Method, device and equipment for three-dimensional occupation matching based on ability psychological motivation and medium

Abstract

The application relates to a three-dimensional occupation matching method, device, equipment and medium based on ability, psychology and motivation. The method comprises the following steps: calculating the skill matching degree of a candidate and a target post according to the resume text of the candidate and the post requirement text; if the skill matching degree is greater than a skill screening threshold, obtaining a psychological characteristic vector according to the historical behavior record of the candidate, combining a post psychological demand vector, and calculating the psychological adaptation degree of the candidate; if the psychological adaptation degree is greater than a psychological screening threshold, identifying the occupation value preference based on the occupation track data of the candidate, calculating the occupation motivation value of the candidate, and weighting and synthesizing the skill matching degree, the psychological adaptation degree and the occupation motivation value to obtain the matching accuracy; if the matching accuracy is greater than a recruitment reference threshold, inputting the matching accuracy and the occupation motivation value into a pre-trained retention prediction model to output the long-term retention rate of the candidate. The method can effectively improve the accuracy of the matching between the post and the employee and the predictability of the long-term retention of the employee.

Description

Technical Field

[0001] This invention belongs to the field of career matching technology, and in particular relates to a three-dimensional career matching method, device, equipment and medium based on ability, psychodynamics. Background Technology

[0002] As the integration of human resource management and artificial intelligence technology continues to deepen, the technology for matching people to jobs, as a core aspect of corporate talent recruitment and allocation, is undergoing a transformation from experience-driven to data-driven approaches.

[0003] Existing job-person matching methods primarily rely on semantic matching between job descriptions and resume texts. For example, the "AI-based Job Competency and Recruitment Job Matching Method" applied for by Zhongce Gaoke uses keyword extraction, entity recognition, and semantic vector similarity calculation to compare information presented in candidates' past educational background, work experience, and skill certificates with job descriptions and qualification requirements. Ultimately, it outputs a quantitative score reflecting the degree of overlap in knowledge and skills, achieving preliminary screening and matching ranking between candidates and positions. This approach, using resume text as the sole data source and explicit abilities as the only evaluation dimension, is technically mature and has low operating costs, thus being widely used in various recruitment platforms and talent management systems.

[0004] However, practice shows that "fit" (psychological fit) and "willingness" (career motivation) equally affect employees' long-term performance and retention rates. Methods based on resume text semantic analysis can essentially only touch upon the surface-level ability dimension of a candidate's "competence," lacking a quantitative channel for quantifying the match between the candidate's intrinsic psychological traits and the psychological requirements of the position. Since a candidate's personality traits, interests, and values ​​cannot be directly extracted from their resume, and the implicit psychological requirements of the position regarding stability and resilience cannot be explicitly obtained from the job description, traditional methods cannot establish a psychological fit measurement mechanism, making it even more difficult to further assess a candidate's career motivation level and long-term retention potential after joining the company. This technological blind spot leads to significant deficiencies in traditional methods for predicting a candidate's actual work performance and job stability after joining the company. Summary of the Invention

[0005] Therefore, it is necessary to provide a person-job matching method that can comprehensively and quantitatively evaluate candidates and positions from three dimensions: competence, psychological fit, and career motivation, in order to address the aforementioned technical issues.

[0006] Firstly, this application provides a three-dimensional occupational matching method based on ability, psychodynamics, and other factors, including:

[0007] Obtain the candidate's resume text and the target position's job requirements text, and calculate the skill matching degree between the candidate and the target position based on the resume text and job requirements text;

[0008] If the skill matching degree is greater than the skill screening threshold, the candidate's stability and stress resistance are analyzed based on the candidate's historical behavior records to obtain the candidate's psychological trait vector, which includes stability tendency score and stress resistance tendency score.

[0009] The psychological needs information of the job requirements text is extracted to obtain the psychological needs vector of the target job, and the psychological fit of the candidate is calculated based on the psychological trait vector and the psychological needs vector.

[0010] If the psychological fit is greater than the psychological screening threshold, the candidate's career trajectory data in the historical behavior record is used to identify the candidate's career value preferences and obtain the candidate's career value preferences, which include the stability preference quantification value and the growth preference quantification value.

[0011] Based on career value preferences and psychological fit, the candidate's career motivation score is calculated, and the skill matching score, psychological fit, and career motivation score are weighted and synthesized to obtain the matching accuracy between the candidate and the target position.

[0012] If the matching accuracy is greater than the recruitment reference threshold, the matching accuracy and career motivation value are input into the pre-trained retention prediction model, and the long-term retention rate of the candidate in the target position is output.

[0013] Furthermore, based on the resume text and job requirement text, the skill match degree between the candidate and the target position is calculated, including:

[0014] The resume text and job requirement text are structured and parsed to obtain candidate skill terms and job candidate skill terms;

[0015] A feature extraction algorithm based on a pre-trained language model is used to map each candidate skill term and job candidate skill term to the standard skill node in the standard skill knowledge graph to obtain standard skill terms and job requirement skill terms. The standard skill terms and job requirement skill terms are then summarized to obtain the candidate's explicit skill set and the job skill requirement set.

[0016] The explicit skill set and the skill requirement set are mapped to a unified skill tag vector space to obtain the candidate skill vector and the job skill requirement vector.

[0017] Calculate the weighted Jaccard similarity between the candidate's skill vector and the job skill requirement vector in the skill tag vector space, and determine the weighted Jaccard similarity as the skill matching degree.

[0018] Furthermore, historical behavioral records include project participation records and collaboration and communication records. If the skill match is greater than the skill screening threshold, stability and stress resistance analysis is performed on the candidate based on their historical behavioral records to obtain the candidate's psychological trait vector, including:

[0019] If the skill matching degree is greater than the skill screening threshold, the project start and end time, project role and final project delivery result of each project are obtained from the project participation record of historical behavior records;

[0020] Based on the project start and end times, project roles, and final project deliverables, stability analysis is conducted on candidates to obtain stability feature vectors. These stability feature vectors include average project tenure, standard deviation of project tenure, project role promotion index, and project completion rate.

[0021] From the collaborative communication records of historical behavior records, obtain the communication time, communication frequency and communication role of each communication record. The communication role includes the active communication role and the passive communication role.

[0022] Based on the proportion of communication times for each communication partner, the communication time, frequency, and role of each communication record, stress resistance analysis is conducted on candidates to obtain stress resistance feature vectors. The stress resistance feature vectors include the proportion of high-frequency communication periods, the proportion of proactive communication, and the diversity of communication partners.

[0023] By concatenating the stability feature vector and the resilience feature vector, a multi-dimensional feature vector is obtained. This multi-dimensional feature vector is then input into an ensemble learning model based on the XGBoost framework, which outputs stability tendency score and resilience tendency score.

[0024] By combining the stability tendency score and the stress resistance tendency score, a psychological trait vector of the candidate is obtained.

[0025] Furthermore, career trajectory data includes reasons for job changes and duration of employment. If the psychological fit exceeds the psychological screening threshold, career value preferences are identified based on the candidate's career trajectory data in historical behavioral records, resulting in the candidate's career value preferences, including:

[0026] The reasons for job switching and the length of service of candidates are obtained from their historical behavior records. The reasons for job switching include actively seeking development, actively seeking salary increase, actively seeking work environment improvement, being laid off, being eliminated due to performance, and being affected by organizational adjustment.

[0027] Based on the length of service of each position held by the candidate, the mean and standard deviation of the length of service are calculated. Based on the mean and standard deviation of the length of service, the reasons for job switching and the total number of positions held, the candidate's stability preference value is calculated.

[0028] Based on the trend of job level changes in each of the candidate's job experiences, the total job level increase from the first job to the last job is calculated, and the candidate's growth preference quantification value is calculated based on the reason for job switching and the total job level increase.

[0029] By combining the stability preference value and the growth preference value, the candidate's career value preference is obtained.

[0030] Furthermore, the pre-trained retention prediction model is obtained through the following steps:

[0031] Obtain the historical match accuracy and historical career motivation values ​​of all past incumbents for the target position, and combine the historical match accuracy and historical career motivation values ​​to obtain a covariate vector;

[0032] Obtain the actual departure time and event indicator variables of all historical incumbents in the target position;

[0033] A training sample set is constructed based on the covariate vectors of all historical employees, the actual departure time, and the event indicator variables.

[0034] By setting the Cox proportional hazards regression model as the algorithmic framework for the retention prediction model, we obtain the initial retention prediction model, whose risk function is:

[0035] ;

[0036] in, Let t represent the employee's turnover risk, and t represent time. As the benchmark risk function, Let the vector of regression coefficients be estimated. Let T be the covariate vector, and T be the transpose.

[0037] Based on the training sample set, the partial likelihood estimation method is used to solve for the regression coefficient vector to be estimated in the initial retention prediction model, and the regression coefficient vector is obtained.

[0038] Substituting the regression coefficient vector into the initial retention prediction model yields the trained retention prediction model.

[0039] Secondly, this application also provides a three-dimensional career matching device based on ability-psychodynamics, including:

[0040] The skills matching degree calculation module is used to obtain the resume text submitted by the candidate and the job requirements text of the target position, and calculate the skills matching degree between the candidate and the target position based on the resume text and the job requirements text.

[0041] The stability and stress resistance analysis module is used to analyze the stability and stress resistance of candidates based on their historical behavior records if the skill matching degree is greater than the skill screening threshold. The module obtains the candidate's psychological trait vector, which includes stability tendency score and stress resistance tendency score.

[0042] The psychological fit calculation module is used to extract the psychological needs information of the job requirements text, obtain the psychological needs vector of the target job, and calculate the psychological fit of the candidate based on the psychological trait vector and the psychological needs vector.

[0043] The preference identification module is used to identify the candidate's career value preferences based on the candidate's career trajectory data in historical behavior records if the psychological fit is greater than the psychological screening threshold. The career value preferences include the stability preference quantification value and the growth preference quantification value.

[0044] The matching accuracy acquisition module is used to calculate the candidate's career motivation value based on career value preferences and psychological fit, and to weight and synthesize the skill matching degree, psychological fit and career motivation value to obtain the matching accuracy between the candidate and the target position.

[0045] The retention rate prediction module is used to input the matching accuracy and career motivation value into a pre-trained retention prediction model if the matching accuracy is greater than the recruitment reference threshold, and output the candidate's long-term retention rate in the target position.

[0046] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores at least one instruction, at least one program, a code set, or an instruction set, and the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by the processor to implement any of the three-dimensional career matching methods based on ability-psychodynamics described in the embodiments of this application.

[0047] Fourthly, this application also provides a computer-readable storage medium storing at least one piece of program code, which is loaded and executed by a processor to implement any of the competency-psychodynamic three-dimensional career matching methods described in the embodiments of this application.

[0048] The aforementioned three-dimensional career matching method, device, equipment, and medium based on ability, psychology, and motivation calculate the skill matching degree between the candidate and the target position based on the candidate's resume and the job requirements. If the skill matching degree is greater than the skill screening threshold, psychological trait vectors related to stability and stress resistance are obtained based on the candidate's historical behavioral records. A job psychological demand vector is also obtained based on the job requirements. The candidate's psychological fit is calculated based on the psychological trait vector and the psychological demand vector. If the psychological fit is greater than the psychological screening threshold, the candidate's career value preferences are identified based on their career trajectory data. Combined with the psychological fit, a career motivation value is calculated. The skill matching degree, psychological fit, and career motivation value are weighted and synthesized to obtain the matching accuracy between the candidate and the target position. If the matching accuracy is greater than the recruitment reference threshold, the matching accuracy and career motivation value are input into a retention prediction model to output the candidate's long-term retention rate in the target position. It can expand the assessment dimensions from a single ability dimension to three dimensions: ability, psychological fit, and career motivation. This effectively solves the core problems of traditional job matching methods, such as the single matching dimension, neglect of psychological fit, and inability to predict retention risks. It significantly improves the accuracy of job matching and the predictability of long-term employee retention. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 This is a flowchart illustrating a three-dimensional career matching method based on ability, psychodynamics, and other methods in one embodiment.

[0051] Figure 2 This is a flowchart illustrating the steps involved in calculating the skill matching degree between a candidate and a target position based on resume text and job requirement text, as shown in one embodiment.

[0052] Figure 3 This is a schematic diagram of the structure of a three-dimensional career matching device based on ability-psychodynamics in one embodiment. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0054] In one embodiment, a three-dimensional career matching method based on ability, psychodynamics, and other factors is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, and further to a system including both a terminal and a server, and can be implemented through the interaction between the terminal and the server. Figure 1 As shown, in this embodiment, the method includes the following steps:

[0055] S101: Obtain the resume text submitted by the candidate and the job requirements text of the target position, and calculate the skill matching degree between the candidate and the target position based on the resume text and the job requirements text.

[0056] Optionally, the resume text includes unstructured information such as the candidate's educational background, work experience, project experience, and professional skills; the job requirements text includes job descriptions, qualification requirements, and professional skills requirements.

[0057] For example, the matching terminal can obtain the resume text submitted by the candidate and the job requirements text of the target position, and perform structured parsing on the resume text to obtain candidate skill terms consisting of nouns, verbs, and gerund combinations. The matching terminal can employ a feature extraction algorithm to obtain the contextual semantic vector representation of each candidate skill term based on the context sentence in which each candidate skill term appears. Based on the semantic vectors of each context and the semantic vectors of each standard skill node in a pre-constructed standard skill knowledge graph, the terminal calculates the explicit skill set reflecting the candidate's mastery of tools and knowledge through cosine similarity. The matching terminal can perform the same preprocessing and feature extraction process on the job requirements text, and additionally introduce a rule-based demand intensity discriminator to identify skill terms appearing after strong constraint phrases such as "must" and "requirement" as mandatory mastery items, thus obtaining a skill requirement set. The matching terminal can map the explicit skill set and the skill requirement set to a unified skill tag vector space to obtain the candidate skill vector and the job skill requirement vector, and then calculate the weighted Jaccard similarity between the two, determining the weighted Jaccard as the skill matching degree.

[0058] S102. If the skill matching degree is greater than the skill screening threshold, the candidate's stability and stress resistance are analyzed based on the candidate's historical behavior records to obtain the candidate's psychological trait vector, which includes stability tendency score and stress resistance tendency score.

[0059] Optionally, the skills screening threshold is a pre-set numerical constant in the matching terminal, used to determine whether the candidate meets the basic requirements of the position in terms of skills. It is dimensionless and ranges from [0,1], consistent with the range of skills matching degree. It can be pre-configured and dynamically adjusted according to factors such as the breadth of skills required for the target position, the urgency of recruitment, and the organization's tolerance for skills matching. Generally speaking, for positions with strict professional skills requirements and skills gaps that are difficult to fill through post-employment training, the skills screening threshold can be set higher; for general positions or positions for which the organization is willing to invest training resources, the skills screening threshold can be appropriately lowered. Stability and stress resistance analysis refers to the complete processing process by which the matching terminal, based on the candidate's historical behavioral records, summarizes the candidate's numerical representation in the two psychological dimensions of stability tendency and stress resistance tendency from behavioral data through a series of quantitative feature extraction and machine learning inference operations.

[0060] For example, if the skill matching degree is greater than the skill screening threshold, the matching terminal can perform stability and stress resistance analysis on the candidate based on the candidate's historical behavior records to obtain the candidate's psychological trait vector, which includes stability tendency score and stress resistance tendency score.

[0061] S103: Extract the psychological needs information of the job requirements text, obtain the psychological needs vector of the target job, and calculate the psychological fit of the candidate based on the psychological trait vector and the psychological needs vector.

[0062] For example, the matching terminal can construct a psychological need keyword library, which includes keywords related to stability and keywords related to resilience. The matching terminal can calculate the matching degree between each sentence in the job requirement text and the keywords in the psychological need keyword library, marking sentences with matching degrees exceeding a matching threshold as psychological need-related sentences. The matching terminal can input these marked psychological need-related sentences into a pre-tuned BERT (Bidirectional Encoder Representations from Transformers) semantic classification model, outputting the stability need strength and resilience need strength, both of which are dimensionless continuous values. The matching terminal can combine the stability need strength and resilience need strength to obtain the psychological need vector of the target job. The matching terminal can place the candidate's psychological trait vector and the target job's psychological need vector in a unified two-dimensional trait vector space, calculate the cosine similarity between them, and determine the cosine similarity as the psychological fit. The essence of cosine similarity is to measure the degree of consistency between two vectors in direction; the closer its value is to 1, the more consistent the direction of the candidate's psychological traits and the job's psychological needs, and the higher the degree of psychological fit.

[0063] Optionally, the construction of the psychological needs keyword database is based on classic research findings on the soft requirements of positions in the fields of organizational behavior and industrial and organizational psychology. It summarizes the implicit psychological expectations of job holders into two dimensions: stability and resilience. Each entry in the psychological needs keyword database is a word or phrase highly related to a specific psychological dimension, accompanied by a dimension tag to indicate its category of psychological need. Keywords covered in the stability dimension include, but are not limited to, long-term, stable, continuous, and steadfast. When these words appear in a job description, they usually imply that the position expects the holder to maintain a relatively long tenure. Keywords covered in the resilience dimension include... Keywords include, but are not limited to, stress resistance, high intensity, fast pace, adaptability, and composure. When these words appear in a job description, they typically suggest a high-pressure work environment or frequent unpredictable events. The matching score can be calculated as follows: For each sentence, count the number of words appearing in the psychological needs keyword database within that sentence, multiply each matched word by its weight coefficient in the database, and divide the weighted sum by the total number of words in the sentence to obtain the original matching score. Since job requirement texts may contain complex semantic structures such as negation or contrast, the matching terminal can adjust the score based on the above word-level statistics. Furthermore, a negation detection mechanism based on dependency parsing is introduced: For a hit keyword, it is analyzed whether the keyword is dominated by negation words such as "not," "without," and "avoid" in the sentence dependency tree. If a negation relationship exists, the matching contribution of the keyword is zeroed out. For words not included in the psychological need keyword database but semantically clearly pointing to stability or resilience, the matching terminal can pre-convert all keywords in the psychological need keyword database into fixed-dimensional word vectors. When the cosine similarity between a word and any keyword in the psychological need keyword database exceeds the similarity threshold, it is considered an implicit hit and included in the matching score. After the above processing, each A sentence is assigned a quantified matching score. The similarity threshold is a pre-set cosine similarity threshold used to filter irrelevant words and prevent words with significantly different meanings from being misidentified. The matching threshold is a numerical threshold used when screening sentences related to psychological needs. It is used to determine whether a sentence should be marked as a sentence related to psychological needs. The matching threshold is usually determined based on the statistical analysis results of historical job description texts. Specifically, a large number of labeled job description samples can be collected, and the matching score distribution of each sentence can be calculated. The optimal threshold value is selected based on maximizing the F1 score (i.e., the harmonic mean of precision and recall).

[0064] Optionally, the BERT semantic classification model is constructed using a "pre-training-fine-tuning" paradigm, with bert-base-chinese as the backbone network. bert-base-chinese is a Chinese pre-trained version released by Google, with a vocabulary size of 21,128 tokens and employing a full-word masking strategy. The bert-base-chinese backbone network consists of 12 stacked Transformer encoder layers, each including a multi-head self-attention sublayer and a feedforward neural network sublayer. The hidden layer dimension is 768, the number of self-attention heads is 12, and the total number of parameters is approximately 110 million. A linear output layer for sequence-level classification, i.e., the classification head, is added above the backbone network. This classification head uses the final hidden state labeled [CLS] by BERT as the semantic representation of the entire input sequence. The structure of the classification head is as follows: first, a dropout layer is connected with a dropout probability of 0.1; then, a fully connected linear layer is connected to map the 768-dimensional hidden state to a 2-dimensional output space, with the 2 dimensions corresponding to the two regression targets of stability requirement strength and resilience requirement strength, respectively. The parameters of the classification head are learned from scratch during fine-tuning, while all parameters of the pre-trained BERT backbone network participate in the fine-tuning and updating. The initial values ​​of all parameters of the backbone network can be directly loaded with the weight values ​​obtained by pre-training bert-base-chinese on a large general Chinese corpus. The entire model performs inference in a forward propagation manner: after the input sentence is segmented and labeled with [CLS] and [SEP], it passes through a 12-layer Transformer encoder, and the final output vector (768 dimensions) at the [CLS] position is taken. After passing through the dropout layer, it is input into the linear layer, and the output is a 2-dimensional real vector as the prediction result. The training process of the BERT semantic classification model is as follows: a training dataset is constructed. The training dataset comes from the psychological demand annotation results of a large number of job description texts by human resources experts. Experts annotate each job description text with two continuous values: stability demand intensity and stress resistance demand intensity, which can take values ​​in the range of [0,1] and [0,10], respectively. During the data preprocessing stage, the matching terminal can use a word segmenter compatible with BERT-base-chinese to segment the job description text, truncating or padding each sentence to a fixed maximum length, typically 512 tokens, and adding [CLS] and [SEP] tags to each input sequence. During training, a supervised regression fine-tuning approach can be used: the preprocessed text sequence is input into the BERT semantic classification model, the final hidden state at the [CLS] position is taken as the full sentence representation, mapped to a 2D output vector by the classification head, and the mean squared error (MSE) loss is calculated by combining it with the labeled stability requirement strength and resilience requirement strength. The sum of the two losses is the total loss.The optimizer can be AdamW, with a learning rate of 2×10. −5 A linear learning rate scheduler combined with a warm-up strategy is employed, with warm-up steps accounting for 6% to 10% of the total training steps. The number of training epochs is set to 3 to 5. Mini-batch stochastic gradient descent can be used during training, with the batch size set to 16 or 32 depending on the available GPU memory. Training terminates when the validation set loss no longer decreases for several consecutive epochs (e.g., 3 epochs). After training, a pre-tuned BERT semantic classification model is obtained.

[0065] Optionally, the mechanism of mini-batch stochastic gradient descent refers to the matching terminal randomly selecting a fixed number of samples (i.e., batch size) from the training dataset in each iteration to form a mini-batch. All samples in the mini-batch are simultaneously input into the model for forward propagation. The mean squared error loss between the predicted and labeled values ​​of each sample is calculated. The average loss of all samples within the batch is taken as the overall loss of that batch. The matching terminal can calculate the gradient of this overall loss with respect to each trainable parameter in the model using the backpropagation algorithm, and use the average of these gradients as an unbiased estimate of the current batch's gradient with respect to the true gradient, which guides the direction of parameter updates. AdamW is Adam (Adaptive Moment of Gradient Descent). AdamW is an improved version of adaptive moment estimation (AMS), whose core lies in decoupling weight decay and L2 regularization in the loss function. During each parameter update, AdamW calculates the gradient of the current batch loss with respect to each parameter, updating both the first-order momentum (exponential moving average of the gradient) and the second-order momentum (exponential moving average of the gradient) separately, and correcting for biases to eliminate the influence of initial estimation biases. The corrected first-order momentum is divided by the sum of the square root of the corrected second-order momentum and a preset small constant to obtain the update increment direction and magnitude for that parameter. Finally, the product of a preset weight decay coefficient and the current value of the parameter is subtracted from this increment direction, resulting in the decoupled weight decay term. The final parameter update is applied to the current parameters; the preset small constant refers to a fixed, small constant given in advance; the warm-up strategy refers to the learning rate gradually increasing linearly from a near-zero minimum value to a preset maximum learning rate within a specific number of steps in the early stages of training; the linear learning rate scheduler is a strategy used to dynamically adjust the learning rate during deep learning model training. Specifically, it linearly interpolates the current learning rate from the maximum value reached at the end of the warm-up phase, according to the remaining total training steps, gradually decaying it to a final value close to zero. After each round of parameter updates, the matching terminal calculates the decay coefficient based on the ratio of the currently executed update steps to the total number of steps, in the form of... Multiply the current learning rate by the decay coefficient to obtain the actual learning rate used in this step.

[0066] S104. If the psychological fit is greater than the psychological screening threshold, the candidate's career trajectory data in the historical behavior record is used to identify the candidate's career value preference and obtain the candidate's career value preference, which includes the stability preference quantification value and the growth preference quantification value.

[0067] Optionally, the psychological screening threshold is a pre-set numerical constant in the matching terminal, used to determine whether the candidate matches the psychological requirements of the target position in terms of psychological traits. It is dimensionless, with a value range of [0,1], consistent with the value range of psychological fit. It can be pre-configured according to the sensitivity of the target position to psychological fit: for positions with high pressure or high collaboration requirements, the psychological screening threshold can be set appropriately higher to ensure that the candidate and the position are highly consistent in terms of psychology; for positions with more lenient psychological requirements, the psychological screening threshold can be appropriately lowered. The identification of career value preferences refers to the complete process after the candidate passes the psychological screening, based on the career trajectory data contained in their historical behavior records, through a series of quantitative statistical calculations and feature extraction operations, to summarize their value orientation in the two dimensions of stability and growth from the candidate's past career experience, and output it in the form of quantitative values.

[0068] For example, if the psychological fit is greater than the psychological screening threshold, the matching terminal can identify the candidate's career value preferences based on the candidate's career trajectory data in the historical behavior record, and obtain the candidate's career value preferences, which include the quantitative values ​​of stability preference and growth preference.

[0069] S105 calculates the candidate's career motivation score based on career value preferences and psychological fit, and then weights and synthesizes the skill matching degree, psychological fit, and career motivation score to obtain the matching accuracy between the candidate and the target position.

[0070] Optionally, weighted synthesis refers to the complete process in which the matching terminal inputs the three dimensions of skill matching degree, psychological fit degree and career motivation value into a pre-trained matching degree fusion model, and the matching degree fusion model performs forward inference calculations and outputs the comprehensive matching accuracy.

[0071] For example, the matching terminal can determine the value of the adjustment parameter based on the nature of the target position. For positions with higher stability requirements, the adjustment parameter takes a larger value; for positions with higher growth requirements, the adjustment parameter takes a smaller value. The matching terminal can then perform a weighted summation of the stability preference quantification value and the growth preference quantification value according to the adjustment parameter. That is, multiply the stability preference quantification value by the adjustment parameter, multiply the growth preference quantification value by 1 and subtract the adjustment parameter, and add the two to obtain the quantification value of the motivation dimension, which represents the candidate's willingness to continuously invest in their career value. The matching terminal can then perform a weighted summation of the quantification value of this motivation dimension and the psychological fit to obtain the career motivation value. The weighted summation formula is as follows: ,in, This is a preset fusion weight coefficient, such as 0.4, which can be adjusted and determined based on algorithm performance and business objectives. For the quantitative value of the dynamic dimension, For psychological compatibility, The career motivation score integrates information on the candidate's intrinsic career value preferences and psychological traits—specifically, the degree of match between these traits and the psychological requirements of the position—to reflect the candidate's long-term commitment to the role. The matching terminal can combine skill matching, psychological fit, and career motivation score to obtain a three-dimensional input vector. This vector is then fed into a pre-trained matching accuracy model for forward inference calculations, yielding a comprehensive assessment of the candidate's fit with the target position.

[0072] Optionally, weighted summation refers to the mathematical operation of linearly combining the quantified values ​​of the motivation dimension and the psychological fit when the matching terminal calculates the career motivation value. The matching degree fusion model is a regression model based on a fully connected neural network architecture, employing a three-layer fully connected feedforward neural network. The input layer has three neurons, used to receive skill matching degree, psychological fit, and career motivation value, respectively. The hidden layer has 16 neurons, and each hidden layer neuron has a weighted connection with all three neurons in the input layer. That is, the input received by each hidden layer neuron is the sum of the three input values ​​multiplied by their respective connection weights, plus the neuron's own bias term, to obtain a linear weighted sum. The linear weighted sum is then input to the ReLU activation function for nonlinear transformation. The ReLU function is defined as... When the weighted sum is positive, the value is output directly; when the weighted sum is negative or zero, zero is output. The output layer has one neuron, which is also connected to all 16 neurons in the hidden layer with weighted full connections. Its input is the sum of the output values ​​of the 16 hidden layer neurons multiplied by their respective connection weights, plus a bias term. This linearly weighted sum is then fed into the Sigmoid activation function, which is defined as... This is used to weight and compress the output layer to the [0,1] interval. The training process of the matching degree fusion model is as follows: The matching terminal collects sample data of recruitment decisions that have been completed in historical recruitment as the training set. Each sample includes a three-dimensional input vector consisting of skill matching degree, psychological fit degree, and career motivation value, and a corresponding label value. The label value is determined based on whether the candidate's actual performance after joining the company meets or exceeds expectations. If the candidate performs well in the performance evaluation after joining the company and stays for more than the preset observation period, the label value is set to 1; if the candidate's performance is substandard or leaves the company early, the label value is set to 0. After the training set is constructed, the matching terminal can use the mini-batch stochastic gradient descent algorithm and the Adam optimizer, using binary cross-entropy as the loss function to iteratively optimize the network parameters so that the matching accuracy of the model output is as close as possible to the true label value. After sufficient iteration until the loss function converges, all weights and bias parameters inside the model are fixed, and the pre-trained matching degree fusion model is obtained.

[0073] Optionally, the preset observation period is a fixed time window parameter used by the matching terminal to define positive samples (label value 1) and negative samples (label value 0) when constructing the training dataset for the matching degree fusion model. For example, 12 months can be set according to factors such as the general pattern of personnel turnover in the industry and the average tenure of each position. The Adam optimizer is a parameter update algorithm. The core technology of this optimizer is to independently maintain the first-order momentum and second-order momentum for each trainable parameter in the model, and adaptively adjust the update step size of each parameter based on these two momentum. Forward inference refers to the data being calculated and passed from the input layer to the output layer along the neural network, only performing prediction calculations. In the training process of the matching degree fusion model, for each training sample, the model outputs a matching accuracy prediction value with a value range of [0,1] after forward inference. The true label value y of the sample is either 0 or 1. The binary cross-entropy loss function takes the predicted value and the true label value as input and is defined as follows: .

[0074] S106. If the matching accuracy is greater than the recruitment reference threshold, the matching accuracy and career motivation value are input into the pre-trained retention prediction model, and the long-term retention rate of the candidate in the target position is output.

[0075] Optionally, the recruitment reference threshold is a pre-set numerical constant in the matching terminal, used to determine whether, after skills screening, psychological screening, and comprehensive matching accuracy calculation, the candidate meets the overall qualification standards for inclusion in the subsequent retention evaluation stage. It is dimensionless, with a value range of [0,1], consistent with the value range of the matching accuracy. It can be pre-configured according to the organization's overall recruitment strategy, the strategic importance of the position, and the organization's strictness in talent selection: for key core positions or positions with low recruitment tolerance, the recruitment reference threshold can be set higher to select the candidates with the highest comprehensive matching degree for retention evaluation; for general positions recruited in batches, the recruitment reference threshold can be appropriately lowered.

[0076] For example, the matching terminal can compare the matching accuracy with the hiring reference threshold. If the matching accuracy is less than the hiring reference threshold, the candidate is determined not to meet the hiring reference criteria, the subsequent processing is terminated, and a "not recommended" conclusion is output. If the matching accuracy is greater than or equal to the hiring reference threshold, the matching terminal can combine the current candidate's matching accuracy and career motivation value to obtain a candidate covariate vector. This candidate covariate vector is then input into a trained Cox proportional hazards regression model to obtain an estimate of the candidate's survival function in the target position. The value of this survival function at any time point t represents the probability that the candidate will continue working in the position for more than t months. The matching terminal can calculate the specific value of the survival function at time T based on a preset retention observation period T, and use this value as the long-term retention rate of the candidate who has worked in the target position for a full T months.

[0077] Optionally, the preset retention observation period T is a fixed time length parameter used by the matching terminal in the Cox proportional hazards regression model inference stage. It is measured in months and is used to represent the period during which the candidate, which the organization is concerned about, will remain in the target position. The physical meaning of T is that in the recruitment decision-making scenario, the organization is usually not concerned about whether the candidate will never leave, but whether the candidate can remain stable in the position within a critical time window. After this time window, the candidate has contributed enough value to the organization, and even if the candidate leaves later, the turnover cost has been reasonably amortized.

[0078] This embodiment provides a three-dimensional career matching method based on ability, psychology, and motivation. It calculates the skill matching degree between the candidate and the target position based on the candidate's resume and the job requirements. If the skill matching degree is greater than a skill screening threshold, it obtains psychological trait vectors related to stability and resilience based on the candidate's historical behavioral records. These vectors, combined with the job psychological needs vector obtained from the job requirements, are used to calculate the candidate's psychological fit. If the psychological fit is greater than a psychological screening threshold, it identifies the candidate's career value preferences based on their career trajectory data. Combined with the psychological fit, it calculates the candidate's career motivation value. The skill matching degree, psychological fit, and career motivation value are weighted and synthesized to obtain the matching accuracy between the candidate and the target position. If the matching accuracy is greater than a recruitment reference threshold, the matching accuracy and career motivation value are input into a pre-trained retention prediction model to output the candidate's long-term retention rate in the target position. It can expand the assessment dimensions from a single ability dimension to three dimensions: ability, psychological fit, and career motivation. This effectively solves the core problems of traditional job matching methods, such as the single matching dimension, neglect of psychological fit, and inability to predict retention risks. It significantly improves the accuracy of job matching and the predictability of long-term employee retention.

[0079] In one embodiment, based on the resume text and the job requirement text, the skill match degree between the candidate and the target position is calculated, including:

[0080] S201 performs structured parsing of the resume text and job requirement text to obtain candidate skill terms and job candidate skill terms.

[0081] Optionally, structured parsing refers to breaking down the resume text and job requirement text separately, extracting and organizing the scattered natural language into a standardized, terminological skills list, and generating candidate skills terms and job candidate skills terms. Candidate skills terms refer to the personal skills extracted from the job seeker's resume; job candidate skills terms refer to the skills required for the job extracted from the hard / soft requirements of the recruitment position.

[0082] For example, the matching terminal can segment the candidate's resume text and the target job requirements text into several independent short sentence units according to punctuation marks such as periods, semicolons, and line breaks. Each short sentence carries relatively complete semantic information. The matching terminal can use a bidirectional maximum matching algorithm to perform Chinese word segmentation processing on each short sentence unit. This bidirectional maximum matching algorithm simultaneously performs maximum length matching from the front and back of the string to be processed, and takes the result with less ambiguity as the final word segmentation. It also combines a general dictionary and a domain-defined dictionary to improve the word segmentation accuracy. After word segmentation, the matching terminal can perform part-of-speech tagging on each word segmentation result, and select words with parts of speech such as nouns, verbs, and gerunds as candidate skill entries. Among them, nouns usually refer to entity-type skills such as tools and technology stacks, such as Java; verbs usually refer to operational skills such as development; and gerunds refer to composite skills such as data analysis. After the above processing, a set of candidate skill entries from the resume text and a set of candidate skill entries from the job requirements text are obtained, both of which exist in the form of basic units at the word level.

[0083] Optionally, the bidirectional maximum matching algorithm is a combination of forward maximum matching and reverse maximum matching, used to solve the problems of inaccurate word segmentation and ambiguous segmentation in single matching; the domain-specific dictionary includes professional terms, tool names, technical frameworks and other words commonly found in major industries such as information technology, finance, manufacturing and medical care, to ensure that these professional terms that may be missing in the general dictionary can be correctly identified rather than incorrectly segmented.

[0084] S202 employs a feature extraction algorithm based on a pre-trained language model to map each candidate skill term and job candidate skill term to a standard skill node in the standard skill knowledge graph, thereby obtaining standard skill terms and job requirement skill terms. The standard skill terms and job requirement skill terms are then summarized to obtain the candidate's explicit skill set and the job skill requirement set.

[0085] Optionally, the feature extraction algorithm based on the pre-trained language model is essentially a complete processing flow that uses the pre-trained language model (i.e., BERT-base-Chinese) as an encoding tool to convert candidate skill terms at the vocabulary level into vector representations that incorporate contextual semantics, and compares them with the standard skill knowledge graph to achieve standardized mapping. The standard skill knowledge graph is a large-scale structured resource library with skill nodes as basic units, covering the skill terminology system of major industries such as information technology, finance, manufacturing, medical care, and education. Each skill node is uniquely identified by a standard name and is accompanied by a semantic vector representation corresponding to that standard name. This semantic vector is pre-encoded during the knowledge graph construction stage by inputting the standard skill name into the BERT-base-Chinese model and taking the final hidden state of its [CLS] position.

[0086] For example, for each candidate skill term received by the input layer, the matching terminal can locate the context position of the candidate skill term in the original short sentence, extract the complete sentence including the candidate skill term, input the complete sentence into the BERT-base-Chinese pre-trained language model, and obtain the semantic vector representation of the candidate skill term in the current context. The matching terminal can calculate the cosine similarity between the context semantic vector and the semantic vector of each standard skill node in the standard skill knowledge graph. The closer the cosine similarity value is to 1, the more consistent the directions of the two vectors are in the semantic space. The matching terminal can select the standard skill node with the highest similarity among all standard skill nodes with similarity exceeding a preset threshold, and determine the standard name of the standard skill node as the standardized mapping result of the candidate skill term. After completing the above mapping for all candidate skill terms, the matching terminal can summarize the mapped standard skill terms into a set; for terms that cannot be mapped to any standard skill node, the matching terminal can determine them as non-skill terms and remove them. For candidate skill terms derived from resume text, after processing, a set of explicit skills for candidates is obtained, reflecting the tools and knowledge skills they have mastered. For candidate skill terms derived from job requirements text, the matching terminal can additionally introduce a rule-based requirement intensity discriminator to retain only skill terms that appear after strong constraint phrases such as "must," "required," and "proficient," while removing skill terms that appear after weak constraint phrases such as "priority," "bonus points," and "understand," thereby obtaining a set of job skill requirements that only includes skills explicitly listed as mandatory in the job requirements.

[0087] Optionally, the preset threshold is a numerical threshold used by the matching terminal to determine whether a candidate skill term can be successfully mapped to a certain standard skill node in the standard skill knowledge graph. It is dimensionless and ranges from [0,1], consistent with the range of cosine similarity. The preset threshold is usually determined based on statistical analysis of the semantic similarity distribution between standard nodes in the standard skill knowledge graph, with the goal of maximizing mapping recall and precision to select the optimal threshold value. The BERT-base-Chinese pre-trained language model consists of 12 stacked Transformer encoder layers. Each encoder layer includes a multi-head self-attention sublayer and a feedforward neural network sublayer. The multi-head self-attention sublayer has 12 heads, and each head has a dimension of 64. That is, the query, key, and value projection dimensions of a single self-attention head are all 64. After the 12 heads are computed in parallel, they are concatenated to obtain a 768-dimensional output. The feedforward neural network sublayers consist of two linear transformations. The intermediate hidden layer has a dimension of 3072, meaning the 768-dimensional input is first mapped to 3072 dimensions, activated by a Gaussian Error Linear Unit (GELU), and then mapped back to 768 dimensions. Each sublayer is followed by residual connections and layer normalization to ensure stable gradient propagation in deep networks. The model input uses WordPiece word segmentation with a vocabulary size of 21,128 tokens, and each input sequence contains at most 512 tokens. The total number of model parameters is approximately 110 million, including the word embedding matrix (approximately 21,128 × 768 ≈ 16.22 million parameters), multi-head self-attention parameters of each Transformer encoder layer (approximately 2.36 million parameters per layer), feedforward neural network parameters (approximately 4.72 million parameters per layer), and normalization parameters of each layer (approximately 1536 parameters per layer). The BERT-base-Chinese pre-trained language model employs a self-supervised learning approach during training. It simultaneously performs two pre-training tasks on a large-scale unlabeled Chinese corpus covering text sources such as Wikipedia, news, and books: The first task is masked language modeling: approximately 15% of the tokens in the input sequence are randomly masked, and the model needs to predict the original vocabulary of the masked tokens based on the unmasked context. This task allows the model to learn the bidirectional dependency relationship between words and their context. The second task is next-sentence prediction: two concatenated text segments are input into the model, and the model needs to determine whether the second segment naturally follows the first. This task allows the model to learn sentence-level semantic coherence and discourse structure. During training, the Adam optimizer is used, with the learning rate decreasing linearly after warm-up, iterating sufficiently until the loss converges.After training, the model's parameters are frozen. The model's function is essentially a general semantic encoder, used to map natural language sentences of arbitrary length into dense vector representations of fixed dimensions, so that semantically similar sentences or words are close to each other in the vector space, while semantically dissimilar sentences are far apart in the vector space.

[0088] Optionally, WordPiece segmentation is a sub-word segmentation algorithm. Its basic idea is to divide a word into several frequently occurring sub-word units, rather than directly using complete words or single characters as the basic units; GELU is a non-linear activation function, mathematically defined as: Where Φ(x) represents the cumulative distribution function of the standard normal distribution, i.e. The physical meaning of GELU is to perform probabilistic gating on the input x according to its value relative to zero. That is, when x is much greater than zero, Φ(x) approaches 1, the function output approaches x, and the pass rate is close to 100%; when x is much less than zero, Φ(x) approaches 0, the function output approaches 0, and the pass rate is close to 0%; when x is near zero, the function scales the input with a probability value between 0 and 1. Layer normalization is a normalization technique that calculates the mean and variance on the feature dimension for each input sample, rather than on the batch dimension. Random masking is a randomization operation performed on the input text in the masking language modeling task during the BERT pre-training stage. Its purpose is to construct the target that the pre-trained model needs to predict. In each training iteration, the model randomly selects about 15% of the tokens in the current input sequence; the selected tokens are the masked targets. For each selected token, the random masking operation is handled in one of three ways: with an 80% probability, the token is replaced with a special [MASK] marker, indicating that the model needs to predict the original word masked at this location based on the context; with a 10% probability, the token is replaced with a randomly selected word, allowing the model to learn that any word may appear even if the token is not masked; or with a 10% probability, the token remains unchanged, requiring the model to still make a prediction at this location, thus learning that noise in the input sequence will not affect the prediction of a specific location. Masked language modeling is one of the core self-supervised training tasks used by the BERT-base-Chinese pre-trained language model. The goal of this task is to enable the model to predict the original word at the masked location based on the unmasked context words in a sentence, thereby learning the bidirectional dependencies between words.

[0089] S203 maps the explicit skill set and skill requirement set to a unified skill tag vector space to obtain the candidate skill vector and the job skill requirement vector.

[0090] Optionally, the dimension of the skill tag vector space is equal to the total number of all standard skill nodes in the standard skill knowledge graph, that is, each dimension in the space uniquely corresponds to a standard skill node in the knowledge graph.

[0091] For example, the matching terminal can map the explicit skill set and the skill requirement set to the same unified skill tag vector space. In the skill tag vector space, the vector representation of any skill term adopts a high-dimensional sparse one-hot encoding method. Specifically, for any dimension d in the skill tag vector space, if the standard skill node corresponding to the dimension is exactly the standard name of the current skill term, the dimension is set to 1, and all other dimensions are set to 0. The matching terminal can generate a one-hot encoded vector for each skill term in the explicit skill set, and accumulate these one-hot encoded vectors dimension by dimension to obtain the candidate's skill vector. Each dimension of the skill vector indicates whether the candidate possesses the standard skill corresponding to that dimension. The matching terminal can generate a one-hot encoded vector for each skill term in the job skill requirement set and accumulate it dimension by dimension to obtain the job skill requirement vector. Each dimension of this vector indicates whether the standard skill is a mandatory skill requirement for the job. Since both sets are mapped to the same vector space, the candidate's skill vector and the job skill requirement vector are completely aligned in dimensional structure, that is, the first dimension corresponds to the same standard skill node, and the second dimension corresponds to the same standard skill node.

[0092] S204. Calculate the weighted Jaccard similarity between the candidate's skill vector and the job skill requirement vector in the skill label vector space, and determine the weighted Jaccard similarity as the skill matching degree.

[0093] For example, the matching terminal can calculate the weighted Jaccard similarity between the candidate's skill vector and the job skill requirement vector in the skill tag vector space, and determine the value of the weighted Jaccard similarity as the skill matching degree between the candidate and the target job. In the vector space, the standard skills corresponding to the non-zero dimensions constitute the candidate's explicit skill set and the job's skill requirement set. Optionally, the weighted Jaccard similarity is calculated as follows: the matching terminal traverses all dimensions of the candidate's skill vector and the job skill requirement vector respectively, records the position index of all non-zero dimensions (i.e., dimensions with a value of 1), and determines the intersection of the explicit skill set and the job skill requirement set based on the position index of the non-zero dimensions, i.e., all standard skills that the candidate possesses and the job requires; at the same time, it determines the union of the two, i.e., all standard skills that the candidate possesses or the job requires. For each standard skill in the intersection, the matching terminal can read the weight coefficient of that standard skill from the standard skill knowledge graph. This weight coefficient is calculated based on the frequency of the skill's appearance as a mandatory requirement for the position in historical recruitment data. Specifically, it is the logarithm of the reciprocal of the frequency of appearance, giving higher weights to skills that are rare in historical data and lower weights to common skills, thus avoiding the dilution effect of common skills on similarity calculation. The matching terminal can read the weight coefficient corresponding to each standard skill in the intersection and calculate the sum of the weight coefficients of all standard skills in the intersection. Then, it can calculate the sum of the weight coefficients of all standard skills in the intersection and divide the former by the latter. The resulting ratio is the weighted Jaccard similarity value, ranging from [0,1]. This value objectively reflects the degree of matching between the candidate and the target position in terms of skills. The higher the ratio, the greater the overlap between the candidate's explicit skill set and the job's skill requirement set, and the higher the weight of the overlapping skills in terms of scarcity and importance.

[0094] In one embodiment, historical behavior records include project participation records and collaboration communication records. If the skill matching degree is greater than the skill screening threshold, stability and stress resistance analysis is performed on the candidate based on their historical behavior records to obtain the candidate's psychological trait vector, including:

[0095] S301. If the skill matching degree is greater than the skill screening threshold, obtain the project start and end time, project role and final project delivery result of each project from the project participation record of historical behavior records.

[0096] Optionally, project participation records are structured behavioral data left by candidates when participating in various projects in their past professional experience. The sources include project files in the enterprise human resources integrated management system, project tracking records in the task management platform, and project summary sections in performance evaluation documents. Each project participation record is based on a single project and includes multiple structured fields. The enterprise human resources integrated management system is a technological achievement that integrates human resources management processes through intelligent technology, including core modules such as recruitment management, employment management, and payroll settlement, and has functions such as intelligent talent matching, electronic contract management, and risk warning.

[0097] For example, if the skill matching degree is greater than the skill screening threshold, the matching terminal can obtain the project participation record subclass from the candidate's historical behavior records, and extract three core fields from the project participation records: project start and end time, project role, and project final delivery result. The matching terminal can arrange all project records in chronological order. Among them, the project start and end time is the specific date from when the candidate started participating in a project to when they ended participating in the project, stored in year-month-day format; the project role is the candidate's role in the project, and possible values ​​include project leader, core developer, and ordinary participant; the project final delivery result is the completion status of the project, and possible values ​​include on-time delivery, delayed delivery, and early delivery.

[0098] S302. Based on the start and end times of each project, the project role, and the final deliverables, conduct a stability analysis on the candidates to obtain a stability feature vector. The stability feature vector includes the average project tenure, the standard deviation of the project tenure, the project role promotion index, and the project completion rate.

[0099] For example, the matching terminal can calculate the duration of service for each project based on the start and end times of each project, which is the time difference between the project end time and the project start time, measured in months. The matching terminal can take the arithmetic mean of the durations of service for all projects to obtain the average project duration, reflecting the typical length of time a candidate spends in a project environment. The matching terminal can calculate the standard deviation of the durations of service for all projects to obtain the standard deviation of project duration, reflecting the dispersion of a candidate's duration of service in different project periods relative to the average. A smaller standard deviation indicates a high degree of consistency in the candidate's time spent in each project, while a larger standard deviation indicates significant fluctuations in the candidate's project participation time. The matching terminal can calculate a project role promotion index based on the project role in each project: mapping project roles to increasing levels in the order of ordinary participants, core developers, and project leaders; calculating whether the role level has increased for adjacent projects sorted by time; and counting the proportion of role level increases out of the total number of adjacent project pairs to reflect the candidate's growth trajectory in project roles. The matching terminal can calculate the project completion rate based on the final delivery results of each project. This rate represents the proportion of projects delivered on time or ahead of schedule out of the total number of projects, reflecting the candidate's reliability and sense of responsibility in the face of project constraints. The matching terminal can combine the average project tenure, standard deviation of project tenure, project role promotion index, and project completion rate in a fixed order to obtain a stability feature vector.

[0100] S303. From the collaborative communication records of historical behavior records, obtain the communication time, communication frequency and communication role of each communication record. The communication role includes the active communication role and the passive communication role.

[0101] Optionally, collaborative communication records are structured behavioral data generated when candidates interact with colleagues, superiors, subordinates, or external partners in their past professional environments. Sources include enterprise email system logs, instant messaging chat logs, meeting minutes from meeting management systems, and communication tracking data from collaborative office platforms. Each collaborative communication record is based on a single communication event or a communication aggregation within a unit of time, and includes multiple structured fields. An enterprise email system is a professional email sending and receiving collaboration platform built by the enterprise itself or hosted by a service provider, using the enterprise's own domain name (xxx@companydomain.com) as its account system, and designed for organizational office work. A meeting management system is a digital tool based on a software platform, designed to manage meeting activities throughout their entire lifecycle. A collaborative office platform is a digital tool integrating task management, document sharing, instant messaging, and other functions.

[0102] For example, the matching terminal can further extract a subclass of collaborative communication records from the candidate's historical behavior records, and extract three core fields from the collaborative communication records: communication time, communication frequency, and communication role. The matching terminal can arrange all collaborative communication records in chronological order and segment the communication records according to the project timeline, assigning each communication record to its corresponding time interval. Specifically, the communication time is the specific date and time of the communication event, stored in year-month-day-hour-minute format; the communication frequency is the number of communication events the candidate participated in within a unit of time (usually with weekly or monthly granularity); and the communication role is the candidate's level of initiative in the communication process, possibly taking the form of an active or passive communication role. An active communication role indicates that the candidate is the initiator or main promoter of the communication, while a passive communication role indicates that the candidate is the respondent or recipient of the information.

[0103] S304. Based on the proportion of communication times for each communication partner, the communication time, frequency, and role of each communication record, conduct stress resistance analysis on the candidate to obtain a stress resistance feature vector. The stress resistance feature vector includes the proportion of high-frequency communication periods, the proportion of proactive communication, and the diversity of communication partners.

[0104] For example, the matching terminal can first identify the candidate's key nodes in each project cycle, namely the two-week period before the project delivery deadline, and calculate the communication frequency per unit time within this period. The matching terminal can calculate the global average of the candidate's communication frequency per unit time across all project cycles, and divide the length of the period during which the communication frequency per unit time exceeds the global average by the total length of the key node period to obtain the percentage of high-frequency communication periods, which reflects the candidate's ability to maintain or increase communication activity when facing time pressure and task urgency. The matching terminal can also count the number of records in which the candidate's role is an initiator of communication, divide this number by the total number of communication records, and obtain the percentage of proactive communication, which reflects the candidate's tendency to proactively initiate information interaction in a collaborative environment. The matching terminal can count the total number of different communication partners involved in all communication records. For each communication partner, it calculates the proportion of communication frequency to the total number of communication frequency. Communication partner diversity is calculated using the Shannon information entropy method: multiplying the proportion of communication frequency for each communication partner by its natural logarithm, taking the negative value, and then summing the results. A higher Shannon information entropy value indicates a more even distribution of the candidate's communication among different partners, reflecting the breadth of the candidate's collaborative network and social adaptability. The matching terminal can combine the proportion of high-frequency communication periods, the proportion of proactive communication, and communication partner diversity in a fixed order to obtain a resilience feature vector.

[0105] S305 concatenates the stability feature vector and the resilience feature vector to obtain a multi-dimensional feature vector, and inputs the multi-dimensional feature vector into an ensemble learning model based on the XGBoost framework to output the stability tendency score and the resilience tendency score.

[0106] Optionally, the ensemble learning model based on the XGBoost framework is a machine learning model built around the XGBoost algorithm. It takes multi-dimensional feature vectors as input, performs ensemble inference through decision trees within the model, and outputs two values ​​in parallel: a stability tendency score and a resilience tendency score. XGBoost (Extreme Gradient Boosting) is an ensemble learning algorithm based on the gradient boosting framework. It uses classification and regression trees as base learners and iteratively builds multiple decision trees. In each iteration, the algorithm fits a new decision tree based on the residual direction between the current overall model's predictions from the previous round and the true labels, ensuring that the tree's predictions compensate for the shortcomings of the preceding models. The newly generated decision tree is weighted with a certain learning rate and added to the overall model, forming the starting point for the next iteration. After multiple iterations, the predictions of all decision trees are summed according to their respective weights to obtain the final output value. During training, the algorithm explicitly adds a regularization term to the loss function, penalizing the complexity of the decision trees, to control model complexity and effectively suppress the risk of overfitting. The ensemble learning model comprises two output branches that share the same input feature space but optimize independently. The stability output branch is implemented as a binary classification task: it uses whether a candidate leaves within a pre-defined observation period as a binary label, it iteratively constructs multiple decision trees, and finally outputs a probability value ranging from [0,1] as a stability tendency score. The closer the value is to 1, the higher the probability that the candidate exhibits a stable behavioral pattern. The resilience output branch is implemented as a regression task: it uses the candidate's stress coping dimension score from performance evaluations during employment as a continuous label, and through the same iterative tree construction process, it finally outputs a continuous value ranging from [0,10] as a resilience tendency score. The higher the value, the stronger the candidate's coping and adaptability under high-pressure situations. During the training phase, each branch independently optimizes its own decision tree set using its own labeled data. During the inference phase, it receives the same input feature vector and outputs two scores in parallel.

[0107] Optionally, the training process of the ensemble learning model is as follows: The matching terminal collects multi-dimensional feature vector data of employees who have joined the company in historical recruitment, along with their corresponding stability and resilience labels, to form a training dataset. The stability label is based on whether the employee left the company within a preset observation period: if the tenure exceeds the observation period, it is labeled as stable (value 1); otherwise, it is labeled as unstable (value 0). The resilience label uses the average stress coping dimension score in each performance evaluation during the employee's tenure as a continuous label, with a value range of [0, 10]. After the training dataset is constructed, the matching terminal can define separate loss functions for the two output branches: the stability branch uses logarithmic loss, and the resilience branch uses mean squared error loss. Using the total loss as the optimization objective, a gradient boosting algorithm is used to iteratively construct a decision tree. In each iteration, the algorithm calculates the residual between the current overall prediction and the true label, and then fits a new decision tree to reduce this residual. The iteration continues until the preset maximum number of trees is reached or the loss on the validation set no longer decreases. At this point, the model training is complete. The preset maximum number of trees refers to the maximum number of decision trees that can be built in the model configuration. After training, the entire decision tree structure and its parameters within the model are fixed and can be directly used in the inference phase to perform forward computation on the multi-dimensional feature vectors of new candidates, outputting stability tendency scores and stress resistance tendency scores. The parameters of the ensemble learning model include structural hyperparameters and tree parameters. Structural hyperparameters refer to the numerical configurations that control the overall complexity of the model, learning strategy, and training behavior, such as the learning rate, the maximum depth of the decision trees, and the L2 regularization coefficient. These are predetermined based on prior analysis of the distribution characteristics of historical data before training begins. Tree parameters include the splitting feature index and splitting threshold of each internal node, and the weight values ​​of each leaf node. These are determined based on the residual analysis between the current overall model and the predicted results of the training samples and the true labels.

[0108] For example, the matching terminal can arrange the four components of the stability feature vector in sequence first and the three components of the resilience feature vector in sequence last, merge them to generate a seven-dimensional multi-dimensional feature vector, and input the multi-dimensional feature vector into an ensemble learning model based on the XGBoost framework to output the stability tendency score and the resilience tendency score in parallel.

[0109] S306 combines the stability tendency score and the stress resistance tendency score to obtain the candidate's psychological trait vector.

[0110] For example, the matching terminal can combine stability tendency scores and stress resistance tendency scores in a fixed order, with stability first and stress resistance second, to obtain a psychological trait vector.

[0111] In one embodiment, career trajectory data includes reasons for job changes and length of service. If the psychological fit is greater than the psychological screening threshold, the candidate's career value preferences are identified based on the candidate's career trajectory data in historical behavior records, resulting in the candidate's career value preferences, including:

[0112] S401: Obtain the candidate's reasons for job switching and length of service from historical behavioral records. Reasons for job switching include proactively seeking development, proactively seeking salary increases, proactively seeking improvements in the work environment, being laid off, being eliminated due to performance issues, and being subject to organizational adjustments.

[0113] For example, if the psychological fit is greater than the psychological screening threshold, the matching terminal can extract career trajectory data, including two core fields: reason for job switching and duration of employment, from the candidate's historical behavior records. The reason for job switching refers to the candidate's actual motivation for each job change. The matching terminal can extract relevant textual descriptions such as resignation application documents, exit interview records, performance evaluation reports, and project handover summaries from the time boundary between the end of the previous employment period and the start of the next employment period by locating the dividing point of each job switching event in the historical behavior records. The matching terminal can categorize these text descriptions into one of several predetermined categories using a combination of keyword matching and semantic classification: "Actively Seeking Development" indicates a candidate voluntarily resigns due to a desire for career growth and promotion; "Actively Seeking Salary Increase" indicates a candidate voluntarily resigns due to salary and benefits; "Actively Seeking Improvement of Work Environment" indicates a candidate voluntarily resigns due to factors such as work environment, team atmosphere, or geographical location; "Involuntary Layoff" indicates a candidate is laid off due to organizational strategic adjustments or business contraction; "Involuntary Performance Elimination" indicates a candidate's employment contract is terminated due to substandard performance; and "Involuntary Organizational Adjustment" indicates a candidate is adjusted due to organizational reasons such as departmental restructuring or job elimination. Tenure refers to the continuous working time of a candidate in each position. The matching terminal can read the start and end dates of each employment experience from the career trajectory data and calculate the difference in months. The matching terminal can arrange all job change reasons and tenure data in chronological order to obtain a structured career trajectory sequence. Optionally, the method based on the combination of keyword matching and semantic classification refers to the matching terminal integrating both keyword matching and semantic classification to classify text. Keyword matching refers to capturing explicit words and phrases related to reasons for leaving the job in the text and performing coarse matching; semantic classification refers to understanding the meaning of the whole sentence rather than just looking at the literal words, making up for the shortcomings of simple keywords and avoiding misjudgment based on just words.

[0114] S402. Based on the length of service of each of the candidate's job experiences, the mean and standard deviation of the length of service are calculated. Based on the mean and standard deviation of the length of service, the reasons for job switching and the total number of job experiences, the candidate's stability preference value is calculated.

[0115] For example, the matching terminal can count the total number of job segments for a candidate, denoted as variable H, with the unit being "segment". For each job segment, the matching terminal can record the duration of the employment as... (h=1,2,…,H), unit is months. The matching terminal can calculate the arithmetic mean of all employment durations to obtain the average employment duration. The calculation formula is: ,in, The average length of service is expressed in months. Based on this average length of service, the standard deviation of all lengths of service is calculated using the following formula: ,in, The standard deviation of all employment durations is given in months. The matching terminal can calculate the coefficient of variation of employment duration based on the mean and standard deviation, using the following formula: ,in, The coefficient of variation (COP) is dimensionless and reflects the dispersion of a candidate's tenure across different periods relative to the average. A smaller COP indicates a high degree of consistency in the candidate's tenure across different periods, reflecting a strong preference for long-term employment. The matching terminal can iterate through all reasons for job changes and count the number of times they are categorized as passive. Passive categories include passive layoffs, passive performance-based layoffs, and passive organizational restructuring. The unit is "this" (time), and the stability preference quantification value is calculated based on the coefficient of variation and the number of times the passive category is used. The calculation formula is as follows: ,in, The value is a dimensionless quantification of stability preference, ranging from [0,1]. The closer the value is to 1, the stronger the candidate's preference for long-term stable employment. α and β are pre-set weighting coefficients used to allocate the contribution of the two feature indicators. Both are dimensionless constants and α+β=1.

[0116] S403. Based on the trend of job level changes in each of the candidate's job experiences, calculate the total job level increase from the first job to the last job, and calculate the candidate's growth preference quantification value based on the reason for job switching and the total job level increase.

[0117] For example, the matching terminal can establish a job level mapping table, mapping common job titles to incremental level values. For instance, assistant is mapped to 1, specialist to 2, supervisor to 3, manager to 4, director to 5, vice president to 6, and so on. The matching terminal can read the job title corresponding to each work experience from historical behavior records, and convert the job title into the corresponding level value by consulting the job level mapping table. Let the level of the h-th work experience be denoted as... The unit is level. The matching terminal can read the level of the first employment experience. and the level of the last job experience The total promotion in job grade is calculated using the following formula: The unit is level, and the total promotion in job level reflects the cumulative extent to which a candidate assumes higher-level responsibilities throughout their career trajectory. The matching terminal can iterate through all reasons for job changes and count the number of times they are categorized as "actively seeking development." The unit is "times," and a growth preference quantification value is calculated based on the total number of job level promotions and the number of times the individual proactively sought development opportunities. The calculation formula is as follows: ,in, This is a quantitative value for growth preference, which is dimensionless and ranges from [0,1] after normalization. The closer it is to 1, the stronger the candidate's pursuit of career advancement and ability improvement. and These are pre-defined weighting coefficients, both of which are dimensionless constants and γ+δ=1. During the calculation, if H=1, meaning the candidate has only one employment experience, If the denominator of an item is zero, the matching terminal can specially process the value of that item as 0.

[0118] S404 combines the stability preference value and the growth preference value to obtain the candidate's career value preference.

[0119] For example, the matching terminal can combine stability preference quantification values ​​and growth preference quantification values ​​in a fixed order of stability first and growth last to generate a career value preference vector.

[0120] In one embodiment, the pre-trained retention prediction model is obtained through the following steps:

[0121] S501: Obtain the historical matching accuracy and historical career motivation values ​​of all historical incumbents of the target position, and combine the historical matching accuracy and historical career motivation values ​​to obtain the covariate vector.

[0122] For example, the matching terminal can retrieve the personnel file data of all past employees in the target position's historical recruitment, and obtain the matching accuracy of each employee at the time of their employment from the personnel file data. and career motivation value and combine and Obtain the covariate vector of all incumbents. .

[0123] S502, obtain the actual departure time and event indicator variables of all historical incumbents of the target position.

[0124] For example, the matching terminal can continue to retrieve the departure information of each employee from the historical employee records of the target position, and obtain the actual departure time and event indicator variable for each historical employee from the departure information. The matching terminal can store the actual departure time t and event indicator variable δ for all historical employees according to their employee identity. The actual departure time refers to the specific time when the employee stopped working in the target position, expressed as the number of months between the start date and the departure date, denoted as variable t, in months. If the employee is still employed at the time of data collection, the actual departure time is recorded as the current length of employment, also in months. The event indicator variable is a binary labeled variable, denoted as δ, with a value of 0 or 1, dimensionless. The assignment rule for the event indicator variable is: if the employee has already left the position, i.e., the actual departure time corresponds to a real departure event, then δ=1; if the employee has not yet left the position at the time of data collection, then δ=0.

[0125] S503 constructs a training sample set based on the covariate vectors of all historical employees, the actual departure time, and the event indicator variables.

[0126] For example, the matching terminal can associate the covariate vector of each historical employee with the actual departure time and event indicator variable of the same employee, outputting a complete training sample. Each training sample is represented in the form of a triple, i.e. Where the subscript r represents the r-th historical incumbent, Let r be the two-dimensional covariate vector of the r-th incumbent. For the r-th employee, the actual time of departure is... Let be the event indicator variable for the r-th incumbent. The matching terminal can aggregate R triplet samples corresponding to all R historical incumbents of the target position to obtain a training sample set specific to the target position. .

[0127] S504 sets the Cox proportional hazards regression model as the algorithmic framework for the retention prediction model, resulting in the initial retention prediction model. The hazard function of the initial retention prediction model is:

[0128] ;

[0129] in, Let t represent the employee's turnover risk, and t represent time. As the benchmark risk function, Let the vector of regression coefficients be estimated. Let T be the covariate vector, and T be the transpose.

[0130] For example, the matching terminal can establish the Cox proportional hazards regression model as the algorithmic framework for retention prediction models. The Cox proportional hazards regression model is a semi-parametric survival analysis model. Its core assumption is that the risk of turnover for any employee at any point in time equals a time-dependent baseline risk function multiplied by an exponential function related to covariates. The matching terminal can set the risk function of the Cox proportional hazards regression model as follows: ,in, The instantaneous risk of an employee leaving the company at time t (in months); The baseline risk function represents the risk of turnover when all covariates take the value of zero, expressed in monthly units. The form of this function is unknown and no parametric distribution assumptions are made. Let be the vector of regression coefficients to be estimated, which is dimensionless. Each component is used to reflect the strength of the influence of the numerical change of the corresponding dimension in the covariate vector on the logarithm of turnover risk. The vector is a covariate vector, dimensionless; T is the vector transpose symbol, used to transpose a column vector. Perform an inner product operation with X in row vector form. The matching terminal can then perform an inner product operation with the model. and All are marked as unknowns. The model framework has been established but parameter estimation has not yet been completed; this is the initial retention prediction model.

[0131] S505, based on the training sample set, uses the partial likelihood estimation method to solve for the regression coefficient vector to be estimated in the initial retention prediction model, and obtains the regression coefficient vector.

[0132] Alternatively, the core idea of ​​partial likelihood estimation is to avoid involving the baseline risk function. In this case, based solely on the occurrence of events in the training data (i.e. The order information of the samples is used to construct a system that is only about... The partial likelihood function is obtained, and the partial likelihood function is estimated by maximizing this function. .

[0133] For example, the matching terminal can match all training samples according to the actual time of departure. Sort by size from smallest to largest and filter out all. The sample, i.e., the sample of those who have left the company, and the time for each sorted sample r of those who have left the company. Risk set at any moment This risk set includes all actual departure times greater than or equal to [the specified time]. The sample, i.e. All samples, regardless These samples are either 0 or 1. Each employee is constantly under observation and at risk of leaving the company. The contribution factor of the departed sample *r* to the partial likelihood function is calculated, and its form is: The numerator is the risk index of the sample itself, the denominator is the sum of the risk indices of all samples in the risk set, and the physical meaning of the contribution factor is... The probability of a resignation occurring at any given time. The partial likelihood function is obtained by multiplying the contribution factors of all resigned samples together. The matching terminal can be maximized using the Newton-Raphson method, which maximizes the logarithmic form of the partial likelihood function, i.e., by calculating the logarithmic partial likelihood function on the logarithmic side. The first derivative (gradient vector) and the second derivative (Hessian matrix), with The zero vector is used as the initial value for iteration, and is continuously updated. The value of is taken until the magnitude of the gradient vector is less than the convergence threshold, at which point... The value of is the estimate that makes the partial likelihood function reach its maximum value, denoted as . The two components of the regression coefficient vector are used to quantify the contribution direction and strength of historical matching accuracy and historical career motivation value to turnover risk, respectively. Optionally, the convergence threshold is a minimal positive real constant, dimensionless, used to determine whether the iterative process has reached sufficient convergence. Its physical meaning is the upper tolerance limit of the gradient vector magnitude, which can be determined through numerical optimization experiments on similar historical tasks. Numerical optimization experiments refer to the experimental process of conducting multiple sets of comparative simulations or offline trials based on sample datasets of similar historical tasks, traversing candidate threshold values, evaluating the algorithm output accuracy and iteration convergence steps corresponding to each threshold, and selecting the threshold that simultaneously meets the accuracy requirements and has an iteration step count within a reasonable range as the final engineering value.

[0134] S506, substitute the regression coefficient vector into the initial retention prediction model to obtain the trained retention prediction model.

[0135] For example, the matching terminal can use the regression coefficient vector Substituting this into the initial retention prediction model framework, the risk function is now updated to... The baseline risk function in the updated model The exact timeframe remains unknown; the matching terminal can use the Breslow estimation method to estimate this function. Optionally, the Breslow estimation method operates as follows: the matching terminal calculates the actual departure time of all samples in the training sample set. Sort in ascending order and filter all. The only time of leaving the company For each such time point, calculate the risk set for that time. The sum of risk indices of all samples Then the baseline cumulative risk increment at that moment The cumulative risk function is calculated by dividing the number of departures at that point in time by the sum of the risk indices. The number of departures is typically 1; if there are ties, the tie count is taken. The matching terminal can sum the baseline cumulative risk increments for all departure points to obtain the complete baseline cumulative risk function. Furthermore, the baseline survival function is derived. At this point, all unknown parameters within the model, i.e., the regression coefficient vector, have been determined. and baseline survival function All have been confirmed. The matching terminal can... and These are collectively packaged into a pre-trained retention prediction model specific to the target position. This retention prediction model can be used... During the inference phase, the retention probability at any point in time is calculated based on the covariate vector of the new candidate, enabling a quantitative prediction of the long-term retention rate of the new candidate.

[0136] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0137] Based on the same inventive concept, this application also provides a capability-based psychodynamic three-dimensional career matching device for implementing the capability-based psychodynamic three-dimensional career matching method described above. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations of one or more capability-based psychodynamic three-dimensional career matching device embodiments provided below can be found in the limitations of the capability-based psychodynamic three-dimensional career matching method described above, and will not be repeated here.

[0138] In one exemplary embodiment, such as Figure 3 As shown, a three-dimensional career matching device 300 based on ability, psychodynamics, and other factors is provided, including:

[0139] The skills matching degree calculation module 301 is used to obtain the resume text submitted by the candidate and the job requirements text of the target position, and calculate the skills matching degree between the candidate and the target position based on the resume text and the job requirements text.

[0140] The stability and stress resistance analysis module 302 is used to perform stability and stress resistance analysis on candidates based on their historical behavior records if the skill matching degree is greater than the skill screening threshold, and to obtain the candidate's psychological trait vector, which includes stability tendency score and stress resistance tendency score.

[0141] The psychological fit calculation module 303 is used to extract the psychological needs information of the job requirements text, obtain the psychological needs vector of the target job, and calculate the psychological fit of the candidate based on the psychological trait vector and the psychological needs vector.

[0142] The preference recognition module 304 is used to identify the candidate's career value preferences based on the candidate's career trajectory data in the historical behavior record if the psychological fit is greater than the psychological screening threshold. The career value preferences include the stability preference quantification value and the growth preference quantification value.

[0143] The matching accuracy acquisition module 305 is used to calculate the candidate's career motivation value based on career value preference and psychological fit, and to weight and synthesize the skill matching degree, psychological fit and career motivation value to obtain the matching accuracy between the candidate and the target position.

[0144] The retention rate prediction module 306 is used to input the matching accuracy and career motivation value into the pre-trained retention prediction model if the matching accuracy is greater than the recruitment reference threshold, and output the long-term retention rate of the candidate in the target position.

[0145] Furthermore, the skill matching calculation module 301 is also used for:

[0146] The resume text and job requirement text are structured and parsed to obtain candidate skill terms and job candidate skill terms;

[0147] A feature extraction algorithm based on a pre-trained language model is used to map each candidate skill term and job candidate skill term to the standard skill node in the standard skill knowledge graph to obtain standard skill terms and job requirement skill terms. The standard skill terms and job requirement skill terms are then summarized to obtain the candidate's explicit skill set and the job skill requirement set.

[0148] The explicit skill set and the skill requirement set are mapped to a unified skill tag vector space to obtain the candidate skill vector and the job skill requirement vector.

[0149] Calculate the weighted Jaccard similarity between the candidate's skill vector and the job skill requirement vector in the skill tag vector space, and determine the weighted Jaccard similarity as the skill matching degree.

[0150] Furthermore, the stability and compressive strength analysis module 302 is also used for:

[0151] If the skill matching degree is greater than the skill screening threshold, the project start and end time, project role and final project delivery result of each project are obtained from the project participation record of historical behavior records;

[0152] Based on the project start and end times, project roles, and final project deliverables, stability analysis is conducted on candidates to obtain stability feature vectors. These stability feature vectors include average project tenure, standard deviation of project tenure, project role promotion index, and project completion rate.

[0153] From the collaborative communication records of historical behavior records, obtain the communication time, communication frequency and communication role of each communication record. The communication role includes the active communication role and the passive communication role.

[0154] Based on the proportion of communication times for each communication partner, the communication time, frequency, and role of each communication record, stress resistance analysis is conducted on candidates to obtain stress resistance feature vectors. The stress resistance feature vectors include the proportion of high-frequency communication periods, the proportion of proactive communication, and the diversity of communication partners.

[0155] By concatenating the stability feature vector and the resilience feature vector, a multi-dimensional feature vector is obtained. This multi-dimensional feature vector is then input into an ensemble learning model based on the XGBoost framework, which outputs stability tendency score and resilience tendency score.

[0156] By combining the stability tendency score and the stress resistance tendency score, a psychological trait vector of the candidate is obtained.

[0157] Furthermore, the preference recognition module 304 is also used for:

[0158] The reasons for job switching and the length of service of candidates are obtained from their historical behavior records. The reasons for job switching include actively seeking development, actively seeking salary increase, actively seeking work environment improvement, being laid off, being eliminated due to performance, and being affected by organizational adjustment.

[0159] Based on the length of service of each position held by the candidate, the mean and standard deviation of the length of service are calculated. Based on the mean and standard deviation of the length of service, the reasons for job switching and the total number of positions held, the candidate's stability preference value is calculated.

[0160] Based on the trend of job level changes in each of the candidate's job experiences, the total job level increase from the first job to the last job is calculated, and the candidate's growth preference quantification value is calculated based on the reason for job switching and the total job level increase.

[0161] By combining the stability preference value and the growth preference value, the candidate's career value preference is obtained.

[0162] Furthermore, the device also includes a model training module, which is used for:

[0163] Obtain the historical match accuracy and historical career motivation values ​​of all past incumbents for the target position, and combine the historical match accuracy and historical career motivation values ​​to obtain a covariate vector;

[0164] Obtain the actual departure time and event indicator variables of all historical incumbents in the target position;

[0165] A training sample set is constructed based on the covariate vectors of all historical employees, the actual departure time, and the event indicator variables.

[0166] By setting the Cox proportional hazards regression model as the algorithmic framework for the retention prediction model, we obtain the initial retention prediction model, whose risk function is:

[0167] ;

[0168] in, Let t represent the employee's turnover risk, and t represent time. As the benchmark risk function, Let the vector of regression coefficients be estimated. Let T be the covariate vector, and T be the transpose.

[0169] Based on the training sample set, the partial likelihood estimation method is used to solve for the regression coefficient vector to be estimated in the initial retention prediction model, and the regression coefficient vector is obtained.

[0170] Substituting the regression coefficient vector into the initial retention prediction model yields the trained retention prediction model.

[0171] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the competence-psychodynamic three-dimensional career matching method as described above.

[0172] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0173] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0174] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A three-dimensional occupational matching method based on ability, psychodynamics, and other characteristics, The method includes: Obtain the resume text submitted by the candidate and the job requirements text of the target position, and calculate the skill matching degree between the candidate and the target position based on the resume text and the job requirements text; If the skill matching degree is greater than the skill screening threshold, the candidate's stability and stress resistance are analyzed based on the candidate's historical behavior records to obtain the candidate's psychological trait vector, which includes stability tendency score and stress resistance tendency score. The psychological needs information of the job requirements text is extracted to obtain the psychological needs vector of the target job, and the psychological fit of the candidate is calculated based on the psychological trait vector and the psychological needs vector. If the psychological fit is greater than the psychological screening threshold, based on the candidate's career trajectory data in the historical behavior record, the candidate's career value preference is identified to obtain the candidate's career value preference, which includes a stability preference quantification value and a growth preference quantification value. Based on the professional value preference and the psychological fit, the candidate's professional motivation value is calculated, and the skill matching degree, the psychological fit degree and the professional motivation value are weighted and synthesized to obtain the matching accuracy between the candidate and the target position. If the matching accuracy is greater than the recruitment reference threshold, the matching accuracy and the career motivation value are input into the pre-trained retention prediction model, and the long-term retention rate of the candidate in the target position is output.

2. The method according to claim 1, characterized in that, The calculation of the skill match between the candidate and the target position based on the resume text and the job requirement text includes: The resume text and the job requirement text are subjected to structured parsing to obtain candidate skill terms and job candidate skill terms; A feature extraction algorithm based on a pre-trained language model is used to map each candidate skill term and each candidate job skill term to a standard skill node in a standard skill knowledge graph, thereby obtaining standard skill terms and job requirement skill terms. The standard skill terms and job requirement skill terms are then summarized to obtain the candidate's explicit skill set and job skill requirement set. The explicit skill set and the skill requirement set are mapped to a unified skill tag vector space to obtain the candidate skill vector and the job skill requirement vector. Calculate the weighted Jaccard similarity between the candidate's skill vector and the job skill requirement vector in the skill tag vector space, and determine the weighted Jaccard similarity as the skill matching degree.

3. The method according to claim 2, characterized in that, The historical behavior records include project participation records and collaboration and communication records. If the skill matching degree is greater than the skill screening threshold, stability and stress resistance analysis is performed on the candidate based on the candidate's historical behavior records to obtain the candidate's psychological trait vector, including: If the skill matching degree is greater than the skill screening threshold, the project start and end time, project role and final delivery result of each project are obtained from the project participation record of the historical behavior record; Based on the start and end times of each project, the project role, and the final delivery results of each project, a stability analysis is performed on the candidate to obtain a stability feature vector. The stability feature vector includes the average project tenure, the standard deviation of project tenure, the project role promotion index, and the project completion rate. From the collaborative communication records in the historical behavior records, obtain the communication time, communication frequency and communication role of each communication record, where the communication role includes active communication role and passive communication role; Based on the proportion of communication times for each communication partner, the communication time, the communication frequency, and the communication role in each communication record, a stress resistance analysis is performed on the candidate to obtain a stress resistance feature vector. The stress resistance feature vector includes the proportion of high-frequency communication periods, the proportion of proactive communication, and the diversity of communication partners. The stability feature vector and the resilience feature vector are concatenated to obtain a multidimensional feature vector, which is then input into an ensemble learning model based on the XGBoost framework to output the stability tendency score and the resilience tendency score. The stability tendency score and the stress resistance tendency score are combined to obtain the psychological trait vector of the candidate.

4. The method according to claim 1, characterized in that, The career trajectory data includes reasons for job changes and duration of employment. If the psychological fit is greater than the psychological screening threshold, based on the candidate's career trajectory data in the historical behavior records, the candidate's career value preferences are identified to obtain the candidate's career value preferences, including: The reasons for job switching and the length of service of the candidates are obtained from the historical behavior records. The reasons for job switching include actively seeking development, actively seeking salary increase, actively seeking work environment improvement, being laid off, being eliminated due to performance, and being subject to organizational adjustment. Based on the duration of each job experience of the candidate, the mean duration of job experience and the standard deviation of job experience are calculated. Based on the mean duration of job experience, the standard deviation of job experience, the reason for job switching and the total number of job experience segments, the stability preference value of the candidate is calculated. Based on the job level change trend of the candidate's various job experiences, calculate the total job level increase from the first job to the last job, and calculate the candidate's growth preference quantification value based on the job switching reason and the total job level increase. By combining the stability preference value and the growth preference value, the candidate's career value preference is obtained.

5. The method according to claim 1, characterized in that, The pre-trained retention prediction model is obtained through the following steps: Obtain the historical matching accuracy and historical career motivation values ​​of all historical incumbents of the target position, and combine the historical matching accuracy and the historical career motivation values ​​to obtain a covariate vector; Obtain the actual departure time and event indicator variables of all historical incumbents in the target position; A training sample set is constructed based on the covariate vectors of all the historical employees, the actual departure time, and the event indicator variables. The Cox proportional hazards regression model is set as the algorithmic framework of the retention prediction model to obtain the initial retention prediction model, the risk function of which is: ； in, Let t represent the employee's turnover risk, and t represent time. As the benchmark risk function, Let the vector of regression coefficients be estimated. Let T be the covariate vector, and T be the transpose. Based on the training sample set, the partial likelihood estimation method is used to solve for the regression coefficient vector to be estimated in the initial retention prediction model, and the regression coefficient vector is obtained. Substituting the regression coefficient vector into the initial retention prediction model yields the trained retention prediction model.

6. A three-dimensional career matching device based on ability, psychodynamics, and characteristics, The device includes: The skills matching degree calculation module is used to obtain the resume text submitted by the candidate and the job requirements text of the target position, and calculate the skills matching degree between the candidate and the target position based on the resume text and the job requirements text. The stability and stress resistance analysis module is used to perform stability and stress resistance analysis on the candidate based on the candidate's historical behavior records if the skill matching degree is greater than the skill screening threshold, and to obtain the candidate's psychological trait vector, which includes stability tendency score and stress resistance tendency score. The psychological fit calculation module is used to extract the psychological needs information of the job requirements text, obtain the psychological needs vector of the target job, and calculate the psychological fit of the candidate based on the psychological trait vector and the psychological needs vector. The preference recognition module is used to identify the candidate's career value preference based on the candidate's career trajectory data in the historical behavior record if the psychological fit is greater than the psychological screening threshold, and to obtain the candidate's career value preference, which includes a stability preference quantification value and a growth preference quantification value. The matching accuracy acquisition module is used to calculate the candidate's career motivation value based on the career value preference and the psychological fit, and to weight and synthesize the skill matching degree, the psychological fit and the career motivation value to obtain the matching accuracy between the candidate and the target position; The retention rate prediction module is used to input the matching accuracy and the career motivation value into a pre-trained retention prediction model if the matching accuracy is greater than the recruitment reference threshold, and output the long-term retention rate of the candidate in the target position.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 5.

Images