Personalized learning recommendation method and system based on large model

Abstract

The invention discloses a personalized learning recommendation method and system based on a large model, belongs to the technical field of deep learning, and aims to solve the technical problem of how to introduce the large model to realize efficient, intelligent and accurate personalized learning experience. Comprising a user behavior analysis module, a content recommendation engine module and a learning path planning module, the system takes a user behavior feature vector, a semantic feature vector of a learning resource and an interaction relationship between the user and the learning resource as input, and learns a low-dimensional embedded vector of the user and the learning resource through a graph neural network; based on the low-dimensional embedding vectors of the user and the learning resources, the learning resources are recommended to the user in a mixed recommendation mode combining content recommendation and collaborative filtering recommendation.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and specifically to a personalized learning recommendation method and system based on a large model. Background Art

[0002] In the wave of digital transformation of education, personalized learning recommendation systems have emerged, aiming to provide learners with tailored learning resources and paths through algorithmic models. Although early recommendation systems have met personalized needs to a certain extent, they are usually limited by static user portraits and simple collaborative filtering mechanisms, making it difficult to adapt to learners' rapidly changing learning needs and preferences. In addition, traditional systems also face challenges in processing large-scale data and real-time recommendations.

[0003] With the rapid development of artificial intelligence technology, especially the advancement of deep learning, reinforcement learning and big data analysis technology, the education industry has ushered in new development opportunities. The application potential of these technologies lies in improving the intelligence level of recommendation systems, enabling them to more accurately capture learners' subtle preferences, update recommendation strategies in real time, and extract valuable information from massive data, thereby providing each learner with a more accurate and efficient personalized learning experience.

[0004] How to introduce large models to achieve efficient, intelligent and accurate personalized learning experience is a technical problem that needs to be solved. Summary of the invention

[0005] The technical task of the present invention is to address the above shortcomings and provide a personalized learning recommendation method and system based on a large model to solve the technical problem of how to introduce a large model to achieve efficient, intelligent and accurate personalized learning experience.

[0006] In a first aspect, the present invention provides a personalized learning recommendation system based on a large model, comprising a user behavior analysis module, a content recommendation engine module, and a learning path planning module;

[0007] The user behavior analysis module is used to collect historical user data of multiple users as sample data, and to train the constructed user behavior analysis model based on the sample data, and to collect user data of target users, and to use the trained user behavior analysis model to predict and output the user behavior feature vector of the target user as a user portrait, with the user data of the target user as input. The user behavior analysis model is a large model constructed based on the Transformer model that introduces the self-attention mechanism;

[0008] The content recommendation engine module is used to perform semantic analysis on learning resources based on a pre-trained semantic analysis model, predict and output the semantic feature vector of the learning resources, construct a knowledge graph with the semantic feature vector of the learning resources as entities, the entities as nodes of the knowledge graph, and the relationships between the entities as edges, learn the low-dimensional embedding vectors of users and learning resources with the user behavior feature vector, the semantic feature vector of the learning resources, and the interaction relationship between the user and the learning resources as inputs through a graph neural network, and recommend learning resources for users through a hybrid recommendation method that combines content recommendation and collaborative filtering recommendation. Among them, the low-dimensional embedding vectors of the user and the learning resources include the user embedding vector and the learning resource embedding vector. The user embedding vector is the representation of the user in the low-dimensional space, capturing the user's behavior patterns, preferences, and relationships with other users and learning resources. The learning resource embedding vector is the representation of the learning resource in the low-dimensional space, capturing the semantic features of the learning resource and its relationships with other resources and users;

[0009] The learning path planning module is used to design a policy network based on a deep reinforcement learning algorithm and dynamically adjust the parameters of the graph neural network according to the user's real-time feedback.

[0010] Preferably, the user data includes user behavior data and context information. The user behavior data includes user login, course selection, learning progress, interaction behavior records, and test score-related data. The context information includes time, device type, location, and learning environment;

[0011] The user behavior analysis model includes an input embedding layer, a position encoding layer, a self-attention layer, a feed-forward network, and a residual connection and normalization layer. The input embedding layer is used to encode the input user data into a vector. The position encoding layer is used to add temporal information to the vector output by the input embedding layer to preserve the order of the behavior sequence. The self-attention layer is used to calculate the correlation weights between each behavior in the sequence and other behaviors. The feed-forward network is used to perform a non-linear transformation on the output of the self-attention layer. The residual connection and layer normalization are used to improve the training stability and model depth of the user behavior analysis model and output the user behavior feature vector.

[0012] Preferably, the semantic analysis model is a Bert model. The Bert model takes the text data of the learning resources as input, predicts and outputs a sentence vector and a token-level vector. Among them, the text content of the learning resources includes course descriptions, textbook contents, and articles. The sentence vector is a semantic representation vector obtained through the CLS label and is used for the overall classification of the learning resources. The token-level vector includes the extracted keywords and topics;

[0013] Correspondingly, the content recommendation engine module is used to take the extracted token-level vector as the semantic feature vector of the learning resource.

[0014] Preferably, based on the low-dimensional embedding vectors of the user and the learning resource, a hybrid recommendation method combining content recommendation and collaborative filtering recommendation is used to recommend learning resources to the user, including the following steps:

[0015] Content recommendation: Calculate the semantic similarity between learning resources through the learning resource embedding vector, recommend semantically similar learning resources to the user, and obtain a recommendation list;

[0016] Collaborative filtering: Calculate the similarity between users through the user embedding vector, recommend learning resources liked by other users to the user, and obtain a recommendation list;

[0017] Hybrid recommendation: Perform weighted summation on the recommendation list obtained through content recommendation and the recommendation list obtained through collaborative filtering to obtain the final recommendation list.

[0018] In a second aspect, a personalized learning recommendation method based on a large model is used to recommend learning resources to a user through a personalized learning recommendation system based on a large model as described in any item of the first aspect. The method includes the following steps:

[0019] User behavior analysis: Collect the historical user data of multiple users as sample data, train the constructed user behavior analysis model based on the sample data, collect the user data of the target user, and use the user data of the target user as input to predict and output the user behavior feature vector of the target user as the user portrait through the trained user behavior analysis model. Among them, the user behavior analysis model is a large model constructed based on the Transformer model introducing the self-attention mechanism;

[0020] Content recommendation: Based on a pre-trained semantic analysis model, perform semantic analysis on learning resources, predict and output the semantic feature vectors of learning resources. Using the semantic feature vectors of learning resources as entities, construct a knowledge graph with entities as nodes of the knowledge graph and the relationships between entities as edges. Taking the user behavior feature vectors, the semantic feature vectors of learning resources, and the interaction relationships between users and learning resources as inputs, learn the low-dimensional embedding vectors of users and learning resources through a graph neural network. Based on the low-dimensional embedding vectors of users and learning resources, recommend learning resources for users through a hybrid recommendation method that combines content recommendation and collaborative filtering recommendation. Among them, the low-dimensional embedding vectors of users and learning resources include user embedding vectors and learning resource embedding vectors. The user embedding vector is the representation of the user in the low-dimensional space, capturing the user's behavior patterns, preferences, and relationships with other users and learning resources. The learning resource embedding vector is the representation of the learning resource in the low-dimensional space, capturing the semantic features of the learning resource and its relationships with other resources and users;

[0021] Learning path planning: Design a policy network based on the deep reinforcement learning algorithm and dynamically adjust the parameters of the graph neural network according to the user's real-time feedback.

[0022] Preferably, the user data includes user behavior data and context information. The user behavior data includes user login, course selection, learning progress, interaction behavior records, and test score-related data. The context information includes time, device type, location, and learning environment;

[0023] The user behavior analysis model includes an input embedding layer, a positional encoding layer, a self-attention layer, a feed-forward network, and a residual connection and normalization layer. The input embedding layer is used to encode the input user data into vectors. The positional encoding layer is used to add temporal information to the vectors output by the input embedding layer to preserve the order of the behavior sequence. The self-attention layer is used to calculate the association weights between each behavior in the sequence and other behaviors. The feed-forward network is used to perform a non-linear transformation on the output of the self-attention layer. The residual connection and layer normalization are used to improve the training stability and model depth of the user behavior analysis model and output the user behavior feature vectors.

[0024] Preferably, the semantic analysis model is a Bert model. The Bert model takes the text data of learning resources as input and predicts and outputs sentence vectors and token-level vectors. Among them, the text content of learning resources includes course descriptions, textbook contents, and articles. The sentence vector is a semantic representation vector obtained through the CLS label and is used for the overall classification of learning resources. The token-level vectors include extracted keywords and topics;

[0025] Correspondingly, when making content recommendations, use the extracted token-level vectors as the semantic feature vectors of learning resources.

[0026] Preferably, a hybrid recommendation method that combines content recommendation and collaborative filtering recommendation based on the low-dimensional embedding vectors of users and learning resources is used to recommend learning resources to users, including the following steps:

[0027] Content recommendation: Calculate the semantic similarity between learning resources through the learning resource embedding vectors, recommend semantically similar learning resources to users, and obtain a recommendation list;

[0028] Collaborative filtering: Calculate the similarity between users through the user embedding vectors, recommend learning resources liked by other users to users, and obtain a recommendation list;

[0029] Hybrid recommendation: Perform weighted summation on the recommendation list obtained through content recommendation and the recommendation list obtained through collaborative filtering to obtain the final recommendation list.

[0030] The personalized learning recommendation system and method based on the large model of the present invention have the following advantages:

[0031] 1. Improve personalized matching degree: By deeply analyzing user behavior, the system realizes the accurate identification of the personalized needs of learners, thereby providing resource recommendations that are more in line with their learning styles and ability levels;

[0032] 2. Enhance the learning experience: Through dynamic learning path planning, the system can design more attractive and challenging learning plans for learners, effectively improving the participation and learning motivation of learners;

[0033] 3. Optimize learning efficiency: The real-time recommendation and adaptive adjustment mechanism ensure that learners can obtain appropriate learning materials at the right time, reduce the time consumption of searching for resources, and improve learning efficiency;

[0034] 4. Promote educational equity: The system provides customized learning recommendations for all learners, regardless of geographical location, economic or social background, helping to narrow the educational gap between different learners;

[0035] 5. Improve resource utilization: Through intelligent recommendation, the system can allocate and utilize educational resources more reasonably, avoid resource waste, and ensure that each learning content can reach the learners who need it;

[0036] 6. Support educational research and development: A large amount of user behavior data and feedback information collected by the system provide a rich empirical basis for educational research, helping educators and researchers to deeply understand the learning process and effects;

[0037] 7. Flexibly respond to diverse learning scenarios: The design of the system takes into account the needs of different learning scenarios, and can provide effective personalized recommendations in both formal educational environments and self-study environments;

[0038] 8. Promote learners' self - growth: By tracking learners' learning progress and achievements, the system can encourage learners to conduct self - reflection and adjustment, thereby promoting the development of their autonomous learning ability and lifelong learning ability;

[0039] 9. Reduce teachers' burden: The system's automated recommendation and planning functions reduce the workload of teachers in resource allocation and learning guidance, enabling teachers to focus more on teaching innovation and individual student tutoring;

[0040] 10. Promote the system's own iteration and improvement: The system is designed with a self - learning and optimization mechanism that can continuously adjust the recommendation algorithm according to user feedback and behavior patterns, achieving continuous performance improvement and function perfection. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for use in the embodiments or the description of the prior art. Obviously, the following - described drawings 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.

[0042] The present invention will be further described below in conjunction with the drawings.

[0043] Figure 1 It is a flowchart of a personalized learning recommendation system based on a large - model for Embodiment 1. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0044] The present invention will be further described below in conjunction with the drawings and specific embodiments, so that those skilled in the art can better understand the present invention and be able to implement it. However, the illustrated embodiments are not intended to limit the present invention. Without conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0045] The embodiments of the present invention provide a personalized learning recommendation system and method based on a large - model, which are used to solve the technical problem of how to introduce a large - model to achieve an efficient, intelligent, and accurate personalized learning experience.

[0046] Embodiment 1:

[0047] A personalized learning recommendation system based on a large - model of the present invention includes a user behavior analysis module, a content recommendation engine module, and a learning path planning module.

[0048] The user behavior analysis module is used to collect the historical user data of multiple users as sample data, and based on the sample data, train the constructed user behavior analysis model. It is also used to collect the user data of the target user, and use the user data of the target user as input to predict and output the user behavior feature vector of the target user as the user portrait through the trained user behavior analysis model. Among them, the user behavior analysis model is a large model constructed based on the Transformer model introducing the self-attention mechanism.

[0049] Among them, the user data includes user behavior data and context information. The user behavior data includes user login, course selection, learning progress, interaction behavior records, and test score-related data. The context information includes time, device type, location, and learning environment.

[0050] As a specific implementation of the user behavior analysis model, the model includes an input embedding layer, a position encoding layer, a self-attention layer, a feed-forward network, and a residual connection and normalization layer. The input embedding layer is used to encode the input user data into a vector. The position encoding layer is used to add temporal information to the vector output by the input embedding layer to retain the order of the behavior sequence. The self-attention layer is used to calculate the association weights between each behavior in the sequence and other behaviors. The feed-forward network is used to perform a non-linear transformation on the output of the self-attention layer. The residual connection and layer normalization are used to improve the training stability and model depth of the user behavior analysis model and output the user behavior feature vector.

[0051] The user behavior analysis model of this embodiment integrates a deep learning model based on Transformer, and uses its self-attention mechanism to analyze the user behavior sequence, capturing the behavior characteristics of the user at different time points and their mutual relationships. When collecting user data, user behavior data is collected and context information such as time, location, and learning environment is introduced to more comprehensively understand user behavior and enhance the accuracy and practicality of the user portrait.

[0052] The content recommendation engine module is used to perform semantic analysis on learning resources based on a pre-trained semantic analysis model, predict and output the semantic feature vectors of learning resources, use the semantic feature vectors of learning resources as entities, construct a knowledge graph with entities as nodes of the knowledge graph and the relationships between entities as edges, use the user behavior feature vectors, the semantic feature vectors of learning resources, and the interaction relationships between users and learning resources as inputs, learn the low-dimensional embedding vectors of users and learning resources through a graph neural network, and recommend learning resources for users based on the low-dimensional embedding vectors of users and learning resources through a hybrid recommendation method that combines content recommendation and collaborative filtering recommendation. Among them, the low-dimensional embedding vectors of users and learning resources include user embedding vectors and learning resource embedding vectors. The user embedding vector is the representation of the user in the low-dimensional space, capturing the user's behavior patterns, preferences, and relationships with other users and learning resources. The learning resource embedding vector is the representation of the learning resource in the low-dimensional space, capturing the semantic features of the learning resource and its relationships with other resources and users.

[0053] As a specific implementation of the semantic analysis model, the Bert model is selected. The Bert model takes the text data of learning resources as input, predicts and outputs sentence vectors and token-level vectors. Among them, the text content of learning resources includes course descriptions, textbook contents, and articles. The sentence vector is the semantic representation vector obtained through the CLS label, used for the overall classification of learning resources. The token-level vectors include the extracted keywords and topics. Correspondingly, the content recommendation engine module is used to use the extracted token-level vectors as the semantic feature vectors of learning resources.

[0054] In this embodiment, pre-trained language models such as BERT are used to perform in-depth semantic analysis on learning materials, extract key concepts and topics, and achieve accurate understanding and representation of learning resources. Construct an educational domain knowledge graph, connect learning resources with entities and relationships in the knowledge graph, and use graph neural network technologies such as graph convolutional networks (GCN) to recommend resources by leveraging the rich relationships between entities.

[0055] Among them, the entities are concepts in the educational domain (such as "mathematics", "Newton's laws", "trigonometric functions"), learning resources (such as "Course A", "Exercise Set B"), knowledge points (such as "derivative", "integral"), and the relationships are semantic associations between entities (such as "belong to", "prerequisite knowledge", "related resources"). The concepts extracted from learning materials (such as "derivative") are linked to the knowledge graph as entities, and resources are connected to entities through "contains" or "explains" relationships.

[0056] In this embodiment, based on the low-dimensional embedding vectors of users and learning resources, learning resources are recommended for users through a hybrid recommendation method that combines content recommendation and collaborative filtering recommendation, including the following steps:

[0057] (1) Content recommendation: Calculate the semantic similarity between learning resources through the embedded vectors of learning resources, recommend semantically similar learning resources to users, and obtain a recommendation list.

[0058] (2) Collaborative filtering: Calculate the similarity between users through the embedded vectors of users, recommend learning resources liked by other users to users, and obtain a recommendation list.

[0059] (3) Hybrid recommendation: Perform weighted summation on the recommendation list obtained through content recommendation and the recommendation list obtained through collaborative filtering to obtain the final recommendation list.

[0060] In this embodiment, this step combines content-based recommendation and collaborative filtering recommendation, utilizes the features extracted by deep learning and user behavior data, and provides more accurate personalized resource recommendations through a hybrid recommendation framework.

[0061] The learning path planning module is used to design a policy network based on the deep reinforcement learning algorithm and dynamically adjust the parameters of the graph neural network according to the real-time feedback of users.

[0062] In the implementation process, during deep reinforcement learning, the Actor-Critic network is applied for policy learning. Among them, the Actor network is responsible for generating learning path recommendations, and the Critic network evaluates the expected effect of the recommended path to achieve dynamic optimization. When adjusting the parameters of the graph neural network, combined with the real-time feedback and learning effectiveness of the learner, techniques such as the policy gradient method or Q-learning in reinforcement learning are used to continuously adjust the learning path to optimize the long-term learning goals of the learner. Here, the adjustment of the parameters of the graph neural network based on deep reinforcement learning is implemented based on the steps of adjusting the parameters of the neural network by deep reinforcement learning in the prior art and will not be specifically described.

[0063] Embodiment 2:

[0064] A personalized learning recommendation method based on a large model of the present invention recommends learning resources to users through the system disclosed in Embodiment 1. This method includes three steps: user behavior analysis, content recommendation, and learning path planning.

[0065] User behavior analysis: Collect the historical user data of multiple users as sample data, train the constructed user behavior analysis model based on the sample data, collect the user data of the target user, and use the user data of the target user as input. Through the trained user behavior analysis model, predict and output the user behavior feature vector of the target user as the user portrait. Among them, the user behavior analysis model is a large model constructed based on the Transformer model introducing the self-attention mechanism.

[0066] Among them, user data includes user behavior data and context information. User behavior data includes data related to user login, course selection, learning progress, interaction behavior records, and test scores. Context information includes time, device type, location, and learning environment.

[0067] As a specific implementation of the user behavior analysis model, the model includes an input embedding layer, a positional encoding layer, a self-attention layer, a feed-forward network, and a residual connection and normalization layer. The input embedding layer is used to encode the input user data into vectors. The positional encoding layer is used to add temporal information to the vectors output by the input embedding layer to preserve the order of the behavior sequence. The self-attention layer is used to calculate the correlation weights between each behavior in the sequence and other behaviors. The feed-forward network is used to perform a non-linear transformation on the output of the self-attention layer. The residual connection and layer normalization are used to improve the training stability and model depth of the user behavior analysis model and output the user behavior feature vector.

[0068] The user behavior analysis model in this embodiment integrates a deep learning model based on Transformer, and uses its self-attention mechanism to analyze the user behavior sequence, capturing the behavior characteristics of the user at different time points and their mutual relationships. When collecting user data, user behavior data is collected and context information such as time, location, and learning environment is introduced to more comprehensively understand user behavior and enhance the accuracy and usability of the user profile.

[0069] Content recommendation: Perform semantic analysis on learning resources based on a pre-trained semantic analysis model, predict and output the semantic feature vectors of learning resources. Using the semantic feature vectors of learning resources as entities, construct a knowledge graph with entities as nodes of the knowledge graph and the relationships between entities as edges. Using the user behavior feature vector, the semantic feature vector of the learning resource, and the interaction relationship between the user and the learning resource as inputs, learn the low-dimensional embedding vectors of the user and the learning resource through a graph neural network. Based on the low-dimensional embedding vectors of the user and the learning resource, recommend learning resources for the user through a hybrid recommendation method that combines content recommendation and collaborative filtering recommendation. Among them, the low-dimensional embedding vectors of the user and the learning resource include the user embedding vector and the learning resource embedding vector. The user embedding vector is the representation of the user in the low-dimensional space, capturing the user's behavior patterns, preferences, and relationships with other users and learning resources. The learning resource embedding vector is the representation of the learning resource in the low-dimensional space, capturing the semantic characteristics of the learning resource and its relationships with other resources and users.

[0070] As a specific implementation of the semantic analysis model, the Bert model is selected. The Bert model takes the text data of learning resources as input and predicts output sentence vectors and token-level vectors. Among them, the text content of learning resources includes course descriptions, teaching materials, and articles. The sentence vector is a semantic representation vector obtained through the CLS label and is used for the overall classification of learning resources. The token-level vector includes the extracted keywords and themes. Correspondingly, the content recommendation engine module uses the extracted token-level vector as the semantic feature vector of the learning resources.

[0071] In this embodiment, pre-trained language models such as BERT are used to perform in-depth semantic analysis on learning materials, extract key concepts and themes, and achieve accurate understanding and representation of learning resources. A knowledge graph in the education field is constructed, connecting learning resources with entities and relationships in the knowledge graph. Through graph neural network technologies such as graph convolutional networks (GCNs), rich relationships between entities are used for resource recommendation.

[0072] Among them, the entities are concepts in the education field (such as "mathematics", "Newton's laws", "trigonometric functions"), learning resources (such as "Course A", "Exercise Set B"), knowledge points (such as "derivatives", "integrals"), and the relationships are semantic associations between entities (such as "belongs to", "prerequisite knowledge", "related resources"). The concepts extracted from learning materials (such as "derivatives") are linked to the knowledge graph as entities, and resources are connected to entities through "contains" or "explains" relationships.

[0073] In this embodiment, based on the low-dimensional embedding vectors of users and learning resources, a hybrid recommendation method that combines content recommendation and collaborative filtering recommendation is used to recommend learning resources to users, including the following steps:

[0074] (1) Content recommendation: Calculate the semantic similarity between learning resources through learning resource embedding vectors, recommend semantically similar learning resources to users, and obtain a recommendation list.

[0075] (2) Collaborative filtering: Calculate the similarity between users through user embedding vectors, recommend learning resources liked by other users to users, and obtain a recommendation list.

[0076] (3) Hybrid recommendation: Perform weighted summation on the recommendation list obtained through content recommendation and the recommendation list obtained through collaborative filtering to obtain the final recommendation list.

[0077] This step in this embodiment combines content-based recommendation and collaborative filtering recommendation, uses features extracted by deep learning and user behavior data, and provides more accurate personalized resource recommendations through a hybrid recommendation framework.

[0078] Learning path planning: Design a policy network based on the deep reinforcement learning algorithm, and dynamically adjust the parameters of the graph neural network according to the real-time feedback of users.

[0079] In the implementation process, during deep reinforcement learning, the Actor-Critic network is applied for policy learning. Among them, the Actor network is responsible for generating learning path recommendations, and the Critic network evaluates the expected effects of the recommended paths to achieve dynamic optimization. When adjusting the parameters of the graph neural network, combined with the real-time feedback and learning effectiveness of the learner, techniques such as the policy gradient method or Q-learning in reinforcement learning are used to continuously adjust the learning path to optimize the long-term learning goals of the learner. Here, the adjustment of the parameters of the graph neural network based on deep reinforcement learning is implemented based on the steps of adjusting the parameters of the neural network by deep reinforcement learning in the prior art, without special explanation.

[0080] The above has introduced in detail the method of the personalized learning recommendation system based on the large model. Specific examples are used in this article to elaborate on the principle and implementation manner of the present invention. The description of the above embodiments is only used to help understand the method and its core idea of the present invention; at the same time, for those of ordinary skill in the art, according to the idea of the present invention, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation to the present invention.

Claims

1. A personalized learning recommendation system based on a large model, characterized in that: Includes user behavior analysis module, content recommendation engine module and learning path planning module; The user behavior analysis module is used to collect historical user data of multiple users as sample data, and to train the constructed user behavior analysis model based on the sample data, and to collect user data of target users, and to use the trained user behavior analysis model to predict and output the user behavior feature vector of the target user as a user portrait, with the user data of the target user as input. The user behavior analysis model is a large model constructed based on the Transformer model that introduces the self-attention mechanism; The content recommendation engine module is used to perform semantic analysis on learning resources based on a pre-trained semantic analysis model, predict and output semantic feature vectors of learning resources, and construct a knowledge graph with semantic feature vectors of learning resources as entities, entities as nodes of knowledge graphs, and relationships between entities as edges; it is used to take user behavior feature vectors, semantic feature vectors of learning resources, and interactive relationships between users and learning resources as inputs, and learn low-dimensional embedding vectors of users and learning resources through graph neural networks; it is used to recommend learning resources to users based on the low-dimensional embedding vectors of users and learning resources, through a hybrid recommendation method combining content recommendation and collaborative filtering recommendation, wherein the low-dimensional embedding vectors of users and learning resources include user embedding vectors and learning resource embedding vectors, the user embedding vector is a representation of the user in a low-dimensional space, capturing the user's behavior patterns, preferences, and relationships with other users and learning resources, and the learning resource embedding vector is a representation of the learning resource in a low-dimensional space, capturing the semantic features of the learning resource and its relationships with other resources and users; The learning path planning module is used to design a strategy network based on a deep reinforcement learning algorithm and dynamically adjust the parameters of the graph neural network according to real-time feedback from users.

2. The large model-based personalized learning recommendation system according to claim 1, characterized in that: The user data includes user behavior data and context information. The user behavior data includes user login, course selection, learning progress, interactive behavior records and test score related data. The context information includes time, device type, location and learning environment. The user behavior analysis model includes an input embedding layer, a position encoding layer, a self-attention layer, a feedforward network, and a residual connection and normalization layer. The input embedding layer is used to encode the input user data into a vector, the position encoding layer is used to add timing information to the vector output by the input embedding layer to retain the order of the behavior sequence, the self-attention layer is used to calculate the association weights of each behavior with other behaviors in the sequence, the feedforward network is used to perform nonlinear transformation on the output of the self-attention layer, and the residual connection and layer normalization are used to improve the training stability and model depth of the user behavior analysis model and output the user behavior feature vector.

3. The large model-based personalized learning recommendation system according to claim 1, characterized in that: The semantic analysis model is a Bert model, which takes the text data of the learning resources as input, predicts and outputs sentence vectors and token-level vectors, wherein the text content of the learning resources includes course descriptions, teaching materials, and articles, the sentence vectors are semantic representation vectors obtained through CLS tags, and are used for the overall classification of the learning resources, and the token-level vectors include extracted keywords and topics; Correspondingly, the content recommendation engine module is used to use the extracted token-level vector as the semantic feature vector of the learning resource.

4. The large model-based personalized learning recommendation system according to claim 1, characterized in that: Based on the low-dimensional embedding vectors of users and learning resources, a hybrid recommendation method combining content recommendation and collaborative filtering recommendation is used to recommend learning resources to users, including the following steps: Content recommendation: Calculate the semantic similarity between learning resources through the learning resource embedding vector, recommend semantically similar learning resources to users, and obtain a recommendation list; Collaborative filtering: Calculate the similarity between users through user embedding vectors, recommend learning resources that other users like, and obtain a recommendation list; Hybrid recommendation: The recommendation list obtained through content recommendation and the recommendation list obtained through collaborative filtering are weighted summed to obtain the final recommendation list.

5. A personalized learning recommendation method based on a large model, characterized in that: The method is used to recommend learning resources to a user through a personalized learning recommendation system based on a large model as described in any one of claims 1 to 4, the method comprising the following steps: User behavior analysis: Collect historical user data of multiple users as samples, train the constructed user behavior analysis model based on the sample data, collect user data of target users, use the user data of target users as input, and use the trained user behavior analysis model to predict and output the user behavior feature vector of the target user as the user portrait. The user behavior analysis model is a large model built based on the Transformer model that introduces the self-attention mechanism; Content recommendation: Based on the pre-trained semantic analysis model, semantic analysis is performed on learning resources, and the semantic feature vectors of learning resources are predicted and output. The semantic feature vectors of learning resources are used as entities, entities are used as nodes of the knowledge graph, and the relationship between entities is used as edges to construct a knowledge graph. The user behavior feature vector, the semantic feature vector of learning resources, and the interactive relationship between users and learning resources are used as inputs. The low-dimensional embedding vectors of users and learning resources are learned through a graph neural network. Based on the low-dimensional embedding vectors of users and learning resources, learning resources are recommended to users through a hybrid recommendation method that combines content recommendation and collaborative filtering recommendation. The low-dimensional embedding vectors of users and learning resources include user embedding vectors and learning resource embedding vectors. The user embedding vector is the representation of the user in a low-dimensional space, which captures the user's behavior patterns, preferences, and relationships with other users and learning resources. The learning resource embedding vector is the representation of the learning resource in a low-dimensional space, which captures the semantic features of the learning resource and its relationship with other resources and users. Learning path planning: Design a strategy network based on a deep reinforcement learning algorithm and dynamically adjust the parameters of the graph neural network based on real-time feedback from users.

6. The large model-based personalized learning recommendation method according to claim 5, characterized in that: The user data includes user behavior data and context information. The user behavior data includes user login, course selection, learning progress, interactive behavior records and test score related data. The context information includes time, device type, location and learning environment. The user behavior analysis model includes an input embedding layer, a position encoding layer, a self-attention layer, a feedforward network, and a residual connection and normalization layer. The input embedding layer is used to encode the input user data into a vector, the position encoding layer is used to add timing information to the vector output by the input embedding layer to retain the order of the behavior sequence, the self-attention layer is used to calculate the association weights of each behavior with other behaviors in the sequence, the feedforward network is used to perform nonlinear transformation on the output of the self-attention layer, and the residual connection and layer normalization are used to improve the training stability and model depth of the user behavior analysis model and output the user behavior feature vector.

7. The large model-based personalized learning recommendation method according to claim 5, characterized in that: The semantic analysis model is a Bert model, which takes the text data of the learning resources as input, predicts and outputs sentence vectors and token-level vectors, wherein the text content of the learning resources includes course descriptions, teaching materials, and articles, the sentence vectors are semantic representation vectors obtained through CLS tags, and are used for the overall classification of the learning resources, and the token-level vectors include extracted keywords and topics; Correspondingly, when recommending content, the extracted token-level vector is used as the semantic feature vector of the learning resource.

8. The large model-based personalized learning recommendation method according to claim 5, characterized in that: Based on the low-dimensional embedding vectors of users and learning resources, a hybrid recommendation method combining content recommendation and collaborative filtering recommendation is used to recommend learning resources to users, including the following steps: Content recommendation: Calculate the semantic similarity between learning resources through the learning resource embedding vector, recommend semantically similar learning resources to users, and obtain a recommendation list; Collaborative filtering: Calculate the similarity between users through user embedding vectors, recommend learning resources that other users like, and obtain a recommendation list; Hybrid recommendation: The recommendation list obtained through content recommendation and the recommendation list obtained through collaborative filtering are weighted summed to obtain the final recommendation list.

Images