A course system architecture optimization method based on hypergraph neural network and genetic algorithm

Abstract

The application belongs to the technical field of educational informatization, and discloses a course system structure optimization method based on a hypergraph neural network and a genetic algorithm, which comprises the following steps: collecting course data and external demand data, preprocessing the data, and constructing a weighted target feature matrix; establishing a course system structure state model according to a course set and a semantic vector of a course text; obtaining a system output feature vector according to a course-feature contribution matrix and a course prerequisite relationship matrix; identifying a weak target feature item according to the system output feature vector and the weighted target feature matrix; constructing a course hypergraph according to the course system structure state model and the weak target feature item, and generating an optimal course structure adjustment strategy; forming a new course system structure state according to the optimal course structure adjustment strategy; and repeatedly executing the system output feature vector calculation to the state updating step according to the new course system structure state, so as to realize continuous optimization of the course system structure.

Description

Technical Field

[0001] This invention belongs to the field of educational informatization technology, specifically relating to a curriculum system architecture optimization method based on hypergraph neural networks and genetic algorithms. Background Technology

[0002] With the development of educational informatization and intelligent technologies, curriculum system construction, curriculum relationship analysis, and dynamic adjustment of curriculum content have gradually become research focuses in the fields of data processing and intelligent optimization. Existing technologies have developed various implementation methods for curriculum system generation, curriculum recommendation, curriculum relationship analysis, and curriculum structure adjustment.

[0003] Some existing patents primarily construct curriculum systems through text processing, knowledge point extraction, or demand matching, enabling them to extract course nodes, divide knowledge points, and organize node relationships. Another type of patent mainly uses user behavior data, individual profiles, or course usage parameters to achieve course recommendation or course library construction, improving course matching efficiency. Yet another type of patent designs curriculum systems based on preset models, teaching processes, or knowledge point dependencies, forming a certain course organization logic. Furthermore, some technical solutions also use statistical analysis, course comparison, conceptual computation, or course group analysis to conduct rationality analysis, knowledge relationship updates, or structural display of the curriculum system.

[0004] However, the existing technologies mentioned above still have the following shortcomings: First, they have limited ability to express various relationships in the curriculum system, making it difficult to uniformly represent prerequisite dependencies, content similarity relationships, and collaborative support relationships between courses; second, they lack a unified model and computable representation of the overall structural state of the curriculum system, resulting in curriculum adjustments usually remaining at the level of local matching, local analysis, or experience-based revision; third, they lack effective technical means to perform global optimization solutions under multiple constraints such as credits, number of courses, prerequisite logic, and resource limitations; and fourth, they lack technical solutions for continuously updating the state of the curriculum system and achieving closed-loop iterative optimization by combining operational data.

[0005] Therefore, it is necessary to propose a curriculum system structure optimization method based on multi-source data semantic representation, hypergraph neural network modeling and optimization solution, so as to improve the curriculum system's structural modeling ability, global optimization ability and dynamic adjustment ability. Summary of the Invention

[0006] This invention aims to address the shortcomings of existing technologies and provides the following solutions:

[0007] A method for optimizing the curriculum architecture based on hypergraph neural networks and genetic algorithms includes the following steps:

[0008] S1. Collect course data and external demand data, preprocess and semantically vectorize them, and construct a weighted target feature matrix;

[0009] S2. Based on the semantic vectors of the course set and course texts, construct the course feature matrix, and calculate the course-feature contribution matrix and the course prerequisite relationship matrix to establish a course system architecture state model;

[0010] S3. Calculate the course aggregation weights based on the course-feature contribution matrix and the course prerequisite relationship matrix, and then obtain the system output feature vector that characterizes the comprehensive response of the curriculum system;

[0011] S4. Based on the system output feature vector and the weighted target feature matrix, calculate the feature deviation vector, and identify weak target feature items and optimize priorities based on the deviation threshold;

[0012] S5. Based on the course system structure state model and the weak target feature items, construct a course hypergraph that includes prerequisite dependencies, content similarity and collaborative support hyperedges, use a hypergraph neural network to extract the course node structure embedding representation, and solve it through a genetic algorithm under multiple constraints to generate the optimal course structure adjustment strategy;

[0013] S6. Update the course set, the course feature matrix, the course-feature contribution matrix, and the prerequisite relationship matrix according to the optimal course structure adjustment strategy to form a new course structure state;

[0014] S7. Based on the new curriculum system architecture state, repeat the steps of calculating the system output feature vector to updating the state, and continuously optimize the curriculum system architecture through closed-loop iteration until the convergence condition is met.

[0015] Preferably, the method for constructing the weighted target feature matrix includes:

[0016] Collect the course data and the external demand data. The course data includes: course text data, course attribute data and course relationship data. The external demand data includes: job description text, skill tag text, task description text and industry documents.

[0017] The course data and the external demand data are subjected to text cleaning, format standardization, word segmentation, terminology normalization, synonym merging, stop word filtering, and abnormal data removal to obtain the processed data.

[0018] Extract the target feature set from the processed data and use the Sentence-BERT pre-trained language model for semantic vectorization to obtain the semantic vector representation of each feature item;

[0019] Construct a target feature matrix and assign weights to the target feature items based on their frequency of occurrence or importance score in the external demand data to obtain the weighted target feature matrix.

[0020] Preferably, the method for constructing the state model of the course system architecture includes:

[0021] Construct the course set and its corresponding attribute parameters, which include: course name, course number, course category, course module, course credits, total course hours, semester offered, and syllabus text;

[0022] The course text data is preprocessed and semantically vectorized to obtain the course feature matrix;

[0023] The course-feature contribution matrix is ​​calculated by the cosine similarity between the course feature vectors and the target feature vectors of the course feature matrix, and the comparability of contribution values ​​is ensured by Softmax normalization.

[0024] Extract the course prerequisite relationship information and construct the course prerequisite relationship matrix, where a matrix element of 1 indicates a prerequisite dependency and a matrix element of 0 indicates no dependency;

[0025] Based on the course set, the course feature matrix, the course-feature contribution matrix, and the course prerequisite relation matrix, construct the course system architecture state model for the k-th iteration.

[0026] Preferably, the method for obtaining the system output feature vector includes:

[0027] The course-feature contribution matrix is ​​weighted and aggregated along the course dimension, and the course aggregation weight is calculated by combining the course credits and hierarchical attributes.

[0028] If a course is a core course, it will be given a higher weight; if it is a supporting course, it will be given a lower weight.

[0029] By combining the course prerequisite relation matrix with the course aggregation weights, the system output feature vector is obtained, which is used to characterize the comprehensive response of the curriculum system in each target feature dimension.

[0030] The system output feature vector is subjected to interval constraint processing to ensure that the output of each target feature dimension is within a reasonable range.

[0031] Preferably, the method for identifying the weak target features includes:

[0032] The system output feature vector is mapped to the same semantic space as the target feature matrix, and the feature deviation matrix is ​​calculated.

[0033] The feature deviation matrix is ​​transformed into a scalar deviation vector, where each component represents the deviation intensity of the corresponding target feature term;

[0034] Based on the deviation intensity, the weak target feature items are filtered according to the deviation threshold, and a sequence of optimization priorities is generated according to the deviation magnitude;

[0035] The mean and variance of the deviation, as well as the change in deviation between two adjacent iterations, are calculated to assess the overall deviation level and local imbalance of the curriculum system, and to provide a basis for dynamic adjustment for subsequent optimization.

[0036] Preferably, the method for generating the optimal curriculum structure adjustment strategy includes:

[0037] The curriculum system is represented as the curriculum hypergraph, which includes a set of course nodes and three types of hyperedges, namely: prerequisite dependency hyperedges, content similarity hyperedges, and collaborative support hyperedges.

[0038] The prerequisite dependency hyperedge is constructed from the course prerequisite relation matrix, the content similarity hyperedge is constructed when the semantic similarity of the course feature vectors exceeds a threshold, and the collaborative support hyperedge is constructed from courses that contribute highly to the same weak target feature item.

[0039] Based on the course hypergraph, a node degree matrix and a hyperedge degree matrix are constructed, and then input into a multilayer hypergraph convolutional neural network to obtain the course node embedding representation;

[0040] Define course structure adjustment actions: adding courses, deleting courses, adjusting course content, adding and deleting prerequisite relationships, and encode the actions to form a chromosome representation;

[0041] The objective functions are set as minimizing deviation and minimizing adjustment cost, and a balance coefficient is introduced;

[0042] Set constraints: prerequisite relationships must be acyclic, total course credits, number of courses, budget, and content adjustment limits;

[0043] Based on the course node embedding representation, a genetic algorithm is used to perform selection, crossover, mutation, and elite retention operations, and fuzzy logic is combined to adaptively adjust the crossover probability, mutation probability, and penalty coefficient to generate the optimal course structure adjustment strategy.

[0044] Preferably, the fitness function of the genetic algorithm includes a constraint violation degree, which is used to penalize infeasible solutions with prerequisite cycles, credit limits, course limits, budget limits, and content adjustment limits.

[0045] The inputs to the fuzzy logic include the overall deviation intensity, the deviation change rate, and the population distribution state. By adjusting the crossover probability, mutation probability, and penalty coefficient through rules, the parameters are adaptively optimized.

[0046] Preferably, the method for obtaining the new curriculum structure state includes:

[0047] The course set, course feature matrix, course-feature contribution matrix, and prerequisite relation matrix are updated according to the optimal course structure adjustment strategy.

[0048] The newly added courses are semantically vectorized and their contribution values ​​are calculated. The corresponding matrix rows and columns are removed for the deleted courses. The prerequisite relationship matrix is ​​updated synchronously.

[0049] After the update, a consistency check is performed on the number of courses, the total credit hours for the courses, the semester in which the courses are offered, and the budget constraints.

[0050] Ensure that the updated new curriculum architecture meets the requirements of computability, implementability, and optimizability.

[0051] Preferably, S7 includes:

[0052] Use the new course system architecture state as the input for the next iteration, and repeat S3 to S6.

[0053] Triggering conditions include fixed training cycles and dynamic conditions, wherein the dynamic conditions include: significant changes in external demand, overall deviation exceeding a threshold, and significant changes in curriculum structure;

[0054] A smooth update mechanism is introduced to spread large-scale adjustments across multiple iterations to avoid drastic fluctuations in the course structure.

[0055] Iterate until the overall deviation converges to a preset threshold, or the weak target feature term stabilizes;

[0056] After convergence, the system enters monitoring mode to continuously observe the operational data of the curriculum system and external demands. If the deviation expands again, closed-loop optimization is triggered again.

[0057] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0058] (1) By constructing a course hypergraph that includes three types of hyperedges: prerequisite dependencies, content similarity, and collaborative support, this invention breaks through the limitation of traditional graph structures that can only represent binary relationships. The hypergraph structure can effectively capture the complex nonlinear relationships between multiple courses that jointly support a certain skill objective, realize the unified modeling of multiple logical relationships within the curriculum system, and make the structural representation more accurate.

[0059] (2) It combines the feature extraction capability of hypergraph neural networks with the global search capability of genetic algorithms. The node embedding representation extracted by hypergraph neural networks can deeply integrate structural information, while the genetic algorithm with fuzzy logic adaptive adjustment can find the balance point between minimizing deviation and minimizing cost under multiple strict constraints such as credit, budget, and prerequisite logic, effectively avoiding the problem of local optima.

[0060] (3) A closed-loop iterative system was established, encompassing demand collection, status assessment, deviation identification, structural optimization, and status updating. By introducing a smooth update mechanism, this invention can not only cope with drastic fluctuations in external industry demand but also ensure the stability of curriculum system adjustments, achieving dynamic and precise alignment between educational supply and industry demand. Attached Figure Description

[0061] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0062] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention;

[0063] Figure 2 This is a flowchart illustrating the control strategy of an embodiment of the present invention. Detailed Implementation

[0064] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0065] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0066] Example 1

[0067] In this embodiment, as Figure 1 , Figure 2 As shown, a method for optimizing the curriculum architecture based on hypergraph neural networks and genetic algorithms includes the following steps:

[0068] S1. Collect course data and external demand data, preprocess and semantically vectorize them, and construct a weighted target feature matrix.

[0069] The method for constructing a weighted target feature matrix includes: collecting course data and external demand data. Course data includes course text data, course attribute data, and course relationship data. External demand data includes job description text, skill tag text, task description text, and industry documents. The course data and external demand data are cleaned, formatted, segmented, normalized, synonyms merged, stop words filtered, and outlier data removed to obtain processed data. The set of target feature items is extracted from the processed data and semantically vectorized using the Sentence-BERT pre-trained language model to obtain the semantic vector representation of each feature item. The target feature matrix is ​​constructed, and weights are assigned based on the frequency or importance of the target feature items in the external demand data to obtain the weighted target feature matrix.

[0070] In this embodiment, this step is used to construct a target feature set for curriculum system architecture optimization, and to transform the target feature set into a computable semantic vector representation, forming the target input in the curriculum system architecture optimization process, providing a unified feature benchmark for subsequent curriculum system architecture state modeling, system output feature calculation, deviation analysis and optimization solution.

[0071] First, external requirement data related to curriculum structure optimization is collected. This external requirement data includes job descriptions, skill tag texts, task descriptions, thematic requirement texts, industry technical documents, or other information that can characterize the curriculum system's target requirements. Due to differences in format, expression, and granularity among these data sources, unified preprocessing is necessary. The preprocessing includes text cleaning, format standardization, word segmentation, terminology normalization, merging of synonyms, stop word filtering, and outlier removal.

[0072] The process involves several steps: First, text data is cleaned using regular expression matching or rule filtering to remove special symbols, repeated punctuation, web page tags, and invalid fields. Second, data format is standardized using encoding conversion and field mapping to convert data from different sources into a unified encoding format and field structure. Third, text is segmented using lexical analysis models or word segmentation algorithms to divide continuous text into word sequences or phrase sequences. Fourth, terms are normalized using term mapping dictionaries or rule replacement methods to map semantically similar but differently expressed terms to a unified expression. Fifth, terms are merged using thesaurus or semantic similarity calculation methods to group terms with similar semantics into the same standard term. Sixth, terms are filtered using stop word lists to remove terms that contribute little to semantic representation. Finally, anomaly detection is performed using rule validation or threshold judgment methods to remove data records with missing content, field conflicts, or obvious invalidity. After these processes, the multi-source, heterogeneous external demand data is transformed into a comparable and aggregatable standardized text set. After data preprocessing, feature terms are extracted from the external demand data to form a target feature term set. Let the target feature term set be:

[0073] ;

[0074] Where B represents the set of target feature terms, Indicates the number of target feature terms. This represents the i-th target feature. Each target feature is used to characterize the target requirements content that the curriculum system intends to cover during the optimization process. Each target feature can be a phrase, a task topic, a skill description, or a technical concept, or it can be a combination of multiple terms.

[0075] To imbue the target feature terms with computable and comparable quantifiable attributes, a Sentence-BERT pre-trained language model is used to semantically vectorize each target feature term, denoted as Ψ. For any target feature term... Inputting it into the pre-trained language model Ψ yields the corresponding semantic vector representation, i.e. .in, Let represent the semantic vector corresponding to the i-th target feature term, and d represent the dimension of the semantic vector. Thus, all target feature terms are mapped to a unified semantic space. For ease of subsequent calculation, the set of target feature vectors is represented as a target feature matrix:

[0076] Thus, all target feature terms are mapped to a unified semantic space. For ease of subsequent calculation, the set of target feature vectors is represented as a target feature matrix:

[0077] ;

[0078] Where T represents the transpose of the matrix; Y represents the target feature matrix. Each row of the matrix corresponds to a semantic vector representation of a target feature item. The target feature matrix comprehensively depicts the target demand state corresponding to the optimization of the curriculum system structure, serving as a reference benchmark for subsequent curriculum system structure modeling and deviation analysis. Considering that different target feature items may have different degrees of influence in the overall curriculum system structure optimization, a target feature weight vector is further constructed. The weight vector is used to characterize the relative importance of each target feature item in the target feature set. Let the target feature weight vector be:

[0079] ;

[0080] Where W represents the target feature weight vector, This represents the weight parameter of the i-th target feature.

[0081] The weighting parameters are determined based on the frequency of occurrence of the target feature items in the external demand data. In implementation, to characterize the relative importance of different target feature items in the overall curriculum structure optimization, a target feature weight vector is further constructed. Let the i-th target feature item... The frequency of occurrence in external demand data is: Then its weight parameters Defined as:

[0082] ;

[0083] The resulting weight parameters satisfy Furthermore, by weighting the target feature matrix according to the target feature weight vector, a weighted target feature matrix can be obtained. The weighted target feature matrix is ​​used for bias vector construction and target optimization calculation in subsequent steps.

[0084] ;

[0085] This step, through unified preprocessing of external demand data, extraction of target features, and semantic vectorization, constructs a set of target features and their semantic representations for the curriculum architecture optimization process. This transforms the originally scattered, heterogeneous, and unstructured demand information into computable objects in a unified semantic space. Consequently, it provides a unified target reference system for subsequent curriculum architecture state modeling, system output feature vector calculation, and deviation analysis, transforming the curriculum architecture optimization process from unstructured processing based on text descriptions to computable processing based on vector representations.

[0086] S2. Based on the semantic vectors of the course set and course texts, construct the course feature matrix and calculate the course-feature contribution matrix and the course prerequisite relationship matrix to establish a course system architecture state model.

[0087] The method for constructing a course system architecture state model includes: constructing a set of courses and corresponding attribute parameters, including: course name, course number, course category, course module, course credits, total course hours, semester offered, and syllabus text; preprocessing and semantically vectorizing the course text data to obtain a course feature matrix; calculating the course-feature contribution matrix using the cosine similarity between the course feature vectors and target feature vectors in the course feature matrix, and ensuring the comparability of contribution values ​​through Softmax normalization; extracting course prerequisite relationship information and constructing a course prerequisite relationship matrix, where a matrix element of 1 indicates prerequisite dependency and 0 indicates no dependency; and constructing the course system architecture state model under the k-th iteration based on the course set, course feature matrix, course-feature contribution matrix, and course prerequisite relationship matrix.

[0088] In this embodiment, this step constructs a course set, a course-feature contribution matrix, and a course prerequisite relationship matrix to represent the course system as a computable, analyzable, and updatable structural state model, which serves as the basis for subsequent system output feature calculation, deviation vector construction, and optimization solution.

[0089] First, basic course data is collected from the curriculum system to be optimized. This basic data includes one or more of the following: course name, course number, course attribute, course category, module to which the course belongs, course credits, total course hours, semester offered, syllabus text, knowledge point description text, experimental project description text, practical component description text, and prerequisite information. Based on this basic data, a course set is constructed. The course set at the k-th iteration is defined as... ,in, Let be the number of courses in the k-th iteration. Let represent the j-th course. All courses together constitute the basic building blocks of the current curriculum. To facilitate unified handling of constraints such as the number of courses, course credits, semester offered, and course category during subsequent optimization, course attribute parameters are configured for each course in the course set. For any course in the course set... Collect the corresponding course text description set. Let the course... The set of course text descriptions is as follows:

[0090] ;

[0091] in, Represented as a course A collection of texts, These represent text units such as course name, course description, syllabus excerpt, key knowledge points, and experimental instructions. In addition, courses are... The set of attribute parameters is as follows:

[0092] ;

[0093] in, Indicates course Credits, Indicates course The semester in which classes begin, Indicates course Course categories, Indicates course The total class hours. Course attribute parameters are used to calculate the total course hours, credit hours, course start cycle, and resource constraints during subsequent optimization. Since the text descriptions of different courses vary in length, granularity, and expression, similar preprocessing operations to S1 are required on the course text description set. These include text cleaning, format standardization, word segmentation, terminology normalization, merging of synonyms, stop word filtering, and removal of abnormal text to form a standardized course text set.

[0094] After completing the course text preprocessing, the same Sentence-BERT pre-trained language model Ψ as in S1 is used to perform semantic vectorization on each course text, thereby obtaining the course feature vector for each course. For any course Its course feature vector is denoted as ,in, Indicates course The corresponding feature vector, d, represents the vector dimension. Therefore, all courses in the course set can be mapped to a set of course feature vectors, which can then be represented as a course feature matrix: Each row corresponds to a semantic feature representation of a course. The course feature matrix is ​​used for subsequent calculation of course-feature contribution relationships and modeling of relationships between courses.

[0095] In obtaining the course feature matrix and target feature matrix Next, the contribution of each course to each objective feature is calculated, thus constructing a course-feature contribution matrix. For any course... and any target feature Course feature vectors are calculated using the cosine similarity method. With target feature vector The degree of matching between them determines the course. For target feature terms Original contribution value:

[0096] ;

[0097] in, Indicates course For target feature terms The original semantic matching value. The larger the matching value, the stronger the course. Content semantics and target feature terms The higher the semantic relevance, the stronger the support of the course for that target feature. To improve the comparability of the contribution values ​​of each course to each target feature and to normalize the contribution of each course across the target feature dimensions, the original contribution values ​​at the k-th iteration are... The course was obtained by using Softmax normalization. For target feature terms Normalized contribution value:

[0098] ;

[0099] in, And for any course ,have .when The closer it is to 1, the better the course. With target feature terms The higher the degree of correlation; when The closer the value is to 0, the lower the correlation between the two. Furthermore, the normalized contribution values ​​of all courses to all target feature terms are combined to construct the course-feature contribution matrix for the k-th iteration. The j-th row in the matrix represents the course. The contribution distribution across all target feature dimensions, where the i-th column of the matrix represents the target feature term. The contribution distribution is supported by all courses. The course-feature contribution matrix realizes the quantitative correlation between course units and target feature units, and is the core foundation for subsequent system output feature vector calculation and deviation analysis.

[0100] After constructing the course-feature contribution matrix, the prerequisite dependencies between courses are further extracted to construct a course prerequisite relationship matrix. For any course... Based on the prerequisite information in the basic course data, extract the set of prerequisite courses, denoted as . .in, Indicates learning courses The set of courses that need to be completed beforehand. Based on the prerequisite course set for each course, construct the course prerequisite relationship matrix for the k-th iteration. Let the matrix elements be... for:

[0101] ;

[0102] in, Indicates course It is a course prerequisite courses Indicates course With courses There are no prerequisite dependencies. Therefore, the course prerequisite matrix can be obtained as follows:

[0103] ;

[0104] The prerequisite relation matrix is ​​used to represent the learning sequence constraints and course dependency structure in the curriculum system.

[0105] Therefore, the state model of the course system architecture under the k-th iteration can be expressed as follows: .in, This represents the structural state model of the curriculum system at the k-th iteration. The structural state model fully describes the course composition, the contribution relationship between courses and target features, and the prerequisite dependencies between courses in the current curriculum system. It is the direct object of subsequent system output feature vector calculation, deviation vector construction, curriculum hypergraph modeling, and curriculum system structure optimization solution.

[0106] This step involves unified modeling of course data, course text data, and course relationship data within the curriculum system. This results in the construction of course sets, course feature matrices, course-feature contribution matrices, and course prerequisite relationship matrices, further forming a state model of the curriculum system structure. This transformed the curriculum system from a raw course list, textual description, and relational records into a structured, computable state representation, laying the foundation for subsequent system output feature vector calculation, deviation analysis, and multi-constraint optimization solutions.

[0107] S3. Based on the course-feature contribution matrix and the course prerequisite relationship matrix, calculate the course aggregation weights, and then obtain the system output feature vector that represents the comprehensive response of the curriculum system.

[0108] The method for obtaining the system output feature vector includes: weighting and aggregating the course-feature contribution matrix along the course dimension, and calculating the course aggregation weight by combining course credits and hierarchical attributes; assigning higher weights to core courses and lower weights to supporting courses; correcting the course aggregation weights by combining the course prerequisite relation matrix to obtain the system output feature vector, which is used to characterize the comprehensive response of the curriculum system in each target feature dimension; and performing interval constraint processing on the system output feature vector to ensure that the output of each target feature dimension is within a reasonable range.

[0109] In this embodiment, the course architecture state model is constructed using S2. Given the current state of the curriculum system, this step aggregates and calculates the curriculum contribution values ​​in the curriculum-feature contribution matrix to obtain the system output feature vector of the curriculum system under the current structural state. This vector is used to characterize the comprehensive response of the current curriculum system to each target feature item, providing a quantitative basis for subsequent deviation vector construction and structural optimization solutions.

[0110] Since the overall output of the curriculum system is not determined by a single course, but rather by the combined effect of multiple courses across different target feature dimensions, a course-feature contribution matrix is ​​needed. Aggregate calculations are performed along the course dimension. Specifically, for any target feature term... The contribution values ​​of all courses to this objective feature are weighted and aggregated to obtain the contribution of the curriculum system to the objective feature. The system output value on is denoted as The system output value is defined as follows:

[0111] ;

[0112] in, Indicates course The course aggregation weights in the k-th iteration satisfy:

[0113] ;

[0114] Among them, course aggregation weight Used to characterize courses The degree of structural influence of a course within the overall curriculum system. To ensure that the system's output feature vector reflects both the course's content contribution and its structural position within the curriculum system, this implementation determines the course aggregation weight based on both course credits and course hierarchical attributes. Let the course... Credits The course level coefficient is Then the course The original structural weights are defined as follows:

[0115] ;

[0116] in, Used to characterize courses Its hierarchical position within the curriculum system. If the course... If it belongs to the core courses, then The value is relatively high; if the course If it belongs to general support courses, then The value is relatively low. In the implementation method, to enhance the sensitivity of the system's output feature vector to the course structure relationship, the course prerequisite relationship matrix can also be combined. The course aggregation weights are adjusted. Specifically, the structural influence coefficient of each course is first calculated based on the course prerequisite relationship matrix. Course Design The structural influence coefficient is defined as:

[0117] ;

[0118] in, Indicates course The number of times a course is a prerequisite for other courses. The higher the value, the more prerequisite courses are required. The stronger its fundamental supporting role in the curriculum system, the better the curriculum will be. The original structural weights are corrected as follows:

[0119] ;

[0120] Then, the corrected course aggregation weights are obtained through normalization:

[0121] ;

[0122] This allows the system's output feature vector to not only reflect the matching relationship between course content and target features, but also the fundamental, supporting, and transmissive role of the course within the curriculum structure. Therefore, the system output values ​​for each target feature item of the curriculum system can be further written as:

[0123] ;

[0124] The curriculum system in terms of target characteristics The system output value is determined by the contribution of all courses to the target feature, and different courses have different degrees of influence on the overall output according to their credit ratio and hierarchical attributes.

[0125] Furthermore, to avoid an imbalance in the overall output caused by an excessively high single contribution value, after obtaining the system output feature vector, interval constraints can be applied to each component of the vector. For any target feature term... Its final output value is denoted as Defined as:

[0126] ;

[0127] Furthermore, the constrained system output feature vector can be obtained. .in, Let represent the system output feature vector after interval constraint processing in the k-th iteration. The i-th component of the vector... This indicates the curriculum system's target characteristics. The overall response level. The system output feature vector fully characterizes the output results of the current course system structure state across all target feature dimensions, and is the core intermediate quantity connecting the course system structure state with the subsequent deviation vector construction. For ease of subsequent matrix operations, the system output feature vector can be written in matrix form. Let the course aggregation weight vector be:

[0128] ;

[0129] The system outputs a feature vector. It can be represented as ,in, , ,therefore, This matrix representation facilitates its integration with the target feature matrix in subsequent steps. Perform deviation comparison and unified calculation.

[0130] This step involves analyzing the course-feature contribution matrix. By performing weighted aggregation, the system output feature vector of the curriculum system under the current structural state was calculated. The system outputs a feature vector to characterize the comprehensive response of the current curriculum system to all target feature items, realizing the state model of the curriculum system structure. This maps the results to the overall output of the curriculum system. This provides a quantitative foundation for subsequent deviation vector construction, weak feature identification, and curriculum system structure optimization, enabling the curriculum system structure optimization process to naturally transition from the modeling stage based on structural representation to the optimization stage based on output deviation.

[0131] S4. Based on the system output feature vector and the weighted target feature matrix, calculate the feature deviation vector, and identify weak target feature items and optimize priorities based on the deviation threshold.

[0132] The method for identifying weak target features includes: mapping the system output feature vector to the same semantic space as the target feature matrix and calculating the feature deviation matrix; converting the feature deviation matrix into a scalar deviation vector, where each component represents the deviation intensity of the corresponding target feature; based on the deviation intensity, filtering weak target features according to the deviation threshold and generating a sequence of optimization priorities according to the deviation magnitude; calculating the mean and variance of the deviation, as well as the change in deviation between adjacent iterations, to assess the overall deviation level and local feature imbalance of the curriculum system, and to provide a dynamic adjustment basis for subsequent optimization.

[0133] In this embodiment, the weighted target feature matrix constructed with S1 As the target input, the system output feature vector after interval constraint processing is the output of S3. As the output of the current curriculum system, this step quantifies the difference between the current structural state and the target state of the curriculum system by constructing a deviation vector between the target features and the system output, and further identifies the target feature items with large deviations, forming the deviation signal required for subsequent curriculum system structure optimization.

[0134] Due to the weighted target feature matrix Each row in the vector corresponds to a weighted semantic vector, while the system outputs a feature vector. Each component in the matrix corresponds to a scalar output value, and the two differ in form. Therefore, it is necessary to first map the system output feature vector to a feature representation space consistent with the target feature matrix. Specifically, for any target feature term... , its system output value As a scaling factor, it is applied to the target feature vector. The target feature term is obtained in the k-th iteration. The output feature vector is denoted as:

[0135] ;

[0136] in, This indicates that the curriculum system in the k-th iteration has the following target feature term. The output feature vector. When When, it indicates that the curriculum system has reached the output level corresponding to the target feature vector on that target feature item; when When, it indicates that the output of the curriculum system on that target characteristic is zero; when When this occurs, it indicates that the curriculum system is in a state of partial response to that objective characteristic.

[0137] Furthermore, the output feature vectors corresponding to all target feature terms are combined to construct the output feature matrix of the curriculum system in the k-th iteration. .in, The i-th row in the matrix represents the curriculum system in terms of target features. The output feature representation. In obtaining the output feature matrix... Then, it is compared with the weighted target feature matrix. By performing difference calculations, the eigenvalue deviation matrix for the k-th iteration is obtained:

[0138] ;

[0139] Will and Substituting the definition into the above equation and rearranging, we get:

[0140] ;

[0141] It can be seen that the deviation corresponding to the i-th target feature term is determined by the target feature weight parameters. and the system output value of the curriculum system on this target characteristic item The decision is made jointly by both parties. The closer When the difference is large, it indicates that the output of the curriculum system on that target feature is closer to the target state; when the difference is large, it indicates that the deviation of the curriculum system on that target feature is more obvious. In order to facilitate the direct use of deviation information in subsequent optimization solutions, it is necessary to further refine the feature deviation matrix. Converted to a scalar bias vector. Specifically, for any target feature term... The scalar deviation value under its k-th iteration is defined as:

[0142] ;

[0143] Furthermore, the scalar deviation values ​​of all target feature terms are combined to obtain the deviation vector in the k-th iteration:

[0144] ;

[0145] in, Its i-th component This indicates the curriculum system's target characteristics. The intensity of the deviation. When When, it indicates that the curriculum system has the following characteristics in terms of objectives. The current output is lower than the target requirement; when When, it indicates that the curriculum system is consistent with the target state in that target characteristic item; when When this occurs, it indicates that the output of the curriculum system on that target feature is higher than the current weight target. To further quantify the overall deviation level of the curriculum system, the deviation vector can be... Calculate its norm. Preferably, the L2 norm is used as the overall deviation strength index, denoted as:

[0146] ;

[0147] in, This represents the overall deviation strength of the curriculum system in the k-th iteration. The larger this value, the greater the overall difference between the current curriculum system structure and the target state; the smaller this value, the closer the current curriculum system structure is to the target state.

[0148] Furthermore, to identify the target feature terms that need to be prioritized for optimization, the bias vector is... The components are sorted and filtered. A preset deviation threshold is set as follows: Then, target feature terms that satisfy the following conditions are defined as weak target feature terms:

[0149] ;

[0150] in, Let represent the set of weak target features in the k-th iteration. The target features in this set correspond to the target dimensions with insufficient output or significant deviation in the current curriculum system, and are the areas that need to be focused on for correction in subsequent curriculum structure adjustments. In some implementations, the weak target features can also be prioritized based on the magnitude of the deviation value. Let the ranking function be . Then, by sorting the set of weak target feature items according to the deviation value from largest to smallest, a priority sequence can be obtained:

[0151] ;

[0152] in, This represents the optimization priority sequence of weak target feature terms in the k-th iteration. Target feature terms that appear earlier in the sequence indicate a larger current deviation and should be given a higher adjustment priority in subsequent optimization solutions. To characterize the distribution of deviations in the curriculum system, the mean and variance of the deviation vector can be further calculated. Let the mean deviation be... The deviation variance is Then we have:

[0153] ;

[0154] ;

[0155] in, Used to characterize the average deviation level of the curriculum system across all target feature dimensions. Used to characterize the degree of dispersion of the deviation distribution between different target features. When When the value is large, it indicates that the deviation distribution among different target feature terms is uneven, and the local feature terms with prominent deviations need to be addressed in the subsequent optimization process; when When the deviation is relatively small, it indicates that the deviation of the curriculum system across various target feature dimensions is relatively balanced, and subsequent optimization can adopt a holistic adjustment strategy.

[0156] Furthermore, to reflect the trend of deviation changes in adjacent iterations, the amount of deviation change can also be calculated. Let the deviation vector in the (k-1)th iteration be... Then the deviation change vector under the k-th iteration is defined as:

[0157] ;

[0158] Wherein, the i-th deviation change component is . Representing target feature terms The change in deviation between two adjacent iterations. When, it represents the target feature term. The deviation is widening; when When, it represents the target feature term. The deviation is decreasing. The deviation change vector can serve as an important basis for dynamically adjusting the optimization intensity in subsequent optimization processes.

[0159] This step involves weighting the target feature matrix. The system output feature vector after interval constraint processing, compared with the output of S3 Alignment and difference calculations were performed to construct the deviation vector of the curriculum system under the current structural state. Furthermore, it identified a set of weak target features. and its priority sequence This enables a quantitative representation of the difference between the current state and the target state of the curriculum system structure, allowing subsequent optimization of the curriculum system structure to be directionally adjusted based on clear deviation signals, thereby driving the curriculum system structure to continuously approach the target state.

[0160] S5. Based on the state model of the curriculum system structure and the weak target features, construct a curriculum hypergraph that includes prerequisite dependencies, content similarity and collaborative support hyperedges. Use a hypergraph neural network to extract the embedded representation of the curriculum node structure, and solve it through a genetic algorithm under multiple constraints to generate the optimal curriculum structure adjustment strategy.

[0161] The method for generating the optimal course structure adjustment strategy includes: representing the course system as a course hypergraph, which includes a set of course nodes and three types of hyperedges: prerequisite dependency hyperedges, content similarity hyperedges, and collaborative support hyperedges; prerequisite dependency hyperedges are constructed from the course prerequisite relation matrix, content similarity hyperedges are constructed when the semantic similarity of course feature vectors exceeds a threshold, and collaborative support hyperedges are constructed from courses that contribute highly to the same weak target feature item; constructing node degree matrices and hyperedge degree matrices based on the course hypergraph and inputting them into a multi-layer hypergraph convolutional neural network to obtain the course node embedding representation; defining course structure adjustment actions: adding courses, deleting courses, adjusting course content, and adding and deleting prerequisite relations, and encoding these actions to form chromosome representations; setting the objective function as minimizing deviation and minimizing adjustment cost, and introducing a balance coefficient; setting constraints: acyclic prerequisite relations, total course credits, number of courses, budget, and content adjustment range constraints; and using a genetic algorithm to perform selection, crossover, mutation, and elite retention operations based on the course node embedding representation, and combining fuzzy logic to adaptively adjust the crossover probability, mutation probability, and penalty coefficient to generate the optimal course structure adjustment strategy.

[0162] The fitness function of the genetic algorithm includes the constraint violation degree, which is used to penalize infeasible solutions with prerequisite cycles, credit limits, course limits, budget limits, and content adjustment limits. The input of fuzzy logic includes the overall deviation intensity, deviation change rate, and population distribution state. The crossover probability, mutation probability, and penalty coefficient are adjusted by rules to achieve adaptive optimization of parameters.

[0163] In this embodiment, this step uses the course architecture state model constructed in S2. As the current state of the curriculum system, the deviation vector obtained from S4 Set of weak target features and priority sequence As the driving information for optimization, the current curriculum system is structurally optimized.

[0164] First, a unified model is constructed to represent the course relationships within the current curriculum system, resulting in a course hypergraph. Let the course hypergraph in the k-th iteration be:

[0165] ;

[0166] in, Represents the set of course nodes. This is a set of superedges. Preferably, the set of course nodes is directly composed of the set of courses in S2, i.e. Each course corresponds to a node in the hypergraph. To enable the course hypergraph to simultaneously express the learning order relationship, content similarity relationship, and collaborative support relationship for the same target feature among courses, three types of hyperedges are constructed in this implementation. The first type is the prerequisite dependency hyperedge. Based on the course prerequisite matrix The relationships within are constructed. For any course... and ,like This indicates the course It is a course The first type is the prerequisite courses, and the corresponding prerequisite dependencies are established in the course hypergraph to represent the knowledge transfer paths and learning order constraints between courses. The second type is content-similar hyperedges. It is constructed based on the semantic similarity between course feature vectors. For any course and Based on the course feature vector obtained in S2 and Calculate the content similarity between the two:

[0167] ;

[0168] when Greater than the preset content similarity threshold At that time, the course and courses Similar hyperedges, categorized under the same content, are used to represent semantically similar, substitutable, or potentially integrated relationships between courses. The third category is collaborative support hyperedges. Based on the course-feature contribution matrix and the set of weak target features Construction is performed. For any weak target feature term... Extract the set of courses with high contribution values ​​to the target feature from the course-feature contribution matrix, denoted as:

[0169] ;

[0170] in, To preset the contribution threshold, Indicates the target feature terms A set of courses that provides strong support. When the set... When there are at least two courses, these courses are constructed as collaborative support hyperedges to represent the joint support relationship of multiple courses for the same target feature. Therefore, the set of hyperedges in the course hypergraph can be represented as:

[0171] ;

[0172] in, Let represent the set of superedges with prior repair relations. Represents a set of superedges with similar content. This represents a set of collaboratively supporting superedges.

[0173] To facilitate subsequent hypergraph convolution operations, the association matrix of the course hypergraph is constructed. Its elements Defined as follows: a value of 1 is taken when a course node v belongs to a hyperedge e, and a value of 0 otherwise. The association matrix is ​​used to record the affiliation relationship between course nodes and various types of hyperedges, and is the foundation for subsequent hypergraph neural network implementation of node feature propagation and high-order structure aggregation.

[0174] After the course hypergraph is constructed, structural embedding learning is performed on the course nodes. In S2, the course feature matrix has already been obtained through semantic vectorization of the course text. This matrix is ​​used as the initial input feature matrix of the hypergraph neural network, denoted as: To characterize the local connectivity and global structural relationships of course nodes in the hypergraph, a node degree matrix is ​​further constructed. and hypermarginality matrix The node degree matrix represents the number of hyperedges each course node participates in, and its diagonal elements are defined as follows:

[0175] ;

[0176] The hyperedge degree matrix is ​​used to represent the number of course nodes connected by each hyperedge, and its diagonal elements are defined as follows:

[0177] ;

[0178] Building upon this, multi-layer hypergraph convolution is employed for representing course nodes. Let the output of the l-th layer hypergraph convolution be... Then we have:

[0179] ;

[0180] in, This is the hyperedge weight matrix. The learnable parameter matrix for the l-th layer is used to perform a linear transformation on the course node features. Its parameters are learned through training on course architecture data to enable the course node embedding representation to better characterize the structural and semantic relationships between courses. The learnable parameter matrix is ​​updated layer by layer during the multi-layer hypergraph convolution process, thereby realizing the mapping and optimization of course node features in different representation spaces. The sigmoid function is used as the non-linear activation function. Through the above operations, the course node features are first aggregated from the nodes to the hyperedges, and then propagated back from the hyperedges to the nodes, achieving a unified modeling of prerequisite relationships, content similarity relationships, and collaborative support relationships. After L layers of hypergraph convolution, the final embedding representation matrix of the course nodes in the k-th iteration is obtained:

[0181] ;

[0182] Where the j-th row vector Indicates course The course structure is embedded in the model. This embedding comprehensively reflects the course's own content semantics, prerequisite dependencies with other courses, adjacency relationships with courses of similar content, and group relationships with supporting courses. It can be used for subsequent course structure adjustment actions and fitness evaluation. After obtaining the course structure embedding, the course structure adjustment actions are further defined. Let the course structure optimization strategy in the k-th iteration be:

[0183] ;

[0184] in, This represents the m-th course structure adjustment action, where M represents the total number of actions. Course structure adjustment actions include five categories: adding courses, deleting courses, adjusting course content, adding prerequisites, and deleting prerequisites. To facilitate combinatorial optimization by the genetic algorithm, these actions are uniformly coded. The code for adding a course is as follows: The first 0 indicates a newly added action type. This indicates that a course identifier will be added later. Indicates the semester in which classes are offered. Indicates credit hours; delete course action code is The first '1' indicates the deletion action type. This indicates a course to be deleted; the action code for adjusting course content is... The first digit 2 indicates the type of content adjustment action. Indicates the course, This represents the adjustment amount of the course feature vector. After adjustment, the course feature vector is updated as follows:

[0185] ;

[0186] Priority relationship adds action coding as This indicates the addition of new courses. For the course Prerequisites; the prerequisite for the relationship deletion action is coded as follows: This indicates that the course has been deleted. As a course Prerequisite courses. Therefore, a single chromosome can be represented as a combination of several structural adjustment actions:

[0187] ;

[0188] in, Let L represent the l-th gene in the chromosome, and L represent the chromosome length. After defining and encoding the actions, establish the objective function for optimizing the curriculum structure. The optimization objective is to minimize the overall deviation of the adjusted curriculum system while controlling the adjustment cost, under the premise of satisfying the constraints of the curriculum structure. Let a certain curriculum structure adjustment strategy be... The corresponding adjusted deviation vector is Then the objective function is defined as:

[0189] ;

[0190] The first term represents the sum of squares of the deviations in each target feature dimension after adjustment, used to characterize the degree to which the adjustment scheme improves the overall curriculum system deviation; the second term... Let λ represent the structural adjustment cost; λ is the balance coefficient, used to explain the relationship between the reduction of balance deviation and the adjustment cost. Further, the adjustment cost can be defined as:

[0191] ;

[0192] in, Indicates the number of newly added courses. Indicates the number of courses deleted. Indicates the number of times the course content has been adjusted. Indicates the number of times the relationship has been adjusted. , , , Let be the cost coefficient for the corresponding action. Therefore, the curriculum structure optimization problem is described as:

[0193] ;

[0194] To ensure the feasibility of the optimization results, multiple constraints are set in addition to the objective function. First, the adjusted prior relation must satisfy the acyclic constraint, that is, the adjusted prior relation matrix... The corresponding course dependency graph must be a directed acyclic graph; if a cycle exists, the adjustment strategy is deemed infeasible. Secondly, constraint 1 requires a minimum total credit hour requirement for the curriculum. The credit limit must be met, where, This sets a preset credit limit. Constraint 2 requires a minimum number of courses after adjustment. To meet the course quantity constraints, among which, This represents the maximum number of courses. Constraint 3 requires that adjustments to the total cost must meet budget constraints, where... This is a preset budget cap. Constraint 4 requires that the extent of adjustments to the course content must meet certain requirements. ,in, This sets the upper limit for course content adjustments. Constraint 5 stipulates that newly added or adjusted courses must meet the semester requirements for the semester in which they are offered. This is the maximum semester number allowed by the curriculum system.

[0195] Based on the above objective function and constraints, a genetic algorithm is used for global optimization. First, an initial population is generated. Individuals in the initial population are randomly combined from several course structure adjustment actions, prioritizing those revolving around the set of weak target features identified in S4. and its priority sequence Construct initial candidate solutions to improve search efficiency. Let the population size be... , No. Generation population is denoted as:

[0196] ;

[0197] For any individual with a chromosome in the population Based on the course structure adjustment actions corresponding to its encoding, the course set, course feature matrix, course-feature contribution matrix, and prerequisite relationship matrix are updated to obtain the predicted state of the adjusted course system. Then calculate the corresponding objective function value. To facilitate the selection of superior individuals using genetic algorithms, the fitness function is defined as follows:

[0198] ;

[0199] in, The constraint violation degree is represented by μ, which is the penalty coefficient. The constraint violation degree includes penalty items such as prerequisite cycles, credit limits, total course limits, budget limits, and content adjustment limits, which are used to prevent infeasible solutions from entering the subsequent population.

[0200] To enable the parameters in the genetic algorithm to automatically change with the optimization state, this embodiment further introduces fuzzy logic to dynamically adjust the key parameters of the genetic algorithm. Specifically, the overall deviation intensity, deviation change trend, and population distribution state of the current curriculum system are used as input variables for the fuzzy logic system. The overall deviation intensity is the deviation vector obtained from S4. The 2-norm representation, i.e. The trend of deviation change is represented by the change in deviation between two adjacent iterations, that is: The population distribution is represented by the variance of the current generation's population fitness, denoted as:

[0201] ;

[0202] in, Indicates the first Mean fitness of the population. (Based on crossover probability) Probability of mutation The constraint penalty coefficient μ is used as the output variable of the fuzzy logic system, and parameter adaptive adjustment is achieved through fuzzy rules. When the overall deviation is large and the decreasing trend is not obvious, the mutation probability is increased. To enhance the ability to explore the solution space; to reduce the probability of mutation when the overall deviation is small and tends to be stable. To improve convergence stability; when the population fitness variance is small, increase the crossover probability. To avoid premature population concentration, the penalty coefficient μ is increased when the proportion of infeasible solutions is high or the constraint violation rate is large, thereby enhancing the suppression of infeasible solutions. Through the above fuzzy logic adjustment process, the genetic algorithm parameters no longer depend on fixed preset values, but can automatically change according to the current optimization state, thus improving the adaptability and stability of the curriculum system structure optimization solution. In some implementations, the fuzzy logic parameter tuning mechanism is embedded in the optimization process as a preferred method for updating genetic algorithm parameters.

[0203] After parameter settings are completed, the genetic algorithm sequentially performs selection, crossover, mutation, and elite retention operations. Selection selects individuals with high fitness from the current population to enter the mating pool; crossover swaps parts of the parent chromosomes to generate new offspring; mutation randomly changes the course identifier, action type, prerequisite relationship, or adjusts parameters in a gene to maintain population diversity; elite retention directly preserves the highest-fitting individuals from the current population to the next generation to avoid losing the optimal solution. The genetic algorithm repeats this process until stopping conditions are met. Stopping conditions include reaching the maximum number of iterations, the optimal fitness no longer significantly improving for several consecutive generations, or the objective function value falling below a preset threshold. Finally, the algorithm outputs the individual with the highest fitness. And decode it into the optimal course structure optimization strategy under the k-th iteration:

[0204] ;

[0205] in, This represents the optimal course structure optimization strategy in the k-th iteration.

[0206] Through the above process, this step ultimately outputs the course hypergraph. Course structure embedding representation matrix Optimal Course Structure Optimization Strategies Adjusted Predicted Curriculum System Status and the adjusted prediction bias vector Among them, the optimal course structure optimization strategy It will serve as the direct input for updating the course architecture state in S6.

[0207] This step transforms the course architecture optimization problem driven by deviation vectors into a combinatorial optimization problem by constructing a course hypergraph and extracting the structural embedding representations of course nodes under multiple constraints. Finally, a genetic algorithm is used to obtain the optimal adjustment strategy, thus realizing the optimization of the course architecture state model. Sum of deviation vectors To course structure optimization strategy The mapping provides executable optimization results for subsequent course system architecture state updates and closed-loop iterations.

[0208] S6. Update the course set, course feature matrix, course-feature contribution matrix, and prerequisite relation matrix according to the optimal course structure adjustment strategy to form a new course structure system state.

[0209] The method for obtaining the new course structure state includes: updating the course set, course feature matrix, course-feature contribution matrix, and prerequisite relation matrix according to the optimal course structure adjustment strategy; performing semantic vectorization and calculating contribution values ​​for newly added courses, removing corresponding matrix rows and columns for deleted courses, and synchronously updating the prerequisite relation matrix; performing consistency checks on the number of courses, total course credits, semester offered, and budget constraints after the update; and ensuring that the updated new course structure state meets the requirements of computability, implementability, and optimizability.

[0210] In this embodiment, in S6, the optimal course structure optimization strategy obtained in S5 is used. As input, the current course architecture state model constructed using S2 As the initial state, the course set, course feature vectors, course-feature contribution matrix, and course prerequisite relation matrix are updated to obtain the course system architecture state model under the (k+1)th iteration. .

[0211] In S5, the optimal course structure optimization strategy is expressed as: , among which, among which, Let M represent the m-th optimal course structure adjustment action, where M represents the total number of actions included in this round of optimization. According to the aforementioned action definition, course structure adjustment actions include adding courses, deleting courses, adjusting course content, adding prerequisites, and deleting prerequisites. This step involves updating each component of the current course structure system state model item by item based on the action type and its parameters.

[0212] First, the course set is updated. In S2, the course set in the k-th iteration is represented as... Optimization strategies for the optimal course structure Each structural adjustment movement within the lesson is analyzed, detailing its movement type and the target audience of the course. If a certain movement... To add new course actions, namely:

[0213] ;

[0214] New courses will be added. The course set includes the corresponding credits, semester offered, course category, course description, and other attributes, forming a new course set. If a certain action To delete a course action, i.e. Then the course Remove from the current course set and create a new course set to be deleted. Therefore, the updated course set is defined as follows: .

[0215] in, This represents the updated course set after the (k+1)th iteration. This update expression indicates that the course set consists of courses retained from the original set and newly added courses, while deleted courses no longer participate in the new course structure state representation. After updating the course set, the course feature vector set and course feature matrix are further updated. For retained courses in the course set, if the optimal course structure optimization strategy does not include corresponding course content adjustment actions, their course feature vectors remain unchanged; if a certain course... The corresponding course content adjustments exist, namely... Then, according to the adjustment rules defined in S5, the course... The course feature vector is updated to obtain ,in, Indicates course The course feature vector in the k-th iteration, Indicates its characteristic adjustment amount, This represents the updated course feature vector. For newly added courses... Then, based on the set of course text descriptions corresponding to that course. The semantic vectorization process is performed again using the same pre-trained language model Ψ as S2 to obtain the course feature vectors of the new courses. ,in, Let represent the feature vector of the newly added course in the (k+1)th iteration. Therefore, the updated set of course feature vectors, described in matrix form, yields the course feature matrix in the (k+1)th iteration:

[0216] ;

[0217] in, .

[0218] After updating the course set and course feature matrix, the course-feature contribution matrix is ​​further updated. In step 2, the course-feature contribution matrix is ​​represented as follows: .in, Indicates course For target feature terms The normalized contribution values. Since the course set changes and course feature vectors may be adjusted, the updated course-feature contribution matrix needs to be recalculated. For the updated course set... Any course and arbitrary target feature terms Based on the updated course feature vector The weighted target feature vector constructed with S1 Recalculate the original contribution value:

[0219] ;

[0220] Furthermore, in this embodiment, the original contribution values ​​are processed using the same normalization method as in S2. Preferably, Softmax normalization is used to obtain the course. and arbitrary target feature terms Updated contribution value:

[0221] ;

[0222] For any course ,have: Therefore, by combining all the updated contribution values, the course-feature contribution matrix for the (k+1)th iteration is constructed. The course-feature contribution matrix fully reflects the contribution distribution of each course to each target feature in the updated curriculum system, and serves as the basis for subsequent system output feature vector calculation and deviation analysis.

[0223] After updating the course-feature contribution matrix, the course prerequisite relation matrix is ​​further updated. In S2, the course prerequisite relation matrix under the k-th iteration is represented as follows: ,in, Indicates course It is a course prerequisite courses This indicates that there is no prerequisite dependency between the two. Since the optimal course structure optimization strategy may include adding courses, deleting courses, adding prerequisites, and deleting prerequisites, the prerequisite relationship matrix needs to be updated synchronously.

[0224] For new courses Add corresponding rows and columns to the course prerequisite relationship matrix. Initially, the prerequisite relationships between newly added courses and other courses can be initialized to zero, meaning that newly added courses do not have prerequisite or successor dependencies by default. Subsequent assignments are only made based on the prerequisite relationship adjustment actions in the optimal course structure optimization strategy. For deleting courses... If so, then remove the corresponding row and column in the prior relation matrix. Adding actions to prior relations... Then, in the updated prior relation matrix, the elements will be... Set to 1; for the prior relationship deletion action Then the corresponding elements in the updated prior relation matrix will be... Set it to 0. Thus, we obtain the course prerequisite relation matrix for the (k+1)th iteration. .

[0225] After updating the prerequisite relation matrix, to ensure that the updated course system structure state model meets the implementability constraints, it is necessary to perform an acyclicity check on the updated prerequisite relation matrix. Specifically, this involves... The corresponding course dependencies are represented as a directed graph, which is then subjected to a topological sort. If the topological sort can traverse all course nodes, the updated course dependency graph is a directed acyclic graph, and the prerequisite relation matrix satisfies the structural constraints. If it cannot traverse all course nodes, the updated prerequisite relations contain cycles, and the current update result does not satisfy the structural constraints. For cases with cycles, a rollback mechanism can be used to undo actions that add to the prerequisite relations that cause the cycle, or in subsequent optimization iterations, relevant penalty terms can be increased to suppress such infeasible solutions.

[0226] Furthermore, after updating the course set, course feature matrix, course-feature contribution matrix, and course prerequisite relation matrix, a consistency check is performed on the updated course system state. The consistency check includes checks on the number of courses, total credits, the range of semesters offered, and budget constraints. Let the total credits of the course system in the (k+1)th iteration be... And should satisfy Let the total number of courses in the (k+1)th iteration be... Then it should satisfy For semesters with newly added or adjusted courses, the following conditions should be met: .

[0227] If all the above consistency checks pass, the updated course architecture state satisfies the structural and implementation constraints and can be used as the initial state for the next iteration. If they fail, the relevant update actions are corrected or rolled back to ensure that the course architecture state model remains computable, implementable, and optimizable. Therefore, the updated course architecture state model for the (k+1)th iteration is represented as follows: ,in, This represents the updated set of courses. This represents the updated course-feature contribution matrix. This represents the updated course prerequisite relation matrix. Course system architecture state model. It is the k-th structural optimization strategy The actual update results after the action are also the basis for recalculating the system output feature vector, constructing a new round of deviation vectors, and starting the next round of course system structure optimization in subsequent steps.

[0228] This step optimizes the optimal curriculum structure obtained by S5. This involves specific updates to the course set, course feature matrix, course-feature contribution matrix, and course prerequisite relationship matrix, thereby realizing the transformation of the course system structure state from... arrive The evolution of the curriculum structure is thus completed. This completes the "execution update" step in the curriculum structure optimization process, enabling the optimization results obtained by S5 to truly apply to the curriculum structure itself and provide a new state basis for subsequent closed-loop iterations.

[0229] S7. Based on the new curriculum system structure state, repeat the steps of calculating the system output feature vector to updating the state, and continuously optimize the curriculum system structure through closed-loop iteration until the convergence condition is met.

[0230] S7 includes: using the new curriculum structure state as the input for the next iteration, repeating S3 to S6; triggering conditions include fixed training cycles and dynamic conditions, with dynamic conditions including: significant changes in external demand, overall deviation exceeding the threshold, and significant changes in curriculum structure; introducing a smooth update mechanism to distribute large-scale adjustments across multiple iterations to avoid drastic fluctuations in the curriculum structure; iterating until the overall deviation converges to the preset threshold, or the weak target feature stabilizes; after convergence, entering monitoring mode to continuously observe the curriculum system operation data and external demand, and re-triggering closed-loop optimization if the deviation expands again.

[0231] In this embodiment, the course architecture state model updated in S6 is used. As the initial state for the next round of optimization, S3 to S6 are repeated, combining new course operation data, external demand data, and information on changes in course structure, thus forming a closed-loop iterative optimization process for the course system structure. The purpose of this step is to enable the course system to continuously correct its structural state as external demands change and internal structure evolves through a continuously running dynamic closed loop, gradually approaching the target characteristic state, and entering monitoring mode after meeting the convergence conditions.

[0232] First, define the state transit relationships in the closed-loop iterative process of the curriculum system. Let the state model of the curriculum system structure obtained after the k-th iteration be... Based on this state model, the system output feature vector at the k-th iteration can be calculated using S3. Furthermore, the corresponding deviation vector can be constructed using S4. Then, the optimization strategy for the course structure can be obtained through S5. Finally, the course architecture state model is updated using S6 to obtain the course architecture state model for the next iteration. Therefore, the closed-loop iterative optimization process of the curriculum system can be represented as:

[0233] ;

[0234] This state transit relationship indicates that each iteration of the curriculum system includes five stages: "state input, output feature calculation, deviation construction, optimization solution, and state update". Furthermore, the updated state of the curriculum system in each round serves as the input state for the next round of optimization, thus forming a closed-loop feedback mechanism.

[0235] To ensure the continuous operation of the closed-loop iterative optimization process, it is necessary to further define the triggering conditions for each iteration. In this implementation, the curriculum structure optimization can be triggered either by a fixed cycle or by dynamic conditions. Fixed cycle triggering means that a preset course operation cycle is used as the iteration unit. After each course operation cycle ends, relevant course operation data is collected again, and a new round of curriculum structure optimization is started. The course operation cycle can correspond to a semester, an academic year, a course group operation cycle, or other preset cycles. Dynamic condition triggering means that when the curriculum system undergoes significant structural changes or target deviations during operation, a new round of iterative optimization is started directly without waiting for the fixed cycle to end. Specifically, the curriculum structure optimization process can be re-executed when any of the following conditions are met: First, the external demand data changes significantly, i.e., the demand text, skill tags, or topic task set on which the target feature set is based in S1 is added, deleted, or undergoes significant semantic shift; Second, the deviation vector obtained in S4... The overall deviation intensity exceeds the preset deviation threshold; third, the course set, course-feature contribution matrix, or prerequisite relationship matrix in the S6-updated course system structure state model undergo significant changes; fourth, the addition of courses, deletion of courses, or reconstruction of prerequisite relationships in the course system reaches the preset change threshold. By combining fixed-period triggering with dynamic condition triggering, the closed-loop iterative optimization of the course system can have both stable periodicity and rapid response capability to abnormal changes and structural mutations.

[0236] After triggering a new round of optimization, it is necessary to re-collect and reorganize the data input required for the next iteration. Specifically, for the (k+1)th iteration, S1 is first re-executed based on the new external demand data to update the target feature set and the weighted target feature matrix. If the external demand data remains unchanged, the weighted target feature matrix from the previous iteration can be used directly. If the external demand data changes significantly, the text cleaning, feature extraction, semantic vectorization, and target feature weight update processes need to be re-executed to obtain a new weighted target feature matrix. Furthermore, the updated course structure state model obtained in S6... The target feature state and the course system structure state are used as the initial state input for S2 in the (k+1)th iteration, and S3 to S6 are executed accordingly. Thus, in each iteration, the target feature state and the course system structure state can be updated synchronously, enabling the closed-loop optimization process to have continuous adaptive capability.

[0237] To quantify the convergence trend of the curriculum system during the closed-loop optimization process, it is necessary to further define the iterative convergence criterion. Preferably, the deviation vector is used. The overall deviation strength is used as the basis for the convergence criterion. Let the norm of the deviation vector in the k-th iteration be:

[0238] ;

[0239] in, This represents the overall deviation strength of the curriculum system in the k-th iteration. Furthermore, the deviation reduction rate between two adjacent iterations is defined as:

[0240] ;

[0241] in, This represents the relative decrease in the overall deviation of the curriculum system from the k-th iteration to the (k+1)-th iteration. A larger value indicates that the optimization of the current curriculum structure has a significant improvement effect; when... When the value is relatively small, it indicates that the overall deviation of the curriculum system has decreased to a limited extent, and the structural state tends to be stable.

[0242] Based on the aforementioned deviation reduction rate, the convergence condition for closed-loop optimization of the curriculum system is further defined. Preferably, when continuous All iterations satisfy Where ε represents the preset convergence threshold, The threshold for the number of consecutive iterations indicates that the curriculum structure optimization process has entered a convergent state. Convergence means that after several rounds of structure updates, the overall deviation of the curriculum system in the target feature dimension has been reduced to a small level, or the improvement obtained from further iterations is less than the preset threshold. At this point, the curriculum structure has reached a relatively stable optimization result.

[0243] In addition to judging convergence based on the overall deviation intensity and deviation reduction rate, further auxiliary judgment can be made by combining the local deviation distribution. Specifically, for the set of weak target feature terms defined in S4: If, in several consecutive iterations, the size of the set of weak target features no longer increases, and the deviation values ​​corresponding to each target feature in the set show a decreasing trend, it indicates that the structural adjustment of the curriculum system in the key weak target dimensions has achieved a stable effect. If the set of weak target features continues to expand, or the deviation values ​​of the core target features continue to rise, it indicates that the current curriculum system has not yet reached a stable state and closed-loop optimization needs to continue. By combining the overall deviation criterion with the local weak feature criterion, the convergence and stability of the curriculum system's closed-loop optimization process can be evaluated more comprehensively.

[0244] After determining that the curriculum system has entered a convergence state, it enters monitoring mode. Specifically, in monitoring mode, only the operational data of the curriculum system and external demand data continue to be collected, and changes in target features and output features of the curriculum system are continuously observed. If any of the following situations occur in monitoring mode, the closed-loop optimization process will be restarted: First, new external demand data causes a significant change in the target feature set; second, the deviation between the current output feature vector of the curriculum system and the target feature state widens again; third, the internal structure of the curriculum system undergoes significant changes, such as the addition of new courses, reorganization of course groups, or large-scale rearrangement of prerequisite relationships; fourth, certain key target feature items are detected to have re-entered a weak state.

[0245] To mitigate the structural fluctuations caused by frequent curriculum system adjustments, this implementation also introduces a smooth update mechanism. Specifically, for the optimal curriculum system structure optimization strategy obtained in S5... If the structural adjustment strategy involves too many actions, or if the adjustment range of a certain type of action is too large, then the entire set of actions will not be executed in a single iteration. Instead, the structural adjustment actions will be distributed across several consecutive iterations based on a preset smoothing coefficient. Let M be the number of actions required in a given optimal adjustment strategy, and the actual number of actions executed after the k-th iteration be... Then we have: ,in, Indicates the smooth update coefficient. This indicates the floor function. When When the value is small, it indicates that only part of the structural adjustment actions are performed in this round, and the remaining actions are postponed to be completed gradually in subsequent iterations; when A value close to 1 indicates that a relatively large proportion of adjustments will be made in this round. By introducing a smooth update mechanism, we can avoid overly drastic structural changes in the curriculum system during a single iteration, thereby improving the stability and feasibility of the curriculum system's evolution.

[0246] Furthermore, to measure the overall evolutionary effect in the closed-loop optimization process, the cumulative improvement of the curriculum structure optimization can be defined. Let the initial deviation strength in the 0th iteration be... The bias intensity in the k-th iteration is Then, the cumulative improvement of the curriculum system relative to the initial state at the k-th iteration is defined as:

[0247] ;

[0248] in, This represents the cumulative deviation improvement of the curriculum system from its initial state to the k-th iteration state. A larger value indicates a more significant overall improvement in the curriculum system relative to its initial state during the closed-loop optimization process. The cumulative improvement allows for a global assessment of the long-term effects of curriculum system optimization. This invention enables the curriculum system construction method proposed in this paper to achieve a fundamental shift from experience-based design to data-driven optimization through multiple iterations and gradual curriculum system refinement, providing a practical and reproducible technical solution for precise talent cultivation in the era of artificial intelligence.

[0249] This step defines the state propagation relationships, triggering conditions, convergence criteria, monitoring mechanisms, and smooth update mechanisms for the closed-loop iterative optimization of the curriculum system. It organizes S3 to S6 into a repeatable, dynamic, closed-loop process, enabling the curriculum system to continuously correct its structural state through multiple iterations and gradually approach the target characteristic state constructed in step 1. This completes the closed-loop iterative process in the curriculum system structure optimization method, ensuring the curriculum system's continuous optimization capability, dynamic adaptability, and structural stability under conditions of changing external demands and internal structural evolution.

[0250] Example 2

[0251] In this embodiment, in the example of curriculum system optimization for the Measurement and Control Technology and Instruments major, the 11 graduation requirements proposed in the training program are first refined into 33 specific indicator points based on job requirements. These 33 indicator points include signal and system theory knowledge, measurement and control system design ability, data analysis ability, experimental operation ability, innovative practice ability, teamwork ability, professional ethics, engineering ethics, and continuous learning ability. The weight of each indicator point reflects its importance in job requirements and industry standards. To this end, statistical analysis is performed on external requirement data to calculate the frequency of each indicator point in job description texts, skill tags, task descriptions, and industry documents. The weight parameters of the indicator points are determined based on the frequency of occurrence, and all weights are normalized so that their sum is 1, forming a 33-dimensional weight vector. Visualizing this weight vector as a bar chart can clearly show the relative importance of the 33 indicator points in job requirements.

[0252] Next, the 33 indicator points were semantically vectorized using the Sentence-BERT base model all-MiniLM-L6-v2, mapping each indicator point to a fixed-dimensional semantic vector. This resulted in a 33 × 384 target feature matrix Y. The vertical axis of the matrix represents the 33 indicator points, and the horizontal axis represents the 384 dimensions of the semantic vector. Matrix elements reflect the feature representation strength of each indicator point in the semantic space. To visually represent the target feature matrix, a heatmap was used. The color intensity indicates the size of the matrix elements; a darker color indicates a larger semantic feature value for that indicator point in that dimension, while a lighter color indicates a smaller contribution. The heatmap clearly shows the distribution of each indicator point in the semantic space, aiding in subsequent contribution analysis and bias calculation, and providing a quantitative basis for optimization iterations.

[0253] Secondly, based on the talent training program and corresponding course syllabus of the 2023 Measurement and Control Engineering major, a model was built for the pre-optimization curriculum system. Sentence-BERT was used to map the syllabi of 72 courses into a 72×384 matrix. Combining the courses and the support matrix of graduate feedback on graduation requirements, a 72×33 course-indicator point contribution matrix was constructed, and a 72×72 prerequisite constraint matrix was built based on the course syllabus. To more intuitively understand the contribution matrix of courses to indicator points, a heatmap representation of the contribution matrix is ​​given. In the Measurement and Control Engineering major, apart from the general education courses in the first semester of freshman year (such as English, Physical Education, Ideological and Political Education, and Advanced Mathematics), most professional courses have strict topological dependencies. For example, the Measurement and Control System section follows a sequence from circuits to analog electronics, digital electronics, then to microcomputer principles and microcontrollers, measurement and control circuits, and measurement and control systems. A prerequisite matrix was constructed, and a heatmap representation of the prerequisite matrix is ​​given to illustrate its description. As can be clearly seen from the diagram, the courses are arranged in ascending order from top to bottom and from left to right according to the semester. When time constraints are introduced, the prerequisite matrix forms an upper triangular representation.

[0254] To fully verify the scientific validity and effectiveness of the curriculum structure optimization method based on hypergraph neural networks and genetic algorithms, a typical external industry demand shock scenario was set up: with the popularization of industrial IoT and artificial intelligence technologies, the industry has put forward more stringent requirements for the "AI tool application ability," "multidisciplinary team collaboration," and "comprehensive engineering project management" of measurement and control graduates. The system first converts these new industry demand texts into target feature vectors and then performs a deep comparison with the system output features of the current training program. The diagnostic results accurately exposed the weaknesses of the original program—specifically, in the "modern tool application (…)" aspect. 16 indicators (5.2) and "Multidisciplinary team collaboration ( 23 Indicator Points 8.2), "Engineering Project Management ( 27 Indicator Points 10.1) and Lifelong Learning ( There is a significant lack of support in aspects such as the 32 indicators (11.2), creating a gap in capacity building.

[0255] To address this, a hypergraph neural network was used to deeply mine and extract the embedded features of the courses within the overall architecture. These features were then incorporated into the optimization framework of a genetic algorithm. Under the strict constraints of ensuring that the total class hours / credits do not exceed the limit and that the prerequisite order of core courses is not disrupted, the genetic algorithm iterates through the population to continuously find the optimal combination of course hour allocation and support weights. Experiments showed that the algorithm's fitness curve converged stably after multiple iterations, demonstrating the model's efficient optimization capability.

[0256] The algorithm ultimately outputs a globally optimal curriculum adjustment strategy. The solution suggests appropriately reducing redundant theoretical hours in some traditional hardware foundation courses and precisely transferring them to key courses such as "Introduction to Artificial Intelligence" and "Comprehensive Design of Measurement and Control Systems." It also mandates the introduction of interdisciplinary teamwork and project management assessments in practical components. Quantitative comparative analysis clearly shows that the algorithm perfectly fills the gaps in previously weak indicators, significantly reducing the target deviation values ​​of typical weak indicators to extremely low levels, fully demonstrating the algorithm's high precision and global perspective in the dynamic optimization of complex curriculum systems.

[0257] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for optimizing the structure of a curriculum system based on hypergraph neural networks and genetic algorithms, characterized in that, Includes the following steps: S1. Collect course data and external demand data, preprocess and semantically vectorize them, and construct a weighted target feature matrix; S2. Based on the semantic vectors of the course set and course texts, construct the course feature matrix, and calculate the course-feature contribution matrix and the course prerequisite relationship matrix to establish a course system architecture state model; S3. Calculate the course aggregation weights based on the course-feature contribution matrix and the course prerequisite relationship matrix, and then obtain the system output feature vector that characterizes the comprehensive response of the curriculum system; S4. Based on the system output feature vector and the weighted target feature matrix, calculate the feature deviation vector, and identify weak target feature items and optimize priorities based on the deviation threshold; S5. Based on the course system structure state model and the weak target feature items, construct a course hypergraph that includes prerequisite dependencies, content similarity and collaborative support hyperedges, use a hypergraph neural network to extract the course node structure embedding representation, and solve it through a genetic algorithm under multiple constraints to generate the optimal course structure adjustment strategy; S6. Update the course set, the course feature matrix, the course-feature contribution matrix, and the prerequisite relationship matrix according to the optimal course structure adjustment strategy to form a new course structure state; S7. Based on the new curriculum system architecture state, repeat the steps of calculating the system output feature vector to updating the state, and continuously optimize the curriculum system architecture through closed-loop iteration until the convergence condition is met.

2. The curriculum structure optimization method based on hypergraph neural network and genetic algorithm according to claim 1, characterized in that, The method for constructing the weighted target feature matrix includes: Collect the course data and the external demand data. The course data includes: course text data, course attribute data, and course relationship data. The external demand data includes: job description text, skill tag text, task description text, and industry documents. The course data and the external demand data are subjected to text cleaning, format standardization, word segmentation, terminology normalization, synonym merging, stop word filtering, and abnormal data removal to obtain the processed data. Extract the target feature set from the processed data and use the Sentence-BERT pre-trained language model for semantic vectorization to obtain the semantic vector representation of each feature item; Construct a target feature matrix and assign weights to the target feature items based on their frequency of occurrence or importance score in the external demand data to obtain the weighted target feature matrix.

3. The curriculum structure optimization method based on hypergraph neural network and genetic algorithm according to claim 2, characterized in that, The methods for constructing the state model of the course architecture include: Construct the course set and its corresponding attribute parameters, which include: course name, course number, course category, course module, course credits, total course hours, semester offered, and syllabus text; The course text data is preprocessed and semantically vectorized to obtain the course feature matrix; The course-feature contribution matrix is ​​calculated by the cosine similarity between the course feature vectors and the target feature vectors of the course feature matrix, and the comparability of contribution values ​​is ensured by Softmax normalization. Extract the course prerequisite relationship information and construct the course prerequisite relationship matrix, where a matrix element of 1 indicates a prerequisite dependency and a matrix element of 0 indicates no dependency; Based on the course set, the course feature matrix, the course-feature contribution matrix, and the course prerequisite relation matrix, construct the course system architecture state model for the k-th iteration.

4. The curriculum structure optimization method based on hypergraph neural network and genetic algorithm according to claim 3, characterized in that, The methods for obtaining the system output feature vector include: The course-feature contribution matrix is ​​weighted and aggregated along the course dimension, and the course aggregation weight is calculated by combining the course credits and hierarchical attributes. If a course is a core course, it will be given a higher weight; if it is a supporting course, it will be given a lower weight. By combining the course prerequisite relation matrix with the course aggregation weights, the system output feature vector is obtained, which is used to characterize the comprehensive response of the curriculum system in each target feature dimension. The system output feature vector is subjected to interval constraint processing to ensure that the output of each target feature dimension is within a reasonable range.

5. The curriculum structure optimization method based on hypergraph neural network and genetic algorithm according to claim 4, characterized in that, The method for identifying the weak target features includes: The system output feature vector is mapped to the same semantic space as the target feature matrix, and the feature deviation matrix is ​​calculated. The feature deviation matrix is ​​transformed into a scalar deviation vector, where each component represents the deviation intensity of the corresponding target feature term; Based on the deviation intensity, the weak target feature items are filtered according to the deviation threshold, and a sequence of optimization priorities is generated according to the deviation magnitude; The mean and variance of the deviation, as well as the change in deviation between two adjacent iterations, are calculated to assess the overall deviation level and local imbalance of the curriculum system, and to provide a basis for dynamic adjustment for subsequent optimization.

6. The curriculum structure optimization method based on hypergraph neural network and genetic algorithm according to claim 3, characterized in that, The method for generating the optimal curriculum structure adjustment strategy includes: The curriculum system is represented as the curriculum hypergraph, which includes a set of course nodes and three types of hyperedges, namely: prerequisite dependency hyperedges, content similarity hyperedges, and collaborative support hyperedges. The prerequisite dependency hyperedge is constructed from the course prerequisite relation matrix, the content similarity hyperedge is constructed when the semantic similarity of the course feature vectors exceeds a threshold, and the collaborative support hyperedge is constructed from courses that contribute highly to the same weak target feature item. Based on the course hypergraph, a node degree matrix and a hyperedge degree matrix are constructed, and then input into a multilayer hypergraph convolutional neural network to obtain the course node embedding representation; Define course structure adjustment actions: adding courses, deleting courses, adjusting course content, adding and deleting prerequisite relationships, and encode the actions to form a chromosome representation; The objective functions are set as minimizing deviation and minimizing adjustment cost, and a balance coefficient is introduced; Set constraints: prerequisite relationships must be acyclic, total course credits, number of courses, budget, and content adjustment limits; Based on the course node embedding representation, a genetic algorithm is used to perform selection, crossover, mutation, and elite retention operations, and fuzzy logic is combined to adaptively adjust the crossover probability, mutation probability, and penalty coefficient to generate the optimal course structure adjustment strategy.

7. The curriculum structure optimization method based on hypergraph neural network and genetic algorithm according to claim 6, characterized in that, The fitness function of the genetic algorithm includes constraint violation degree, which is used to penalize infeasible solutions with prerequisite cycles, credit limits, course limits, budget limits, and content adjustment limits. The inputs to the fuzzy logic include the overall deviation intensity, the deviation change rate, and the population distribution state. By adjusting the crossover probability, mutation probability, and penalty coefficient through rules, the parameters are adaptively optimized.

8. The curriculum structure optimization method based on hypergraph neural network and genetic algorithm according to claim 6, characterized in that, The methods for obtaining the new curriculum architecture state include: The course set, course feature matrix, course-feature contribution matrix, and prerequisite relation matrix are updated according to the optimal course structure adjustment strategy. The newly added courses are semantically vectorized and their contribution values ​​are calculated. The corresponding matrix rows and columns are removed for the deleted courses. The prerequisite relationship matrix is ​​updated synchronously. After the update, a consistency check is performed on the number of courses, the total credit hours for the courses, the semester in which the courses are offered, and the budget constraints. Ensure that the updated new curriculum architecture meets the requirements of computability, implementability, and optimizability.

9. The curriculum structure optimization method based on hypergraph neural network and genetic algorithm according to claim 1, characterized in that, S7 includes: Use the new course system architecture state as the input for the next iteration, and repeat S3 to S6. Triggering conditions include fixed training cycles and dynamic conditions, wherein the dynamic conditions include: significant changes in external demand, overall deviation exceeding a threshold, and significant changes in curriculum structure; A smooth update mechanism is introduced to spread large-scale adjustments across multiple iterations to avoid drastic fluctuations in the course structure. Iterate until the overall deviation converges to a preset threshold, or the weak target feature term stabilizes; After convergence, the system enters monitoring mode to continuously observe the operational data of the curriculum system and external demands. If the deviation expands again, closed-loop optimization is triggered again.

Images