Learning Content Recommendation Method and System Based on Knowledge Tracing

Abstract

This application relates to the technical field of content recommendation, and discloses a learning content recommendation method and system based on knowledge tracing. The method includes: obtaining all the knowledge points that the learner needs to learn and constructing an initial knowledge graph; collecting and preprocessing knowledge state data to obtain a knowledge map; performing feature matching on the knowledge map and the knowledge graph to obtain multiple candidate learning paths; calculating the path comprehensive score of each candidate learning path through a multi-feature fusion path analysis model; selecting multiple target learning paths and generating recommended learning content for the learner; collecting the state feedback data of the learner, and updating and optimizing the initial knowledge graph and the multi-feature fusion path analysis model through a probabilistic multi-objective reinforcement learning algorithm to obtain a target knowledge graph and a target path analysis model. This application improves the accuracy of learning content recommendation based on knowledge tracing.

Description

Technical Field

[0001] This application relates to the technical field of content recommendation, and particularly to a learning content recommendation method and system based on knowledge tracing. Background Art

[0002] In today's information society, the education industry is undergoing profound changes, and personalized learning has gradually become an important means to improve the quality of education. However, the traditional teaching mode often fails to meet the personalized needs of each learner, resulting in low teaching efficiency and uneven knowledge mastery among learners. To address these challenges, a learning content recommendation method based on knowledge tracing has emerged. Its core lies in using big data and artificial intelligence technologies to track the knowledge state of learners in real time and dynamically adjust learning content to achieve a personalized learning experience.

[0003] Currently, although some personalized learning recommendation systems have been put into use, most systems still face many problems in actual applications. These systems often lack a comprehensive and dynamic assessment of the knowledge state of learners, making it difficult to accurately grasp the learning progress and knowledge mastery of learners, resulting in the recommended learning content not matching the actual needs. Secondly, existing methods usually only focus on the optimization of a single goal, such as learning efficiency or knowledge mastery, and fail to simultaneously balance and optimize multiple goals, ignoring the interests and personalized needs of learners. In addition, traditional systems lack an effective dynamic adjustment mechanism when dealing with learner feedback and adjusting the learning path, resulting in the difficulty of timely updating and optimizing the recommended learning path. Summary of the Invention

[0004] This application provides a learning content recommendation method and system based on knowledge tracing, which is used to improve the accuracy of learning content recommendation based on knowledge tracing.

[0005] In a first aspect, this application provides a learning content recommendation method based on knowledge tracing. The learning content recommendation method based on knowledge tracing includes:

[0006] Obtain all the knowledge points that the learner needs to learn, classify the knowledge points and construct a relationship graph to obtain an initial knowledge graph;

[0007] Collect and preprocess the knowledge state data of the learner to obtain the knowledge map of the learner;

[0008] Perform feature matching between the knowledge map and the knowledge graph to obtain multiple candidate learning paths;

[0009] Input the multiple candidate learning paths into a pre-set multi-feature fusion path analysis model respectively to calculate the learning path scores, and obtain the comprehensive path scores of each candidate learning path;

[0010] Select multiple target learning paths according to the comprehensive path score and the preset target value, and generate the recommended learning content for the learner according to the multiple target learning paths;

[0011] Collect the status feedback data of the learner, and update and optimize the initial knowledge graph and the multi-feature fusion path analysis model through a probabilistic multi-objective reinforcement learning algorithm to obtain a target knowledge graph and a target path analysis model.

[0012] In a second aspect, the present application provides a learning content recommendation system based on knowledge tracing. The learning content recommendation system based on knowledge tracing includes:

[0013] A construction module, configured to obtain all the knowledge points that the learner needs to learn, classify the knowledge points, and construct a relationship graph to obtain an initial knowledge graph;

[0014] An acquisition module, configured to collect and preprocess the knowledge state data of the learner to obtain the knowledge map of the learner;

[0015] A matching module, configured to perform feature matching between the knowledge map and the knowledge graph to obtain multiple candidate learning paths;

[0016] A calculation module, configured to input the multiple candidate learning paths into a preset multi-feature fusion path analysis model respectively to calculate the learning path score, and obtain the comprehensive path score of each candidate learning path;

[0017] A selection module, configured to select multiple target learning paths according to the comprehensive path score and the preset target value, and generate the recommended learning content for the learner according to the multiple target learning paths;

[0018] An optimization module, configured to collect the status feedback data of the learner, and update and optimize the initial knowledge graph and the multi-feature fusion path analysis model through a probabilistic multi-objective reinforcement learning algorithm to obtain a target knowledge graph and a target path analysis model.

[0019] In the technical solution provided by this application, by collecting and preprocessing the knowledge state data of learners, using online test data, homework grading data, and classroom question data, and combining technologies such as data cleaning, normalization, and principal component feature extraction, an accurate knowledge map is constructed. The K-means clustering algorithm and Bayesian network are adopted to classify knowledge points and analyze causal relationships, and an initial knowledge graph is constructed. It can dynamically generate and adjust the learning path to ensure that the recommended learning path highly matches the actual needs of learners, improving learning efficiency and effect. Through the probabilistic multi-objective reinforcement learning algorithm, the present invention can continuously update and optimize the initial knowledge graph and path analysis model. According to the state feedback data of learners, it dynamically adjusts the learning content and path, providing personalized learning resources and recommended content. Learners can make efficient progress on a learning path that better meets their own needs and interests, improving learning satisfaction and effect. By continuously collecting the state feedback data of learners and using the feedback information for reinforcement learning and model update, it is ensured that the knowledge graph and path analysis model always remain consistent with the actual situation of learners. By regularly evaluating and generating learning effect reports, the recommendation system is continuously improved and optimized, improving the accuracy and personalization level of learning effects, and ensuring that learners can continuously obtain the best learning experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are 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.

[0021] Figure 1 It is a schematic diagram of an embodiment of the learning content recommendation method based on knowledge tracing in the embodiments of this application;

[0022] Figure 2 It is a schematic diagram of an embodiment of the learning content recommendation system based on knowledge tracing in the embodiments of this application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0023] The embodiments of the present application provide a learning content recommendation method and system based on knowledge tracing. The terms "first", "second", "third", "fourth", etc. (if any) in the description, claims and the above-mentioned drawings of the present application are used to distinguish similar objects, and do not have to be used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments described here can be implemented in an order other than that illustrated or described here. In addition, the term "comprising" or "having" and any variation thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device comprising a series of steps or units does not have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0024] For ease of understanding, the specific process of the embodiments of the present application will be described below. Please refer to Figure 1 , an embodiment of the learning content recommendation method based on knowledge tracing in the embodiments of the present application includes:

[0025] Step S101, obtain all the knowledge points that the learner needs to learn, classify the knowledge points and construct a relationship graph to obtain an initial knowledge graph;

[0026] It can be understood that the execution subject of the present application can be a learning content recommendation system based on knowledge tracing, or a terminal or a server. Specifically, no limitation is made here. The embodiments of the present application will be described by taking the server as the execution subject as an example.

[0027] Specifically, all the knowledge points that the learner needs to learn are obtained. These knowledge points are classified by the K-means clustering algorithm. Utilizing the efficiency and accuracy of the K-means clustering algorithm, a large number of knowledge points are reasonably classified according to their characteristics, and the classified knowledge points are obtained. The causal relationship analysis of the classified knowledge points is carried out through the Bayesian network. As a probability model, the Bayesian network can reveal the causal relationship between knowledge points, and a relationship information set is obtained. Based on the Neo4j graph database, node connections are made for the relationship information set and the classified knowledge points, and a node connection graph is obtained. As an efficient graph database, Neo4j can handle complex node relationships. By making node connections for the relationship information set and the knowledge points, a complete node connection graph is formed. Weight assignment is performed on the node connection graph to obtain a weighted relationship graph. The importance of the relationships between knowledge points is highlighted. By assigning different weights, important relationships are made more prominent, and a weighted relationship graph is formed. Then, the weighted relationship graph is visualized to obtain a graphical knowledge graph. The complex relationship graph is displayed in a visual way, making the knowledge graph more intuitive and easy to understand. Hierarchical division is carried out on the graphical knowledge graph to obtain a hierarchical knowledge graph. To make the knowledge graph more structured, through hierarchical division of the knowledge graph, the relationships between knowledge points at each level become clearer, and consistency verification is carried out on the hierarchical knowledge graph to ensure the logic and consistency of the knowledge graph, and finally an initial knowledge graph is obtained.

[0028] Step S102: Collect and preprocess the knowledge state data of the learner to obtain the knowledge map of the learner;

[0029] Specifically, knowledge state data of learners is collected to obtain initial knowledge state data, which includes online test data, homework grading data, and classroom question data. The online test data reflects the learners' mastery of different knowledge points; the homework grading data provides the learners' performance in practical applications; the classroom question data records the learners' performance in classroom interactions. These data comprehensively reflect the learners' knowledge state. The initial knowledge state data is cleaned by removing invalid data and handling outliers to obtain the cleaned knowledge state data. The cleaned knowledge state data is normalized to eliminate the dimensional differences between different data, making the data comparable. Through normalization, the normalized knowledge state data is obtained, enabling the analysis and processing of data from different sources on the same scale. Principal component feature extraction is performed on the knowledge state data. Principal component feature extraction is a dimensionality reduction technique that reduces the dimensionality of data, reduces data complexity, and retains the main information of the data by extracting the main features of the data. Through principal component feature extraction, knowledge state feature data is obtained, which can better reflect the learners' knowledge state. The knowledge state feature data is mapped to a knowledge map through the BKT model. The BKT model (Bayesian Knowledge Tracing) is a probability-based knowledge tracing model that can establish a knowledge map of learners by tracking and predicting the learners' knowledge state. By inputting the knowledge state feature data into the BKT model, the model can predict the learners' mastery of each knowledge point based on the learners' performance on different knowledge points, and obtain the learners' knowledge map. The knowledge map is an intuitive display of the learners' knowledge state, reflecting the learners' mastery of different knowledge points and knowledge structure.

[0030] Step S103: Perform feature matching on the knowledge map and the knowledge graph to obtain multiple candidate learning paths;

[0031] Specifically, by extracting nodes from the learner's knowledge map, each knowledge node therein is identified, and these nodes represent the specific knowledge points of the learner in the knowledge structure. Similarly, by extracting nodes from the knowledge graph, graph nodes are obtained, and these nodes represent the key knowledge points in the entire knowledge domain. By extracting nodes from the two graphs, the correspondence between the learner's current knowledge state and the complete knowledge structure can be clarified. Calculate the similarity between the knowledge nodes and the graph nodes to obtain an initial similarity matrix. The similarity calculation can be based on various methods, such as cosine similarity, Euclidean distance, or content-based matching algorithms. The initial similarity matrix reflects the similarity degree between each knowledge node and the graph node. Perform threshold screening on the initial similarity matrix, set a similarity threshold, and screen out the similarity pairs higher than this threshold to obtain a target similarity matrix. Through threshold screening, irrelevant node pairs with low similarity are removed, and those node pairs with high similarity are retained to ensure the accuracy of feature matching. Conduct feature matching analysis on the target similarity matrix, and perform detailed feature matching based on the node pairs with high similarity to obtain feature matching results. Generate paths based on the feature matching results. By analyzing the connectivity between knowledge nodes, multiple initial learning paths are generated, and these paths represent multiple possible learning routes from the current knowledge state to the target knowledge structure. Calculate the weights of the multiple initial learning paths. Based on the importance and connectivity of each knowledge node in the path, calculate the weight of each initial learning path to obtain multiple weighted learning paths. The weight calculation can consider factors such as the learning difficulty of knowledge points and the relationship strength between knowledge points, so that the path weight can reflect the quality of the learning path. Use the dynamic programming algorithm to screen the multiple weighted learning paths. The dynamic programming algorithm screens out the best combination of learning paths through a comprehensive evaluation of the paths to obtain multiple candidate learning paths. The candidate learning paths are optimized and screened learning routes that can effectively guide the learner to gradually master the complete knowledge structure from the current knowledge state.

[0032] Step S104: Input the multiple candidate learning paths into a pre-set multi-feature fusion path analysis model respectively to calculate the learning path score, and obtain the comprehensive path score of each candidate learning path;

[0033] Specifically, feature extraction is performed on multiple candidate learning paths to obtain multiple path features for each candidate learning path. These path features may include key knowledge points in the learning path, path coherence, learning time, and difficulty coefficients of knowledge points, etc. Feature normalization and feature vector encoding are performed on the multiple path features of each candidate learning path. The normalization process can eliminate the dimensional differences between different features, making the feature data more unified and comparable; feature vector encoding is to convert the normalized feature data into vector form for subsequent model input and calculation. Through feature normalization and feature vector encoding, a path feature encoding vector for each candidate learning path is obtained. The path feature encoding vectors of each candidate learning path are input into a preset multi-feature fusion path analysis model, which includes a multi-head self-attention mechanism, multiple deep neural networks, and a fully connected layer. The multi-head self-attention mechanism analyzes the path feature encoding vectors. The multi-head self-attention mechanism can capture the correlations and mutual influences between different features and generate multiple path feature attention vectors, which can highlight the key parts of the features in the path. The multiple path feature attention vectors are input into multiple deep neural networks for feature relationship and feature weight analysis. The deep neural network has non-linear modeling capabilities. By analyzing the attention vectors, it can deeply explore the complex relationships and importance between various features, and obtain the feature relationships and feature weights of each path feature attention vector. These analysis results reflect the roles and contributions of path features in the overall learning path scoring. In the fully connected layer, learning path scoring calculation and weighted fusion are performed on the multiple path feature attention vectors according to the feature relationships and feature weights. The fully connected layer can generate the final comprehensive score for each candidate learning path through comprehensive analysis and weighted processing of each feature. The comprehensive score not only considers the independent contributions of each feature but also the interactive effects between them.

[0034] Step S105: Select multiple target learning paths according to the path comprehensive score and a preset target value, and generate recommended learning content for the learner based on the multiple target learning paths;

[0035] Specifically, multiple target learning paths are selected based on the comprehensive path score and the preset target value. By analyzing the comprehensive scores of each candidate learning path, the paths with scores higher than the preset target value are selected as the target learning paths. The learning content of the multiple target learning paths is transformed. Each knowledge point or learning task on the path is converted into specific learning content. The learning content can include video explanations, courseware materials, practice questions, reading materials, etc. In this way, the abstract learning path is transformed into specific learning content, obtaining multiple first learning contents, enabling learners to directly carry out learning. The multiple first learning contents are organized. The transformed learning content is structured to have a reasonable learning sequence and logic, obtaining multiple second learning contents. These second learning contents are arranged according to the front-back relationship of knowledge points and the difficulty level of learning tasks, enabling learners to learn step by step, avoiding content chaos and repetition, and improving learning efficiency. After the content organization is completed, the learning styles and interests of the learners are obtained to ensure that the recommended learning content is more personalized and meets the needs of the learners. The learning styles of learners can be visual, auditory, kinesthetic, etc. Different learning styles have different preferences for the presentation form of learning content; while the interests of learners can guide the selection and arrangement of content, making the learning process more interesting and attractive. The multiple second learning contents are integrated to generate the recommended learning content for the learners. The organized learning content is matched and adjusted with the learning styles and interests of the learners. By integrating different types of learning resources, the recommended learning content not only meets the requirements of the learning path but also can meet the personalized needs of the learners. The integrated learning content can include various forms of learning resources, such as videos, texts, charts, and interactive exercises, providing a comprehensive and diverse learning experience.

[0036] Step S106: Collect the status feedback data of the learner, and update and optimize the initial knowledge graph and the multi-feature fusion path analysis model through the probability multi-objective reinforcement learning algorithm to obtain the target knowledge graph and the target path analysis model.

[0037] Specifically, based on the recommended learning content, the state feedback data of the learner is collected. These data include the learner's performance after completing various learning tasks, such as test scores, homework completion, learning time, and interaction participation. The knowledge mastery vector and learning task feature vector of the learner are defined according to the state feedback data. The knowledge mastery vector reflects the learner's mastery of different knowledge points. By analyzing the test scores and homework completion, the mastery of each knowledge point by the learner is quantified; the learning task feature vector describes the characteristics of each learning task, including task difficulty, task type, and task completion time. The knowledge mastery vector and learning task feature vector are input into the probabilistic multi-objective reinforcement learning algorithm for probability calculation and multi-objective analysis to obtain multi-objective feedback information. The probabilistic multi-objective reinforcement learning algorithm is a multi-objective optimization method based on a probability model that can consider multiple optimization objectives simultaneously. By analyzing and calculating the input vector, the performance prediction and optimization suggestions of the learner on different learning paths are obtained. The multi-objective feedback information is used for reinforcement learning to obtain the reinforcement learning result. Reinforcement learning is a machine learning method that continuously optimizes strategies through a trial-and-error and feedback mechanism. By performing reinforcement learning on the multi-objective feedback information, the learning path and learning content recommendation strategy can be gradually optimized to improve the learning effect. The reinforcement learning result includes the optimized knowledge point mastery strategy and learning task arrangement plan, and these results will be used as the basis for updating the knowledge graph and path analysis model. According to the reinforcement learning result, the initial knowledge graph is updated to obtain the target knowledge graph. The knowledge graph update reflects the learner's latest learning performance and knowledge mastery in the knowledge graph. By adding new knowledge points, updating the knowledge point relationships, and adjusting the knowledge point weights, the knowledge graph can more accurately and dynamically reflect the learner's knowledge state. The obtained target knowledge graph can provide a more accurate basis for subsequent learning path recommendations. According to the reinforcement learning result, the multi-feature fusion path analysis model is updated to obtain the target path analysis model. The path analysis model update adjusts the feature weights and path scoring mechanism in the path analysis model according to the reinforcement learning result, so that the model can better reflect the learner's actual learning situation and optimization suggestions. By updating the model, the accuracy and effectiveness of path analysis and recommendation are improved.

[0038] In the embodiments of the present application, by collecting and preprocessing the knowledge state data of learners, using online test data, homework grading data, and classroom question data, and combining technologies such as data cleaning, normalization, and principal component feature extraction, an accurate knowledge map is constructed. The K-means clustering algorithm and Bayesian network are adopted to classify knowledge points and analyze causal relationships, and an initial knowledge graph is constructed. It can dynamically generate and adjust the learning path to ensure that the recommended learning path highly matches the actual needs of learners, improving learning efficiency and effect. Through the probabilistic multi-objective reinforcement learning algorithm, the present invention can continuously update and optimize the initial knowledge graph and path analysis model. According to the state feedback data of learners, it dynamically adjusts the learning content and path, providing personalized learning resources and recommended content. Learners can make efficient progress on a learning path that better meets their own needs and interests, improving learning satisfaction and effect. By continuously collecting the state feedback data of learners and using the feedback information for reinforcement learning and model update, it is ensured that the knowledge graph and path analysis model are always consistent with the actual situation of learners. By regularly evaluating and generating learning effect reports, the recommendation system is continuously improved and optimized to improve the accuracy and personalization level of learning effects, ensuring that learners can continuously obtain the best learning experience.

[0039] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0040] (1) Obtain all the knowledge points that learners need to learn, and classify the knowledge points through the K-means clustering algorithm to obtain the classified knowledge points;

[0041] (2) Analyze the causal relationships of the classified knowledge points through the Bayesian network to obtain a set of relationship information;

[0042] (3) Based on the Neo4j graph database, connect the nodes of the set of relationship information and the classified knowledge points to obtain a node connection graph;

[0043] (4) Assign weights to the node connection graph to obtain a weighted relationship graph, and graphically represent the weighted relationship graph to obtain a graphical knowledge graph;

[0044] (5) Divide the graphical knowledge graph into layers to obtain a layered knowledge graph, and perform consistency verification on the layered knowledge graph to obtain an initial knowledge graph.

[0045] Specifically, all the knowledge points that the learner needs to learn are obtained. The knowledge points are classified by the K-means clustering algorithm. The K-means algorithm is a clustering method whose goal is to divide data points into k clusters, such that the data points within the same cluster are as similar as possible, while the data points between different clusters are as different as possible. In specific implementation, the knowledge points are vectorized, that is, each knowledge point is represented as a feature vector X i . Initialize k cluster centers, and through iterative calculation of the distances from each knowledge point to each cluster center, assign the knowledge points to the nearest cluster center, and update the positions of the cluster centers until the clustering result converges or reaches the predetermined number of iterations. The mathematical expression is:

[0046]

[0047] where J is the objective function, representing the total loss of clustering, k is the number of clusters, n i is the number of knowledge points in the i-th cluster, is the j-th knowledge point in the i-th cluster, μ i is the cluster center of the i-th cluster, Indicates the distance between the knowledge points and the cluster centers. Through the K-means clustering algorithm, a set of classified knowledge points is obtained. For example, knowledge points related to algebra are clustered into one category, knowledge points related to geometry are clustered into another category, and so on. The causal relationship analysis of the classified knowledge points is carried out through a Bayesian network. A Bayesian network is a probabilistic graphical model that can represent the causal relationship between variables. Construct a Bayesian network structure, where nodes represent knowledge points and edges represent the causal relationship between knowledge points. Through the statistical analysis of the learner's historical learning data, the conditional probability distribution of each edge is estimated to obtain a set of relationship information. Based on the Neo4j graph database, the nodes of the relationship information set and the classified knowledge points are connected to obtain a node connection graph. Neo4j is an efficient graph database that is good at processing complex node and edge relationships. In specific implementation, each knowledge point is used as a node, and each causal relationship is used as an edge and stored in the Neo4j database to form a node connection graph. This node connection graph intuitively shows the causal relationship structure between each knowledge point. Weight assignment is performed on the node connection graph to obtain a weighted relationship graph. The weight assignment determines the weight of the edge according to the conditional probability distribution in the Bayesian network. The weighted relationship graph is visualized to obtain a visualized knowledge graph. Visualization is to display the complex weighted relationship graph in an intuitive way. Visualization tools such as Gephi and D3.js can be used to layout and draw the nodes and edges to make the structure of the knowledge graph clearer and easier to understand. Hierarchical division is performed on the visualized knowledge graph to obtain a hierarchical knowledge graph. Hierarchical division is to better organize and display the hierarchical relationship between knowledge points, and can be implemented by using hierarchical clustering algorithms or community discovery algorithms. The hierarchical clustering algorithm divides the knowledge points into different levels, so that the upper-level knowledge points can be used as the basis for the lower-level knowledge points. Consistency verification is performed on the hierarchical knowledge graph to obtain an initial knowledge graph. Consistency verification is to ensure the logicality and rationality of the graph structure, including checking whether the hierarchical relationship between knowledge points conforms to the actual teaching logic, and whether the weight of the edge accurately reflects the association strength between knowledge points. Through consistency verification, unreasonable connections and relationships are excluded, and the graph structure is optimized to finally obtain the initial knowledge graph.

[0048] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0049] (1) Collect knowledge state data of the learner to obtain initial knowledge state data, where the initial knowledge state data includes: online test data, homework grading data, and classroom question data;

[0050] (2) Clean the initial knowledge state data to obtain the cleaned knowledge state data, and normalize the cleaned knowledge state data to obtain the normalized knowledge state data;

[0051] (3) Extract the principal component features from the knowledge state data to obtain the knowledge state feature data;

[0052] (4) Map the knowledge state feature data through the BKT model to obtain the knowledge map of the learner.

[0053] Specifically, collect the knowledge state data of the learner to obtain the initial knowledge state data. The initial knowledge state data includes online test data, homework grading data, and classroom question data. The online test data includes the scores and completion times of the learner in the online test, which reflect the learner's understanding and mastery of certain knowledge points. The homework grading data reflects the learner's performance in applying knowledge in practice, including the scores and completion status of each homework. The classroom question data records the correct rate and answering time of the learner's answers in the classroom, which can reflect the learner's immediate response ability and knowledge point mastery. Clean the initial knowledge state data to remove invalid data and handle outliers to improve the data quality and reliability. Delete the duplicate items and missing values in the online test data, homework grading data, and classroom question data. Handle the outliers, such as marking and correcting the records with scores significantly lower than the average level or higher than the reasonable range. Through data cleaning, obtain the cleaned knowledge state data. Normalize the cleaned knowledge state data to eliminate the differences in different data sources and dimensions, making the data comparable. Extract the principal component features from the knowledge state data. Principal component feature extraction (PCA) is a dimensionality reduction technique that extracts the main features in the data by transforming high-dimensional data into a low-dimensional space. The implementation steps of PCA include: calculating the covariance matrix of the data, performing eigenvalue decomposition, and selecting the eigenvectors corresponding to the top k largest eigenvalues as the principal components.

[0054] The specific formula is:

[0055]

[0056] where C is the covariance matrix, X i is the i-th sample, and μ is the sample mean. Obtain the eigenvector matrix through eigenvalue decomposition, and select the top k eigenvectors to form the principal component matrix P. Map the knowledge state feature data through the BKT model to obtain the knowledge map of the learner.

[0057] The BKT (Bayesian Knowledge Tracing) model is a knowledge tracing model based on Bayesian networks, used to estimate learners' mastery of knowledge points. The basic assumption of the BKT model is that the knowledge state of learners is a hidden binary variable (mastered or not mastered), and the estimation of the knowledge state is updated through observed learning behaviors (such as test scores and homework scores). The core formula of the BKT model is:

[0058]

[0059] where P(L t |O t ) is the probability that the learner masters knowledge point L t after observing O t at time t, P(O t |L t ) is the probability of observing O t when the learner masters L t , P(L t ) is the prior probability that the learner masters L t at time t, and P(O t ) is the marginal probability of observing O t .

[0060] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0061] (1) Extract nodes from the knowledge map to obtain knowledge nodes, and extract nodes from the knowledge graph to obtain graph nodes;

[0062] (2) Calculate the similarity between the knowledge nodes and the graph nodes to obtain an initial similarity matrix, and perform threshold screening on the initial similarity matrix to obtain a target similarity matrix;

[0063] (3) Perform feature matching analysis on the target similarity matrix to obtain a feature matching result, and generate paths based on the feature matching result to obtain multiple initial learning paths;

[0064] (4) Calculate the weights of the multiple initial learning paths to obtain multiple weighted learning paths, and use the dynamic programming algorithm to screen the multiple weighted learning paths to obtain multiple candidate learning paths.

[0065] Specifically, node extraction is performed on the knowledge map to obtain knowledge nodes. A knowledge map is a graph structure that reflects the current knowledge state of a learner, where each node represents a specific knowledge point. By analyzing the knowledge map, all knowledge points are extracted as knowledge nodes. These knowledge nodes include both the knowledge points that the learner has mastered and those that have not been mastered, reflecting the overall knowledge structure of the learner. Similarly, node extraction is performed on the knowledge graph to obtain graph nodes. A knowledge graph is a more comprehensive knowledge structure graph that contains all the knowledge points to be learned and their interrelationships. By analyzing the knowledge graph, all knowledge points are extracted as graph nodes, and these nodes represent all the key knowledge points in the learning field. The similarity between the knowledge nodes and the graph nodes is calculated to obtain an initial similarity matrix. Similarity calculation is used to evaluate the degree of similarity between the learner's knowledge nodes and the graph nodes. Multiple methods can be used for similarity calculation, such as content-based similarity, context-based similarity, etc. For example, cosine similarity is calculated by comparing the cosine values of two node vectors to measure their similarity. The obtained initial similarity matrix is a two-dimensional matrix, where each element represents the similarity between a knowledge node and a graph node. Threshold screening is performed on the initial similarity matrix to obtain a target similarity matrix. Threshold screening is used to remove those node pairs with low similarity and only retain the node pairs with similarity higher than a certain threshold. By setting a reasonable similarity threshold, irrelevant or unimportant node pairs can be filtered out to ensure the accuracy and reliability of the matching results. Feature matching analysis is performed on the target similarity matrix to obtain a feature matching result. To find the best matching relationship between the learner's knowledge nodes and the graph nodes, multiple methods can be used for feature matching, such as the Hungarian algorithm, the maximum matching algorithm, etc. Through feature matching analysis, it is determined which graph node each knowledge node best matches, thereby obtaining a feature matching result. The feature matching result shows the specific matching relationship between the learner's current knowledge state and the ideal knowledge structure. Based on the feature matching result, multiple initial learning paths are generated. Specific learning paths are designed according to the feature matching result, and these paths represent multiple possible routes from the current knowledge state to the target knowledge structure. Algorithms such as depth-first search and breadth-first search can be used to generate multiple initial learning paths by analyzing the connectivity between knowledge nodes. These initial learning paths contain all possible learning steps from the current knowledge point to the target knowledge point, providing multiple learning options for selection. Weight calculation is performed on the multiple initial learning paths to obtain multiple weighted learning paths. To evaluate the quality of each learning path, weight calculation can be based on multiple factors, such as learning difficulty, importance of knowledge points, learning time, etc. By weighting the knowledge points in each learning path and calculating the total weight of the path, multiple weighted learning paths are obtained. The dynamic programming algorithm is used to screen the multiple weighted learning paths to obtain multiple candidate learning paths.Dynamic programming is an efficient optimization algorithm that decomposes a problem into sub-problems and gradually solves for the optimal solution. During the path screening process, the dynamic programming algorithm is used to comprehensively evaluate multiple weight learning paths and screen out the optimal combination of learning paths. The obtained candidate learning paths are the best learning plans that have been optimized and screened, and can effectively guide learners to gradually master the target knowledge structure from the current knowledge state.

[0066] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0067] (1) Extract features from multiple candidate learning paths respectively to obtain multiple path features of each candidate learning path;

[0068] (2) Perform feature normalization and feature vector encoding on the multiple path features of each candidate learning path to obtain the path feature encoding vector of each candidate learning path;

[0069] (3) Input the path feature encoding vectors of each candidate learning path into a pre-set multi-feature fusion path analysis model respectively. The multi-feature fusion path analysis model includes: a multi-head self-attention mechanism, multiple deep neural networks, and a fully connected layer;

[0070] (4) Perform multi-head self-attention mechanism analysis on the path feature encoding vectors to generate multiple path feature attention vectors;

[0071] (5) Input the multiple path feature attention vectors into multiple deep neural networks for feature relationship and feature weight analysis to obtain the feature relationship and feature weight of each path feature attention vector;

[0072] (6) In the fully connected layer, calculate the path comprehensive score of each candidate learning path and perform weighted fusion according to the feature relationship and feature weight, and output the path comprehensive score of each candidate learning path.

[0073] Specifically, feature extraction is performed on multiple candidate learning paths to obtain multiple path features for each candidate learning path. Each learning path can have multiple features, such as the number of knowledge points in the learning path, the difficulty of each knowledge point, the total learning time of the path, the proportion of key knowledge points in the path, etc. These features are extracted by detailed analysis of each candidate learning path. For example, assume there are three candidate learning paths A, B, and C. Path A includes 5 knowledge points, with a total learning time of 10 hours and a proportion of key knowledge points of 0.6; Path B includes 4 knowledge points, with a total learning time of 8 hours and a proportion of key knowledge points of 0.5; Path C includes 6 knowledge points, with a total learning time of 12 hours and a proportion of key knowledge points of 0.7. The features of each path can be represented as a vector. For example, the feature vector of path A is [5, 10, 0.6], the feature vector of path B is [4, 8, 0.5], and the feature vector of path C is [6, 12, 0.7]. Feature normalization and feature vector encoding are performed on the multiple path features of each candidate learning path to obtain the path feature encoding vector of each candidate learning path. The normalization process is to eliminate the dimensional differences between different features and make the features comparable. A common method is min-max normalization, which converts the feature values to the range of 0 to 1. For example, assume that the number of knowledge points for all paths is between 1 and 10, the total learning time is between 5 and 20 hours, and the proportion of key knowledge points is between 0 and 1. Through the normalization process, the feature vector of path A can be converted to [0.5, 0.33, 0.6], the feature vector of path B is [0.4, 0.2, 0.5], and the feature vector of path C is [0.6, 0.47, 0.7]. The normalized feature vectors are further subjected to feature vector encoding to generate encoding vectors suitable for input into the model. The path feature encoding vectors of each candidate learning path are input into a pre-set multi-feature fusion path analysis model. The multi-feature fusion path analysis model includes a multi-head self-attention mechanism, multiple deep neural networks, and a fully connected layer. The path feature encoding vectors are analyzed through the multi-head self-attention mechanism. The multi-head self-attention mechanism can capture the relationships and interactions between different features, and generate multiple path feature attention vectors by weighted calculation of the input vectors through multiple attention heads. For example, after the feature vector of path A is analyzed by the multi-head self-attention mechanism, three attention vectors are generated, namely [0.45, 0.35, 0.55], [0.5, 0.4, 0.6], and [0.55, 0.45, 0.65]. The multiple path feature attention vectors are input into multiple deep neural networks for feature relationship and feature weight analysis. The deep neural network deeply explores the complex relationships and importance between features through multiple layers of non-linear transformations. Each path feature attention vector is processed in the deep neural network to obtain the feature relationship and feature weight of each vector.For example, after being processed by a deep neural network, the feature relationship of the first attention vector of path A is [0.48, 0.38, 0.58], and the feature weights are [0.6, 0.4, 0.5]; the feature relationship of the second attention vector is [0.52, 0.42, 0.62], and the feature weights are [0.7, 0.5, 0.6]; the feature relationship of the third attention vector is [0.57, 0.47, 0.67], and the feature weights are [0.8, 0.6, 0.7]. In the fully connected layer, based on the feature relationship and feature weights, learning path score calculation and weighted fusion are performed on multiple path feature attention vectors to output the path comprehensive score of each candidate learning path. The fully connected layer comprehensively evaluates the advantages and disadvantages of each learning path through weighted calculation of each feature and generates the final score. For example, the comprehensive score of path A can be obtained by performing weighted fusion calculation on the feature relationships and feature weights of all attention vectors. Suppose the final score is 85 points; the comprehensive score of path B is 78 points; the comprehensive score of path C is 90 points. These scores reflect the comprehensive effects of different learning paths and provide a scientific basis for the selection of learning paths.

[0074] In a specific embodiment, the process of executing step S105 may specifically include the following steps:

[0075] (1) Select multiple target learning paths according to the path comprehensive score and a preset target value;

[0076] (2) Perform learning content conversion on the multiple target learning paths to obtain multiple first learning contents;

[0077] (3) Organize the multiple first learning contents to obtain multiple second learning contents;

[0078] (4) Obtain the learning styles and interests of the learners, and perform content integration on the multiple second learning contents to generate the recommended learning contents for the learners.

[0079] Specifically, multiple target learning paths are selected based on the comprehensive path score and the preset target value. The comprehensive path score reflects the effectiveness and adaptability of different learning paths, while the preset target value is the set minimum acceptance standard. By screening out the learning paths with scores higher than the preset target value, it is ensured that the selected learning paths have high learning effectiveness and matching degree. For example, suppose there are three learning paths A, B, and C. The score of path A is 85 points, the score of path B is 78 points, the score of path C is 90 points, and the preset target value is 80 points. Then, paths A and C are selected as the target learning paths. The learning content of multiple target learning paths is transformed to obtain multiple first learning contents. The knowledge points and learning tasks on each learning path are transformed into specific learning contents. These learning contents can include video explanations, courseware materials, practice questions, reading materials, etc. In this way, the abstract learning paths are made concrete, enabling learners to directly engage in learning. For example, path A contains five knowledge points: linear equations, quadratic equations, geometric basics, introduction to calculus, and introduction to probability theory. After transformation, the learning content of path A can include video explanations of linear equations, practice questions of quadratic equations, courseware materials of geometric basics, reading materials of introduction to calculus, and class notes of introduction to probability theory. The multiple first learning contents are organized to obtain multiple second learning contents. The transformed learning contents are structured to have a reasonable learning sequence and logic. By organizing the learning contents, it is ensured that learners learn in an order from easy to difficult and step by step, avoiding chaos and repetition of the content and improving learning efficiency. For example, the learning content organization of path A can first arrange the video explanations of linear equations, then the practice questions of quadratic equations, followed by the courseware materials of geometric basics, then the reading materials of introduction to calculus, and finally the class notes of introduction to probability theory. Through such organization, learners can learn in a reasonable order and gradually master the knowledge points. After the content organization is completed, the learning styles and interests of the learners are obtained. Ensure that the recommended learning content is more personalized and meets the needs of the learners. The learning styles of learners can be visual, auditory, kinesthetic, etc. Different learning styles have different preferences for the presentation forms of learning contents; while the interests of learners can guide the selection and arrangement of the content, making the learning process more interesting and attractive. For example, a visual learner may prefer to watch videos and charts, while a kinesthetic learner may prefer to learn through hands-on practice and interactive exercises. The multiple second learning contents are integrated to generate the recommended learning content for the learners. The organized learning contents are matched and adjusted with the learning styles and interests of the learners. By integrating different types of learning resources, the recommended learning content meets both the requirements of the learning path and the personalized needs of the learners.For example, assume that the learner is a visual learner who is particularly interested in geometry. During the content integration process, diagrams and video explanations of geometric fundamentals can be added while reducing the amount of text-based materials to ensure that the recommended learning content better aligns with the learner's preferences and needs.

[0080] In a specific embodiment, the process of executing step S106 may specifically include the following steps:

[0081] (1) Collect the learner's status feedback data based on the recommended learning content;

[0082] (2) Define the learner's knowledge mastery vector and learning task feature vector according to the status feedback data;

[0083] (3) Input the knowledge mastery vector and the learning task feature vector into a probabilistic multi-objective reinforcement learning algorithm for probabilistic calculation and multi-objective analysis to obtain multi-objective feedback information;

[0084] (4) Perform reinforcement learning on the multi-objective feedback information to obtain a reinforcement learning result;

[0085] (5) Update the initial knowledge graph according to the reinforcement learning result to obtain a target knowledge graph;

[0086] (6) Update the path analysis model of the multi-feature fusion path analysis model according to the reinforcement learning result to obtain a target path analysis model.

[0087] Specifically, based on the recommended learning content, the status feedback data of the learner is collected. The status feedback data includes test scores, homework completion, learning time, interaction participation, etc. According to the status feedback data, the knowledge mastery vector and the learning task feature vector of the learner are defined. The knowledge mastery vector reflects the learner's mastery of each knowledge point and can be quantified through data such as test scores and homework scores. The learning task feature vector describes the characteristics of the learning task, including task type, task difficulty, learning time, etc. The knowledge mastery vector and the learning task feature vector are input into the probabilistic multi-objective reinforcement learning algorithm for probability calculation and multi-objective analysis. The probabilistic multi-objective reinforcement learning algorithm is an optimization method that can consider multiple objectives simultaneously. Through probability calculation and multi-objective analysis, it evaluates the learner's performance on different learning paths and optimization suggestions. Reinforcement learning is performed on the multi-objective feedback information to obtain the reinforcement learning result. Reinforcement learning is a machine learning method that continuously optimizes strategies through trial and error and feedback mechanisms. Through reinforcement learning of the multi-objective feedback information, the learning path and the learning content recommendation strategy can be gradually optimized to improve the learning effect. For example, the algorithm can adjust the learning path and content recommendation strategy according to the learner's learning performance and multi-objective feedback information, optimize the learning plan, and finally obtain the reinforcement learning result, including the updated knowledge point mastery strategy and learning task arrangement. According to the reinforcement learning result, the initial knowledge graph is updated to obtain the target knowledge graph. The learner's latest learning performance and knowledge mastery are reflected in the knowledge graph. By adding new knowledge points, updating the relationships between knowledge points, and adjusting the weights of knowledge points, the knowledge graph can more accurately and dynamically reflect the learner's knowledge state. For example, after the user has learned the introduction to probability theory and shows a high mastery of this knowledge point, advanced probability theory content can be added to the knowledge graph, and at the same time, the weights and relationships of related knowledge points can be adjusted to obtain a target knowledge graph that is more suitable for the user. According to the reinforcement learning result, the path analysis model of the multi-feature fusion path analysis model is updated to obtain the target path analysis model. The update of the path analysis model is to adjust the feature weights and path scoring mechanism in the path analysis model according to the reinforcement learning result, so that the model can better reflect the learner's actual learning situation and optimization suggestions. For example, the algorithm can adjust the weights of different features in the path analysis model according to the learner's learning performance and reinforcement learning result, such as increasing the weight of high-participation tasks and reducing the weight of low-difficulty tasks, and optimizing the path scoring mechanism, and finally obtain the target path analysis model.

[0088] The above describes the learning content recommendation method based on knowledge tracing in the embodiments of the present application. Next, the learning content recommendation system based on knowledge tracing in the embodiments of the present application will be described. Please refer to Figure 2 , an embodiment of the learning content recommendation system based on knowledge tracing in the embodiments of the present application includes:

[0089] The construction module 201 is used to obtain all the knowledge points that the learner needs to learn, classify the knowledge points and construct a relationship graph to obtain an initial knowledge graph;

[0090] The acquisition module 202 is used to collect and preprocess the knowledge state data of the learner to obtain the knowledge map of the learner;

[0091] The matching module 203 is used to perform feature matching between the knowledge map and the knowledge graph to obtain multiple candidate learning paths;

[0092] The calculation module 204 is used to input multiple candidate learning paths into a preset multi-feature fusion path analysis model respectively to calculate the learning path score, and obtain the comprehensive path score of each candidate learning path;

[0093] The selection module 205 is used to select multiple target learning paths according to the comprehensive path score and a preset target value, and generate recommended learning content for the learner according to the multiple target learning paths;

[0094] The optimization module 206 is used to collect the state feedback data of the learner, and update and optimize the initial knowledge graph and the multi-feature fusion path analysis model through a probabilistic multi-objective reinforcement learning algorithm to obtain a target knowledge graph and a target path analysis model.

[0095] Through the collaborative cooperation of the above-mentioned various components, by collecting and preprocessing the knowledge state data of the learner, using online test data, homework grading data and classroom questioning data, and combining technologies such as data cleaning, normalization and principal component feature extraction, an accurate knowledge map is constructed. The K-means clustering algorithm and the Bayesian network are adopted to classify the knowledge points and analyze the causal relationship, and an initial knowledge graph is constructed. It can dynamically generate and adjust the learning path to ensure that the recommended learning path highly matches the actual needs of the learner, improving the learning efficiency and effect. Through the probabilistic multi-objective reinforcement learning algorithm, the present invention can continuously update and optimize the initial knowledge graph and the path analysis model. According to the state feedback data of the learner, dynamically adjust the learning content and path, and provide personalized learning resources and recommended content. The learner can make efficient progress on a learning path that better meets his own needs and interests, improving the satisfaction and effect of learning. By continuously collecting the state feedback data of the learner and using the feedback information for reinforcement learning and model update, it is ensured that the knowledge graph and the path analysis model are always consistent with the actual situation of the learner. By regularly evaluating and generating learning effect reports, continuously improving and optimizing the recommendation system, improving the accuracy and personalization level of the learning effect, and ensuring that the learner can continuously obtain the best learning experience.

[0096] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the above-described systems, systems, and units can refer to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0097] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs that can store program codes.

[0098] The above is the case. The above embodiments are only used to illustrate the technical solutions of the present application and are not intended to limit them. Although the present application 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 for some of the technical features. These modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the various embodiments of the present application.

Claims

1. A learning content recommendation method based on knowledge tracking, characterized in that: The learning content recommendation method based on knowledge tracking includes: Acquire all the knowledge points that the learner needs to learn, and classify and construct a relationship graph for the knowledge points to obtain an initial knowledge graph; specifically include: acquire all the knowledge points that the learner needs to learn, and classify the knowledge points through the K-means clustering algorithm to obtain the classified knowledge points; perform causal relationship analysis on the classified knowledge points through the Bayesian network to obtain a relationship information set; perform node connection on the relationship information set and the classified knowledge points based on the Neo4j graph database to obtain a node connection graph; assign weights to the node connection graph to obtain a weighted relationship graph, and graph the weighted relationship graph to obtain a graphical knowledge graph; hierarchically divide the graphical knowledge graph to obtain a hierarchical knowledge graph, and perform consistency check on the hierarchical knowledge graph to obtain an initial knowledge graph; The learner's knowledge state data is collected and preprocessed to obtain the learner's knowledge map; specifically including: collecting the learner's knowledge state data to obtain initial knowledge state data, the initial knowledge state data includes: online test data, homework score data and classroom question data; cleaning the initial knowledge state data to obtain cleaned knowledge state data, and normalizing the cleaned knowledge state data to obtain normalized knowledge state data; extracting principal component features from the knowledge state data to obtain knowledge state feature data; mapping the knowledge state feature data to a knowledge map through a BKT model to obtain the learner's knowledge map; Performing feature matching on the knowledge map and the knowledge graph to obtain multiple candidate learning paths; specifically comprising: performing node extraction on the knowledge map to obtain knowledge nodes, and performing node extraction on the knowledge graph to obtain graph nodes; performing similarity calculation on the knowledge nodes and the graph nodes to obtain an initial similarity matrix, and performing threshold screening on the initial similarity matrix to obtain a target similarity matrix; performing feature matching analysis on the target similarity matrix to obtain feature matching results, and performing path generation on the feature matching results to obtain multiple initial learning paths; performing weight calculation on the multiple initial learning paths to obtain multiple weighted learning paths, and using a dynamic programming algorithm to screen the multiple weighted learning paths to obtain multiple candidate learning paths; The multiple candidate learning paths are respectively input into a preset multi-feature fusion path analysis model to perform learning path score calculation to obtain a comprehensive path score for each candidate learning path; specifically comprising: extracting features from the multiple candidate learning paths to obtain multiple path features for each candidate learning path; performing feature normalization and feature vector encoding on the multiple path features of each candidate learning path to obtain a path feature encoding vector for each candidate learning path; respectively inputting the path feature encoding vector of each candidate learning path into a preset multi-feature fusion path analysis model, the multi-feature fusion path analysis model comprising: a multi-head self-attention mechanism, multiple deep neural networks and a fully connected layer; performing multi-head self-attention mechanism analysis on the path feature encoding vector to generate multiple path feature attention vectors; inputting the multiple path feature attention vectors into the multiple deep neural networks to perform feature relationship and feature weight analysis to obtain the feature relationship and feature weight of each path feature attention vector; in the fully connected layer, performing learning path score calculation and weighted fusion on the multiple path feature attention vectors according to the feature relationship and the feature weight, and outputting a comprehensive path score for each candidate learning path; Selecting multiple target learning paths according to the comprehensive scores of the paths and the preset target values, and generating recommended learning content for the learner according to the multiple target learning paths; specifically comprising: selecting multiple target learning paths according to the comprehensive scores of the paths and the preset target values; performing learning content conversion on the multiple target learning paths to obtain multiple first learning content; performing content organization on the multiple first learning content to obtain multiple second learning content; obtaining the learning style and interests of the learner, and performing content integration on the multiple second learning content to generate recommended learning content for the learner; The state feedback data of the learner is collected, and the initial knowledge graph and the multi-feature fusion path analysis model are updated and optimized through a probabilistic multi-objective reinforcement learning algorithm to obtain a target knowledge graph and a target path analysis model; specifically comprising: collecting the state feedback data of the learner based on the recommended learning content; defining the learner's knowledge mastery vector and the learning task feature vector according to the state feedback data; inputting the knowledge mastery vector and the learning task feature vector into a probabilistic multi-objective reinforcement learning algorithm for probability calculation and multi-objective analysis to obtain multi-objective feedback information; performing reinforcement learning on the multi-objective feedback information to obtain a reinforcement learning result; performing knowledge graph update on the initial knowledge graph according to the reinforcement learning result to obtain a target knowledge graph; performing path analysis model update on the multi-feature fusion path analysis model according to the reinforcement learning result to obtain a target path analysis model.

2. A learning content recommendation system based on knowledge tracking, characterized in that: The learning content recommendation system based on knowledge tracking includes: A construction module is used to obtain all the knowledge points that learners need to learn, and classify and construct a relationship graph for the knowledge points to obtain an initial knowledge graph; specifically, it includes: obtaining all the knowledge points that learners need to learn, and classifying the knowledge points through the K-means clustering algorithm to obtain the classified knowledge points; performing causal relationship analysis on the classified knowledge points through the Bayesian network to obtain a relationship information set; performing node connection on the relationship information set and the classified knowledge points based on the Neo4j graph database to obtain a node connection graph; assigning weights to the node connection graph to obtain a weighted relationship graph, and graphically converting the weighted relationship graph to obtain a graphical knowledge graph; hierarchically dividing the graphical knowledge graph to obtain a hierarchical knowledge graph, and performing consistency verification on the hierarchical knowledge graph to obtain an initial knowledge graph; The acquisition module is used to collect and preprocess the knowledge state data of the learner to obtain the knowledge map of the learner; specifically, it includes: collecting the knowledge state data of the learner to obtain initial knowledge state data, and the initial knowledge state data includes: online test data, homework score data and classroom question data; cleaning the initial knowledge state data to obtain cleaned knowledge state data, and normalizing the cleaned knowledge state data to obtain normalized knowledge state data; extracting principal component features from the knowledge state data to obtain knowledge state feature data; mapping the knowledge state feature data to a knowledge map through a BKT model to obtain the knowledge map of the learner; A matching module is used to perform feature matching on the knowledge map and the knowledge graph to obtain multiple candidate learning paths; specifically comprising: performing node extraction on the knowledge map to obtain knowledge nodes, and performing node extraction on the knowledge graph to obtain graph nodes; performing similarity calculation on the knowledge nodes and the graph nodes to obtain an initial similarity matrix, and performing threshold screening on the initial similarity matrix to obtain a target similarity matrix; performing feature matching analysis on the target similarity matrix to obtain feature matching results, and performing path generation on the feature matching results to obtain multiple initial learning paths; performing weight calculation on the multiple initial learning paths to obtain multiple weighted learning paths, and using a dynamic programming algorithm to screen the multiple weighted learning paths to obtain multiple candidate learning paths; A calculation module is used to input the multiple candidate learning paths into a preset multi-feature fusion path analysis model to perform learning path score calculation, and obtain a comprehensive path score for each candidate learning path; specifically comprising: extracting features from the multiple candidate learning paths to obtain multiple path features of each candidate learning path; normalizing and encoding the multiple path features of each candidate learning path to obtain a path feature encoding vector of each candidate learning path; inputting the path feature encoding vector of each candidate learning path into a preset multi-feature fusion path analysis model, the multi-feature fusion path analysis model comprising: a multi-head self-attention mechanism, multiple deep neural networks and a fully connected layer; performing multi-head self-attention mechanism analysis on the path feature encoding vector to generate multiple path feature attention vectors; inputting the multiple path feature attention vectors into the multiple deep neural networks to perform feature relationship and feature weight analysis to obtain a feature relationship and feature weight of each path feature attention vector; in the fully connected layer, performing learning path score calculation and weighted fusion on the multiple path feature attention vectors according to the feature relationship and the feature weight, and outputting a comprehensive path score for each candidate learning path; A selection module is used to select multiple target learning paths according to the comprehensive scores of the paths and the preset target values, and generate recommended learning content for the learner according to the multiple target learning paths; specifically comprising: selecting multiple target learning paths according to the comprehensive scores of the paths and the preset target values; performing learning content conversion on the multiple target learning paths to obtain multiple first learning contents; performing content organization on the multiple first learning contents to obtain multiple second learning contents; obtaining the learning style and interests of the learner, and performing content integration on the multiple second learning contents to generate recommended learning content for the learner; The optimization module is used to collect the state feedback data of the learner, and update and optimize the initial knowledge graph and the multi-feature fusion path analysis model through a probabilistic multi-objective reinforcement learning algorithm to obtain a target knowledge graph and a target path analysis model; specifically comprising: collecting the state feedback data of the learner based on the recommended learning content; defining the learner's knowledge mastery vector and learning task feature vector according to the state feedback data; inputting the knowledge mastery vector and the learning task feature vector into a probabilistic multi-objective reinforcement learning algorithm for probability calculation and multi-objective analysis to obtain multi-objective feedback information; performing reinforcement learning on the multi-objective feedback information to obtain a reinforcement learning result; performing knowledge graph update on the initial knowledge graph according to the reinforcement learning result to obtain a target knowledge graph; performing path analysis model update on the multi-feature fusion path analysis model according to the reinforcement learning result to obtain a target path analysis model.

Images