Course recommendation method and system based on large language model decoupling modeling

Abstract

The invention relates to a course recommendation method and system based on large language model decoupling modeling, and the method comprises the steps: defining a prompt template containing a generation task and a required output format, and generating a course sequence from the format prompt template and a course name through an LLM model; curriculum names and corresponding relation descriptions in the curriculum sequence are extracted, a curriculum relation graph is constructed, the curriculum relation graph is input into an LLM model, and knowledge-enhanced embedded vectors are obtained; acquiring a historical record sequence of the learner, converting the historical record sequence into a historical feature sequence, calculating a dot product between historical feature embedding of the target learner and historical embedding of other learners based on the historical feature sequence, and acquiring a set of the learner with the highest score; and obtaining a historical feature sequence of similar learners based on the set of the learners with the highest score, integrating the historical feature sequence with an interaction sequence of the target learner, inputting an integration result into an LLM model, and obtaining a future course embedding vector.

Description

Technical Field

[0001] The present invention relates to the technical field of machine learning, and in particular to a course recommendation method and system based on decoupled modeling of large language models. Background Art

[0002] The goal of a course recommendation system is to help learners efficiently screen through a vast amount of courses based on historical interactions and behavioral characteristics, thereby providing personalized course selection suggestions. This can not only reduce the decision-making burden caused by information overload, but also improve learning efficiency and promote continuous learning. To meet the growing demand for accurate prediction of the next course, more and more researchers are dedicated to developing course recommendation frameworks for intelligent tutoring systems and MOOC platforms. Current course recommendation models can be roughly divided into three categories. (1) Collaborative filtering-based models use user interaction matrices to infer learners' preference patterns through neighborhood-based similarity calculations. (2) Deep learning-based architectures employ complex neural networks to model the multi-relational dependencies between learners and course resources. (3) Reinforcement learning-based models utilize experiential learning to iteratively improve and optimize their recommendation strategies, thereby gradually enhancing their ability to provide personalized and effective course recommendations. Despite these significant research achievements, there are still some challenges in actual online education applications that deserve further exploration.

[0003] First, online education systems face continuous challenges caused by long-tail distributions and cold-start problems, mainly due to insufficient historical behavior data. The effectiveness of personalized course recommendations fundamentally depends on accurately modeling user interaction patterns to predict future engagement. However, in practical applications, most learners only interact with a limited number of courses, resulting in inherent interaction sparsity. The scarcity of historical data severely hinders the capture of learners' preferences, making effective model training challenging.

[0004] Second, most existing recommendation models mainly rely on course identifiers to model user history, which limits the effective extraction and utilization of course relationships in actual educational applications. This limitation arises because courses usually exhibit various structural relationships that contribute to a more comprehensive understanding of user behavior and play a crucial role in improving recommendation effectiveness. For example, before learning "Data Structures", it is recommended to first learn "Discrete Mathematics", "Fundamentals of Computers", and "Fundamentals of Programming Languages" to build a solid theoretical and practical foundation. Therefore, overly simplistic course identifier representations ignore course relevance and structural relationships, thereby weakening the predictive ability of future course recommendation models.

[0005] The latest advancements in large language models (LLMs) have demonstrated remarkable semantic representation capabilities and cross-domain knowledge retention abilities, revolutionizing the learning paradigm in the field of AI research. However, the comprehensive potential of LLMs in educational recommendation systems remains largely untapped, particularly in the context of course recommendation architectures. Additionally, despite their excellent performance in natural language processing tasks, directly applying LLMs to practical course recommendation scenarios seems unrealistic. On the one hand, LLMs struggle to meet the personalized needs of learners. They typically default to recommending popular courses rather than providing truly personalized suggestions. On the other hand, the effectiveness of LLMs-based recommendations is further hindered by the sparsity of interaction data. Summary of the Invention

[0006] To address the problems existing in the above-mentioned prior art, the objective of the present invention is to propose a course recommendation method and system based on decoupled modeling of large language models for a dual-channel architecture in course recommendation: an LLM-enhanced learner and course decoupled modeling framework. It has two synergistic components: (1) Knowledge-enhanced LLM-based course modeling that integrates potential relationships between different courses to obtain enhanced embedding vectors; (2) Interaction-enhanced LLM-based learner modeling that utilizes more historical interactions from similar learners to simulate cold-start learner interactions.

[0007] To achieve the above objective, the present invention provides the following solutions:

[0008] A course recommendation method based on decoupled modeling of large language models, comprising:

[0009] Define a prompt template containing generation tasks and the required output format, and use the LLM model to generate a course sequence from the format prompt template and course names;

[0010] Extract the course names and corresponding relationship descriptions in the course sequence, construct a course relationship graph, and input the course relationship graph into the LLM model to obtain knowledge-enhanced embedding vectors;

[0011] Obtain the historical record sequence of the learner, convert the historical record sequence into a historical feature sequence, and based on the historical feature sequence, calculate the dot product between the historical feature embedding of the target learner and the historical embeddings of other learners to obtain a set of learners with the highest scores;

[0012] Based on the set of learners with the highest scores, obtain the historical feature sequences of similar learners, integrate them with the interaction sequence of the target learner, and input the integration result into the LLM model to obtain future course embedding vectors.

[0013] Optionally, generating the course sequence includes:

[0014] S i = LLMP,c i = c i ,s i ;

[0015] Among them, c i represents the name of the course, s i represents the generated course sequence, that is, the course relationship description includes prerequisite conditions and advanced paths, and P represents the format hint template.

[0016] Optionally, constructing the course relationship graph includes:

[0017] Regarding each course in the course name as a central node, constructing a directed graph around the central node, and taking the directed graph as the course relationship graph Among them, represents the set of courses related to each course, represents establishing directed edges between courses based on the relationship description to form an edge set.

[0018] Optionally, the edge from node v i to node v j in the directed graph is represented as: Among them, node v i is the prerequisite course of node v j , and node v j is the advanced course of node v i .

[0019] Optionally, obtaining the knowledge-enhanced embedding vector includes:

[0020]

[0021] Among them, M c is a specific merging strategy on the final hidden state of the course LLM to generate an enhanced representation of the course, T i is the simplified text representation of the course relationship graph, is the real number field, and d is the dimension of the hidden layer.

[0022] Optionally, obtaining the set of the highest-scoring learners includes:

[0023]

[0024] Among them, j = 1, 2,..., |U|, K is a hyperparameter, T is the transpose, and H i represents the historical feature sequence of user u i , and H j represents the historical feature sequence of user u j .

[0025] Optionally, obtaining the future course embedding vector includes:

[0026]

[0027] wherein, represents the historical feature sequence of K learners similar to the target learner u i e i,n+1 represents the embedding vector of the course predicted by the LLM for the future interaction of u i and M u represents the user LLM.

[0028] Optionally, the method further includes:

[0029] Taking the embedding of generating the next recommended course based on the historical interaction embedding of the target learner as the training objective, and obtaining the predicted interest score for each course:

[0030]

[0031] wherein, W and b are trainable parameters;

[0032] Determining the minimized prediction response according to the predicted interest score:

[0033]

[0034] Optimizing the learning process of the model through the cross-entropy loss between the minimized prediction response and the true response:

[0035]

[0036] wherein, y i is the true response.

[0037] To achieve the above object, the present invention also provides a course recommendation system based on decoupled modeling of a large language model, including:

[0038] A course modeling unit, configured to define a prompt template including a generation task and a required output format, and use the LLM model to generate a course sequence from the format prompt template and the course name;

[0039] Extracting the course name and the corresponding relationship description in the course sequence, constructing a course relationship graph, and inputting the course relationship graph into the LLM model to obtain a knowledge-enhanced embedding vector;

[0040] A learner modeling unit is used to obtain a historical record sequence of a learner, convert the historical record sequence into a historical feature sequence, and based on the historical feature sequence, calculate the dot product between the historical feature embedding of the target learner and the historical embedding of other learners to obtain a set of learners with the highest score;

[0041] The preference prediction unit is used to obtain the historical feature sequence of similar learners based on the set of highest-scoring learners, integrate it with the interaction sequence of the target learner, input the integration result into the LLM model, and obtain the future course embedding vector.

[0042] The beneficial effects of the present invention are:

[0043] The proposed model consistently achieves significant performance improvements on both the MOOCCube and MOOCCourse datasets. The most significant improvements are seen on the MOOCCube dataset, where performance improvements of up to 7.3% on the HR@5 metric, up to 8.9% on the NDCG@5 metric, and up to 12.1% on the NDCG@10 metric are achieved over the state-of-the-art baselines. In contrast, on the MOOCCourse dataset, performance improvements of up to 11.6% on the NDCG@5 metric and 6.8% on the NDCG@10 metric are achieved.

[0044] Compared with traditional recommendation models, our LLM-based model performs well on both datasets. In addition, our model shows better improvement over the state-of-the-art baseline HHCoR. This enhancement is attributed to the effective use of LLM's rich world knowledge, semantic representation ability, and reasoning ability.

[0045] Compared with the LLM in the zero-shot setting, our model consistently outperforms the state-of-the-art GPT-4o on both datasets. Notably, on the MOOCCube dataset, our model achieves up to 151% performance improvement on the HR@5 metric, 140% performance improvement on the HR@10 metric, 104% performance improvement on the NDCG@5 metric, and 80% performance improvement on the NDCG@10 metric. This confirms that our model can effectively address the challenges inherent in directly applying LLM to course recommendation tasks, specifically, they fail to meet the needs of personalized learners and the adverse effects of sparse interaction data.

[0046] Compared with the fine-tuned LLM, the proposed model shows obvious advantages, which further verifies the effectiveness of the dual-channel curriculum and learner modeling approach, which fully utilizes the comprehensive potential of LLM to enhance personalized recommendation capabilities.

[0047] Therefore, the results confirm that the model of the present invention effectively alleviates the limitations brought by the sparsity of learner interaction data and the isolation of courses, thus achieving more accurate and effective course recommendations. Description of the Drawings

[0048] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0049] Figure 1 It is a flowchart of a course recommendation method based on decoupled modeling of large language models according to an embodiment of the present invention. Detailed Embodiments

[0050] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0051] To make the above objects, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the drawings and specific embodiments.

[0052] The rapid development of artificial intelligence (AI) is driving changes in various fields of society. Similarly, in the field of education and learning, AI is redefining the traditional paradigm by providing innovative tools to enhance the learning experience and outcomes. For example, the emergence of massive open online courses (MOOCs) has contributed to the creation of a vast repository of shared educational content, effectively addressing the issue of unequal educational opportunities. However, such abundant resources also pose a major challenge, namely how to effectively identify the courses that best meet individual needs and goals among numerous courses. This challenge has prompted researchers to explore the use of AI technologies to develop personalized course recommendation systems, aiming to promote a more efficient and tailored learning experience and ultimately build a student-centered intelligent education ecosystem. Although previous course recommendation models have achieved encouraging results, they still face limitations in practical applications due to the sparsity of learner-course interaction data and the isolation of course items. The remarkable success of large language models (LLMs) in various fields has inspired the exploration of integrating them into course recommendation systems to further optimize the personalized recommendation function.

[0053] As Figure 1As shown, this embodiment discloses a course recommendation method based on decoupled modeling of large language models, including: defining a format prompt template containing generation tasks and required output formats, and using the LLM model to generate a course sequence from the format prompt template and course names; extracting the course names and corresponding relationship descriptions in the course sequence, constructing a course relationship graph, and inputting the course relationship graph into the LLM model to obtain knowledge-enhanced embedding vectors; obtaining the historical record sequence of the learner, converting the historical record sequence into a historical feature sequence, and based on the historical feature sequence, calculating the dot product between the historical feature embedding of the target learner and the historical embeddings of other learners to obtain a set of learners with the highest scores; based on the set of learners with the highest scores, obtaining the historical feature sequences of similar learners, integrating them with the interaction sequence of the target learner, and inputting the integration result into the LLM model to obtain future course embedding vectors.

[0054] Specifically: Figure 1 Figure 5 shows the overall framework of the LLM-enhanced learner and course decoupled modeling model, which contains two key components: LLM-based knowledge-enhanced course modeling and LLM-based interaction-enhanced learner modeling. First, use the LLM to generate relevant course knowledge sequences for each course. By combining each course with the generated enhanced knowledge, a course relationship graph is constructed and used as the input to the course LLM to obtain enhanced course embeddings. Second, according to the historical sequence of the learner, learners with similar behaviors are filtered out, where the historical records are represented using the enhanced embeddings obtained from course modeling. Then, the learner-course records of similar learners are concatenated to the end of the target learner sequence. This enhanced sequence is provided as input to the learner LLM to generate course recommendations. Finally, the predicted embeddings are combined with the target embeddings and passed through a prediction head to obtain the final output. The detailed steps are outlined as follows:

[0055] Furthermore, generating the course sequence includes:

[0056] S i = LLMP,c i = c i ,s i ;

[0057] where c i represents the name of the course, s i represents the generated course sequence, that is, the course relationship description includes prerequisites and progression paths, and P represents the format prompt template.

[0058] Furthermore, constructing the course relationship graph includes:

[0059] Regarding each course in the course name as a central node, a directed graph is constructed around the central node, and the directed graph is used as the course relationship graph Among them, represents the set of courses related to each course, represents establishing a directed edge between courses based on the relationship description to form an edge set.

[0060] Furthermore, in the directed graph, the edge pointing from node v i to node v j is represented as: Among them, node v i is the prerequisite course of node v j , and node v j is the advanced course of node v i .

[0061] Furthermore, obtaining the knowledge-enhanced embedding vectors includes:

[0062]

[0063] Among them, M c is a specific merging strategy on the final hidden state of the course LLM to generate an enhanced representation of the course, and T i is the simplified text representation of the course relationship graph.

[0064] Specifically: Course modeling based on LLM:

[0065] First, use the course modeling unit based on LLM to generate embedding vectors for each course.

[0066] (1) Relationship establishment: To directly solve the problem of course isolation, this embodiment uses an LLM with extensive world knowledge as an extractor of course structure relationships to enrich the connections between courses from the perspective of natural language. This method not only provides valuable and detailed information but also reveals potential inter-course relationships, helping learners plan their learning paths more effectively. Specifically, this embodiment first defines a prompt template P that includes a generation task and the required output format. Then, use the prompt and the given course name c i to call the LLM to generate a course sequence S i , that is, S i =LLM(P, c i =c i , s i . Among them, c i represents the name of the course, and s iRepresents the generated course relationship description, including prerequisite courses and advanced paths. Finally, the generated results will be reviewed by domain experts, and their feedback will be used to iteratively improve the prompt template P and model parameters, thereby enhancing the accuracy of the course sequence and ensuring consistency with the established teaching logic.

[0067] Define the generation task and the required output format, including:

[0068] The prompt template P for the generation task and the required output format is as follows: "Role: As a professional education consulting expert, be able to correctly select the prerequisite courses and subsequent courses for a course:

[0069] Rule:

[0070] 1. Based on the given course name, accurately select the prerequisite courses for this course from the following courses, that is, the courses that need to be completed before taking this course; at the same time, correctly select the subsequent courses for this course, that is, the courses that can be continued after taking this course.

[0071] 2. The return must be in a standard JSON format and must include the "prerequisite" and "subsequent" fields. If a course has no prerequisite courses, then the value of "prerequisite" is an empty JSONArray. If a course has no subsequent courses, then the value of "subsequent" is also an empty JSONArray.

[0072] 3. The values of the prerequisite courses and subsequent courses for this course must be selected from the following Choosen List. Absolutely no course names can be generated, otherwise severe punishment will be received.

[0073] ChoosenList:

[0074] {course_names};

[0075] OutputFormat:

[0076] {{

[0077] "prerequisite": ["coursename1", "coursename2",.....],

[0078] "subsequent": ["coursename1", "coursename2",.....]

[0079] }}

[0080] ThinkStepbyStep:

[0081] 1. Select the prerequisite courses of the target course from the ChoosenList according to the target course given by the user, and assemble them into a list;

[0082] 2. Select the follow-up courses of the target course from the ChoosenList according to the target course given by the user, and assemble them into a list;

[0083] 3. Assemble them correctly into a JSON for return.

[0084] (2) Relationship graph: From the structured course sequence S generated in the previous step i In this embodiment, the course name c i and its corresponding relationship description s i are extracted. The relationship description is the expression of other courses related to the course name c i and the relationships between them. These elements are the basis for constructing the formal course relationship graph. Each course c i is regarded as a central node, and a directed graph is constructed around this node Among them, represents the set of courses related to c i where |V c | ≤ 5 is a constraint designed to adapt to the computational complexity of the LLM. Based on the extracted relationship description, this embodiment establishes directed edges between courses to form an edge set The edge pointing from node v i to node v j can be expressed as This means that v i is the prerequisite course of v j , and v j is the advanced course of v i . The adjacency matrix of graph G c is represented as The adjacency matrix is the digital mapping of graph G c , describing the relationships between each vertex (i.e., course) in the graph. The simplified text representation T c of graph G i encapsulates the node, edge, and adjacency matrix information. Among them, if there is a relationship between node v i and node v j , then Otherwise, it is 0. Through the above steps, a comprehensive course relationship graph is constructed to capture the structural interconnection between courses, providing a solid foundation for subsequent recommendation tasks.

[0085] (3) Course Encoding: Through the above steps, this embodiment derives the form graph G i of course c c and the simplified text representation T i , which encapsulates node, edge, and adjacency matrix information. This embodiment uses an LLM with excellent text understanding ability as a feature extractor. Input T i into the parameterized LLM to generate a knowledge-enhanced embedding vector e i for course c c :

[0086]

[0087] where M c refers to a specific merging strategy on the final hidden state of the course LLM to generate an enhanced representation of the course.

[0088] Further, obtaining the set of the highest-scoring learners includes:

[0089]

[0090] where j = 1, 2, …, |U| and K is a hyperparameter.

[0091] Further, obtaining the future course embedding vectors includes:

[0092]

[0093] where represents the historical feature sequence of K learners similar to the target learner u i , and e i,n+1 represents the embedding vector of the course predicted by the LLM for the future interaction of u i .

[0094] Specifically: Learner Modeling Based on LLM:

[0095] Using the knowledge extraction function of the LLM, each course is added an extended course sequence containing rich relationship information. The enhanced embedding vector generated by the knowledge-enhanced course modeling unit lays a solid foundation for subsequent learner preference prediction. The interaction-enhanced learner modeling unit is introduced in detail below.

[0096] (1) Filtering Similar Learners: To address the problem of sparse interaction data between learners and courses in the course recommendation system, this embodiment implements a filtering mechanism to identify similar learners for the target user, thereby capturing groups with cognitive characteristics and behavior patterns similar to the target learner. Specifically, first, through course modeling based on the LLM, the historical record sequence of the learner R = R1, R2, …, R |U|Converted into a historical feature sequence H = H1, H2, …, H |U| , the historical record sequence of the learner is a sequence of course names, indicating the courses that the learner has studied. Each learner u i is associated with a historical record R i = r i,1 , r i,2 , …, r i,n , where the record consists of an ordered sequence of previously participated courses arranged in chronological order, where n represents the number of previously studied courses, and r i,j ∈C represents the j-th course participation event of learner u i , and C represents the set containing all courses. And H i = e i,1 , e i,2 , …, e i,n , e i,j represents the enhanced embedding vector of course c i . For the target learner u i , in this embodiment, the dot product between the historical feature embedding of u i and the historical embeddings of other learners u j is calculated to determine the set U i of the top K highest-scoring learners:

[0097]

[0098] where j = 1, 2, …, |U|, and K is a hyperparameter. This indicates that these learners exhibit closely consistent preferences and behavioral patterns.

[0099] (2) Preference prediction: Integrate the historical feature sequences of the similar learners obtained above with the interaction sequence of the target learner and use it as the input of the LLM. Then, the LLM uses the historical features of the target learner and the information from the similar learners to predict the future course embedding:

[0100]

[0101] where, is the historical feature sequence of the K learners similar to the target learner u i . e i,n+1 is the embedding vector of the course predicted by the LLM for the future interaction of u i , specifically corresponding to the (n + 1)-th interaction. Combining the filtered historical records of similar learners as supplementary information to enrich the interaction data of the target learner can more comprehensively characterize the interest characteristics of the learner and ultimately help to more accurately recommend courses.

[0102] Furthermore, the method further includes:

[0103] Using the embedding of the next recommended course generated based on the historical interactions of the target learner as the training objective, obtain the predicted interest score for each course:

[0104]

[0105] where W and b are trainable parameters;

[0106] Determine the minimized predicted response according to the predicted interest score:

[0107]

[0108] Optimize the learning process of the model by the cross-entropy loss between the minimized predicted response and the true response:

[0109]

[0110] where y i is the true response.

[0111] Specifically: Training and optimization:

[0112] Considering the complexity of the LLM, a parameter-efficient fine-tuning method based on LoRA is adopted during training to reduce the overall cost. The training objective of the course recommendation model is to generate the embedding of the next recommended course based on the historical interaction embedding of the target learner. Specifically, based on the historical interaction embedding of u i , the predicted interest score of course c i is obtained by the following formula:

[0113]

[0114] where W and b are trainable parameters. is the corresponding probability of predicting course c i .

[0115] Finally, optimize the learning process by minimizing the cross-entropy loss between the predicted response and the true response y i , defined as follows:

[0116]

[0117] By iteratively updating the loss in each epoch, the optimal weighted parameters can be obtained, so as to make accurate predictions for the next course.

[0118] This embodiment also provides a course recommendation system based on decoupled modeling of large language models, including: a course modeling unit, which is used to define a format prompt template including generation tasks and required output formats, and use the LLM model to generate a course sequence from the format prompt template and course names; extract the course names and corresponding relationship descriptions in the course sequence, construct a course relationship graph, and input the course relationship graph into the LLM model to obtain knowledge-enhanced embedding vectors; a learner modeling unit, which is used to obtain the historical record sequence of the learner, convert the historical record sequence into a historical feature sequence, and based on the historical feature sequence, calculate the dot product between the historical feature embedding of the target learner and the historical embeddings of other learners to obtain a set of learners with the highest scores; a preference prediction unit, which is used to obtain the historical feature sequences of similar learners based on the set of learners with the highest scores, integrate them with the interaction sequence of the target learner, input the integration result into the LLM model, and obtain future course embedding vectors.

[0119] The above-described embodiments are only descriptions of the preferred embodiments of the present invention, and do not limit the scope of the present invention. Without departing from the design spirit of the present invention, various deformations and improvements made by those of ordinary skill in the art to the technical solutions of the present invention shall fall within the protection scope determined by the claims of the present invention.

Claims

1. A course recommendation method based on large language model decoupling modeling, characterized in that: include: Define a template including a generation task and a required output format prompt, and use the LLM model to generate a course sequence from the format prompt template and the course name; Extracting course names and corresponding relationship descriptions in the course sequence, constructing a course relationship graph, inputting the course relationship graph into the LLM model, and obtaining a knowledge-enhanced embedding vector; Obtain a historical record sequence of the learner, convert the historical record sequence into a historical feature sequence, calculate the dot product between the historical feature embedding of the target learner and the historical embedding of other learners based on the historical feature sequence, and obtain a set of learners with the highest score; Based on the set of highest-scoring learners, historical feature sequences of similar learners are obtained and integrated with the interaction sequence of the target learner, and the integration result is input into the LLM model to obtain future course embedding vectors.

2. The course recommendation method based on large language model decoupling modeling according to claim 1 is characterized in that: Generating the course sequence includes: S i =LLM(P,c i )={(c i ,s i )}; Among them, c i Indicates the name of the course, s i It represents the generated course sequence, that is, the course relationship description includes prerequisites and advancement paths, and P represents the format prompt template.

3. The course recommendation method based on large language model decoupling modeling according to claim 1 is characterized in that: Constructing the course relationship diagram includes: Each course in the course name is regarded as a central node, a directed graph is constructed around the central node, and the directed graph is used as the course relationship graph G c =(V c ,ε c );in, represents the set of courses associated with each course, It indicates that directed edges are established between courses based on the relationship description to form an edge set.

4. The course recommendation method based on large language model decoupling modeling according to claim 3 is characterized in that: In the directed graph, from node v i Points to node v j The edge is represented by: e ij =(v i ,v j )∈ε c ; Among them, node v i is node v j Prerequisite course, node v j is node v i Advanced courses.

5. The course recommendation method based on large language model decoupling modeling according to claim 1 is characterized in that: Obtaining the knowledge-enhanced embedding vector includes: Among them, M c is a specific merging strategy on the final hidden state of the course LLM to generate an enhanced representation of the course, T i is a simplified textual representation of the course relationship diagram. is the real number field, and d is the dimension of the hidden layer.

6. The course recommendation method based on large language model decoupling modeling according to claim 1 is characterized in that: Obtaining the set of highest scoring learners includes: Where j = {1, 2, ..., |U|}, K is a hyperparameter, T is the transpose, H i Represents user u i The historical characteristic sequence, H j Represents user u j sequence of historical features.

7. The course recommendation method based on large language model decoupling modeling according to claim 1 is characterized in that: Obtaining the future course embedding vector includes: in, Represents the target learner u i Similar historical feature sequences of K learners, e i,n+1 Indicates LLM is u i The embedding vector of the course for future interaction prediction, M u Indicates user LLM.

8. The course recommendation method based on large language model decoupling modeling according to claim 1 is characterized in that: The method further includes: Generate the embedding of the next recommended course based on the historical interaction embedding of the target learner as the training target, and obtain the predicted interest score of each course: Among them, W and b are trainable parameters; Based on the predicted interest scores, determine the minimized predicted response: The learning process of the model is optimized by minimizing the cross entropy loss between the predicted response and the true response as described: Among them, y i For real response.

9. The course recommendation system based on large language model decoupling modeling is characterized by: include: A course modeling unit, used for defining a template including a generated task and a required output format prompt, and generating a course sequence from the format prompt template and the course name using an LLM model; Extracting course names and corresponding relationship descriptions in the course sequence, constructing a course relationship graph, inputting the course relationship graph into the LLM model, and obtaining a knowledge-enhanced embedding vector; A learner modeling unit is used to obtain a historical record sequence of a learner, convert the historical record sequence into a historical feature sequence, and based on the historical feature sequence, calculate the dot product between the historical feature embedding of the target learner and the historical embedding of other learners to obtain a set of learners with the highest score; The preference prediction unit is used to obtain the historical feature sequence of similar learners based on the set of highest-scoring learners, integrate it with the interaction sequence of the target learner, input the integration result into the LLM model, and obtain the future course embedding vector.

Images