achieve efficient identification and warning of brain drain, but also provide visual and interpretable causal paths to support enterprises in carrying out targeted management interventions, greatly improving management efficiency and prevention ability.

[0180] The present invention uses a Markov logic network model to construct a causal reasoning structure between employee characteristics. By introducing a semantic combination operator, it automatically generates a set of first-order logic rules from four types of characteristics: behavior, psychology, organization, and external environment, avoiding the subjectivity and limitations of manual rule design. At the same time, it constructs a logical rule weight allocation mechanism in combination with the attention mechanism to guide the model to dynamically learn the importance weights of rules during the training process, forming a logical weight set W, realizing the adaptive enhancement of causal rules and the elimination and update of invalid rules. This mechanism significantly improves the model's adaptability to individual differences among employees and the modeling depth of causal chains, providing a more interpretable basis for brain drain warning for enterprise managers.

[0181] The present invention proposes to encode the key rule paths in the Markov logic network as Snake control paths, and constructs an energy function that includes three items: path structure continuity, feature adaptability, and logical consistency. In each iteration, the Snake path adjusts its structure according to the feedback of historical employee turnover characteristics and the gradient of the energy function, dynamically updating the positions and logical states of control points, making the model's inference path closer to the real turnover mechanism. Especially in the scenario of high correlation among multiple variables, the Snake path can enhance the logical consistency of the path by introducing fusion rule nodes, solving the problem that static rules are difficult to express cross-influences.

[0182] By mapping the optimized Snake control path to the subgraph inference structure in the Markov logic network and limiting the inference scope within the key causal path, the present invention significantly improves the model calculation efficiency and real-time response ability. At the same time, the system generates a risk score vector R for each employee during the inference stage and outputs a turnover risk explanation report in combination with the activated path nodes, including key logical rules, triggered paths, and turnover probability values, supporting enterprises to implement differential interventions and retention measures based on traceable causal logic.

[0183] The above are only the preferred specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, making equivalent substitutions or changes, should be covered by the protection scope of the present invention.

Claims

1. A method for analyzing the loss of core talents in an enterprise based on a Markov logic network, characterized in that It includes the following steps: S1. Collect a multi-dimensional data set related to the loss of core talents in the enterprise; S2. Preprocess the multi-dimensional data set to generate a structured multi-dimensional data set; S3. Based on the structured multi-dimensional data set, perform feature engineering, extract and vectorize the key feature variables affecting the loss of core talents, and form a feature set for the loss of core talents in the enterprise; S4. Use the feature set for the loss of core talents in the enterprise to construct an initial improved Markov logic network model; S5. Encode the key causal rule paths in the initial improved Markov logic network model as Snake control paths. The Snake control paths are used to express the causal interaction relationships between various logical rules in the risk prediction of the loss of core talents in the enterprise. Use the Snake optimization algorithm to iteratively optimize the encoded Snake control paths to form an optimized causal path structure; S6. Update the improved Markov logic network model according to the optimized Snake control path, and use the real-time collected data related to the loss of core talents in the enterprise to perform online inference and risk assessment on the updated improved Markov logic network model, and output the risk assessment results of the loss of each core talent.

2. The method for analyzing the loss of core talents of an enterprise based on a Markov logic network according to claim 1, wherein The S1 includes the following steps: S11. Set a data collection time window, and collect the employee basic information data set from the enterprise internal management system within the time window. The employee basic information data set includes employee number, employment time, position level, department where the employee is located, and current on-the-job status; S12. Collect the salary data set related to the employee salary structure. The employee salary data set includes the monthly salary, year-end bonus, total welfare value of the employee, and their relative deviation in the average level of the same position; S13. Collect the employee career development path data set. The employee career development path data set includes the sequence of rank changes during the employee's tenure; S14. Collect the employee satisfaction survey data set. The employee satisfaction survey data set is based on the enterprise satisfaction scale to score the employees in four dimensions: work engagement, leadership relationship, work recognition, and organizational belonging, construct a satisfaction vector, and calculate the comprehensive satisfaction score; S15. Collect the employee work environment evaluation data set. The employee work environment evaluation data set includes the subjective evaluation scores of the employees on office resources, team collaboration atmosphere, work load balance, and remote work support, construct a work environment vector, and then calculate the comprehensive environment satisfaction index; S16. Uniformly encode and align the time of the employee basic information dataset, employee salary dataset, employee career development path dataset, employee satisfaction survey dataset, and employee work environment evaluation dataset to construct an integrated multi-dimensional dataset D across time nodes raw .

3. The method for analyzing the loss of core talents in an enterprise based on a Markov logic network according to claim 1, wherein The S2 includes the following steps: S21. Perform a field integrity check on the data cube D raw Eliminate the records with missing primary key fields and retain the data samples with complete primary key fields; S22. Perform outlier detection on the continuous numerical variables in the multidimensional dataset D raw using the outlier identification rule based on the quartile method, and define the upper threshold U k and the lower threshold L k for each variable value x i,k in the i-th sample. If it satisfies x i,k >U k or x i,k <L k , it is marked as an outlier; S23. Correct the variables detected as outliers, and replace the outliers by filling them with the distribution median; S24. Fill in the missing values for the missing numerical or categorical variables in the multidimensional dataset D raw For the missing numerical variables, fill them with the mean value, and for the missing categorical variables, adopt the mode filling strategy; S25. Perform normalization on all continuous numerical variables, scale the variable values to the interval [0, 1], reorganize the data after outlier correction, missing value filling, and standardization into a structured multi-dimensional dataset D struct , and align the sample time dimension.

4. A method for analyzing the loss of core talents in an enterprise based on a Markov logic network according to claim 1, characterized in that The S3 includes the following steps: S31. Based on the structured multi-dimensional dataset D struct Extract the employee behavior characteristics, where the employee behavior characteristics include employee attendance frequency, overtime hours, job transfer frequency, and promotion cycle, and construct a behavior feature vector S32. Based on the structured multi-dimensional dataset D struct Extract the psychological characteristics of employees from the employee satisfaction survey data and construct a psychological characteristic vector S33. Based on the organizational structure and position trajectory information in the structured multi-dimensional dataset D struct Extract organizational features and construct an organizational feature vector The organizational feature vector includes the average turnover rate of employees in the department Department size The current position level l i And the position mobility index S34. Extract external environmental features based on the work environment evaluation information in the structured multi-dimensional dataset D struct and construct an external environmental feature vector S35. Cascade and combine the employee behavior feature vector psychological feature vector organizational feature vector and the external environment feature vector to form a comprehensive feature vector in the enterprise core talent loss feature set: Among them, f i represents the comprehensive feature vector of the i-th employee; S36. Uniformly encode and store the comprehensive feature vectors of all employees, and construct the enterprise's core talent loss feature set F = {f1, f2, …, f N}, where N represents the total number of employee samples in the structured multi-dimensional data set D struct .

5. The enterprise core talent loss analysis method based on Markov logic network according to claim 4, characterized in that The S4 includes the following steps: S41. Based on the enterprise core talent loss feature set F, construct the node set V = {X1, X2, …, X M}, where each node X j corresponds to a feature dimension of the comprehensive feature vector of employees in the enterprise core talent loss feature set. The feature dimensions include employee behavior characteristics, psychological characteristics, organizational characteristics, and external environment characteristics, and are used to characterize the multi-dimensional factors affecting employees' turnover tendency; S42. Introduce a semantics-driven feature combination operator Γ(·), and according to the causal logic relationship of the loss of core talents, perform nested combination on the potential interaction relationships between nodes to generate a set of first-order logic rules Φ. Each rule φ in the set of first-order logic rules l is composed of the following form: Among them, X a , X b represents the employee promotion frequency and the job load characteristic node. The significant loss risk (X t ) represents the target node of the potential loss tendency. The rule indicates that there is a loss risk when the promotion is restricted and the load is high. Based on the feature combination operator Γ(X a , X b ) is automatically generated; S43. The first-order logic rule set Φ is weighted using the weight assignment method guided by the attention mechanism to construct a weight set W, and the initial value of each weight w in the weight set is: l is: Among them, α i,l represents the relevance attention weight of employee i and rule φ l , represents whether the employee feature f i satisfies rule φ l ; S44. Construct the first-order logic rule set Φ, node set V, and weight set W into the initial improved Markov logic network model M=(V,Φ,W). The joint probability distribution of the improved Markov logic network model is defined as the conditional probability of modeling the loss risk under different combinations of employee characteristics: Among them, Y is the employee turnover label set, n l (F, Y) represents the rule φ l The number of matches consistent with the turnover result under the feature set F, and Z is the normalization factor; S45. Construct a dynamic weight update module based on the feedback mechanism of talent heterogeneity sensitivity to adjust the weight set W; Among them, represents the weight of the logical rule after being updated in the (t + 1)-th round of iteration, η is the learning rate, and δ l (f i ) represents the historical deviation degree of employee i from the rule φ l , and λ is the adjustment factor; S46. Perform the logical rule reconstruction process after each round of training iteration, and retain the rule set Φ with stability higher than the threshold and prediction contribution greater than the threshold * , and introduce new rules in combination with the churn path differences between samples to form an incremental rule set ΔΦ. Finally, update the initial improved Markov logic network model to M * = (V, Φ * ∪ ΔΦ, W * ).

6. The enterprise core talent loss analysis method based on Markov logic network according to claim 5, characterized in that The S5 includes the following steps: S51. Extract the key logical rule paths from the improved Markov logic network model M * to construct a set of key causal paths For each path P k represents a sequence of logical rules connected by an inference path, which is used to represent the causal propagation chain between employee characteristics; S52. Encode each causal path P k into a Snake control path where s k,t is the feature node or logical state corresponding to the t-th control point on the path, and the Snake control path is used to model the node sequence and rule jump in the brain drain reasoning process; S53. Construct the energy function of the Snake path The energy function consists of a path structure term, a data adaptation term, and a logical consistency term: Among them, represents the local continuity of the path, represents the matching error between the control point and the feature data f i of, represents the consistency measure between the control point and the original first-order logic rule set Φ * ∪ΔΦ, where α, β, and γ are weight coefficients; S54. Based on the feature set F and the churn label Y, use the gradient descent method to optimize each Snake path to minimize the energy function and the update strategy is as follows: where η1 is the path step size, is the control point position in the n-th round of optimization; S55. When three consecutive control points in the Snake control path meet the following conditions in three consecutive iterations: the behavior feature subvector in the employee feature vector The attendance frequency of the organization is lower than 70% of the sample mean, and the organizational feature subvector Corresponding job mobility index 130% above the average level for the position, and the psychological characteristics sub-vector When the organizational sense of belonging score is lower than the median, the path structure adjustment is triggered; S56. Map each optimized Snake control path to an optimized causal path to form a set of optimized causal path structures 7. The method for analyzing the loss of core talents in an enterprise based on a Markov logic network according to claim 6, characterized in that, The path structure adjustment rules are as follows: If the following conditions are met and then insert the current control point into the behavior - organizational rule node If the following conditions are met simultaneously then replace the original control point logic rule with a joint rule node that integrates the three factors of behavior, psychology, and organization The said rule node and will be temporarily inserted into the current path control point position and participate in the subsequent recalculation of the energy function and the dynamic evolution of the path structure.

8. The enterprise core talent loss analysis method based on Markov logic network according to claim 6, characterized in that The S6 includes the following steps: S61. Map the optimized set of causal path structures to the improved Markov logic network model M * as the inference subgraph structure constraint condition, restricting the inference scope of the improved Markov logic network model to be executed only within the path structure; S62. Construct a dynamic employee feature input set using the latest real-time collected structured multi-dimensional data set. Each feature vector in the dynamic employee feature input set represents the behavior, psychology, organization, and environmental feature status of the i-th employee in the current time window; S63. Input the feature vector F real into the improved Markov logic network model M * and combine with the causal path structure to perform conditional probability inference and obtain the employee turnover probability output set Each in the employee turnover probability output set represents the turnover risk value of employee i under the current features; S64. For all churn risk values Values are normalized and a churn risk score vector R = {r1, r2, …, r N} is generated, where r i ∈ [0, 1], which is used to represent the employee churn probability level, and the risk level status of each employee is marked according to the set threshold θ: Determination of the likelihood level of employee turnover: r i ≥0.8 indicates high risk and triggers a red alert; Employee turnover probability level determination: 0.5 ≤ r i <0.8 indicates medium risk and triggers a yellow warning; Determination of the likelihood level of employee turnover: r i <0.5 indicates low risk and does not trigger an early warning; Output the corresponding causal path explanation information for each high-risk or medium-risk employee, including the set of trigger logic rules Critical path and the sequence of logic rule nodes activated in the current reasoning to generate a risk explanation report for managers.

Images