Human resource information processing system and method

Abstract

The invention discloses a human resource information processing system and method, and relates to the technical field of human resource management information processing.A prototype measurement mechanism is introduced into an embedding space, a limited number of labeled samples are aggregated into stable category center representation, dependence on large-scale historical data is reduced, and the accuracy of human resource information processing is improved. Representative high-loss risk and low-loss risk category prototypes can also be formed in sample scarcity scenes such as medium and small enterprises, new business departments or new posts; characteristic embedding is carried out through cooperation of a pre-trained neural network model and transfer learning, the method effectively inherits the general characterization capability in large-scale human resource data, and then a prototype calculation module and two-norm-based normalization processing are superposed, so that feature vectors of different employees or candidates are more uniform in the numerical scale, and the accuracy of the human resource data is improved. The interference of the abnormal sample modulus length difference on the subsequent distance measurement is obviously weakened, so that the numerical stability during the similarity calculation in the high-dimensional embedding space is improved.

Description

Technical Field

[0001] This invention relates to the field of human resource management information processing technology, and in particular to a human resource information processing system and method. Background Technology

[0002] In modern enterprise recruitment management, human resource information processing systems mostly support job allocation and employment decisions. A key aspect is predicting candidate hiring outcomes and the risk of short-term turnover after onboarding. These systems, based on data analysis, process information such as candidates' historical behavioral characteristics, job suitability, and probationary period performance to enable companies to make more rational hiring decisions and identify high-risk candidates in advance. With the popularization of artificial intelligence technology, machine learning models are widely used to learn patterns from past recruitment and employment data, thereby predicting whether candidates are suitable for target positions and their stability after onboarding. This approach relies on large amounts of structured data, such as candidate resumes, interview evaluation records, probationary period assessment results, and work hours statistics, to achieve relatively accurate modeling and decision support. Ideally, these human resource information processing systems can assist human resource departments in optimizing recruitment processes and reducing operational costs caused by incorrect hiring or frequent turnover.

[0003] In recent years, predictive models based on deep learning and ensemble learning have been applied to recruitment and employment decision-making scenarios. These models use complex algorithms to process multidimensional features to improve the accuracy of candidate matching and turnover risk prediction. For example, random forests or gradient boosting trees are used to integrate multiple weak learners, or neural networks are used to model the nonlinear relationship between candidate characteristics and job requirements. However, in recruitment scenarios for SMEs, newly established business departments, or emerging positions, the limited amount of historical data available for model training poses challenges. Small sample data cannot cover diverse candidate characteristics and employment patterns, making model parameter estimation unstable and prone to overfitting. In addition, data sparsity may amplify the impact of noise, reduce the model's generalization ability in real recruitment environments, and make it difficult to identify unsuitable candidates or those with a high risk of turnover in a timely and reliable manner.

[0004] To address the issue of small sample sizes, existing solutions have explored techniques such as transfer learning and data augmentation in recruitment prediction and human resource analysis. Transfer learning methods utilize large-scale human resource or recruitment datasets from the source domain for pre-training, and then fine-tune them to suit the small-sample recruitment scenarios of the target company, thus mitigating the impact of insufficient data. Data augmentation techniques attempt to expand the training set by synthesizing minority candidate samples or introducing generative models. However, in practical applications, these methods are easily constrained by the distributional differences between the source domain and the target recruitment scenario, and the knowledge transfer effect is often not stable enough. At the same time, the authenticity and diversity of synthesized candidate data are difficult to guarantee, which may introduce potential biases and affect the reliability of the prediction results on candidate matching degree and turnover risk.

[0005] Therefore, how to construct a human resource information processing system and method that can make full use of limited data and has good generalization ability in small-sample recruitment scenarios, so as to improve the accuracy and reliability of recruitment decisions, has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] In view of the aforementioned existing problems, the present invention is proposed.

[0007] This invention provides a human resources information processing system and method to solve the problems of existing employee turnover prediction models being unstable and severely overfitting in small sample scenarios, making it difficult to support reliable recruitment and retention decisions.

[0008] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0009] In a first aspect, embodiments of the present invention provide a human resources information processing system, comprising:

[0010] The data acquisition module is configured to acquire characteristic data of employees or candidates from a human resources database for recruitment decisions, the characteristic data including at least one attribute related to employee status and / or risk of leaving the company.

[0011] The embedding module is configured to map the feature data to the embedding space to generate the corresponding embedding vector, and the mapping process is a nonlinear transformation.

[0012] The prototype calculation module is configured to calculate prototype vectors for multiple categories based on a support set, wherein the support set includes the embedding vectors and corresponding category labels, and the category labels at least represent high-churn-risk categories and low-churn-risk categories;

[0013] The prediction module is configured to output the category prediction result of the query sample based on the distance between the embedding vector of the query sample and multiple prototype vectors, wherein the distance is calculated based on a metric function;

[0014] The recruitment decision module is configured to generate corresponding recruitment decision information based on the predicted results of the categories, which is used to assist the human resources department in determining whether to hire the candidates or whether to take retention measures for existing employees.

[0015] As a preferred embodiment of the human resources information processing system of the present invention, the embedding module includes a pre-trained neural network model, which is initialized from a large-scale human resources dataset through transfer learning to improve the feature representation ability in small-sample recruitment scenarios.

[0016] In a preferred embodiment of the human resources information processing system of the present invention, the prototype calculation module is configured to perform embedding vector normalization processing on the embedding vectors in the support set, and then calculate the mean of the embedding vectors according to the category to generate each prototype vector.

[0017] The prototype calculation module is configured as follows:

[0018] Normalization based on the second norm is performed on each embedding vector in the support set to constrain the magnitude of the embedding vector within a preset range, thereby reducing the impact of scale differences between different samples on subsequent measurement results. After normalization, the embedding vectors are arithmetically averaged by category, and the resulting average vector is used as the prototype vector of each category, so that each prototype vector centrally represents the typical turnover risk characteristics of the corresponding employee or candidate in the embedding space.

[0019] As a preferred embodiment of the human resources information processing system of the present invention, it further includes an update module configured to fine-tune the embedding module using an incremental learning algorithm and update the support set and the prototype vector when newly labeled employee or candidate feature data is obtained, in order to adapt to changes in data distribution under small sample recruitment scenarios.

[0020] As a preferred embodiment of the human resources information processing system of the present invention, the prediction module calculates the distance between the embedding vector of the query sample and each of the prototype vectors through the distance calculation of the metric function, and determines the category prediction result based on the minimum distance principle. The metric function includes Euclidean distance or cosine similarity.

[0021] The prediction module is configured as follows:

[0022] After obtaining the embedding vector of the query sample, the Euclidean distance and / or cosine similarity between the embedding vector and the prototype vector of each category are measured, and a probability normalization function with a scaling coefficient is used to convert the distance or similarity corresponding to each category into a predicted probability distribution. The prediction module also outputs the category prediction result of the query sample according to the category with the smallest distance or the largest predicted probability, and forms recruitment decision information with risk score based on the result, which is used to distinguish between high-risk and low-risk categories and provide a quantitative basis for recruitment or retention strategies.

[0023] As a preferred embodiment of the human resources information processing system described in this invention, the pre-trained neural network model is a multilayer perceptron network or a convolutional neural network structure.

[0024] Secondly, the present invention provides a human resources information processing method, including,

[0025] Step S1: Obtain employee or candidate characteristic data from the human resources database for recruitment decisions, the characteristic data including at least one attribute related to employee status and / or turnover risk;

[0026] Step S2: Map the feature data into the embedding space through a nonlinear transformation to generate the corresponding embedding vector;

[0027] Step S3: Based on the support set containing the embedded vector and its category label, calculate prototype vectors for multiple categories, wherein the category label represents at least high churn risk categories and low churn risk categories;

[0028] Step S4: Based on the distance between the embedding vector of the query sample and the multiple prototype vectors, perform distance calculation using a metric function and output the category prediction result of the query sample;

[0029] Step S5: Generate corresponding recruitment decision information based on the category prediction results.

[0030] In a preferred embodiment of the human resources information processing method of the present invention, the mapping in step S2 is implemented by a pre-trained neural network model, which is initialized from a large-scale human resources dataset through transfer learning.

[0031] In a preferred embodiment of the human resources information processing method of the present invention, step S3 includes: normalizing the embedding vectors in the support set, and calculating the mean of the embedding vectors according to the category to generate each prototype vector.

[0032] As a preferred embodiment of the human resources information processing method of the present invention, it further includes: when newly labeled employee or candidate feature data is obtained, using an incremental learning algorithm to fine-tune the neural network model used to perform step S2, and updating the support set and the prototype vector to adapt to the data distribution changes in small sample recruitment scenarios.

[0033] The beneficial effects of this invention are as follows: The human resources information processing system and method provided by this invention constructs an end-to-end technical solution to address the small sample problem in recruitment decision-making and employee turnover risk prediction. By introducing a prototype measurement mechanism in the embedding space, a limited number of labeled samples are aggregated into stable category center representations, reducing the dependence on large-scale historical data. In sample-scarce scenarios such as small and medium-sized enterprises, new business departments, or new positions, representative high-turnover-risk and low-turnover-risk category prototypes can also be formed.

[0034] This invention utilizes a pre-trained neural network model combined with transfer learning for feature embedding. It effectively inherits the general representation capabilities of large-scale human resource data. Furthermore, by adding a prototype calculation module and L2-based normalization, the feature vectors of different employees or candidates become more uniform in numerical scale, significantly reducing the interference of abnormal sample modulus differences on subsequent distance measurements. This improves the numerical stability of similarity calculations in high-dimensional embedding spaces. Based on this, in the prediction stage, this invention uses Euclidean distance and / or cosine similarity to measure the relationship between query samples and prototypes of each category. A probability normalization function with a scaling coefficient is used to convert the distance or similarity into a risk probability distribution. This not only provides classifications for high and low churn risk but also generates continuous risk scores, allowing human resource departments to set different threshold strategies according to business needs, achieving refined management of candidate recruitment and employee retention. In addition, this invention introduces an incremental learning module to dynamically update the embedding model and category prototypes after acquiring new labeled samples. This enables the system to continuously adapt to changes in data distribution, avoiding the problem of rapid performance degradation in traditional one-time training models over time and with business changes. By integrating data acquisition, embedding mapping, prototype calculation, measurement prediction, and recruitment decision output, this invention can still provide stable, interpretable, and quantifiable talent loss risk assessment results even under small sample conditions, thereby reducing the costs caused by incorrect hiring and frequent turnover, and improving the overall efficiency and reliability of recruitment decisions and human resource management. Attached Figure Description

[0035] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation on the scope of this application.

[0036] Figure 1 This is a schematic diagram of the framework of the human resources information processing system in the embodiment.

[0037] Figure 2 This is a flowchart illustrating the human resources information processing method in the embodiment. Detailed Implementation

[0038] 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.

[0039] All terms used in this application (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0040] For example, the terms “first” and “second” used in this application are only used to distinguish and describe similar objects, to differentiate the first object from another object, and are not used to describe a specific order or sequence, nor should they be interpreted as indicating or implying relative importance.

[0041] This application proposes a human resources information processing system, combining... Figure 1 As shown, the system includes:

[0042] The data acquisition module is configured to acquire characteristic data of employees or candidates from the human resources database for recruitment decisions. The characteristic data includes at least one attribute related to employee status and / or risk of turnover.

[0043] The embedding module is configured to map feature data to the embedding space to generate corresponding embedding vectors. The mapping process is a non-linear transformation.

[0044] The prototype calculation module is configured to calculate prototype vectors for multiple categories based on a support set. The support set includes embedding vectors and corresponding category labels, and the category labels at least represent high-churn-risk categories and low-churn-risk categories.

[0045] The prediction module is configured to output the category prediction result of the query sample based on the distance between the embedding vector of the query sample and multiple prototype vectors, where the distance is calculated based on a metric function;

[0046] The recruitment decision module is configured to generate corresponding recruitment decision information based on category prediction results, which helps the human resources department determine whether to hire candidates or take retention measures for existing employees.

[0047] In this embodiment, the human resources database can be understood as a structured data set stored in the enterprise's existing human resources management system. It typically includes employee basic information, job information, performance records, attendance statistics, historical resignation records, and candidate resume analysis results. The data acquisition module retrieves these fields periodically or as needed as the original source of feature data through pre-configured data table field mapping or interface calls. Specifically, attributes related to employee status and turnover risk can include whether the employee is in a probationary period, length of employment, job level, performance appraisal results, leave and overtime statistics, historical job transfer records, and past history of frequent job-hopping. The feature dimensions can be optionally set to between dozens and hundreds in the engineering implementation to cover the main influencing factors while avoiding excessively high dimensions that would lead to excessively long training times. Furthermore, high and low turnover risk category labels can be automatically generated based on whether resignations have occurred within the historical observation window. For example, employees who voluntarily resign or are involuntarily dismissed during the statistical period are marked as high-risk, while employees who remain employed without warning records are marked as low-risk. The statistical period can be set from three months to one year according to the enterprise's business rhythm and adjusted by the human resources department based on experience. Optionally, in the candidate scenario, the system can map candidate features that are homologous to the aforementioned employee features to the same feature space, and indirectly infer the probability of a candidate entering the high-risk category based on the degree of similarity with historically high-risk employees, thereby forming recruitment decision information that includes hiring and retention recommendations. For cases where feature data contains missing fields, the data acquisition module can employ a simple missing value imputation strategy without adding other modules. For example, for numerical fields, the median or mean of employees in the same position can be used as a substitute; for categorical fields, the most frequently occurring value can be used as a substitute. This ensures that the embedding module operates stably while maintaining complete input dimensions. When the missing proportion exceeds a preset upper limit, the system can mark the employee or candidate as having insufficient data quality and provide only a conservative risk warning.

[0048] In one embodiment, the embedding module includes a pre-trained neural network model, which is initialized from a large-scale human resources dataset through transfer learning to improve the feature representation capability in small-sample recruitment scenarios.

[0049] In one embodiment, the prototype calculation module is configured to normalize the embedding vectors in the support set, and then calculate the mean of the embedding vectors by category to generate each prototype vector.

[0050] The prototype calculation module is configured as follows:

[0051] Normalization based on the L2 norm is performed on each embedding vector in the support set to constrain the magnitude of the embedding vector within a preset range, thereby reducing the impact of scale differences between different samples on subsequent measurement results. After normalization, the embedding vectors are arithmetically averaged by category, and the resulting average vector is used as the prototype vector of each category. This allows each prototype vector to centrally represent the typical turnover risk characteristics of the corresponding employee or candidate in the embedding space, thereby improving the separability and measurement stability of different categories in the embedding space.

[0052] The normalization process and the calculation methods for prototype vectors of each category are as follows:

[0053] In the prototype computation module, when the embedding vectors from the support set are received, a normalization process based on the L2 norm is performed on the embedding vector of each sample to ensure that the vector magnitudes of different samples are on a uniform scale. For the i-th sample with class index c in the support set, its embedding vector is denoted as z. c,i The prototype computation module performs the normalization operation shown in the following formula on the embedded vector:

[0054]

[0055] in, z represents the normalized embedding vector of the i-th sample with category index c. c,i Let |z| represent the original embedding vector of the i-th sample with class index c. c,i ||2 represents the embedding vector z c,i The L2 norm of the vector is given by: c represents the category index used to distinguish different employee or candidate categories in the support set; i represents the sample index used to number the support samples within the category with category index c; ε represents a small positive constant used to avoid zero denominators and improve numerical stability; and ||·||2 represents the operator for calculating the L2 norm of the vector. By performing the above L2 norm-based normalization on the embedding vector, the prototype calculation module reduces the interference of the difference in the magnitude of the embedding vectors between different samples on the subsequent distance metric, making the numerical range of the metric function more stable in the high-dimensional embedding space.

[0056] After normalizing all embedding vectors in the support set, the prototype calculation module performs an arithmetic mean on the normalized embedding vectors according to each category to generate prototype vectors for each category; for the category with category index c, its prototype vector p c The calculation is expressed as:

[0057]

[0058] Where, p c N represents the prototype vector of the category with category index c in the embedding space. cThis indicates the number of support samples in the support set that belong to category index c. Let represent the normalized embedding vector of the i-th supporting sample with category index c, where i represents the sample index that numbers the supporting samples within the category, and c represents the category index used to distinguish between different categories such as high-churn-risk and low-churn-risk categories. By arithmetically averaging the normalized embedding vectors within the same category, the resulting category prototype vector centrally reflects the typical turnover risk characteristics of employees or candidates in that category in the embedding space, providing a stable category center representation for the prediction module to determine the category based on the distance between the query sample and each category prototype vector.

[0059] Specifically, the above implementation details the steps of the prototype calculation module. By introducing a normalization operation based on the L2 norm, the module weakens the impact of differences in the feature vector magnitudes of different employees or candidates at the numerical level, making the subsequent distance calculation process more stable on a numerical scale and less susceptible to the influence of amplified or reduced similarity of abnormal samples. On this basis, the system performs an arithmetic average of the normalized sample features in the same category, thereby forming a prototype representation that represents the overall distribution center of the category in the embedding space. This prototype forms a significant spatial distribution difference between high-churn-risk and low-churn-risk categories, which helps to improve the reliability of category determination of query samples in small sample scenarios, and provides more robust feature support and discrimination basis for recruitment decisions and employee retention strategies.

[0060] Specifically, the support set, in actual deployment, can consist of historical employees or candidate samples from completed processes. During the initialization phase, the system constructs an initial support set based on existing labeled data, and continuously adds new labeled samples through an update module during subsequent operation, ensuring that the prototype vectors for each category always reflect the latest data distribution trend. In small-sample scenarios, the number of support samples for each category is typically several to dozens. To avoid unstable prototype estimation due to insufficient category samples, this embodiment sets the lower limit for the number of support samples for each category to at least 3-5. When the number of effective support samples for a certain category falls below the lower limit, the prototype calculation module can temporarily use the prototype vector from the previous period or adopt the known prototype vector closest to that category as a degenerate replacement, ensuring that the system can still provide continuous prediction results even when labeled data is insufficient. Furthermore, the small positive constants in the normalization process can be selected on an appropriate order of magnitude based on the dimension and numerical range of the embedded vector. Typical values ​​can range from 10 to 10 to avoid division by zero errors without excessively affecting the preservation of vector direction information. The dimension of the embedded vector can be set to common sizes such as 16, 32, or 128 dimensions based on the neural network structure. Those skilled in the art can choose a suitable size based on a trade-off between computational resources and model complexity. Optionally, if some enterprises cannot clearly distinguish between high and low churn risk in historical data and can only provide a coarse-grained risk stratification, the prototype calculation module can also configure the number of categories to two or more without changing the overall processing flow. It can still generate prototypes for each category using the same normalization and arithmetic averaging strategies, thus ensuring compatibility with more business grading strategies while maintaining consistent input and output interfaces. For cases where there are abnormal samples or obvious labeling errors in the support set, the prototype calculation module in this embodiment can perform a simple rule check on the embedded vector magnitude and category label before calculation, and remove samples whose magnitude is far beyond the normal range or whose labels conflict with the basic business rules from the current batch of prototype calculation, so as to reduce the interference of a single abnormal sample on the category prototype.

[0061] In one embodiment, it also includes an update module, configured to fine-tune the embedding module using an incremental learning algorithm and update the support set and prototype vector when newly labeled employee or candidate feature data is obtained, in order to adapt to changes in data distribution under small sample recruitment scenarios.

[0062] In one embodiment, the prediction module calculates the distance between the embedding vector of the query sample and each prototype vector by using a distance metric function, and determines the category prediction result based on the minimum distance principle. The metric function includes Euclidean distance or cosine similarity.

[0063] The prediction module is configured as follows:

[0064] After obtaining the embedding vector of the query sample, the Euclidean distance and / or cosine similarity between the embedding vector and the prototype vector of each category are measured, and the distance or similarity corresponding to each category is converted into a predicted probability distribution using a probability normalization function with a scaling coefficient. The prediction module also outputs the category prediction result of the query sample according to the category with the smallest distance or the highest predicted probability, and forms recruitment decision information with risk score based on the result, which is used to distinguish between high-risk and low-risk categories and provide a quantitative basis for recruitment or retention strategies.

[0065] In the prediction module, after obtaining the human resources feature data of the query sample, the corresponding query sample embedding vector is obtained through the embedding module. Then, the distance or similarity is calculated with the prototype vector of each category based on the selected metric function. The category prediction result is then derived from the distance or similarity to distinguish between high-risk and low-risk categories.

[0066] When the metric function is Euclidean distance, the prediction module is based on the normalized embedding vector of the query sample. With each category's prototype vector p c The Euclidean distance between them is calculated and expressed as:

[0067]

[0068] in, This represents the Euclidean distance between the query sample and the class prototype vector with class index c. The superscript (E) indicates that the distance is measured using the Euclidean distance metric. This represents the embedding vector of the query sample after embedding and normalization. The superscript (q) indicates that the vector corresponds to the query sample. c Let c represent the prototype vector of the category with category index c in the embedding space, ||·||2 represent the L2 norm of the vector, and c represent the category index used to distinguish between high-churn-risk categories and low-churn-risk categories; the prediction module can obtain the predicted category based on the minimum distance principle, for example:

[0069]

[0070] in, This indicates the predicted category index given for the query sample. This indicates the Euclidean distance across all category indexes c. The category index corresponding to the minimum value; to obtain the probability output that can be used for risk threshold setting, the prediction module can also convert the Euclidean distance into category probabilities through a softmax form with a scaling factor:

[0071]

[0072] in, This represents the normalized embedding vector for a given query sample. Under the condition that the query sample belongs to the category with category index c, y represents the random variable of the category label of the query sample, α represents the positive scaling coefficient used to adjust the strength of the influence of Euclidean distance on the probability distribution, c′ represents the category index used when summing all candidate categories in the denominator, and exp(·) represents the exponential operation; at this time, the predicted category can be represented as the category index that maximizes the predicted probability, which is consistent with the minimum distance principle in terms of monotonicity;

[0073] When the metric function is cosine similarity, the prediction module can be based on the normalized embedding vector of the query sample. With each category's prototype vector p c The similarity of the included angle between them is calculated, and the cosine similarity expression is:

[0074]

[0075] in, This represents the cosine similarity between the query sample and the prototype vector of the category with category index c. This represents the inner product of the normalized embedding vector and the prototype vector. Let ||p| represent the L2 norm of the normalized embedding vector of the query sample. c ||2 represents the L2 norm of the category prototype vector, and ε represents a small positive constant introduced to avoid the denominator approaching zero, with the same meaning as ε in the normalization process; the prediction module can select the prediction category based on the principle of maximum similarity:

[0076]

[0077] in, This indicates the predicted category index given for the query sample. This indicates that the cosine similarity is calculated across all category indexes c. The category index corresponding to the maximum value; to obtain a smoother risk score in recruitment scenarios, the prediction module can also convert cosine similarity into probability using softmax:

[0078]

[0079] Wherein, β represents the positive scaling coefficient used to adjust the strength of the influence of cosine similarity on the probability distribution, and its value affects the steepness of the distribution of probability differences between categories; through the above conversion of distance or similarity to probability, the system can give a predicted probability close to 1 for high-churn-risk categories and a predicted probability close to 0 for low-churn-risk categories, which makes it easier for the recruitment decision module to output hiring suggestions or retention measures based on preset thresholds.

[0080] Specifically, this implementation closely integrates two commonly used metrics, Euclidean distance and cosine similarity, with the category prediction process in human resources scenarios. By calculating the Euclidean distance between the query sample and each category prototype, the system can directly determine the category based on the minimum distance principle. Simultaneously, it utilizes probability transformation with a scaling factor to achieve fine-grained differentiation between high and low churn risk. The introduction of cosine similarity, targeting vector direction information, is more suitable for processing normalized embedding representations, highlighting category prototypes with similar directions even when sample modulus differs. Furthermore, by combining maximum similarity-based decision-making with risk probability output formed using the softmax method, the model can not only make hard-decision category judgments but also provide continuous risk scores. This facilitates setting risk thresholds for different business scenarios in human resource management, triggering candidate screening rules or employee early warning strategies, thereby improving the stability and usability of churn risk identification under small sample conditions.

[0081] Furthermore, the scaling coefficient in the probability normalization function can be set based on historical data verification results. Its value is typically chosen to be between one and several tens, used to adjust the sensitivity of distance or similarity to the probability distribution. When the scaling coefficient is small, the output probability is smoother, suitable for lenient risk stratification. When the scaling coefficient is large, the probability difference between different categories is amplified, more suitable for recruitment scenarios requiring explicit screening of high-risk individuals. In this embodiment, the preset risk threshold can be set by retrospectively analyzing historical recruitment and attrition results, combined with the experience of the human resources department. A typical range is between 0.5 and 0.9. Different companies can choose a more conservative or aggressive threshold configuration based on their acceptable risk level. For example, in positions requiring strict control of labor costs, the high attrition risk threshold can be set lower, classifying candidates with a predicted probability slightly above the neutral level as high-risk to prompt further manual review. In positions with greater flexibility in staffing, the threshold can be increased, triggering the interception rule only when the predicted probability is significantly higher. Optionally, when calculating Euclidean distance and cosine similarity simultaneously, the prediction module can internally perform a simple weighting of the two metrics or select one as the primary criterion. When the category judgments given by the two metrics are inconsistent, the system can prioritize the metric that is more consistent with historical validation results to reduce the impact of judgment fluctuations on business decisions. When the predicted probabilities of all categories of a query sample are close and cannot be clearly distinguished, the prediction module can mark the sample as uncertain and output a neutral risk score. The recruitment decision module can then prompt for manual review or postponement of the decision, thereby avoiding overly arbitrary automatic judgments in high-uncertainty areas while ensuring feasibility.

[0082] In one embodiment, the pre-trained neural network model is a multilayer perceptron network or a convolutional neural network structure.

[0083] This application proposes a human resources information processing method, applied to recruitment decisions and employee turnover risk prediction, combining... Figure 2 As shown, the method includes:

[0084] Step S1: Obtain employee or candidate characteristic data from the human resources database for recruitment decisions. The characteristic data includes at least one attribute related to employee status and / or turnover risk.

[0085] Step S2: Map the feature data into the embedding space through a nonlinear transformation to generate the corresponding embedding vector;

[0086] Step S3: Based on the support set containing the embedding vectors and their category labels, calculate the prototype vectors for multiple categories, where the category labels at least represent high-churn-risk categories and low-churn-risk categories;

[0087] Step S4: Based on the distance between the embedding vector of the query sample and multiple prototype vectors, use a metric function to calculate the distance and output the category prediction result of the query sample;

[0088] Step S5: Generate corresponding recruitment decision information based on the category prediction results;

[0089] In one embodiment, the mapping in step S2 is implemented through a pre-trained neural network model, which is initialized from a large-scale human resources dataset through transfer learning.

[0090] In one embodiment, step S3 includes: normalizing the embedding vectors in the support set, and calculating the mean of the embedding vectors by category to generate each prototype vector.

[0091] In one embodiment, the method further includes: when newly labeled employee or candidate feature data is obtained, using an incremental learning algorithm to fine-tune the neural network model used to perform step S2, and updating the support set and prototype vector to adapt to changes in data distribution under small sample recruitment scenarios.

[0092] Optionally, the incremental learning algorithm can be triggered at fixed time periods or in batches based on accumulated samples during engineering implementation. For example, a fine-tuning process can be initiated daily, weekly, or after a certain number of newly labeled samples to avoid excessive system resource consumption due to overly frequent model updates. Furthermore, a small learning rate and a limited number of iterations can be used during fine-tuning to mitigate the drastic impact of new data on existing parameters. This allows the model to maintain stable performance on existing samples while adapting to the latest data distribution. For cases where the number of samples in each class changes significantly, the update module can apply a simple moving average strategy when generating new prototype vectors, weighting the prototypes calculated in the current batch with historical prototypes according to a certain ratio, thereby smoothing the migration path of the class center. In this embodiment, when no new reliable labels are obtained within a certain time window, the update module can keep the neural network parameters and prototype vectors unchanged, only recording timestamp information in the support set management. An update can then be performed centrally when a new batch of labels arrives, reducing the impact of fragmented updates on system maintainability. Similarly, if abnormal situations such as training failure, abnormally increased loss, or significant decline in evaluation indicators occur during incremental training, the update module can roll back to the model parameters and prototype vectors of the previous version and record alarm information for operation and maintenance personnel to analyze, ensuring that the system always uses a stable version of the model for online prediction while the human resources business continues to operate.

[0093] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0094] Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are meant to be within the scope of this application and form different embodiments. For example, all the embodiments above can be used in any combination. The information disclosed in this background section is intended only to enhance the understanding of the general background of this application and should not be construed as an admission or in any way implying that such information constitutes prior art known to those skilled in the art.

Claims

1. A human resources information processing system, characterized in that, include: The data acquisition module is configured to acquire characteristic data of employees or candidates from a human resources database for recruitment decisions, the characteristic data including at least one attribute related to employee status and / or risk of leaving the company. The embedding module is configured to map the feature data to the embedding space to generate the corresponding embedding vector, and the mapping process is a nonlinear transformation. The prototype calculation module is configured to calculate prototype vectors for multiple categories based on a support set, wherein the support set includes the embedding vectors and corresponding category labels, and the category labels at least represent high-churn-risk categories and low-churn-risk categories; The prediction module is configured to output the category prediction result of the query sample based on the distance between the embedding vector of the query sample and multiple prototype vectors, wherein the distance is calculated based on a metric function; The recruitment decision module is configured to generate corresponding recruitment decision information based on the predicted results of the categories, which is used to assist the human resources department in determining whether to hire the candidates or whether to take retention measures for existing employees.

2. The human resources information processing system as described in claim 1, characterized in that, The embedding module includes a pre-trained neural network model, which is initialized from a large-scale human resources dataset through transfer learning to improve feature representation capabilities in small-sample recruitment scenarios.

3. The human resources information processing system as described in claim 1, characterized in that, The prototype calculation module is configured to normalize the embedding vectors in the support set, and then calculate the mean of the embedding vectors by category to generate each prototype vector. The prototype calculation module is configured as follows: Normalization based on the second norm is performed on each embedding vector in the support set to constrain the magnitude of the embedding vector within a preset range, thereby reducing the impact of scale differences between different samples on subsequent measurement results. After normalization, the embedding vectors are arithmetically averaged by category, and the resulting average vector is used as the prototype vector of each category, so that each prototype vector centrally represents the typical turnover risk characteristics of the corresponding employee or candidate in the embedding space.

4. The human resources information processing system as described in claim 1, characterized in that, It also includes an update module, configured to fine-tune the embedding module using an incremental learning algorithm when newly labeled employee or candidate feature data is obtained, and update the support set and the prototype vector to adapt to changes in data distribution under small sample recruitment scenarios.

5. A human resources information processing system as described in claim 1, characterized in that, The prediction module calculates the distance between the embedding vector of the query sample and each of the prototype vectors using the distance calculation of the metric function, and determines the category prediction result based on the minimum distance principle. The metric function includes Euclidean distance or cosine similarity. The prediction module is configured as follows: After obtaining the embedding vector of the query sample, the Euclidean distance and / or cosine similarity between the embedding vector and the prototype vector of each category are measured, and a probability normalization function with a scaling coefficient is used to convert the distance or similarity corresponding to each category into a predicted probability distribution. The prediction module also outputs the category prediction result of the query sample according to the category with the smallest distance or the largest predicted probability, and forms recruitment decision information with risk score based on the result, which is used to distinguish between high-risk and low-risk categories and provide a quantitative basis for recruitment or retention strategies.

6. A human resources information processing system as described in claim 2, characterized in that, The pre-trained neural network model is a multilayer perceptron network or a convolutional neural network structure.

7. A human resources information processing method, applied to a human resources information processing system according to any one of claims 1 to 6, characterized in that, include: Step S1: Obtain employee or candidate characteristic data from the human resources database for recruitment decisions, the characteristic data including at least one attribute related to employee status and / or turnover risk; Step S2: Map the feature data into the embedding space through a nonlinear transformation to generate the corresponding embedding vector; Step S3: Based on the support set containing the embedded vector and its category label, calculate prototype vectors for multiple categories, wherein the category label represents at least high churn risk categories and low churn risk categories; Step S4: Based on the distance between the embedding vector of the query sample and the multiple prototype vectors, perform distance calculation using a metric function and output the category prediction result of the query sample; Step S5: Generate corresponding recruitment decision information based on the category prediction results.

8. The human resources information processing method as described in claim 7, characterized in that, The mapping in step S2 is achieved through a pre-trained neural network model, which is initialized from a large-scale human resources dataset through transfer learning.

9. A human resources information processing method as described in claim 7, characterized in that, Step S3 includes: normalizing the embedding vectors in the support set, and calculating the mean of the embedding vectors by category to generate each prototype vector.

10. A human resources information processing method as described in claim 7, characterized in that, Also includes: When newly labeled employee or candidate feature data is obtained, an incremental learning algorithm is used to fine-tune the neural network model used to perform step S2, and the support set and the prototype vector are updated to adapt to the data distribution changes in the small sample recruitment scenario.

Images