Job seeker portrait analysis method and system based on multiple agents

Abstract

The invention provides a job seeker portrait analysis method and system based on multiple agents, and relates to intelligent recruitment, and the method comprises the steps: a data processing agent collects multi-source static attribute data and historical interaction data of a job seeker, and carries out the preprocessing of the data, and obtains the preprocessing data; the job seeker clustering agent performs threshold preliminary screening according to the number of active delivery behaviors on the basis of the preprocessed data, and further divides job seekers through a silence subclass merging and representative selection mechanism; a label identification agent extracts page interaction, a behavior sequence and equipment association data of the job seekers, inputs the implicit demand label identification model, and outputs implicit demand labels of all silent subclass job seekers; the questionnaire intelligent agent identifies a main demand tag based on a hidden demand tag, generates a scenarized questionnaire in combination with a multimedia material library and a generative video model, and realizes adaptive pushing through a decision logic tree; and the adjustment feedback agent pre-adjusts the enterprise post information based on the questionnaire feedback data.

Description

TECHNICAL FIELD

[0001] The present application relates to intelligent recruitment, in particular a job seeker portrait analysis method and system based on multi-agent. BACKGROUND

[0002] With the rapid development of the new generation of information technology industry, especially the breakthrough in the field of artificial intelligence and emerging software services, the digital recruitment platform as a key application scenario in the field of human resources is facing the urgent need for technological upgrading. Among them, in the digital recruitment scene, job seeker portrait analysis as the core link of matching supply and demand is facing multi-dimensional technical bottlenecks, and it is difficult to meet the needs of "accurate matching and efficient interaction" of both parties.

[0003] 1. The governance ability of the multi-source data of the job seeker is insufficient, which leads to low data processing quality and affects the accuracy of subsequent portrait analysis.

[0004] 2. The existing method mostly divides the job seekers by a single threshold such as the number of active delivery behaviors, which leads to significant differences in the potential job seeker willingness and demand of the job seekers, resulting in the missed opportunity of accurate reach of high-value silent groups by enterprises.

[0005] 3. The implicit demand of the job seeker not directly expressed in the prior art cannot be accurately identified, which leads to a large deviation between the output demand and the real demand of the job seeker, greatly reducing the matching degree between enterprises and job seekers.

[0006] Therefore, the present application provides a job seeker portrait analysis method and system based on multi-agent, which solves the problems of insufficient multi-source data governance ability, low precision of job seeker division, and poor recognition of job seeker demand in the prior art, greatly improving the recruitment efficiency of enterprises and the job seeking experience of job seekers. SUMMARY

[0007] In order to solve the above technical problems, the purpose of the present application is to provide a job seeker portrait analysis method and system based on multi-agent, which solves the problems of insufficient multi-source data governance ability, low precision of job seeker division, and poor recognition of job seeker demand in the prior art, greatly improving the recruitment efficiency of enterprises and the job seeking experience of job seekers.

[0008] In order to achieve the above purpose, the present application provides the following technical solution: a job seeker portrait analysis method based on multi-agent, the method comprising: The data processing agent collects multi-source static attribute data and historical interaction data of the job seeker, fills in the missing values through attribute similarity clustering algorithm, and filters and removes the interaction data noise based on attribute importance quantification combined with time threshold, to obtain preprocessed data; The job seeker clustering agent uses preprocessed data to perform initial screening based on the number of active job applications. Through the merging of silent subclasses and the selection of representatives, job seekers are divided into four categories: actively seeking jobs, high-potential jobs, information gathering jobs, and deeply observing jobs. The tag recognition agent extracts job seekers' page interactions, behavior sequences, and device-related data, inputs them into a latent demand tag recognition model trained through hyperparameter iteration optimization, and outputs latent demand tags for each silent subclass of job seekers. The questionnaire agent performs hierarchical weight analysis based on implicit demand tags, identifies the main demand tags, generates scenario-based questionnaires by combining a multimedia material library and a generative video model, and achieves adaptive push through a decision logic tree. The adjustment feedback agent pre-adjusts the company's job information based on questionnaire feedback data and sends a confirmation request to the company.

[0009] Preferably, the step of filling missing values ​​using an attribute similarity clustering algorithm and filtering out noise in interactive data based on attribute importance quantification and time thresholds to obtain preprocessed data includes: Based on various multi-source static attribute data, and based on a preset attribute similarity clustering algorithm, job seekers with similar attributes are grouped into the same group, and the missing values ​​of the corresponding multi-source static attribute data are filled in by the distribution characteristics of the multi-source static attribute data within the corresponding group. Based on the historical interaction data, and using the attribute importance quantification method, the importance quantification value of different interaction pages or behavior sequences relative to the overall behavior set is calculated. Combined with the timestamp of the behavior, and based on the time threshold filtering strategy, interaction data noise caused by short-term and random user behavior is removed to obtain the corresponding preprocessed data.

[0010] Preferably, based on preprocessed data, a threshold-based initial screening is performed according to the number of proactive application behaviors. Through a mechanism of merging silent subcategories and selecting representatives, job seekers are divided into four categories: actively seeking employment, high-potential candidates, information gathering candidates, and those deeply observing the market. S1. Based on the preprocessed multi-source static attribute data and historical interaction data, the number of times each job seeker actively applied for jobs is counted; job seekers whose number of active applications exceeds the threshold for the first application are classified as active job seekers, and the remaining job seekers are initially classified as silent job seekers. S2. For each job seeker in the silent job seeker category, calculate the difference in the number of proactive applications between each job seeker, determine the job seekers with the smallest and largest differences in the number of proactive applications for each job seeker, and associate them with the first job seeker and the second job seeker, respectively. S3. Based on the first and second job seekers associated with each job seeker, classify each job seeker and their associated first job seeker into the same silent subclass, forming the first set of silent subclasses. In the first set of silent subclasses, if the absolute value of the difference between the number of active applications of a job seeker and the first job seeker and the difference between the number of active applications of the first job seeker and the second job seeker is less than a preset difference threshold, then classify the job seeker into the silent subclass corresponding to the second job seeker. S4. Randomly select one job seeker from the first batch of silent subclasses as the representative of each silent subclass; S5. Based on the preset job silence similarity calculation formula, calculate the job silence similarity between each representative in the first batch of silent subclasses, find the two representatives with the highest job silence similarity, merge the silent subclasses to which the two representatives belong to form a new silent subclass. A random number between 0 and 1 is generated. If the random number is greater than the random threshold, the representatives of the other silent subclasses will not change. The new representative of the silent subclass will be the job seeker whose job-seeking silence is closest to that of the two original representatives. Conversely, if the result is not met, S4 is executed again to randomly select representatives until the number of clusters of the remaining silent subclasses meets the requirements, thus forming the high-potential class, information gathering class, and deep observation class corresponding to the silent job seeker class.

[0011] Preferably, the training process of the implicit demand label recognition model trained through hyperparameter iterative optimization includes: Step 1: Collect historical silent data for each of the three categories of silent job seekers: high-potential, information gathering, and deep observation. The historical silent data includes historical page interaction data, historical behavior sequence data, and historical device association data. According to a preset division ratio, the corresponding historical silent data is divided into training set and test set, and the historical silent data is annotated with the labeled content. Step 2: Construct an implicit demand label recognition model and determine the hyperparameter set corresponding to the implicit demand label recognition model and the value range of each hyperparameter. Step 3: Based on the value range of each hyperparameter, randomly generate several hyperparameter groups; Step 4: Input the data in the training set into the implicit demand label recognition model corresponding to each hyperparameter group, and output the historical prediction data corresponding to each hyperparameter group. Step 5: Based on historical prediction data and corresponding standard content, and using the preset hyperparameter group rationality calculation formula, calculate the rationality of each hyperparameter group and arrange them in descending order to form a rationality ranking list for each hyperparameter group. Step 6: Based on the reasonableness ranking list, select all hyperparameter groups ranked before the preset first ranking and place them in the update library; select all hyperparameter groups ranked after the preset second ranking and place them in the backup library. Step 7: Select several hyperparameter groups from the backup library as representative groups, and select the hyperparameter group with the highest rationality from the update library as the invariant group; Step 8: Pair the representative group with the remaining hyperparameter groups in the update library and randomly swap several hyperparameters to form a new hyperparameter group. Integrate the new hyperparameter group with the invariant group to form the hyperparameter group corresponding to the next loop. Repeat step 4 until the number of loops meets the preset loop threshold. At this time, output the implicit demand label recognition model corresponding to the invariant group as the pre-trained implicit demand label recognition model.

[0012] Preferably, the questionnaire agent performs hierarchical weight analysis based on implicit demand tags to identify key demand tags, including: The questionnaire agent extracts and identifies the implicit demand tags of silent job seekers output by the agent. The tags are stratified into three levels: career development, job security, and work environment. The high-frequency tags corresponding to each level are counted, and cross-level strong correlation tags are mined based on the counted high-frequency tags and included in the candidate pool as candidate tags. Based on the job seekers in the candidate pool, calculate the tag coverage and behavior intensity value corresponding to each candidate tag in the candidate pool, and determine the basic weight corresponding to each candidate tag in the candidate pool. Based on the three subcategories of the silent job seeker category—high-potential, information gathering, and deep observation—a subcategory adaptation coefficient is set for each candidate tag in the candidate pool. Based on the subcategory adaptation coefficient and basic weight of each candidate tag, the adjustment weight of each candidate tag under the corresponding subcategory is determined as the final dynamic weight of the corresponding candidate tag. Based on a preset benchmark threshold, candidate tags with a final dynamic weight higher than the benchmark threshold are selected from the candidate pool and used as the main tags to form a set of main tags to be selected. A hierarchical diversity test is performed on each candidate main tag. After the diversity test, the candidate main tags are sorted from high to low according to dynamic weight, and the top five candidate main tags are selected as the main implicit requirement tags.

[0013] Preferably, the step of generating a scenario-based questionnaire by combining a multimedia resource library with a generative video model includes: The questionnaire agent, based on the identified main implicit demand tags and a pre-set multimedia material library, matches core scene elements related to each main implicit demand tag from the multimedia material library to form a core scene element set. Based on the core scene element set, the various elements in the core scene element set are combined through a pre-trained generative video model to form an interactive storyline script that includes characters, simple plots and potential conflict points. The script immerses job seekers in an interactive storyline, automatically pausing at key plot points and popping up embedded questions corresponding to the current key request turning point and the main implicit demand tags. Based on the job seeker's answers to the embedded questions, the presentation order, depth, and specific content of subsequent embedded questions are dynamically determined through a pre-set decision logic tree, thus achieving an adaptive questionnaire flow.

[0014] Preferably, the implementation of the adaptive questionnaire flow includes: Several embedded questions at the beginning of the questionnaire are set as decision nodes, and a decision logic tree is pre-defined to map different combinations of answers to subsequent embedded questions. When the decision logic tree determines that a new embedded question needs to be loaded, it retrieves the ontology-standardized question library in real time to generate an embedded question set that conforms to the current context. It has a preset maximum number of questions threshold and estimates and displays the remaining time for job seekers to answer questions in real time.

[0015] A second aspect of the present invention also provides a job seeker profiling analysis system based on multiple agents, including a data processing agent, a job seeker clustering agent, a tag recognition agent, a questionnaire agent, and an adjustment feedback agent; The data processing intelligent agent is used to collect multi-source static attribute data and historical interaction data of job seekers, fill missing values ​​through attribute similarity clustering algorithm, and filter out noise in interaction data based on attribute importance quantification and time threshold to obtain preprocessed data. The job seeker clustering intelligence is used to perform initial screening based on preprocessed data and the number of active application behaviors, and to classify job seekers into active job seekers, high-potential job seekers, information-gathering job seekers, and deeply observing job seekers through silent subclass merging and representative selection mechanisms. The tag recognition intelligent agent is used to extract job seekers' page interactions, behavior sequences and device-related data, input the implicit demand tag recognition model trained by hyperparameter iteration optimization, and output the implicit demand tags of each silent subclass of job seekers. The questionnaire agent is used to perform hierarchical weight analysis based on implicit demand tags, identify the main demand tags, generate scenario-based questionnaires by combining a multimedia material library and a generative video model, and achieve adaptive push through a decision logic tree. The adjustment feedback agent is used to pre-adjust the company's job information based on questionnaire feedback data and output a confirmation request to the company.

[0016] Compared with the prior art, the beneficial effects of the present invention are: This invention presents a multi-agent-based job seeker profiling analysis system that effectively addresses three major technical bottlenecks in digital recruitment scenarios through highly specialized agent division of labor and collaboration. This solution integrates five core agents, each with unique functions yet working collaboratively to form a complete processing ecosystem. The data processing agent is responsible for collecting and cleaning raw data; the job seeker clustering agent performs preliminary classification; the tag recognition agent uncovers deeper needs; the questionnaire agent enables interactive verification; and the adjustment feedback agent completes final calibration and output. This specialized division of labor and streamlined operation significantly improves the intelligence level of recruitment platforms. Compared to the simple matching model of traditional recruitment platforms based on "keywords + tags," this system achieves intelligent processing throughout the entire process through multi-agent collaboration.

[0017] Traditional recruitment platforms typically rely on a single dimension (such as the number of applications submitted) to roughly segment job seekers. This method fails to capture the diversity and complexity of the job seeker group, causing companies to miss out on a large number of high-value, unseen talents. The job seeker clustering intelligence in this invention achieves refined segmentation of the job seeker group through a multi-stage process involving threshold screening, merging of unseen subclasses, and representative selection mechanisms, significantly improving the accuracy and practicality of the classification.

[0018] The core innovation of this system lies in its in-depth analysis of the silent job seeker group. Traditional methods typically categorize low-activity job seekers simply as an "inactive group," while this system establishes a correlation network by calculating the difference in the number of proactive job applications among job seekers, thereby identifying subgroups with similar characteristics but different behaviors. Specifically, the system first filters out the active job seeker group based on the criterion that the number of proactive job applications exceeds a threshold; then, it conducts in-depth analysis of the remaining silent job seeker group, establishing correlations by calculating behavioral differences among job seekers.

[0019] The representative selection and subclass merging mechanism is another innovation of this invention. The system randomly selects representatives from each silent subclass and calculates the job-seeking silence similarity among different representatives. Subclasses with high similarity are merged, and this process is iterated until a stable classification result is formed. Ultimately, the system accurately classifies silent job seekers into three categories: high-potential, information-gathering, and deeply observing. This refined classification helps recruiting companies develop differentiated outreach strategies for different groups, significantly improving talent acquisition efficiency. This invention's four-quadrant classification of job seekers surpasses the simple classification based on a single behavioral indicator used by traditional recruitment platforms, providing a technological possibility for value mining of silent job seeker groups. Practice shows that this refined classification can help companies identify those "inactive but high-quality" talents, broadening and deepening the talent pool, and effectively solving the structural contradictions of "looking but not applying" and "applying but not responding" in recruitment.

[0020] In the recruitment field, the implicit needs that job seekers fail to express directly are a key factor affecting the accuracy of job matching. Traditional technologies mainly rely on keyword matching and explicit label filtering, which cannot effectively identify the true intentions behind job seekers' surface behavior. The label recognition agent in this invention achieves in-depth mining and accurate identification of job seekers' implicit needs through multi-dimensional data fusion and adaptive model optimization. This optimization mechanism enables the model to adapt to the behavioral characteristics of different job seeker groups, avoiding overfitting or underfitting problems and ensuring the accuracy of the recognition results. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0022] Figure 1 This is a schematic diagram of a multi-agent-based job seeker profiling analysis method.

[0023] Figure 2 This is a schematic diagram of a job seeker profiling and analysis system based on multi-agent intelligence. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be described in detail below. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other implementation methods obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0025] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0026] Example 1: This invention provides, for example Figure 1 The multi-agent-based job seeker profiling analysis method shown includes the following steps: The data processing agent collects multi-source static attribute data and historical interaction data of job seekers, fills in missing values ​​through an attribute similarity clustering algorithm, and filters out noise in the interaction data based on attribute importance quantification combined with time thresholds to obtain preprocessed data. In this embodiment, the data processing agent collects publicly available resume data of job seekers in batches through API interfaces (such as education, years of work experience, and skill certificates), and obtains social network information (such as GitHub projects and LinkedIn contacts) through the authorization of job seekers. It uses semantic parsing technology to extract structured attribute fields to form corresponding multi-source static attribute data.

[0027] For historical interaction data, a front-end event tracking SDK is deployed to capture the fine-grained behavioral sequences of job seekers on the recruitment platform (such as the duration of a single page stay and whether they revisit the same job posting after submitting their resume), and device fingerprinting technology is used to link the behavioral data of the same user across different devices.

[0028] It should be noted that the process of filling missing values ​​using an attribute similarity clustering algorithm and filtering out noise in interactive data based on attribute importance quantification and time thresholds to obtain preprocessed data includes: Based on various multi-source static attribute data, and using a preset attribute similarity clustering algorithm, job seekers with similar attributes are grouped into the same group. Furthermore, missing values ​​in the corresponding multi-source static attribute data are filled in using the distribution characteristics of the multi-source static attribute data within each group. In this embodiment, the clustering process of the attribute similarity clustering algorithm is as follows: Define the job seeker's attribute vector, which includes numerical attributes (such as years of work experience), categorical attributes (such as industry), and sequential attributes (such as skill tag sequence). The improved Gower similarity coefficient was used to calculate the similarity between two job seekers i and j; The similarity calculation formula is as follows: ; in, For numerical attribute similarity, ; For categorical attribute similarity, ; For sequence-type attribute similarity; These are the weight values ​​of the corresponding attributes; Job seekers were divided into k clusters using a clustering algorithm. This maximizes the average similarity within the cluster.

[0029] For clusters Missing attributes of a job seeker i : like If it is a numeric type, then the fill value is a cluster. All non-missing elements The weighted average of the values, with weights of 1 / 2. .

[0030] like If it is a categorical type, then the fill value is a cluster. All non-missing elements The mode of the value.

[0031] Based on the historical interaction data, and using an attribute importance quantification method, the importance quantification value of different interaction pages or behavior sequences relative to the overall behavior set is calculated. Combined with the timestamp of the behavior, a time threshold filtering strategy is used to remove noise data generated by short-term, random user behavior. In this embodiment, the importance quantification value of behavior b is defined. : ;in, For overall importance, assessment is performed through the mutual information between behavior b and key events. For personalized influence, it represents the degree to which the probability of a key event occurring increases in the short term after behavior b occurs. This is the corresponding adjustment coefficient.

[0032] Define the time decay weight of behavior b ; The noise judgment rule is as follows: If satisfied If it is detected as noise, it will be filtered out. This is the preset noise threshold.

[0033] The job seeker clustering agent uses preprocessed data to perform initial screening based on the number of active job applications. Through the merging of silent subclasses and the selection of representatives, job seekers are divided into four categories: actively seeking jobs, high-potential jobs, information gathering jobs, and deeply observing jobs. It should be noted that, based on preprocessed data, a threshold screening is performed according to the number of proactive applications. Through a mechanism of merging silent subcategories and selecting representatives, job seekers are divided into four categories: actively seeking employment, high-potential applicants, information gatherers, and those in a state of deep observation. S1. Based on the preprocessed multi-source static attribute data and historical interaction data, the number of times each job seeker actively applied for jobs is counted; job seekers whose number of active applications exceeds the threshold for the first application are classified as active job seekers, and the remaining job seekers are initially classified as silent job seekers. S2. For each job seeker in the silent job-seeking category, calculate the difference in the number of proactive applications among them, and determine the job seekers with the smallest and largest differences in the number of proactive applications, respectively denoted as the first job seeker and the second job seeker, and associate them with the job seekers mentioned above; In this embodiment, for each job seeker p in the silent category, calculate the absolute difference in the number of proactive applications between p and all other job seekers q in the category, denoted as p. ; Find two key neighbors for job seeker p (the first job seeker and the second job seeker): The first job seeker is the nearest neighbor. :satisfy The smallest, i.e., the job seeker whose number of applications is closest to p.

[0034] The second job seeker is the furthest neighbor. :satisfy The largest, i.e., the job seeker whose number of applications differs most from p.

[0035] Compare job seeker p with its nearest neighbors They are grouped into the same initial subclass. After iterating through all job seekers, an initial set of subclasses is formed. .

[0036] In addition, this embodiment also makes minor adjustments to the subclass members: For each job seeker p, calculate an adjustment decision value. ; once If the value is less than the preset adjustment threshold, then job seeker p will be moved to its farthest neighbor. The subclass in which it belongs. Otherwise, p is retained in the atomic class.

[0037] S3. Based on the first and second job seekers associated with each job seeker, classify each job seeker and their associated first job seeker into the same silent subclass, forming the first set of silent subclasses. In the first set of silent subclasses, if the absolute value of the difference between the number of active applications of a job seeker and the first job seeker and the difference between the number of active applications of the first job seeker and the second job seeker is less than a preset difference threshold, then classify the job seeker into the silent subclass corresponding to the second job seeker. S4. Randomly select one job seeker from the first batch of silent subclasses as the representative for each silent subclass; in this embodiment, the selection strategy for the representative is as follows: For each subclass Calculate the centrality index for each member p in the class. : ; In the formula, represents the centrality value; the smaller the sum of distances from a member to all other members within the class, the greater its centrality value. (Selecting subclasses) Centrality The largest member, acting as the representative of that subclass. .

[0038] S5. Based on the preset job silence similarity calculation formula, calculate the job silence similarity between each representative in the first batch of silent subclasses, find the two representatives with the highest job silence similarity, merge the silent subclasses to which the two representatives belong to form a new silent subclass. In this embodiment, all data corresponding to each representative in the first batch of silent subclasses is quantized into an n-dimensional feature vector, such as the representative. and representatives The corresponding feature vector is: ; ; Representative and representatives The formula for calculating the job-seeking silence similarity between two individuals is: ; In the formula, For vectors and dot product, and They are vectors and The length of the module.

[0039] A random number between 0 and 1 is generated. If the random number is greater than the random threshold, the representatives of the other silent subclasses will not change. The new representative of the silent subclass will be the job seeker whose job-seeking silence is closest to that of the two original representatives. Conversely, if the result is not met, S4 is executed again to randomly select representatives until the number of clusters of the remaining silent subclasses meets the requirements, thus forming the high-potential class, information gathering class, and deep observation class corresponding to the silent job seeker class.

[0040] The tag recognition agent extracts job seekers' page interactions, behavior sequences, and device-related data, inputs them into a latent demand tag recognition model trained through hyperparameter iteration optimization, and outputs latent demand tags for each silent subclass of job seekers. It should be noted that the training process of the implicit demand label recognition model trained through hyperparameter iterative optimization includes: Step 1: Collect historical silent data for each of the three categories of silent job seekers: high-potential, information gathering, and deep observation. The historical silent data includes historical page interaction data, historical behavior sequence data, and historical device association data. According to a preset division ratio, the corresponding historical silent data is divided into training set and test set, and the historical silent data is annotated with the labeled content. Step 2: Construct an implicit demand label recognition model and determine the hyperparameter set corresponding to the implicit demand label recognition model and the value range of each hyperparameter. Step 3: Based on the value range of each hyperparameter, randomly generate several hyperparameter groups; Step 4: Input the data in the training set into the implicit demand label recognition model corresponding to each hyperparameter group, and output the historical prediction data corresponding to each hyperparameter group. Step 5: Based on historical prediction data and corresponding standard content, and using a preset hyperparameter group rationality calculation formula, calculate the rationality of each hyperparameter group, and arrange them in descending order to form a rationality ranking list for each hyperparameter group; in this embodiment, the hyperparameter group rationality calculation formula is: ; In the formula, The accuracy of the training set. For the F1 scores of the training set, The mean absolute error of the training set. A random penalty term ranging from 0 to 1.

[0041] Step 6: Based on the reasonableness ranking list, select all hyperparameter groups ranked before the preset first ranking and place them in the update library; select all hyperparameter groups ranked after the preset second ranking and place them in the backup library. Step 7: Select several hyperparameter groups from the backup library as representative groups, and select the hyperparameter group with the highest rationality from the update library as the invariant group; Step 8: Pair the representative group with the remaining hyperparameter groups in the update library, and randomly swap several hyperparameters to form a new hyperparameter group. Integrate the new hyperparameter group with the invariant group to form the hyperparameter group corresponding to the next loop. Repeat step 4 until the number of loops meets the preset loop threshold. At this point, the implicit demand label recognition model corresponding to the invariant group is output as the pre-trained implicit demand label recognition model. In this embodiment, the process of forming the new hyperparameter group is as follows: After pairing the representative with the remaining hyperparameter groups in the update library, a random intersection point is selected, and some hyperparameters after the intersection point are swapped to generate corresponding offspring, forming new hyperparameter groups. For each hyperparameter in the offspring, mutation probability is used to determine whether to mutate; if the mutation probability is greater than a preset probability threshold, mutation is performed.

[0042] During mutation, a random perturbation is applied within the range of values ​​for the hyperparameter. For continuous hyperparameters, a Gaussian perturbation is commonly used (i.e., adding Gaussian noise with a mean of 0 and a standard deviation of σ, and then truncating to the range). For discrete hyperparameters, they can be randomly reset to a value within the range.

[0043] The questionnaire agent performs hierarchical weight analysis based on implicit demand tags, identifies the main demand tags, generates scenario-based questionnaires by combining a multimedia material library and a generative video model, and achieves adaptive push through a decision logic tree. It should be noted that the questionnaire agent performs hierarchical weight analysis based on implicit demand tags to identify key demand tags, including: The questionnaire agent extracts and identifies the implicit demand tags of silent job seekers output by the agent. The tags are stratified into three levels: career development, job security, and work environment. The high-frequency tags corresponding to each level are counted, and cross-level strong correlation tags are mined based on the counted high-frequency tags and included in the candidate pool as candidate tags. Based on the job seekers in the candidate pool, calculate the tag coverage and behavior intensity value corresponding to each candidate tag in the candidate pool, and determine the basic weight corresponding to each candidate tag in the candidate pool. Based on the three subcategories of the silent job seeker category—high-potential, information gathering, and deep observation—a subcategory adaptation coefficient is set for each candidate tag in the candidate pool. Based on the subcategory adaptation coefficient and basic weight of each candidate tag, the adjustment weight of each candidate tag under the corresponding subcategory is determined as the final dynamic weight of the corresponding candidate tag. Based on a preset benchmark threshold, candidate tags with a final dynamic weight higher than the benchmark threshold are selected from the candidate pool and used as the main tags to form a set of main tags to be selected. A hierarchical diversity test is performed on each candidate main tag. After the diversity test, the candidate main tags are sorted from high to low according to dynamic weight, and the top five candidate main tags are selected as the main implicit requirement tags.

[0044] It should also be noted that the process of generating scenario-based questionnaires by combining a multimedia resource library with a generative video model includes: The questionnaire agent, based on the identified main implicit demand tags and a pre-set multimedia material library, matches core scene elements related to each main implicit demand tag from the multimedia material library to form a core scene element set. In this embodiment, a multimodal requirement-element mapping is achieved by setting a three-dimensional mapping matrix, and the specific element matching degree is calculated using multi-dimensional methods: This forms a core set of scene elements; , , This represents the corresponding weight value.

[0045] Specifically, semantic relevance uses the BERT model to calculate the cosine similarity between the tag text and the element description text; scene fit is based on the knowledge graph to calculate the association strength between the element and the professional scene; sentiment consistency analyzes the matching degree between the sentiment polarity (positive / neutral / negative) of the element and the expected sentiment of the tag; at the same time, a time decay factor is introduced to dynamically adjust the element weights, and the element library weights are updated based on real-time user feedback.

[0046] Based on a core set of scene elements, a pre-trained generative video model combines these elements to form an interactive storyline script containing characters, simple plot points, and potential conflict points. In this embodiment, a three-level rendering optimization strategy is employed during script generation: a base layer pre-renders static backgrounds and general actions, completed before the video generation process begins or during the initialization phase. This layer primarily renders elements that are unlikely to change in subsequent interactions, such as static backgrounds like streets, offices, and natural scenery; and general recurring actions like walking and sitting. Utilizing the parallel computing capabilities of the GPU, these general scenes and actions are pre-rendered in batches and stored in a "pre-rendered material pool." When video generation is needed, these high-quality materials are directly called, avoiding redundant calculations.

[0047] The dynamic layer generates personalized plot branches in real time and runs continuously. Based on the user's primary implicit needs (such as "pursuing work-life balance"), it dynamically generates personalized plot branches and scene content against the background of the base layer. For example, it generates a scene of "the protagonist calmly leaving work at dusk" for a user with the aforementioned tags, rather than a scene of "working overtime late at night." It employs computationally efficient video generation models. These models are based on text instructions (i.e., needs tags) and focus on generating specific actions, facial expressions, or adjustments to local scenes for foreground characters without requiring full-frame re-rendering, significantly reducing computational overhead. The interaction layer dynamically inserts specific user information at decision points. In the video stream generated by the dynamic layer, at preset key decision points (such as when the protagonist faces two choices), the video pauses. At this time, the interaction layer seamlessly overlays the user's personal information (such as name, selected answer) onto the current video frame in real time using visual elements (such as customized text prompts, virtual identification, etc.).

[0048] The script immerses job seekers in an interactive storyline, automatically pausing at key plot points and popping up embedded questions corresponding to the current key request turning point and the main implicit demand tags. Based on the job seeker's answers to the embedded questions, the presentation order, depth, and specific content of subsequent embedded questions are dynamically determined through a pre-set decision logic tree, thus achieving an adaptive questionnaire flow.

[0049] The implementation of the adaptive questionnaire flow includes: Several embedded questions at the beginning of the questionnaire are set as decision nodes, and a decision logic tree is pre-defined to map different answer combinations to subsequent embedded questions. In this embodiment, a three-level decision node architecture is adopted. The first-level node sets 3 to 5 basic questions to determine the job seeker's main needs. The second-level node delves into the details of specific fields based on the answers of the first-level node. The third-level node is used for cross-validation and priority ranking.

[0050] A weighted multidimensional logic tree construction method is adopted, and the node selection probability is calculated using the formula: Node selection probability = Basic weight × Timeliness factor × Personalization coefficient. Here, the basic weight is preset by domain experts, the timeliness factor is dynamically adjusted based on the recent usage effect of the issue, and the personalization coefficient is calculated based on the user profile matching degree.

[0051] When the decision logic tree determines that a new embedded question needs to be loaded, it retrieves the ontology-standardized question library in real time to generate an embedded question set that conforms to the current context. In this embodiment, the ontology-standardized question library includes the following dimensions: The system employs a semantic dimension based on knowledge graph-based question classification; a contextual dimension considering the relevance to specific professional scenarios; a difficulty dimension assessing the cognitive complexity of the question; and a timeliness dimension defining the effective timeframe for the question. The question generation score is calculated as follows: Semantic Relevance × 0.4 + Contextual Matching × 0.3 + Difficulty Suitability × 0.2 + Timeliness Compliance × 0.1. Each parameter is calculated in real-time, and the question with the highest score is selected for loading.

[0052] The system has a preset maximum question count threshold and estimates and displays the remaining completion time corresponding to the job seeker's answers in real time. In this embodiment, the prediction formula for the remaining completion time is: Remaining time prediction = Σ (Question baseline time × Personalization coefficient) + Buffer time; where the question baseline time is set based on the question type and length, the personalization coefficient is dynamically adjusted according to the user's answering speed and accuracy, and the buffer time reserves 10% to 15% of flexible time.

[0053] The adjustment feedback agent pre-adjusts the company's job information based on questionnaire feedback data and sends a confirmation request to the company.

[0054] This embodiment also discloses a job seeker profiling analysis system based on multiple agents, including a data processing agent, a job seeker clustering agent, a tag recognition agent, a questionnaire agent, and an adjustment feedback agent; The data processing intelligent agent is used to collect multi-source static attribute data and historical interaction data of job seekers, fill missing values ​​through attribute similarity clustering algorithm, and filter out noise in interaction data based on attribute importance quantification and time threshold to obtain preprocessed data. The job seeker clustering intelligence is used to perform initial screening based on preprocessed data and the number of active application behaviors, and to classify job seekers into active job seekers, high-potential job seekers, information-gathering job seekers, and deeply observing job seekers through silent subclass merging and representative selection mechanisms. The tag recognition intelligent agent is used to extract job seekers' page interactions, behavior sequences and device-related data, input the implicit demand tag recognition model trained by hyperparameter iteration optimization, and output the implicit demand tags of each silent subclass of job seekers. The questionnaire agent is used to perform hierarchical weight analysis based on implicit demand tags, identify the main demand tags, generate scenario-based questionnaires by combining a multimedia material library and a generative video model, and achieve adaptive push through a decision logic tree. The adjustment feedback agent is used to pre-adjust the company's job information based on questionnaire feedback data and output a confirmation request to the company.

[0055] Optionally, in this embodiment, those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing the hardware related to the terminal device. The program can be stored in a computer-readable storage medium, which may include: flash drive, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0056] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0057] If the integrated units in the above embodiments are implemented as software functional units and sold or used as independent products, they can be stored in the aforementioned computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause one or more electronic devices to execute all or part of the steps of the methods described in the various embodiments of this application.

[0058] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0059] In the several embodiments provided in this application, it should be understood that the disclosed application can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection of units or modules may be electrical or other forms.

[0060] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0061] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0062] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A job seeker profiling analysis method based on multi-agent intelligence, characterized in that, The method includes: The data processing intelligent agent collects multi-source static attribute data and historical interaction data of job seekers, fills missing values ​​through attribute similarity clustering algorithm, and filters out noise in interaction data based on attribute importance quantification and time threshold to obtain preprocessed data. The job seeker clustering agent uses preprocessed data to perform initial screening based on the number of times a job seeker actively submits applications. Through a mechanism of merging silent subclasses and selecting representatives, job seekers are divided into four categories: actively seeking jobs, high-potential jobs, information gathering jobs, and deeply observing jobs. The tag recognition agent extracts job seekers' page interactions, behavior sequences, and device-related data, inputs them into a latent demand tag recognition model trained through hyperparameter iteration optimization, and outputs latent demand tags for each silent subclass of job seekers. The questionnaire agent performs hierarchical weight analysis based on implicit demand tags, identifies the main demand tags, generates scenario-based questionnaires by combining a multimedia material library and a generative video model, and achieves adaptive push through a decision logic tree. The adjustment feedback agent pre-adjusts the company's job information based on questionnaire feedback data and sends a confirmation request to the company.

2. The job seeker profiling analysis method based on multi-agent technology according to claim 1, characterized in that, The process of filling missing values ​​using an attribute similarity clustering algorithm and filtering out noise in interactive data based on attribute importance quantification and time thresholds to obtain preprocessed data includes: Based on various multi-source static attribute data, and based on a preset attribute similarity clustering algorithm, job seekers with similar attributes are grouped into the same group, and the missing values ​​of the corresponding multi-source static attribute data are filled by the distribution characteristics of the multi-source static attribute data within the corresponding group. Based on the historical interaction data, and using the attribute importance quantification method, the importance quantification value of different interaction pages or behavior sequences relative to the overall behavior set is calculated. Combined with the timestamp of the behavior, and based on the time threshold filtering strategy, noisy data caused by short-term and random user behavior is removed.

3. The job seeker profiling analysis method based on multi-agent technology according to claim 2, characterized in that, Based on preprocessed data, a threshold-based initial screening is performed according to the number of proactive application behaviors. Through a mechanism of merging silent subcategories and selecting representatives, job seekers are divided into four categories: actively seeking employment, high-potential candidates, information gathering candidates, and those deeply observing the situation. S1. Based on the preprocessed multi-source static attribute data and historical interaction data, the number of times each job seeker actively applied for jobs is counted; job seekers whose number of active applications exceeds the threshold for the first application are classified as active job seekers, and the remaining job seekers are initially classified as silent job seekers. S2. For each job seeker in the silent job seeker category, calculate the difference in the number of active application behaviors between each job seeker, determine the job seekers with the smallest and largest differences in the number of active application behaviors for each job seeker, and associate them with the first job seeker and the second job seeker, respectively. S3. Based on the first and second job seekers associated with each job seeker, classify each job seeker and their associated first job seeker into the same silent subclass, forming the first set of silent subclasses. In the first set of silent subclasses, if the absolute value of the difference between the number of active applications of a job seeker and the first job seeker and the difference between the number of active applications of the first job seeker and the second job seeker is less than a preset difference threshold, then classify the job seeker into the silent subclass corresponding to the second job seeker. S4. Randomly select one job seeker from the first batch of silent subclasses as the representative of each silent subclass; S5. Based on the preset job silence similarity calculation formula, calculate the job silence similarity between each representative in the first batch of silent subclasses, find the two representatives with the highest job silence similarity, merge the silent subclasses to which the two representatives belong to form a new silent subclass. A random number between 0 and 1 is generated. If the random number is greater than the random threshold, the representatives of the other silent subclasses will not change. The new representative of the silent subclass will be the job seeker whose job-seeking silence is closest to that of the two original representatives. Conversely, if the result is not met, S4 is executed again to randomly select representatives until the number of clusters of the remaining silent subclasses meets the requirements, thus forming the high-potential class, information gathering class, and deep observation class corresponding to the silent job seeker class.

4. The job seeker profiling analysis method based on multi-agent technology according to claim 3, characterized in that, The training process of the implicit demand label recognition model trained through hyperparameter iterative optimization includes: Step 1: Collect historical silent data for each of the three categories of silent job seekers: high-potential, information gathering, and deep observation. The historical silent data includes historical page interaction data, historical behavior sequence data, and historical device association data. According to a preset division ratio, the corresponding historical silent data is divided into training set and test set, and the historical silent data is annotated with the labeled content. Step 2: Construct an implicit demand label recognition model and determine the hyperparameter set corresponding to the implicit demand label recognition model and the value range of each hyperparameter. Step 3: Based on the value range of each hyperparameter, randomly generate several hyperparameter groups; Step 4: Input the data in the training set into the implicit demand label recognition model corresponding to each hyperparameter group, and output the historical prediction data corresponding to each hyperparameter group. Step 5: Based on historical prediction data and corresponding standard content, and using the preset hyperparameter group rationality calculation formula, calculate the rationality of each hyperparameter group and arrange them in descending order to form a rationality ranking list for each hyperparameter group. Step 6: Based on the reasonableness ranking list, select all hyperparameter groups ranked before the preset first ranking and place them in the update library; select all hyperparameter groups ranked after the preset second ranking and place them in the backup library. Step 7: Select several hyperparameter groups from the backup library as representative groups, and select the hyperparameter group with the highest rationality from the update library as the invariant group; Step 8: Pair the representative group with the remaining hyperparameter groups in the update library and randomly swap several hyperparameters to form a new hyperparameter group. Integrate the new hyperparameter group with the invariant group to form the hyperparameter group corresponding to the next loop. Repeat step 4 until the number of loops meets the preset loop threshold. At this time, output the implicit demand label recognition model corresponding to the invariant group as the pre-trained implicit demand label recognition model.

5. The job seeker profiling analysis method based on multi-agent technology according to claim 4, characterized in that, The questionnaire agent performs hierarchical weight analysis based on implicit demand tags to identify key demand tags, including: The questionnaire agent extracts and identifies the implicit demand tags of silent job seekers output by the agent. The tags are stratified into three levels: career development, job security, and work environment. The high-frequency tags corresponding to each level are counted, and cross-level strong correlation tags are mined based on the counted high-frequency tags and included in the candidate pool as candidate tags. Based on the job seekers in the candidate pool, calculate the tag coverage and behavior intensity value corresponding to each candidate tag in the candidate pool, and determine the basic weight corresponding to each candidate tag in the candidate pool. Based on the three subcategories of the silent job seeker category—high-potential, information gathering, and deep observation—a subcategory adaptation coefficient is set for each candidate tag in the candidate pool. Based on the subcategory adaptation coefficient and basic weight of each candidate tag, the adjustment weight of each candidate tag under the corresponding subcategory is determined as the final dynamic weight of the corresponding candidate tag. Based on a preset benchmark threshold, candidate labels with a final dynamic weight higher than the benchmark threshold are selected from the candidate pool and used as the main labels to form a set of main labels to be selected. A hierarchical diversity test is performed on each candidate main tag. After the diversity test, the candidate main tags are sorted from high to low according to dynamic weight, and the top five candidate main tags are selected as the main implicit requirement tags.

6. The job seeker profiling analysis method based on multi-agent technology according to claim 5, characterized in that, The process of generating scenario-based questionnaires by combining a multimedia resource library with a generative video model includes: The questionnaire agent, based on the identified main implicit demand tags and a pre-set multimedia material library, matches core scene elements related to each main implicit demand tag from the multimedia material library to form a core scene element set. Based on the core scene element set, the various elements in the core scene element set are combined through a pre-trained generative video model to form an interactive storyline script that includes characters, simple plots and potential conflict points. The script immerses job seekers in an interactive storyline, automatically pausing at key plot points and popping up embedded questions corresponding to the current key request turning point and the main implicit demand tags. Based on the job seeker's answers to the embedded questions, the system dynamically determines the presentation order, depth, and specific content of subsequent embedded questions through a pre-set decision logic tree, thus achieving an adaptive questionnaire flow.

7. The job seeker profiling analysis method based on multi-agent technology according to claim 6, characterized in that, The implementation of the adaptive questionnaire flow includes: Several embedded questions at the beginning of the questionnaire are set as decision nodes, and a decision logic tree is pre-defined to map different combinations of answers to subsequent embedded questions. When the decision logic tree determines that a new embedded question needs to be loaded, it retrieves the ontology-standardized question library in real time to generate an embedded question set that conforms to the current context. It has a preset maximum number of questions threshold and estimates and displays the remaining time for job seekers to answer questions in real time.

8. A job seeker profiling analysis system based on multi-agent intelligence, implementing the job seeker profiling analysis method based on multi-agent intelligence as described in any one of claims 1 to 7, characterized in that, include: This includes data processing agents, job seeker clustering agents, tag recognition agents, survey agents, and adjustment feedback agents; The data processing intelligent agent is used to collect multi-source static attribute data and historical interaction data of job seekers, fill missing values ​​through attribute similarity clustering algorithm, and filter out noise in interaction data based on attribute importance quantification and time threshold to obtain preprocessed data. The job seeker clustering intelligence is used to perform initial screening based on preprocessed data and the number of active application behaviors, and to classify job seekers into active job seekers, high-potential job seekers, information-gathering job seekers, and deeply observing job seekers through silent subclass merging and representative selection mechanisms. The tag recognition intelligent agent is used to extract job seekers' page interactions, behavior sequences and device-related data, input the implicit demand tag recognition model trained by hyperparameter iteration optimization, and output the implicit demand tags of each silent subclass of job seekers. The questionnaire agent is used to perform hierarchical weight analysis based on implicit demand tags, identify the main demand tags, generate scenario-based questionnaires by combining a multimedia material library and a generative video model, and achieve adaptive push through a decision logic tree. The adjustment feedback agent is used to pre-adjust the company's job information based on questionnaire feedback data and output a confirmation request to the company.

Images