Intelligent education method and system based on student behavior sequence recommendation
By adopting recommendation algorithms based on student behavior sequences and knowledge graph reasoning on the AI-assisted education platform, combined with the improved self-attention mechanism of the sparse attention module, the shortcomings of the existing platform in personalized learning recommendations are solved, and efficient and accurate personalized learning path recommendations are achieved.
Patent Information
- Application Number
- CN202510194536.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-21
- Publication Date
- 2025-06-10
AI Technical Summary
The existing AI-assisted education platform has shortcomings in combining student behavior sequence analysis and knowledge graph for personalized learning recommendations. It lacks the ability to build knowledge graphs and real-time supervision of students' learning status, resulting in limited accuracy and real-time nature of personalized learning recommendations.
The recommendation algorithm based on student behavior sequence is adopted, combined with SASRec and the improved self-attention mechanism introduced by the sparse attention module, model students' behavior sequences, and capture the correlation between knowledge points through knowledge graph reasoning to generate personalized learning paths for students.
The accuracy and real-time nature of personalized learning recommendations are realized, learning efficiency and understanding ability are improved, and the dynamic update mechanism ensures that the map reflects the latest knowledge progress and enhances the pertinence and effectiveness of learning.
Smart Images

Figure CN120125392A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of intelligent education, and particularly relates to an intelligent education method and system based on student behavior sequence recommendation. Background Art
[0002] With the rapid development of artificial intelligence technology, AI + education has gradually become a research hotspot in the global education field. Some leading international education projects have begun to use AI technology, especially the application of large models, to customize students' learning paths, automatically correct homework, and provide real-time learning feedback. In recent years, student behavior sequence recommendation algorithms have gradually attracted the attention of researchers. By analyzing students' behavior data during the learning process, such as the frequency of accessing course resources, the time to complete questions, and the degree of knowledge mastery, etc., it provides new possibilities for personalized teaching. Although existing platforms have unique advantages in some aspects, there are still deficiencies in personalized learning recommendation by combining student behavior sequence analysis and knowledge graphs. In response to the pain point that students urgently need one-on-one tutoring during the programming teaching process, a certain existing platform provides real-time, personalized, and heuristic programming tutoring services for students based on a large model. It can solve most of the students' problems. For a few difficult problems that cannot be solved, the "Ask Teacher for Help" button can be clicked, and the teacher or teaching assistant will receive an in-site notice and an offline email notice, and then will come in time to provide targeted guidance to the students, thus realizing the on-demand service of the division of labor and cooperation between AI and teachers. Teachers can use functions such as class management, course management, and student management, can customize the tutoring functions and modes of their own classes (such as enabling or disabling the code generation ability), and can see the whole process, multi-dimensional, and fine-grained learning behavior data of students using the code, so that they can freely carry out teaching experiments and provide more targeted high-quality education services for students.
[0003] Under the guidance of prompt engineering and agent frameworks, the existing platform combines retrieval-augmented generation (RAG) technology to achieve targeted AI empowerment of teaching materials. The core functions include an AI knowledge base, a learning assistant, an AI teaching toolbox, a Math Copilot mathematics tool, etc. In practical applications, the platform can not only assist students in reading literature, provide real-time question-answering support, but also help teachers generate customized exercise questions and outlines, effectively improving the teaching efficiency of teachers and students.
[0004] Existing AI-assisted education platforms all have certain defects, mainly reflected in the imperfect functions and low comprehensiveness:
[0005] For example, it mainly focuses on tutoring in programming education, which to a certain extent limits the breadth of its educational field. Although it provides real-time, personalized, and heuristic tutoring services in programming teaching, compared with other comprehensive learning platforms, the limitations of its educational content are relatively obvious. More critically, it lacks the function of constructing a knowledge graph. As a powerful knowledge representation tool, the knowledge graph can clearly show the associations and hierarchies between knowledge points, helping students build a complete knowledge system. However, due to the failure to introduce this function, students may have difficulty comprehensively grasping the internal connections between knowledge points during the programming learning process, thus affecting their learning effect and in-depth understanding. In addition, it fails to make full use of students' behavioral data (such as learning paths, question completion times, and error distributions) for personalized learning recommendations. Such behavioral sequence data is of great significance for accurate recommendation algorithms and dynamic teaching path optimization.
[0006] Although some platforms have achieved remarkable results in the AI empowerment of teaching materials, there are still deficiencies in the construction of knowledge graphs. They fail to make full use of the advantages of modern information technology to systematically and structurally organize a vast amount of educational resources. This makes it difficult for students to find a learning path suitable for themselves among numerous learning materials and to deeply understand the associations between knowledge points. In addition, there are also certain shortcomings in the supervision of students' learning status. The platform lacks a mechanism for real-time tracking of students' learning progress, timely feedback on learning effects, and timely discovery and solution of learning problems. At the same time, it does not combine students' behavioral data with knowledge graph reasoning, restricting the accuracy and real-time nature of personalized learning recommendations. This deficiency affects the application potential of the platform in dynamic teaching path optimization.
[0007] Currently, there is still a lack of a comprehensive intelligent education construction and optimization method that can cover from knowledge graph construction to homework correction, real-time Q&A, and fully integrate students' behavioral data with knowledge graph reasoning to provide systematic support for personalized learning and dynamic teaching path optimization. Summary of the Invention
[0008] The technical problem to be solved by the present invention is to provide a smart education method and system based on student behavior sequence recommendation to achieve educational intelligence in view of the above-mentioned deficiencies of the prior art.
[0009] To solve the above technical problem, the technical solution adopted by the present invention is: On the one hand, the present invention provides a smart education method based on student behavior sequence recommendation, including:
[0010] Obtain student behavior sequence data;
[0011] Model the student's behavior sequence by using an improved self-attention mechanism introducing a sparse attention module;
[0012] Obtain learning resource data and construct a knowledge graph;
[0013] Perform knowledge graph reasoning to capture the association relationships between knowledge points;
[0014] Generate a personalized learning path for students based on the modeling results of the student behavior sequence and the results of knowledge graph reasoning.
[0015] Preferably, the student behavior sequence data is the learning behavior of students on the platform; the learning behavior sequence of each student can be expressed as:
[0016]
[0017] where, represents the i-th learning behavior of student u;
[0018] The learning resource data includes various learning resources on the platform, and each resource is associated with a specific knowledge point;
[0019] The knowledge graph is a structured knowledge representation form that includes the association relationships between knowledge points; each node in the knowledge graph represents a knowledge point, and the edge represents the association between knowledge points; the knowledge graph is represented as:
[0020] G=(V,E)
[0021] where, V is the set of knowledge points, and E is the association relationship between knowledge points.
[0022] Preferably, the improved self-attention mechanism with a sparse attention module is used to model the student behavior sequence, including:
[0023] Given a student behavior sequence First, map each behavior to an embedding vector e i ∈R d , where d is the embedding dimension; then, use the improved self-attention mechanism with the sparse attention module introduced in Diffuser to calculate the importance weights of each behavior;
[0024]
[0025] where, Q, K, and V are the query matrix, key matrix, and value matrix obtained through linear transformation respectively, and d k is the dimension of the key matrix K; SparseSparse represents the sparsification operation.
[0026] Preferably, the knowledge graph reasoning is used to capture the association relationships between knowledge points and recommend a learning path that matches the student's current knowledge mastery; the graph neural network GNN is used to reason about the knowledge graph; for a given knowledge graph, first initialize an embedding vector $\mathbb{R}$ for each knowledge point $v \in V$. v $\in \mathbb{R}$ d ; then, update the embedding representation of the knowledge points through a multi-layer graph convolutional network GCN:
[0027]
[0028] where $N(v)$ represents the neighbor nodes of the knowledge point $v$, $W^{(l)}$ is the weight matrix of the $l$-th layer, and $\sigma$ is the activation function.
[0029] Preferably, the embedding vector of each knowledge point is generated in the following ways:
[0030] Content-based embedding: Use the name and description information of the knowledge point to generate an embedding vector through an NLP model;
[0031] Graph-structure-based embedding: Use a graph embedding algorithm to generate an embedding vector to capture the structural relationships between knowledge points;
[0032] Preferably, combining the results of student behavior sequence analysis and knowledge graph reasoning to generate a personalized learning path for students, including:
[0033] Student knowledge mastery assessment: According to the student's behavior sequence and the knowledge graph reasoning results, evaluate the student's mastery of each knowledge point; use a linear model to calculate the mastery score $s$ of the student for the knowledge point $v$ v :
[0034] $s$ v $ = w$ T $h$ v $ + b$
[0035] where $w$ and $b$ are learnable parameters;
[0036] Learning path generation: According to the student's mastery score of the knowledge points, generate an optimal learning path from the knowledge graph; use the shortest path algorithm to find the shortest path from the knowledge points currently mastered by the student to the target knowledge point; each node in the path represents a knowledge point, and the edge represents the prerequisite relationship between knowledge points;
[0037] Resource recommendation: According to the generated learning path, recommend learning resources related to each knowledge point; select the most suitable resources according to the difficulty, type of the resources and the student's preferences.
[0038] On the other hand, the present invention also provides an intelligent education system based on a student behavior sequence recommendation algorithm, including a front-end display layer, a gateway cluster layer, a Springboot layer, a middle platform service layer, a basic service layer, a storage technology layer, and a data integration layer;
[0039] The front-end display layer supports multi-terminal interaction and data visualization, providing users with a convenient learning experience;
[0040] The gateway cluster layer realizes traffic management and security control through Gateway, and cooperates with the authentication and permission management of the underlying Springboot layer to ensure the stability and availability of the system;
[0041] The middle platform service layer is the core of the system, including a user management module, a course management module, a knowledge graph construction module, and an intelligent recommendation module;
[0042] The user management module is responsible for the creation, login, permission allocation, information editing, status management, and data analysis of user accounts, ensuring the effective management and security of platform users;
[0043] The course management module is responsible for the creation, editing, release, enrollment, learning progress tracking, grade management, and evaluation feedback of courses, ensuring the efficient management of course content and teaching quality;
[0044] The knowledge graph construction module constructs a knowledge graph based on learning resource data, dynamically associating students, courses, and knowledge points;
[0045] The intelligent recommendation module, based on the knowledge graph constructed by the knowledge graph construction module, customizes personalized learning paths for students through the SASRec behavior sequence recommendation algorithm;
[0046] The basic service layer uses a microservice framework for service discovery, and tools for configuration service and traffic control provide efficient communication and traffic control for microservices;
[0047] The storage technology layer combines multiple databases to achieve efficient multi-type data storage;
[0048] The data integration layer realizes efficient data transmission and message queue management, ensuring smooth data integration and asynchronous processing between different systems and modules.
[0049] Preferably, the knowledge graph construction module constructs a knowledge graph including:
[0050] Collect text data through the open-source OCR engine Tesseract library, which supports reading multiple file formats; use the open-source model MinerU to intelligently extract knowledge content from learning resources, supporting redundant information; through data cleaning and extraction, convert the knowledge content extracted from learning resources into plain text format to lay the foundation for the construction of the knowledge graph; utilize the LightRAG framework and the APIs it provides, combined with locally deployed large models for AI knowledge extraction and graph construction, construct a knowledge graph that can reflect the internal relationships between knowledge points, and visualize the knowledge graph to help teachers and students intuitively view the relationships between various knowledge points.
[0051] Preferably, based on the knowledge graph constructed by the knowledge graph construction module, the intelligent recommendation module customizes personalized learning paths for students through the SASRec behavior sequence recommendation algorithm, including:
[0052] Arrange the learning behaviors of students in chronological order to form a behavior sequence; extract key features from the behavior data;
[0053] Convert the knowledge points and relationships in the knowledge graph into vector representations;
[0054] Dynamically update the knowledge graph based on the learning behavior data of students;
[0055] Use the relationships between knowledge points in the knowledge graph to determine the logical order of the learning path; recommend suitable knowledge points or learning resources according to the students' knowledge mastery; dynamically adjust the difficulty of the recommended content according to the difficulty level of the knowledge points;
[0056] Adopt the SASRec algorithm to capture the long-term dependencies in the student behavior sequence through an improved self-attention mechanism, and predict the possible content that students may learn next; recommend learning content that matches their interests and learning goals according to the students' historical behavior sequences;
[0057] Determine the end point of the learning path according to the students' learning goals; generate a learning path from the current knowledge point to the target knowledge point in combination with the relationships between knowledge points in the knowledge graph; dynamically adjust the order of knowledge points or recommended content in the path according to the students' learning progress and mastery;
[0058] Recommend the knowledge points that students may be interested in next according to the prediction results of the SASRec algorithm; recommend suitable learning resources in combination with the attributes of knowledge points in the knowledge graph; recommend learning content with moderate difficulty according to the students' knowledge mastery.
[0059] Preferably, the specific method for the SASRec algorithm to capture the long-term dependencies in the student behavior sequence through an improved self-attention mechanism and predict the possible content that students may learn next is:
[0060] Each learning behavior is first transformed into a vector representation of a fixed dimension through an embedding layer. These vectors are processed by an improved self-attention mechanism and a feed-forward neural network as inputs and then input to the output layer. The self-attention mechanism captures the dependencies between behaviors, while the feed-forward neural network further performs point-wise non-linear transformations on the features of each behavior to extract deep-level information to improve the prediction effect. Finally, the output layer generates predictions for the student's future learning behaviors or learning effects.
[0061] The beneficial effects produced by adopting the above technical solutions are as follows: The intelligent education method and system based on student behavior sequence recommendation provided by the present invention, through the construction of a knowledge graph identified by AI autonomously and an innovative knowledge fusion method, structurally and hierarchically display the curriculum knowledge system, clearly presenting the knowledge framework, which is conducive to realizing the construction of the knowledge graph framework of the whole-platform curriculum, efficiently associating prerequisite courses and subsequent courses, and providing students with a clear and complete curriculum learning path. It supports real-time response and dynamic update. By associating relevant resources, students can accurately obtain knowledge, improving learning efficiency and comprehension ability. The dynamic update mechanism ensures that the graph reflects the latest knowledge progress, enhancing the pertinence and effectiveness of learning, and helping students maintain the coherence and depth of learning in a constantly changing learning environment.
[0062] Adopting a recommendation algorithm based on student behavior sequences, combining SASRec (Self-Attention Sequence Recommendation) and a self-attention mechanism with a sparse attention module, can deeply analyze the student's behavior data and adjust the learning path according to real-time learning data, providing personalized learning suggestions for students. This technology combined with knowledge graph reasoning can more accurately recommend learning content and optimize the learning path in real time, further enhancing the student's learning experience and effect. Through the intelligent learning situation analysis module, teachers can obtain detailed reports on students' learning progress, learning status, and learning problems in real time, so as to precisely adjust teaching strategies and optimize the teaching process.
[0063] Compared with the prior art, the significant advantage of the present invention lies in its innovative integration of multi-dimensional functions such as knowledge graph construction, recommendation algorithms for student behavior sequences, and intelligent learning path optimization, with a wide coverage and strong comprehensiveness. By structurally displaying the knowledge system, personalized adjustment of learning strategies, intelligent generation of homework, and using the knowledge graph to optimize teaching and learning paths, the present invention not only greatly improves learning efficiency and teaching quality, but also reduces the burden on teachers, enhances the pertinence and effectiveness of learning, and provides strong support for cultivating cross-type composite talents, promoting the process of intelligent and high-quality development of the education industry. BRIEF DESCRIPTION OF THE DRAWINGS
[0064] Figure 1 It is a flowchart of the intelligent education method based on student behavior sequence recommendation provided in Embodiment 1 of the present invention;
[0065] Figure 2 This is the structural block diagram of the intelligent education system for student behavior sequence recommendation provided in Embodiment 2 of the present invention; Figure 3 This is the schematic diagram of AI autonomous recognition and extraction of the knowledge graph provided in Embodiment 2 of the present invention;
[0066] Figure 4 This is the schematic diagram of the behavior sequence recommendation process based on SASRec provided in Embodiment 2 of the present invention;
[0067] Figure 5 This is the schematic diagram of the student behavior sequence modeling process provided in Embodiment 2 of the present invention. Detailed implementation manners
[0068] The following combines the accompanying drawings and embodiments to further describe in detail the specific implementation manners of the present invention. The following embodiments are used to illustrate the present invention, but are not used to limit the scope of the present invention.
[0069] Embodiment 1:
[0070] In this embodiment, an intelligent education method based on student behavior sequence recommendation, as Figure 1 shown, includes the following steps:
[0071] Step S1: Obtain student behavior sequence data;
[0072] The student behavior sequence data is the learning behavior of students on the platform, including clicks, learning duration, homework completion status, test scores, etc.; The learning behavior sequence of each student can be expressed as:
[0073]
[0074] wherein, represents the i-th learning behavior of student u;
[0075] Step S2: Model the behavior sequence of students by using an improved self-attention mechanism (Self-Attention) introducing a sparse attention module;
[0076] In order to extract useful information from the learning behavior sequence of students, an improved self-attention mechanism (Self-Attention) introducing the sparse attention module in Diffuser is used to model the behavior sequence of students; The self-attention mechanism can capture long-term dependencies in the sequence and can dynamically adjust the degree of attention to historical behaviors according to the current learning task;
[0077] The sparse attention module improves the model's ability to capture key behaviors by reducing the attention to irrelevant behaviors, thereby improving the accuracy of learning path recommendation.
[0078] The model adopts an improved self-attention mechanism (Self-Attention) that incorporates the sparse attention module in Diffuser to model the student's behavior sequence. This model can efficiently process long sequence data and dynamically capture the long-term dependencies and short-term trends in the student's behavior sequence. Specifically, the sparse attention module reduces the model's attention to irrelevant behaviors by introducing sparsity, thereby improving the model's efficiency and accuracy.
[0079] The sparse attention module focuses on those behaviors that are most critical for learning path recommendations by dynamically adjusting the attention weights. For example, behaviors such as a student's clicks, learning duration, and homework completion during the learning process will be assigned higher weights, while those irrelevant behaviors (such as accidental clicks or short-term learning behaviors) will be ignored.
[0080] In the sparse attention module, the model reduces its attention to irrelevant behaviors through a certain sparsification method (such as Top-k selection or sparse matrix calculation).
[0081] Given a student's behavior sequence First, each behavior is mapped to an embedding vector e i ∈R d , where d is the embedding dimension; then, the improved self-attention mechanism that incorporates the sparse attention module in Diffuser is used to calculate the importance weights of each behavior;
[0082]
[0083] where Q, K, and V are the query matrix, key matrix, and value matrix obtained through linear transformation respectively, and d k is the dimension of the key matrix K; SparseSparse represents the sparsification operation; through the improved self-attention mechanism, the model can dynamically focus on the historical behaviors most relevant to the student's current learning task; the improved self-attention mechanism incorporates the sparse attention module in Diffuser, which can focus on key behaviors and ignore noise features;
[0084] Step S3: Obtain learning resource data and construct a knowledge graph;
[0085] The learning resource data includes various learning resources on the platform (such as videos, documents, exercises, etc.), and each resource is associated with specific knowledge points;
[0086] The knowledge graph is a structured knowledge representation form that contains the association relationships between knowledge points; each node in the knowledge graph represents a knowledge point, and the edges represent the associations between knowledge points (such as prerequisite relationships, dependency relationships, etc.). The knowledge graph is represented as:
[0087] G=(V, E)
[0088] where V is the set of knowledge points and E is the association relationship between knowledge points;
[0089] For example, the application of a certain mathematical formula may depend on the student's mastery of basic mathematical concepts. In this way, the knowledge graph forms a comprehensive and structured knowledge network.
[0090] In addition, the knowledge graph can also be dynamically adjusted and updated according to factors such as the student's learning records, behavior data, learning progress, etc., reflecting the student's current knowledge mastery and the knowledge areas to be learned.
[0091] Step S4: Perform knowledge graph reasoning to capture the association relationships between knowledge points;
[0092] Knowledge graph reasoning is used to capture the association relationships between knowledge points and recommend learning paths that match the student's current knowledge mastery; a graph neural network (Graph Neural Network, GNN) is used to reason about the knowledge graph; for a given knowledge graph, first initialize an embedding vector h v ∈R d for each knowledge point v ∈ V; then, update the embedding representation of the knowledge point through a multi-layer graph convolutional network (GCN):
[0093]
[0094] where N(v) represents the neighbor nodes of the knowledge point v, W(l) is the weight matrix of the l-th layer, and σ is the activation function; through multi-layer graph convolutional operations, the model can capture the high-order association relationships between knowledge points;
[0095] The embedding vector of each knowledge point is generated in the following way:
[0096] Content-based embedding: Use information such as the name and description of the knowledge point to generate an embedding vector through an NLP model (such as BERT); the specific formula is as follows:
[0097] h v =BERT(knowledge point name + knowledge point description)
[0098] Graph Structure-based Embedding: Use graph embedding algorithms (such as Node2Vec, GraphSAGE) to generate embedding vectors and capture the structural relationships between knowledge points. The specific formula is as follows:
[0099] h v = GraphSAGE(G, v)
[0100] where G is the knowledge graph and v is the current knowledge point.
[0101] Step S5: Generate a personalized learning path for the student based on the results of the student behavior sequence modeling and the knowledge graph reasoning, specifically including:
[0102] Student Knowledge Mastery Assessment: Evaluate the student's mastery of each knowledge point based on the student's behavior sequence and the knowledge graph reasoning results. Use a simple linear model to calculate the mastery score s of the student for the knowledge point v v :
[0103] s v = w T h v + b
[0104] where w and b are learnable parameters;
[0105] Learning Path Generation: Generate an optimal learning path from the knowledge graph based on the student's mastery score for the knowledge point. Use a shortest path algorithm (such as Dijkstra's algorithm) to find the shortest path from the knowledge point currently mastered by the student to the target knowledge point. Each node in the path represents a knowledge point, and the edge represents the prerequisite relationship between knowledge points;
[0106] Resource Recommendation: Recommend learning resources related to each knowledge point based on the generated learning path. Select the most suitable resources according to the difficulty, type (such as video, document, exercise, etc.) of the resources and the student's preferences.
[0107] Example 2:
[0108] In this example, a smart education system recommended based on the student behavior sequence, as Figure 2 shown, includes a front-end display layer, a gateway cluster layer, a Springboot layer, a middle platform service layer, a basic service layer, a storage technology layer, and a data integration layer;
[0109] The front-end display layer uses technologies such as the Vue framework, the ElementUI component library, and echarts to support multi-terminal interaction and data visualization, providing a convenient learning experience for users;
[0110] The gateway cluster layer realizes traffic management and security control through Gateway, and cooperates with the authentication and permission management of the underlying Springboot layer to ensure the stability and availability of the system;
[0111] The middle platform service layer is the core of the system, including a user management module, a course management module, a knowledge graph construction module, and an intelligent recommendation module;
[0112] The user management module is responsible for the creation, login, permission assignment, information editing, status management, and data analysis of user accounts to ensure the effective management and security of platform users;
[0113] The course management module is responsible for the creation, editing, publishing, enrollment, learning progress tracking, grade management, and evaluation feedback of courses to ensure the efficient management of course content and teaching quality;
[0114] The knowledge graph construction module constructs a knowledge graph based on learning resource data, and dynamically associates students, courses, and knowledge points;
[0115] The intelligent recommendation module, based on the knowledge graph constructed by the knowledge graph construction module, customizes personalized learning paths for students through the SASRec behavior sequence recommendation algorithm;
[0116] The basic service layer uses a microservice framework for service discovery, and tools for service configuration and traffic control provide efficient communication and traffic control for microservices;
[0117] The storage technology layer combines multiple databases (MySQL, Redis, and ES) to achieve efficient multi-type data storage;
[0118] The data integration layer uses tools such as Kafka and RabbitMQ to achieve efficient data transmission and message queue management, ensuring smooth data integration and asynchronous processing between different systems and modules.
[0119] In this embodiment, the specific method for the knowledge graph construction module to construct a knowledge graph is as follows:
[0120] Text data is collected through the open-source OCR engine Tesseract library, which supports reading various file formats such as PDF, DOCX, and PPTX. The open-source model MinerU is used to intelligently extract knowledge content from learning resources (such as PDF files, textbooks, courseware, etc.), supporting code and formula recognition, and removing redundant information such as headers and footers. Through data cleaning and extraction, the knowledge content extracted from learning resources is converted into a pure text format, laying a foundation for the construction of a knowledge graph. Using the LightRAG framework and its provided APIs, combined with the locally deployed Qwen large model and Zhipu glm-4-flash, AI knowledge extraction and graph construction are carried out to build a knowledge graph that can reflect the internal relationships between knowledge points. This graph will be visually displayed through the Neo4j database or HTML to help teachers and students intuitively view the relationships between various knowledge points.
[0121] In this embodiment, the algorithm implementation of AI autonomous recognition and extraction of the knowledge graph is as Figure 3 shown. This embodiment processes and analyzes a large amount of document data, including multiple key steps such as chunking, embedding, entity and relationship extraction, deduplication, and data storage. According to the above steps, AI autonomous recognition is performed to generate a knowledge graph.
[0122] The teacher uploads the teaching materials and lesson plans of a certain course on the platform to support the generation of a knowledge graph. The input documents will be split into smaller text chunks, which will be stored in a key-value (KV) storage system for subsequent processing. Then, in this embodiment, the EB model is applied to perform an embedding operation on each text chunk, converting it into vector-form data for easy knowledge entity extraction and relationship extraction, and the generated embedding vectors are stored in a vector database for subsequent querying and retrieval.
[0123] Next, entity and relationship extraction are carried out. Use the locally deployed Tongyi Qianwen or call large language models such as the Zhipu Qingyan api to identify entities and the relationships between them from the text chunks. In this process, it is necessary to perform deduplication processing on the entity and relationship data and update the description information of the entities as needed to ensure a reduction in the repetition rate of knowledge points. This step of entity and relationship extraction requires multiple iterations to improve the accuracy of the extracted knowledge graph.
[0124] As the uploaded resources are updated, it is necessary to continuously re-embed the updated data to reflect the latest information changes. This step is an embodiment of the dynamic update mechanism of the knowledge graph.
[0125] At different stages of the entire process, indexes will be created or updated to improve search and retrieval efficiency and provide structured knowledge data for the AI intelligent learning companion.
[0126] In this embodiment, the intelligent recommendation module constructs a knowledge graph based on the knowledge graph construction module, and through the SASRec behavior sequence recommendation algorithm, customizes a personalized learning path for students, such as Figure 4 as shown, including:
[0127] Arrange the learning behaviors of students in chronological order to form a behavior sequence; extract key features from the behavior data;
[0128] Convert the knowledge points and relationships in the knowledge graph into vector representations;
[0129] Dynamically update the knowledge graph based on the learning behavior data of students;
[0130] Use the knowledge point relationships in the knowledge graph (such as prerequisite relationships, dependency relationships) to determine the logical order of the learning path; recommend suitable knowledge points or learning resources according to the student's knowledge mastery; dynamically adjust the difficulty of the recommended content according to the difficulty level of the knowledge points;
[0131] Adopt the SASRec algorithm to capture the long-term dependencies in the student behavior sequence through an improved self-attention mechanism (Self-Attention), and predict the content that the student may learn next; recommend learning content that matches the student's interests and learning goals according to the student's historical behavior sequence;
[0132] Determine the end point of the learning path according to the student's learning goals (such as passing a certain exam, mastering a certain course); combine the knowledge point relationships in the knowledge graph to generate a learning path from the current knowledge point to the target knowledge point; dynamically adjust the knowledge point order or recommended content in the path according to the student's learning progress and mastery;
[0133] Recommend the knowledge points that the student may be interested in next according to the prediction results of the SASRec algorithm; recommend suitable learning resources in combination with the knowledge point attributes in the knowledge graph; recommend learning content with moderate difficulty according to the student's knowledge mastery level.
[0134] To further optimize the modeling of the student behavior sequence, the present invention adopts SASRec (Self-Attentive Sequential Recommendation) as the basic framework of the recommendation system, and combines sparse injection technology for transformation, introducing a sparse injection attention module based on Diffuser. SASRec is essentially an efficient sequence recommendation system that can capture the dependencies between behaviors through the self-attention mechanism and is suitable for the sequence modeling of student learning behaviors. However, the self-attention mechanism of traditional SASRec pays insufficient attention to sparse but key features when dealing with long sequences. Therefore, we strengthen the modeling ability of these key behaviors through sparse injection technology.
[0135] SASRec (Self-Attention-based Sequential Recommendation) is a sequential recommendation algorithm based on the self-attention mechanism, mainly applied to process and predict data based on time series or behavioral sequences. In an educational scenario, students' learning behavior data, such as learning time, learning progress, learning path, answering questions situation, etc., can be regarded as a time series. Through the self-attention mechanism, SASRec can capture the behavioral patterns and potential interests of students during the learning process, so as to provide personalized learning recommendations for students.
[0136] The main advantage of SASRec is that it can pay attention to the long-term dependence relationship of students' learning behaviors, and consider the dynamics and variability of the time series when making recommendations, generating more accurate personalized recommendation results.
[0137] The core idea of sparse injection is to enhance the influence of sparse but important features by assigning higher weights to them or directly injecting independent channels. In the analysis of students' learning behaviors, features such as "homework correct rate" and "effective video viewing rate" are sparse but highly predictive. By introducing the improved Transformer structure in the Diffuser model, a sparse attention module is added on the basis of the traditional self-attention mechanism. This module dynamically enhances the sparse key features in the sequence while suppressing the interference of irrelevant behaviors, thus significantly improving the expression ability of the model.
[0138] In the specific implementation, the input sequence of the model is first vectorized by the embedding layer for behaviors, and then input into the improved self-attention mechanism. This mechanism introduces the sparse attention module in Diffuser, which can focus on key behaviors and ignore noise features. Then, the feed-forward neural network further extracts deep information from the behavioral features at each time step to ensure that the influence of key behaviors can run through the entire prediction process. Finally, the output layer generates predictions for students' future learning behaviors or learning effects, such as recommending the next learning behavior or predicting exam scores.
[0139] This system can not only analyze students' daily learning behavior data (such as listening time, number of practice times, homework completion rate, etc.), but also dynamically adjust the personalized learning path planning. Through the combination of sparse injection Diffuser and the recommendation algorithm SASRec, the model captures the patterns of behavioral sequences while highlighting the role of key behaviors, significantly improving the accuracy and robustness of predictions. This provides a powerful technical solution for personalized learning recommendations in the field of education, and at the same time provides theoretical support for the optimal allocation of teaching resources and personalized tutoring.
[0140] In this embodiment, as Figure 5As shown, the SASRec algorithm captures long-term dependencies in the student behavior sequence through an improved self-attention mechanism, and the specific method for predicting the content that students may learn next is as follows:
[0141] The learning behaviors of students are a complex time series, and each behavior (such as attending classes, reading, doing exercises, taking exams, etc.) represents different stages in the learning process. In this context, the Transformer model in deep learning can effectively model these behavior sequences to predict students' future learning behaviors or learning effects.
[0142] Each learning behavior is first converted into a vector representation of a fixed dimension through an embedding layer, and these vectors are processed as inputs through an improved self-attention mechanism (Self-Attention Layer) and a feed-forward neural network (Feed Forward Network, FFN) and input into the output layer; the self-attention mechanism captures the dependencies between behaviors, such as the association between "review frequency" and "exam scores", while the feed-forward neural network further performs point-wise non-linear transformations on the features of each behavior to extract deep-level information to improve the prediction effect; finally, the output layer generates a prediction of the students' future learning behaviors or learning effects.
[0143] In this embodiment, the working process of the intelligent recommendation module is as follows:
[0144] a. Collection of student learning behavior data
[0145] The intelligent recommendation module first collects students' learning behavior data through various data sources (such as students' behavior records on the learning platform, homework submission situations, quiz results, etc.). These data contain various behaviors of students during the learning process, such as watching videos, completing homework, participating in discussions, clicking on certain course content, etc.
[0146] b. Dynamic update of the knowledge graph
[0147] Based on the students' learning behavior data, the knowledge graph will be dynamically updated. For example, when a student completes a learning task or answers a question correctly, the system will automatically update the node status of the student in the knowledge graph according to the student's behavior, marking their mastery of a certain knowledge point. In addition, if the student's learning path is not reasonable or there are omissions, the system will automatically adjust the recommended learning path to ensure that students can learn more efficiently.
[0148] c. The SASRec algorithm analyzes learning behaviors
[0149] Based on the historical learning behavior data of students, the SASRec algorithm analyzes the behavior sequences of students and extracts the implicit learning patterns in the sequences through the self-attention mechanism. For example, some students may tend to learn a certain basic mathematical knowledge point first and then learn higher-order knowledge; while other students may show different learning paths. The SASRec algorithm can predict the future learning needs and preferences of students according to their personalized learning patterns, and then recommend suitable learning content.
[0150] d. Personalized learning path recommendation
[0151] Combining the results of knowledge graph and SASRec algorithm analysis, the intelligent recommendation module customizes a personalized learning path for each student. This learning path is not only adjusted according to the student's current mastery, but also takes into account multi-dimensional factors such as the student's interests, learning speed, and learning difficulty. Based on this personalized recommendation, students can learn in the most suitable way for themselves, improving learning efficiency and effectiveness.
[0152] e. System output and feedback
[0153] The system will display the personalized learning path to the students through the recommendation interface, and suggest that the students gradually learn according to the recommended learning path. At the same time, the students' learning progress will be continuously fed back into the system, and the system will continue to optimize the knowledge graph and learning recommendations, forming a closed loop to ensure that students can always receive the most suitable learning resources for their current situation.
[0154] Students can not only obtain a personalized learning path, but also ensure the progression and suitability of the learning content at each stage. The knowledge graph provides a structured background of knowledge for the recommendation, while SASRec provides a deep understanding of the behavior sequences for the recommendation. This combination ensures the accuracy of the recommendation system and the personalized needs of students' learning.
[0155] The present invention focuses on the following key areas:
[0156] 1. Knowledge graph construction
[0157] Making full use of artificial intelligence technology, extracting valuable information from massive educational data, constructing a complete and accurate knowledge graph to help teachers and students understand the internal logic and structure of knowledge, and improving learning and teaching efficiency. In this process, not only promoting the digital transformation of educational content, but also promoting the deep integration of the physical education system and digital technology through accurate data analysis, forming a new model of efficient and sustainable intelligent education.
[0158] 2. Behavior sequence analysis to achieve educational intelligence
[0159] By analyzing the behavioral sequences of students during the programming process (such as code submission history, number of modifications, etc.), a basis is provided for teaching feedback and optimization. This innovative approach not only makes the education process more intelligent and precise, but also pushes the deep integration of physical education and digital technology to a new level, realizing the comprehensive digital transformation of education content and evaluation methods.
[0160] This invention proposes a comprehensive and systematic smart education construction and optimization method, deeply innovates the education and teaching process, improves teaching quality and learning efficiency, and provides students with accurate and personalized learning experience. At the same time, it actively explores the cross-application of AI technology in professional education, contributes to the digital transformation of cultivating compound talents, and promotes the intelligent transformation and high-quality development of the education field.
[0161] The key points of the present invention are the following aspects:
[0162] (1) Combining AI to display knowledge system in a structured and hierarchical manner
[0163] The present invention combines prompt engineering to realize a knowledge graph construction method for artificial intelligence to autonomously identify knowledge points and the correlation between knowledge points, and displays a complex curriculum system in a structured and hierarchical manner. This method not only reveals the intrinsic connection between knowledge points, but also constructs a clear knowledge framework, so that teachers and students can see the whole picture of the entire knowledge system at a glance. In this way, students can understand and learn each knowledge point more systematically, thereby improving learning efficiency and learning effect.
[0164] (2) Dynamic Update Mechanism
[0165] The knowledge graph has a powerful dynamic update capability and can reflect the latest knowledge progress in real time. When the teacher uploads new teaching materials, courseware and other resources and confirms the release, the knowledge graph on the student side of the course will be updated in real time to ensure that students can always access the most cutting-edge knowledge. At the same time, teachers can also flexibly edit a certain knowledge node in the knowledge graph to review and improve the specific knowledge content of the node. This flexible operation method not only improves the practicality and credibility of the knowledge graph, but also ensures that the platform courses receive the latest knowledge content in real time, always keep pace with the development of the times, and thus enhance the adaptability and scalability of the system.
[0166] (3) Student behavior sequence recommendation algorithm and intelligent learning path optimization
[0167] The present invention adopts a recommendation algorithm based on student behavior sequence, combined with SASRec (self-attention sequence recommendation) and the self-attention mechanism improved by introducing a sparse attention module. These algorithms can perform in-depth analysis of students' behavior data, adjust the learning path according to real-time data in the learning process, and provide personalized learning suggestions. This technology, combined with knowledge graph reasoning, can more accurately recommend learning content and optimize learning paths in real time, further improving students' learning experience and effects. The application of this algorithm not only improves the accuracy of personalized recommendations, but also enhances the system's ability to respond to students' dynamic learning needs. In addition, the system also generates a detailed learning situation analysis report through learning situation analysis, helping teachers to understand students' learning status in real time, grasp each student's learning progress and mastery. This function provides teachers with a comprehensive analysis view, supports teachers to adjust teaching strategies and methods in a targeted manner, optimizes the teaching process, and further improves the quality of education and teaching effects.
[0168] (4) Multi-agent orchestration model and machine learning algorithm
[0169] Using the optimized multi-agent collaborative algorithm model, questions are automatically generated according to knowledge points or ability assessment requirements, and AI automatic grading and question analysis are realized. This function not only greatly improves the efficiency of homework generation, but also ensures the pertinence and diversity of questions, thereby better meeting students' learning needs. Through AI grading and analysis, students can understand their learning situation in a timely manner, adjust their learning strategies, and improve their learning effects.
[0170] (5) Learning status supervision
[0171] The system monitors students' learning status in real time by recording data such as students' learning time, learning progress, correctness of questions, number of code submissions, and modification history. This function helps teachers to find students' learning problems in a timely manner and take targeted teaching measures. Through comprehensive monitoring and analysis of students' learning status, teachers can understand each student's learning progress more accurately, thereby providing more personalized teaching support and further improving the quality of education.
[0172] (6) Technical support based on domestic large models
[0173] This embodiment relies on the domestic Qwen2.5 large model and Zhipu technology to realize functions such as knowledge point extraction and AI intelligent learning companion. By making full use of the advantages of domestic large models, it can provide students with a more efficient and convenient learning experience, while promoting the intelligent process of the education industry and promoting the deep integration of education and technology.
[0174] In summary, the system of the present invention can not only analyze the daily learning behavior data of students (such as the length of listening to lectures, the number of practice times, the completion rate of homework, etc.), but also dynamically adjust the personalized learning path planning. Through the combination of the sparse injection Diffuser and the recommendation algorithm SASRec, while capturing the patterns of behavior sequences, the model highlights the role of key behaviors, significantly improving the accuracy and robustness of prediction. This provides a powerful technical solution for personalized learning recommendations in the field of education, and at the same time provides theoretical support for the optimal allocation of teaching resources and personalized tutoring.
[0175] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present invention.
Claims
1. A smart education method based on student behavior sequence recommendation, characterized by: include: Obtain student behavior sequence data; An improved self-attention mechanism with sparse attention module is used to model students’ behavior sequence; Obtain learning resource data and build a knowledge graph; Perform knowledge graph reasoning to capture the relationship between knowledge points; Based on the results of student behavior sequence modeling and knowledge graph reasoning, a personalized learning path is generated for students.
2. The intelligent education method based on student behavior sequence recommendation according to claim 1, characterized in that: The student behavior sequence data is the learning behavior of students on the platform; the learning behavior sequence of each student is expressed as: in, represents the i-th learning behavior of student u; The learning resource data includes various types of learning resources on the platform, each resource being associated with a specific knowledge point; The knowledge graph is a structured knowledge representation form, which contains the association relationship between knowledge points. Each node in the knowledge graph represents a knowledge point, and the edge represents the association between knowledge points. The knowledge graph is represented as: G=(V,E) Among them, V is the set of knowledge points, and E is the association relationship between knowledge points.
3. The intelligent education method based on student behavior sequence recommendation according to claim 2 is characterized by: The improved self-attention mechanism using the sparse attention module is used to model the student's behavior sequence, including: Given a student's behavior sequence First, each behavior Mapped to an embedding vector e i ∈R d , where d is the embedding dimension; then, the improved self-attention mechanism introduced into the sparse attention module in Diffuser is used to calculate the importance weight of each behavior; Among them, Q, K, and V are respectively the query matrix, key matrix, and value matrix obtained by linear transformation, d k is the dimension of the key matrix K; SparseSparse represents the sparsification operation.
4. The intelligent education method based on student behavior sequence recommendation according to claim 3 is characterized by: The knowledge graph reasoning is used to capture the association between knowledge points and recommend learning paths that match students' current knowledge mastery; Use graph neural network GNN to reason about knowledge graphs; for a given knowledge graph, first initialize an embedding vector h for each knowledge point v∈V v ∈R d ; Then, the embedded representation of the knowledge point is updated through the multi-layer graph convolutional network GCN: Among them, N(v) represents the neighbor nodes of knowledge point v, W(l) is the weight matrix of the lth layer, and σ is the activation function.
5. The intelligent education method based on student behavior sequence recommendation according to claim 4 is characterized by: The embedding vector of each knowledge point is generated in the following way: Content-based embedding: Use the name and description information of knowledge points to generate embedding vectors through the NLP model; Graph structure-based embedding: Use graph embedding algorithms to generate embedding vectors to capture the structural relationships between knowledge points.
6. The intelligent education method based on student behavior sequence recommendation according to claim 4 is characterized by: The above method combines the results of student behavior sequence analysis and knowledge graph reasoning to generate a personalized learning path for students, including: Student knowledge mastery assessment: Based on the student's behavior sequence and knowledge graph reasoning results, evaluate the student's mastery of each knowledge point; use a linear model to calculate the student's mastery score s for the knowledge point v v : s v =w T h v +b Among them, w and b are learnable parameters; Learning path generation: Generate an optimal learning path from the knowledge graph based on the student's mastery of the knowledge points; use the shortest path algorithm to find the shortest path from the student's current knowledge point to the target knowledge point; each node in the path represents a knowledge point, and the edge represents the prerequisite relationship between the knowledge points; Resource recommendation: Based on the generated learning path, recommend learning resources related to each knowledge point; select the most suitable resource based on the difficulty, type and student preferences of the resource.
7. A smart education system based on student behavior sequence recommendation, implemented based on the method of claim 1, characterized in that: Including front-end display layer, gateway cluster layer, Springboot layer, middle-end service layer, basic service layer, storage technology layer and data integration layer; The front-end display layer supports multi-terminal interaction and data visualization, providing users with a convenient learning experience; The gateway cluster layer implements traffic management and security control through Gateway, and cooperates with the authentication and permission management of the underlying Springboot layer to ensure the stability and availability of the system; The middle platform service layer is the core of the system, including user management module, course management module, knowledge graph construction module and intelligent recommendation module; The user management module is responsible for user account creation, login, authority allocation, information editing, status management and data analysis to ensure effective management and security of platform users; The course management module is responsible for course creation, editing, publishing, registration, learning progress tracking, performance management and evaluation feedback, ensuring efficient management of course content and teaching quality; The knowledge graph construction module constructs a knowledge graph based on learning resource data, dynamically associating students, courses and knowledge points; The intelligent recommendation module is based on the knowledge graph constructed by the knowledge graph construction module, and uses the SASRec behavior sequence recommendation algorithm to tailor personalized learning paths for students; The basic service layer uses the microservice framework to perform service discovery, configure services and flow control tools to provide efficient communication and flow control for microservices; The storage technology layer combines multiple databases to achieve efficient multi-type data storage; The data integration layer implements efficient data transmission and message queue management, ensuring smooth integration and asynchronous processing of data between different systems and modules.
8. The intelligent education system based on student behavior sequence recommendation according to claim 7 is characterized by: The knowledge graph construction module constructs the knowledge graph including: Text data is collected through the open source OCR engine testract library, which supports reading of multiple file formats; the open source model MinerU is used to intelligently extract knowledge content from learning resources and supports redundant information; through data cleaning and extraction, the knowledge content extracted from learning resources is converted into plain text format, laying the foundation for the construction of knowledge graphs; the LightRAG framework and the API it provides are used in combination with the locally deployed large model for AI knowledge extraction and graph construction, to build a knowledge graph that can reflect the internal relationship of knowledge points, and the knowledge graph is visualized to help teachers and students intuitively view the relationship between each knowledge point.
9. The intelligent education system based on student behavior sequence recommendation according to claim 8, characterized in that: The intelligent recommendation module is based on the knowledge graph constructed by the knowledge graph construction module, and uses the SASRec behavior sequence recommendation algorithm to tailor personalized learning paths for students, including: Arrange students' learning behaviors in chronological order to form a behavior sequence; extract key features from the behavior data; Convert knowledge points and relationships in the knowledge graph into vector representations; Dynamically update the knowledge graph based on students’ learning behavior data; Use the relationship between knowledge points in the knowledge graph to determine the logical order of the learning path; recommend appropriate knowledge points or learning resources based on the students' knowledge mastery; dynamically adjust the difficulty of the recommended content based on the difficulty level of the knowledge points; The SASRec algorithm is used to capture the long-term dependencies in the student behavior sequence through an improved self-attention mechanism to predict what the student may learn next; based on the student's historical behavior sequence, the learning content that matches their interests and learning goals is recommended; Determine the end point of the learning path based on the student's learning goals; generate a learning path from the current knowledge point to the target knowledge point based on the knowledge point relationship in the knowledge graph; dynamically adjust the order of knowledge points or recommended content in the path based on the student's learning progress and mastery; Based on the prediction results of the SASRec algorithm, we recommend knowledge points that students may be interested in next. Based on the attributes of knowledge points in the knowledge graph, we recommend suitable learning resources. Based on the students' knowledge mastery, we recommend learning content of moderate difficulty.
10. The intelligent education system based on student behavior sequence recommendation according to claim 9, characterized in that: The SASRec algorithm captures the long-term dependencies in the student behavior sequence through an improved self-attention mechanism, and the specific method for predicting what the student may learn next is: Each learning behavior is first converted into a vector representation of fixed dimension through an embedding layer. These vectors are input into the output layer after being processed by an improved self-attention mechanism and a feedforward neural network. The self-attention mechanism captures the dependencies between behaviors, while the feedforward neural network further performs point-by-point nonlinear transformation on the features of each behavior, extracting deep information to improve the prediction effect. Finally, the output layer generates a prediction of the student's future learning behavior or learning effect.
Citation Information
Cited By
AI-enabled intelligent teaching personalized service method and system
CN120542747A
An AI-enabled intelligent teaching personalized service method and system
CN120542747B
Educational resource sequence recommendation method and related device
CN120780915A
Education-oriented AI working table application capability construction and resource matching system
CN120852114A
Special management method and system for education research platform
CN120876177A
