Method and device for recommending exercises combining hypergraph knowledge modeling and automatic hint optimization

Abstract

The application belongs to the field of cross-course exercise recommendation, and relates to an exercise recommendation method and device combining hypergraph knowledge modeling and automatic prompt optimization, wherein a cross-course heterogeneous hypergraph and a knowledge point final migration correlation matrix from a source course to a target course are obtained according to exercises, knowledge points, an exercise and knowledge point correlation matrix and a knowledge point prerequisite dependency relationship of the source course and the target course; a first knowledge state vector is obtained according to the user's answer record in each course and the knowledge point final migration correlation matrix; an exercise recommendation score is determined according to the first knowledge state vector, the target course exercise and knowledge point correlation matrix and the knowledge point prerequisite dependency relationship; and the preset number of exercises with the highest recommendation score are input into a large language model with a double-layer prompt word optimization mechanism to obtain a recommendation result. The application uses cross-course heterogeneous hypergraph knowledge modeling to obtain more accurate user knowledge states, and uses automatic prompt optimization to mine user preferences and make recommendations, thereby improving the adaptability of the exercise recommendation result.

Description

Technical Field

[0001] This invention belongs to the field of cross-course exercise recommendation, specifically involving an exercise recommendation method and apparatus that combines hypergraph knowledge modeling and automatic prompt optimization. Background Technology

[0002] With the widespread application of online intelligent tutoring platforms and the continuous expansion of online learning scenarios, massive amounts of historical course learning data and exercise resources have been rapidly accumulated. In the early stages of new courses, faced with a large number of exercise resources of varying quality, learners often lack the ability to comprehensively discern the difficulty, knowledge matching, and content quality of exercises. This can easily lead to problems such as blindly doing exercises or repeating exercises, making it difficult to quickly and accurately select exercises that meet their own learning needs, resulting in low practice efficiency.

[0003] For example, patent CN117743699A discloses a method and system for recommending exercises based on DKT and Thompson's sampling algorithm. It establishes a set of similar users for new users based on course foundation, professional relevance, learning ability, and knowledge point interest, and performs initial recommendations using Thompson's sampling algorithm and user-based collaborative filtering. After accumulating a certain number of practice records, it uses a deep knowledge tracking model to predict the accuracy of answers and makes recommendations based on knowledge point relationships. Patent CN120804307A discloses a method and device for personalized exercise recommendation based on a knowledge graph-driven large model. It determines the level of knowledge point mastery based on students' historical answer performance, constructs a personal profile knowledge graph, and combines a course exercise database, knowledge point relationships, and a large language model to filter and recommend personalized exercises. However, these methods do not fully consider the complex knowledge relationships between different courses and neglect the student's cognitive information contained in past course learning records.

[0004] It is evident that existing exercise recommendation methods do not fully utilize users' existing course learning records, nor do they adequately consider the complex knowledge relationships between and within courses. Furthermore, when modeling the interaction between a specific user and exercises, they fail to adequately consider the impact of the user's performance on knowledge-related exercises. In addition, existing exercise recommendation methods based on large language models typically use pre-designed prompts to generate recommendations or iteratively adjust prompts based on recommendation performance. However, the optimization strategies used to guide these prompt adjustments are usually pre-defined and do not update themselves based on the actual optimization results. This limits the optimization of prompts to the pre-defined strategy content, thus affecting the suitability of the recommended exercises. Therefore, a new exercise recommendation method is needed to accurately recommend exercises that match learners' individual learning needs. Summary of the Invention

[0005] To address the aforementioned technical issues, this application provides a method and apparatus for recommending exercises that combines hypergraph knowledge modeling and automatic prompting optimization. Based on cross-course hypergraph knowledge modeling, it obtains a more accurate understanding of users' knowledge points in the early stages of the target course. Based on the automatic prompting optimization mechanism, it provides users with exercise resources that better match their cognitive level and exercise preferences.

[0006] The first aspect of this application provides a problem recommendation method that combines hypergraph knowledge modeling and automatic suggestion optimization, including: Obtain the exercises, knowledge points, exercise-knowledge point association matrix, knowledge point prerequisites and dependencies, user's answer records in each course, user representation, exercise representation of each exercise in each course, and user's initial course knowledge state vector in the source course for each course. The multiple courses include the source course and the target course. For multiple courses, the knowledge points and exercises of multiple courses are used as nodes. Based on the association matrix between each exercise and knowledge point and the prerequisite dependency relationship of each knowledge point, the prerequisite dependency hyperedge relationship and knowledge point similarity hyperedge relationship between knowledge points in each course are determined. The association hyperedge relationship between each exercise and knowledge point in each course is also determined. The cross-course prerequisite knowledge point hyperedge relationship between the knowledge points in the source course and the knowledge points in the target course is determined. Based on each hyperedge relationship, the cross-course heterogeneous hypergraph association matrix and hyperedge weight matrix are obtained. Based on the cross-course heterogeneous hypergraph association matrix and hyperedge weight matrix, the node degree matrix is ​​obtained. Based on the cross-course heterogeneous hypergraph association matrix, the hyperedge degree matrix is ​​obtained. Finally, based on each cross-course prerequisite knowledge point hyperedge relationship, the initial transfer association matrix between the knowledge points of the source course and the knowledge points of the target course is determined. Based on the cross-course heterogeneous hypergraph association matrix, hyperedge weight matrix, node degree matrix, and hyperedge degree matrix, the initial feature representation of each node is iteratively updated in multiple rounds to obtain the higher-order node representation of each node. Based on the initial transfer association matrix, the association weights of the higher-order node representations corresponding to each knowledge point are corrected to obtain the final transfer association weights of knowledge points from each source course to the target course. For each course, based on the course's exercise-knowledge point association matrix and the course's answer records, the historical interaction context representation of the user with each exercise is determined, and the user representation, the exercise representation, and the historical interaction context representation are combined to obtain the user's historical interaction enhancement discriminability vector for each exercise. Based on the discrimination vector enhanced by the user's historical interaction with each exercise in the source course, the difficulty vector of each exercise, the association matrix between exercises and knowledge points, and the answer records, the initial course knowledge state vector of the source course is optimized, and the post-training course knowledge state vector of the source course is obtained. Based on the post-training course knowledge state vector and the final transfer association weight of knowledge points from each source course to the target course, the initial course knowledge state vector of the target course is determined. The initial course knowledge state vector of the target course is obtained by optimizing the discrimination vector of the user's historical interaction with each exercise in the target course, the difficulty vector of each exercise in the target course, the association matrix between the exercises and knowledge points in the target course, and the user's answer record in the target course. Based on the association matrix between exercises and knowledge points in the target course, the first knowledge state vector, and the prerequisite dependencies of knowledge points, the recommendation score of each exercise in the target course relative to the user is determined. Then, based on the large language model, a preset number of exercises with the highest recommendation scores and multiple task prompt words are processed to obtain the exercise recommendation results for different target tasks. Among them, the task prompt words are obtained by iterative optimization based on the user's interaction with the historical exercise recommendation results.

[0007] In some embodiments of this application, the process of processing a preset number of exercises with the highest recommendation scores and multiple task prompts based on a large language model to obtain exercise recommendation results for different target tasks includes: obtaining task prompts for multiple target tasks and recommended task input information; and processing multiple task prompts, recommended task input information, and a preset number of exercises based on a large language model to obtain exercise recommendation results for each target task.

[0008] In some embodiments of this application, the exercise recommendation method combining hypergraph knowledge modeling and automatic prompt optimization further includes: determining the recommendation fitness corresponding to each task prompt word based on the user's interaction with each exercise recommendation result; selecting a preset number of task prompt words with the highest recommendation fitness to obtain a parent task prompt word set, and obtaining prompt word rewriting feedback for each task prompt word in the parent task prompt word set, wherein the prompt word rewriting feedback includes: the parent task prompt words of the task prompt words in the parent task prompt word set and their exercise recommendation results, and a task optimization strategy for rewriting the parent task prompt words. The algorithm generates a set of prompts for the next generation of tasks, including prompt words, task prompt words in the parent task prompt word set and their recommended exercise results, the recommendation fitness of the parent task prompt words and the task prompt words in the parent task prompt word set, and the difference in fitness between the two. Based on a large language model, the algorithm processes the parent task prompt word set, the prompt word rewriting feedback of each task prompt word, and the input information of the recommended task to obtain the next generation task optimization strategy prompt word set. Based on a large language model, the algorithm processes the next generation task optimization strategy prompt word set and the parent task prompt word set to obtain the next generation task prompt words for multiple target tasks, which are used to generate the exercise recommendation results for each target task in the next iteration.

[0009] In some embodiments of this application, the initial course knowledge state vector of the source course is optimized based on the discrimination vector enhanced by the user's historical interactions with each exercise in the source course, the difficulty vector of each exercise, the association matrix between exercises and knowledge points, and the answer records. This yields the trained course knowledge state vector of the source course. The initial course knowledge state vector of the target course is determined based on the trained course knowledge state vector and the final transfer association weights of knowledge points from each source course to the target course. For the source course, the cognitive diagnostic model to be trained optimizes the initial course knowledge state vector of the source course based on the discrimination vector enhanced by the user's historical interactions with each exercise in the source course, the difficulty vector of each exercise, the association matrix between exercises and knowledge points, and the answer records. The initial course knowledge state vector, the difficulty vector of each exercise in the source course, and the association matrix between the exercises and knowledge points in the source course are used to determine the predicted correct answer probability of the user relative to each exercise in the source course. Based on the predicted correct answer probability of each exercise in the source course and the binary cross-entropy loss function value determined by the user's answer record in the source course, the cognitive diagnostic model to be trained is trained to obtain the trained cognitive diagnostic model and output the trained course knowledge state vector of the source course. Based on the trained course knowledge state vector of the source course and the final transfer association weight of each knowledge point from the source course to the target course, the initial course knowledge state vector of the target course is determined.

[0010] In some embodiments of this application, the process of optimizing the initial course knowledge state vector of the target course based on the discrimination vector of the user's historical interaction with each exercise in the target course, the difficulty vector of each exercise in the target course, the association matrix between the exercises and knowledge points in the target course, and the user's answer records in the target course, to obtain the first knowledge state vector of the target course, includes: for the target course, the cognitive diagnostic model to be trained determines the predicted correct answer probability of the user relative to each exercise in the target course based on the discrimination vector of the user's historical interaction with each exercise in the target course, the initial course knowledge state vector of the target course, the difficulty vector of each exercise in the target course, and the association matrix between the exercises and knowledge points in the target course; the cognitive diagnostic model to be trained is trained based on the predicted correct answer probability of each exercise in the target course and the binary cross-entropy loss function value determined by the user's answer records in the target course, to obtain the trained cognitive diagnostic model and output the first knowledge state vector.

[0011] In some embodiments of this application, for each course, based on the course's exercise-knowledge point association matrix and the course's answer records, the historical interaction context representation of the user with each exercise is determined, and the user representation, the exercise representation, and the historical interaction context representation are combined to obtain the user's historical interaction enhancement discrimination vector for each exercise. This includes: for each course, based on the course's exercise-knowledge point association matrix, determining the set of relevant knowledge point answer records for the user and each exercise in the course's answer records, determining the user's historical answer representation for each question in the relevant knowledge point answer record set, and determining the attention weight of each question relative to the exercise; for each exercise, based on the historical answer representation and attention weight of each question, determining the exercise's historical interaction context representation; determining the initial interaction discrimination vector of the user representation with respect to each exercise representation, and correcting the initial interaction discrimination vector based on the historical interaction context representation to obtain the user's historical interaction enhancement discrimination vector for each exercise.

[0012] In some embodiments of this application, determining the initial interactive discrimination vector of the user representation for each exercise representation, and correcting the initial interactive discrimination vector according to the historical interaction context representation to obtain the historically enhanced discrimination vector of the user for each exercise includes: determining the initial interactive discrimination vector of the user representation for each exercise representation, and obtaining the historically corrected vector according to the historical interaction context representation; and fusing the initial interactive discrimination vector and the historically corrected vector based on a gating mechanism to obtain the historically enhanced discrimination vector of the user for each exercise.

[0013] In some embodiments of this application, the recommended score of each exercise in the target course relative to the user is determined based on the association matrix between the exercises and knowledge points of the target course, the first knowledge state vector, and the prerequisite dependency relationship of the knowledge points. This includes: determining the suitability of each exercise in the target course relative to the user based on the association matrix between the exercises and knowledge points of the target course and the first knowledge state vector, and determining the user's level of prerequisite knowledge for the exercises based on the prerequisite dependency relationship of the knowledge points; for each exercise, the recommended score of the exercise is determined based on the suitability of the exercise and the level of prerequisite knowledge.

[0014] A second aspect of this application provides a problem recommendation device combining hypergraph knowledge modeling and automatic prompting optimization, comprising: a data acquisition module for acquiring the problems, knowledge points, problem-knowledge point association matrices, knowledge point prerequisite dependencies, user answer records in each course, user representation, problem representations of each problem in each course, and the user's initial course knowledge state vector in the source course, wherein the multiple courses include source courses and target courses; and an initial migration association matrix determination module for determining the prerequisite dependency hyperedge relationships between knowledge points in each course, using the knowledge points and problems of the multiple courses as nodes, based on the problem-knowledge point association matrices and knowledge point prerequisite dependencies. The system establishes a hyper-edge relationship framework, which includes: similarity relationships between knowledge points; hyper-edge relationships between exercises and knowledge points within each course; and cross-course prerequisite knowledge point hyper-edge relationships between knowledge points in the source course and knowledge points in the target course. Based on these hyper-edge relationships, it obtains a cross-course heterogeneous hypergraph association matrix and a hyper-edge weight matrix. It then obtains a node degree matrix based on the cross-course heterogeneous hypergraph association matrix and hyper-edge weight matrix, and a hyper-edge degree matrix based on the cross-course heterogeneous hypergraph association matrix. Finally, it determines the initial transfer association matrix between knowledge points in the source course and the target course based on the cross-course prerequisite knowledge point hyper-edge relationships. The final transfer association weight determination module is used to determine the transfer association weights based on the cross-course heterogeneous hypergraph association matrix, hyper-edge weight matrix, node degree matrix, and hyper-edge degree matrix. The system iterates and updates the initial feature representation of each node multiple times to obtain the higher-order node representation of each node. Based on the initial transfer association matrix, it corrects the association weights of the higher-order node representations corresponding to each knowledge point, thus obtaining the final transfer association weights of knowledge points from the source course to the target course. The discrimination vector determination module, for each course, determines the historical interaction context representation of the user with each exercise based on the course's exercise-knowledge point association matrix and the course's answer records. It then combines the user representation, the exercise representation, and the historical interaction context representation to obtain the discrimination vector enhanced by the user's historical interaction with each exercise. The initial course knowledge state vector determination module determines the discrimination vector based on the user's historical interaction enhancement with each exercise in the source course. The initial course knowledge state vector of the source course is optimized using the discrimination vector, the difficulty vector of each exercise, the association matrix between exercises and knowledge points, and the answer records. This yields the post-training course knowledge state vector of the source course. Based on the post-training course knowledge state vector and the final transfer association weights of knowledge points from each source course to the target course, the initial course knowledge state vector of the target course is determined. The first knowledge state vector determination module is used to optimize the initial course knowledge state vector of the target course based on the discrimination vector enhanced by the user's historical interaction with each exercise in the target course, the difficulty vector of each exercise in the target course, the association matrix between exercises and knowledge points in the target course, and the user's answer records in the target course, to obtain the first knowledge state vector of the target course.The exercise recommendation module determines the recommendation score of each exercise in the target course relative to the user based on the association matrix between exercises and knowledge points, the first knowledge state vector, and the prerequisite dependencies of knowledge points. It then processes a preset number of exercises with the highest recommendation scores and multiple task prompts using a large language model to obtain exercise recommendation results for different target tasks. The task prompts are iteratively optimized based on the user's interaction with historical exercise recommendation results.

[0015] A third aspect of this application provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the exercise recommendation method combining hypergraph knowledge modeling and automatic prompting optimization as provided in the first aspect of this application.

[0016] The beneficial effects of this invention are that, since user learning behavior data and answer records are relatively scarce in the early stages of a course, traditional exercise recommendation methods, which only analyze limited data within the new course, cannot accurately characterize users' cognitive levels and learning preferences. Furthermore, in existing exercise recommendation methods based on large language models, the optimization strategies used to guide task prompt word adjustments are usually pre-set, and the optimization of task prompt words is easily limited by the content of the pre-set strategies, resulting in insufficient targeting of exercise recommendation results. Therefore, the exercise recommendation method of this application, which combines hypergraph knowledge modeling and automatic prompt optimization, constructs a cross-course heterogeneous hypergraph based on the semantic relationships of knowledge points. This method can fully characterize various cross-course knowledge association relationships, such as single-source single-objective transfer relationships and multi-source single-objective transfer relationships, as well as intra-course knowledge association relationships such as prerequisite dependencies, concept similarity, and exercise-knowledge point associations. It also updates task prompt words and task optimization strategy prompt words used to optimize task prompt words based on user interaction with the exercise recommendation results. Compared to existing cross-course exercise recommendation methods, the exercise recommendation method provided in this application, which combines hypergraph knowledge modeling and automatic prompt optimization, can more fully characterize the complex knowledge relationships between and within courses, improve the effect of knowledge state transfer, and enhance the ability to represent the user's interaction with different exercises by combining the user's performance on knowledge-related questions, thereby obtaining a more accurate knowledge state. As a result, in the exercise recommendation stage, it can more fully obtain recommendation results that are more in line with the user's cognitive level based on the first knowledge state vector. At the same time, by updating the task prompt words and their task optimization strategy prompt words according to the actual recommendation effect, the limitation of the preset optimization strategy on the adjustment of task prompt words is reduced, thereby obtaining task prompt words that are more adapted to user needs during exercise recommendation, and thus obtaining exercise recommendation results that are more in line with user needs.

[0017] In the above method, the knowledge state is further updated through knowledge state transfer and a small number of answer records of the target course to obtain the first knowledge state vector. The first knowledge state vector reflects the user's mastery of multiple knowledge points in the target course. The organic combination of the first knowledge state vector and the association matrix between exercises and knowledge points can map the user's mastery of knowledge points to the exercise level, thereby determining the user's fit with the exercises. Furthermore, the user's prior knowledge is determined by combining the prerequisite dependencies of knowledge points, and the recommended score of the exercise is determined accordingly. Based on this, exercises with higher recommended scores are used as candidate exercises for exercise recommendation using a large language model, and the exercise recommendation results are obtained by combining automatically optimized task prompts. Therefore, the above method revolves around the user's mastery of knowledge points, comprehensively considers the user's fit with the exercises and the prior knowledge base, so that the determined exercise recommendation score can reflect the matching degree between the exercise and the user's cognitive level. Furthermore, the task prompts are optimized based on the user's interaction with the recommendation results to obtain exercise recommendation results that simultaneously meet the user's cognitive level and learning preferences.

[0018] Therefore, the above methods can improve the knowledge state transfer effect through cross-course heterogeneous hypergraph knowledge modeling, enhance the characterization of the interaction between users and exercises through knowledge-related question performance, obtain more accurate cognitive diagnosis results, and optimize the large language model through automatic prompts to generate exercise recommendation results that are more in line with users' learning preferences, thereby providing users with more suitable exercise recommendation results. Attached Figure Description

[0019] Figure 1 A flowchart illustrating a problem recommendation method combining hypergraph knowledge modeling and automatic suggestion optimization provided in an embodiment of this application; Figure 2 A flowchart illustrating a method for determining exercise recommendation results, provided as an embodiment of this application; Figure 3 A flowchart illustrating a process for determining a user's historical interaction enhancement discriminative vector for each exercise, provided in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of an exercise recommendation device provided in an embodiment of this application. Detailed Implementation

[0020] 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.

[0021] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this disclosure should have the ordinary meaning understood by one of ordinary skill in the art to which this disclosure pertains. The terms "first," "second," and similar terms used in the embodiments of this disclosure do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0022] To accurately recommend exercises that match the learning needs of learners, embodiments of this application provide an exercise recommendation method. See also Figure 1 The exercise recommendation method combining hypergraph knowledge modeling and automatic prompting optimization provided in this application includes: S102. Obtain the exercises, knowledge points, exercise-knowledge point association matrix, knowledge point prerequisite dependencies, user's answer records in each course, user representation, exercise representation of each exercise in each course, and user's initial course knowledge state vector in the source course, wherein the multiple courses include the source course and the target course. S104. For multiple courses, the knowledge points and exercises of multiple courses are used as nodes. Based on the association matrix between each exercise and knowledge point and the prerequisite dependency relationship of each knowledge point, the prerequisite dependency hyperedge relationship and knowledge point similarity hyperedge relationship between knowledge points in each course are determined. The association hyperedge relationship between each exercise and knowledge point in each course is also determined. The cross-course prerequisite knowledge point hyperedge relationship between the knowledge points in the source course and the knowledge points in the target course is determined. Based on each hyperedge relationship, the cross-course heterogeneous hypergraph association matrix and hyperedge weight matrix are obtained. Based on the cross-course heterogeneous hypergraph association matrix and hyperedge weight matrix, the node degree matrix is ​​obtained. Based on the cross-course heterogeneous hypergraph association matrix, the hyperedge degree matrix is ​​obtained. Based on each cross-course prerequisite knowledge point hyperedge relationship, the initial transfer association matrix between the knowledge points of the source course and the knowledge points of the target course is determined. S106. Based on the cross-course heterogeneous hypergraph association matrix, hyperedge weight matrix, node degree matrix, and hyperedge degree matrix, the initial feature representation of each node is iteratively updated in multiple rounds to obtain the higher-order node representation of each node. Based on the initial transfer association matrix, the association weight of the higher-order node representation corresponding to each knowledge point is corrected to obtain the final transfer association weight of each knowledge point from the source course to the target course. S108. For each course, based on the course's exercise and knowledge point association matrix and the course's answer records, determine the historical interaction context representation between the user and each exercise, and combine the user representation, each exercise representation, and the historical interaction context representation to obtain the user's historical interaction enhancement discriminability vector for each exercise. S110. Based on the discrimination vector enhanced by the user's historical interaction with each exercise in the source course, the difficulty vector of each exercise, the association matrix between the exercise and the knowledge point, and the answer record, optimize the initial course knowledge state vector of the source course to obtain the post-training course knowledge state vector of the source course. Based on the post-training course knowledge state vector and the final transfer association weight of the knowledge points from each source course to the target course, determine the initial course knowledge state vector of the target course. S112. Based on the discrimination vector enhanced by the user's historical interaction with each exercise in the target course, the difficulty vector of each exercise in the target course, the association matrix between the exercises and knowledge points in the target course, and the user's answer records in the target course, optimize the initial course knowledge state vector of the target course to obtain the first knowledge state vector of the target course.

[0023] Specifically, the first knowledge state vector may include multiple elements, with each element representing the user's level of mastery of a knowledge point.

[0024] S114. Based on the association matrix between exercises and knowledge points of the target course, the first knowledge state vector, and the prerequisite dependency relationship of knowledge points, determine the recommendation score of each exercise in the target course relative to the user, and process the preset number of exercises with the highest recommendation scores and multiple task prompt words based on the large language model to obtain the exercise recommendation results for different target tasks; wherein, the task prompt words are obtained by iterative optimization based on the user's interaction with the historical exercise recommendation results.

[0025] Because user learning behavior data and answer records are relatively scarce in the early stages of a course, traditional exercise recommendation methods, which only analyze limited data within the new course, struggle to accurately characterize users' cognitive levels and learning preferences, resulting in insufficient targeting of exercise recommendation results. Therefore, the exercise recommendation method in this application, combining hypergraph knowledge modeling and automatic prompting optimization, constructs a cross-course heterogeneous hypergraph based on the semantic relationships of knowledge points. This fully characterizes various cross-course knowledge relationships, such as single-source single-objective transfer relationships and multi-source single-objective transfer relationships, as well as intra-course knowledge relationships such as prerequisite dependencies, concept similarity, and exercise-knowledge point associations. Compared to existing cross-course exercise recommendation methods, the exercise recommendation method provided in this application, combining hypergraph knowledge modeling and automatic prompting optimization, can more fully characterize complex knowledge relationships between and within courses, improving the effectiveness of knowledge state transfer. Simultaneously, by combining user performance on knowledge-related questions, it enhances the representation ability of user interactions with different exercises, thereby obtaining a more accurate knowledge state. This allows for a more comprehensive recommendation of results that better match the user's cognitive level based on the first knowledge state vector during the exercise recommendation stage. In the above method, the knowledge state is further updated through knowledge state transfer and a small number of answer records from the target course, resulting in a first knowledge state vector. This first knowledge state vector reflects the user's mastery of multiple knowledge points in the target course. The organic combination of the first knowledge state vector and the association matrix between exercises and knowledge points maps the user's mastery of knowledge points to the exercise level, thereby determining the user's fit with the exercises. Furthermore, by combining the prerequisite dependencies of knowledge points, the user's prior knowledge is determined, and the recommended score for the exercises is accordingly determined. Therefore, the above method revolves around the user's mastery of knowledge points, comprehensively considering the user's fit with the exercises and their prerequisite knowledge base, so that the determined recommended score for the exercises reflects the degree of matching between the exercises and the user's cognitive level. Thus, the above method can improve the knowledge state transfer effect through cross-course heterogeneous hypergraph knowledge modeling and enhance the characterization of the interaction between the user and the exercises through knowledge-related question performance, obtaining more accurate cognitive diagnostic results and providing users with more suitable exercise recommendations.

[0026] See Figure 2 In some embodiments of this application, step S114, which involves processing a preset number of exercises with the highest recommendation scores based on a large language model and multiple task prompts to obtain exercise recommendation results for different target tasks, includes: S202. Obtain task prompts and recommended task input information for multiple target tasks; S204. Based on a large language model, process multiple task prompts, recommended task input information, and a preset number of exercises to obtain exercise recommendation results for each target task.

[0027] Specifically, step S204 includes: processing multiple task prompts, recommended task input information, and a preset number of exercises based on a large language model to obtain exercise recommendation results, learning preference summaries, and recommended explanation texts for each target task.

[0028] As an example, the exercise recommendation results, learning preference summaries, and recommendation explanation texts for each target task are as follows: ; in, This indicates that for users The Summary of learning preferences for each target task This indicates that for users The Recommended exercises for each target task. This indicates that for users The Recommended explanation text for each target task. This indicates the process of calling the large language model. Indicates the first The task prompt for the target task. This represents the input information for the recommendation task. This represents a set of candidate exercises with a predetermined number of questions.

[0029] In some embodiments of this application, the exercise recommendation method combining hypergraph knowledge modeling and automatic suggestion optimization further includes: Based on user interaction with the recommended results for each exercise, determine the recommendation fitness level corresponding to each task prompt word; Select a preset number of task prompt words with the highest recommended fitness to obtain a set of parent task prompt words. Then, obtain the prompt word rewriting feedback for each task prompt word in the set of parent task prompt words. The prompt word rewriting feedback includes: the parent task prompt words and their exercise recommendation results for the task prompt words in the set of parent task prompt words, the task optimization strategy prompt words used to rewrite the parent task prompt words, the task prompt words and their exercise recommendation results in the set of parent task prompt words, the recommended fitness of the parent task prompt words and the task prompt words in the set of parent task prompt words, and the difference in fitness between the two. Based on the large language model, the set of prompt words for the parent task, the feedback on the rewriting of prompt words for each task, and the input information of the recommended task are processed to obtain the set of prompt words for the next generation task optimization strategy. By processing the next-generation task optimization strategy prompt word set and the parent task prompt word set based on the large language model, the next-generation task prompt words for multiple target tasks are obtained, which are then used to generate the exercise recommendation results for each target task in the next iteration.

[0030] In some embodiments of this application, step S110 includes: For the source course, the cognitive diagnostic model to be trained determines the predicted probability of a user answering each question in the source course based on the discrimination vector of the user's historical interaction enhancement for each question in the source course, the initial course knowledge state vector of the source course, the difficulty vector of each question in the source course, and the association matrix between the questions and knowledge points in the source course. The cognitive diagnostic model to be trained is trained based on the predicted correct answer probability of each exercise in the source course and the binary cross-entropy loss function value determined by the user's answer record in the source course. The trained cognitive diagnostic model is obtained and the trained course knowledge state vector of the source course is output. Based on the post-training course knowledge state vector of the source courses and the final transfer association weights of knowledge points from each source course to the target course, the initial course knowledge state vector of the target course is determined.

[0031] In some embodiments of this application, step S112 includes: For the target course, the cognitive diagnostic model to be trained determines the predicted probability of the user answering each question in the target course based on the discrimination vector of the user's historical interaction enhancement for each question in the target course, the initial course knowledge state vector of the target course, the difficulty vector of each question in the target course, and the association matrix between the questions and knowledge points in the target course. The cognitive diagnostic model to be trained is trained based on the predicted correct answer probability of each exercise in the target course and the binary cross-entropy loss function value determined by the user's answer record in the target course. The trained cognitive diagnostic model is then obtained and the first knowledge state vector is output.

[0032] See Figure 3 In some embodiments of this application, step S108 includes: S302. For each course, based on the course's exercise and knowledge point association matrix, determine the set of relevant knowledge point answer records for the user and each exercise in the course's answer records, determine the user's historical answer representation for each question in the relevant knowledge point answer record set, and determine the attention weight of each question relative to the exercises.

[0033] Specifically, for the target course, the historical answer representation of the questions can be, for example, as follows: ; in, Target Course The The question is for the user Regarding the history questions, For the first The title representation of each question, For users For the The answers to each question and These are trainable parameters.

[0034] Specifically, for the target course, the attention weight of the questions relative to the exercises is: ; in, Target Course The The question is relative to the first one. Attention weights for each exercise, softmax h (·) represents the softmax function. For the first The exercise representation of each problem. For the first The title representation of each question; and These are trainable parameters; d emb This represents the vector dimension in attention computation, consistent with the vector dimension represented in the exercises.

[0035] S304. For each exercise, determine the historical interaction context representation of the exercise based on the historical answer representation and attention weight of each exercise.

[0036] Specifically, for the target course, the historical interaction context of the exercises is represented as follows: ; in, Target Course The The exercise is for the user Historical interaction context representation, Target Course The The question is relative to the first one. Attention weights for each exercise Target Course The The question is for the user The answer to the historical question is as follows.

[0037] S306. Determine the initial interactive discrimination vector of the user representation for each exercise representation, and modify the initial interactive discrimination vector according to the historical interaction context representation to obtain the historical interaction enhanced discrimination vector of the user for each exercise.

[0038] In some embodiments of this application, step S306 includes: Determine the initial interactive discriminative vector for each exercise representation of the user representation, and obtain the historical correction vector based on the historical interactive context representation; By fusing the initial interactive discrimination vector and the historical correction vector based on the gating mechanism, a discrimination vector enhanced by the user's historical interaction with each exercise is obtained.

[0039] Specifically, in the target course, for users Exercises, and their initial interactive discrimination vector for: ; in, Indicates user The initial interactive discrimination vector with the exercises; Indicate learner In the target course Users in the middle said, The exercise is represented as an exercise. and It is a trainable parameter matrix; Represents the Hadamard product; s (·) represents the sigmoid function.

[0040] For the target course, historical correction vector for: ; in, Target Course The The exercise is for the user Historical interaction context representation, and These are trainable parameters.

[0041] By fusing the initial interactive discrimination vector and the historical correction vector based on the gating mechanism, a discrimination vector enhanced by the user's historical interaction with each exercise is obtained.

[0042] Specifically, for the target course, the discrimination vector of the historical interaction enhancement of the exercises is: ; ; in, Indicate target course The The exercise is for the user The historical interaction-enhanced discriminative vector, [·||·] denotes a gated vector; [·||·] denotes vector concatenation. and For trainable parameters, Indicates user With the The initial interactive discrimination vector for each exercise. This is the historical correction vector.

[0043] In some embodiments of this application, step S114, which determines the recommendation score of each exercise in the target course relative to the user based on the association matrix between exercises and knowledge points of the target course, the first knowledge state vector, and the prior knowledge dependency relationship, includes: Based on the association matrix between exercises and knowledge points in the target course and the first knowledge state vector, determine the degree of suitability of each exercise in the target course for the user, and combine the prerequisite dependency relationship of knowledge points to determine the degree of prerequisite knowledge possessed by the user for the exercises. For each exercise, the recommended score is determined based on the exercise's suitability and the level of prior knowledge required.

[0044] Specifically, the knowledge point prerequisite dependency relationship indicates the dependency relationship between different knowledge points for a user to learn. That is, before learning a knowledge point, a user should first learn the previous knowledge point that the knowledge point depends on. The user's answer record in the target course can include the questions that the user has already answered in the target course.

[0045] In some embodiments of this application, the degree of adaptation includes: the degree of matching in terms of coverage of weak knowledge points. Determining the degree of adaptation of each exercise in the target course relative to the user, based on the association matrix between exercises and knowledge points in the target course and the first knowledge state vector, includes: Select multiple weak knowledge points from the first knowledge state vector whose mastery level is lower than the preset mastery level threshold; For each exercise in the target course, the total number of knowledge points tested in the exercise is obtained based on the correlation matrix of the target course. The degree of coverage and matching of weak knowledge points in the exercise is determined by the ratio of the number of weak knowledge points tested in the exercise to the total number of knowledge points tested in the exercise.

[0046] Specifically, the coverage and matching degree of the weak knowledge points in the exercises is as follows: ; in, This indicates that for users Regarding the target course The The degree of matching in terms of coverage of weak knowledge points in each exercise. Indicate target course The Do these exercises test knowledge points from the target course? , Indicates user In the target course A collection of weak knowledge points Indicate target course The total number of knowledge points in the target course. The Each exercise tests the knowledge points of the target course. In the case of Take 1, in the target course The These exercises do not test knowledge points from the target course. In the case of Take 0. The larger the value, the better the exercise covers the learner's weaker areas in the target course.

[0047] In some embodiments of this application, the degree of adaptation includes: the degree of matching of exercise difficulty. Determining the degree of adaptation of each exercise in the target course relative to the user, based on the association matrix between exercises and knowledge points in the target course and the first knowledge state vector, includes: For each exercise in the target course, the ratio of the total mastery of all knowledge points tested in the exercise to the total number of knowledge points tested in the exercise is determined based on the association matrix between the exercise and the knowledge points and the first knowledge state vector. The degree of matching of the exercise difficulty is determined based on the absolute difference between the preset exercise difficulty and the ratio.

[0048] Specifically, the difficulty level of the exercises is matched as follows: ; in, This indicates that for users Regarding the target course The The matching degree of difficulty of each exercise. Indicate target course The Does the first exercise test the first...? One knowledge point, Indicates user In the target course For the The degree of mastery of each knowledge point Indicate target course The total number of knowledge points, Indicates the preset difficulty level of the exercises. This represents the preset target difficulty offset, used to control the degree to which the difficulty of the exercises deviates from the average level of mastery of the relevant knowledge points by users. It can be manually preset or adjusted according to the teaching requirements of the target course, the range of exercise difficulty markings, or recommendation strategies. The larger the value, the more closely the difficulty of the exercises matches the learner's current cognitive level.

[0049] In some embodiments of this application, the degree of fit between each exercise in the target course and the user is determined based on the association matrix between exercises and knowledge points of the target course and the first knowledge state vector. Furthermore, the degree of prior knowledge the user possesses for the exercises is determined by combining the prior knowledge dependencies of the knowledge points. Based on the correlation matrix between exercises and knowledge points in the target course and the first knowledge state vector, determine the degree of suitability of each exercise in the target course for the user. For each exercise in the target course, the set of knowledge points tested by the exercise is determined based on the association matrix, and the set of prerequisite knowledge points for the set of knowledge points tested by the exercise is determined based on the prerequisite knowledge point dependencies. The total mastery level of all knowledge points in the set of prerequisite knowledge points is determined based on the first knowledge state vector. The user's level of prior knowledge for the exercises is determined by the ratio of their overall mastery of all knowledge points to the total number of knowledge points in the prerequisite knowledge set.

[0050] Specifically, the user's prerequisite knowledge for the exercises is as follows: ; in, Indicates user For target courses Exercises The level of prerequisite knowledge possessed. Indicates the first Exercises Prerequisite knowledge points The first in One knowledge point , Indicates user In the target course For the The degree of mastery of each knowledge point. The degree of prerequisite knowledge possessed. The larger the value, the more prerequisite knowledge the user already possesses that will enable them to complete the exercise. (This refers to the set of prerequisite knowledge points.) Explain the exercises when the set is empty. If there are no prerequisite knowledge points, in this case, the degree of prerequisite knowledge possessed. Take 1.

[0051] In some embodiments of this application, determining the recommended score of an exercise based on the degree of fit between the exercise and the user and the level of prior knowledge includes: determining the recommended score of the exercise based on the weighted sum of the degree of fit between the exercise and the user and the level of prior knowledge.

[0052] The following is combined Figure 4The present application provides an example of a problem recommendation method that combines hypergraph knowledge modeling and automatic suggestion optimization. The apparatus involved in the problem recommendation method consists of the following four modules, which are described below: The text embedding construction module receives knowledge-point related text, exercise-related text, and exercise-knowledge-point association information from the "Source Course Resource Database" and the "Target Course Resource Database." It then uses a text embedding model to semantically encode the knowledge-point related text and exercise-related text, obtaining knowledge-point related text embedding vectors and exercise-related text embedding vectors. Based on the exercise-knowledge-point association information, it constructs an exercise-knowledge-point association matrix. The knowledge-point related text embedding vectors, exercise-related text embedding vectors, and exercise-knowledge-point association matrix of the source and target courses are then passed to the "Cross-Course Knowledge Modeling Module." The exercise-knowledge-point association matrix of the source and target courses is passed to the "Cross-Course Cognitive Diagnosis Module," and the exercise-knowledge-point association matrix of the target course is passed to the "Exercise Recommendation Module."

[0053] The cross-course knowledge modeling module receives knowledge-related text from the "source course resource database" and the "target course resource database." It analyzes the complex knowledge relationships between courses using a large language model to construct cross-course hyperedges in a cross-course heterogeneous hypergraph. Based on the knowledge-related text embedding vectors and exercise-knowledge-point association matrices output by the "text embedding construction module," it constructs intra-course hyperedges in the cross-course heterogeneous hypergraph. Next, it constructs a hypergraph association matrix, hyperedge weight matrix, node degree matrix, and hyperedge degree matrix based on the cross-course heterogeneous hypergraph. Then, based on the knowledge-related text embedding vectors and exercise-related text embedding vectors output by the "text embedding construction module," it obtains the initial feature representations of knowledge-point nodes and exercise nodes in the cross-course heterogeneous hypergraph. Based on the above matrices and the initial feature representations of knowledge-point nodes and exercise nodes, it obtains higher-order knowledge-point representations through hypergraph convolution, generating the final transfer association matrix from source course knowledge points to target course knowledge points. This module then passes the final transfer association matrix and the initial feature representations of exercise nodes to the "cross-course cognitive diagnosis module."

[0054] The cross-course cognitive diagnosis module receives learners' answer records in the source course and a small number of answer records in the target course from the "Learner's Historical Learning Records." It also receives the final transfer association matrix and initial feature representations of exercise nodes from the "Cross-Course Knowledge Modeling Module," as well as the exercise-knowledge point association matrices for the source and target courses from the "Text Embedding Construction Module." This module first constructs a cognitive diagnosis model with enhanced discrimination. Specifically, it selects a set of historical answer records related to the knowledge of each learner's answer record, constructs a historical interaction context representation for each answer record, and uses this representation to correct the interactive discrimination used to represent the student-exercise interaction relationship in the cognitive diagnosis model, thus enhancing the accuracy of the cognitive diagnosis results. Then, it inputs the learner's source course answer records into the cognitive diagnosis model to obtain the learner's source course knowledge state. Combined with the final transfer association matrix, it obtains the learner's initial knowledge state for the target course. Finally, it inputs a small number of answer records from the target course into the cognitive diagnosis model to obtain the updated target course knowledge state. This module passes the updated knowledge status of the target course to the "Exercise Recommendation Module".

[0055] The exercise recommendation module receives the target course exercise-knowledge point association matrix from the "Text Embedding Construction Module," the target course knowledge status from the "Cross-Course Cognitive Diagnosis Module," learner answer records in the source course and a small number of answer records in the target course from the "Learner's Historical Learning Records," and exercise resources from the "Target Course Resource Database." First, a candidate exercise set is determined based on the target course knowledge status and the exercise-knowledge point association matrix. Then, a two-layer prompt word optimization mechanism is used, leveraging a large language model to simultaneously optimize task prompt words and task optimization strategy prompt words. Specifically, this module generates initial task prompts by combining the input information of the recommendation task, task optimization strategy prompts, and prompt generation style constraints. Based on the input information of the recommendation task, such as answer records, and the candidate exercise set, it uses the task prompts to generate a learning preference summary, exercise recommendation results, and recommendation explanations, and calculates the fitness score of each task prompt. It updates the task optimization strategy prompts based on feedback from historical prompts, parent task prompts, and the input information of the recommendation task, and then rewrites the parent task prompts with the task optimization strategy prompts to generate child task prompts. This iterative process is repeated until a preset number of iterations or fitness convergence condition is reached, and the task prompt with the highest fitness is selected as the optimal task prompt. Finally, based on the input information of the recommendation task and the candidate exercise set, the optimal task prompt is used to generate exercise recommendation results.

[0056] The specific work process is as follows: (a) The working content of the "Text Embedding Building Module" is as follows: This module receives text related to knowledge points, exercises, and exercise-knowledge point associations from the "Source Course Resource Database" and text related to knowledge points, exercises, and exercise-knowledge point associations from the "Target Course Resource Database." For each knowledge point, the related text includes the knowledge point name and definition; for each exercise, the related text includes the exercise question and its explanation. This module encodes these texts using a text embedding model to obtain their corresponding text embedding vectors. Based on the exercise-knowledge point association information, an exercise-knowledge point association matrix is ​​constructed.

[0057] Suppose that the set of knowledge points for the source course and the set of knowledge points for the target course are represented as follows: , ,in, n S This indicates the number of knowledge points in the source course. n T This indicates the number of knowledge points in the target course. For knowledge points in the source course... The text containing the knowledge point name and the text containing the knowledge point definition are respectively denoted as... , Regarding the target course knowledge points The text containing the knowledge point name and the text containing the knowledge point definition are respectively denoted as... , .

[0058] The source course exercise set and the target course exercise set are respectively represented as follows: and ,in, m S and m T These represent the number of exercises in the source course and the number of exercises in the target course, respectively. (The last part, "for the source course exercises," appears to be incomplete and lacks context.) The exercise question text and the exercise solution text are respectively denoted as , Based on the association information between exercises and knowledge points, the association vector of knowledge points is obtained. Among them, if the exercises Examine the knowledge points of the source course ,but ,otherwise Based on the knowledge point association vectors of all exercises in the source course, the exercise-knowledge point association matrix of the source course is obtained. Similarly, for the exercises in the target course... The exercise question text and the exercise solution text are respectively denoted as , Based on the association information between exercises and knowledge points, the association vector of knowledge points is obtained. Based on the knowledge point association vectors of all exercises in the target course, the exercise-knowledge point association matrix of the target course is obtained. .

[0059] To effectively extract semantic information from text related to knowledge points and exercises, the texts are input into a text embedding model to obtain corresponding embedding vectors. For the source course knowledge points... Each name and its definition text are input into a text embedding model to obtain the corresponding name embedding vector. and defining embedding vectors As shown in formulas (1) and (2).

[0060] (1) (2) in, F (·) represents a text embedding model, which can be Word2Vec, BERT, etc.

[0061] Similarly, the target course knowledge points are obtained through text embedding models. Name embedding vector and defining embedding vectors Source Course Exercises Title embedding vector and parsing embedding vectors Target Course Exercises Title embedding vector and parsing embedding vectors The text embedding vectors and the exercise-knowledge point association matrix of the source course and the target course obtained above are passed to the "Cross-Course Knowledge Modeling Module", and the exercise-knowledge point association matrix of the source course and the target course is passed to the "Cross-Course Cognitive Diagnosis Module".

[0062] (II) The work content of the "Cross-Course Knowledge Modeling Module" is as follows: Step 1: Construction of a cross-course heterogeneous hypergraph This module receives the knowledge point name embedding vectors, knowledge point definition embedding vectors, exercise question embedding vectors, exercise analysis embedding vectors, and exercise-knowledge point association matrices from the source and target courses, as well as the knowledge point-related texts from the "Source Course Resource Database" and the "Target Course Resource Database," and uses them to construct a cross-course heterogeneous hypergraph between the source and target courses. ,in, This represents a set of nodes, including source course knowledge points, target course knowledge points, source course exercises, and target course exercises. This represents the set of superedges.

[0063] To enable the simultaneous learning of complex knowledge relationships both between and within courses during knowledge transfer, this module constructs cross-course hyperedges and intra-course hyperedges within the cross-course heterogeneous hypergraph. Cross-course hyperedges describe the knowledge foundation required for learners to master the knowledge points of the target course from one or more knowledge points in the source course, thus providing a basis for the transfer of knowledge states formed in the source course to the target course. Intra-course hyperedges represent prerequisite dependencies, conceptual similarities, and exercise-knowledge point relationships within the source and target courses, allowing the subsequent hypergraph-based knowledge state transfer process to utilize both inter-course and intra-course knowledge relationships. Therefore, ,in, This represents the set of cross-course hyperedges, which is a set of hyperedges representing transfer relationships from a single source and a single goal. and the set of superedges for multi-source single-target migration relationships constitute, Denotes the set of superedges within a course, which is derived from the set of prerequisite-dependent superedges within the course. Set of concept-similar hyperedges And the set of superedges related to exercises and knowledge points Composition. The specific construction process of cross-course hyperedges and intra-course hyperedges is as follows.

[0064] 1) Cross-course hyperborder construction process Based on the source course knowledge point related text and target course knowledge point related text received from the "source course resource database" and "target course resource database", this module uses a large language model to analyze the names and definitions of the source course knowledge points and target course knowledge points, determines whether one or more knowledge points in the source course can be used as cross-course prerequisite knowledge required to master the target course knowledge points, analyzes the corresponding relation weights, and constructs cross-course hyperedges based on the analysis results.

[0065] Specifically, regarding the target course knowledge points This module uses its name and definition as the text of the knowledge points to be analyzed, and collects the knowledge points from the source course. K S The names and definitions of each knowledge point in the text are used as a collection of candidate source course knowledge point texts. These are then analyzed using a large language model to determine which knowledge points will be selected as target course knowledge points. The source course knowledge points of the prerequisite knowledge form A collection of prerequisite knowledge points across courses Meanwhile, the large language model will provide the overall relationship weight between the cross-course prerequisite knowledge point set and the target course knowledge point. The degree of prerequisite association between different source course knowledge points and target course knowledge points varies. Therefore, the large language model will provide the contribution weight of each source course knowledge point and target course knowledge point in the cross-course prerequisite knowledge point set. The specific process is shown in formula (3).

[0066] (3) Among them, LLM pre (·) indicates the process of using a large language model to identify prerequisite knowledge points across courses; This represents the name and definition text of all knowledge points in the source course; and These represent the target course knowledge points. Name text and definition text; This represents the overall relational weights of the cross-course prerequisite knowledge point set output by the large language model; Represents a set The set of contribution weights of knowledge points from each source course.

[0067] Based on the cross-course prerequisite knowledge point set Based on the number of knowledge points, cross-course hyperedges are divided into two types: single-source single-objective transfer relationship hyperedges and multi-source single-objective transfer relationship hyperedges. The specific construction process is as follows.

[0068] a) Single-source, single-target migration relationship hyperedge when That is, the target course knowledge points When the set of prerequisite knowledge points across courses contains only one source course knowledge point, it means that the target course knowledge point is composed of only a single source course knowledge point. Based on this knowledge base, a single-source, single-target migration relationship hyperedge is constructed. And add this hyperedge to the set of single-source, single-target migration relationship hyperedges. In addition, the hyperedge weight is defined as the overall relation weight output by the large language model, i.e. ,and ∈[0, 1]. In a single-source, single-objective scenario, the source course knowledge points The contribution weight is Source course knowledge points To the target course knowledge points The effective transfer weight across courses is .

[0069] b) Multi-source single-target migration relationship hyperedge when That is, the target course knowledge points When the set of prerequisite knowledge points for a cross-course course contains knowledge points from two or more source courses, this means that the target course knowledge points... It requires knowledge points from two or more source courses to form its knowledge base, therefore, a multi-source single-objective transfer relationship hyperedge needs to be constructed. And add this hyperedge to the set of multi-source single-objective migration relationship hyperedges. In addition, the hyperedge weight is defined as the overall relation weight output by the large language model, i.e. In multi-source, single-objective scenarios, the source course knowledge points To the target course knowledge points The effective transfer weight across courses is ,in .

[0070] 2) Hyperedge construction process within the course Hyperedges within the same course are divided into prerequisite-dependent hyperedges. Conceptually similar hyperedge Exercises - Knowledge Point Related Hyperedge The specific construction process is as follows.

[0071] a) First fix the dependent hyperedge Prerequisite dependency hyperedges are used to characterize the prerequisite dependencies between knowledge points within the same course. For the same course C ∈{ S , T Experts marked the prerequisite dependencies between knowledge points according to the course syllabus. If knowledge points... It is a knowledge point The prerequisite knowledge points are then used to construct a prerequisite dependency hyperedge. And add it to the set of pre-condition dependent superedges. At the same time, this superedge weight .

[0072] b) Conceptually similar hyperedges Intra-course concept similarity hyperedges are used to characterize the semantic similarity relationships between knowledge points within the same course. For the same course C ∈{ S , T Any two knowledge points in} and First, based on the name embedding vectors of two knowledge points, the similarity between their knowledge point names is calculated using a cosine similarity function. Based on the definition embedding vectors of the two knowledge points, the similarity of their knowledge point definitions is calculated using the cosine similarity function. The specific process is shown in formulas (4) and (5).

[0073] (4) (5) Where cos(·) is the cosine similarity function; and These represent knowledge points. and The name embedding vector is obtained from the knowledge point name embedding vector of the source or target course passed by the "Text Embedding Builder"; and These represent knowledge points. and The embedding vector is defined by the knowledge points of the source or target course passed from the "Text Embedding Builder" and is obtained by defining the embedding vector.

[0074] Next, through the knowledge points and Knowledge point name similarity Similarity to knowledge point definitions To obtain knowledge points and final semantic similarity The specific process is shown in formula (6).

[0075] (6) in, and These are non-negative weighting coefficients, and .

[0076] if Then construct knowledge points and Conceptual similarity hyperedge And add it to the set of concept-similar hyperedges. Meanwhile, the weight of this hyperedge is determined by the final semantic similarity. To obtain, that is .

[0077] c) Exercises - Knowledge Point Related Hyperedges The exercise-knowledge point relationship hyperedge is used to characterize whether exercises within the same course test a specific knowledge point. For courses... C ∈{ S , T Exercises in} Based on the "text embedding building block" passed in, the exercise-knowledge point association matrix Q C The set of knowledge points examined is as follows: Based on the set of knowledge points tested in this exercise, construct a hyperedge relating the exercise and the knowledge points. And add it to the set of superedges relating exercises and knowledge points. The superedge weight .

[0078] Step 2: Construction of Hypergraph Matrix Representation In obtaining a cross-course heterogeneous hypergraph Then, based on the association relationships between nodes and hyperedges and the weights of various hyperedges, we construct a matrix representation for subsequent hypergraph convolution calculations, including the hypergraph association matrix. Hyperedge weight matrix W Node degree matrix D and hypermarginality matrix D .

[0079] For ease of matrix representation, the set of nodes is denoted as... ,in, This represents the number of nodes in the hypergraph. The set of hyperedges is denoted as... ,in, This indicates the number of hyperedges in the hypergraph.

[0080] Based on the association relationships between nodes and hyperedges, the hypergraph incidence matrix is ​​obtained. .in, , These represent the total number of nodes and the total number of hyperedges, respectively. (In the hypergraph incidence matrix...) In this diagram, each row corresponds to a node, and each column corresponds to a hyperedge. H v,ε Represents a node v With super-edge e The relationships between nodes. If nodes v Belongs to superedge e Then H v,ε =1; if node v Not a hyperedge e Then H v,ε =0. The hypergraph incidence matrix represents whether there is an association between nodes and hyperedges, and the strength of the association between different hyperedges is represented by the hyperedge weight. For any hyperedge... e ∈ Let its weights be expressed as According to the hyperedge set Construct the hyperedge weight matrix , among which, diag( ) represents the constructor for a diagonal matrix.

[0081] Based on hypergraph correlation matrix and hyperedge weight matrix W The node degree matrix D is obtained. and hypermarginality matrix D First, nodes v of degree d (v ) is related to the node v The sum of the weights of all connected hyperedges is defined as shown in formula (7).

[0082] (7) Super Edge e degree d ( e ) is the number of nodes contained in the superedge, as defined in formula (8).

[0083] (8) This yields the node degree matrix D. and hypermarginality matrix D As shown in formulas (9) and (10).

[0084] (9) (10) Step 3: Construction of the initial migration association matrix To transfer the knowledge state of the source course to the target course, we use cross-course hyperedges. Hypergraph Independence Matrix and hyperedge weight matrix W Construct the initial migration association matrix .matrix The rows represent the source course knowledge points, the columns represent the target course knowledge points, and the matrix elements represent the initial transfer association weights between the source course knowledge points and the target course knowledge points. Representing source course knowledge points With the target course knowledge points The initial migration association weights between them.

[0085] Based on the knowledge points of the source course With the target course knowledge points For example, the initial migration association weights between them Source course knowledge points and target course knowledge points The cross-course effective transfer weight is calculated as shown in formula (11).

[0086] (11) in, and These represent the effective transfer weights across courses in single-source, single-objective and multi-source, multi-objective scenarios, respectively. The resulting... As the initial migration correlation matrix.

[0087] Step 4: Hypergraph Convolutional Representation Learning Obtain the hypergraph incidence matrix Hyperedge weight matrix W Node degree matrix D and hypermarginality matrix D Then, we update the node representations using hypergraph convolution.

[0088] For any node Its initial characteristics are represented as Specifically, if the node is a course C ∈{ S , T} knowledge point nodes The initial feature representation of this node is obtained by weighted aggregation of its corresponding knowledge point name embedding vector and knowledge point definition embedding vector, i.e. ,in, and These represent knowledge points. Name embedding vector and definition embedding vector, and These are non-negative weighting coefficients. If the node is a course... C ∈{ S , T Exercise nodes The initial feature representation of this node is obtained by weighted aggregation of its corresponding exercise question embedding vector and exercise parse embedding vector, i.e. ,in, and Representing the exercises The question embedding vector and the parsed embedding vector, and These are non-negative weighting coefficients.

[0089] Based on the initial feature representations of the nodes, the initial layer node representation matrix is ​​obtained as follows: ,in, Represents a node v i The corresponding initial feature representation, F (0) As input to the first layer of hypergraph convolution. For the ... l Layer node representation matrix F (l) Perform hypergraph convolution operation to obtain the first... l +1 level nodes represent matrix F (l+1) The process is shown in formula (12).

[0090] (12) in, s (·) represents the sigmoid function; Indicates the first lThe trainable parameter matrix of the hypergraph convolution.

[0091] go through L After convolution of the hypergraph, the higher-order node representations of the source course knowledge point nodes and the target course knowledge point nodes are obtained, as shown in formulas (13) and (14).

[0092] (13) (14) in, Representing source course knowledge points The higher-order node representation; Indicates the target course knowledge points The higher-order node representation of a knowledge point, after hypergraph convolution, includes not only the textual semantic information of the knowledge point node itself, but also the related node information connected to the knowledge point through cross-course single-source single-objective transfer relationship hyperedges, multi-source single-objective transfer relationship hyperedges, as well as in-course prerequisite dependency hyperedges, concept similarity hyperedges, and exercise-knowledge point association hyperedges.

[0093] Step 5: Generation of the final migration association matrix Based on the initial migration correlation matrix A 0 The hypergraph representations of the source course knowledge points and the target course knowledge points are used to generate the final transfer association matrix. First, based on the higher-order node representations of the source course knowledge points and the target course knowledge points, the node representation association weights between the two are obtained. Then, the initial transfer association weights are corrected to obtain the corrected transfer association weights. The specific process is shown in formula (15).

[0094] (15) in, It is a trainable parameter matrix; Representing source course knowledge points With the target course knowledge points The nodes between them represent the association weights; Source course knowledge points With the target course knowledge points The initial migration association weights.

[0095] To facilitate subsequent knowledge state transfer, Normalization was performed according to the knowledge points of the target course to obtain the final transfer association weights. As shown in formula (16).

[0096] (16) in, Representing source course knowledge points To the target course knowledge points The final migration association weights. The resulting final migration association matrix. The initial feature representations of the exercise nodes of the source course and the target course are passed to the "cross-course cognitive diagnosis module".

[0097] (III) The work content of the "Cross-Curriculum Cognitive Diagnosis Module" is as follows: Step 1: Constructing a Historical Interaction Context Representation Hypothetical Course C ∈{ S , T},in, S Indicates the source course, T Indicates the target course. (In the course) C In China, for learners u One of the exercises mentioned in the answer record We start from the learners u answer Selected from previous answer records The relevant historical answer records. If the historical answer records involve exercises... Exercises If at least one knowledge point is shared, then that historical answer record is considered the learner's answer. u In the exercises The learner's historical answer records. u In the exercises The relevant historical answer records set is defined as shown in formula (17).

[0098] (17) in, Indicate learner u For exercises The answer result, This indicates that the answer is correct. This indicates an incorrect answer; , Representing learners u Answer the exercises , Time, Indicate learner u Answer the exercises Time precedes answering exercises Time; , Representing the exercises Exercises The knowledge point association vector is the exercise-knowledge point association matrix Q passed by the "text embedding building module". C get, Representing exercises With Exercises Share at least one knowledge point.

[0099] For learners u One of the exercises mentioned in the answer record. Corresponding historical answer records ∈ Construct the corresponding historical response representation As shown in formula (18).

[0100] (18) in, Representing exercises The exercise representation is obtained from the initial feature representation of the exercise nodes in the "Cross-Course Knowledge Modeling Module"; [·||·] represents vector concatenation; and These are trainable parameters.

[0101] To differentiate the impact of different historical response records on student performance, learners u One of the exercises mentioned in the answer record The exercise representation is used as the query, and the corresponding exercise representation from the relevant historical answer records is used as the key. The attention score is normalized using the softmax function to obtain the exercises from the relevant historical answer records. Compared to exercises Attention weights The specific process is shown in formula (19).

[0102] (19) Among them, softmax h (·) indicates that the exercise is solved using the softmax function. The corresponding collection of historical answer records The attention scores for each history exercise are normalized so that the sum of the attention weights for each relevant history answer record is 1. , Representing the exercises and exercises The exercise representation is obtained from the initial feature representation of the exercise nodes in the "Cross-Course Knowledge Modeling Module"; and These are trainable parameters; d emb This represents the vector dimension in attention computation, consistent with the vector dimension represented in the exercises.

[0103] Finally, based on attention weights, a weighted sum of relevant historical response representations is obtained to determine the learner's response.u A certain exercise mentioned in the answer record Historical interaction context representation The specific process is shown in formula (20).

[0104] (20) Step 2: Construction of Historical Interaction Enhanced Discriminative Representation After obtaining the historical interaction context representation, we first calculate the initial interaction discrimination vector based on the learner representation and the exercise representation. Then, we use the historical interaction context representation to correct the initial interaction discrimination vector to obtain the historical interaction enhanced discrimination vector, and calculate the predicted probability of the learner answering the exercise correctly based on this vector.

[0105] First, for learners u and exercises Its initial interactive discriminant vector The calculation process is shown in formula (21).

[0106] (twenty one) in, Indicate learner u With Exercises The initial interactive discriminant vector; Indicate learner u In the course C Learners in the middle indicated; Representing exercises The exercises are represented as follows; and It is a trainable parameter matrix; Represents the Hadamard product; s (·) represents the sigmoid function.

[0107] Then, based on the historical interaction context, it represents Obtain the historical correction vector The specific process is shown in formula (22).

[0108] (twenty two) in, and These are trainable parameters.

[0109] Next, the initial interactive discrimination vector and the historical correction vector are fused through a gating mechanism to obtain the historical interactive enhanced discrimination vector. The specific process is shown in formulas (23) and (24).

[0110] (twenty three) (twenty four) in, [·||·] denotes a gated vector; [·||·] denotes vector concatenation. and These are trainable parameters; This represents the interactive discriminative vector that enhances historical interactions.

[0111] After obtaining the interactive discrimination vector enhanced by historical interactions, the module multiplies it dimension-by-dimensionally with the learner's initial answer advantage vector on the questions, and obtains the probability of correct answer through a forward fully connected layer. The specific process is shown in formula (25).

[0112] (25) in, Indicate learner u Answer the exercises correctly The predicted probability; Represents the interactive discriminative vector for enhanced historical interaction; Representing exercises The knowledge point association vector is the exercise-knowledge point association matrix Q passed by the "text embedding building module". C get; Indicate learner u In the course C The state of knowledge within; Representing exercises The difficulty vector; Indicate learner u For exercises The initial response advantage vector; PL C (·) indicates a course C The forward fully connected layer.

[0113] Step 3: Source Course Knowledge Status Diagnosis Based on the content of steps 1 and 2, a discrimination enhancement cognitive diagnostic model is constructed, and the discrimination enhancement cognitive diagnostic model of the source course is trained by using a binary cross-entropy loss function based on the learner's answer records in the "learner's historical learning record". The loss function is defined as shown in formula (26).

[0114] (26) in, Representing learners u A collection of answer records indexes from the source course; Indicate learner u Answer source course exercises The result; Indicate learner uAnswer the source course exercises correctly. The predicted probability is obtained by optimizing the above loss function. u Knowledge status in the source course ,in, Indicate learner u For the source course knowledge points The level of mastery. This source course knowledge status serves as input for generating the initial knowledge status of subsequent target courses.

[0115] Step 4: Generating the initial knowledge state of the target course Through the final transfer association matrix of the "Cross-Course Knowledge Modeling Module," learners' knowledge state in the source course is transferred to the target course, resulting in the initial knowledge state of the target course. Specifically, this is applied to the knowledge points of the target course. To achieve the final migration association matrix The transfer correlation weights from knowledge points in each source course to knowledge points in the target course are used as weighting coefficients to weight and aggregate the learner's knowledge state in the source courses, thus obtaining the learner's knowledge state in the target course. The initial level of mastery. The specific process is shown in formula (27).

[0116] (27) in, Representing source course knowledge points To the target course knowledge points The final migration association weight; Indicate learner u For the source course knowledge points The degree of mastery; This represents the target course knowledge points obtained by transferring the learner's knowledge state from the source course. The initial level of mastery.

[0117] The learner's initial knowledge state of the target course is composed of their initial mastery of all the knowledge points in the target course. .

[0118] Step 5: Update the knowledge status of the target course Initial knowledge status of the target course The knowledge state in the cognitive diagnostic model constructed using steps 1 and 2 is initialized, and the cognitive diagnostic model constructed using steps 1 and 2 is trained based on a small number of learner's answer records in the target course from the "learner's historical learning record". Let the learner... u The index set of a small number of answer records for the target course is The loss function of the cognitive diagnostic model for enhanced discrimination in the target course is defined as follows.

[0119] (28) in, Indicate learner u Answer the exercises for the target course The result; Indicate learner u Answer the exercises for the target course correctly. The predicted probability.

[0120] By optimizing the loss function of the cognitive diagnostic model that enhances discrimination in the target course, learners are obtained. u Updated target course knowledge status .

[0121] (iv) The working content of the "Exercise Recommendation Module" is as follows: This module combines the target course knowledge status output by the "Cross-Course Cognitive Diagnosis Module" with the target course exercise-knowledge point association matrix Q from the "Text Embedding Construction Module". T The learner's answer records in the source course and a small number of answer records in the target course, as well as the exercise resources in the target course resource database, are used to generate exercise recommendation results that match the learner's cognitive level and learning preferences through a two-layer prompt word optimization mechanism. This mechanism utilizes a large language model to simultaneously optimize task prompt words and task optimization strategy prompt words.

[0122] Step 1: Generating a set of candidate exercises based on the knowledge status of the target course Receive learners' updated target course knowledge status from the "Cross-Course Cognitive Diagnosis Module". Target course exercises from the "Text Embedding Building Block" - Knowledge Point Association Matrix Q T In addition, there are exercise resources in the "Target Course Resource Database". To avoid the large language model directly generating exercises from the complete question bank that are inconsistent with the learner's cognitive level, we use the learner's updated target course knowledge status from the "Cross-Course Cognitive Diagnosis Module". And the target course exercises - knowledge point association matrix Q T The candidate recommendation score for each exercise in the "Target Course Resource Database" is calculated. The candidate recommendation score for an exercise is calculated based on its coverage of weak knowledge points, its difficulty level, and the learner's prior knowledge. For exercises in the target course... The degree of coverage of its weak knowledge points Difficulty matching level and the learner's prior knowledge level The specific calculation process is as follows.

[0123] 1) Coverage of weak knowledge points

[0124] The coverage of weak knowledge points measures the proportion of weak knowledge points that learners currently need to consolidate, as covered by the exercises in the target course. This is first determined based on the learner's updated knowledge status of the target course in the "Cross-Course Cognitive Diagnosis Module". Identify the set of weak knowledge points of learners in the target course. ,in, Indicate learner u For the target course knowledge points The degree of mastery, This indicates the threshold for identifying weak knowledge points. If... If the set is empty, the set of knowledge points with the lowest level of knowledge in the target course will be selected as the set of knowledge points to be strengthened, ensuring that the candidate exercise selection and recommendation generation process can still be executed. Next, based on the learner's set of weak knowledge points in the target course... The correlation matrix of knowledge points in the target course exercises Q T Obtain the exercises for the target course. Coverage of weak knowledge points The specific calculation method is shown below.

[0125] (29) in, Representing the target course exercises Does it test the knowledge points of the target course? The correlation matrix Q between the target course exercises and knowledge points T get. The larger the value, the better the exercise covers the learner's weaker areas in the target course.

[0126] 2) Difficulty matching level

[0127] Difficulty matching level This is used to measure the degree of match between the difficulty of the target course exercises and the learner's current knowledge level. First, the learner's... u Exercises for the target course The average level of mastery of the knowledge points tested is then combined with the difficulty of the exercises and the set difficulty offset to obtain the target course exercises. Difficulty matching level The specific process is as follows.

[0128] (30) in, Indicates the difficulty level of the exercises; Representing the target course exercises Does it test the knowledge points of the target course? The correlation matrix Q between the target course exercises and knowledge points T get; Indicate learner u For the target course knowledge points The degree of mastery is indicated; This represents the target difficulty offset, used to control the degree to which the difficulty of candidate exercises deviates from the learner's average mastery of the relevant knowledge points. It can be manually preset or adjusted according to the teaching requirements of the target course, the range of exercise difficulty markings, or recommendation strategies. The larger the value, the more closely the difficulty of the exercises matches the learner's current cognitive level.

[0129] 3) The learner's prior knowledge level

[0130] learner's prior knowledge level This is used to measure the extent to which learners possess the prerequisite knowledge required to complete the exercises in this target course. It is based on the target course exercise-knowledge point association matrix Q, passed from the "text embedding building block". T The set of knowledge points examined is as follows: Based on the prerequisites and dependencies within the target course, the exercises for the target course are obtained. The set of knowledge points to be examined The corresponding set of prerequisite knowledge points By calculating the students' prior knowledge set... The average mastery of the knowledge points in the middle school is used to obtain the exercises for the target course. The level of prior knowledge possessed by learners . The larger the value, the more prerequisite knowledge the learner has acquired to support their completion of the exercise. The calculation process is shown below.

[0131] (31) Overall coverage of weak knowledge points Difficulty matching level and the learner's prior knowledge level Obtain the exercises for the target course. Candidate recommendation score The specific process is as follows.

[0132] (32) in, l CW , l DF and l PMThe weighting coefficients, representing the coverage of weak knowledge points, the degree of difficulty matching, and the learner's prior knowledge, are manually set and satisfy the following conditions: l CW + l DF + l PM =1.

[0133] Based on the candidate recommendation scores, the top-ranked exercises are selected from the "Target Course Resource Database". K cand The exercises were selected as candidate exercises, resulting in a candidate exercise set. .

[0134] Step 2: Initial task prompt generation By describing the recommendation task and combining task optimization strategy prompts and prompt generation style constraints, initial task prompts are generated to guide the large language model in generating learning preference summaries and target course top-level suggestions. N Exercise recommendations and explanations.

[0135] In the initial round 0, given the input information for the recommendation task... Disc The initial task optimization strategy prompt word set and prompt generation style constraint set are generated through a large language model. The initial task prompt words are used to guide the large language model in generating learning preference summaries and target course Top-level suggestions. N The exercise recommendation list and explanations, along with initial task optimization strategy prompts, guide the large language model on how to generate task prompts. Initial task prompts. The specific generation process is as follows.

[0136] (33) Among them, LLM init (·) indicates the large language model invocation process used to generate initial task prompts; Disc To provide input for the recommendation task, the task description requires the large language model to generate a summary of learning preferences and a Top-Level list of target courses that match the learner's cognitive level and learning preferences, based on information such as the learner's answer records for the source and target courses, their knowledge status in the target course, and the set of candidate exercises for the target course. N Exercise recommendations and explanations; Indicates the first i Initial task optimization strategy prompts can be set manually or generated by a large language model; ThinkStyle i Indicates the first i The prompt generation style constraints include, for example, emphasizing priority for weak knowledge points, emphasizing difficulty suitability, emphasizing complete explanation structure, or emphasizing clear summary of learning preferences, etc. This indicates the number of prompts in the initial task prompt set.

[0137] Step 3: Evaluation of Task Prompts In the f In each round of iteration, based on the prompts for each task... Recommended task input information and candidate exercise set Generate learning preference summaries and target course top-level summaries through large language models. N The exercise recommendation results and explanations are presented, and the fitness scores for each task cue are calculated. To evaluate the recommendation effectiveness of different task prompt words.

[0138] First, based on the set of learners' weak knowledge points in the target course obtained in step 1... The updated target course knowledge status of learners in the "Cross-Course Cognitive Diagnosis Module" The learner's answer records in the source course and a small number of answer records in the target course are used to construct the input information for the recommended task. ,in, Record u This represents a set of learners' answer records in the source course and a small number of answer records in the target course; Syllabus T This represents the syllabus for the target course, which provides constraints on course organization, such as the chapter structure and teaching sequence.

[0139] Then, the first f Round i Task prompt words Recommended task input information and candidate exercise set Inputting into a large language model yields a summary of learning preferences and a Top-ranked target course. N The recommended exercises and their explanations are presented below.

[0140] (34) in, This represents a summary of learning preferences output by the large language model, used to illustrate learners' learning preferences in terms of question type, difficulty level, and practice pace. Indicates the target course Top- N Exercise recommendations; This refers to the recommended explanatory text, which describes the learner's weak knowledge points covered by the recommended exercises, their relationship with the learner's learning preferences, and the coverage of prerequisite knowledge points for the learner's weak knowledge points; LLM rec(·) indicates the generation of a summary of learning preferences and a Top-level target course. N The process of calling the large language model for exercise recommendation results and recommendation explanations.

[0141] We use commonly used recommendation task evaluation metrics as the core evaluation criteria for task prompt word adaptability, and apply them to evaluate the recommendation effectiveness of different task prompt words. Specifically, in the first... f In the first round, based on the first i Top-level task prompts generate target courses. N Exercise Recommendation Results The recommendation results are then output to learners; subsequently, interactive feedback data, such as actual clicks and responses from learners regarding these recommendations, is collected to form recommendation feedback data. This is based on the target course Top- N Exercise Recommendation Results Based on the recommendation feedback data collected after the recommendation results are output, commonly used recommendation task evaluation metrics NDCG@N, Recall@N, Hit@N, and MAP@N are calculated. By combining the evaluation metrics of each recommendation task, the fitness score of the prompt word in this round of recommendation feedback is obtained. The calculation process of the fitness score is shown below.

[0142] (35) in, , , , The weight coefficients are non-negative and satisfy the following conditions: .

[0143] Step 4: Generating task optimization strategy prompts and sub-task prompts Task optimization strategy prompts are used to generate or rewrite task prompts, making them more suitable for guiding the large language model to mine learning preferences and generate exercise recommendations that better match learners' cognitive levels and learning preferences. In round 0, the initial task optimization strategy prompts... Used to generate initial task prompts; in subsequent rounds, task optimization strategy prompts are used to rewrite the top performers with the highest fitness scores across all rounds. This is a task prompt. The first task prompt word is called the parent task prompt word. The new task prompt word obtained by rewriting the parent task prompt word through the task optimization strategy prompt word is called the child task prompt word, and it will be used as the task prompt word for the next round.

[0144] To guide the large language model in finding effective prompt word rewriting directions, we first construct a historical prompt word rewriting feedback set to record prompt word rewriting feedback that has been evaluated up to the current round. Each prompt word rewriting feedback record stores the parent task prompt word and its task output, the prompt word used to rewrite the parent task's prompt word as an optimization strategy, the rewritten child task prompt word and its task output, the fitness of the parent and child task prompt words, and the difference in fitness between them. f The set of feedback for rewriting the history prompts of the wheel is denoted as It's important to note that since the initial task prompt for round 0 is derived from the initial task optimization strategy prompt, and not rewritten from the parent task prompt, in round 0, It is constructed from the initial task prompt, the corresponding initial task optimization strategy prompt, the task output, and the fitness score.

[0145] In the f After evaluating the task prompts in each round, we select the top-performing prompts from all generated rounds based on their fitness scores. The first task prompt word, as the first f +1 set of parent task prompts to be rewritten Pat (φ+1) .

[0146] Get the first f Wheel history prompt rewriting feedback set and the f +1 set of parent task prompts to be rewritten Pat (φ+1) Then, it is compared with the input information of the recommended task. Disc In the common input large language model, the first generation is generated. f +1 round of task optimization strategy prompts The details are as follows.

[0147] (36) Among them, LLM opt (·) indicates the large language model call process for generating prompts for the next round of task optimization strategies; Indicates the first f +1 set of parent task prompts to be rewritten Pat (φ+1) The Middle i Parent generation task prompt words; Indicates the first f +1 round of task optimization strategy prompts The Middle i The task optimization strategy prompts are for... Pat(φ+1) The Middle i Parent generation task prompts Generated.

[0148] Get the first f +1 round of task optimization strategy prompt Then, it was compared with the first f +1 set of parent task prompts to be rewritten Pat (φ+1) In the common input large language model, the first generation is generated. f +1 round of mission prompts The specific process is as follows.

[0149] (37) Among them, LLM mut (·) indicates the large language model invocation process of rewriting the parent task prompt word according to the task optimization strategy prompt word; Indicates the first f +1 round of mission prompts The i Task prompt words.

[0150] No. f The task prompt words for the +1 round will have their fitness scores calculated according to step 3; then, a new round of task optimization strategy prompt words and task prompt words will be generated according to step 4. The above process continues to iterate until the preset number of iteration rounds or fitness convergence conditions are reached.

[0151] Step 5: Determining the optimal task prompt words and generating recommendation results After the maximum set number of iterations or when the fitness score reaches the set convergence condition, the task prompt word with the highest fitness is selected from all task prompt words generated in all iterations as the optimal task prompt word. This optimal task prompt word is then used in conjunction with the recommended task input information. and candidate exercise set Generate learning preference summaries and target course top-level summaries through large language models. N Exercise recommendations and explanations.

[0152] It should be noted that the determination of the optimal task prompt is based on the fitness evaluation results when the preset maximum number of rounds or fitness convergence condition is met. After the optimal task prompt is used for subsequent recommendations, interactive feedback data such as learners' clicks and responses to recommended exercises can be collected, and the recommendation performance of the optimal task prompt in subsequent recommendation processes can be evaluated based on the interactive feedback data. When further prompt optimization is needed, task optimization strategy prompts and offspring task prompts are generated according to steps 3 and 4. In subsequent rounds, if the newly generated task prompts obtain higher fitness scores, the optimal task prompt is updated.

[0153] In summary, the exercise recommendation method and apparatus combining hypergraph knowledge modeling and automatic prompting optimization provided in this application utilizes a large language model to analyze the semantic relationships of knowledge points between courses and constructs a cross-course heterogeneous hypergraph. This fully characterizes various cross-course knowledge relationships, such as single-source single-objective transfer relationships and multi-source single-objective transfer relationships, as well as intra-course knowledge relationships such as prerequisite dependencies, concept similarity, and exercise-knowledge point associations. Compared to existing cross-domain cognitive diagnostic methods, the exercise recommendation method and apparatus combining hypergraph knowledge modeling and automatic prompting optimization proposed in this application, by combining a large language model to analyze knowledge point-related text between courses and constructing a cross-course heterogeneous hypergraph, more fully characterizes the complex knowledge relationships between and within courses, improving the effect of knowledge state transfer. This allows for a more accurate understanding of the user's knowledge state at the beginning of the course, thereby recommending exercises that better meet the user's needs.

[0154] This application's embodiments introduce a historical interaction-enhanced discrimination modeling method into the cognitive diagnostic model. This enhances the model's ability to model the differences in answering advantages among different user-practice combinations for the same knowledge point, thereby improving the accuracy of cognitive diagnostic results and consequently, the accuracy of exercise recommendations. Compared to existing cognitive diagnostic models, this application's embodiments, when constructing the interactive discrimination vector to characterize the non-linear dependency relationship between users and exercises, introduce answering information from other questions related to each exercise's knowledge. Based on this knowledge-related answering information, a historical interaction context representation is constructed. This historical interaction context representation is then used to correct the interactive discrimination vector, thereby enhancing the model's ability to model the differences in answering advantages among different user-practice combinations for the same knowledge point. This results in more accurate cognitive diagnostic results, leading to the recommendation of exercises that better match the user's cognitive level.

[0155] This application introduces an automatic prompting optimization method based on a two-layer prompt word optimization mechanism. It iteratively mines learner learning preferences using optimal task prompt words and recommends exercise resources that match learners' cognitive levels and learning preferences. Compared to existing recommendation methods based on automatic prompting optimization, this application not only optimizes the task prompt words used to mine learning preferences and make recommendations, but also optimizes the task optimization strategy prompt words used to generate task prompt words. This reduces the limitations imposed by preset task optimization strategy prompt words on task prompt word adjustments, making task prompt words more suitable for guiding the large language model to recommend exercise resources that better match learners' cognitive levels and learning preferences, thus providing learners with personalized exercise recommendation services.

[0156] This application embodiment also provides an exercise recommendation device, including: The data acquisition module is used to acquire the exercises, knowledge points, exercise-knowledge point association matrix, knowledge point prerequisite dependencies, user answer records in each course, user representation, exercise representation of each exercise in each course, and user's initial course knowledge state vector in the source course. The multiple courses include source courses and target courses. The initial transfer association matrix determination module is used for multiple courses, taking the knowledge points and exercises of multiple courses as nodes. Based on the association matrix between each exercise and knowledge point and the prerequisite dependencies of each knowledge point, it determines the prerequisite dependency hyperedge relationships and similarity hyperedge relationships between knowledge points within each course, the association hyperedge relationships between exercises and knowledge points within each course, the cross-course prerequisite knowledge point hyperedge relationships between knowledge points in the source course and knowledge points in the target course, and obtains the cross-course heterogeneous hypergraph association matrix and hyperedge weight matrix based on each hyperedge relationship. It also obtains the node degree matrix based on the cross-course heterogeneous hypergraph association matrix and hyperedge degree matrix, and finally determines the initial transfer association matrix between the knowledge points of the source course and the knowledge points of the target course based on the cross-course prerequisite knowledge point hyperedge relationships. The final migration association weight determination module is used to perform multiple rounds of iterative updates on the initial feature representation of each node based on the cross-course heterogeneous hypergraph association matrix, hyperedge weight matrix, node degree matrix, and hyperedge degree matrix to obtain the higher-order node representation of each node. Based on the initial migration association matrix, the module corrects the association weight of the higher-order node representation corresponding to each knowledge point to obtain the final migration association weight of each knowledge point from the source course to the target course. The discrimination vector determination module is used to determine the historical interaction context representation between the user and each exercise for each course based on the course's exercise-knowledge point association matrix and the course's answer records. It then combines the user representation, the exercise representation, and the historical interaction context representation to obtain the discrimination vector enhanced by the user's historical interaction with each exercise. The initial course knowledge state vector determination module is used to optimize the initial course knowledge state vector of the source course based on the discrimination vector enhanced by the user's historical interaction with each exercise in the source course, the difficulty vector of each exercise, the association matrix between exercises and knowledge points, and the answer records. This results in the post-training course knowledge state vector of the source course. Based on the post-training course knowledge state vector and the final transfer association weights of knowledge points from each source course to the target course, the initial course knowledge state vector of the target course is determined. The first knowledge state vector determination module is used to optimize the initial course knowledge state vector of the target course based on the discrimination vector enhanced by the user's historical interaction with each exercise in the target course, the difficulty vector of each exercise in the target course, the association matrix between the exercises and knowledge points in the target course, and the user's answer records in the target course, so as to obtain the first knowledge state vector of the target course. The exercise recommendation module is used to determine the recommendation score of each exercise in the target course relative to the user based on the association matrix between exercises and knowledge points, the first knowledge state vector, and the prerequisite dependencies of knowledge points. It also processes a preset number of exercises with the highest recommendation scores and multiple task prompt words based on a large language model to obtain exercise recommendation results for different target tasks. Among them, the task prompt words are obtained through iterative optimization based on the user's interaction with the historical exercise recommendation results.

[0157] It should be noted that the information interaction and execution process between the above-mentioned devices / modules are based on the same concept as the method embodiments of this application. Their specific functions and technical effects can be found in the method embodiments section, and will not be repeated here. Those skilled in the art will understand that, for ease of description and brevity, the division of the above-mentioned functional units and modules is only used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the modules in the above-mentioned devices can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0158] This application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the exercise recommendation method that combines hypergraph knowledge modeling and automatic prompting optimization as provided in this application.

[0159] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the exercise recommendation method combining hypergraph knowledge modeling and automatic prompting optimization as provided in this application.

[0160] The processor mentioned above can be a central processing unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), or field-programmable gate arrays (FPGAs). The memory can be a programmable gate array (GGEA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor. In some embodiments, the memory can be an internal storage unit of the device, such as the device's hard drive or RAM. In other embodiments, the memory can be an external storage device of the device, such as a plug-in hard drive, smart media card (SMC), secure digital card (SD), flash card, etc. Furthermore, the memory can include both internal and external storage units of the device. The memory is used to store the operating system, applications, bootloader, data, and other programs, such as the program code of a computer program. The memory can also be used to temporarily store data that has been output or will be output.

[0161] This application provides a computer program product that, when run on a terminal device, enables the terminal device to implement the steps described in the various method embodiments.

[0162] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of protection of this application is limited to these examples; under the concept of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of one or more embodiments of this application as described above, which are not provided in detail for the sake of brevity.

[0163] One or more embodiments in this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of this application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments in this application should be included within the protection scope of this application.

Claims

1. A problem recommendation method combining hypergraph knowledge modeling and automatic suggestion optimization, characterized in that, include: Obtain the exercises, knowledge points, exercise-knowledge point association matrix, knowledge point prerequisite dependencies, user's answer records in each course, user representation, exercise representation of each exercise in each course, and the user's initial course knowledge state vector in the source course, wherein the multiple courses include the source course and the target course. For the multiple courses, the knowledge points and exercises of the multiple courses are used as nodes. Based on the association matrix between each exercise and knowledge point and the prerequisite dependency relationship of each knowledge point, the prerequisite dependency hyperedge relationship and knowledge point similarity hyperedge relationship between knowledge points in each course are determined. The association hyperedge relationship between each exercise and knowledge point in each course is also determined. The cross-course prerequisite knowledge point hyperedge relationship between the knowledge points in the source course and the knowledge points in the target course is determined. Based on each hyperedge relationship, the cross-course heterogeneous hypergraph association matrix and the hyperedge weight matrix are obtained. Based on the cross-course heterogeneous hypergraph association matrix and the hyperedge weight matrix, the node degree matrix is ​​obtained. Based on the cross-course heterogeneous hypergraph association matrix, the hyperedge degree matrix is ​​obtained. Based on each cross-course prerequisite knowledge point hyperedge relationship, the initial transfer association matrix between the knowledge points of the source course and the knowledge points of the target course is determined. Based on the cross-course heterogeneous hypergraph association matrix, the hyperedge weight matrix, the node degree matrix, and the hyperedge degree matrix, the initial feature representation of each node is iteratively updated in multiple rounds to obtain the higher-order node representation of each node. Based on the initial migration association matrix, the association weights of the higher-order node representations corresponding to each knowledge point are corrected to obtain the final migration association weights of the knowledge points from the source course to the target course. For each of the courses, based on the association matrix between the exercises and knowledge points of the course and the answer records of the course, the historical interaction context representation of the user with each exercise is determined. Combining the user representation, the representation of each exercise, and the historical interaction context representation, a discriminative vector for the user's historical interaction with each exercise is obtained. Based on the discrimination vector enhanced by the user's historical interaction with each exercise in the source course, the difficulty vector of each exercise, the association matrix between the exercise and the knowledge point, and the answer record, the initial course knowledge state vector of the source course is optimized to obtain the post-training course knowledge state vector of the source course. Based on the post-training course knowledge state vector and the final transfer association weight of the knowledge points from each source course to the target course, the initial course knowledge state vector of the target course is determined. The initial course knowledge state vector of the target course is optimized based on the user's historical interaction enhancement of each exercise in the target course, the difficulty vector of each exercise in the target course, the association matrix between the exercises and knowledge points in the target course, and the user's answer records in the target course, to obtain the first knowledge state vector of the target course. Based on the association matrix between exercises and knowledge points, the first knowledge state vector, and the prerequisite dependencies of knowledge points in the target course, the recommendation score of each exercise in the target course relative to the user is determined. Then, based on the large language model, a preset number of exercises with the highest recommendation scores and multiple task prompt words are processed to obtain exercise recommendation results for different target tasks. Among them, the task prompt words are obtained by iterative optimization based on the user's interaction with the historical exercise recommendation results.

2. The exercise recommendation method according to claim 1, characterized in that, The process of using a large language model to process a preset number of exercises with the highest recommendation scores and multiple task prompts yields exercise recommendation results for different target tasks, including: Obtain task prompts and recommended task input information for multiple target tasks; Based on the large language model, the multiple task prompts, the recommended task input information, and the preset number of exercises are processed to obtain the exercise recommendation results for each target task.

3. The exercise recommendation method according to claim 2, characterized in that, Also includes: Based on the user's interaction with the recommended results of each exercise, the recommendation fitness corresponding to each task prompt word is determined; Select a preset number of task prompt words with the highest recommended fitness to obtain a set of parent task prompt words, and obtain prompt word rewriting feedback for each task prompt word in the set of parent task prompt words. The prompt word rewriting feedback includes: the parent task prompt words and their exercise recommendation results for the task prompt words in the set of parent task prompt words, the task prompt words and their exercise recommendation results in the set of parent task prompt words, the recommended fitness of the parent task prompt words and the task prompt words in the set of parent task prompt words, and the fitness difference between the two. Based on the large language model, the set of prompt words for the parent task, the prompt word rewriting feedback for each task prompt word, and the input information of the recommended task are processed to obtain the set of prompt words for the next generation task optimization strategy. Based on the large language model, the next-generation task optimization strategy prompt word set and the parent task prompt word set are processed to obtain next-generation task prompt words for multiple target tasks, which are then used to generate exercise recommendation results for each target task in the next iteration.

4. The exercise recommendation method according to claim 1, characterized in that, The process of optimizing the initial course knowledge state vector of the source course based on the user's historical interaction enhancement discrimination vector for each exercise in the source course, the difficulty vector of each exercise, the association matrix between exercises and knowledge points, and the answer records, to obtain the post-trained course knowledge state vector of the source course, and determining the initial course knowledge state vector of the target course based on the post-trained course knowledge state vector and the final transfer association weights of knowledge points from each source course to the target course, includes: For the source course, the cognitive diagnostic model to be trained determines the predicted probability of the user answering each question in the source course based on the discrimination vector of the user's historical interaction enhancement for each question in the source course, the initial course knowledge state vector of the source course, the difficulty vector of each question in the source course, and the association matrix between the questions and knowledge points in the source course. The cognitive diagnostic model to be trained is trained based on the predicted correct answer probability of each exercise in the source course and the binary cross-entropy loss function value determined by the user's answer record in the source course, so as to obtain the trained cognitive diagnostic model and output the trained course knowledge state vector of the source course. The initial course knowledge state vector of the target course is determined based on the post-training course knowledge state vector of the source course and the final transfer association weight of the knowledge points from each source course to the target course.

5. The exercise recommendation method according to claim 1, characterized in that, The process of optimizing the initial course knowledge state vector of the target course based on the user's historical interaction enhancement of each exercise in the target course, the difficulty vector of each exercise in the target course, the association matrix between exercises and knowledge points in the target course, and the user's answer records in the target course, to obtain the first knowledge state vector of the target course, includes: For the target course, the cognitive diagnostic model to be trained determines the predicted probability of the user answering each question in the target course based on the discrimination vector of the user's historical interaction enhancement for each question in the target course, the initial course knowledge state vector of the target course, the difficulty vector of each question in the target course, and the association matrix between the questions and knowledge points in the target course. The cognitive diagnostic model to be trained is trained based on the predicted correct answer probability of each exercise in the target course and the binary cross-entropy loss function value determined by the user's answer record in the target course, and the trained cognitive diagnostic model is obtained and the first knowledge state vector is output.

6. The exercise recommendation method according to claim 1, characterized in that, For each of the courses, based on the association matrix between exercises and knowledge points of the course and the answer records of the course, a historical interaction context representation of the user with each exercise is determined. Then, by combining the user representation, the representation of each exercise, and the historical interaction context representation, a discriminative vector for enhancing the user's historical interaction with each exercise is obtained, including: For each of the courses, based on the association matrix between the exercises and knowledge points of the course, a set of relevant knowledge point answer records of the user and each exercise of the course is determined in the answer records of the course, the historical answer representation of the user for each question in the relevant knowledge point answer record set is determined, and the attention weight of each question relative to the exercise is determined; For each of the exercises, the historical interaction context representation of the exercise is determined based on the historical answer representation and attention weight of each exercise; Determine the initial interactive discriminative vector of the user representation for each exercise representation, and modify the initial interactive discriminative vector according to the historical interaction context representation to obtain the historical interaction-enhanced discriminative vector of the user for each exercise.

7. The exercise recommendation method according to claim 6, characterized in that, The step of determining the initial interactive discriminative vector of the user representation for each of the exercise representations, and correcting the initial interactive discriminative vector based on the historical interaction context representation to obtain the user's historical interaction-enhanced discriminative vector for each exercise includes: Determine the initial interactive discriminative vector of the user representation for each exercise representation, and obtain the historical correction vector based on the historical interactive context representation; Based on the gating mechanism, the initial interactive discrimination vector and the historical correction vector are fused to obtain the historical interaction-enhanced discrimination vector for each exercise.

8. The exercise recommendation method according to claim 1, characterized in that, The step of determining the recommendation score of each exercise in the target course relative to the user based on the association matrix between exercises and knowledge points, the first knowledge state vector, and the prior knowledge dependency relationship of the target course includes: Based on the association matrix between exercises and knowledge points of the target course and the first knowledge state vector, determine the degree of suitability of each exercise in the target course for the user, and combine the prerequisite dependency relationship of the knowledge points to determine the degree of prerequisite knowledge possessed by the user for the exercises. For each exercise, the recommended score for that exercise is determined based on the exercise's suitability and the level of prior knowledge required.

9. A problem recommendation device combining hypergraph knowledge modeling and automatic prompting optimization, characterized in that, include: The data acquisition module is used to acquire the exercises, knowledge points, exercise-knowledge point association matrix, knowledge point prerequisite dependencies, user answer records in each of the multiple courses, user representation, exercise representation of each exercise in each of the courses, and the user's initial course knowledge state vector in the source course, wherein the multiple courses include the source course and the target course. The initial migration association matrix determination module is used to, for the multiple courses, take the knowledge points and exercises of the multiple courses as nodes, and determine the prerequisite dependency hyperedge relationships and similarity hyperedge relationships between knowledge points in each course, the association hyperedge relationships between exercises and knowledge points in each course, the cross-course prerequisite knowledge point hyperedge relationships between knowledge points in the source course and knowledge points in the target course, obtain the cross-course heterogeneous hypergraph association matrix and hyperedge weight matrix based on the hyperedge relationships, obtain the node degree matrix based on the cross-course heterogeneous hypergraph association matrix and the hyperedge weight matrix, obtain the hyperedge degree matrix based on the cross-course heterogeneous hypergraph association matrix, and determine the initial migration association matrix between the knowledge points of the source course and the knowledge points of the target course based on the cross-course prerequisite knowledge point hyperedge relationships. The final migration association weight determination module is used to perform multiple rounds of iterative updates on the initial feature representation of each node based on the cross-course heterogeneous hypergraph association matrix, the hyperedge weight matrix, the node degree matrix, and the hyperedge degree matrix to obtain the higher-order node representation of each node, and to correct the association weight of the higher-order node representation corresponding to each knowledge point based on the initial migration association matrix to obtain the final migration association weight of each knowledge point from the source course to the target course. The discrimination vector determination module is used to determine the historical interaction context representation of the user with each exercise based on the association matrix between the exercises and knowledge points of the course and the answer records of the course, and to obtain the discrimination vector enhanced by the historical interaction of the user with each exercise by combining the user representation, the representation of each exercise and the historical interaction context representation. The initial course knowledge state vector determination module is used to optimize the initial course knowledge state vector of the source course based on the user's historical interaction enhancement discrimination vector for each exercise in the source course, the difficulty vector of each exercise, the association matrix between exercises and knowledge points, and the answer records, to obtain the post-training course knowledge state vector of the source course. Based on the post-training course knowledge state vector and the final transfer association weight of knowledge points from each source course to the target course, the initial course knowledge state vector of the target course is determined. The first knowledge state vector determination module is used to optimize the initial course knowledge state vector of the target course based on the user's historical interaction enhancement discrimination vector for each exercise in the target course, the difficulty vector of each exercise in the target course, the association matrix between exercises and knowledge points in the target course, and the user's answer records in the target course, so as to obtain the first knowledge state vector of the target course. The exercise recommendation module is used to determine the recommendation score of each exercise in the target course relative to the user based on the association matrix between exercises and knowledge points, the first knowledge state vector, and the prior knowledge dependency relationship of the target course. It then processes a preset number of exercises with the highest recommendation scores and multiple task prompt words based on a large language model to obtain exercise recommendation results for different target tasks. The task prompt words are obtained through iterative optimization based on the user's interaction with historical exercise recommendation results.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the exercise recommendation method as described in any one of claims 1 to 8, which combines hypergraph knowledge modeling and automatic prompting optimization.

Images