An employment supply and demand accurate matching device based on a five-layer dynamic structure

Abstract

The application discloses a kind of based on five-layer dynamic structure's employment supply and demand accurate matching device, and the present application relates to human resource service technical field.It includes data acquisition module, operation matching module, result output module, operation matching module covers semantic enhancement layer, knowledge partition layer, label matching layer, dynamic tuning layer, feedback closed loop layer;The information feature of the present application is regularized after multi-stage semantic enhancement, relies on three-level theme tree and index routing to improve search efficiency;Through physical environment and psychological characteristics double-path label matching, all-around evaluation adaptation degree, still can dynamically adjust search operator weight and post recall quantity, optimization matching order.Combined with user job-seeking to the whole cycle behavior data completes stability test, drives model iteration, with parameter rollback and demand drift detection capability, can adapt to market labor demand change, matching result dimension is comprehensive, precision is high, effectively improves employment matching success rate and employment stability.

Description

Technical Field

[0001] This invention relates to the field of human resources service technology, specifically to a precise matching device for employment supply and demand based on a five-layer dynamic structure. Background Technology

[0002] The current job market faces long-standing challenges in matching supply and demand, including information asymmetry, limited matching dimensions, and poor dynamic adaptability. Existing matching systems largely rely on static searches using keywords related to education and skills, achieving only superficial text matching. They neglect unstructured factors such as work environment and psychological needs, resulting in significant discrepancies between matching results and job seekers' actual needs, leading to low job stability and a persistently high rate of mismatch between job seekers and employers.

[0003] In traditional solutions, vector retrieval and keyword retrieval have fixed weights, making it impossible to dynamically adjust retrieval strategies based on the completeness of information for different users and the characteristics of job data. Furthermore, they lack effective handling of semantically ambiguous expressions, leading to the omission of information for some job positions or job seekers. Simultaneously, existing systems lack a feedback loop based on user behavior data throughout the entire job search to resignation cycle, making it impossible to dynamically optimize the matching model based on employment stability. They also lack a demand drift detection mechanism, making it difficult to cope with dynamic changes in industry employment needs, and long-term matching accuracy is prone to decline.

[0004] Furthermore, existing matching frameworks are mostly single-layered, unable to achieve fine-grained routing retrieval from industry to atomic tasks, and lack a dual-path matching mechanism of physical and psychological tags, making it difficult to quantify the implicit fit between users and positions. Therefore, there is an urgent need for a multi-dimensional, dynamically optimized, and self-iterative precise matching solution for employment supply and demand to address the above pain points and improve matching accuracy and employment stability. Summary of the Invention

[0005] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a precise employment supply and demand matching device based on a five-layer dynamic structure, which solves the problems mentioned in the background section.

[0006] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: a precise employment supply and demand matching device based on a five-layer dynamic structure, comprising: The data acquisition module is used to collect job search information and job information of users with employment matching needs, and to collect user characteristic data, job requirement data, explicit physical environment tags, job requirement psychological tags, matching environment requirement tags, implicit psychological characteristic tags, and full-cycle behavior data of users with employment matching needs from job seeking to leaving their jobs. The computation and matching module is coupled to the data acquisition module and includes, from top to bottom, a semantic enhancement layer, a knowledge partitioning layer, a tag matching layer, a dynamic optimization layer, and a feedback loop layer. The semantic enhancement layer is used to construct user feature vectors and job feature vectors, and executes a five-stage semantic enhancement process to transform unstructured and colloquial raw data into structured feature vectors and calculate semantic similarity. The knowledge partitioning layer is used to construct a three-level topic tree of industry-job type-atomic task, and configure HNSW physical sub-indexes for each level of the three-level topic tree, and then retrieve according to the routing of each level; The label matching layer is used to establish a dual-path matching mechanism between explicit physical environment labels and implicit psychological feature labels, and to calculate the comprehensive label matching score. The dynamic optimization layer is used to dynamically adjust the vector retrieval weight and keyword retrieval weight based on semantic similarity and the dispersion of retrieval results, while optimizing the number of candidate positions recalled and optimizing the job matching ranking in real time. The feedback loop layer is used to calculate the employment stability score based on the user's behavior data throughout the entire lifecycle from job seeking to leaving the job, driving model iteration and supporting parameter rollback and demand drift detection. The results output module is coupled to the calculation and matching module and is used to output the final job ranking and multi-dimensional matching indicators.

[0007] Preferably, from the job search information of users with employment matching needs, the characteristics of these users are obtained, and the user feature vector is denoted as A = {a}. i}, where, i=1,2,...,n,a i Let i be the adaptive feature value for the i-th individual, where i represents the individual adaptive feature index and n represents the number of individual adaptive features. From the job information of users with employment matching needs, obtain the job requirements and denote the job requirement vector as B={b}. j}, where j=1,2,......,m, b j Let m be the job requirement value for the j-th position, where j represents the job requirement feature index and m represents the number of job requirement features.

[0008] Preferably, the five-stage semantic enhancement process in the semantic enhancement layer is as follows: Metadata extraction: Extract metadata from job search information and job posting information of users with employment matching needs, including individual matching type, educational background, work experience, job matching environment requirements, potential assessment value, as well as the industry, job type, skill requirements, physical environment requirements, work intensity, and communication method requirements of the job. Time and space standardization: In job search information and job posting information of users with employment matching needs, fuzzy time descriptions are transformed into standard time intervals, and fuzzy spatial descriptions are transformed into geographical coordinate ranges. Professional field mapping: In the job search information and job information of users with employment matching needs, non-standard personal descriptions and job titles are matched and mapped with the preset standard industry-job dictionary to generate corresponding standard industry codes and job codes, and establish a one-to-one mapping relationship between non-standard expressions and standard codes. Search granularity selection: Based on the job search information and job posting information of users with employment matching needs, the search granularity is dynamically selected, which is divided into two types: job level and task level, and is uniformly represented by the search granularity identifier L; Where L=1 represents a job-level index and L=2 represents a task-level index; The job level is determined as follows: When the job search information and job posting information of users with employment matching needs include: complete job title, basic job attributes, and no job descriptions, their needs are determined as overall job requirements, and the retrieval granularity is selected as job level, in which case L=1; The task-level judgment method is as follows: When the job search information and job posting information of users with employment matching needs include: no complete job title, single job task, or job description with responsibilities, their needs are determined to be single job task, and the retrieval granularity is selected as task level, in which case L=2. Multi-hop path planning: For job search information and job posting information of users with employment matching needs that cannot be matched by a single keyword or a single round of search, the information is broken down into multi-hop path step-by-step search tasks, forming a search path sequence, and each step of the search operation is executed in sequence. After completing the five-stage semantic enhancement, through: Calculate the semantic similarity X between user feature vector A and job feature vector B. A，B In the formula: i = 1, 2, ..., e, j = 1, 2, ..., e, e ≤ n and e ≤ m; |a i -b j | represents the user feature value 'a' for employment matching needs under the same corresponding feature dimension. i With the job requirement characteristic value b j The absolute difference; w i，j This is represented as the user feature value 'a' for employment matching needs under the same corresponding feature dimension. i With the job requirement characteristic value b j The pre-defined correlation ratio between them.

[0009] Preferably, both industry codes and job codes are taken from a pre-defined standard industry-job dictionary; the standard industry-job dictionary pre-establishes the correspondence between industry names and industry codes, and between job names and job codes; During the matching and mapping process, keywords are extracted from non-standard personal descriptions and job titles, and their similarity is compared with the keywords corresponding to industry names and job titles in the standard industry-job dictionary. The industry names and job titles with the highest similarity are selected. Then, according to the correspondence of the standard industry-job dictionary, the corresponding standard industry code and job code are output.

[0010] The preferred multi-hop path step-by-step retrieval task is broken down as follows: First jump: Search for relevant industries based on individual suitability type and job suitability environment requirements, and designate them as candidate industries, while forming and outputting a set of candidate industries; The second step: Within the candidate industries, search for job types according to skill requirements, define them as candidate job types, and simultaneously assemble and output a set of candidate job types; The third step: break down the candidate job into multiple atomic tasks, extract the execution conditions of each atomic task, and filter and match them with the user's own limitations in terms of employment adaptation needs. Select the atomic tasks that the user can complete independently, and then form and output a set of atomic tasks. Atomic tasks refer to the smallest unit of work, such as typing or checking documents. The fourth step: Based on the physical environment requirements and the job fit environment requirements, filter out the jobs whose physical environment requirements do not meet the job fit environment requirements, and designate the remaining jobs as the final candidate jobs. At the same time, the final candidate job set is formed and output.

[0011] Preferably, the three-level theme tree of industry-job type-atomic task is divided into industry layer, job type layer, and atomic task layer; The industry layer contains industry names corresponding to several industry categories, and industry indexes are configured accordingly; the job type layer is a derivative layer of the industry layer, with each industry covering job names corresponding to several job types, and job indexes are configured accordingly; the atomic task layer is a derivative layer of the job type layer, with each job type broken down into atomic task names corresponding to several atomic tasks, and atomic task indexes are configured accordingly. During retrieval, the system routes results in a hierarchical structure: first, it matches the industry index to locate the corresponding industry sub-index; then, it matches the job index to locate the corresponding job sub-index; and finally, it matches the atomic task index to retrieve the corresponding job position for the atomic task.

[0012] Preferably, the HNSW physical sub-index parameters are: the maximum number of edges and the number of candidate constructs for the corresponding level index, denoted as M and EF respectively; The initial values ​​of the maximum number of edges M and the number of candidates EF are set manually in advance; they are then dynamically adjusted according to the retrieval accuracy.

[0013] The preferred method for dynamically fine-tuning the maximum number of edges M and the number of candidate edges EF is as follows: Based on the ranking benchmark score SP output by the dynamic optimization layer and the corresponding preset matching threshold, the search accuracy is determined. The method is as follows: the ranking benchmark score SP is used as the matching degree between users with employment matching needs and job positions, and their matching thresholds are compared. The number of job positions with a matching degree greater than or equal to the matching threshold (TP) and the number of job positions with a matching degree less than the matching threshold (FP) are counted. This is achieved through: The search precision rate HM was calculated. Compare the retrieval precision HM with the preset low precision threshold ML and high precision threshold MH: if HM < ML, it means that M or EF needs to be increased, and the corresponding adjustment formula is as follows. and ; When HM > MH, it means that M or EF needs to be reduced, and the corresponding adjustment formula is: and In the formula, M1 and EF1 are the maximum number of edges and the number of candidate edges after dynamic fine-tuning, respectively; h is the preset adjustment step size coefficient. If ML≤HM≤MH, it means that M or EF does not need to be adjusted.

[0014] Preferably, explicit physical environment tags are obtained from the job information of users with employment matching needs, and the explicit physical environment tag vector is denoted as C={c s}, where s = 1, 2, ..., x, c s Let x be the feature value of the s-th explicit physical environment tag, where s represents the index of the explicit physical environment tag and x represents the number of explicit physical environment tags. In the job search information of users with employment matching needs, implicit psychological feature labels are obtained, and the implicit psychological feature label vector is denoted as E={e t}, where t=1,2,......,y,e t Let y be the feature value of the t-th implicit psychological feature label, where t represents the index of the implicit psychological feature label and y represents the number of implicit psychological feature labels. From the job search information of users with employment matching needs, obtain the matching environment requirement tags, and denote the matching environment requirement tag vector as D={d s}, d s Let be the requirement value of the s-th adaptation environment requirement tag, where the explicit physical environment tag vector and the tags in the adaptation environment requirement tag vector correspond one-to-one; From the job information of users with employment matching needs, obtain the psychological tags for job requirements, and denote the vector of psychological tags for job requirements as F={f t}, f t Let be the requirement value of the psychological label for the t-th job requirement, where the implicit psychological feature label vector and the labels in the job requirement psychological label vector correspond one-to-one.

[0015] The preferred dual-path matching mechanism is as follows: Path 1: Explicit physical environment tag matching: pass: Calculate the physical label matching score P between the explicit physical environment label vector and the adaptive environment requirement label vector. C，D ; In the formula: c s and d s When 's' are the same, it indicates that the explicit physical environment tag and the adaptation environment requirement tag correspond; P(c s d s P(c) represents the matching function between explicit physical environment labels and adaptive environment requirement labels. When the feature value of the explicit physical environment label meets the requirement value of the adaptive environment requirement label, then P(c) is a matching function. s d s The value of P(c) is 1; otherwise, P(c) is 1. s d s The value is 0; λ1 s This represents the weight parameter between the s-th corresponding explicit physical environment label and the adaptation environment requirement label, where all λ1 s The sum is 1; Path 2: Implicit Mental Feature Label Matching pass: The psychological label matching score P between the implicit psychological feature label vector and the job requirement psychological label vector is calculated. E，F ; In the formula: e t and f t When t is the same, it indicates a correspondence between implicit psychological characteristic labels and job requirement psychological labels; λ2 t λ² represents the weight parameter between the t-th corresponding implicit psychological characteristic label and the job requirement psychological label, where all λ² t The sum is 1; The physical label matching score and psychological label matching score obtained from the dual-path matching mechanism are analyzed by... A weighted fusion is performed to determine the comprehensive label matching score P1; where: w P Match the score weight parameters to the preset physical labels.

[0016] Preferably, the vector retrieval weight and keyword retrieval weight are dynamically adjusted based on the semantic similarity value: when the semantic similarity is lower than a preset threshold, the vector retrieval weight is increased and the keyword retrieval weight is decreased; when the semantic similarity is higher than the preset threshold, the vector retrieval weight is decreased and the keyword retrieval weight is increased, in order to adapt to different matching scenarios. Specifically, through: , The dynamically adjusted vector retrieval weight w1 and keyword retrieval weight w2 are calculated; where: w 10 Δw represents the preset initial weights for vector retrieval; X(max) represents the preset weight adjustment range. A，B This is the preset maximum semantic similarity threshold.

[0017] Preferably, the optimization method for the number of candidate positions recalled is as follows: Extract the K0 candidate positions initially recalled from the knowledge partitioning layer, calculate the semantic similarity between each candidate position and the user vector of employment matching needs, and form a semantic similarity set; based on the similarity value set, calculate the standard deviation of the semantic similarity data in the set; divide the obtained standard deviation by the preset theoretical maximum value of the semantic similarity standard deviation to complete the normalization process, and obtain the semantic dispersion (DDI) of the recall results; adjust the initial recall quantity according to the semantic similarity dispersion: based on the initial recall quantity, when the semantic similarity dispersion increases, reduce the number of recalled positions; conversely, increase the number of recalled positions. The formula for optimizing the number of recalled positions is as follows: In the formula: K0 is the preset initial recall quantity; σ is the preset dispersion adjustment coefficient.

[0018] Preferably, the dynamically adjusted vector retrieval weights and keyword retrieval weights are retrieved independently, and the corresponding candidate positions and their respective matching scores S1 and S2 under independent retrieval are re-output; then, the matching scores under independent retrieval corresponding to the vector retrieval weights and keyword retrieval weights are weighted and fused to obtain the ranking benchmark score SP for the candidate positions; the weighted fusion formula is: .

[0019] Preferably, the full-cycle behavior data vector is denoted as U={u g}, g=1,2,......,z,u g For users seeking employment to meet their job matching needs, the behavioral indicator values ​​are defined as follows: g is the behavioral indicator index, z is the number of behavioral indicators, and u represents the number of behavioral indicators in the full-cycle behavioral data. g This is the dimensionless index value after normalization.

[0020] Preferably, the employment stability score D of users with relevant employment matching needs is obtained by weighting and summing various behavioral indicators throughout the entire job search to job resignation cycle according to preset weights; the formula is as follows: ;γ g Let γ be the preset weight parameter corresponding to the g-th behavioral indicator, and all γ g The sum is 1.

[0021] Preferably, the model iteration involves iteratively updating the weights of each behavioral indicator based on the direction and magnitude of changes in the employment stability score; the formula is: In the formula: γ1 g For the updated weight parameters; γ g The preset weight parameters are used to define the behavioral indicators before the update iteration; ρ is the preset iteration step size. The gradient represents the direction and magnitude of changes in the employment stability score.

[0022] Preferably, the employment stability score gradient is obtained based on a pre-constructed employment stability score calculation formula: the weights of each behavioral indicator are determined according to the scoring formula, and the correlation mapping relationship between the system weight parameters and behavioral indicators is combined. The partial derivatives of the employment stability score with respect to each weight parameter are solved by chain derivative method, and all partial derivatives are combined to form the employment stability score gradient.

[0023] Preferably, the parameter rollback mechanism is set as follows: After completing the iterative update of the weight parameters, the employment stability score corresponding to users with relevant employment matching needs is recalculated; the changes in employment stability scores before and after the iteration are compared. If the relative decrease in employment stability score compared to before the iteration update is greater than a preset threshold, then γ1 is adjusted. g Restore the weight parameters γ to their original values ​​before the iterative update. g .

[0024] The preferred method for demand drift detection is as follows: The method for detecting demand drift involves periodically comparing the current job demand vector with the previous period's job demand vector, and then averaging the absolute values ​​of the differences between them to calculate the job demand drift amount. The formula is as follows: In the formula, ΔB represents the job demand drift, and b1 j b0 represents the updated demand value for the j-th job position. j This represents the demand value for the j-th job position before the update; When the job demand drift exceeds the preset drift threshold, the model is retrained to update the parameters of each level in the model. Preferably, the result output module uses the ranking benchmark score output by the dynamic optimization layer as the first ranking criterion, and arranges the candidate positions in descending order from high to low to perform a preliminary ranking; then, it uses the comprehensive tag matching score output by the tag matching layer as the second ranking criterion, and arranges them in descending order from high to low. Based on the preliminary ranking, the ranking order of the candidate positions is corrected to determine the final arrangement order of the suitable positions. Each candidate job posting output includes multi-dimensional matching indicators such as semantic similarity, physical tag matching score, psychological tag matching score, and comprehensive tag matching score; and simultaneously outputs dynamically adjusted retrieval operator weights and the number of candidate job postings recalled.

[0025] (III) Beneficial Effects This invention provides a precise job supply and demand matching device based on a five-layer dynamic structure. Compared with existing technologies, it has the following advantages: Through the coordinated operation of the five-layer dynamic structure, the entire process of matching employment supply and demand is optimized. The semantic enhancement layer, through five-stage processing, effectively eliminates the ambiguity in the descriptions of job seeker information and job information, constructs more accurate feature vectors, significantly improves the accuracy of semantic similarity calculation, and provides a reliable data foundation for subsequent matching.

[0026] The three-level topic tree and HNSW physical sub-index constructed by the knowledge partitioning layer enable hierarchical retrieval from industry to atomic task. Through hierarchical routing and dynamic index optimization, the efficiency and accuracy of job retrieval are significantly improved. It supports rapid retrieval of large-scale job data and can adapt to different retrieval granularity requirements, ensuring the high efficiency of the matching process.

[0027] The dual-path matching mechanism of the tag matching layer takes into account both the matching needs of explicit physical environment and implicit psychological characteristics. By weightedly fusing physical and psychological tag matching scores, it breaks through the limitation of traditional matching that only focuses on hard skills, and can more comprehensively assess the suitability of users and positions, effectively reducing the risk of employment instability caused by environmental and psychological mismatch.

[0028] The dynamic optimization layer can dynamically adjust the weights of vector and keyword retrieval, as well as the number of retrieved results, based on semantic similarity and retrieval dispersion. This mechanism overcomes the limitations of a single retrieval method, optimizes matching strategies according to data distribution characteristics, and significantly improves the rationality of job ranking and the recall quality of candidate jobs.

[0029] The feedback loop layer calculates employment stability scores based on users' full-cycle behavioral data, driving iterative model optimization. It also supports parameter rollback and demand drift detection, continuously improving model adaptability and responding promptly to changes in market demand, ensuring the long-term stability and adaptability of matching results, and achieving system self-improvement.

[0030] The results output module uses dual sorting criteria to determine the job matching order and simultaneously outputs multi-dimensional matching indicators and optimization parameters. This provides users with a clear reference for job priority and facilitates subsequent tracking of matching logic and optimization direction, improving the transparency and interpretability of the matching results and enhancing user trust. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of a module of an employment supply and demand precision matching device based on a five-layer dynamic structure according to the present invention.

[0032] Figure 2 This is a flowchart illustrating the tag matching layer in this invention. Detailed Implementation

[0033] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] Please see Figure 1 and Figure 2 As shown, the embodiments of the present invention provide the following technical solutions: As an embodiment of the present invention: This invention relates to a precise job supply and demand matching device based on a five-layer dynamic structure, comprising: The data acquisition module is used to collect job search information and job information of users with employment matching needs, and to collect user characteristic data, job requirement data, and full-cycle behavior data of users with employment matching needs from job seeking to leaving their jobs. Users with job matching needs refer to the employment group whose individual adaptation differences mean that individuals have different compatibility with general job standards in terms of physiological functions, perception abilities, behavior patterns, and psychological states. Therefore, they have personalized adaptation needs for job environment, work methods, task intensity, etc. in the employment scenario. In this embodiment, the unstructured data corresponding to the user characteristic data, job requirement data, and full-cycle behavior data of employment matching needs are all converted into structured data; the numerical data are all normalized, and the normalization process adopts max-min standardization or Z-score standardization to eliminate the difference in the units of measurement between different indicators; the categorical data are all encoded and converted, that is, the categorical data is mapped to standardized numerical codes. Full-cycle behavioral data includes, but is not limited to, probation retention rate, work output achievement rate, enterprise satisfaction, on-the-job duration, and task completion accuracy, which are obtained by normalization after being filled in by the enterprise.

[0035] The computation and matching module is coupled to the data acquisition module and includes, from top to bottom, a semantic enhancement layer, a knowledge partitioning layer, a tag matching layer, a dynamic optimization layer, and a feedback loop layer. The specific structure is as follows: The semantic enhancement layer is used to construct user feature vectors and job feature vectors, and to calculate semantic similarity; In the job search information of users with employment matching needs, the characteristics of users with employment matching needs are obtained. In this embodiment, the individual matching characteristics specifically include individual matching type, educational background, work experience, job matching environment requirements, and potential assessment value. Individual fit types are categorized into fit differences based on dimensions such as physiology, perception, movement, psychology, and cognition, and are used to accurately match corresponding job types; individual fit levels refer to the standardized grading of the degree of individual fit differences, and are used to quantify the difficulty of matching and fit requirements. In this embodiment, individual adaptation type can also be characterized as social adaptation feature, representing user social preferences and interaction patterns, and is divided into two categories: extroverted and introverted. Among them, extroverts prefer high-frequency interpersonal communication, teamwork, and public expression; introverts prefer independent work, low social intensity, and primarily written communication; during the matching process, extroverted users are given priority for customer service positions; introverted users are given priority for copywriting positions. Let the user feature vector be denoted as A = {a}. i}, where, i=1,2,...,n,a i Let i be the adaptive feature value for the i-th individual, where i represents the individual adaptive feature index and n represents the number of individual adaptive features. In the job information of users with employment matching needs, the job requirements are obtained. In this embodiment, the job requirements include the industry, job type, skill requirements, physical environment requirements, work intensity, and communication method requirements. Let the job requirement vector be denoted as B = {b} j}, where j=1,2,......,m, b j Let m be the job requirement value for the j-th position, where j represents the job requirement feature index and m represents the number of job requirement features.

[0036] pass: ; Calculate the semantic similarity X between user feature vector A and job feature vector B. A，B ; In the formula: i=1,2,...,e, j=1,2,...,e, e≤n and e≤m; |a i -bj | represents the user feature value 'a' for employment matching needs under the same corresponding feature dimension. i With the job requirement characteristic value b j The absolute difference; For example: User characteristic value a for job matching needs i The characteristic value b represents the commuting distance requirement and the job requirement. j Commuting distance is the same corresponding feature dimension; w i，j This is represented as the user feature value 'a' for employment matching needs under the same corresponding feature dimension. i With the job requirement characteristic value b j The pre-defined correlation ratio between them; In this embodiment, w i，j The value range is [0,1], and the sum of the correlation ratio coefficients of all feature dimensions under the same model is 1; among them, the weight of core hard skills features is usually set to 0.3-0.6, the weight of general ability and soft skills features is usually set to 0.1-0.3, and the weight of personal preference features is usually set to 0.05-0.2. The knowledge partitioning layer is used to construct a three-level topic tree of industry-job type-atomic task, and configure HNSW physical sub-indexes for each level of the three-level topic tree, and then retrieve according to the routing of each level; The three-level theme tree of industry-job type-atomic task is divided into industry layer, job type layer, and atomic task layer; The industry layer contains the industry names corresponding to several industry categories, and the industry index in the industry layer is denoted as p1; In this embodiment, the number of industry categories is denoted as u1; The job type layer is a derivative layer of the industry layer. Each industry covers several job types and their corresponding job names. The job type index in the job type layer is denoted as p2. In this embodiment, the number of job types in an industry category is denoted as u2, where the number of job types differs across different industry categories; The atomic task layer is a derivative layer of the job type layer. Each job type is broken down into several atomic task names, and the atomic task index in the atomic task layer is recorded as p3. In this embodiment, the number of atomic tasks in a job is denoted as u3; the number of atomic tasks is different in different jobs. The HNSW physical sub-index parameters are: the maximum number of edges and the number of candidate edges to be constructed for the corresponding level index, denoted as M and EF respectively; In this embodiment, HNSW is a hierarchical navigation small-world graph, which is a commonly used algorithm for vector retrieval, so it will not be described in detail here; the maximum number of edges is the maximum number of neighbors that each node can connect to in each layer; the number of candidates is the size of the neighbor candidate pool of each node when building the graph; During retrieval, routing is performed hierarchically by industry level, job type level, and atomic task level; First, match the industry index p1 to locate the corresponding industry sub-index; then match the job index p2 to locate the corresponding job seed index; finally, match the atomic task index p3 to retrieve the corresponding job position for the atomic task.

[0037] The label matching layer is used to establish a dual-path matching mechanism between explicit physical environment labels and implicit psychological feature labels, and to calculate the comprehensive label matching score. In the job information of users with employment matching needs, explicit physical environment tags are obtained. In this embodiment, the explicit physical environment tags specifically include job matching environment, commuting distance, environmental comfort, etc. Let the explicit physical environment tag vector be denoted as C={c s}, where s = 1, 2, ..., x, c s Let x be the feature value of the s-th explicit physical environment tag, where s represents the index of the explicit physical environment tag and x represents the number of explicit physical environment tags. In the job search information of users with employment matching needs, implicit psychological feature tags are obtained. In this embodiment, the implicit psychological feature tags specifically include social anxiety level, attention duration, stress tolerance, etc. Let the implicit psychological feature label vector be denoted as E={e}. t}, where t=1,2,......,y,e t Let y be the feature value of the t-th implicit psychological feature label, where t represents the index of the implicit psychological feature label and y represents the number of implicit psychological feature labels. The dual-path matching mechanism is as follows: Path 1: Explicit physical environment tag matching: Obtain the matching environment requirement tag from the job search information of users with employment matching needs; Let the vector of labels for adapting to environmental requirements be denoted as D={d}. s}, d s This represents the required value for the s-th adaptation environment requirement tag. Among them, the labels in the explicit physical environment label vector and the adaptive environment requirement label vector correspond one-to-one; pass: ; Calculate the physical label matching score P between the explicit physical environment label vector and the adaptive environment requirement label vector. C，D ; Among them, P C，D The larger the value, the higher the matching degree between the explicit physical environment label vector and the adaptive environment requirement label vector; In the formula: cs and d s When the 's' in the same string is the same, it indicates that the explicit physical environment tag and the adaptation environment requirement tag correspond to each other. P(c) s d s P(c) represents the matching function between explicit physical environment labels and adaptive environment requirement labels. When the feature value of the explicit physical environment label meets the requirement value of the adaptive environment requirement label, then P(c) is a matching function. s d s The value of P(c) is 1; otherwise, P(c) is 1. s d s The value is 0; In this embodiment, the requirement value of the adaptation environment requirement label is a range value, including the minimum and maximum values ​​of the environment adaptation requirement; When the feature value of the explicit physical environment tag is greater than or equal to the minimum value of the environment adaptation requirement and less than or equal to the maximum value of the environment adaptation requirement, it means that the feature value of the explicit physical environment tag meets the requirement value of the adaptation environment requirement tag. λ1 s This represents the weight parameter between the s-th corresponding explicit physical environment label and the adaptation environment requirement label, where all λ1 s The sum is 1.

[0038] Path 2: Implicit Mental Feature Label Matching Extract psychological tags related to job requirements from job information of users with employment matching needs; Let the psychological label vector of job requirements be denoted as F = {f}. t}, f t Let t be the required value of the psychological label for the t-th job position; Among them, the implicit psychological characteristic label vector and the labels in the job requirement psychological label vector correspond one-to-one; pass: ; Calculate the psychological label matching score P between the implicit psychological feature label vector and the job requirement psychological label vector. E，F ; In the formula: e t and f t When 't' is the same, it indicates that the implicit psychological characteristic label and the job requirement psychological label correspond. λ2 t λ² represents the weight parameter between the t-th corresponding implicit psychological characteristic label and the job requirement psychological label, where all λ² t The sum is 1.

[0039] The physical label matching score and the psychological label matching score obtained from the dual-path matching mechanism are weighted and fused to determine the comprehensive label matching score; The weighted processing is as follows: ; In the formula: P1 represents the overall score for label matching; w P Match score weight parameters to preset physical labels; In this embodiment, w P The value range is [0.4, 0.6].

[0040] The dynamic optimization layer is used to dynamically adjust the weight of the retrieval operator and the number of candidate positions recalled based on the matching semantic features and the aggregation status of the retrieval results, thereby optimizing the job matching and ranking effect. The dynamic adjustment method for retrieval operator weights is as follows: The retrieval operator includes vector retrieval weights and keyword retrieval weights; pass: ; ; Calculate the dynamically adjusted vector retrieval weight w1 and keyword retrieval weight w2; In the formula: w 10 The initial weights for the preset vector retrieval; Δw is the preset weight adjustment range; X (max) A，B The preset maximum semantic similarity threshold; When semantic similarity X A，B When the value is low, increase the vector retrieval weight w1; When there are many technical terms related to skills, increase the keyword search weight w2.

[0041] The dynamic adjustment method for the number of candidate positions recalled is as follows: Extract the K0 candidate positions initially recalled from the knowledge partitioning layer, calculate the semantic similarity between each candidate position and the user vector of employment matching needs, and form a semantic similarity set. Based on the set of similarity values, the standard deviation of the semantic similarity data within the set is calculated; the obtained standard deviation is divided by the preset theoretical maximum value of the semantic similarity standard deviation, and after normalization, the semantic dispersion (DDI) of the recall results is obtained. pass: ; Calculate the optimized number of recall positions, K; The number of recalled positions represents the total number of candidate matching positions initially extracted from the position resource database through the retrieval operator; In the formula: K0 is the preset initial recall quantity; σ is the preset dispersion adjustment coefficient; In this embodiment, σ is used to control the discrete tolerance of the similarity calculation between the user feature vector and the job feature vector, avoiding large fluctuations in the matching results due to small deviations in individual features. Its value range is 0.1≤σ≤0.5; When σ is small, such as 0.1-0.2, the model is more sensitive to feature differences and the matching results are more accurate, but the recall rate is relatively low. When σ is large, such as 0.3-0.5, the model has a higher tolerance for feature differences, which can expand the matching range and is suitable for scenarios where there is insufficient supply of job positions in the market and unclear user needs.

[0042] When DDI increases, K will automatically decrease.

[0043] After dynamic adjustment, the vector retrieval weight and keyword retrieval weight are retrieved independently, and the corresponding candidate positions and their matching scores under each independent retrieval are re-output; Then, the matching scores under independent searches corresponding to the vector retrieval weight and the keyword retrieval weight are weighted and fused to obtain the ranking benchmark score for the candidate positions; The weighted fusion formula is as follows: ; In the formula; SP is the ranking benchmark score for candidate positions; S1 and S2 are the matching scores for independent searches, corresponding to the vector retrieval weight and the keyword retrieval weight, respectively.

[0044] The feedback loop layer is used to calculate the employment stability score based on the full-cycle behavioral data of users from job seeking to leaving their jobs, driving model iteration and supporting parameter rollback and demand drift detection. The full-cycle behavior data vector is denoted as U={u g}, g=1,2,......,z,u g The behavioral metrics values ​​for users from job seeking to job resignation, where g is the behavioral metric index and z is the number of behavioral metrics; Behavioral indicators u in full-cycle behavioral data g The dimensionless index value after normalization; The formula for scoring job stability is as follows: ; In the formula: D represents the employment stability score for users with relevant employment matching needs; γ g Let γ be the preset weight parameter corresponding to the g-th behavioral indicator, and all γ g The sum is 1; The model iteration method is as follows: Update the weight parameter γ based on the employment stability score D. g The iterative formula is as follows: ; In the formula: γ1 g These are the updated weight parameters; γ g The preset weight parameters are used to update the behavioral indicators before the iteration. ρ is the preset iteration step size; This represents the gradient of employment stability scores, indicating the direction and magnitude of changes in employment stability scores. In this embodiment, the employment stability score gradient is obtained based on a pre-constructed employment stability score calculation formula: the weights of each behavioral indicator are determined according to the scoring formula, and the correlation mapping relationship between the system weight parameters and the behavioral indicators is combined. The partial derivatives of the employment stability score with respect to each weight parameter are solved by chain derivative method, and all partial derivatives are combined to form the employment stability score gradient.

[0045] The parameter rollback mechanism is set as follows: After completing the iterative update of the weight parameters, the employment stability score corresponding to users with relevant employment matching needs is recalculated; the changes in employment stability scores before and after the iteration are compared. If the relative decrease in employment stability score compared to before the iteration update is greater than a preset threshold, then γ1 is adjusted. g Restore the weight parameters γ to their original values ​​before the iterative update. g ; The parameter rollback mechanism effectively avoids the decline in matching performance caused by abnormal adjustments to weight parameters.

[0046] The demand drift detection method is as follows: Periodically extract the job requirement vector and denote the current job requirement vector as B1={b1}. j}, b1 j This represents the updated job demand value for the j-th position, while the previous period's job demand vector is denoted as B0 = {b0}. j}, b0 j This represents the demand value for the j-th job position before the update; Subsequently passed: ; Calculate the job demand drift ΔB; Among them, the larger the value of the job demand drift ΔB, the more obvious the change in job demand; Compare the job demand drift ΔB with the preset drift threshold: When ΔB exceeds the preset drift threshold, the model is retrained, and the parameters of each layer in the model are updated. In this embodiment, the model is retrained as follows: Collect user job search data, job demand data, and behavioral data after demand drift occurs; Update user characteristic tags, job requirement tags, and implicit psychological characteristic tags, and add new feature dimensions; Based on the updated data, initialize the weight parameters w of the matching model. i,j The feature weights that are still valid in the original model are retained as initial values; The model was trained by mixing historical matching data with newly collected data. The model parameters were dynamically adjusted based on the employment stability score and the retrieval accuracy HM until the matching effect returned to the normal range. The original model is replaced with the retrained model, and the matching effect and stability score are continuously monitored. When abnormal changes occur, the update process is triggered again.

[0047] The results output module, coupled with the calculation and matching module, is used to output the final ranking of suitable positions and multi-dimensional matching indicators, as detailed below: The candidate positions are initially ranked based on the ranking benchmark score output by the dynamic optimization layer, arranged in descending order from high to low. Then, the comprehensive tag matching score output by the tag matching layer is used as the second ranking basis, also arranged in descending order from high to low. Based on the initial ranking, the order of candidate positions is adjusted to determine the final arrangement order of suitable positions. Each candidate job posting output includes multi-dimensional matching indicators such as semantic similarity, physical tag matching score, psychological tag matching score, and comprehensive tag matching score; and simultaneously outputs dynamically adjusted retrieval operator weights and the number of candidate job postings recalled.

[0048] This embodiment achieves accurate job matching for users with employment needs through a dual-path tag matching mechanism. Dynamically optimizing the weights of retrieval operators and the number of retrieved results, combined with full-cycle behavioral data-driven model iteration, effectively reduces matching bias and improves job matching accuracy and employment stability.

[0049] As a second embodiment of the present invention: In specific implementation, compared to Embodiment 1, the technical solution of this embodiment differs from that of Embodiment 1 only in this embodiment. The semantic enhancement layer transforms unstructured, colloquial raw data into structured feature vectors through a five-stage semantic enhancement process. The five-stage semantic enhancement process is as follows: Metadata extraction: Extract metadata from job search information and job posting information of users with employment matching needs, including individual matching type, educational background, work experience, job matching environment requirements, potential assessment value, as well as the industry, job type, skill requirements, physical environment requirements, work intensity, and communication method requirements of the job.

[0050] Time and space standardization: In job search information and job posting information of users with employment matching needs, fuzzy time expressions are transformed into standard time intervals T, and fuzzy spatial expressions are transformed into geographic coordinate ranges G. Where T = [t0, t1], t0 is the start timestamp, and t1 is the end timestamp; G = (x min ,x max ,y min ,y max ), x min x max These represent the minimum and maximum longitude values ​​within the geographic coordinate range, respectively, y min y max These represent the minimum and maximum latitude values ​​within the geographic coordinate range, respectively.

[0051] Professional field mapping: In the job search information and job information of users with employment matching needs, non-standard personal descriptions and job titles are matched and mapped with the preset standard industry-job dictionary to generate corresponding standard industry codes and job codes, and establish a one-to-one mapping relationship between non-standard expressions and standard codes. The industry code and job code are both taken from the pre-set standard industry-job dictionary; the standard industry-job dictionary pre-establishes the correspondence between industry name and industry code, and job name and job code. During the matching and mapping process, keywords are extracted from non-standard personal descriptions and job titles, and their similarity is compared with the keywords corresponding to industry names and job titles in the standard industry-job dictionary. The industry names and job titles with the highest similarity are selected. Then, according to the correspondence of the standard industry-job dictionary, the corresponding standard industry code and job code are output.

[0052] Search granularity selection: Based on the job search information and job posting information of users with employment matching needs, the search granularity is dynamically selected, which is divided into two types: job level and task level, and is uniformly represented by the search granularity identifier L; The retrieval granularity L is used to control the precision of the knowledge partitioning layer routing, where L=1 represents the job-level index and L=2 represents the task-level index. Its role in routing is as follows: When L=1, the knowledge partitioning layer directly uses the complete job posting as the matching unit. During routing, it matches user characteristics with the overall job profile, which is suitable for scenarios with clear job search goals and high requirements for hard skills. The matching results are highly relevant overall, which can reduce invalid applications. When L=2, the knowledge partitioning layer breaks down the job into core work task units for routing and matching. This is suitable for scenarios where users have vague job requirements and strong skill transferability. It can expand the matching range and uncover suitable job positions that users have not explicitly expressed. Among them, the retrieval granularity L determines the node division granularity of the topic tree of the knowledge partitioning layer. The larger the value of L, the finer the node division, and the more focused the feature dimensions covered by a single node. When filtering routes, low-relevance partitions can be removed more accurately, reducing the workload of subsequent matching calculations. The smaller the value of L, the coarser the node division, and the broader the feature dimensions covered by a single node. This can reduce the omission of effective positions due to over-filtering. The job level is determined as follows: When the job search information and job posting information of users with employment matching needs include: complete job title, basic job attributes, and no job descriptions, their needs are determined as overall job requirements, and the retrieval granularity is selected as job level, in which case L=1; The task-level judgment method is as follows: When the job search information and job posting information of users with employment matching needs include: no complete job title, single job task, or job description with responsibilities, their needs are determined to be single job task, and the retrieval granularity is selected as task level, in which case L=2. In this embodiment, the basic attributes of a job include industry, job type, work location, working hours, salary range, full-time / part-time, etc.; a single job task refers to a work link in the job, and the responsibilities qualifiers include only doing, only responsible for, specializing in, not doing, avoiding, only needing, etc.

[0053] Multi-hop path planning: For job search information and job posting information of users with employment matching needs that cannot be matched by a single keyword or a single round of search, the information is broken down into multi-hop path step-by-step search tasks, forming a search path sequence P, and each step of the search operation is executed in sequence. The multi-hop path step-by-step retrieval task is broken down as follows: First jump: Search for relevant industries based on individual suitability type and job suitability environment requirements, and designate them as candidate industries, while forming and outputting a set of candidate industries; The second step: Within the candidate industries, search for job types according to skill requirements, define them as candidate job types, and simultaneously assemble and output a set of candidate job types; The third step: break down the candidate job into multiple atomic tasks, extract the execution conditions of each atomic task, and filter and match them with the user's own limitations in terms of employment adaptation needs. Select the atomic tasks that the user can complete independently, and then form and output a set of atomic tasks. Atomic tasks refer to the smallest unit of work, such as typing or checking documents. In this embodiment, the execution conditions of the atomic task include limb range of motion, sensory ability, and operational precision requirements; The fourth step: Based on the physical environment requirements and the job adaptation environment requirements, filter out the jobs whose physical environment requirements do not meet the job adaptation environment requirements, and set the remaining jobs as the final candidate jobs. At the same time, the final candidate job set is formed and output. In this embodiment, the retrieval results of each hop are used as the input for the next hop in sequence, and the process is executed sequentially and converges step by step. After completing the five-stage semantic enhancement, the semantic similarity between the user feature vector and the job feature vector is calculated.

[0054] This embodiment solves the challenges of processing non-standard expressions and ambiguous spatiotemporal information through a five-stage semantic enhancement process, while supporting multi-hop path retrieval and dynamic granularity selection. It can accurately match user capabilities with job tasks, adapt to colloquial job postings, and significantly improve the structure of matched information and retrieval efficiency.

[0055] As an embodiment of the present invention: In specific implementation, compared with Embodiment 1 and Embodiment 2, the technical solution of this embodiment is to combine the solutions of Embodiment 1 and Embodiment 2. The only difference between the technical solution of this embodiment and Embodiment 1 and Embodiment 2 is that in this embodiment, the initial values ​​of the maximum number of edges M and the number of candidate EF are pre-set manually; and then dynamically fine-tuned according to the retrieval accuracy. The dynamic fine-tuning methods for the maximum number of edges M and the number of candidate edges EF are as follows: Based on the ranking benchmark score SP output by the dynamic optimization layer and the corresponding preset matching threshold, the search accuracy is determined. The method is as follows: the ranking benchmark score SP is used as the matching degree between users with employment matching needs and job positions, and their matching thresholds are compared. The number of job positions with a matching degree greater than or equal to the matching threshold (TP) and the number of job positions with a matching degree less than the matching threshold (FP) are counted. This is achieved through: The search precision rate HM was calculated. The retrieval precision (HM) is compared with the preset low precision threshold (ML) and high precision threshold (MH): When HM < ML, it means that M or EF needs to be increased, and the corresponding adjustment formula is as follows: ; ; When HM > MH, it means that M or EF needs to be reduced, and the corresponding adjustment formula is as follows: ; ; In the formula; M1 and EF1 are the maximum number of edges and the number of candidate edges after dynamic fine-tuning, respectively; h is the preset adjustment step size coefficient; When ML≤HM≤MH, it means that M or EF does not need to be adjusted; This embodiment dynamically fine-tunes the HNSW index parameters, optimizing index performance in real time based on retrieval accuracy, thus resolving the issue of poor adaptability of static indexes when job requirements change. This effectively improves the accuracy and stability of vector retrieval, ensuring efficient operation of the matching process under large-scale job data.

[0056] As an embodiment of the present invention: In specific implementation, compared with Embodiment 1, Embodiment 2 and Embodiment 3, the technical solution of this embodiment is to combine the solutions of Embodiment 1, Embodiment 2 and Embodiment 3.

[0057] This embodiment achieves end-to-end collaboration of semantic enhancement, hierarchical indexing, dual-path matching, dynamic optimization, and feedback loop. It can comprehensively address the diverse matching scenarios of users with employment matching needs, taking into account information organization, retrieval efficiency, matching accuracy, and long-term stability, resulting in stronger adaptability.

[0058] It should be stated that all user data collected in this application was collected with the user's consent and authorization, and the use of user data is legal and compliant, and the use and processing of user data comply with the relevant laws, regulations and standards of the relevant regions.

[0059] Furthermore, any content not described in detail in this specification is existing technology known to those skilled in the art.

[0060] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0061] The above formulas are all dimensionless calculations. Dimensionless calculation involves introducing a reference benchmark, such as the maximum, minimum, standard deviation, or theoretical extreme value of a physical quantity, to transform the original physical quantity into a dimensionless relative value. This value is usually mapped to a specific interval, such as [0,1] or [-1,1], which eliminates the influence of units while preserving the relative size relationship of the physical quantities. The formula is derived from software simulation based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0062] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

[0063] The various embodiments of this application have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A precise matching device for employment supply and demand based on a five-layer dynamic structure, characterized in that, include: The data acquisition module is used to collect job search information and job information of users with employment matching needs, and to collect user characteristic data, job requirement data, explicit physical environment tags, job requirement psychological tags, matching environment requirement tags, implicit psychological characteristic tags, and full-cycle behavior data of users with employment matching needs from job seeking to leaving their jobs. The matching module, from top to bottom, includes: The semantic enhancement layer is used to construct user feature vectors and job feature vectors. It generates structured feature vectors through five stages of semantic enhancement and calculates semantic similarity. The five-stage semantic enhancement process in the semantic enhancement layer is as follows: Metadata extraction: Extracting corresponding metadata from job-seeking information and job information of users with employment matching needs; Time and space normalization: Converting fuzzy time expressions into standard time intervals and fuzzy spatial expressions into geographic coordinate ranges; Professional domain mapping: Matching non-standard personal descriptions and job names against a preset standard industry-job dictionary to generate corresponding standard industry codes and job codes, establishing a one-to-one mapping relationship between non-standard expressions and standard codes; Retrieval granularity selection: Dynamically selecting the retrieval granularity, divided into job-level and task-level types, uniformly represented by the retrieval granularity identifier L; where L=1 represents a job-level index and L=2 represents a task-level index; Multi-hop path planning: For information that cannot be matched by a single keyword or single-round retrieval, it is broken down into multi-hop path step-by-step retrieval tasks, forming a retrieval path sequence, and executing each step of the retrieval operation sequentially. The knowledge partitioning layer is used to build a three-level topic tree of industry-job type-atomic task, configure HNSW physical sub-indexes and retrieve by route at each level; The label matching layer is used to establish a dual-path matching mechanism between explicit physical environment labels and implicit psychological feature labels, and to calculate the comprehensive label matching score. The dynamic optimization layer is used to dynamically adjust the vector retrieval weight and keyword retrieval weight based on semantic similarity and the dispersion of retrieval results, while optimizing the number of candidate positions recalled and optimizing the job matching ranking in real time. The feedback loop layer is used to calculate the employment stability score based on the full-cycle behavioral data of users with employment matching needs from job seeking to leaving their jobs. This drives model iteration and supports parameter rollback and demand drift detection. The parameter rollback mechanism is as follows: after completing the iterative update of the weight parameters, the employment stability score corresponding to the relevant users with employment matching needs is recalculated; the change in employment stability score before and after the iteration is compared, and when the relative decrease in employment stability score compared to before the iteration update is greater than a preset threshold, the weight parameters after the iteration update are restored to the weight parameters before the iteration update. The demand drift detection method is to periodically compare the current job demand vector with the job demand vector of the previous period, and calculate the mean of the absolute values ​​of the differences between them to obtain the job demand drift amount; when the job demand drift amount exceeds the preset drift threshold, the model is triggered to retrain and update the parameters of each layer in the model. The results output module is used to output the final ranking of suitable positions and multi-dimensional matching indicators.

2. The employment supply and demand precision matching device based on a five-layer dynamic structure according to claim 1, characterized in that: Based on the difference between user feature values ​​and job requirement values ​​under the same feature dimension, and combined with preset dimension weights, the semantic similarity between user feature vectors and job feature vectors is calculated comprehensively.

3. The employment supply and demand precision matching device based on a five-layer dynamic structure according to claim 1, characterized in that: The three-level theme tree of industry-job type-atomic task is divided into industry layer, job type layer, and atomic task layer. The industry layer contains the industry names corresponding to several industry categories and is configured with industry indexes. The job type layer is a derivative layer of the industry layer. Each industry covers the job names corresponding to several jobs and is configured with job indexes. The atomic task layer is a derivative layer of the job type layer. Each job type is broken down into the atomic task names corresponding to several atomic tasks and is configured with atomic task indexes. During retrieval, the routing is performed hierarchically by industry layer, job type layer, and atomic task layer: first, the industry index is matched to locate the corresponding industry sub-index; then, the job type index is matched to locate the corresponding job type sub-index; finally, the atomic task index is matched to retrieve the corresponding job position for the atomic task; the HNSW physical sub-index parameters are: the maximum number of edges and the number of candidate constructions for the corresponding hierarchical index, and their initial values ​​are set manually in advance.

4. The employment supply and demand precision matching device based on a five-layer dynamic structure according to claim 1, characterized in that: The dual-path matching mechanism is as follows: When the feature value of the explicit physical environment label meets the requirement value of the adaptive environment requirement label, the value is 1; otherwise, the value is 0. The physical label matching score between the explicit physical environment label vector and the adaptive environment requirement label vector is calculated by weighting the judgment results of the preceding conditions with preset weights. Calculate the relative difference between implicit psychological feature labels and job requirement psychological labels, and combine the weighted sum of the relative differences with preset weights to calculate the psychological label matching score between the implicit psychological feature label vector and the job requirement psychological label vector. The physical label matching score and the psychological label matching score are weighted and fused according to preset weights to obtain the comprehensive label matching score.

5. The employment supply and demand precision matching device based on a five-layer dynamic structure according to claim 1, characterized in that: The dynamic adjustment method for vector retrieval weight and keyword retrieval weight is based on the dynamic allocation of semantic similarity values: when the semantic similarity is lower than a preset threshold, the vector retrieval weight is increased and the keyword retrieval weight is decreased. When the semantic similarity is higher than a preset threshold, the weight of vector retrieval is reduced and the weight of keyword retrieval is increased. The optimization method for the number of candidate positions recalled is as follows: extract the K0 candidate positions initially recalled from the knowledge partitioning layer, calculate the semantic similarity between each candidate position and the user vector of employment matching needs, and form a semantic similarity set. Based on the set of similarity values, the standard deviation of the semantic similarity data within the set is calculated. The obtained standard deviation is divided by the preset theoretical maximum value of the semantic similarity standard deviation to complete the normalization process, and the semantic dispersion (DDI) of the recall results is obtained. The initial recall quantity is adjusted based on the semantic similarity dispersion: with the initial recall quantity as the baseline, when the semantic similarity dispersion increases, the number of recalled positions is reduced; conversely, the number of recalled positions is increased. After the vector retrieval weights and keyword retrieval weights are dynamically adjusted, they are retrieved independently, and the corresponding candidate positions and their matching scores under each independent retrieval are re-output. Then, the matching scores under the independent retrieval corresponding to the vector retrieval weights and keyword retrieval weights are weighted and fused to obtain the ranking benchmark score SP for the candidate positions.

6. The employment supply and demand precision matching device based on a five-layer dynamic structure according to claim 1, characterized in that: The employment stability score of users with relevant employment matching needs is obtained by summing the various behavioral indicators throughout the entire job search to job resignation cycle according to preset weights. The model iterates by updating the weights of each behavioral indicator based on the direction and magnitude of changes in the employment stability score.

7. The employment supply and demand precision matching device based on a five-layer dynamic structure according to claim 6, characterized in that: The employment stability score gradient is obtained based on a pre-constructed employment stability score calculation formula: the weights of each behavioral indicator are determined according to the scoring formula, and the correlation mapping relationship between the system weight parameters and behavioral indicators is combined. The partial derivatives of the employment stability score with respect to each weight parameter are solved by chain derivation, and all partial derivatives are combined to form the employment stability score gradient.

8. The employment supply and demand precision matching device based on a five-layer dynamic structure according to claim 1, characterized in that: The output module uses the ranking benchmark score output by the dynamic optimization layer as the first ranking criterion, and arranges the candidate positions in descending order from high to low to perform a preliminary ranking. Then, it uses the comprehensive tag matching score output by the tag matching layer as the second ranking criterion, and arranges them in descending order from high to low. Based on the preliminary ranking, the order of the candidate positions is adjusted to determine the final arrangement order of the suitable positions. Each candidate job posting output includes multi-dimensional matching indicators such as semantic similarity, physical tag matching score, psychological tag matching score, and comprehensive tag matching score; and simultaneously outputs dynamically adjusted retrieval operator weights and the number of candidate job postings recalled.

Images