Personnel scheduling method and system based on neural network

Through neural network technology, personnel scheduling is evaluated, combined with resumes, work records, soft skills and behavioral characteristics, the problem of subjective evaluation bias is solved, the accuracy and efficiency of evaluation is improved, and multi-dimensional evaluation results are achieved.

CN118982272AInactive Publication Date: 2024-11-19ZHEJIANG SUPENG INFORMATION TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410994185.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-24
Publication Date
2024-11-19
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The existing personnel scheduling methods rely on subjective assessments by the human resources department or managers, which are biased and inefficient, making it difficult to effectively separate subjective and objective factors, resulting in the accuracy and fairness of the evaluation results being affected.

Method used

Using a neural network-based personnel scheduling method, the subject's resume and work records are collected and preprocessed, and the information identification model and evaluation model are used to integrate and evaluate text features and tabular data, combining soft skills and behavioral characteristics to achieve the effect of Multi-dimensional evaluation of the subjects reduces the impact of subjective bias.

Benefits of technology

Improve the objectivity and accuracy of the evaluation results, reduce the random error of manual intervention, realize a comprehensive evaluation of subjects, improve the efficiency of talent assessment, reduce the work burden of human resource management, and adapt to different enterprises and positions. need.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118982272A_ABST
    Figure CN118982272A_ABST
Patent Text Reader

Abstract

The invention relates to the field of data analysis, and relates to a personnel scheduling method and system based on a neural network, and the method comprises the steps: collecting a resume and a work record of a subject; the resumes and the work records of the subjects are subjected to coarse classification preprocessing, and text feature vectors and table data are output; inputting the text feature vector to a pre-constructed information identification model to output a background data feature vector; integrating and preprocessing the background data feature vector and the table data and then outputting an evaluation feature vector; and inputting the evaluation feature vector to a pre-constructed evaluation model to output a score. According to the method, subjective and objective factors are successfully mixed, the influence of subjective deviation in the evaluation process is remarkably diluted, and the objectivity and accuracy of an evaluation result are improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of data analysis, and more specifically, to a personnel scheduling method and system based on neural network. Background Art

[0002] In the current data analysis field about personnel scheduling, personnel scheduling relies on the results of talent assessment, which mainly relies on the subjective assessment of the human resources department or direct managers, which may be biased and inefficient.

[0003] The Chinese patent with the existing authorization announcement CN112581037B discloses a background investigation method and system for multi-dimensional talent evaluation, including the following steps: establishing an evaluation index questionnaire, screening indicators, forming talent evaluation indicators, establishing a talent questionnaire, matching talent information and generating a talent evaluation report. A background investigation system for multi-dimensional talent evaluation is also disclosed, including establishing an evaluation index questionnaire module, screening index module, forming a talent evaluation index module, establishing a talent questionnaire module, matching talent information module and generating a talent evaluation report module. The background investigation method and system for multi-dimensional talent evaluation realizes quantitative evaluation of target talents, while avoiding the influence of human subjective will to the greatest extent, improving the accuracy of target talent evaluation results, and shortening time and reducing costs.

[0004] Although the above-mentioned disclosed technology reduces the influence of human subjective will, in actual operation, subjective and objective factors are in a separated state. For example, when evaluating soft skills, expert evaluation itself is a subjective evaluation, and the grading evaluation is also to make the subjective evaluation as objective as possible, but the essence of subjective evaluation remains unchanged. Subjective evaluation is indispensable in the talent evaluation stage of personnel scheduling. Therefore, it is necessary to disassemble, separate and reorganize the subjective and objective factors, and use objective factors to dilute the influence of subjective factors as much as possible. If the subjective and objective are still in a separated state, the subjective factors will have a great impact on the talent evaluation, and then cause deviations in personnel scheduling. Summary of the invention

[0005] In order to solve the above problems, the present invention provides a personnel scheduling method based on a neural network, the method comprising:

[0006] Collect subjects' resumes and work records;

[0007] Perform rough preprocessing on the subjects’ resumes and work records to output text feature vectors and table data;

[0008] Input text feature vector to the pre-built information recognition model to output background data feature vector;

[0009] The background data feature vector is integrated with the table data and preprocessed to output the evaluation feature vector;

[0010] Input the evaluation feature vector to the pre-built evaluation model output score.

[0011] Furthermore, the method for constructing the information recognition model includes:

[0012] The information recognition model includes an input layer, a word embedding layer, a convolution layer, a hidden layer, and an output layer; the input layer is a text feature vector, in which the structured values ​​in the text feature vector are represented by a unique hot code; the word embedding layer uses word embedding technology, the input dimension of the word embedding layer is V, V is a preset value, and the output dimension of the word embedding layer is T, T is a preset value; the input dimension size of the convolution layer is the number of samples × the sequence length G at input × the output dimension of the word embedding layer; the number of convolution kernels is 32, the convolution kernel size is 3 × the output dimension of the word embedding layer, the sliding strategy is 1 word interval sliding, the activation function is ReLU, and the expression formula of the output sequence length size G′ is: G is the sequence length at input, the convolution kernel size is kj, Indicates rounding down.

[0013] Furthermore, the specific construction method of the hidden layer includes:

[0014] The hidden layer uses LSTM network, and there are T / 2 hidden units in the hidden layer. The hidden units include input end, cell state, input gate, forget gate, cell candidate value, cell state update, output gate and hidden state; the input dimension of the input end is the output dimension of the word embedding layer + structured numerical dimension; each time step of the cell state has T / 2 cell state values; the input gate multiplies the input sequence and the hidden state at the previous moment, and then calculates the input gate output value through the sigmoid function, and the input gate output value ∈ [0, 1]; the input sequence of the forget gate and the hidden state at the previous moment are multiplied The cell candidate value is the amount that needs to be added to the cell state to calculate the input information, and the cell candidate value is obtained by transforming the input sequence and the hidden state at the previous moment through the tanh function; the cell state is updated according to the input gate, the forget gate and the cell candidate value; the input sequence of the output gate is dot-multiplied with the hidden state at the previous moment, and then the sigmoid function is used to calculate the output value of the output gate to control the amount of information output by the cell state; the hidden state is dot-multiplied by the output gate and the cell state to obtain the hidden state at this moment.

[0015] Furthermore, the training method of the information recognition model includes:

[0016] Extract J historical recognition logs from the database. One historical recognition log includes one set of text feature vectors and the corresponding background data feature vectors. Divide the J historical recognition logs into a training set and a validation set. The training set is C historical recognition logs and the validation set is P historical recognition logs, with a ratio of 4:1. The optimizer is the Adam optimizer with an initial learning rate of 0.001. Import the training set into the model and use the square error loss function. The model predicts the probability and label value. Update the value of each parameter according to the optimizer. Reduce the learning rate by half every 5-10 rounds, calculate the loss function on the validation set, and observe for 100-150 rounds to see if the loss function converges. If so, terminate the training.

[0017] Furthermore, the method of integrating and preprocessing the background data feature vector and the table data includes:

[0018] Calculate the mean and standard deviation of each feature in the background data feature vector, subtract the mean from each feature in the background data feature vector, and divide by the standard deviation; verify the index or ID of the background data feature vector and the tabular data to identify each subject, and use the Pandas library to merge the background data feature vector and the tabular data; use feature selection techniques based on statistical methods to screen out redundant or irrelevant features after the background data feature vector and the tabular data are merged, and integrate the remaining features into a complete feature matrix, that is, the evaluation feature vector.

[0019] Furthermore, the method for constructing the evaluation model includes:

[0020] The set of all evaluation feature vectors is used as the input of the evaluation model, the evaluation model uses the score predicted for each group of evaluation feature vectors as the output, uses the actual score corresponding to each group of evaluation feature vectors as the prediction target, and uses minimizing the sum of the first prediction accuracies of all predicted scores as the training target;

[0021] Among them, the calculation formula for the first prediction accuracy is: γ x =(vf x -ur x ) 2 , where x is the group number of each group of evaluation feature vectors, γ x is the first prediction accuracy, vf x Evaluate the predicted score corresponding to the feature vector for group x, ur x is the actual score corresponding to the x-th group of evaluation feature vectors; the evaluation model is trained until the sum of the first prediction accuracies reaches convergence and the training is stopped; the evaluation model is a multilayer perceptron or a convolutional neural network.

[0022] Furthermore, it also includes:

[0023] Conduct questionnaire structure, indicator quantification and behavioral data analysis on soft skill assessment and behavioral characteristics to obtain soft skill data and behavioral data;

[0024] Input soft skill data and behavioral data into the subjective quantitative model to output soft ability assessment;

[0025] Both soft skill assessment and behavioral characteristics are covered as soft competency assessment.

[0026] Furthermore, the questionnaire structure, quantitative indicators, and behavioral data analysis methods include:

[0027] Structured questionnaire and quantitative indicators: A structured questionnaire was designed using the Likert scale to quantify soft skill assessment and behavioral characteristics through clear questions and scoring criteria;

[0028] Behavioral data analysis: Using data from the company's internal work reports and task management systems, the behavioral characteristics of the subjects are evaluated through data mining and natural language processing techniques.

[0029] Furthermore, the method for constructing the subjective quantification model includes:

[0030] Using the soft skill data and the behavior data as analysis data, pre-collecting the soft ability evaluations corresponding to the c groups of analysis data, and converting the analysis data and the corresponding soft ability evaluations into a corresponding set of first feature vectors;

[0031] Each group of first eigenvectors is used as the input of a subjective quantification model. The subjective quantification model takes a group of soft capability evaluations corresponding to each group of analysis data as output, and takes the actual soft capability evaluations corresponding to each group of analysis data as prediction targets, where the actual soft capability evaluations are pre-collected soft capability evaluations corresponding to the analysis data. Minimizing the sum of prediction errors of all analysis data is used as a training target. The subjective quantification model is trained until the sum of prediction errors reaches convergence, and the training is stopped. The subjective quantification model is a random forest model.

[0032] A personnel dispatching system based on a neural network, the system comprising:

[0033] Collection module: collect subjects’ resumes and work records;

[0034] Quantification module: Conduct questionnaire structure, indicator quantification and behavioral data analysis on soft skill assessment and behavioral characteristics to obtain soft skill data and behavioral data;

[0035] Hybrid module: input soft skill data and behavioral data into the subjective quantitative model to output soft ability evaluation;

[0036] Regeneration module: Cover both soft skill assessment and behavioral characteristics into soft competency assessment;

[0037] Preprocessing module: Perform rough preprocessing on the subject's resume and work records to output text feature vectors and table data;

[0038] Background generation module: inputs text feature vector to the pre-built information recognition model to output background data feature vector;

[0039] Integration module: integrates and preprocesses the background data feature vector with the table data to output the evaluation feature vector;

[0040] Final scoring module: Input the evaluation feature vector to the pre-built evaluation model to output the score.

[0041] The technical effects and advantages of a neural network-based personnel scheduling method and system provided by the present invention are as follows:

[0042] The present invention firstly performs a rough decomposition of the resume and work record of the subject, and then uses the machine model of the dual-model algorithm combination to separately decompose and classify the unstructured text data, and then integrates the decomposed unstructured text data and the structured text data into an evaluation feature vector that can be quickly received by the machine learning model. Finally, the evaluation feature vector after the subjective evaluation is diluted is evaluated by the evaluation model to obtain the final score, which will be used as the basis for personnel scheduling. Through subjective evaluation processing, the influence of personal bias on the evaluation result is reduced, which is more fair and true. An end-to-end automated process is adopted to avoid random errors introduced by human intervention, and the model effect is more stable and controllable. The present invention combines the resume information, work records, soft skill evaluation and behavioral characteristics of the subjects, and the method can achieve a comprehensive and multi-dimensional evaluation of the subjects; this evaluation not only includes the hard skills (such as professional skills and work experience) of the subjects, but also covers soft skills (such as communication skills and teamwork) and behavioral characteristics (such as work attitude and adaptability); using pre-built information recognition models and evaluation models, it is possible to automatically process the resumes and work records of a large number of subjects and quickly generate evaluation results, which greatly improves the efficiency of talent evaluation and reduces the workload in human resource management; by using the Likert scale to quantitatively evaluate soft skills and behavioral characteristics, disassembling the resumes and work records of the subjects, and then analyzing the fragmented data, and then integrating and performing secondary analysis, as well as processing and analyzing using a neural network model, the subjective and objective factors are successfully mixed, the influence of subjective bias in the evaluation process is significantly diluted, and the objectivity and accuracy of the evaluation results are improved; it has a high degree of adaptability and can be customized and expanded according to the needs of different enterprises and positions. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a flow chart of a personnel scheduling method based on a neural network in Example 1;

[0044] Figure 2This is a flow chart of a personnel scheduling method based on a neural network in Example 2;

[0045] Figure 3 A connection diagram of a personnel scheduling system based on a neural network;

[0046] Figure 4 is a schematic diagram of an electronic device of the present invention;

[0047] Figure 5 Schematic diagram of the storage medium of the present invention. DETAILED DESCRIPTION

[0048] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0049] Embodiment 1:

[0050] See also Figure 1 As shown, a personnel scheduling method based on a neural network described in this embodiment includes:

[0051] Collect subjects' resumes and work records;

[0052] Perform rough preprocessing on the subjects’ resumes and work records to output text feature vectors and table data;

[0053] Input text feature vector to the pre-built information recognition model to output background data feature vector;

[0054] The background data feature vector is integrated with the table data and preprocessed to output the evaluation feature vector;

[0055] Input the evaluation feature vector to the pre-built evaluation model output score.

[0056] The subject's resume includes basic information, professional skills, soft skill assessment and behavioral characteristics; the subject's resume is obtained through questionnaires, China Higher Education Student Information and Career Center and internal company work reports; basic information includes basic information such as age, education and length of service, professional skills include professional abilities such as professional knowledge and technical ability, soft skills include communication skills and teamwork, and behavioral characteristics include work attitude, learning ability and adaptability.

[0057] Work records include work experience, company name, position, working hours, job description and project reports; work records are obtained from internal work reports of the company; among them, project reports are detailed reports on projects that the subjects participated in at work and study, including project name, description, technology stack, role, work content and project results.

[0058] The method for rough classification pretreatment comprises:

[0059] The unstructured text data in the subject's resume and work records are segmented, stop words are removed, and stems are extracted. Word2Vec is used to convert the unstructured text data into text feature vectors in vector form, and the structured text data is organized into a table to obtain tabular data; unstructured text is a large amount of descriptive language and is the main carrier area of ​​subjective evaluation, such as the manager's performance evaluation of the subject:

[0060] "A002 has strong technical capabilities and is proficient in multiple programming languages ​​and development frameworks. He has made great contributions to the company's R&D. However, he is sincere and needs to learn how to flexibly mediate conflicts of interest within the team. Improving this will benefit his leadership potential."

[0061] Soft skills self-evaluation in resume:

[0062] "I have strong communication skills and have represented the company in many important customer negotiations, using my flexible expression skills to solve various problems. However, I have a straightforward personality and tend to touch other people's sensitive nerves when working in a team. I need to improve my ability to control my emotional expression."

[0063] In the case of structured data, data annotation is done mechanically, for example:

[0064] Table 1 Subject resume table (structured data)

[0065] Basic Information Professional skills Soft skills Behavioral characteristics Age: 30 Major: Computer Science Communication skills: Excellent Work attitude: Positive Education: PhD Skills: Python development Teamwork: Good Learning ability: strong Working experience: 8 years Resilience: Fast

[0066] Table 2 Subject work record table (structured data)

[0067]

[0068]

[0069] When the subjective and objective factors are separated and no further processing is carried out, inefficiency will occur. For example, in the past common sense cognition, excellent communication skills = excellent teamwork, but the subjective evaluation of emotional expression has led to a decline in the evaluation of teamwork. If the separation is carried out and no further processing is carried out, objectively, the evaluation results of excellent teamwork and poor teamwork will appear at the same time. In general, this difference is confirmed by experts, so inefficient behavior will occur. The cognition of expert confirmation itself may be subjectively biased. For example, the psychological cognition of the first impression will affect the subsequent evaluation. If the expert first sees the soft skill self-evaluation in the unstructured text data, then the first impression will be that this person has low emotional intelligence, and the subsequent evaluation will include his own subjective guesses. However, in the actual comprehensive evaluation, when conducting a comprehensive evaluation from a bird's eye view, it will be known that this person does not have low emotional intelligence, but is a matter-of-fact person. However, the amount of data and time required to obtain this comprehensive evaluation from a bird's eye view will be very long, and the efficiency is too low. Therefore, the unstructured text data and structured text data here need to be separated and adaptively processed to prepare for the subsequent classification of the machine learning model.

[0070] It should be noted that Word2Vec is a model for natural language processing that can learn the vector representation of words, that is, convert words into word vectors. It is a fairly mature natural language processing model on the market.

[0071] The method for constructing the information recognition model includes:

[0072] The information recognition model includes an input layer, a word embedding layer, a convolution layer, a hidden layer, and an output layer; the input layer is a text feature vector, in which the structured values ​​in the text feature vector are represented by a unique hot code; the word embedding layer uses word embedding technology, the input dimension of the word embedding layer is V, V is a preset value, and the output dimension of the word embedding layer is T, T is a preset value; the input dimension size of the convolution layer is the number of samples × the sequence length G at input × the output dimension of the word embedding layer; the number of convolution kernels is 32, the convolution kernel size is 3 × the output dimension of the word embedding layer, the sliding strategy is 1 word interval sliding, the activation function is ReLU, and the expression formula of the output sequence length size G′ is: G is the sequence length at input, the convolution kernel size is kj, Indicates rounding down;

[0073] Convolutional neural networks can automatically learn implicit semantic dependencies such as vocabulary, phrases, and structures. Multi-scale sliding windows can capture semantic information of different granularities. The hidden layer learns high-level abstract features of text to support deep sentiment judgment. After testing, it can achieve high classification accuracy on large data sets.

[0074] The specific construction method of the hidden layer includes:

[0075] The hidden layer uses LSTM network, and there are T / 2 hidden units in the hidden layer. The hidden units include input end, cell state, input gate, forget gate, cell candidate value, cell state update, output gate and hidden state; the input dimension of the input end is the output dimension of the word embedding layer + structured numerical dimension; each time step of the cell state has T / 2 cell state values; the input gate multiplies the input sequence and the hidden state at the previous moment, and then calculates the input gate output value through the sigmoid function, and the input gate output value ∈ [0, 1]; the input sequence of the forget gate and the hidden state at the previous moment are multiplied The cell candidate value is the amount that needs to be added to the cell state to calculate the input information, and the cell candidate value is obtained by transforming the input sequence and the hidden state at the previous moment through the tanh function; the cell state is updated according to the input gate, the forget gate and the cell candidate value; the input sequence of the output gate is dot-multiplied with the hidden state at the previous moment, and then the sigmoid function is used to calculate the output value of the output gate to control the amount of information output by the cell state; the hidden state is dot-multiplied by the output gate and the cell state to obtain the hidden state at this moment.

[0076] The use of LSTM network can well capture the long-term dependencies in time series; LSTM controls the inflow and output of information through input gate, forget gate and output gate, and can learn which information needs to be remembered and which information needs to be forgotten. The Cell state value of each time step can remember the important information of the previous input sequence, effectively solving the short-term memory problem. At the same time, the forget gate can selectively forget irrelevant information. The input gate can control the update degree of the current input sequence to the Cell state to avoid the direct overwriting of old information. The forget gate can control the preservation degree of old information in the Cell state. The Cell candidate value can extract the nonlinear association of the input sequence through the tanh nonlinear transformation. The output gate can effectively control the amount of information output by the current Cell state to the hidden layer to avoid the problem of unnecessary information emerging in the hidden layer. The hidden state can learn a higher level of feature expression in the sequence by capturing the dependency of the input sequence at different time steps. The structural design of the hidden layer helps the model extract effective features and rules from the sequence data and improve the effect of information recognition.

[0077] It is particularly important to note that the above moments belong to time steps, which are composed of moments. Because the text feature vector contains not only numerical values ​​but also text descriptions, the data width that a computer system can process at the same time is generally only 64 bits, while text descriptions usually exceed 64 bits. Therefore, there will be a time difference when inputting, that is, time attributes appear, so moments and time steps need to be considered in the design.

[0078] The training method of the information recognition model includes:

[0079] Extract J history recognition logs from the database. One history recognition log includes one set of text feature vectors and corresponding background data feature vectors. Divide the J history recognition logs into a training set and a validation set. The training set is C history recognition logs and the validation set is P history recognition logs. The ratio is 4:1. The optimizer is Adam optimizer with an initial learning rate of 0.001. Import the training set into the model and use the square error loss function. The loss function ls formula is:

[0080]

[0081] Where d is the true label value, b is the model prediction probability, and g is the historical identification log group number.

[0082] It is important to note that machine learning classification problems generally use probabilistic selection for directional output. For example, after layers of calculation and screening, the probability that item 1 belongs to category 1 is 20%, the probability that item 2 belongs to category 1 is 30%, and the probability that item 3 belongs to category 1 is 40%. When outputting category 1, item 3 is selected as the final output, that is, item 3 belongs to category 1, and the model predicts a probability of 40%. The true label value is given an objective probability or manual assignment based on the actual situation to adjust the prediction direction of the model.

[0083] The model predicts probability and label value, and updates the value of each parameter according to the optimizer. The learning rate is reduced by half every 5-10 rounds, and the loss function on the validation set is calculated. After 100-150 rounds, the loss function is observed to see if it converges. If it converges, the training ends.

[0084] The trained information recognition model will classify and decompose the text feature vectors into background data feature vectors. The basic information and work records of the subjects contain a large amount of unstructured data. It would be quite difficult to directly use these for modeling, and problems such as data sparsity may easily occur. The background data feature vectors automatically learn implicit relationships through the information recognition model, and can extract high-level abstract features of the overall background of the subjects. These high-order background features are more beneficial to subsequent tasks such as evaluation and prediction.

[0085] The method for integrating and preprocessing the background data feature vector and the table data includes:

[0086] Calculate the mean and standard deviation of each feature in the background data feature vector, subtract the mean from each feature in the background data feature vector, and divide by the standard deviation; verify the index or ID of the background data feature vector and the tabular data to identify each subject, and use the Pandas library to merge the background data feature vector and the tabular data; use feature selection techniques based on statistical methods to screen out redundant or irrelevant features after the background data feature vector and the tabular data are merged, and use feature selection techniques based on statistical methods such as variance selection method, chi-square test and mutual information method to integrate the remaining features into a complete feature matrix, that is, the evaluation feature vector.

[0087] Calculating the mean and standard deviation of each feature and performing standardization can effectively solve the impact caused by the inconsistency of the original feature value range, which is beneficial to subsequent modeling. Using indicators to identify different subjects can accurately match background feature vectors and table data, retaining the integrity of sample information. Pandas performs feature integration and uses its efficient computing power to simplify the data processing process. Statistical feature selection can automatically filter out redundant features with low feature importance or strong correlation, reduce unnecessary information, and improve modeling efficiency. Features from different sources and types are integrated to extract their inherent correlation, which helps to discover implicit knowledge. The generated evaluation feature vector reflects the sample information value, reduces the risk of overfitting, and is beneficial to model interpretability. In this way, the process is highly automated in subsequent modeling, easy to deploy and expand, and the final task of subjective dilution is completed at the same time.

[0088] It is worth noting that the Pandas library can also be used to solve the calculation problem of calculating the mean and standard deviation of each feature in the background data feature vector. The Pandas library is one of the four basic libraries for machine learning.

[0089] The method for constructing the evaluation model includes:

[0090] The set of all evaluation feature vectors is used as the input of the evaluation model, and the evaluation model takes the predicted score for each group of evaluation feature vectors as the output, the actual score corresponding to each group of evaluation feature vectors as the prediction target, and minimizing the sum of the first prediction accuracies of all predicted scores as the training target.

[0091] Among them, the calculation formula for the first prediction accuracy is: γ x =(vf x -ur x ) 2 , where x is the group number of each group of evaluation feature vectors, γ x is the first prediction accuracy, vf x Evaluate the predicted score corresponding to the feature vector for group x, ur xis the actual score corresponding to the x-th group of evaluation feature vectors; the evaluation model is trained until the sum of the first prediction accuracies reaches convergence and the training is stopped; the evaluation model is a multilayer perceptron or a convolutional neural network.

[0092] It is particularly noted that: after the previous integration and preprocessing, the subjective factors are digitized, that is, objectified, and then the obtained evaluation feature vector is modeled and reorganized using a machine learning model to achieve complete automated dilution. In this way, the impact of subjective factors in the evaluation obtained will be greatly reduced, and the expected tasks can be completed with high efficiency. Finally, the scores will be used as the basis for personnel scheduling. For example, those with higher scores will become managers, and those with lower scores will become manager candidates and have a better chance of being competent for important talent positions.

[0093] Example 2

[0094] like Figure 2 As shown, this embodiment is further improved on the basis of the embodiment 1. The difference is that the embodiment 1 provides a personnel scheduling method based on a neural network. In actual application, although the subject resume uses objective data to mix subjective data, which greatly reduces the influence of subjective factors, the soft skill evaluation and behavioral characteristics in the subject resume are still dominated by subjective factors. In view of the important influence of the subject resume on the subsequent scoring, a single subjective soft skill evaluation and behavioral characteristics may cause an objective evaluation. Therefore, it is necessary to mix the soft skill evaluation and behavioral characteristics subjectively and objectively. Based on this, a personnel scheduling method based on a neural network also includes:

[0095] Conduct questionnaire structure, indicator quantification and behavioral data analysis on soft skill assessment and behavioral characteristics to obtain soft skill data and behavioral data;

[0096] Input soft skill data and behavioral data into the subjective quantitative model to output soft ability assessment;

[0097] Overlay the soft skills assessment and behavioral characteristics in the subject’s resume and work record into a soft competency assessment.

[0098] The questionnaire structure, quantitative indicators, and behavioral data analysis methods include:

[0099] Structured questionnaire and quantitative indicators: A structured questionnaire was designed using the Likert scale to quantify soft skill assessment and behavioral characteristics through clear questions and scoring criteria.

[0100] Likert scale questionnaire example:

[0101] Rating settings: 1: strongly disagree; 2: disagree; 3: neutral; 4: agree; 5: strongly agree.

[0102] Example Question:

[0103] Communication skills: How often you express yourself clearly in team meetings (1-5).

[0104] Teamwork: How often you proactively help others on team projects (1-5).

[0105] Work attitude: Frequency of maintaining a positive attitude when facing stress (1-5).

[0106] Behavioral data analysis: Using data from the company's internal work reports and task management systems, the behavioral characteristics of the subjects are evaluated through data mining and natural language processing techniques.

[0107] Example behavioral data:

[0108] Work attitude: Task completion time and sentiment analysis in work logs.

[0109] Learning ability: number of training courses attended and activity on the company’s internal knowledge sharing platform.

[0110] Resilience: The speed and quality of response to emergency tasks.

[0111] The method for constructing the subjective quantification model includes:

[0112] The soft skill data and the behavior data are used as analysis data, the soft ability evaluations corresponding to the c groups of analysis data are collected in advance, and the analysis data and the corresponding soft ability evaluations are converted into a corresponding set of first feature vectors.

[0113] Each group of first eigenvectors is used as the input of the subjective quantitative model. The subjective quantitative model takes a group of soft ability evaluations corresponding to each group of analysis data as output, and takes the actual soft ability evaluations corresponding to each group of analysis data as the prediction target. The actual soft ability evaluations are the pre-collected soft ability evaluations corresponding to the analysis data. The training target is to minimize the sum of the prediction errors of all analysis data. The calculation formula of the prediction error is Z K =(a K -μ K ) 2 , where Z K is the prediction error, K is the group number of the first eigenvector corresponding to the analyzed data, α K is the soft capability evaluation corresponding to the Kth group of analysis data, μ K is the actual soft capability evaluation corresponding to the Kth group of analysis data; the subjective quantification model is trained until the sum of the prediction errors reaches convergence and the training is stopped; the subjective quantification model is specifically a random forest model.

[0114] Example 3

[0115] like Figure 3As shown, a personnel scheduling system based on a neural network includes:

[0116] Collection module: collect subjects’ resumes and work records;

[0117] Quantification module: Conduct questionnaire structure, indicator quantification and behavioral data analysis on soft skill assessment and behavioral characteristics to obtain soft skill data and behavioral data;

[0118] Hybrid module: input soft skill data and behavioral data into the subjective quantitative model to output soft ability evaluation;

[0119] Regeneration module: Cover both soft skill assessment and behavioral characteristics into soft competency assessment;

[0120] Preprocessing module: Perform rough preprocessing on the subject's resume and work records to output text feature vectors and table data;

[0121] Background generation module: inputs text feature vector to the pre-built information recognition model to output background data feature vector;

[0122] Integration module: integrates and preprocesses the background data feature vector with the table data to output the evaluation feature vector;

[0123] Final scoring module: Input the evaluation feature vector to the pre-built evaluation model to output the score.

[0124] Example 4

[0125] See also Figure 4 As shown, according to another aspect of the present application, an electronic device 500 is also provided. The electronic device 500 may include one or more processors and one or more memories. The memories store computer readable codes, and when the computer readable codes are run by one or more processors, a personnel scheduling method based on a neural network as described above may be executed.

[0126] The system according to the embodiment of the present application can also be used Figure 4 The electronic device architecture shown in FIG. Figure 4 As shown, the electronic device 500 may include a bus 501, one or more CPUs 502, ROM 503, RAM 504, a communication port 505 connected to a network, an input / output 506, a hard disk 507, etc. The storage device in the electronic device 500, such as ROM 503 or hard disk 507, may store a personnel scheduling method based on a neural network provided by the present application. Furthermore, the electronic device 500 may also include a user interface 508. Of course, Figure 4 The architecture shown is only exemplary and can be omitted according to actual needs when implementing different devices. Figure 4One or more components of an electronic device are shown.

[0127] Example 5

[0128] See also Figure 5 As shown, it is a computer-readable storage medium according to an embodiment of the present application. Computer-readable instructions are stored on the computer-readable storage medium. When the computer-readable instructions are executed by the processor, a personnel scheduling method based on a neural network according to an embodiment of the present application described with reference to the above figures can be executed. The storage medium includes, but is not limited to, for example, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and cache memory (cache), etc. Non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc.

[0129] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

[0130] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A personnel scheduling method based on neural network, Characterized in that the method include: Collect subjects' resumes and work records; Perform rough preprocessing on the subjects’ resumes and work records to output text feature vectors and table data; Input text feature vector to the pre-built information recognition model to output background data feature vector; The background data feature vector is integrated with the table data and preprocessed to output the evaluation feature vector; Input the evaluation feature vector to the pre-built evaluation model output score.

2. A personnel scheduling method based on a neural network according to claim 1, It is characterized in that The method for constructing the information recognition model includes: The information recognition model includes an input layer, a word embedding layer, a convolution layer, a hidden layer, and an output layer; the input layer is a text feature vector, in which the structured values ​​in the text feature vector are represented by a unique hot code; the word embedding layer uses word embedding technology, the input dimension of the word embedding layer is V, V is a preset value, and the output dimension of the word embedding layer is T, T is a preset value; the input dimension size of the convolution layer is the number of samples × the sequence length G at input × the output dimension of the word embedding layer; the number of convolution kernels is 32, the convolution kernel size is 3 × the output dimension of the word embedding layer, the sliding strategy is 1 word interval sliding, the activation function is ReLU, and the output sequence length size G' is expressed as: G is the sequence length at input, the convolution kernel size is kj, Indicates rounding down.

3. A personnel scheduling method based on neural network according to claim 2, It is characterized in that The specific construction method of the hidden layer includes: The hidden layer uses LSTM network, and there are T / 2 hidden units in the hidden layer. The hidden units include input end, cell state, input gate, forget gate, cell candidate value, cell state update, output gate and hidden state; the input dimension of the input end is the output dimension of the word embedding layer + structured numerical dimension; each time step of the cell state has T / 2 cell state values; the input gate multiplies the input sequence and the hidden state at the previous moment, and then calculates the input gate output value through the sigmoid function, and the input gate output value ∈ [0, 1]; the input sequence of the forget gate and the hidden state at the previous moment are multiplied The cell candidate value is the amount that needs to be added to the cell state to calculate the input information, and the cell candidate value is obtained by transforming the input sequence and the hidden state at the previous moment through the tanh function; the cell state is updated according to the input gate, the forget gate and the cell candidate value; the input sequence of the output gate is dot-multiplied with the hidden state at the previous moment, and then the sigmoid function is used to calculate the output value of the output gate to control the amount of information output by the cell state; the hidden state is dot-multiplied by the output gate and the cell state to obtain the hidden state at this moment.

4. A personnel scheduling method based on neural network according to claim 2, It is characterized in that The training method of the information recognition model includes: Extract J historical recognition logs from the database. One historical recognition log includes one set of text feature vectors and the corresponding background data feature vectors. Divide the J historical recognition logs into a training set and a validation set. The training set is C historical recognition logs and the validation set is P historical recognition logs, with a ratio of 4:

1. The optimizer is the Adam optimizer with an initial learning rate of 0.

001. Import the training set into the model and use the square error loss function. The model predicts the probability and label value. Update the value of each parameter according to the optimizer. Reduce the learning rate by half every 5-10 rounds, calculate the loss function on the validation set, and observe for 100-150 rounds to see if the loss function converges. If so, terminate the training.

5. A personnel scheduling method based on neural network according to claim 1, It is characterized in that The method for integrating and preprocessing the background data feature vector and the table data includes: Calculate the mean and standard deviation of each feature in the background data feature vector, subtract the mean from each feature in the background data feature vector, and divide by the standard deviation; verify the index or ID of the background data feature vector and the tabular data to identify each subject, and use the Pandas library to merge the background data feature vector and the tabular data; use feature selection techniques based on statistical methods to screen out redundant or irrelevant features after the background data feature vector and the tabular data are merged, and integrate the remaining features into a complete feature matrix, that is, the evaluation feature vector.

6. A personnel scheduling method based on neural network according to claim 5, It is characterized in that The method for constructing the evaluation model includes: The set of all evaluation feature vectors is used as the input of the evaluation model, the evaluation model uses the score predicted for each group of evaluation feature vectors as the output, uses the actual score corresponding to each group of evaluation feature vectors as the prediction target, and uses minimizing the sum of the first prediction accuracies of all predicted scores as the training target; Among them, the calculation formula for the first prediction accuracy is: γ x =(vf x -ur x ) 2 , where x is the group number of each group of evaluation feature vectors, γ x is the first prediction accuracy, vf x Evaluate the predicted score corresponding to the feature vector for group x, ur x is the actual score corresponding to the x-th group of evaluation feature vectors; the evaluation model is trained until the sum of the first prediction accuracies reaches convergence and the training is stopped; the evaluation model is a multilayer perceptron or a convolutional neural network.

7. A personnel scheduling method based on neural network according to claim 1, It is characterized in that Also includes: Conduct questionnaire structure, indicator quantification and behavioral data analysis on soft skill assessment and behavioral characteristics to obtain soft skill data and behavioral data; Input soft skill data and behavioral data into the subjective quantitative model to output soft ability assessment; Both soft skill assessment and behavioral characteristics are covered as soft competency assessment.

8. A personnel scheduling method based on neural network according to claim 7, It is characterized in that The questionnaire structure, quantitative indicators, and behavioral data analysis methods include: Structured questionnaire and quantitative indicators: A structured questionnaire was designed using the Likert scale to quantify soft skill assessment and behavioral characteristics through clear questions and scoring criteria; Behavioral data analysis: Using data from the company's internal work reports and task management systems, the behavioral characteristics of the subjects are evaluated through data mining and natural language processing techniques.

9. A personnel scheduling method based on neural network according to claim 7, It is characterized in that The method for constructing the subjective quantification model includes: Using the soft skill data and the behavior data as analysis data, pre-collecting the soft ability evaluations corresponding to the c groups of analysis data, and converting the analysis data and the corresponding soft ability evaluations into a corresponding set of first feature vectors; Each group of first eigenvectors is used as the input of a subjective quantification model. The subjective quantification model takes a group of soft capability evaluations corresponding to each group of analysis data as output, and takes the actual soft capability evaluations corresponding to each group of analysis data as prediction targets, where the actual soft capability evaluations are pre-collected soft capability evaluations corresponding to the analysis data. Minimizing the sum of prediction errors of all analysis data is used as a training target. The subjective quantification model is trained until the sum of prediction errors reaches convergence, and the training is stopped. The subjective quantification model is a random forest model.

10. A personnel scheduling system based on neural network, It is characterized in that A neural network-based personnel scheduling method according to any one of claims 1 to 8 is implemented, and the system comprises: Collection module: collect subjects’ resumes and work records; Quantification module: Conduct questionnaire structure, indicator quantification and behavioral data analysis on soft skill assessment and behavioral characteristics to obtain soft skill data and behavioral data; Hybrid module: input soft skill data and behavioral data into the subjective quantitative model to output soft ability evaluation; Regeneration module: Cover both soft skill assessment and behavioral characteristics into soft competency assessment; Preprocessing module: Perform rough preprocessing on the subject's resume and work records to output text feature vectors and table data; Background generation module: inputs text feature vector to the pre-built information recognition model to output background data feature vector; Integration module: integrates and preprocesses the background data feature vector with the table data to output the evaluation feature vector; Final scoring module: Input the evaluation feature vector to the pre-built evaluation model to output the score.