Adaptive recruitment decision-making system and method based on recruitment behavior data

Abstract

The present invention discloses an adaptive recruitment decision system based on recruitment behavior data to solve the problem that the existing technology does not take the portrait of the recruiter and the time factor of the employment strategy into consideration. Therefore, it is difficult for the decision model to learn historical experience and dynamic changes, which makes it difficult for the decision mechanism to adjust adaptively. The present invention adopts a behavior sequence method to select a batch of typical samples of recruitment information and candidate information, and make decisions on these samples based on the evaluation criteria of experts. The sequence after the decision is regarded as a historical decision sequence, and is input into the model together with the job information to be recruited and the candidate information to be evaluated to obtain the prediction result. The prediction result obtained in this way is consistent with the historical decision sequence. In this way, without changing the model parameters themselves, by inputting different behavior sequences, changes in the evaluation results will be brought about, which has the effect of dynamic adaptation.

Description

Technical Field

[0001] The present invention belongs to the field of data processing technology, and in particular relates to an adaptive recruitment decision-making system and method based on recruitment behavior data. Background Art

[0002] With the development of information technology and the widespread use of digital systems, employers' job postings and candidates' resumes are increasingly stored in digital form. The accumulation of data has also fueled the development of data processing and machine learning technologies in the recruitment field, such as the creation and application of systems for resumé parsing and matching job postings with resumes. Employers can train recruitment decision-making mechanisms into models using decision-making data, rather than conveying them to each recruiter through written regulations, to achieve consistency and professionalism in hiring decisions. Currently, recruitment decision-making models are primarily limited to matching job postings with resumes. The basic process involves first parsing the text to extract structured information, then performing a semantic similarity match between the job requirements and the skills and work experience described in the resume to assess whether the candidate meets the job requirements.

[0003] Similarity matching methods based solely on text content ignore the differences between recruiters and the dynamic influence of employers' hiring decision-making mechanisms that change over time. On the one hand, recruiters' understanding of hiring policies and their own experience vary, leading to different hiring decisions for the same position and candidate. On average, decisions made by experienced recruiters are better and more aligned with the employer's hiring strategy than those made by junior and mid-level recruiters. When new recruiters join the hiring decision-making team, they need to train them on their past decision-making experience to align with hiring goals. This training process is time-consuming and labor-intensive, and requires the new recruits to gradually accumulate experience. Furthermore, employers frequently adjust their hiring decisions due to overall or partial strategic adjustments and external market conditions. Similarity matching methods based on text content fail to incorporate the recruiter's profile and the temporal nature of hiring strategies. Therefore, the decision-making model struggles to learn from historical experience and dynamic changes, making it difficult for the decision-making mechanism to adapt adaptively. Summary of the Invention

[0004] In order to solve the defects of the existing technology, the present invention provides an adaptive recruitment decision-making system and method based on recruitment behavior data.

[0005] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0006] An adaptive recruitment decision-making system based on recruitment behavior data includes a data storage module and a decision model training module. The decision model training module includes a decision model and a data pre-processing module. The data storage module is used to store input data, which includes job recruitment information, candidate information, historical decision behavior sequences, and candidate review status. A relational database or a document database that supports nested structures is used.

[0007] The data pre-processing module is used to transform the unstructured portion of the original information into a structured one. It converts the original text of the basic information into a tree structure with nested attributes, recorded in a schema structure format in a key-value format: each key represents the path from the tree to a leaf node as jsonpath, and the value represents the field type on the leaf node. The pre-processing module incorporates a knowledge graph for data augmentation and information error correction, adding augmented fields to the main structure in a structured format. It also includes vectorized embedding, which embeds tokens using a dictionary lookup (lookup table) and adds positional encodings. It also includes tokenization, which accepts token sequences as input. The original transformer architecture processes text sequences.

[0008] As an optimal technical solution of the present invention, the vectorized embedding includes selecting different vectorized embedding methods for tokens according to their types, including discrete type, numerical type, string type, and advanced type. The discrete type is to map the token value into the numerical range of [0,1024) by mapping, and then expand it into a vector by dictionary search. The numerical type is normalized to the interval [-1,1] according to mu-law, and then evenly distributed to the value interval [1024,2048), and then expanded into a vector by dictionary search, where the mu-law formula is

[0009]

[0010] The present invention also provides an adaptive recruitment decision-making method based on recruitment behavior data. The self-supervised pre-training model adopts a model structure consistent with the decision-making model, but only resume data is used as training data. In order to simulate the decision-making behavior sequence, the following processing steps are included:

[0011] S1: Extract the most recent work experience and other information during this period from the resume as the candidate's last job recruitment information; remove other resume information from the most recent work experience and use it as the candidate's resume information when applying for the last job; randomly select and fill in the process except for the process; the decision behavior value is "Joining the job",

[0012] S2: Cluster all collected resumes based on the similarity of job titles in the most recent work experience;

[0013] S3: For each type of resume, group them according to the company where they worked most recently. Within each group, sort them by the start time of their most recent job. This way, each group is a decision sequence containing historical decision-making behaviors.

[0014] S4: Since the decision sequence simulated based on the resume only has positive samples, that is, only entry information, and no elimination information, negative sampling technology is used in each decision sequence of batch training.

[0015] Furthermore, a continuous training step based on real behavioral data: If the companies and positions in the simulated data are significantly different from those of the employer, the prediction effect of the pre-trained model cannot meet the employer's recruitment decision-making standards. Based on the pre-trained model, continue to use real decision-making behavior data for fine-tuning; after the data volume and diversity are sufficient and the decision-making effect meets the expectations, the model parameters can be fixed, and only the model can be used for prediction.

[0016] Furthermore, the decision model prediction process: When the employer uses a pre-trained model or a model fine-tuned based on its own decision data to make predictions, it follows the following process:

[0017] 1) Data storage: Upload candidate information and job recruitment information through the data collection system and store it; collect and store the decision-making behavior of recruiters through the process management system;

[0018] 2) Data processing: The data processing process is consistent with the data processing during training;

[0019] 3) Model prediction: The processed data is input into the model, the probabilities of various possible decision behaviors are output, and the action with the highest probability is selected as the recommended decision behavior for output.

[0020] Furthermore, in step 3, the conditional probability of the output of the decision model is expressed as

[0021] p θ(action|job,candidate,stage,history), where the conditional probability in the above formula expresses the probability of the next decision action made by a candidate applying for a job at the stage of the process, under the premise of following a certain recruitment behavior sequence history.

[0022] Furthermore, in step 3, the loss function of the decision model is defined as:

[0023]

[0024] Among them, y(action (i) ) is a masking function, that is, when action (i) When y(action (i) )=1; otherwise, y(action (i) )=0.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] 1) Adjusting recruitment decision strategies: When employers need to adjust recruitment evaluation criteria, the traditional approach is to develop standards for various local requirements. However, this makes it difficult to cover the diversity of input information, resulting in a loss of scientific consistency. Using a behavioral sequence approach, a group of representative samples of recruitment and candidate information can be selected and decisions made based on these samples using expert evaluation criteria. The decision sequence is treated as a historical decision sequence and input into the model along with information about the positions to be filled and the candidates to be evaluated to generate predictions. This method ensures that the predictions are consistent with the historical decision sequence. Thus, by inputting different behavioral sequences, the evaluation results can be varied without changing the model parameters, achieving a dynamic and adaptive effect.

[0027] 2) Training and Assessment of Recruiters: Recruiters understand recruitment evaluation criteria based on their own experience. Due to varying experiences, their understanding of evaluation criteria can be inconsistent, leading to individualized evaluation results. Training is equivalent to using the behavioral sequence data of recruitment experts to guide recruiters so that they can produce consistent evaluation results. Assessment is equivalent to collecting each recruiter's own behavioral sequence and then calculating the performance differences between these behavioral sequences and those of the recruitment experts. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a flow chart of the adaptive recruitment decision-making method based on recruitment behavior data of the present invention.

[0029] Figure 2It is a list of fields included in the recruitment information of the present invention.

[0030] Figure 3 This is a field list diagram of candidate information of the present invention.

[0031] Figure 4 It is a schematic diagram of the training process of the decision model of the present invention. DETAILED DESCRIPTION

[0032] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0033] In order to achieve the purpose of the present invention, the following embodiments are described:

[0034] Data Storage Module: Decision input data consists of four components: job opening information, candidate information, historical decision-making behavior sequences, and the candidate's current stage in the job process. The output of the decision is the recruiter's next decision regarding the candidate, such as whether to move a candidate who passed the written test stage into the interview stage or to eliminate a candidate who passed the written test stage.

[0035] Job information is represented by "job" and candidate information is represented by "candidate." Generally, these two pieces of data are stored separately in relational databases using multiple tables, or in document databases using formats that support nested structures, such as JSON. The candidate's current stage in the recruitment process, such as the written test or interview, is represented by "stage." The recruiter's next decision regarding the candidate is represented by "action." The recruiter's past decisions are represented by "history," a sequence of actions: history = [(job, candidate, stage, action), ...]. Each element of the sequence is a four-tuple, recording the recruiter's actions at a specific point in time, moving a resume from one process to another within the recruitment process for a specific position. Action sequences are like log records and can be stored in files or document databases.

[0036] 1. Decision model training module:

[0037] a. Decision model: The output of the decision model is expressed as the following conditional probabilities:

[0038] p θ (action|job,candidate,stage,history),;

[0039] The conditional probability above expresses the probability of a candidate applying for a particular job taking the next action at a given stage in the process, given a certain history of recruitment behaviors. For the same job, candidate, and stage, the action will vary depending on the history. Here, represents the parameters of the decision model, which can be trained by inputting a large number of personalized behavior sequences to learn a variety of different recruitment criteria.

[0040] History represents a sequence of behavioral data. Objective data can be used to represent a specific recruitment criterion: the evaluation behavior for a set of job openings and a set of candidates. This behavior sequence can, on the one hand, express the individuality of recruitment criteria, and on the other, the dynamic changes in recruitment strategies. This can be further illustrated using the following scenario: adjusting recruitment decision strategies. For example, when an employer adjusts its strategy and needs to invest in a specific market and recruit a large number of people, it may relax its evaluation criteria. This change in evaluation criteria can be represented by a behavior sequence.

[0041] Without changing the model parameters themselves, inputting different behavior sequences can lead to changes in the evaluation results, achieving a dynamic and adaptive effect. Training and assessing recruiters: Recruiters understand recruitment evaluation criteria based on their own experience. Due to differences in experience, their understanding of evaluation criteria may be inconsistent, leading to personalized evaluation results.

[0042] Training is equivalent to using the behavior sequence data of recruitment experts to guide recruiters to produce evaluation results consistent with the recruitment experts. Assessment is equivalent to collecting each recruiter's own behavior sequence and then calculating the performance difference between these behavior sequences and the behavior sequence of the recruitment experts. The loss function of the decision model is defined as:

[0043]

[0044] Among them, y(action (i) ) is a masking function, that is, when action (i) When y(action (i) )=1; otherwise, y(action (i) )=0.

[0045] Model architecture such as Figure 1As shown in the figure, the job recruitment information and candidate information, after the data preprocessing step, will be arranged into a sequence according to the order of each field; the historical behavior decision sequence is spliced into the current decision behavior sequence as a subsequence based on the expansion of the historical recruitment information and historical candidate information. Therefore, the decision data as a whole is a sequence structure, so any sequence model can be used for modeling. From the perspective of model expression and expansion, the subject can choose the transformer architecture in deep learning. Next, we will further explain how to pre-process and vectorize the data input into the transformer. 2. Data pre-processing module:

[0046] The input data of the decision model needs to go through the following data processing preparations before entering the model:

[0047] i. Information structuring:

[0048] The main purpose of information extraction is to convert the unstructured part of the original information into structured information without losing information as much as possible, and then input the more clearly structured information into the decision model to improve the accuracy, training efficiency and prediction efficiency of the decision model. For example, the original job recruitment information may be represented by text, which needs to be parsed to identify the type, condition combination and nested structure of each field in the job requirements; the original resume in the candidate information may be

[0049] Written in the form of documents such as PDF / Word / pictures, it needs to go through OCR technology and resume file parsing technology to extract relevant structured information.

[0050] a. Job recruitment information:

[0051] After structured processing, job postings are transformed from raw text into a tree structure with nested attributes, typically recorded in JSON format. However, this tree structure is inconvenient for subsequent model building. Therefore, we propose an equivalent flat schema structure representation in key-value format: each key represents the path from the tree to a leaf node, i.e., JSONPath, and the value represents the field type of the leaf node. The schema is as follows:

[0052]

[0053] An example of data filled in this format is as follows:

[0054]

[0055]

[0056] The above representation has two advantages:

[0057] First, the key is represented by a JSON path, which can fully retain the tree structure information;

[0058] Secondly, when the expanded fields are vectorized and embedded (embedding) and certain vectorized operations are performed, they can be run in parallel, thereby speeding up training and prediction efficiency.

[0059] b. Candidate Information:

[0060] Candidate information includes resumes, as well as information about written tests and interviews. Written test and interview information can be represented using final scores, or by leveraging interview parsing techniques to extract interview responses for a richer representation of information. Resumes are structured and extracted using the same approach as described above for job postings, then represented using a flattened tree structure. In particular, since work experience and project experience in resumes are represented as arrays of variable length, they are represented in the flattened schema as follows:

[0061] { ...

[0063] "resume.works[].company":companyName, / / Company name, companyName type;

[0064] The parent type of the companyName type is string, which is represented by the companyName type.

[0065] "resume.works[].job":jobName, / / Job name, jobName type; the jobName type is consistent with the jobName type in the "Job Recruitment Information" schema ...

[0067] } ...

[0069] An example of data filled in this format is as follows:

[0070] {

[0071] "resume.works[1].company":"Alibaba"

[0072] "resume.works[1].job":"Sales Manager" ...

[0074] "resume.works[5].company":"Tencent"

[0075] "resume.works[5].job":"Sales Specialist" ...

[0077] }

[0078] In the above example, to keep the sequence lengths aligned, the array can be truncated. For example, only the five most recent work experiences are considered, and any missing work experience is padded to 5.

[0079] ii. Knowledge graph information enhancement:

[0080] Raw data may lack common sense knowledge, requiring additional common sense data sources, such as knowledge graphs, for information enhancement to provide the most complete input data possible for the decision model. For example, the original job recruitment requirements may include "students graduating from 985, 211, and other universities are preferred." However, the phrases "985" and "211" are actually a way to categorize universities, which is external knowledge. Knowledge graphs are needed to fill in this missing information. Resumes also include work experience at a certain company at a certain time. Companies actually have profile attributes such as industry, size, and region. This information is likely not included in the original resume or may be filled in incorrectly. Knowledge graphs are also needed for information enhancement and error correction. Information enhancement involves adding expanded fields to the main structure in a structured format.

[0081] iii. Tokenization:

[0082] The transformer accepts a sequence of tokens as input. The original transformer architecture processes text sequences, with tokenization being equivalent to word segmentation. The recruitment decision-making behavior sequence considered in the technical solution has a multi-layered structure, requiring further explanation of the specific token processing:

[0083] 1) Position and candidate information are represented in a flattened key-value format. Each field is considered a token. The token type is the field type, and its content consists of two parts: the field name and the field content. The order of the token sequence follows the order of the corresponding keys in the schema.

[0084] 2) The process stage information stage is considered as a token, whose content is the process stage name and type is a discrete category. The value range is selected from the process stage set, for example, {"delivery", "first interview", ..., "onboarding", "elimination"};

[0085] 3) The decision action is considered as a token, the content of which is the name of the process stage, the type is a discrete category, and the value range is consistent with the stage.

[0086] iv. Vectorized embedding:

[0087] The original transformer uses a dictionary lookup (lookup table) to vectorize and embed tokens, and superimposes positional encoding. The vectorization embedding part of the transformer is modified to adapt to the multi-layer structure of the decision action sequence. Different vectorization embedding methods are used for tokens according to their type, for example:

[0088] a. Discrete: Token values are mapped into the range [0, 1024) and then expanded into vectors using a dictionary lookup.

[0089] b. Numeric type integer or float: Normalized to the range [-1, 1] according to the MU-Law, then evenly distributed to the range [1024, 2048), and then expanded into a vector using a dictionary lookup. The MU-Law formula is as follows;

[0090]

[0091] c. String type: Encode the text into a vector by adding a text sequence encoder. A simplified version of the Bertencoder can be used as the text sequence encoder. The model parameters of the text encoder can be trained together with the model parameters of the main transformer, or the parameters can be fixed to improve the overall training speed.

[0092] d. Advanced types, such as companyName and jobName: can be processed in the same way as string types, or entity vectors can be embedded in the knowledge graph in advance, and then the entity vector embedding results can be directly referenced by dictionary lookup. When directly referencing entity vector embedding, the number of embedding parameters is fixed and does not change with transformer training. The information processing flow of recruitment information and candidate information, such as Figure 2 and Figure 3 As shown,

[0093] The decision model training method provided in this embodiment is as follows: If the employer has not used a digital recruitment process before and lacks the necessary decision data accumulation for training, a pre-trained decision model can be used for cold start, and then subsequent fine-tuning training can be performed on its own decision behavior data.

[0094] Specifically, based on the self-supervised pre-training method, the self-supervised pre-training model adopts the same model structure as the decision model, but only resume data is used for training data. In order to simulate the decision-making behavior sequence, the following processing steps are adopted:

[0095] i. Extract the most recent work experience and other information within this period from the resume as the candidate's last job recruitment information; remove other resume information from the most recent work experience and use it as the candidate's resume information when applying for the last job; randomly select and fill in the process except {"Joining", "Elimination"}; the decision action value is "Joining", such as Figure 4 ;

[0096] ii. Cluster all collected resumes based on the similarity of job titles in their most recent work experience;

[0097] iii. For each type of resume, group them by the company where they most recently worked, and sort them by the start time of their most recent job. This way, each group is a decision sequence containing historical decision-making behaviors.

[0098] iv. Since the decision sequence simulated based on resumes only contains positive samples, that is, only entry information and no elimination information, negative sampling technology can be used in each decision sequence of batch training.

[0099] Continuous training based on real behavioral data: If the companies and positions in the simulated data are significantly different from those of the employer, the prediction results of the pre-trained model may not meet the employer's recruitment decision-making standards. In this case, fine-tuning can be performed using real decision-making behavior data based on the pre-trained model.

[0100] Once the data volume and diversity are sufficient and the decision-making effect meets expectations, the model parameters can be fixed and only the model can be used for prediction.

[0101] Furthermore, the decision model prediction process: When the employer uses a pre-trained model or a model fine-tuned based on its own decision data to make predictions, it follows the following process:

[0102] 1) Data storage: Upload candidate information and job recruitment information through the data collection system and store it; collect and store the decision-making behavior of recruiters through the process management system;

[0103] 2) Data processing: The data processing process is consistent with the data processing during training;

[0104] 3) Model prediction: The processed data is input into the model, the probabilities of various possible decision behaviors are output, and the action with the highest probability is selected as the recommended decision behavior for output.

[0105] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. An adaptive recruitment decision-making method based on recruitment behavior data, including a data storage module and a decision model training module, the decision model training module includes a decision model and a data pre-processing module, the data storage module is used to store input data, the data contains job recruitment information, candidate information, historical decision behavior sequence, candidate review status; a relational database or a document database that supports nested structures is used; the data pre-processing module is used to convert the unstructured part of the original information into a structured one; the original text of the basic information is converted into a tree structure with nested attributes, recorded in a schema structure format, in a key-value form: each key represents the path from the tree to the leaf node as jsonpath, and the value represents the field type on the leaf node; the pre-processing module introduces a knowledge graph for data enhancement and information error correction, and adds the amplified fields to the main structure in a structured format; it also includes vectorized embedding, the original transformer The token is vectorized and embedded through a dictionary lookup lookup table, and the position code is superimposed; in the tokenization process, the transformer accepts a sequence with tokens as elements as input; the original transformer architecture processes text sequences; the vectorized embedding includes different vectorized embedding methods for tokens according to their types, including discrete, numerical, string, and advanced types. The discrete type is to map the token value into the range of [0,1024) by mapping, and then expand it into a vector by dictionary lookup. The numerical type is normalized to the interval [-1,1] according to mu-law, and then evenly distributed to the interval [1024,2048), and then expanded into a vector by dictionary lookup; the self-supervised pre-training model adopts a model structure consistent with the decision model, but only resume data is needed for training data to simulate the decision behavior sequence, which is characterized by, The following processing steps are included: S1: Extract the most recent work experience and other information within this period from the resume as the candidate's recruitment information for the previous job; remove other information from the most recent work experience and use it as the candidate's resume information when applying for the previous job; randomly select and fill in the process except for "Joining" and "Elimination"; the decision behavior value is "Joining", S2: Cluster all collected resumes based on the similarity of job titles in the most recent work experience; S3: For each type of resume, group them according to the company where they worked most recently. Within each group, sort them by the start time of their most recent job. This way, each group is a decision sequence containing historical decision-making behaviors. S4: Since the decision sequence simulated based on the resume only has positive samples, that is, only entry information and no elimination information, negative sampling technology is applied to each decision sequence of batch training.

2. The adaptive recruitment decision-making method based on recruitment behavior data according to claim 1, characterized in that: Decision model prediction process: Employers use pre-trained models or models fine-tuned based on their own decision data to make predictions, following the following process: 1) Data storage: Upload candidate information and job recruitment information through the data collection system and store it; collect and store the decision-making behavior of recruiters through the process management system; 2) Data processing: The data processing process is consistent with the data processing during training; 3) Model prediction: The processed data is input into the model, the probabilities of various possible decision behaviors are output, and the action with the highest probability is selected as the recommended decision behavior for output.

Images