A talent label portrait intelligent analysis system based on big data

Abstract

The present invention discloses a talent label portrait intelligent analysis system based on big data, which relates to the field of intelligent analysis technology. It realizes unified access and processing of multi-source heterogeneous talent data through a vector module, and combines BERT embedding with the label semantic evolution mechanism to effectively solve the problems of inconsistent label expression and semantic drift, thereby improving the accuracy of label normalization and portrait structure consistency. The system integrates the label structure relationship with the ability score through a graph construction module to generate a structured original portrait vector, and constructs a semantic collaborative network between labels to provide deep semantic support for portrait calculation. The matching module improves the accuracy of person-job matching through vector alignment and similarity calculation of the job portrait and the candidate portrait, and triggers the label offset analysis and reconstruction scoring mechanism when the matching value is insufficient, thereby realizing dynamic optimization of the portrait and enhancing the system's self-learning and label compensation capabilities.

Description

Technical Field

[0001] The present invention relates to the field of intelligent analysis technology, and in particular to a talent label portrait intelligent analysis system based on big data. Background Art

[0002] Currently, with the development of information-based and digital recruitment management platforms, talent profile modeling, skill and ability assessment, and job profile matching have become key requirements for intelligent recruitment. Against this backdrop, talent profiling technology has gradually evolved into a key implementation method for labeling, structuring, and semanticization. For the construction and analysis of talent label portraits in these areas, current mainstream methods rely primarily on rule-matching-based label extraction or shallow analysis based on keyword extraction. This can lead to inaccuracies in talent portraits, ultimately affecting the evolution of the graph structure and causing the portrait structure to become rigid. Furthermore, for semantic labels such as synonyms in multi-source corpora, existing systems ignore dynamic recognition and semantic fusion, resulting in redundancy in the construction of portrait dimensions. Summary of the Invention

[0003] In view of the above-mentioned problems existing in the prior art, the present application provides a talent label portrait intelligent analysis system based on big data.

[0004] The present disclosure provides a talent label portrait intelligent analysis system based on big data, including:

[0005] The vector module obtains talent data of multiple candidates based on multi-source channels, and constructs candidate label association vectors through dimensionality reduction processing. , and combined with the dynamic nature of label semantics, update the candidate label association vector in real time ;

[0006] The graph construction module will associate the vector with the updated candidate labels. , analyze the ability score of each candidate under the corresponding label dimension to generate the original portrait vector of each candidate , and construct the label graph structure;

[0007] Matching module, by pre-obtaining the portrait vector of the position to be matched , and compare it with the original portrait vector of each candidate talent Perform cosine similarity calculation to obtain the matching value SIM. If the matching value SIM does not exceed the preset threshold, execute the label offset analysis instruction to extract the unqualified labels of the corresponding candidate talents and optimize them to generate a new portrait vector for each candidate talent. ;

[0008] Recommendation module, based on the new profile vector of each candidate , build a corresponding candidate talent profile adaptation ranking table for all positions to be matched to realize the job of job suggestion prompts.

[0009] Optionally, the vector module includes a vector construction unit and a semantic update unit;

[0010] The vector construction unit uses big data technology to obtain talent data of each candidate from multiple sources, including but not limited to resume systems, education platforms, recruitment platforms, project records, behavior logs and social media content. The talent data is then subjected to text parsing, field regularization, missing value filling and noise removal. The BERT embedding model is used to extract label words from the talent data and derive the word vector corresponding to each label word. The word vector corresponding to each label word is matched with the job standard label vector in the job vocabulary by using BERT-CLS to map the word vector corresponding to each label word to the corresponding job standard label in the job vocabulary. The word vector corresponding to each label word is then subjected to dimensionality reduction processing to construct a candidate label association vector. .

[0011] Optionally, the semantic update unit updates the job standard labels in the job vocabulary in real time based on the dynamic nature of the label semantics. Specifically, the semantic change trend of each job standard label in the job vocabulary in the past period is statistically analyzed to construct the label semantic drift entropy coefficient Ht. Based on the value of the label semantic drift entropy coefficient Ht, the semantic stability of the corresponding label in each time period is determined. If the semantics of the corresponding label is determined to be unstable in each time period, the corresponding label is marked as a pseudo-label, and the label set to be replaced is determined. By analyzing the label direction offset between the pseudo-label and each job standard label in the job vocabulary, the label direction offset value Dj is obtained to complete the judgment of the label transformation capability. The label set to be replaced refers to the job standard label vector other than the pseudo-label in the job vocabulary; specifically:

[0012] ;

[0013] in, is the word vector of the g-th pseudo label, is the standard label vector of each position in the label set to be replaced; is the cosine similarity of the angle between the pseudo label and the standard label vector space of each position in the label set to be replaced;

[0014] Extract the job standard label corresponding to the minimum label direction offset value Dj, and fuse it with the corresponding pseudo label to form a parallel extended label. Return the parallel extended label to the vector construction unit and re-match it with the word vector corresponding to each label word in the talent data using BERT-CLS to update the candidate label association vector. .

[0015] Optionally, the graph construction module includes a feature unit, a scoring unit, and a construction unit;

[0016] Feature unit, according to the updated candidate label association vector and talent data, identify the feature items of all candidate talents, analyze the response strength of each feature item under different label dimensions, and use semantic alignment to project each feature item under all label dimensions to obtain the relationship matrix V.

[0017] Optionally, the scoring unit analyzes the ability score of each candidate in the corresponding label dimension based on the relationship matrix V and using the TF-IDF algorithm to obtain the total ability score tss of each candidate in the corresponding label dimension, specifically:

[0018]

[0019] in, Score the overall ability of the corresponding candidate under the i-th label dimension, is the response value of the rth feature item of the corresponding candidate talent under the i-th label dimension, is the importance weight of the rth feature item, r is the feature item number, and m is the total number of feature items;

[0020] The total ability score tss of each candidate talent under all label dimensions is counted to generate the original portrait vector of each candidate talent .

[0021] Optionally, update the candidate label association vector output from the semantic update unit Perform co-occurrence statistics to obtain a co-occurrence matrix. Each element in the co-occurrence matrix represents the number of co-occurrences Yc of two labels in all candidates. Based on the co-occurrence matrix, calculate the association intensity factor RAF between labels;

[0022] Construct a unit, use the label as the node, the association strength factor RAF as the edge between the nodes, and the original portrait vector of each candidate talent As the numerical representation of the label map structure, used to draw the label map structure.

[0023] Optionally, the matching module includes a matching unit and an optimization unit;

[0024] Matching unit, pre-acquires the job portrait vector to be matched and combined with the original portrait vector of each candidate talent , obtain the matching value SIM, the matching value SIM is obtained according to the following formula:

[0025] ;

[0026] in, is the matching value of the a-th candidate talent relative to the b-th position portrait vector to be matched, a is the number of the candidate talent, b is the number of the position to be matched, is the original portrait vector of the a-th candidate talent, is the portrait vector of the bth position to be matched.

[0027] Optionally, the optimization unit compares the matching value SIM with a preset threshold. If the matching value SIM exceeds the preset threshold, it indicates that the current candidate talent is relatively poor relative to the corresponding job profile vector to be matched. If the match is qualified, the current candidate will be included in the candidate talent portrait adaptation ranking table corresponding to the qualified matching position; if the matching value SIM does not exceed the preset threshold, the label offset analysis instruction will be executed.

[0028] Optionally, receive label deviation analysis instructions to analyze the deviation degree of each label dimension of the corresponding candidate talent in the corresponding matching position to obtain the difference , if the difference is satisfied >0, then perform reconstruction score update:

[0029] ;

[0030] in, is the total ability score adjusted by the i-th label dimension in the original portrait vector of the a-th candidate relative to the b-th position to be matched, To optimize the fusion coefficient;

[0031] After performing the reconstruction score update, the original portrait vector of the corresponding candidate talent is regenerated , and marked as the new portrait vector of the corresponding candidate talent .

[0032] Optional, suggestion module, which converts the new profile vector of the corresponding candidate talent Return to the matching unit to regenerate the candidate talent portrait adaptation ranking table, and extract the candidate talent with the largest corresponding matching value SIM value in the regenerated candidate talent portrait adaptation ranking table, and use the candidate talent as the recommended candidate for the corresponding position to be matched.

[0033] The beneficial effects of the present invention are:

[0034] The vector module enables unified access and processing of multi-source, heterogeneous talent data. Combined with BERT embedding and a label semantic evolution mechanism, it effectively addresses label inconsistency and semantic drift, thereby improving the accuracy of label normalization and profile structural consistency. The graph construction module integrates label structural relationships with competency scores to generate structured raw profile vectors. It also constructs a semantic collaboration network between labels, providing deep semantic support for profile computation. The matching module improves the accuracy of person-job matching by aligning and calculating similarity between job profiles and candidate profiles. When the matching value is insufficient, it triggers a label offset analysis and reconstruction scoring mechanism to dynamically optimize profiles and enhance the system's self-learning and label compensation capabilities. The recommendation module generates job matching ranking results based on the new profile vectors, enhancing the personalization and effectiveness of job recommendations. The overall system boasts comprehensive profile semantic representation, timely dynamic label updates, high-precision person-job matching, and adaptable job recommendations, further improving talent selection efficiency and the intelligence of organizational recruitment decisions.

[0035] The system normalizes labels with the same semantics but different expressions from different sources into the standard label library for positions through B, thereby enhancing the comparability of the portraits and the basis for position adaptation. The semantic update unit introduces a dynamic calculation mechanism for the label semantic drift entropy coefficient Ht, which can accurately monitor the changing trends of label semantics in different periods. When the label semantics are unstable, it is automatically identified as a pseudo-label, and label replacement recommendations and fusion generation are performed based on the direction offset value Dj to ensure the continuous consistency of the label system and the semantic structure of the position. Overall, this module improves the accuracy and timeliness of label semantic recognition, making the portrait construction more dynamically adaptable, and providing a label foundation with stable semantics, clear dimensions, and strong structural evolution capabilities for the subsequent establishment of label map structures and position matching.

[0036] By performing full co-occurrence statistics on the updated candidate tag association vectors, constructing a co-occurrence matrix and calculating the association intensity factor (RAF) between tags based on this, it is possible to quantify the combination frequency and degree of synergy of different tags in talent profiles, accurately reflecting the semantic coupling relationship between tags, and overcoming the problems of unclear label edge weights or static structure in traditional graph construction. Secondly, the construction unit uses tags as nodes and RAFs as edges to achieve dynamic construction and continuous evolution of the tag graph structure, allowing the system to adjust the graph structure in real time according to changes in the semantic relationships of tags, improving the true reflection of tag relationship modeling and system interpretability.

[0037] In traditional job-to-person matching systems, when candidates don't fully match the job profile, they are often directly eliminated due to substandard matching values. This ignores the fact that some candidates, despite deviations in individual ability dimensions, may have overall abilities close to the job requirements. This can cause excellent talents to be misjudged as unsuitable, thus affecting matching accuracy and talent utilization efficiency. To address this issue, the system introduces a label offset analysis and reconstruction scoring mechanism through an optimization unit. When the matching value SIM does not reach the threshold, it identifies the label dimensions in the profile that deviate significantly from the job profile, performs a weighted adjustment on their scores, and constructs a new profile vector. Subsequently, the recommendation module regenerates the ranking table based on the optimized profile results, ensuring that potential talents that are relatively suitable for the position are discovered within an acceptable adjustment range. This mechanism effectively improves the system's ability to identify and recommend suboptimal but malleable candidates, enhances the intelligent controllability of the profile system and the robustness of job adaptation, and achieves an upgrade from static matching to dynamic optimization matching, ultimately improving the accuracy of job recommendations and the efficiency of talent utilization. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the technical solutions in the present application or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application.

[0039] Figure 1 This is a system block diagram of the present invention. DETAILED DESCRIPTION

[0040] In order to make the purpose, technical solutions and advantages of this application clearer, the technical solutions of this application will be described clearly and completely below in conjunction with the drawings in this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of them. Based on the embodiments in this application, ordinary technicians in this field can easily make the following inventions without creative work.

[0041] All other embodiments obtained under the premise of the above-mentioned embodiment fall within the scope of protection of this application.

[0042] Example 1

[0043] like Figure 1 As shown, the present invention provides a talent label portrait intelligent analysis system based on big data, including:

[0044] The vector module obtains talent data of multiple candidates based on multi-source channels, and constructs candidate label association vectors through dimensionality reduction processing. , and combined with the dynamic nature of label semantics, update the candidate label association vector in real time ;

[0045] The graph construction module will associate the vector with the updated candidate labels. , analyze the ability score of each candidate under the corresponding label dimension to generate the original portrait vector of each candidate , and construct the label graph structure;

[0046] Matching module, by pre-obtaining the portrait vector of the position to be matched , and compare it with the original portrait vector of each candidate talent Perform cosine similarity calculation to obtain the matching value SIM. If the matching value SIM does not exceed the preset threshold, execute the label offset analysis instruction to extract the unqualified labels of the corresponding candidate talents and optimize them to generate a new portrait vector for each candidate talent. ;

[0047] Recommendation module, based on the new profile vector of each candidate , build a corresponding candidate talent profile adaptation ranking table for all positions to be matched to realize the job of job suggestion prompts.

[0048] In the embodiment of the present invention, a vector module is used to uniformly model multi-source candidate data and construct a candidate label association vector after dimensionality reduction, effectively solving problems such as heterogeneous multi-channel data, label expression differences, and semantic redundancy.

[0049] The introduction of a dynamic tag semantic update mechanism can identify and correct tag drift, ensuring the timeliness and industry adaptability of the portrait dimension;

[0050] The graph construction module structures the candidate's ability labels, forming a label graph and original portrait vectors, providing a structural basis for subsequent job matching; the matching module uses cosine similarity to achieve high-precision matching between candidate and job portraits, and at the same time introduces a portrait optimization mechanism. When the SIM value is lower than the threshold, it can automatically identify and optimize unqualified label dimensions, thereby improving the adaptability between portraits and jobs; the recommendation module can generate a job adaptation ranking table based on the new portrait vector, realize dynamic matching recommendations between candidates and jobs, and further improve the accuracy and intelligence of job recommendations.

[0051] Example 2

[0052] like Figure 1 As shown, specifically: the vector module includes a vector construction unit and a semantic update unit;

[0053] The vector construction unit uses big data technology to obtain talent data of each candidate from multiple sources, including but not limited to resume systems, education platforms, recruitment platforms, project records, behavior logs and social media content. The talent data is then subjected to text parsing, field regularization, missing value filling and noise removal. The BERT embedding model is used to extract label words from the talent data and derive the word vector corresponding to each label word. The word vector corresponding to each label word is matched with the standard label vector of the position in the position vocabulary by using BERT-CLS (equivalent to identifying and merging field items with the same semantics but different representations in different sources), so as to map the word vector corresponding to each label word to the standard position label in the corresponding position vocabulary, and perform dimensionality reduction on the word vector corresponding to each label word to construct a candidate label association vector. .

[0054] The BERT embedding model is a bidirectional Transformer-structured language model proposed by Google. It can model the semantics of words or sentences from context and is used to convert label words or candidate phrases into high-dimensional semantic vector representations.

[0055] Using BERT-CLS matching is one of the semantic matching methods;

[0056] The vector module is used to solve the problems of confusing data sources and invisible talent characteristics. It converts text information, behavioral data, etc. into digital form (vectors). This vectorization processing can also discover hidden abilities. For example, if a person does not explicitly write "project management", the behavioral data and language structure may show that he has done it.

[0057] Specifically, label words refer to keywords used to express certain abilities, job characteristics, technical directions or behavioral attributes. They constitute the semantic units for scoring each dimension in the talent portrait.

[0058] Word vectors convert natural language units such as label words into computable dense vector representations through mathematical modeling, so that computational operations such as comparison, clustering, and matching can be performed in the model.

[0059] The semantic update unit updates the job standard labels in the job vocabulary in real time based on the dynamic nature of label semantics. Specifically, it collects statistics on the semantic change trends of each job standard label in the job vocabulary over the past period to construct the label semantic drift entropy coefficient Ht. The label semantic drift entropy coefficient Ht is obtained by the following formula:

[0060] ;

[0061] in, is the label semantic drift entropy coefficient of the i-th label, is the proportion of label semantic changes in the kth period (i.e., the proportion of label semantic changes in this period), A small positive number to prevent log(0), usually ; n is the total length of the past period, k is the number of each period in the past period, and i is the number of the label in the candidate label association vector; is the entropy calculation term with base 2, which indicates the degree of uncertainty of semantic change;

[0062] The tag semantic drift entropy coefficient Ht is used to measure the degree of fluctuation of a tag's semantics over a period of time. This formula is essentially an information entropy formula that reflects the "uncertainty of the tag semantic change trend";

[0063] Specifically, if the label semantic drift entropy coefficient Ht is large, it means that the label semantics changes frequently in the time dimension, the direction is scattered, and the trend is unstable; if the label semantic drift entropy coefficient Ht is close to 0, it means that the label semantics remain basically unchanged in each time period and the semantics tend to be stable.

[0064] By collecting statistics on the semantic change trends of standard labels for various positions in the past period, we can automatically identify the semantic change trends of labels and improve the long-term accuracy of the system. Label updates can be propagated in the graph to ensure the overall consistency of the system, and retain the label evolution path and version management records to improve the interpretability of the system.

[0065] Based on the value of the label semantic drift entropy coefficient Ht, the semantic stability of the corresponding label in each time period is determined. If the semantics of the corresponding label is determined to be unstable in each time period, the corresponding label is marked as a pseudo-label, and the label set to be replaced is determined. By analyzing the label direction offset between the pseudo-label and the standard labels of each position in the position vocabulary, the label direction offset value Dj is obtained to complete the judgment of the label transformation capability. Among them, the label set to be replaced refers to the standard label vector of the position in the position vocabulary except the pseudo-label; specifically:

[0066] ;

[0067] in, is the word vector of the g-th pseudo label, indicating the semantic difference, is the standard label vector of each position in the label set to be replaced; It is the cosine similarity of the angle between the pseudo label and the standard label vector space of each position in the label set to be replaced. The larger the value, the closer it is (the maximum is 1). It means converting similarity into difference, so that the smaller the value, the closer the direction; g is the number of the pseudo label;

[0068] This formula is used to measure "which candidate label is closest to the direction of the original label" in the semantic vector space; its essence is the "replacement ability judgment" of label semantics; the smaller the label direction offset value Dj, the more suitable the corresponding job standard label in the label set to be replaced is as the new semantic representative of the current pseudo-label.

[0069] Extract the job standard label corresponding to the minimum label direction offset value Dj, and fuse it with the corresponding pseudo label to form a parallel extended label, thereby generating a new label;

[0070] It should be noted that the original pseudo-label and the original job standard label to be fused are retained, and only a new label is generated based on the fusion;

[0071] Regarding the dynamic nature of label semantics, take "full-stack engineer" as an example: in the early stages, this label may represent the ability to perform both front-end and back-end programming tasks. However, in the later stages, new capabilities such as cloud native technologies, automated deployment, container orchestration, and security architecture may need to be mastered. If the definition of this label in the original portrait system is not updated accordingly, it will cause label mapping distortion or delayed capability assessment. Therefore, it is necessary to consider the dynamic nature of label semantics and the impact it will have on later labeling. This part avoids the rigidity of the portrait structure and better adapts to actual needs such as job updates, label expansion, and talent label migration.

[0072] Return the parallel expanded labels to the vector construction unit and re-match them with the word vectors corresponding to each label word in the talent data using BERT-CLS to update the candidate label association vector .

[0073] In an embodiment of the present invention, first, the vector construction unit can extract candidate data from multiple sources and use the BERT embedding model to obtain word vectors for label words, thereby automatically matching and normalizing labels with different expressions but consistent semantics (such as "full-stack engineer" and "Fullstack Developer") in the semantic space, thereby avoiding label redundancy and semantic distortion problems.

[0074] Secondly, the semantic update unit introduces the label semantic drift entropy coefficient Ht to dynamically monitor the semantic change trends of labels in the job vocabulary over different time periods. It can automatically identify pseudo-labels when semantic drift occurs and recommend relatively better alternative expressions. For example, if the "AI Architect" label tends to "AI Platform Architecture Design" in the context of new technological development, the system will automatically analyze its directional offset value Dj and fuse it to form a new parallel extended label "AI Architect - Platform Direction", thereby achieving an evolutionary update of the label map. Through the above mechanism, the system can continuously optimize the candidate label association vector, improve the accuracy, update capability and system adaptability of talent label portraits, and provide a solid foundation for job matching and recommendation.

[0075] Example 3

[0076] like Figure 1 As shown, specifically: the graph construction module includes feature units, scoring units and construction units;

[0077] Feature unit, according to the updated candidate label association vector And talent data, identify the feature items of all candidate talents, analyze the response strength of each feature item under different label dimensions, and use semantic alignment to project each feature item under all label dimensions to obtain the relationship matrix V. The expression of the relationship matrix V is:

[0078] ;

[0079] Among them, M is the number of label dimensions and K is the number of feature items.

[0080] Semantic alignment is used for multi-label projection and label-feature mapping. Semantic alignment involves projecting input items such as feature items, keywords, and label words into a multi-label semantic space, allowing them to be represented responsively across multiple label dimensions.

[0081] The scoring unit analyzes the ability score of each candidate in the corresponding label dimension based on the relationship matrix V and the TF-IDF algorithm to obtain the total ability score tss of each candidate in the corresponding label dimension, specifically:

[0082] ;

[0083] in, Score the overall ability of the corresponding candidate under the i-th label dimension, is the response value of the rth feature item of the corresponding candidate under the i-th label dimension, that is, the support of the feature item for the label, is the importance weight of the rth feature item, r is the feature item number, and m is the total number of feature items;

[0084] The total ability score (TSS) is used to form the "label strength vector" in the subsequent talent profile. It is the core numerical expression of the label profile. It is the weighted sum of the expression strengths of multiple feature dimensions on the current label. The result is the candidate's score under this label semantics. The higher the value, the stronger the ability score.

[0085] The total ability score tss of each candidate talent under all label dimensions is counted to generate the original portrait vector of each candidate talent ;

[0086] Among them, the importance weight of the rth feature item Obtained by the following formula:

[0087] ;

[0088] in, is the frequency of occurrence of the rth feature item in the talent data of the corresponding candidate talent, is the total number of candidate talents, is the number of talent data containing the rth feature item, that is, the document frequency. Used to lower the weight of high-frequency labels so that common words do not interfere with the true discrimination;

[0089] In an embodiment of the present invention, the graph construction module of the present invention achieves a structured, semantic, and quantifiable expression of talent profile construction through the collaborative design of feature units, scoring units, and construction units, with significant technical benefits. First, the feature unit automatically identifies the candidate's feature items based on a joint analysis of candidate label association vectors and talent data, and projects them across different label dimensions through semantic alignment, constructing a response matrix V that reflects the label-feature relationship, thereby enhancing the semantic relevance of the feature expression.

[0090] Secondly, the scoring unit uses the TF-IDF method to assign importance weights to each feature item, and combines its response values under different label dimensions to calculate the ability score TSS under each label dimension, thereby generating a complete original portrait vector, ensuring the quantifiable and differentiated expression of the portrait. This design effectively improves the expression accuracy and semantic integrity of the talent portrait for job labels, and provides accurate and structured portrait input for subsequent map construction and job matching.

[0091] Example 4

[0092] like Figure 1 As shown, specifically: the updated candidate label association vector output in the semantic update unit Perform co-occurrence statistics to obtain a co-occurrence matrix. Each element in the co-occurrence matrix represents the number of co-occurrences Yc of two labels in all candidates. Based on the co-occurrence matrix, calculate the association intensity factor RAF between labels, specifically:

[0093] ;

[0094] in, is the number of co-occurrences of the i-th label and the j-th label in all candidates, is the correlation strength factor between the i-th label and the j-th label, is the independent occurrence frequency of the i-th label, is the independent occurrence frequency of the jth label, i and j are the label numbers in the candidate label association vector;

[0095] Since labels themselves do not exist in isolation, their semantic utility must be understood in conjunction with "structure". Only through the individual's rating and the group's structure can the true semantic combination effect of the labels be restored. Therefore, when constructing the label graph structure, the association strength factor (RAF) between labels must be considered.

[0096] The RAF (association strength factor) between tags reflects the strength of association between tags. A higher value indicates a more frequent combination. It is used as the basis for edge weights in constructing the tag graph structure. The more relevant the tags, the thicker the edges. The system considers that the tags may be combined to form a composite capability, reflecting the semantic synergy and collocation frequency between tags.

[0097] Construct a unit, use the label as the node, the association strength factor RAF as the edge between the nodes, and the original portrait vector of each candidate talent As the numerical representation of the label map structure, the label map structure is drawn, in which the original portrait vector of each candidate talent It is used for external function matching, and the association intensity factor RAF is used for the structural understanding and evolution of the label map structure within the system.

[0098] The matching module includes a matching unit and an optimization unit;

[0099] Matching unit, pre-acquires the job portrait vector to be matched and combined with the original portrait vector of each candidate talent , obtain the matching value SIM, the matching value SIM is obtained according to the following formula:

[0100] ;

[0101] in, is the matching value of the a-th candidate talent relative to the b-th position portrait vector to be matched, a is the number of the candidate talent, b is the number of the position to be matched, is the original portrait vector of the a-th candidate talent, is the portrait vector of the bth position to be matched; is the modulus of the original portrait vector of the a-th candidate talent, is the modulus of the bth job portrait vector to be matched.

[0102] In an embodiment of the present invention, global co-occurrence statistics are performed using semantically updated candidate tag association vectors to construct a co-occurrence matrix. Based on this, the association intensity factor (RAF) between tags is calculated, achieving a quantitative expression of the semantic synergy between tags. A higher RAF value indicates a more frequent joint occurrence between tags, effectively revealing the combination preferences between tags and the semantic collocation patterns of job positions. The construction unit uses tags as nodes, RAF as edge weights, and portrait vectors as numerical representations of tags to draw a tag graph structure, thereby providing support for the structured relationships between tags.

[0103] Secondly, the matching module measures the degree of ability adaptation by calculating the cosine similarity SIM of the original portrait vectors between candidates and positions, quantifies and standardizes the matching results, and supports ranking screening. This mechanism not only realizes the quantitative alignment of multi-dimensional semantics, but also provides a structural and explanatory basis for job recommendations and person-job matching, thereby improving the transparency and credibility of the intelligent recommendation system.

[0104] Example 5

[0105] like Figure 1 As shown, specifically: the optimization unit compares the matching value SIM with the preset threshold. If the matching value SIM exceeds the preset threshold, it indicates that the current candidate talent is relatively poor relative to the corresponding job profile vector to be matched. If the match is qualified, the current candidate will be included in the candidate talent portrait adaptation ranking table corresponding to the qualified matching position; if the matching value SIM does not exceed the preset threshold, the label offset analysis instruction will be executed.

[0106] Because the profile itself may not be fully expressive, the candidate's label ability score is not absolutely accurate. There may be information missing in the feature modeling stage, insufficient semantic coverage in the label projection stage, or ambiguity or redundancy in the candidate's description. This will cause the ability profile to not fully cover or express the job profile. Therefore, if the matching value SIM does not exceed the preset threshold, consider executing the label offset analysis instruction;

[0107] Receive label deviation analysis instructions, analyze the deviation degree of each label dimension of the corresponding candidate talent in the corresponding matching position, and obtain the difference :

[0108] ;

[0109] is the total ability score of the original portrait vector of the a-th candidate under the i-th label dimension, is the total ability score under the i-th label dimension in the b-th job profile vector to be matched, is the difference between the i-th label dimension of the original portrait vector of the a-th candidate and the b-th position to be matched; if the difference is satisfied >0, then perform reconstruction score update:

[0110] ;

[0111] in, is the total ability score adjusted by the i-th label dimension in the original portrait vector of the a-th candidate relative to the b-th position to be matched, To optimize the fusion coefficient, it is used as an adjustment factor when performing reconstruction scoring to control the adjustment range and prevent the label ability score from being over-corrected. The value range is between 0 and 1.

[0112] Candidates may have potential in the corresponding label dimension, but they are not fully expressed during the portrait generation stage. At this time, we need to try to guide the portrait to better align with the semantic structure of the job portrait, and achieve the fitting and optimization of the semantic dimension. In other words, the score update does not change the reality, but strengthens the expression. This is similar to the principle of "semantic embedding optimization" in natural language processing. Instead of directly overwriting the old value (to avoid losing the original feature), the difference is used as incremental information for fusion correction. Controlling the magnitude of the correction is essentially about performing a "weighted sliding update of the score" to ensure that the profile evolves smoothly rather than suddenly. This is because the goal of optimization is not to cover the job profile, but to "bridge the capability gap" and make the profile converge towards the target profile without destroying its underlying structure.

[0113] After performing the reconstruction score update, the original portrait vector of the corresponding candidate talent is regenerated , and marked as the new portrait vector of the corresponding candidate talent .

[0114] The recommendation module converts the new portrait vector of the corresponding candidate talent into Return to the matching unit to regenerate the candidate talent portrait adaptation ranking table, and extract the candidate talent with the largest corresponding matching value SIM value in the regenerated candidate talent portrait adaptation ranking table, and use the candidate talent as the recommended candidate for the corresponding position to be matched.

[0115] In an embodiment of the present invention, the system automatically determines whether the candidate and the job profile are matched by comparing the matching value SIM with a preset threshold, ensuring that the recommended candidate has actual job competency and improving the reliability of job recommendations; when the matching value is insufficient, the system can trigger a label offset analysis mechanism to calculate the ability difference between the candidate and the job under each label dimension, identify the specific ability labels with deficiencies, and then perform reconstruction scoring updates on the dimensions with deviations to generate new label ability scores, thereby achieving "local reinforcement" and "ability repair" of the candidate portrait and enhancing the system's adaptive adjustment capabilities. Secondly, after updating the portrait, the system can re-rank and output relatively better recommended candidates based on the new portrait, forming a recommendation feedback loop.

[0116] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A talent label portrait intelligent analysis system based on big data, characterized by: include: The vector module obtains talent data of multiple candidates based on multi-source channels, and constructs candidate label association vectors through dimensionality reduction processing. , and combined with the dynamic nature of label semantics, update the candidate label association vector in real time ; The graph construction module will associate the vector with the updated candidate labels. , analyze the ability score of each candidate under the corresponding label dimension to generate the original portrait vector of each candidate , and construct the label graph structure; Matching module, by pre-obtaining the portrait vector of the position to be matched , and compare it with the original portrait vector of each candidate talent Perform cosine similarity calculation to obtain the matching value SIM. If the matching value SIM does not exceed the preset threshold, execute the label offset analysis instruction to extract the unqualified labels of the corresponding candidate talents and optimize them to generate a new portrait vector for each candidate talent. ; Recommendation module, based on the new profile vector of each candidate , build a corresponding candidate talent profile adaptation ranking table for all positions to be matched to realize the job of job suggestion prompts.

2. The talent tag portrait intelligent analysis system based on big data according to claim 1 is characterized by: The vector module includes a vector construction unit and a semantic update unit; The vector construction unit uses big data technology to obtain talent data of each candidate from multiple sources, including resume systems, education platforms, recruitment platforms, project records, behavior logs and social media content. The talent data is then parsed, fields are regularized, missing values are filled and noise is removed. The BERT embedding model is used to extract label words from the talent data and derive the word vector corresponding to each label word. The word vector corresponding to each label word is matched with the standard label vector in the job vocabulary by using BERT-CLS to map the word vector corresponding to each label word to the standard job label in the corresponding job vocabulary. The word vector corresponding to each label word is then subjected to dimensionality reduction processing to construct a candidate label association vector. .

3. The big data-based talent tag portrait intelligent analysis system according to claim 2, characterized in that: The semantic update unit updates the job standard labels in the job vocabulary in real time based on the dynamic nature of label semantics. Specifically, the semantic change trend of each job standard label in the job vocabulary in the past period is statistically analyzed to construct the label semantic drift entropy coefficient Ht. Based on the value of the label semantic drift entropy coefficient Ht, the semantic stability of the corresponding label in each time period is determined. If the semantics of the corresponding label is determined to be unstable in each time period, the corresponding label is marked as a pseudo-label, and the label set to be replaced is determined. By analyzing the label direction offset between the pseudo-label and each job standard label in the job vocabulary, the label direction offset value Dj is obtained to complete the judgment of the label transformation capability. Among them, the label set to be replaced refers to the job standard label vector other than the pseudo-label in the job vocabulary; specifically: ； in, is the word vector of the g-th pseudo label, is the standard label vector of each position in the label set to be replaced; is the cosine similarity of the angle between the pseudo label and the standard label vector space of each position in the label set to be replaced; Extract the job standard label corresponding to the minimum label direction offset value Dj, and fuse it with the corresponding pseudo label to form a parallel extended label. Return the parallel extended label to the vector construction unit and re-match it with the word vector corresponding to each label word in the talent data using BERT-CLS to update the candidate label association vector. .

4. The big data-based talent tag portrait intelligent analysis system according to claim 3, characterized in that: The graph construction module includes feature units, scoring units and construction units; Feature unit, according to the updated candidate label association vector and talent data, identify the feature items of all candidate talents, analyze the response strength of each feature item under different label dimensions, and use semantic alignment to project each feature item under all label dimensions to obtain the relationship matrix V.

5. The talent tag portrait intelligent analysis system based on big data according to claim 4 is characterized by: The scoring unit analyzes the ability score of each candidate in the corresponding label dimension based on the relationship matrix V and the TF-IDF algorithm to obtain the total ability score tss of each candidate in the corresponding label dimension, specifically: ； in, Score the overall ability of the corresponding candidate under the i-th label dimension, is the response value of the rth feature item of the corresponding candidate talent under the i-th label dimension, is the importance weight of the rth feature item, r is the feature item number, and m is the total number of feature items; The total ability score tss of each candidate talent under all label dimensions is counted to generate the original portrait vector of each candidate talent .

6. The talent tag portrait intelligent analysis system based on big data according to claim 5 is characterized by: The updated candidate label association vector output from the semantic update unit Perform co-occurrence statistics to obtain a co-occurrence matrix. Each element in the co-occurrence matrix represents the number of co-occurrences Yc of two labels in all candidates. Based on the co-occurrence matrix, calculate the association intensity factor RAF between labels; Construct a unit, use the label as the node, the association strength factor RAF as the edge between the nodes, and the original portrait vector of each candidate talent As the numerical representation of the label map structure, used to draw the label map structure.

7. The talent tag portrait intelligent analysis system based on big data according to claim 6, characterized in that: The matching module includes a matching unit and an optimization unit; Matching unit, pre-acquires the job portrait vector to be matched and combined with the original portrait vector of each candidate talent , obtain the matching value SIM, the matching value SIM is obtained according to the following formula: ； in, is the matching value of the a-th candidate talent relative to the b-th position portrait vector to be matched, a is the number of the candidate talent, b is the number of the position to be matched, is the original portrait vector of the a-th candidate talent, is the portrait vector of the bth position to be matched.

8. The talent tag portrait intelligent analysis system based on big data according to claim 7, characterized in that: The optimization unit compares the matching value SIM with the preset threshold. If the matching value SIM exceeds the preset threshold, it indicates that the current candidate talent is relatively poor relative to the corresponding job profile vector to be matched. If the match is qualified, the current candidate will be included in the candidate talent portrait adaptation ranking table corresponding to the qualified matching position; if the matching value SIM does not exceed the preset threshold, the label offset analysis instruction will be executed.

9. The big data-based talent tag portrait intelligent analysis system according to claim 8, characterized in that: Receive label deviation analysis instructions, analyze the deviation degree of each label dimension of the corresponding candidate talent in the corresponding matching position, and obtain the difference , if the difference is satisfied >0, then perform reconstruction score update: ； in, is the total ability score adjusted by the i-th label dimension in the original portrait vector of the a-th candidate relative to the b-th position to be matched, To optimize the fusion coefficient; is the difference between the i-th label dimension in the original portrait vector of the a-th candidate and the b-th position to be matched; After performing the reconstruction score update, the original portrait vector of the corresponding candidate talent is regenerated , and marked as the new portrait vector of the corresponding candidate talent .

10. The talent tag portrait intelligent analysis system based on big data according to claim 9, characterized in that: The recommendation module converts the new portrait vector of the corresponding candidate talent into Return to the matching unit to regenerate the candidate talent portrait adaptation ranking table, and extract the candidate talent with the largest corresponding matching value SIM value in the regenerated candidate talent portrait adaptation ranking table, and use the candidate talent as the recommended candidate for the corresponding position to be matched.

Images