Industrial demand-oriented dynamically adaptive education decision-making auxiliary system

Abstract

The invention provides an industry demand-oriented dynamically adaptive education decision-making auxiliary system, and relates to the technical field of education informationization, the system comprises a data acquisition layer, a data processing layer and an event detection layer, the data acquisition layer is used for periodically and synchronously acquiring original data from a local policy database and a recruitment market database; the data processing layer adopts a locally deployed large model to perform entity recognition, relationship extraction and structured processing on the collected original data based on a preset educational knowledge graph to obtain structured semantic information; through a multi-source data real-time sensing mechanism and an RAG-fine tuning collaborative reasoning framework, dynamic capture of industrial demands is realized, when the ability popularity index suddenly increases, the system can complete event judgment, trend analysis and adjustment suggestion generation in a short time, the response period is short, the problem of education supply lag in the emerging technical field is effectively solved, and the system has a wide application prospect. And employment competitiveness of students is guaranteed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of educational informatization, and in particular to an education decision-making assistance system oriented to dynamic adaptation of industrial demand. BACKGROUND

[0002] The technical means for supporting professional settings and curriculum adjustment in the education field mainly include three types of known technologies: one is an education data statistics system, which is a basic module of an education management platform, can collect data such as university teaching resources, curriculum settings, teacher structure, and graduation employment, update and output reports according to the academic year or semester, and is used for teaching evaluation and management decision-making. However, this system relies on historical data and lacks real-time sensing capability, and the analysis dimension only stays at the professional level, which cannot realize structured diagnosis at the curriculum-ability level and cannot respond to changes in the industrial end. The second is a post demand analysis tool, which extracts and counts hot abilities and post skills rankings through keyword extraction based on recruitment platform data, and can preliminarily reflect the industry skill preference. However, the technology is limited to shallow statistics and has not established a structured association with specific courses in the education system, so it cannot provide a mapping path of ability-course-major. The third is a special post matching model, which uses a collaborative filtering algorithm to recommend posts or courses based on historical data, and has a personalized recommendation effect. However, the reasoning logic relies on "user / item similarity" and does not mine the causal relationship between post ability and curriculum system, so it cannot dynamically perceive the demand for structural adjustment caused by emerging abilities or policies.

[0003] The existing technology has the following defects in supporting education decision-making, which leads to the lag of university professional training and industrial demand adaptation: on the one hand, the lag is serious, and the annual graduate employment data is used as the basis for decision-making. When emerging technologies emerge, the response period lags at least one training period, affecting the students' employment competitiveness. On the other hand, the granularity is coarse, and the analysis stays at the macro level of majors such as "computer" and "automation", which cannot identify structural defects in course content, such as the problem that the "machine learning" course does not cover the Transformer architecture. At the same time, the static decision-making mechanism relies on periodic teaching reform evaluation and has no connection with real-time external data sources, so it cannot quickly adjust when new professional demand increases. In addition, the regional imbalance problem is prominent, and the professional settings and enrollment planning use a unified template without considering regional industrial differences. Therefore, the present application proposes an education decision-making assistance system oriented to dynamic adaptation of industrial demand to solve the problems existing in the prior art. SUMMARY

[0004] In view of the above problems, the application provides an education decision-making auxiliary system for dynamic adaptation to industrial needs, which realizes dynamic capture of industrial needs through a multi-source data real-time sensing mechanism and a RAG-fine-tuning collaborative reasoning framework.

[0005] To achieve the object of the application, the application realizes the following technical scheme: an education decision-making auxiliary system for dynamic adaptation to industrial needs, comprising a data acquisition layer, a data processing layer and an event detection layer, the data acquisition layer is used for regularly synchronously acquiring original data from a local policy database and a recruitment market database; the data processing layer adopts a locally deployed large model, performs entity recognition, relation extraction and structured processing on the acquired original data based on a preset education knowledge graph to obtain structured semantic information, wherein the education knowledge graph is a semantic network for structurally representing the relationships among courses, capabilities and posts, and the data processing layer integrates a collaborative reasoning framework of RAG retrieval and domain fine-tuning large model.

[0006] The event detection layer detects education industry events based on the structured semantic information through a preset skill demand surge detection algorithm and an education policy analysis algorithm to generate a trend analysis report and a risk assessment report, wherein the skill demand surge detection algorithm calculates a capability heat index based on recruitment data, and the education policy analysis algorithm searches for policies and analyzes the influence of the policies through a large model reasoning cycle.

[0007] Further improvement lies in that the original data of the data acquisition layer comprises regional education policies, industrial planning files in the local policy database, and post demand information and post capability requirement information in the recruitment market database; and the data acquisition layer synchronizes data in a monthly / quarterly cycle to ensure data timeliness.

[0008] Further improvement lies in that the entity recognition and relation extraction of the data processing layer comprise recognizing professional, course, capability, post and policy entities, extracting course-capability, capability-post and policy-professional relationships, and supporting automatic deduplication and merging of multiple types of entities and multiple relationships.

[0009] Further improvement lies in that in the data processing layer, the execution process of the collaborative reasoning framework of RAG retrieval and domain fine-tuning large model is as follows:

[0010] The domain fine-tuning large model analyzes user problems, extracts key entities and knowledge gaps, and outputs retrieval demand signals.

[0011] The retrieval module performs directional retrieval in the educational knowledge graph based on the retrieval demand signal, and generates a structured knowledge package;

[0012] The domain fine-tuned large model fuses the structured knowledge package and internal reasoning capability, verifies logical consistency, and supplements reasoning details;

[0013] When there is a knowledge gap, a supplementary retrieval signal is generated to trigger secondary retrieval until there is no missing key information;

[0014] A conclusion with a credibility label is generated, and the retrieval knowledge trace and internal reasoning verification part are labeled.

[0015] Further improvement lies in that, in the event detection layer, the quantitative standard of the skill demand surge detection algorithm is that, compared with the current month / quarter and the previous period or the current period and the same period of the previous year, the increase of 20% is the first grade, the increase of 50% is the second grade, and the increase of more than 100% is the key attention grade; or when a new ability appears in the top 50 of the ability heat index, it is determined that there is a skill demand surge event.

[0016] Further improvement lies in that, in the event detection layer, the cycle search logic of the educational policy analysis algorithm is that, based on the professional name and the training program, the initial policy is searched through fuzzy keyword matching to extract the core target, the related industry, the talent training expression and the preliminary question point; the supporting policies and local implementation rules are searched according to the preliminary question point, the policy information database is updated and the question point is corrected until there is no missing key information; based on the complete policy information, the influence on the industry and the profession is analyzed, and the number structure of the profession and the adjustment suggestion of the training program are generated.

[0017] Further improvement lies in that it further comprises an education resource constraint optimization model, the education resource constraint optimization model comprising a constraint modeling module, a course importance quantification module, a multi-constraint solving module and a dynamic iteration module; the constraint modeling module sets hard constraints and soft constraints, the hard constraints comprising total learning hours ≤ upper limit of training program, core course learning hours ≥ 64 hours, general education course credit ratio ≤ 20%, and practice course learning hour ratio ≥ 25%; the soft constraints comprising course-skill matching degree ≥ 80% and core course priority weight being 1.5 times of non-core course; the course importance quantification module is used to calculate course importance score; the multi-constraint solving module takes Max (Σ course importance score S × learning hours) as the objective function, and adjusts course credit / learning hour through linear programming algorithm; the dynamic iteration module recalculates the course importance score and optimizes the credit / learning hour allocation when the skill demand increase is ≥ 20% or the policy is updated.

[0018] Further improvement lies in that, the calculation formula of the course importance quantification module is:

[0019] S = (skill demand intensity x matching degree) x course attribute coefficient + policy adaptation score,

[0020] Among them, the skill demand intensity is the increase in the frequency of occurrence of the corresponding skills in industry recruitment, the matching degree is the degree of overlap between course content and target skills, the course attribute coefficient of core courses is 1.2, non-core courses are 0.8, and courses that conform to policy guidance are given an additional 0.2 points.

[0021] Further improvements include a regional adaptive module, which adopts a data layer-feature layer-algorithm layer-output layer architecture. The data layer stores regional industrial geographic distribution data and employment feedback data. The regional industrial geographic distribution data includes the GDP share of leading industries, industry growth rate, and policy orientation. The employment feedback data includes local employment rate, professional matching rate, employment gap rate, salary premium, and employment stability. The feature layer calculates regional demand intensity characteristics F1, employment absorption capacity characteristics F2, and regional carrying capacity constraint characteristics F3. The algorithm layer calculates the theoretical demand scale Stheoretical based on F1 and F2, and corrects it with F3 to obtain the actual enrollment scale Sreal. The output layer outputs regional and professional suggestions for expanding / reducing enrollment, demand heatmaps, and sensitivity analysis.

[0022] A further improvement lies in the following: the formula for calculating the regional demand intensity characteristic F1 is:

[0023] F1 = Industrial linkage degree × Industrial growth coefficient × Policy weight;

[0024] The formula for calculating the employment absorption capacity characteristic F2 is as follows:

[0025] F2 = (Local employment rate × Job matching rate) × (1 + Employment gap rate) × Salary premium coefficient.

[0026] The beneficial effects of this invention are as follows:

[0027] 1. This invention achieves dynamic capture of industry demand through a multi-source data real-time perception mechanism and the RAG-fine-tuning collaborative reasoning framework. When the capability heat index suddenly increases, the system can complete event judgment, trend analysis and adjustment suggestion generation in a short time. The response cycle is short, which effectively solves the problem of lagging education supply in emerging technology fields and ensures students' employment competitiveness.

[0028] 2. Based on the curriculum-ability-job ternary relationship of the educational knowledge graph and the RAG semantic reasoning ability, the system can accurately map job skill requirements to the university curriculum system, identify the content deficiencies of specific courses, break through the limitations of traditional technical professional-level macro analysis, realize precise diagnosis at the course level, provide direct basis for curriculum syllabus revision, and promote the adaptation of education from professional alignment with industry to curriculum alignment with job.

[0029] 3. The regional adaptive module of this invention outputs differentiated enrollment suggestions by quantifying regional industry characteristics and employment dynamics, thus solving the regional imbalance problem caused by the uniform template. The education resource constraint optimization model quantifies the importance of courses based on skill requirements, enabling limited resources to be concentrated on high-value courses. At the same time, the local deployment mode ensures data security, and the trend and risk assessment module provides a scientific basis for decision-making, comprehensively improving the efficiency of education resource allocation and the adaptability to regional industries. Attached Figure Description

[0030] Figure 1 This is a block diagram illustrating the overall system architecture of the present invention.

[0031] Figure 2 This is a schematic diagram of the business process of the present invention;

[0032] Figure 3 This is a timing diagram of the key modules of the present invention;

[0033] Figure 4 This is a loop logic diagram for policy text analysis in this invention;

[0034] Figure 5 This is the flowchart of the RAG-fine-tuning collaborative reasoning of the present invention. Detailed Implementation

[0035] To enhance understanding of the present invention, the present invention will be further described in detail below with reference to embodiments. These embodiments are only used to explain the present invention and do not constitute a limitation on the scope of protection of the present invention.

[0036] Example 1

[0037] according to Figure 1 , 2 As shown in Figures 3, 4, and 5, this embodiment proposes an education decision support system that dynamically adapts to industry needs. It adopts a three-layer architecture, including a data acquisition layer, a data processing layer, and an event detection layer. The data acquisition layer is responsible for collecting raw data from local policy databases and recruitment market databases. The data processing layer performs entity recognition, relation extraction, and structured processing on the collected data, and uses a locally deployed DeepSeek large-scale model for intelligent analysis. The event detection layer detects events in the education industry based on structured data and generates trend analysis and risk assessment reports. For example... Figure 1 .

[0038] After a user enters their major, the system automatically retrieves relevant policy texts and recruitment data from the local database. The data processing layer then performs entity recognition and relationship extraction to construct structured semantic information. Subsequently, the event detection layer analyzes the extracted data to identify potential industry events and trend changes, ultimately outputting structured JSON content containing event identification results, trend judgments, and data insights for easy integration and automation. Figure 2 .

[0039] When a user initiates an analysis request, the system's data acquisition module acquires relevant policy and recruitment data in parallel. Subsequently, the data processing module performs entity recognition and relationship extraction operations to extract key semantic information. Next, the event detection module identifies industry events based on the processing results and generates trend analysis. Finally, the system returns structured analysis results containing events, trends, and data insights. Figure 3 .

[0040] Data Acquisition Layer: Regularly synchronizes policy and recruitment data via a local database interface to ensure data timeliness and completeness. Data Processing Layer: Employs a locally deployed DeepSeek large-scale model, relying on the knowledge graph built within the educational institution, to perform entity recognition and relation extraction, supporting multiple entity types and various relationships, and automatically deduplicating and merging data. Event Detection Layer: Based on rules and large-scale model reasoning, detects events such as sudden increases in skill demand, policy releases, and technological breakthroughs, generating trend summaries and risk assessments.

[0041] Based on recruitment market data, a knowledge graph is used to calculate a skill demand index for job postings. This index is summarized monthly or quarterly. When the summarized data shows an increase in the demand for existing skills compared to previous data (20% as one level, 50% as another, and over 100% requires close attention), or when new skills appear in the top 50 of the skill demand index, it is defined as a "surge in skill demand" event. Educational institutions crawl a large number of policies and store them in a policy database for analysis. During analysis, the reasoning capabilities of a large model are utilized. The model first searches for policies using fuzzy keyword matching based on the major's name and curriculum. Then, based on the search results, it further considers and generates sub-questions. Based on these sub-questions and existing information, it automatically and iteratively searches the policy database to analyze the potential impact of policies on the major. The analysis results include not only how the curriculum of the major should be adjusted (e.g., a university significantly increases AI+ courses in the humanities due to the AI ​​boom), but also adjustments to the number of majors (e.g., a university responds to policy by establishing a new "Low-Altitude Technology and Engineering" major). All analysis results are output in JSON format for easy subsequent system integration and automated processing.

[0042] The core objective of the policy text analysis (including loop search logic) prompt is to deepen the understanding of the policy by iteratively retrieving the policy text and related information, ultimately outputting suggestions for adjusting the professional quantity structure and training program. The specific process steps and output requirements are as follows:

[0043] Initial policy analysis and information extraction: Based on the analyzed major name, conduct an initial policy search and extract the core policy objectives, relevant industry sectors, and talent cultivation-related statements.

[0044] Policy loop search based on points of doubt: Based on the "initial points of doubt" in the initial analysis, retrieve relevant supporting policies, local implementation rules, policy interpretation documents, etc.

[0045] After each search, the policy information database is updated, and points of doubt are added or corrected until no key information is missing.

[0046] Policy Impact Analysis (Based on Complete Policy Information): Combining all retrieved policy texts, analyze the short-term (1-2 years) and medium-term (3-5 years) impacts on related industries: changes in industry size (expansion / contraction / transformation); changes in job demand (new job types, increases or decreases in the number of jobs); and updates in skill requirements (new technologies and skills to be mastered).

[0047] Analysis of the transmission impact of policies on majors: Based on industry impact, this study analyzes the direct and indirect impacts on related majors: changes in the matching degree between majors and industries, changes in the fit between core professional courses and job skills, and changes in the employment competitiveness of professional graduates.

[0048] Recommendations for adjusting the quantity and structure of professional programs: Based on the impact of policies on industries and majors, propose recommendations for adding / removing majors and expanding / reducing enrollment: Recommended majors to be added and the reasons (e.g., matching the needs of emerging industries), recommended majors to be removed and the reasons (e.g., shrinking industry demand), and recommended majors to expand / reduce enrollment and the extent of such expansion (based on job gap / surplus data).

[0049] Recommendations for adjusting the training program: Based on changes in skill requirements, suggestions for adjusting the curriculum system and practical components are proposed: courses that need to be added (such as those incorporating new technologies and content required by policies), courses that need to be deleted or optimized (such as content that is out of touch with industry needs), and adjustments to practical teaching (such as increasing the proportion of corporate internships and project training).

[0050] Each search builds upon the questions raised in the previous analysis, progressively improving the coverage of policy information; when the "Updated Questions" section is empty, the analysis output phase begins; all analyses must include policy references to ensure the recommendations are traceable. Figure 4 .

[0051] Multi-source real-time sensing mechanism

[0052] The core of the collaborative working mechanism between RAG retrieval and domain-specific fine-tuning of the large model is the two-way cooperation between "external knowledge retrieval" and "domain-specific deep reasoning." This not only compensates for the large model's lack of real-time / domain-specific knowledge but also leverages the domain adaptability of the fine-tuned model to improve the accuracy and logic of reasoning (especially in causal reasoning scenarios in the education field). Its main algorithm is as follows:

[0053] Skill Demand Sudden Change Detection Algorithm: Based on recruitment data, this algorithm uses knowledge graph analysis to identify the correlation between job positions and required skills, thereby calculating a skill popularity index. It can be used to compare changes in demand over previous months / quarters, or to analyze changes in skill demand based on the spring / autumn recruitment cycle and the skill popularity index for the same period this year / last year. If the required skill level for a job position increases by 100%, it can be considered a significant event, requiring educational institutions to respond and adjust as quickly as possible.

[0054] Educational policy analysis algorithm: Based on a large model, it automatically analyzes the search and majors for policies that may be related to the majors by using the reasoning ability of the large model. Based on the search information, it continues to break down sub-problems and iteratively search for new policies that may affect the majors until the upper limit of the rounds is reached or no more new effective information is introduced.

[0055] The RAG-fine-tuned collaborative reasoning framework's step-by-step execution process and interaction logic can be refined as follows: Taking an education-related problem (e.g., "The impact of expanding enrollment in a certain major on industry talent supply") as an example, the collaborative mechanism consists of five core steps: The domain fine-tuning model first receives the user's problem and analyzes its core through pre-trained domain adaptation capabilities (e.g., education policies, industry supply and demand logic): clarifying the entities involved (e.g., "a certain major," "industry talent supply"), relationships (e.g., "expansion → supply change"), and reasoning requirements (e.g., causal relationship analysis). Simultaneously, it outputs "retrieval demand signals": extracting key entities and knowledge gaps to be verified in the problem (e.g., current enrollment scale of the major, industry talent gap data, relevant policy constraints, etc.) to guide the retrieval module in targeted queries.

[0056] Educational Knowledge Graph Retrieval and Knowledge Filtering: The retrieval enhancement module, based on the "retrieval demand signal" output by the fine-tuning model, performs targeted retrieval of relevant entities and relationships within the educational knowledge graph. At the entity level, it extracts the curriculum system, historical enrollment data, and corresponding industry job demand data for the major. At the relationship level, it retrieves predefined relationships such as "major size - industry supply" and "policy - enrollment restrictions." The retrieval module performs preliminary filtering on the returned results (e.g., removing outdated data and low-relevance information) to form a "structured knowledge package" (containing entity attributes, relationship triples, key data, etc.).

[0057] Knowledge fusion and reasoning initiation: The fine-tuning model receives the "structured knowledge package" returned by the retrieval module and combines it with its learned domain causal reasoning ability (such as the logical chain of "expanded enrollment → increased number of graduates → short-term oversupply → long-term need to match industry growth rate") to integrate external and internal knowledge: verify the consistency between the internal reasoning logic and the retrieved knowledge (such as whether the model's preset "industry talent gap" matches the real-time data in the knowledge graph); supplement reasoning details (such as combining the "ability matching degree between this major and industry positions" in the knowledge graph to correct the "actual impact weight of expanded enrollment on supply").

[0058] Dynamic Interaction: Reasoning Completion and Secondary Retrieval - If the fine-tuning model discovers a knowledge gap during reasoning (such as "lack of growth forecast data for the industry in the next 3 years"), it will generate a "supplementary retrieval signal" to trigger the retrieval module in reverse: the retrieval module supplements the missing information from the knowledge graph or related data sources (such as industry report databases) based on new needs; the secondary retrieval results are then input into the fine-tuning model to improve the reasoning chain.

[0059] Conclusion Generation and Credibility Verification: The model is fine-tuned to integrate all knowledge and reasoning processes, generating the final answer (e.g., "Expansion of enrollment in this major should be considered in light of the industry's annual growth rate constraint of ≤5%, and it is recommended that the average annual expansion not exceed 8%, otherwise it will lead to a talent surplus for 1-2 years"). A "credibility label" is output simultaneously: this labels which parts of the conclusion rely on retrieved knowledge (traceable) and which rely on internal model reasoning (requiring verification with domain common sense), improving the interpretability of the results. A detailed flowchart is shown below. Figure 5 .

[0060] Core interaction logic:

[0061] The collaboration between the two is not a one-way chain, but a closed-loop interaction where "retrieval serves reasoning, and reasoning guides retrieval": Positive support: The retrieval module provides the fine-tuning model with "real-time, specific knowledge" (such as the latest professional-job matching data in the education knowledge graph), preventing the model from relying on outdated or generalized internal knowledge that could lead to reasoning bias. Reverse driving force: The fine-tuning model, through its domain-specific causal reasoning capabilities, identifies the "relevance boundaries" and "gaps" of the retrieved knowledge (such as determining whether a policy document is directly related to the current issue or whether regional industry difference data needs to be supplemented), dynamically optimizing the direction and granularity of the retrieval. Target alignment: Both work together to provide "accurate answers to questions in the education field." Retrieval ensures the "substance" (accuracy) of the knowledge, while the fine-tuning model ensures the "depth" (domain logic adaptation) of the reasoning, ultimately achieving a "1+1>2" effect (better understanding domain reasoning than pure RAG and closer to real-time knowledge than pure fine-tuning models). This collaborative mechanism makes "retrieval" an "external memory bank" for the fine-tuning model and the "fine-tuning model" an "intelligent processor" for the retrieved knowledge, making it particularly efficient in scenarios such as education that require strong domain logic and real-time data support.

[0062] Educational resource-constrained optimization model

[0063] The core objective of the educational resource constraint optimization model (a multi-constraint solver focusing on course categories / class hours / credits) is to achieve a precise match between curriculum design and skill requirements, and professional training objectives, under the constraints of limited educational resources (total class hour limit, teacher capacity, venue capacity, etc.), by quantifying course importance and dynamically allocating credits / class hours (especially prioritizing resource allocation for core professional courses). Its working mechanism can be divided into four steps: "constraint modeling → importance quantification → multi-objective solution → dynamic iteration," as detailed below:

[0064] Constraint modeling: Define the "unbreakable boundaries" by first identifying the hard and soft constraints affecting course credit / hour allocation, and then define the optimization range for the solver.

[0065] Hard Constraints (Rigid Conditions That Must Be Met) - Total Resource Constraints: Total professional class hours ≤ the upper limit stipulated in the training program (e.g., total undergraduate class hours ≤ 2600 class hours); total credits ≤ graduation requirements (e.g., undergraduate ≥ 160 credits). Course Attribute Constraints: Core Professional Courses: Class hours ≥ minimum standard (e.g., single core course class hours ≥ 64 class hours), credits ≥ 3 credits (ensuring depth of core knowledge); General Education Courses: Credit percentage ≤ 20% (avoiding crowding out professional course resources); Practical Courses: Class hour percentage ≥ 25% (meeting the requirements for application-oriented talent training). Logical Constraints: Prerequisite courses must be completed before subsequent courses (e.g., "Advanced Mathematics" must be offered before "Engineering Mechanics," affecting the timing of class hour allocation).

[0066] Soft constraints (flexible conditions that can be dynamically adjusted) - Skill matching constraints: the degree of matching between course content and popular industry skills (e.g., the matching degree between "Python programming" course and "data analysis ability" must be ≥80%); Priority constraints: the resource allocation priority of core courses > non-core courses (e.g., the weight of adjusting the class hours of core courses is 1.5 times that of ordinary courses).

[0067] Course Importance Measurement: Linking "Skill Demand" with "Course Value"; Quantifying "Changes in Skill Demand" into "Course Resource Allocation Weights" through quantitative indicators to ensure resources are tilted towards high-demand, high-value courses: Core Indicator: Course Importance Score (S) Calculation Formula: S = (Skill Demand Intensity × Matching Degree) × Course Attribute Coefficient + Policy Adaptability Score Skill Demand Intensity: The frequency of occurrence of the corresponding skills in industry recruitment (e.g., if the demand for "Machine Learning" skills increases by 30%, the intensity value of the corresponding course will increase by 30%); Matching Degree: The degree of overlap between course content and target skills (core courses, due to their more focused content, typically have a matching degree ≥ 0.8, while non-core courses may have a matching degree ≤ 0.5); Course Attribute Coefficient: Core courses are set at 1.2, and non-core courses are set at 0.8 (strengthening the priority of core courses); Policy Adaptability Score: If the course conforms to policy guidance such as "New Engineering" or "AI+", an additional 0.2 points are added.

[0068] When the skill demand for "Artificial Intelligence Applications" increases by 50%, the importance score (S) of its corresponding core course "Introduction to AI" will increase accordingly (due to the increased intensity of skill demand), and thus receive higher priority in resource allocation.

[0069] Multi-constraint solution: Achieve "optimal resource allocation" through algorithms - Based on the above constraints and important quantification results, use linear programming or heuristic algorithms (such as genetic algorithms) to solve for the optimal solution and output the credit / hour allocation scheme for each course: Objective function: Maximize total value: Max(Σ course importance score S × hours), that is, give more hours to high-importance courses to ensure that resource input matches value output. Solution Process: Input: Course list (including core / non-core identifiers), current credits / hours for each course, skill requirement data, and constraints; Algorithm Iteration: Under the premise of satisfying hard constraints, gradually adjust the credits / hours of each course: For courses with S≥0.9 (high importance, mostly core courses): prioritize increasing credits (e.g., from 64 credits to 72 credits) and increasing credits (e.g., from 3 credits to 4 credits); For courses with S≤0.5 (low importance, mostly non-core courses): reduce credits (e.g., from 48 credits to 32 credits) and decrease credits (e.g., from 2 credits to 1 credit), freeing up resources for high-demand courses; Output: Optimized course credit / hour allocation table (including the specific values ​​for each course and the verification results of whether all constraints are satisfied).

[0070] Dynamic Iteration: Real-time Adjustment Based on Skill Requirements and Policy Changes - The model is not a static output, but rather achieves dynamic adaptation through a closed loop of "monitoring → recalculation → optimization": When industry skill requirements change (e.g., a skill requirement increases by ≥20%) or policies are updated (e.g., a new "digital literacy" course requirement is added), the model is triggered to recalculate the course importance score S; the solver, based on the new S value, reallocates credits / hours within the original constraint framework (the adjustment range for core courses must meet their minimum standards); a comparison report before and after the adjustment is output to assist teaching administrators in decision-making.

[0071] This model, through the logic of "constraint delimitation, importance quantification, and algorithm solution," avoids the subjective arbitrariness of resource allocation and can dynamically respond to changes in skill demand. When the demand for a certain skill increases, the corresponding core courses will gain more class hours / credits through increased importance, while non-core courses will "make way" through resource compression, ultimately achieving the optimization goal of "concentrating limited resources on high-value courses."

[0072] Region Adaptive Module

[0073] This module is an enrollment scale optimization algorithm based on industrial geographic distribution and employment data. The core objective of the regional adaptive module (enrollment scale optimization algorithm based on industrial geographic distribution and employment data) is to break away from the "one-size-fits-all" enrollment scale model. By quantifying regional industrial characteristics and employment dynamics, it provides differentiated optimization solutions for professional enrollment scales in different geographical units (such as provinces, cities, and districts / counties), ultimately achieving a dynamic balance between "regional industrial demand, talent training scale, and employment absorption capacity". Its core logic can be broken down into a four-layer architecture: "data layer - feature layer - algorithm layer - output layer," which can be further expanded as follows: Data layer: Constructing a regional industry-employment dynamic database. The accuracy of the algorithm depends on multi-dimensional, high-granularity regional data collection, which needs to cover the two core dimensions of "industry geographic distribution" and "employment feedback," and be structured and stored by region (e.g., using prefecture-level cities as the smallest unit) and major (e.g., classified according to the "Directory of Undergraduate Majors in General Higher Education Institutions"). Key indicators of industry geographic distribution data (reflecting "what kind of talent the region needs"): Regional leading industries: The top 3 industries in a region (e.g., electronic information in Shenzhen, digital economy in Hangzhou) and their GDP share (industry concentration); Industry scale: Number of enterprises (by scale: large-scale enterprises / SMEs), links in the industrial chain (e.g., Shanghai's chip industry focuses on the design link, Hefei focuses on the manufacturing link); Industry growth: Revenue growth rate and fixed asset investment growth rate in the past 3 years (predicting future talent demand potential); Policy guidance: Regional industrial planning (e.g., "Beijing-Tianjin-Hebei New Energy Vehicle Industry Belt," "Chengdu-Chongqing Electronic Information Industry Cluster"), key support areas.

[0074] Key indicators of employment feedback data (reflecting "how many professionals a region can absorb") include: Local employment rate: the proportion of graduates of a particular major employed in the region (e.g., the proportion of computer science graduates from Wuhan employed in Hubei); Professional relevance rate: the degree to which graduates' jobs match their majors (e.g., the proportion of clinical medicine graduates employed in regional medical institutions); Employment gap rate: the difference between the annual demand for a particular industry in the region and the supply of local graduates in that major (positive numbers indicate a gap, negative numbers indicate a surplus); Salary premium: the ratio of the average salary of graduates of a particular major in the region to the average salary in the region (the higher the premium, the more urgent the region's demand for professionals in that major); Employment stability: the average length of time graduates remain in their first job in the region (reflecting the industry's long-term demand for talent).

[0075] Feature layer: Extracting regions - professional matching features

[0076] The raw data is transformed into quantifiable "region-major fit characteristics," which serve as the core variables for algorithm input, focusing on characterizing "regional demand intensity for a particular major" and "employment absorption capacity." The formula for calculating the regional demand intensity characteristic (F1) is: F1 = Industry correlation × Industry growth coefficient × Policy weight; Industry correlation: The degree of matching between a major and the region's leading industries (e.g., the correlation between "Materials Science and Engineering" and Tangshan's steel industry = 0.9, and the correlation with Sanya's tourism industry = 0.2); Industry growth coefficient: The compound annual growth rate of the industry in the region over the past 3 years (e.g., the growth rate of the artificial intelligence industry in a region is 30%, coefficient = 1.3; the growth rate of the traditional textile industry is 5%, coefficient = 1.05); Policy weight: If the major conforms to the regional industrial plan, the weight = 1.2 (otherwise = 1.0).

[0077] The formula for calculating the employment absorption capacity characteristic (F2) is: F2 = (local employment rate × job matching rate) × (1 + employment gap rate) × salary premium coefficient; Local employment rate × job matching rate: reflects the actual absorption quality of talents in this profession in the region (e.g., if the local employment rate of a certain profession is 80% and the job matching rate is 70%, then the base value = 0.56); Employment gap rate: if the talent gap rate of this industry in the region is 20%, then this item = 1.2 (amplifying demand); if there is a surplus of 10%, then this item = 0.9 (suppressing demand); Salary premium coefficient: if the salary is 10% higher than the regional average, the coefficient = 1.1; if it is lower than 10%, the coefficient = 0.9 (reflecting the scarcity of talents).

[0078] Regional carrying capacity constraints (F3) reflect the “ceiling” of regional educational resources, such as: the existing faculty capacity of universities (how many students can be enrolled in a certain major, which is limited by the number of teachers and the size of laboratories); regional educational equity constraints (such as the need to maintain a minimum enrollment scale for basic majors in remote areas to avoid “zero enrollment”).

[0079] Algorithm Layer: Dynamic Enrollment Scale Optimization Model - Based on the above characteristics, the algorithm calculates the optimal enrollment scale adjustment range for each region and major through a dynamic balance formula of "demand-supply". The core logic is "determining demand based on employment absorption capacity, growth based on industry potential, and upper limit based on resource constraints": Basic enrollment scale calculation; Taking the current enrollment scale of a certain major in a region as the benchmark (S0), the theoretical demand scale (Stheoretical) is calculated by combining F1 (demand intensity) and F2 (employment absorption capacity): Stheoretical = S0 × (F1 × 0.4 + F2 × 0.6) (employment absorption capacity has a higher weight to avoid "training equals unemployment"). For example, if a computer science major in a certain region currently enrolls 100 students, F1 = 1.3 (strong industry demand), F2 = 1.2 (good employment absorption), then Stheoretical = 100 × (1.3 × 0.4 + 1.2 × 0.6) = 100 × 1.24 = 124 students (theoretically, an additional 24 students need to be enrolled).

[0080] Adjustment of constraints: If Sli exceeds the regional educational resource carrying capacity limit (e.g., the maximum number of teachers supported is 110), then the final adjustment will be Sactual = 110 people (subject to F3 constraint); if Sli is lower than the basic guarantee scale (e.g., the minimum number of clinical medical professionals in remote areas needs to be recruited is 50, while Sli = 40 people), then Sactual = 50 people (taking fairness into consideration).

[0081] Dynamic iteration mechanism: The algorithm updates industry-employment data (such as corporate recruitment demand and graduate employment reports) every quarter, recalculates F1 and F2, and adjusts the enrollment scale recommendations; when a regional industry undergoes a sudden change (such as the introduction of a large new energy company to a certain area), it triggers "emergency iteration" and quickly adjusts the enrollment scale prediction for the corresponding major (such as "Electrical Engineering and Automation").

[0082] Output layer: Regionally differentiated enrollment suggestions: The algorithm finally outputs structured optimization schemes by region and major, including: Direction of enrollment scale adjustment: expand enrollment, reduce enrollment, maintain the status quo; Adjustment basis: clearly marked "Because the new energy industry in a certain region grows by 30%+ and the employment gap is 25%, it is recommended to expand enrollment" and "Because the local employment rate of a certain major is less than 50%, it is recommended to reduce enrollment".

[0083] Decision Support Tool - Regional Heat Map: Uses color to indicate the "demand intensity" of each major (red for high demand, blue for low demand); Sensitivity Analysis: If the regional industry growth rate decreases by 10%, how much should the enrollment scale be adjusted accordingly (e.g., "If the growth rate decreases by 10%, the enrollment expansion rate will decrease from 15% to 8%"). The core value of this module lies in avoiding the disconnect between "nationally unified enrollment quotas" and "regional industry differences" through the logic of "geographic anchoring - data-driven - dynamic adaptation".

[0084] This invention employs a three-layer decoupled distributed system architecture comprising a data acquisition layer, a data processing layer, and an event detection layer. Each layer has a clearly defined functional boundary: the acquisition layer focuses on data acquisition and updating, the processing layer performs entity recognition and relationship extraction, and the detection layer is responsible for event perception and trend analysis. This design enables the system to have good module independence, facilitating separate upgrades and maintenance of each layer, while also supporting on-demand functional expansion, such as integrating new data sources or replacing large model versions, thereby meeting the complex and ever-changing application scenarios required by the education industry.

[0085] This large language model, fine-tuned for the system integration domain and linked with educational knowledge graphs through a Retrieval Augmentation (RAG) mechanism, demonstrates higher accuracy and semantic understanding capabilities in entity recognition and relation extraction tasks. Compared to traditional keyword- or rule-based methods, this large model can semantically analyze professional terminology, ability descriptions, and course content in complex contexts, uncovering implicit causal logic and relationships, effectively improving analysis quality. It is particularly suitable for processing unstructured data such as recruitment requirement texts and policy documents.

[0086] This invention supports local deployment of models and data, fully adapting to the high requirements of education systems for data privacy protection and policy compliance. All processing can be completed on internal servers of universities or education authorities, avoiding the risk of data leakage during transmission and retrieval. Simultaneously, local deployment significantly improves system efficiency, especially when facing large volumes of real-time analysis requests, enabling rapid response and stable operation, ensuring system availability and security.

[0087] The system incorporates a trend monitoring and risk identification module. Based on knowledge graphs and large-scale model reasoning capabilities, it can dynamically track factors such as industry events, changes in skill popularity, and policy adjustments. When events such as "a surge in the popularity of a certain skill" or "a major change in policy direction" are identified, the system automatically generates structured analysis results, including the cause of the event, the scope of its impact, and the policy basis. This assists educational institutions in scientifically formulating strategies such as curriculum adjustments and enrollment planning, and promotes the shift of educational resources from "experience-based allocation" to "data-driven" precise allocation.

[0088] Example 2

[0089] according to Figure 1 , 2 As shown in Figures 3, 4, and 5, this embodiment proposes an educational decision support system that dynamically adapts to industry needs. Taking "Adjustment of Computer Science Curriculum and Enrollment Optimization in a University" as an example, the system operation flow is as follows:

[0090] Data Acquisition Layer: Synchronize local "Digital Economy Industry Policies" and "IT Industry Recruitment Data" monthly to obtain data such as the surge in demand for "Large Model Development Capabilities" (year-on-year increase of 120%) and "Ministry of Education Digital Literacy Curriculum Requirements";

[0091] Data Processing Layer: DeepSeek large model identifies entities such as "large model development capability" and "digital literacy", associates them with "computer science courses" in the education knowledge graph, extracts relationships such as "Python programming - large model development" and "digital literacy - general education courses", and generates structured knowledge packages; the RAG module supplements knowledge such as "core skills of large model development (such as the LangChain framework)" and "digital literacy course credit standards";

[0092] Event Detection Layer: Identify events such as "a significant surge in demand for large-scale model development capabilities" and analyze the impact of policies on computer science majors (requiring the addition of digital literacy courses).

[0093] Educational resource constraint optimization model: Calculate the importance of "large model development related courses" S=(1.2×0.9)×1.2+0.2=1.528, and suggest increasing its class hours from 64 to 72 and credits from 3 to 4, while reducing the class hours of "traditional C++ courses";

[0094] Regional Adaptive Module: Based on the data of the "Eastern Digital Economy Industrial Belt" where the school is located (F1=1.4, F2=1.3), calculate S_theory=200×(1.4×0.4+1.3×0.6)=238. Considering the faculty constraints (F3=240), it is recommended to increase the enrollment of the computer science major by 19% (from 200 to 238 students).

[0095] Output a JSON format report, including course adjustment suggestions, enrollment size recommendations, policy basis, and risk assessment.

[0096] This education decision support system, dynamically adapted to industry needs, utilizes a multi-source data real-time perception mechanism and the RAG-fine-tuning collaborative reasoning framework to dynamically capture industry demands. When a skill popularity index surges, the system can quickly determine the event, analyze trends, and generate adjustment suggestions, demonstrating a short response cycle. This effectively addresses the issue of lagging educational supply in emerging technology fields, ensuring students' employment competitiveness. Furthermore, relying on the curriculum-skill-job ternary relationship of educational knowledge graphs and RAG semantic reasoning capabilities, the system can accurately map job skill requirements to university curriculum systems, identify specific course content deficiencies, and overcome the limitations of traditional technical professional-level macro-analysis. This enables precise course-level diagnosis, providing direct evidence for curriculum syllabus revision and promoting education adaptation from professional-industry alignment to curriculum-job alignment. Meanwhile, the regional adaptive module outputs differentiated enrollment suggestions by quantifying regional industry characteristics and employment dynamics, solving the regional imbalance caused by the uniform template. The education resource constraint optimization model quantifies the importance of courses based on skill requirements, enabling limited resources to be concentrated on high-value courses. At the same time, the local deployment mode ensures data security, and the trend and risk assessment module provides a scientific basis for decision-making, comprehensively improving the efficiency of education resource allocation and the adaptability to regional industries.

[0097] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An educational decision support system dynamically adapted to industry needs, comprising a data acquisition layer, a data processing layer, and an event detection layer, characterized in that: The data acquisition layer is used to periodically and synchronously collect raw data from local policy databases and recruitment market databases; the data processing layer adopts a locally deployed large model, and performs entity recognition, relation extraction and structuring processing on the collected raw data based on a preset educational knowledge graph to obtain structured semantic information. The educational knowledge graph is a semantic network that structurally represents the relationship between courses, abilities and positions. The data processing layer integrates a collaborative reasoning framework of RAG retrieval and domain fine-tuning large model. The event detection layer, based on structured semantic information, detects events in the education industry through a preset skill demand surge detection algorithm and an education policy analysis algorithm, generating trend analysis reports and risk assessment reports. The skill demand surge detection algorithm calculates a capability popularity index based on recruitment data, while the education policy analysis algorithm searches for policies and analyzes their impact through large-scale model reasoning.

2. The educational decision support system dynamically adapted to industry needs as described in claim 1, characterized in that: The raw data of the data acquisition layer includes: regional education policies and industry planning documents in the local policy database, as well as job demand information and job competency requirements information in the recruitment market database; the data acquisition layer synchronizes data on a monthly / quarterly basis to ensure data timeliness.

3. The educational decision support system dynamically adapted to industry needs as described in claim 1, characterized in that: The entity recognition and relationship extraction of the data processing layer includes: recognizing entities such as majors, courses, abilities, positions, and policies; extracting relationships such as courses-abilities, abilities-positions, and policies-majors; and supporting automatic deduplication and merging of multiple types of entities and multiple relationships.

4. The educational decision support system dynamically adapted to industry needs as described in claim 3, characterized in that: In the data processing layer, the execution flow of the collaborative reasoning framework for RAG retrieval and domain fine-tuning of the large model is as follows: Domain fine-tuning of the large model analyzes user questions, extracts key entities and knowledge gaps, and outputs search demand signals; The retrieval module performs targeted searches within the educational knowledge graph based on the retrieval demand signal, generating a structured knowledge package; The domain fine-tuning large model integrates the structured knowledge package with internal reasoning capabilities to verify logical consistency and supplement reasoning details; When a knowledge gap exists, a supplementary retrieval signal is generated to trigger a secondary retrieval until no key information is missing. Generate conclusions with credibility tags, and label the knowledge tracing and internal reasoning verification parts.

5. The educational decision support system dynamically adapted to industry needs according to claim 1, characterized in that: In the event detection layer, the quantitative standard for the skill demand surge detection algorithm is as follows: compare the current month / quarter with the previous period or the current period with the same period of the previous year. An increase of 20% is the first level, 50% is the second level, and more than 100% is the key attention level; or when a new skill appears in the top 50 of the skill demand index, it is judged as a skill demand surge event.

6. The educational decision support system dynamically adapted to industry needs as described in claim 1, characterized in that: In the event detection layer, the iterative search logic of the education policy analysis algorithm is as follows: based on the major name and training program, search for the initial policy through keyword fuzzy matching and extract the core objectives, related industries, talent training descriptions and preliminary questions; based on the preliminary questions, retrieve supporting policies and local implementation rules, update the policy information database and correct the questions until no key information is missing; Based on a comprehensive analysis of the impact of policy information on industries and professions, suggestions for adjusting the number and structure of professions and training programs are generated.

7. The educational decision support system dynamically adapted to industry needs according to claim 1, characterized in that: It also includes an educational resource constraint optimization model, which comprises a constraint modeling module, a course importance measurement module, a multi-constraint solution module, and a dynamic iteration module. The constraint modeling module sets hard and soft constraints. Hard constraints include total class hours ≤ upper limit of the training program, core course class hours ≥ 64 class hours, general education course credit ratio ≤ 20%, and practical course class hour ratio ≥ 25%. Soft constraints include course-skill matching degree ≥ 80%, and the priority weight of core courses being 1.5 times that of non-core courses. The course importance measurement module is used to calculate course importance scores. The multi-constraint solution module uses Max(Σ course importance score S × class hours) as the objective function and adjusts course credits / class hours through a linear programming algorithm. The dynamic iteration module recalculates course importance scores and optimizes credit / class hour allocation when skill demand increases by ≥ 20% or policies are updated.

8. The educational decision support system dynamically adapted to industry needs according to claim 7, characterized in that: The calculation formula for the important metric modules of the course is as follows: S = (Skill Demand Intensity × Matching Degree) × Course Attribute Coefficient + Policy Adaptability Score Among them, the skill demand intensity is the increase in the frequency of occurrence of the corresponding skills in industry recruitment, the matching degree is the degree of overlap between course content and target skills, the course attribute coefficient of core courses is 1.2, non-core courses are 0.8, and courses that conform to policy guidance are given an additional 0.2 points.

9. The educational decision support system dynamically adapted to industry needs according to claim 1, characterized in that: It also includes a regional adaptive module, which adopts a data layer-feature layer-algorithm layer-output layer architecture. The data layer stores regional industrial geographic distribution data and employment feedback data. The regional industrial geographic distribution data includes the GDP share of the leading industries, industry growth rate, and policy orientation. The employment feedback data includes local employment rate, professional matching rate, employment gap rate, salary premium, and employment stability. The feature layer calculates the regional demand intensity feature F1, the employment absorption capacity feature F2, and the regional carrying capacity constraint feature F3. The algorithm layer calculates the theoretical demand scale Stheoretical based on F1 and F2, and corrects it with F3 to obtain the actual enrollment scale Sreal. The output layer outputs regional and professional suggestions for expanding / reducing enrollment, demand heatmaps, and sensitivity analysis.

10. The educational decision support system dynamically adapted to industry needs according to claim 9, characterized in that: The formula for calculating the regional demand intensity characteristic F1 is as follows: F1 = Industrial linkage degree × Industrial growth coefficient × Policy weight; The formula for calculating the employment absorption capacity characteristic F2 is as follows: F2 = (Local employment rate × Job matching rate) × (1 + Employment gap rate) × Salary premium coefficient.

Images