Talent background investigation method based on multi-source data evaluation

Abstract

The invention relates to the technical field of background investigation, and particularly discloses a talent background investigation method based on multi-source data evaluation, which comprises the steps of constructing a dynamic talent knowledge graph, performing semantic alignment processing on multi-source cross-modal data, evaluating data quality based on a multi-dimensional confidence model, and constructing a confidence closed-loop correction mechanism. Performing intelligent correction on the low-confidence data, and reversely updating the credibility weight of the data source according to a correction result; an event-driven incremental update architecture is adopted, and dynamic events of multi-source data are captured in real time through a distributed event awareness engine. According to the method, the dynamic knowledge graph is used as a unified carrier, and the multi-source data is upgraded from fragmented storage to a semantic association network, so that the talent background survey data is converted from static piling to a dynamically evolved knowledge network, and the accuracy and timeliness of evaluation are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of background investigation, and in particular to a talent background investigation method based on multi-source data evaluation. Background Art

[0002] In modern corporate recruitment and talent management, talent background checks are a critical step in assessing a candidate's overall quality, work ability, and suitability. Their accuracy and comprehensiveness directly impact the quality of a company's talent decisions. With the rapid development of information technology, the data sources available for talent background checks are becoming increasingly diverse, encompassing resume information from recruitment platforms, public social media activity, past employer reviews, and professional qualification certification data.

[0003] However, current talent background data integration processes suffer from significant flaws. For one thing, traditional data integration often relies on simple splicing or shallow correlation, failing to deeply explore the inherent logical connections between disparate data sources. For example, mechanically listing a candidate's educational background, work experience, and skill certificates makes it impossible to identify underlying patterns through data fusion, such as the predictive value of the match between a specific skill certificate and specific work experience on job performance. Consequently, the integrated data struggles to form a systematic picture of talent capabilities.

[0004] On the other hand, data quality control mechanisms are missing or incomplete. Multi-source data collection is prone to missing data, errors, and duplications. For example, social media data can be fragmented and subjective, and past employer evaluations can be biased due to varying standards. Existing technologies lack a quantitative assessment system for data integrity, accuracy, and consistency, often relying solely on raw data for subsequent analysis. This results in evaluations that are distorted by noisy data and can lead to misjudgments or omissions.

[0005] Furthermore, data updates lack real-time performance. Talent information is constantly changing, such as new project experience and skill development records. Traditional data integration methods rely on periodic batch processing, making it difficult to capture and integrate new data in real time. This causes background check results to lag behind the talent's actual situation, affecting companies' assessment of candidates' current status. Summary of the Invention

[0006] The purpose of the present invention is to provide a talent background investigation method based on multi-source data evaluation to solve the technical problem that insufficient data fusion depth results in the integrated data being unable to accurately reflect the true status of the talent, affecting the accuracy and timeliness of the background investigation.

[0007] The purpose of the present invention can be achieved through the following technical solutions: This method, based on multi-source data evaluation, uses knowledge graphs as its core carrier. Through the innovative design of "semantic fusion, quality quantification, and real-time evolution," it addresses the pain points of traditional multi-source data integration, such as "broken associations, fuzzy quality, and delayed updates." Its technical focus and creativity are reflected in the following three aspects: 1. Core Carrier: Construction and Semantic Fusion of Dynamic Talent Knowledge Graph Three-layer structured design of entity-attribute-relationship: Entity layer: Abstract candidates, academic qualifications, projects, skills, etc. into independent entities, assign a unique identifier to each entity (such as a global ID based on the ID card hash), and ensure entity matching across data sources (such as associating "education records on the China Higher Education Student Information and Career Center" and "education qualifications on the recruitment platform resume" to the same "education entity").

[0008] Attribute layer: Defines standardized attributes based on the industry's common capability model (e.g., "programming skills" includes "language type," "proficiency level," and "number of project applications"), and supports custom extensions (e.g., adding a "large model training experience" attribute for AI positions), solving the problem of "inconsistent titles and confusing standards" in multi-source data.

[0009] Relationship layer: Automatically mine implicit semantic associations between entities through graph neural networks (GNNs), for example: When the overlap between the "keywords of the candidate's published papers" and the "core technical requirements of the recruitment position" is ≥70%, the relationship label "highly matched in professional field" is automatically generated; When "project budget completion rate" is greater than 90% and "customer satisfaction" is greater than 4.5 / 5, the associated weight of "cost control ability" and "project management ability" will be strengthened (weight increased by 20%).

[0010] Semantic alignment mechanism for cross-modal data: Design unified semantic mapping rules for cross-modal data such as text (letters of recommendation), images (certificates), and numerical values ​​(performance scores): Text data is converted into semantic vectors through the BERT model and matched with attribute vectors such as "job description" and "skill requirements" in the knowledge graph (for example, the attribute value of "leading core system development" in a recommendation letter is increased for "project leadership"). After OCR recognition and blockchain verification, the image data extracts key information such as "certificate number" and "validity period" and maps them to the attributes of the "skill certification" entity; After normalization, the numerical data is used as a quantitative indicator of the entity attributes (for example, 85% of the “project cycle compliance rate” corresponds to 3.5 / 5 points of the “execution ability” attribute).

[0011] 2. Quality Control: Dynamic Quantitative Evaluation System Based on Confidence: Abandoning the traditional binary judgment of "pass / fail", a quantitative evaluation model centered on "data confidence" is constructed to achieve dynamic quality control: Three-dimensional confidence calculation model: Data confidence is calculated from the three dimensions of "data source reliability, data timeliness, and logical consistency" (values ​​range from 0 to 1): Data source reliability: Assign an initial weight to each data source (e.g., 0.9 for government databases, 0.5 for social media), and dynamically adjust it based on historical verification pass rates (e.g., for every 10% decrease in the matching rate between a platform's data and authoritative data, the weight is reduced by 0.1); Data timeliness: Design a decay function so that the confidence level is multiplied by 0.8 for every six months the data is generated (e.g., the confidence level of performance data from two years ago is automatically reduced to the initial value of 0.64). Logical consistency: The rationality of the data is verified through the knowledge graph rule engine (for example, if the person has "2 years of work experience" but "3 years of management experience", the confidence level will be directly deducted by 0.4).

[0012] Confidence-driven closed-loop correction mechanism: Low-confidence data (<0.6) automatically triggers correction: missing values ​​are filled based on the attribute distribution of similar entities in the knowledge graph (such as "average number of projects for the same position"); conflicting values ​​are verified and corrected by calling high-confidence data sources (such as social security records); The correction result is fed back to the data source weight (for example, if the "salary data" provided by a headhunter has a deviation of more than 20% for three consecutive times, its reliability weight will be permanently reduced by 0.2).

[0013] 3. Real-time Evolution: Event-driven Incremental Updates to the Knowledge Graph Breaking through the traditional "scheduled batch update" model, we can achieve real-time dynamic evolution of the knowledge graph: Distributed event perception engine: Deploy lightweight monitoring nodes to capture dynamic changes in multi-source data in real time: Structured data sources (such as enterprise HR systems): Use the rule engine to identify events such as "project acceptance" and "skill certification pass" to trigger data updates; Unstructured data sources (such as professional forums): Use NLP technology to extract information such as "candidate-published patents" and "industry award mentions" to generate incremental data.

[0014] Incremental update algorithm: Update only changed data: New information is located in the knowledge graph through entity matching, and only related attributes and relationships are updated (for example, after adding PMP certification, only the "certification status" and associated weight of the "project management skills" entity are updated); Conflict resolution: When new and old data conflict (e.g., inconsistent "time of leaving"), prioritize the data with higher confidence and retain conflicting records in the graph (marking the confidence difference); Freshness mark: After each update, "data freshness" is generated (calculated based on the difference between the update time and the current time), so that subsequent evaluation models can prioritize data with freshness greater than 0.8.

[0015] The beneficial effects of this invention include: using a dynamic knowledge graph as a unified carrier to upgrade multi-source data from "fragmented storage" to a "semantic association network"; using a confidence metric model to achieve precise data quality control rather than simple screening; and leveraging event-driven incremental updates to ensure that the knowledge graph reflects the latest status of talent in real time. These three factors work together to transform talent background investigation data from a "static stack" to a "dynamically evolving knowledge network," significantly improving the accuracy and timeliness of assessments. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The present invention will be further described below with reference to the accompanying drawings.

[0017] Figure 1 It is a flow chart of a talent background investigation method based on multi-source data evaluation according to the present invention. DETAILED DESCRIPTION

[0018] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0019] See also Figure 1 As shown, the present invention is a talent background investigation method based on multi-source data evaluation, with a three-layer structured design of entity-attribute-relationship: Entity layer: Abstract candidates, academic qualifications, projects, skills, etc. into independent entities, assign a unique identifier to each entity (such as a global ID based on the ID card hash), and ensure entity matching across data sources (for example, associating "education records on the China Higher Education Student Information and Career Center" and "education qualifications on the recruitment platform resume" to the same "education entity"); The entity layer is the basic component of the knowledge graph and can be formally expressed as: ε={e1,e2,...,e n}; Among them, each entity e i By unique identifier ID (e i ) and entity type Type (e i ) definition, and satisfies: ID (e i)≠ID(e j ), i≠j; For example: Candidate entity: e1=(ID=Hash(ID number), Type=Candidate); Education entity: e2=(ID=Hash(graduation certificate number), Type=Education).

[0020] Attribute layer: Defines standardized attributes based on the industry's common capability model (e.g., "programming skills" includes "language type," "proficiency level," and "number of project applications"), and supports custom extensions (e.g., adding a "large model training experience" attribute for AI positions), solving the problem of "inconsistent titles and confusing standards" in multi-source data.

[0021] The attribute layer defines a set of attributes for each entity, which can be expressed as: A(e i ) = {a1 （i） , a1 （i） ,...,a m （i）}; Each attribute a j （i） is a triple: a j （i） =(Name j （i） ,Value j （i） ,Type j （i） ), respectively represent the attribute name, which is taken from the predefined ontology library; the attribute value, which can be a number, text, Boolean value, etc.; the attribute type, which defines the value range and constraints.

[0022] Relationship layer: Automatically mine implicit semantic associations between entities through graph neural networks (GNN); The relationship layer defines the association between entities and can be expressed as: R={r1,r2,...,r k}; Each relation r l is a four-tuple; r l ={e i ,e j ,RelType l ,Weight l}, e i , e j Are two entities associated, RelType l ,Weight l Represent the relationship type and relationship weight respectively.

[0023] The complete mathematical expression of the knowledge graph: G=(ε,A,R).

[0024] The cross-modal data semantic alignment mechanism aims to transform data from different data sources and formats (text, images, numerical values, etc.) into knowledge representations in a unified semantic space, resolving the issues of "data format incompatibility and broken semantic associations" in traditional integration methods. This mechanism achieves deep fusion of multi-source heterogeneous data through three key steps: data preprocessing → feature extraction → semantic mapping, ultimately constructing a unified talent knowledge graph. Specifically, it includes: Text data cleaning: Segment text, remove stop words, identify entities (such as names, company names), and extract key skill descriptions (such as "proficient in using Python for deep learning model development") from texts such as resumes and recommendation letters.

[0025] Image data enhancement: Perform OCR recognition on image data such as certificate scans and conference speech video screenshots, extract text information and perform semantic verification (such as verifying the authenticity of certificates through blockchain).

[0026] Normalization of numerical data: Standardize numerical data such as performance scores and project budgets to unify the dimensions (e.g., converting scores from 0-100 into capability values ​​in the range of 0-1).

[0027] Text data vectorization: Use the pre-trained BERT model to convert text into semantic vectors to capture contextual associations (such as the semantic association between "deep learning" and "neural network").

[0028] Image feature extraction: CNN (convolutional neural network) is used to extract features from images of certificates and work results to generate multi-dimensional feature vectors (such as certificate type and degree of visualization of project results).

[0029] Numerical data encoding: Mapping numerical data to points in a high-dimensional feature space (e.g., mapping "project budget completion rate 95%" to the specific value of the "cost control capability" attribute).

[0030] Build a mapping dictionary: Establish cross-modal entity association rules (such as "Python" code repository contribution record → "Python programming ability" attribute in the knowledge graph).

[0031] Dynamic weight adjustment: Automatically adjust the association weight based on the credibility and timeliness of the data source (for example, the weight of educational information in the government database is 0.9, and the weight of social media skill tags is 0.4).

[0032] Hidden relationship mining: Discovering potential relationships through graph neural networks (GNNs), such as: The correlation between the candidate's GitHub code contribution frequency and the "work engagement" attribute; The semantic matching degree between keywords extracted from industry conference speech videos and recruitment requirements.

[0033] Construct evaluation indicators: quantify the alignment effect through indicators such as entity linking accuracy and relationship prediction recall.

[0034] Feedback correction mechanism: If an alignment error is found (such as incorrectly associating "UI design" with "front-end development"), the mapping rules are automatically adjusted and the encoder is retrained.

[0035] Incremental learning: When new data is added, only the affected local graph is updated (for example, when adding "Metaverse Project Experience", only the associated weights of the relevant skill nodes need to be adjusted).

[0036] The confidence-based dynamic quantitative assessment system aims to address the problems of "extensive quality assessment and lack of dynamic adjustment" in traditional data integration. This system converts data quality into a quantifiable "confidence" indicator (ranging from 0 to 1) and establishes a complete process that includes multi-dimensional assessment, closed-loop correction, and dynamic updates. This system enables precise control of multi-source data, including: 1. Multi-dimensional evaluation model: Data source reliability assessment: Initial weighting is assigned based on the authority and historical performance of the data source (e.g., 0.9 for government databases, 0.5 for social media), and dynamically adjusted through ongoing verification. For example, for every 10% decrease in the matching rate between the data provided by a recruitment platform and that of an authoritative institution, its weighting is automatically reduced by 0.1.

[0037] Data timeliness assessment: Design a time decay mechanism so that the older the data is, the lower its confidence level. For example, if the confidence level of data within three months is 1, it will decrease by 0.1 for every three months beyond that, and the confidence level of data older than one year will drop below 0.5.

[0038] Logical consistency assessment: Verify the rationality of data through the built-in rules of the knowledge graph. For example: Whether there is a conflict between working experience and academic qualifications (e.g., "obtaining a doctorate at the age of 22" may be problematic); Whether the skill description matches the project experience (e.g., claiming to be "proficient in big data" but having no relevant project records).

[0039] 2. Confidence calculation and grading: The evaluation results of the three dimensions are comprehensively calculated into the final confidence level, which is divided into four levels: High confidence (0.8-1): Data from a reliable source, with strong timeliness and logical consistency (such as government-issued academic certificates); Medium confidence (0.6-0.8): Data with minor flaws but can be supported by other data (such as user-filled skill tags, which need to be verified with project experience); Low confidence (0.4-0.6): Data with obvious problems or unreliable sources (such as anonymous social media reviews); Untrustworthy (<0.4): Data with logical contradictions or extremely unreliable sources (such as forged certificate scans).

[0040] 3. Closed-loop correction mechanism: Automatic correction: For low-confidence data, the system prioritizes corrections through the following methods: Missing value filling: based on the attribute distribution of similar entities in the knowledge graph (such as "average salary for the same position"); Conflicting value verification: calling high-confidence data sources (such as social security records) for comparison; Outlier smoothing: Identify and correct data that deviates significantly from the normal range through statistical methods.

[0041] Manual intervention: Data that still does not meet the standards after automatic correction or involves critical decision-making (such as core skills certification) will be pushed to the manual review node.

[0042] Reverse feedback: Correction results are used to reversely update the reliability weight of the data source. For example, if the resume information provided by a headhunting company was repeatedly verified to be exaggerated, its data source weight would be reduced from 0.7 to 0.5.

[0043] 4. Dynamic updates and continuous learning: Real-time data monitoring: Continuously track changes in the confidence of received data (such as a decrease in timeliness due to the passage of time).

[0044] Incremental learning: When new data is accessed or knowledge graph rules are updated, only the confidence of the affected data is re-evaluated rather than a full recalculation.

[0045] Adaptive adjustment: Dynamically optimize the confidence calculation model based on actual application feedback (such as the match between background check results and the candidate's actual performance).

[0046] Event-driven incremental updates of the knowledge graph aim to break through the limitations of the traditional "batch update" model. By capturing dynamic events related to talent (such as skill updates, project progress, and qualification changes) in real time, only the affected parts of the knowledge graph are updated in a targeted manner, rather than rebuilding the entire graph. This ensures data timeliness while significantly reducing computing resource consumption. This mechanism achieves the dynamic evolution of the knowledge graph through a complete process of "event perception - incremental processing - conflict resolution - freshness marking", specifically including: Structured Data Source Event Monitoring: For structured data sources like enterprise HR systems, social security databases, and professional certification platforms, we pre-set key event rules (e.g., "Project status changes to 'Completed'," "New skill certification record added," "Salary adjustment record generated"). When a data source triggers a rule, the engine captures the event in real time and extracts key information (including event type, associated entity ID, occurrence time, and data content). Unstructured data source event mining: Unstructured data such as social media (such as LinkedIn dynamics, GitHub submission records), industry forums, and conference videos are mined for hidden events using natural language processing and keyword recognition technology. For example: Identify the "Complete development of a core module for an open source project" event from GitHub code commit records; Extract the "publishing new technology patent" event from the transcript of the industry conference speech video; Capture the "mastery of an emerging technology (such as AIGC tools)" event from a candidate's blog post.

[0047] Event filtering and prioritization: Filter captured events to exclude invalid information (such as duplicate submissions and incorrect annotations), and sort them according to event importance (such as "winning a top industry award" takes higher priority than "updating personal profile picture"), giving priority to high-priority events.

[0048] Entity matching and positioning: Based on the associated entity ID in the event (such as the candidate's unique identifier and project number), the corresponding entity node (such as "Candidate A" and "Project B") is accurately located in the knowledge graph to avoid traversing the entire graph. Local attribute and relationship updates: Only modify attributes and relationships related to the event, without touching irrelevant data. For example: When the "Candidate Obtains PMP Certification" event occurs, only the "Project Management Qualification" attribute of the candidate entity and the relationship weight with "Project Management Capability" are updated. There is no need to update unchanged attributes such as education background and years of work experience. When the "a project is completed early" event occurs, only the "status" and "cycle" attributes of the project entity and the strength of the association with the "project experience" of the participating candidates are adjusted. Creation of new entities and relationships: If the event involves new knowledge (e.g., a candidate participated in an unrecorded overseas project), a new entity ("Overseas Project C") is created in the knowledge graph, and a "Participation" relationship is established with the candidate, while project-related attributes (e.g., project scope, budget size) are associated.

[0049] Data source priority rules: When new event data conflicts with existing data in the knowledge graph (for example, the "time of leaving" in the social security record is May 2023, while the certificate provided by the former employer is March 2023), the acceptance standard is determined based on the data source credibility priority (such as government agency data > official enterprise data > personal self-reported data), and the conflict record and reason annotation are retained in the graph. Timestamp verification: For repeated events from the same data source (such as inconsistent salary data submitted twice), the event data with the latest timestamp will be used as the basis, and historical change traces will be automatically recorded (such as "Salary adjusted to X in October 2023, and updated to Y in March 2024"). Manual intervention trigger: When the conflict involves core evaluation indicators (such as the validity of key skills certification) and the data source priority is the same (such as two conflicting letters of recommendation from different companies), the manual review process is automatically triggered. The review result serves as the final update basis and is synchronized to the conflict record.

[0050] Freshness tag: After each update, a "last updated timestamp" is added to the relevant entities and relationships, and a "freshness tag" is generated based on the difference between the timestamp and the current time (such as "updated within 1 week", "updated 1-3 months", "not updated for more than 3 months"). Dynamic weight adjustment: In subsequent talent evaluations, the system automatically adjusts the data weight based on the freshness tag. For example: Project results data with a freshness rating of "updated within 1 week" will have a 20% higher weight in the "Project Execution Ability" assessment; The weight of skill descriptions with a freshness of "not updated for more than 3 months" is reduced by 50%, and a prompt is provided to supplement the latest verification information. Expired data processing: Low-importance data (such as early internship experience) that has not been updated for more than one year and has no associated new events will be automatically marked as "archived status" and will not be included in the evaluation scope by default, but query permissions will be retained.

[0051] The above is a detailed description of an embodiment of the present invention. However, the content is only a preferred embodiment of the present invention and should not be considered to limit the scope of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the present invention.

Claims

1. A talent background investigation method based on multi-source data evaluation, characterized in that: The following steps are involved: Construct a dynamic talent knowledge graph, which includes an entity layer, an attribute layer, and a relationship layer; the entity layer defines core entities and assigns unique identifiers; The attribute layer defines standardized attributes based on the industry's general capability model and supports custom extensions; The relation layer mines implicit semantic associations between entities and quantifies relation weights through graph neural networks; Perform semantic alignment on multi-source cross-modal data, convert text data into semantic vectors through pre-trained language models and match them with knowledge graph attribute vectors, map image data to knowledge graph attributes after OCR recognition and verification, and normalize numerical data as quantitative indicators of entity attributes; Data quality is assessed based on a multi-dimensional confidence model, including data source reliability, data timeliness, and logical consistency. Data source reliability is dynamically weighted based on historical verification pass rates, data timeliness is calculated using a decay function, and logical consistency is verified using a knowledge graph rule engine. Build a confidence closed-loop correction mechanism to intelligently correct low-confidence data, and use the correction results to reversely update the data source credibility weight; It adopts an event-driven incremental update architecture, captures dynamic events of multi-source data in real time through a distributed event perception engine, synchronizes only the changed data and performs incremental updates to the knowledge graph, while also performing conflict resolution and freshness marking.

2. The talent background investigation method based on multi-source data evaluation according to claim 1 is characterized in that: Said core entities include candidates, qualifications, projects and skills; The step of mining implicit semantic associations between entities at the relational layer includes: When the overlap between the keywords of the candidate's published papers and the core technical requirements of the recruitment position reaches the preset threshold, a professional field matching relationship label is automatically generated, and the relationship weight is dynamically adjusted according to the overlap.

3. The talent background investigation method based on multi-source data evaluation according to claim 1 is characterized in that: The initial weight distribution rule for the data source reliability is: The initial weight of the government database is not less than 0.9; The initial weight of social media is no higher than 0.6, and its weight is reduced by 0.1-0.3 for every 10% decrease in the historical verification pass rate of the data source.

4. The talent background investigation method based on multi-source data evaluation according to claim 1 is characterized in that: The decay function of the data timeliness satisfies: when the data generation time exceeds 3 months, the confidence is multiplied by a decay coefficient of 0.8-0.

95.

5. The talent background investigation method based on multi-source data evaluation according to claim 1 is characterized in that: Data contradictions are detected through the knowledge graph rule engine. When a preset contradiction type exists, the data confidence level drops directly to below 0.

3.

6. The talent background investigation method based on multi-source data evaluation according to claim 1 is characterized in that: The knowledge graph incremental update step includes: New data is located in the knowledge graph related nodes through the entity matching algorithm, and only relevant attributes and relationships are updated. When data contradictions occur, high-confidence data sources are prioritized and the conflict details are marked.

7. The talent background investigation method based on multi-source data evaluation according to claim 1 is characterized in that: The freshness mark is calculated based on the difference between the data update time and the current time. Historical data with a freshness lower than 0.3 will have its weight reduced by more than 50% in subsequent evaluations.

8. The talent background investigation method based on multi-source data evaluation according to claim 1 is characterized in that: The step of dynamically adjusting the weight of the confidence model includes: Low confidence data automatically triggers corrections: Missing values ​​are filled based on the attribute distribution of similar entities in the knowledge graph; contradictory values ​​are verified and corrected by calling high-confidence data sources; The correction result is fed back to the data source weight.

Images