Educational resource recommendation method and system based on user personalized customization

By constructing an educational knowledge graph and user status representation, the limitations of traditional educational resource recommendation methods are solved, and personalized, transparent and efficient educational resource recommendations are achieved to adapt to the learning needs of different users.

CN118296244BActive Publication Date: 2024-11-01GUANGHE XINZHI (BEIJING) TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410729458.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-06-06
Publication Date
2024-11-01
Estimated Expiration
2044-06-06

AI Technical Summary

Technical Problem

The traditional educational resource recommendation method has large workload, subjective deviation, singularity, lack of personalization and diversity, and is difficult to promote to new fields or user groups, and has high maintenance costs and lacks knowledge levels and correlation mining mechanisms.

Method used

By constructing an educational knowledge graph, extracting the matching of the keywords of educational resources and the knowledge graph, calculating the co-occurrence frequency and semantic similarity, performing matrix decomposition to obtain implicit representations of resources and concepts, combining user historical behavior data and knowledge tracking model, generate User status representation, build an educational resource recommendation model to achieve personalized recommendation.

Benefits of technology

It realizes the structured organization of educational resources and knowledge points, provides transparent and interpretable personalized recommendations, reduces computing and storage overhead, has cross-domain migration capabilities, dynamically captures user learning status, optimizes recommendation strategies, and adapts to different users. Learning needs.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118296244B_ABST
    Figure CN118296244B_ABST
Patent Text Reader

Abstract

The present invention provides a method and system for recommending educational resources based on user personalized customization, which relates to the field of educational service technology, including summarizing all educational resources and extracting resource keywords to match with pre-built educational knowledge graphs, constructing a resource concept bipartite graph, obtaining implicit representations of resources and concepts, mapping resources and concepts into low-dimensional implicit representations using matrix decomposition technology, and calculating the inner product to obtain the association strength. Then, by fusing user historical behavior data, learning progress, and knowledge mastery, a knowledge mastery feature vector is generated, the learning progress feature vector and the knowledge mastery feature vector are spliced, and user state representation is generated by feature fusion in combination with user basic attributes. Based on the user state and resource concept bipartite graph, a reinforcement learning recommendation model is constructed, and the cumulative reward is optimized through a quality evaluation function to output personalized best recommended resources.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of educational service technology, and in particular to a method and system for recommending educational resources based on user personalized customization. Background Art

[0002] With the rapid development of artificial intelligence such as deep learning, future educational resource recommendations will develop in the direction of intelligence, personalization, structure and explainability. Technologies such as semantic recommendation based on knowledge graphs and strategy optimization based on reinforcement learning will promote the efficient use of educational resources and advance the digitalization and intelligence of the education field.

[0003] Traditional educational resource recommendation methods have many limitations. First, the workload of manually labeling resource topics and knowledge points is huge, and there are subjective biases. Second, recommendations based on rules or collaborative filtering are often monotonous and lack personalization and diversity. In addition, models trained based on historical data are difficult to generalize to new fields or user groups and have poor generalization capabilities. Educational resources and knowledge points usually exist in knowledge islands, lacking structured and associated mining mechanisms, and, over time, the cost of maintaining knowledge bases with traditional methods is very high.

[0004] Therefore, a method is needed to organically organize educational resources and knowledge points, and fully consider the knowledge hierarchy and relevance when making recommendations. It should be able to integrate mechanisms such as knowledge tracking and reinforcement learning to make the recommendation process more transparent and explainable while meeting the personalized needs of users at different levels. Summary of the invention

[0005] The embodiments of the present invention provide a method and system for recommending educational resources based on user personalization, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] Provided is a method for recommending educational resources based on user personalization, including:

[0008] Summarize all educational resources, extract keywords for each educational resource, and match them with the pre-built educational knowledge graph, build a resource concept mapping set, organize it into a bipartite graph structure, determine the edge weights by calculating the co-occurrence frequency, build a co-occurrence matrix, use a non-negative matrix, combine the structural constraints of the educational knowledge graph, decompose the co-occurrence matrix to obtain implicit representation of resources and implicit representation of concepts, calculate the inner product between the implicit representation of resources and the implicit representation of concepts, determine the resource concept association strength, and obtain a bipartite graph of resource concepts;

[0009] Collect user historical behavior data, construct a behavior sequence, match the behavior sequence with the educational knowledge graph, identify the weight of each knowledge point, calculate the completion degree of each knowledge point, and generate a learning progress feature vector; use the knowledge tracking model to calculate the user's mastery probability of each knowledge point and generate a knowledge mastery feature vector; splice the learning progress feature vector and the knowledge mastery feature vector, combine the user's basic attributes, and generate a user state representation through feature fusion;

[0010] Based on the user state representation, the state space is determined, and a subset of candidate resources is selected from the resource concept bipartite graph to determine the resource space; based on the state space and the resource space, an educational resource recommendation model is constructed, and the educational resource recommendation model calculates the cumulative score through a quality evaluation function constructed based on feedback data, and determines the best recommended resource by selecting the maximum value of the cumulative score.

[0011] In an optional implementation, keywords are extracted from each educational resource and matched with a pre-built educational knowledge graph, and constructing a resource concept mapping set includes:

[0012] Extract keywords from each of the educational resources, find literal matching concepts corresponding to the keywords in the educational knowledge graph, and establish a primary mapping set;

[0013] Based on the literal matching concept, the neighbor concept of the literal matching concept is found in the educational knowledge graph, and the semantic similarity value between the educational resource and the neighbor concept is calculated. When the semantic similarity value is higher than a preset adoption threshold, the mapping relationship between the educational resource and the neighbor concept is added to the primary mapping set to construct a resource concept mapping set, wherein the neighbor concept includes a superordinate concept and a subordinate concept.

[0014] In an optional implementation, the resource concept mapping set is organized into a bipartite graph structure, the edge weight is determined by calculating the co-occurrence frequency, a co-occurrence matrix is ​​constructed, and the co-occurrence matrix is ​​decomposed by using a non-negative matrix in combination with the structural constraints of the educational knowledge graph to obtain implicit resource representation and implicit concept representation, and the resource concept association strength is determined by calculating the inner product between the implicit resource representation and the implicit concept representation, and the resource concept bipartite graph includes:

[0015] The resource concept mapping set is split and combined into a bipartite graph structure according to the classification of resource nodes and concept nodes, the edge weights of resources and concepts are calculated by a word frequency weighting method, and all edge weights are organized into a co-occurrence matrix;

[0016] Decomposing the co-occurrence matrix into two non-negative matrices, adding regularization terms based on the structural constraints of the educational knowledge graph, constructing a decomposition error function, minimizing the decomposition error, using the alternating least squares method to obtain the optimal solution of the two non-negative matrices, and determining the implicit representation of resources and the implicit representation of concepts;

[0017] Based on each pair of the implicit representation of the resource and the implicit representation of the concept, the inner product is calculated to determine the resource concept association strength, and the resources are arranged in descending order. According to the preset preferred index, the optimal association result is determined to obtain a bipartite graph of resource concepts.

[0018] In an optional implementation, the edge weights of resources and concepts are calculated using a word frequency weighting method, and the formula is as follows:

[0019] ;

[0020] in, r Represents resources, c Represents the concept, ω rc represents the edge weights of resources and concepts, N rc Representation Resources r With concept c The number of co-occurrences, N r Representation Resources r Number of occurrences, R Represents a resource collection, | R | indicates the total number of resources, R c Representation and Concept c The associated resource collection, | R c |Representation and Concept c The number of associated resources;

[0021] Combined with the regularization term, the decomposition error is minimized, and the formula is as follows:

[0022] ;

[0023] Among them, U represents the first non-negative matrix after decomposition, V represents the second non-negative matrix after decomposition, W represents the co-occurrence matrix, and F represents the Frobenius norm. λ 1 represents the first regularization coefficient, u r Represents the corresponding resources in matrix U r The characteristic vector of M r Representation and Resources r The set of all concepts associated with j express Mr The first in the collection j The feature vector of a resource, λ 2 represents the second regularization coefficient, C represents a concept set, v c Represents the corresponding concept in the matrix V c The characteristic vector of M c Representation and Concept c The collection of all associated resources, v i express M c The first in the collection i The feature vector of a resource.

[0024] In an optional implementation, collecting user historical behavior data, constructing a behavior sequence, matching the behavior sequence with the educational knowledge graph, identifying the weight of each knowledge point, calculating the completion degree of each knowledge point, and generating a learning progress feature vector; using a knowledge tracking model, calculating the user's mastery probability of each knowledge point, and generating a knowledge mastery feature vector includes:

[0025] Arrange the user's historical behavior data in chronological order to construct a behavior sequence, use a multi-layer long short-term memory network to model the behavior sequence, determine the temporal dependency, and determine the user's historical behavior state characteristics based on the last hidden state of the multi-layer long short-term memory network and the temporal dependency;

[0026] Based on the educational knowledge graph, the weight of the knowledge point is determined, the behavior sequence is matched with the educational knowledge graph, the knowledge point corresponding to each behavior is determined, and the completion degree of each knowledge point is determined in combination with the weight of the knowledge point. Based on the knowledge point association structure in the educational knowledge graph, the completion degree of the associated knowledge points is aggregated, and the overall completion degree corresponding to the behavior sequence is recursively calculated to generate a learning progress feature vector;

[0027] Based on each knowledge point, using the knowledge tracking model, determine the transition process of the user's mastery of the knowledge point, wherein the mastery state is determined by the initial mastery probability, the transition probability, the error rate and the guessing rate;

[0028] Based on the user's answer sequence on each knowledge point, the user's knowledge mastery rate is determined by the error rate and the guessing rate. Combined with the mastery probability at the previous moment, the mastery probability of the user mastering the knowledge point is recursively calculated. The expectation maximization algorithm is used to calculate the user's mastery probability sequence for the knowledge point, and the mastery probability corresponding to the last step of the mastery probability sequence is extracted to determine the mastery feature. The above steps are repeated to calculate the user's mastery features for all knowledge points and generate a knowledge mastery feature vector.

[0029] In an optional implementation, calculating the knowledge mastery feature vector includes:

[0030] Master probability calculation and recursive calculation, the formula is as follows:

[0031] ;

[0032] in, k a t Indicates that at time step t Whether users have mastered the knowledge points a status, k a t-1 Indicates that at time step t- 1. Whether the user has mastered the knowledge points a status, o 1 t From time step 1 to t The user answer sequence, o 1 t-1 From time step 1 to t- 1 user answer sequence, o t Indicates that at time step t The answer result of Indicates the answer sequence o 1 t In the case of user at time step t About knowledge points a The probability of mastering Indicates that if the user is at time step t Master the knowledge points a The probability of giving the correct answer, Indicates the answer sequence o 1 t-1 In the case of user at time step t About knowledge points a The probability of mastering Indicates the answer sequence o 1 t-1 In the case of user at time step t-1 About knowledge points a The probability of mastering Represents the probability that a user moves from not mastering to mastering after one practice.

[0033] In an optional implementation, based on the user state representation, a state space is determined, and a subset of candidate resources is selected from the resource concept bipartite graph to determine a resource space; based on the state space and the resource space, an educational resource recommendation model is constructed, and the educational resource recommendation model calculates a cumulative score through a quality evaluation function constructed based on feedback data, and determines the best recommended resource by selecting the maximum value of the cumulative score, including:

[0034] Based on the user state representation, analyzing the user's learning progress and knowledge mastery, and determining the state space;

[0035] Selecting a subset of candidate resources matching the user state representation from the resource concept bipartite graph to determine a resource space;

[0036] Pairing the state space with the resource space to determine a state resource response set, and determining a quality evaluation function based on user feedback to construct an educational resource recommendation model, wherein the user feedback includes user instant feedback and long-term feedback on educational effects;

[0037] At the initial moment, based on the current user state representation, a first state is determined, and a first recommended resource set corresponding to the first state is generated through the educational resource recommendation model; recommended resources are selected from the first recommended resource set according to a preset selection quantity, and based on the user feedback, the first state is updated to a second state, and the next moment is entered;

[0038] Iterate in sequence until the preset termination state is reached;

[0039] Based on the recommended resources selected at each moment, a recommended path is constructed, based on the quality evaluation function, the cumulative score of each recommended path is calculated respectively, the recommended path with the highest cumulative score is selected, and the best recommended resource is determined.

[0040] According to a second aspect of the embodiments of the present invention,

[0041] Provide an educational resource recommendation system based on user personalization, including:

[0042] The first unit is used to summarize all educational resources, extract keywords for each educational resource, and establish a match with a pre-constructed educational knowledge graph, construct a resource concept mapping set, organize it into a bipartite graph structure, determine the edge weight by calculating the co-occurrence frequency, construct a co-occurrence matrix, use a non-negative matrix, combine the structural constraints of the educational knowledge graph, decompose the co-occurrence matrix to obtain resource implicit representation and concept implicit representation, calculate the inner product between the resource implicit representation and the concept implicit representation, determine the resource concept association strength, and obtain a resource concept bipartite graph;

[0043] The second unit is used to collect user historical behavior data, construct a behavior sequence, match the behavior sequence with the educational knowledge graph, identify the weight of each knowledge point, calculate the completion degree of each knowledge point, and generate a learning progress feature vector; use the knowledge tracking model to calculate the user's mastery probability of each knowledge point and generate a knowledge mastery feature vector; splice the learning progress feature vector and the knowledge mastery feature vector, combine the user's basic attributes, and generate a user state representation through feature fusion;

[0044] The third unit is used to determine the state space based on the user state representation, and select a subset of candidate resources from the resource concept bipartite graph to determine the resource space; based on the state space and the resource space, an educational resource recommendation model is constructed, and the educational resource recommendation model calculates the cumulative score through a quality evaluation function constructed based on feedback data, and determines the best recommended resource by selecting the maximum value of the cumulative score.

[0045] According to a third aspect of the embodiments of the present invention,

[0046] An electronic device is provided, comprising:

[0047] processor;

[0048] a memory for storing processor-executable instructions;

[0049] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0050] A fourth aspect of the embodiments of the present invention is:

[0051] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0052] By constructing an educational knowledge graph, educational resources and knowledge points are organized into a structured form, which can fully explore the correlation between knowledge. Graph-based resource recommendations not only take into account the user's knowledge level and mastery, but also recommend related resources on the knowledge chain, which is conducive to systematic and progressive learning. At the same time, the interpretability and visualization characteristics of the knowledge graph make the recommendation process more transparent, and users can clearly understand the logical relationship between knowledge points, improving the initiative and consciousness of learning.

[0053] Using matrix decomposition technology to map resources and concepts to a low-dimensional implicit representation space can effectively reduce computing and storage overhead and improve recommendation efficiency. At the same time, implicit representation can also capture the deep semantic information of resources and concepts, laying the foundation for accurate matching. By calculating the inner product between the implicit representation of resources and the implicit representation of concepts, the strength of the association between them can be quantified, providing a basis for personalized recommendations. This technology is not restricted by domains and has the ability to migrate and generalize across domains, which is conducive to the widespread application of knowledge recommendation.

[0054] This embodiment can dynamically capture the user's learning status and continuously optimize the recommendation strategy by integrating reinforcement learning, knowledge tracking and other mechanisms. The quality evaluation function built based on feedback data can guide the intelligent agent to continuously try new recommendation solutions to find the best resource matching method. At the same time, by modeling the user's learning progress and knowledge mastery, the recommendation difficulty can be adaptively adjusted, which is conducive to teaching students in accordance with their aptitude and maximizing their learning potential. This technical solution fully integrates the concepts of personalized recommendation and data-driven decision-making, providing an efficient and reliable solution for lifelong learning scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 A flowchart of a method for recommending educational resources based on user personalization according to an embodiment of the present invention;

[0056] Figure 2 It is a structural diagram of an educational resource recommendation system based on user personalization according to an embodiment of the present invention. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0058] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0059] Figure 1 FIG. 1 is a flow chart of a method for recommending educational resources based on user personalization according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0060] S101. Summarize all educational resources, extract keywords for each educational resource, and establish a match with the pre-constructed educational knowledge graph, construct a resource concept mapping set, organize it into a bipartite graph structure, determine the edge weight by calculating the co-occurrence frequency, construct a co-occurrence matrix, use a non-negative matrix, combine the structural constraints of the educational knowledge graph, decompose the co-occurrence matrix, obtain the implicit representation of resources and the implicit representation of concepts, determine the resource concept association strength by calculating the inner product between the implicit representation of the resource and the implicit representation of the concept, and obtain the resource concept bipartite graph;

[0061] In an optional implementation, keywords are extracted from each educational resource and matched with a pre-built educational knowledge graph, and constructing a resource concept mapping set includes:

[0062] Extract keywords from each of the educational resources, find literal matching concepts corresponding to the keywords in the educational knowledge graph, and establish a primary mapping set;

[0063] Based on the literal matching concept, the neighbor concept of the literal matching concept is found in the educational knowledge graph, and the semantic similarity value between the educational resource and the neighbor concept is calculated. When the semantic similarity value is higher than a preset adoption threshold, the mapping relationship between the educational resource and the neighbor concept is added to the primary mapping set to construct a resource concept mapping set, wherein the neighbor concept includes a superordinate concept and a subordinate concept.

[0064] The resource concept mapping set is organized into a bipartite graph structure. The edge weight is determined by calculating the co-occurrence frequency, and a co-occurrence matrix is ​​constructed. The co-occurrence matrix is ​​decomposed by using a non-negative matrix and combining the structural constraints of the educational knowledge graph to obtain implicit resource representation and implicit concept representation. The resource concept association strength is determined by calculating the inner product between the implicit resource representation and the implicit concept representation. The resource concept bipartite graph includes:

[0065] The resource concept mapping set is split and combined into a bipartite graph structure according to the classification of resource nodes and concept nodes, the edge weights of resources and concepts are calculated by a word frequency weighting method, and all edge weights are organized into a co-occurrence matrix;

[0066] Decomposing the co-occurrence matrix into two non-negative matrices, adding regularization terms based on the structural constraints of the educational knowledge graph, constructing a decomposition error function, minimizing the decomposition error, using the alternating least squares method to obtain the optimal solution of the two non-negative matrices, and determining the implicit representation of resources and the implicit representation of concepts;

[0067] Based on each pair of the implicit representation of the resource and the implicit representation of the concept, the inner product is calculated to determine the resource concept association strength, and the resources are arranged in descending order. According to the preset preferred index, the optimal association result is determined to obtain a bipartite graph of resource concepts.

[0068] The resource concept mapping set refers to the set that associates educational resources with concepts in the knowledge graph, which specifically includes two parts:

[0069] Literal matching mapping: Find concepts with the same literal meaning in the knowledge graph based on resource keywords and establish a one-to-one correspondence.

[0070] Semantic extension mapping: Based on the neighbor concepts (superordinate concepts and subordinate concepts) of the literal matching concept, the semantic similarity is calculated, and a mapping relationship is established if it exceeds the threshold.

[0071] This mapping set includes the direct correspondence between resources and concepts, as well as the indirect associations captured through semantic extension, providing basic data for the subsequent construction of a bipartite graph.

[0072] The bipartite graph structure refers to the binary classification of resource concept mapping sets according to "resource nodes" and "concept nodes", and then the mapping relationship between the two node sets is connected as an edge to form a special network structure, namely a bipartite graph. In the bipartite graph, the weight of the edge is obtained by calculating the co-occurrence frequency of resources and concepts in the corpus. This bipartite graph structure clearly depicts the complex correspondence between resources and concepts.

[0073] Implicit representation refers to the use of non-negative matrix decomposition technology to project high-dimensional sparse bipartite graph network data into a low-dimensional dense vector space. Specifically, the resource set and concept set are mapped to the row vector and column vector of the matrix respectively, and the original network is approximately reconstructed in the form of matrix product. In this way, the implicit representation of resources and concepts in the low-dimensional vector space is obtained, which can effectively capture their semantic and structural information. This implicit representation not only reduces the computational storage overhead, but also provides high-quality feature representation for subsequent reasoning and matching operations.

[0074] Through the process of extracting resource concept relationships through natural language processing, building a weighted bipartite network, and learning implicit representations, we have achieved structured modeling of educational resources and knowledge concepts, laying the foundation for accurate matching and recommendation.

[0075] Exemplarily, when extracting keywords from educational resources and establishing a mapping relationship with the knowledge graph, it is first necessary to extract keywords for each educational resource. For each educational resource (such as courseware, video, text, etc.), use natural language processing technology (such as keyword extraction, named entity recognition, etc.) to extract core keywords from the content.

[0076] The extracted keywords are literally matched in the pre-built educational knowledge graph to find the corresponding concept entities and establish a primary one-to-one mapping set between resources and concepts:

[0077] Use graph databases or knowledge base query languages ​​(such as Cypher, SPARQL, etc.) to search pre-built educational knowledge graphs;

[0078] Construct a query statement, use the extracted keywords as query conditions, and search for concept nodes in the graph whose labels or attribute values ​​exactly match the keywords;

[0079] For the set of matching concept nodes obtained by the query, a primary one-to-one mapping relationship is established with the resource to form a mapping pair <resource, concept>;

[0080] All mapping pairs are grouped into a primary mapping set.

[0081] Based on the literal matching concepts, find their neighbor concepts (superordinate and subordinate concepts) in the graph:

[0082] Traverse each concept node c in the primary mapping set;

[0083] Use the graph database query language to find the node set neighbor(c) that is directly connected to concept c;

[0084] In neighbor(c), the super-concept set superConcepts(c) and the sub-concept set subConcepts(c) of c are identified according to the concept hierarchy relationship.

[0085] Superordinate concepts are usually broader and more abstract concepts;

[0086] Subordinate concepts are usually more detailed and specific concepts within a conceptual category;

[0087] Finally, new mapping pairs are established between the concepts in superConcepts(c) and subConcepts(c) and the original resources respectively and added to the primary mapping set.

[0088] The above steps use the hierarchical inheritance and semantic association between concepts in the knowledge graph to achieve the expansion of neighbor concepts based on literal matching concepts. The expanded mapping set captures richer resource-concept semantic association information. It should be noted that the specific implementation details of the above steps may vary depending on the specific knowledge graph model and query language. In addition, when searching for neighbor concepts, the expansion radius can be set to control the scope and degree of expansion.

[0089] Compute the semantic similarity between a resource and its neighbor concepts:

[0090] Perform text preprocessing on educational resources, such as word segmentation, stop word removal, word form restoration, etc., to obtain word sequences;

[0091] Use pre-trained word vector models (such as Word2Vec, GloVe, etc.) or sentence vector models (such as Doc2Vec, BERT, etc.) to map resource text into vector representations of fixed dimensions;

[0092] For each neighbor concept, obtain its text description in the knowledge graph (such as label, attribute value, etc.);

[0093] Map the concept text description into a vector representation through the same vector model;

[0094] The cosine similarity between the resource vector and the concept vector is calculated as their semantic similarity score.

[0095] If the similarity is higher than the threshold, the resource-concept mapping is added to the primary mapping set. Set a semantic similarity threshold, such as 0.7, and traverse the semantic similarity scores of each pair of resource-neighbor concepts. If the score is higher than the preset threshold, the resource-concept pair is added to the primary mapping set.

[0096] The primary mapping pairs obtained based on literal matching are retained in the mapping set, and the new mapping pairs obtained by semantic similarity expansion are added to the mapping set. The two mapping pairs are merged to obtain a complete resource concept mapping set containing both literal matching and semantic matching relationships.

[0097] It should be noted that when calculating semantic similarity, multiple semantic models can be integrated to improve the accuracy of similarity calculation. In addition, the setting of semantic similarity threshold needs to be adjusted according to the specific data distribution to achieve a balance between mapping coverage and accuracy.

[0098] The above method can utilize the structural information of the domain knowledge graph and integrate natural language processing technology to automatically discover the potential semantic associations between resource content and concepts, laying the foundation for building a high-quality resource concept knowledge network.

[0099] The mapping set is organized into a bipartite graph structure, and a weighted bipartite graph is first constructed.

[0100] The mapping set is separated into resource nodes and concept nodes, the complete resource concept mapping set is traversed, the resource part of the mapping pair is extracted into a non-repeating resource node set R, and the concept part of the mapping pair is extracted into a non-repeating concept node set C.

[0101] Combine resource nodes and concept nodes in a bipartite graph structure to create an empty bipartite graph G, with the resource node set R as one part of the nodes of G and the concept node set C as another part of the nodes of G.

[0102] The weight of the resource concept edge is calculated by word frequency weighting:

[0103] For each pair of resources r and Concept c , use the word frequency weighted method formula to calculate the edge weights between them ω rc .

[0104] Combining all edge weights into a co-occurrence matrix W, create | R |row| C | column matrix W, traverse each edge in the bipartite graph G ( r,c,ω rc ),Will ω rc The value of is assigned to W[ r,c ].

[0105] Implicit Representation Learning via Non-Negative Matrix Factorization:

[0106] Non-negative matrix factorization (NMF) is a commonly used collaborative filtering algorithm for recommendation systems. It predicts missing values ​​of ratings by decomposing the user-item rating matrix into implicit feature matrices of users and items. Applying NMF to the decomposition of the resource-concept co-occurrence matrix can obtain low-dimensional implicit representations of resources and concepts.

[0107] For example, given the co-occurrence matrix W , The goal of NMF is to find the first non-negative matrix and the second non-negative matrix , so that W ≈ UV T .

[0108] Among them, U and V represent the representation of resources and concepts in K-dimensional latent space respectively, and the column vector u r and v c The implicit features corresponding to resource r and concept c.

[0109] The optimization goal is to minimize the decomposition error:

[0110] ;

[0111] Among them, |·|| F Represents the Frobenius norm of a matrix.

[0112] Through this non-negative matrix decomposition method, we can obtain a compact representation of resources and concepts in a low-dimensional vector space, that is, an implicit representation vector, which retains the structure and semantic information in the original network while reducing computational and storage overhead. This low-dimensional vector representation can be applied to subsequent tasks such as correlation calculation and personalized recommendation.

[0113] Exemplarily, the co-occurrence matrix is ​​decomposed to obtain an implicit representation:

[0114] Decompose the co-occurrence matrix W into two non-negative matrices U and V. Use the non-negative matrix factorization (NMF) algorithm to approximately decompose W into U×V T The product form of , U is a matrix with |R| rows and k columns, V is a matrix with k rows and |C| columns, and k is the specified implicit space dimension.

[0115] Introduce the knowledge graph structure regularization term and construct the decomposition error objective function:

[0116] ;

[0117] When decomposing the error objective function, it consists of three parts:

[0118] The first part represents the reconstruction error of matrix decomposition, that is, the square of the Frobenius norm between the original matrix and the matrix product after decomposition. The optimization goal is to minimize this error so that the matrix product after decomposition can be as close to the original matrix as possible.

[0119] The second part is the regularization term of the matrix U, where R stands for The resource collection is equivalent to the collection of row groups in the regularization term. The second part encourages row vectors belonging to the same row group to be close to their central vectors, thereby promoting consistency within the row group. The hyperparameter λ 1 controls the strength of the regularization term in the second part.

[0120] The third part is the regularization term for the matrix V, which is similar to the second term. C Represents a concept set, which is equivalent to a set of column groups in the regularization term. The third part encourages column vectors belonging to the same column group to be close to their central vectors, promotes consistency within the column group, and uses hyperparameters λ 2 controls the strength of the regularization term in the third part.

[0121] The three parts are reconstruction error, row grouping regularization term and column grouping regularization term;

[0122] Two hyperparameters are introduced into the model to balance the influence of reconstruction error and group consistency regularization term. By adjusting the two hyperparameters, the degree of attention paid by the model to matrix decomposition and grouping structure can be flexibly controlled, so that the model can adapt to different data characteristics and application requirements.

[0123] Use the alternating least squares method to solve the minimum value of the error function. First, initialize U and V as non-negative random matrices, then alternately fix one matrix and use the least squares method to solve the other matrix:

[0124] Fix V and optimize the objective function using the least squares method for U;

[0125] Fix U and optimize the objective function using the least squares method for V;

[0126] The iteration is performed until convergence, and the optimal solutions of U and V are obtained as implicit representation vectors of resources and concepts.

[0127] Calculate the association strength and generate a bipartite graph of resource concepts:

[0128] Implicitly represent each pair of resources u r and concept implicit representation v c , calculate their inner product as the association strength;

[0129] Association Strength ;

[0130] Arrange all resource-concept pairs in descending order of association strength;

[0131] Set the optimization index N and select the top N best associations as result pairs. For example, if N = 1000, the top-1000 strongest associations of each resource / concept are selected.

[0132] Based on the result pairs, the final weighted resource concept bipartite graph G is generated:

[0133] The resource node set is the resource part of the result pair with duplicates removed;

[0134] The concept node set is the concept part of the result pair with duplicates removed;

[0135] The edge weight is the corresponding association strength value;

[0136] In this way, we obtain a low-dimensional implicit representation that describes the semantic association between resources and concepts, and quantify their correlation based on the inner product calculation to generate a weighted resource concept relationship network. This network representation not only integrates the structured information of the knowledge graph, but also retains rich implicit semantic signals, which can provide strong support for subsequent personalized recommendations, similarity retrieval and other tasks.

[0137] This implementation plan realizes intelligent modeling and representation of educational resources and knowledge concepts through advanced technical means such as natural language processing, graph model construction and representation learning. It can overcome the defects of traditional manual annotation, objectively and efficiently capture the relationship between resource content and knowledge graph concepts, and lay the foundation for intelligent resource organization.

[0138] Based on the weighted bipartite graph network structure, it not only takes into account the direct correspondence between resources and concepts, but also mines implicit associations from a semantic level, making correlation calculations more comprehensive and accurate. Through non-negative matrix decomposition, it learns the implicit vector representation of resource concepts and quantifies the strength of association, providing strong support for personalized matching and diversified recommendations.

[0139] The knowledge graph makes the system well explainable and transparent. Graph-based recommendations not only give results, but also present reasoning paths to improve user trust. The low-dimensional vectors obtained by representation learning have cross-domain migration capabilities, which helps to quickly integrate new knowledge and support the sustainable development of the system. This implementation technology makes full use of the advantages of artificial intelligence semantic modeling and representation learning, and knowledge-based and structured discrete resources, providing strong technical support for intelligent education resource management and recommendation.

[0140] S102. Collect historical behavior data of users, construct behavior sequences, match the behavior sequences with the educational knowledge graph, identify the weight of each knowledge point, calculate the completion degree of each knowledge point, and generate a learning progress feature vector; use the knowledge tracking model to calculate the probability of the user mastering each knowledge point, and generate a knowledge mastery feature vector; splice the learning progress feature vector and the knowledge mastery feature vector, combine the user's basic attributes, and generate a user state representation through feature fusion;

[0141] In an optional implementation, collecting user historical behavior data, constructing a behavior sequence, matching the behavior sequence with the educational knowledge graph, identifying the weight of each knowledge point, calculating the completion degree of each knowledge point, and generating a learning progress feature vector; using a knowledge tracking model, calculating the user's mastery probability of each knowledge point, and generating a knowledge mastery feature vector includes:

[0142] Arrange the user's historical behavior data in chronological order to construct a behavior sequence, use a multi-layer long short-term memory network to model the behavior sequence, determine the temporal dependency, and determine the user's historical behavior state characteristics based on the last hidden state of the multi-layer long short-term memory network and the temporal dependency;

[0143] Based on the educational knowledge graph, the weight of the knowledge point is determined, the behavior sequence is matched with the educational knowledge graph, the knowledge point corresponding to each behavior is determined, and the completion degree of each knowledge point is determined in combination with the weight of the knowledge point. Based on the knowledge point association structure in the educational knowledge graph, the completion degree of the associated knowledge points is aggregated, and the overall completion degree corresponding to the behavior sequence is recursively calculated to generate a learning progress feature vector;

[0144] Based on each knowledge point, using the knowledge tracking model, determine the transition process of the user's mastery of the knowledge point, wherein the mastery state is determined by the initial mastery probability, the transition probability, the error rate and the guessing rate;

[0145] Based on the user's answer sequence on each knowledge point, the user's knowledge mastery rate is determined by the error rate and the guessing rate. Combined with the mastery probability at the previous moment, the mastery probability of the user mastering the knowledge point is recursively calculated. The expectation maximization algorithm is used to calculate the user's mastery probability sequence for the knowledge point, and the mastery probability corresponding to the last step of the mastery probability sequence is extracted to determine the mastery feature. The above steps are repeated to calculate the user's mastery features for all knowledge points and generate a knowledge mastery feature vector.

[0146] For example, collect user historical behavior data and construct a behavior sequence:

[0147] In the user interaction log of the learning platform, various learning behavior records of users are collected, such as the record of watching videos: recording the video ID and viewing time; the record of answering questions: recording the question ID, answer selection, and whether it is correct; the record of consulting materials: recording the material ID and viewing time;

[0148] Construct a behavior sequence for each user. Arrange all behavior records in the chronological order of the behavior. Map each behavior record to a behavior ID to form the user's behavior sequence.

[0149] Matching behavior sequences with educational knowledge graphs:

[0150] Using the pre-built educational knowledge graph, each node in the knowledge graph represents a knowledge point, and the edge represents the precedence relationship between knowledge points.

[0151] Match each behavior ID with the corresponding knowledge point ID, such as the knowledge point corresponding to the video ID and the question ID. A behavior-knowledge point mapping table can be established in advance.

[0152] To determine the weight of each knowledge point in the graph, indicators such as node centrality and PageRank can be used to quantify the weight, and the weight value reflects the importance of the knowledge point.

[0153] Based on user behavior, the completion of each knowledge point is calculated including:

[0154] Completed: If relevant questions are answered correctly;

[0155] Partially completed: such as watching some related videos;

[0156] Unfinished: There is no related behavior for the corresponding knowledge point;

[0157] Generate a learning progress feature vector:

[0158] Based on the weights and completion of knowledge points, the weights are used as weighting coefficients, the weights of completed knowledge points are summed up, and then divided by the sum of the weights of all knowledge points to obtain the overall completion score of the behavior sequence (between 0 and 1).

[0159] When encoding the overall degree of completion into a feature vector, the score is first divided into multiple intervals at equal intervals, and then multi-hot vectors corresponding to different intervals are used as feature representations of the learning progress.

[0160] Use the knowledge tracking model to calculate the probability of mastering knowledge points: For each knowledge point, model the student's mastery state transition process, set parameters such as initial mastery probability, transition probability, guessing rate and error rate, and based on the student's answer sequence on this knowledge point, use the expectation maximization algorithm to calculate the mastery probability sequence of the knowledge point, and extract the mastery probability of the last step of the sequence as the mastery feature of the knowledge point.

[0161] To generate the knowledge mastery feature vector, it is necessary to repeatedly calculate the students' mastery features of all knowledge points, and concatenate the mastery probabilities of all knowledge points to finally generate the knowledge mastery feature vector.

[0162] Feature fusion generates user state representation:

[0163] Concatenate the learning progress feature vector and the knowledge mastery feature vector;

[0164] Combined with basic user attributes (such as age, learning stage, etc.);

[0165] Through neural networks or other fusion mechanisms, all features are fused to generate a user state vector representation;

[0166] For example, personalized learning resource recommendation is achieved through a learner state representation method based on an LSTM network:

[0167] Use LSTM network to model the learner's historical behavior sequence and extract state features at each time step;

[0168] Calculate learning progress characteristics: Based on the learner's behavior sequence, estimate the completion degree of each knowledge point;

[0169] Calculate the knowledge mastery level characteristics: Based on the learner's answers in exercises and tests, assess their mastery of each knowledge point;

[0170] Combine learning progress, knowledge mastery level and other user attribute features to generate a mixed state representation through a fully connected layer;

[0171] Historical behavior sequence modeling:

[0172] The learner's historical behavior contains rich information about interest preferences and learning patterns. In order to capture the temporal dependencies of the behavior sequence, this paper uses the LSTM network to model it. The learner's behavior records are arranged in chronological order to form a behavior sequence { x 1 , x 2 , ..., x t},in x t Indicates t The behavior vector can be the ID embedding of the corresponding resource, the one-hot encoding of the interaction type, etc.

[0173] At each time step t ,Will x t Input into a multi-layer LSTM network. LSTM uses a gating mechanism to adaptively decide whether to retain or forget historical information, and can effectively process long sequence data. t For LSTM t The hidden state of the step, then the state update formula is:

[0174] ;

[0175] ;

[0176] ;

[0177] ;

[0178] ;

[0179] ;

[0180] Among them, f t 、i t , o t They are forget gate, input gate and output gate respectively. is the candidate memory cell state, c t is the current memory cell state, σis the sigmoid activation function. LSTM parameters can be learned end-to-end using the back-propagation over time algorithm.

[0181] At the last step of the sequence t , taking the hidden state as the state feature of the learner's historical behavior, recorded as ,in D 1 is the dimension of the hidden state. It encodes the temporal pattern and semantic information of the learner's behavior sequence, which can be used for subsequent state fusion.

[0182] Learning progress feature calculation:

[0183] Learning progress reflects the learner's completion of the course knowledge points and is an important basis for recommendation decisions. In order to estimate the learner's learning progress for each knowledge point, this paper designs a progress calculation method based on behavior sequence.

[0184] First, the knowledge points of the course are organized hierarchically to construct a knowledge tree. Each node of the tree represents a knowledge point, and the parent-child relationship represents the prerequisite dependency between knowledge points. Then, the learner's behavior sequence is matched with the knowledge tree to identify the knowledge point corresponding to each behavior. c , define its completion p c for:

[0185] ;

[0186] in, X c Representation and knowledge points c The set of matching behaviors, X Represents the set of all behaviors. w ( x t ) indicates behavior x t This paper sets different weights according to the behavior type, for example, the weight of completing an exercise is higher than that of viewing a video.

[0187] By aggregating the completion of leaf nodes from bottom to top, the completion of each internal knowledge point can be recursively calculated. Finally, a learning progress feature vector is obtained. ,in D 2 is the total number of knowledge points. Each element of the vector corresponds to the completion of a knowledge point, reflecting the learner's overall learning progress in the course.

[0188] Knowledge mastery level feature calculation:

[0189] The answers of learners in formative assessments such as exercises and tests can objectively measure their mastery of knowledge points. This paper uses the Bayesian Knowledge Tracing (BKT) model to dynamically estimate the mastery probability of each knowledge point based on the learner's answer sequence.

[0190] BKT is a classic cognitive diagnosis model that describes the evolution of learners’ knowledge state through hidden Markov processes. c , defining its mastery status , which means that the learner is at time step t Have you mastered this knowledge point?

[0191] state The transition probability is determined by three parameters:

[0192] P ( L 0 ): Learners have mastered the knowledge points at the initial moment c The probability of

[0193] P ( T ): the probability that the learner moves from the unmastered state to the mastered state after each step of practice;

[0194] P ( S ): The probability that learners will still make mistakes in practice after mastering the knowledge points (error rate);

[0195] Another important parameter is P ( G ), which represents the probability (guessing rate) of a learner who has not mastered the knowledge point guessing the correct answer in the exercise. Based on the above four parameters, BKT uses Bayesian inference to calculate the learner's probability at time step t Master the knowledge points c The posterior probability of:

[0196] ;

[0197] in, represents until the time step t The answer sequence of Based on whether the learners have mastered the knowledge points c , can be obtained by P ( S )or P ( G ) calculated. It can be recursively calculated based on the posterior probability and transition probability of the previous moment:

[0198] ;

[0199] Through learners in the knowledge point c All the answer data on the , can use the expectation maximization (EM) algorithm to estimate the parameters of the BKT model. After obtaining the parameters, we can calculate the learner's understanding of the knowledge point c The master probability sequence This paper takes the mastery probability of the last step as the learner's mastery of the knowledge point. c mastery level characteristics.

[0200] Repeat the above process to calculate the learner's mastery level of all knowledge points and obtain a knowledge mastery feature vector Each element of the vector takes a value between [0, 1], indicating the probability of mastering the corresponding knowledge point. s mas It characterizes the learner's knowledge status from the perspective of cognitive diagnosis and is an important supplement to the learning progress characteristics.

[0201] State feature fusion:

[0202] After obtaining the learner's historical behavior status, learning progress, and knowledge mastery level features, this paper generates the final learner status representation through feature fusion. Specifically, the three feature vectors are concatenated with other user attribute features (such as personal information, learning preferences, etc.) into a long vector:

[0203] ;

[0204] in, is the user attribute feature vector. The concatenated state vector The dimension is high and contains heterogeneous feature information. In order to extract the high-level semantics of the state vector and reduce the redundancy between features, this paper uses a fully connected neural network to transform it:

[0205] h 0 =s;

[0206] ;

[0207] u=h L ;

[0208] Among them, W l and b l For the l The weight matrix and bias vector of the fully connected layer, σ is the ReLU activation function, L is the number of layers in the network. The output of the network This is the final representation of the user state, dimension D' Much smaller than the dimension of the original state vector D .

[0209] The parameters of the fully connected network can be learned end-to-end through the back-propagation algorithm, so that u can fully integrate the state features from different sources and extract semantic information related to the downstream recommendation task. The learning goal is to minimize the recommendation error, that is, the difference between the learner's preference for candidate resources predicted based on the state representation u and the actual preference. By jointly training the state representation network and the recommendation model, the two can adapt to each other and improve the overall recommendation performance.

[0210] By combining the user's historical behavior sequence with the knowledge graph, the user's learning status can be fully characterized. Features are extracted from the two dimensions of learning behavior and knowledge mastery, and then integrated to generate a representation of the user's overall learning status, thus providing strong support for personalized education decisions. Mapping learning behavior to the knowledge graph can quantify the user's completion of each knowledge point and measure the learning progress as a whole. At the same time, the knowledge tracking model is used to estimate the user's mastery of each knowledge point and understand the user's knowledge structure in a fine-grained manner. Combining these two aspects of features, a semantically rich user portrait vector is finally generated, which can accurately describe the current learning status.

[0211] By using recurrent neural networks to model user historical behaviors and capture temporal dependencies, and modeling behavior sequences through long short-term memory networks, we can automatically mine implicit temporal patterns between behaviors, and output a hidden state vector that integrates all historical information at the last moment as a dynamic representation of user behavior. This modeling method is closer to the real process of learning, and can reveal the deep motivation of user learning from the perspective of temporal evolution, laying the foundation for generating accurate personalized strategies.

[0212] S103. Based on the user state representation, determine the state space, and at the same time select a subset of candidate resources from the resource concept bipartite graph to determine the resource space; based on the state space and the resource space, construct an education resource recommendation model, the education resource recommendation model calculates the cumulative score through a quality evaluation function constructed based on feedback data, and determines the best recommended resource by selecting the maximum value of the cumulative score.

[0213] In an optional implementation, based on the user state representation, a state space is determined, and a subset of candidate resources is selected from the resource concept bipartite graph to determine a resource space; based on the state space and the resource space, an educational resource recommendation model is constructed, and the educational resource recommendation model calculates a cumulative score through a quality evaluation function constructed based on feedback data, and determines the best recommended resource by selecting the maximum value of the cumulative score, including:

[0214] Based on the user state representation, analyzing the user's learning progress and knowledge mastery, and determining the state space;

[0215] Selecting a subset of candidate resources matching the user state representation from the resource concept bipartite graph to determine a resource space;

[0216] Pairing the state space with the resource space to determine a state resource response set, and determining a quality evaluation function based on user feedback to construct an educational resource recommendation model, wherein the user feedback includes user instant feedback and long-term feedback on educational effects;

[0217] At the initial moment, based on the current user state representation, a first state is determined, and a first recommended resource set corresponding to the first state is generated through the educational resource recommendation model; recommended resources are selected from the first recommended resource set according to a preset selection quantity, and based on the user feedback, the first state is updated to a second state, and the next moment is entered;

[0218] Iterate in sequence until the preset termination state is reached;

[0219] Based on the recommended resources selected at each moment, a recommended path is constructed, based on the quality evaluation function, the cumulative score of each recommended path is calculated respectively, the recommended path with the highest cumulative score is selected, and the best recommended resource is determined.

[0220] Exemplarily, the state space and resource space are determined. First, the learning progress is mapped into discrete states. The user state representation includes a learning progress feature vector. The learning progress feature vector is divided into K intervals. For example, the progress of 0-1 is divided into K intervals such as 0-0.2, 0.2-0.4, ..., 0.8-1.0. Each interval represents a discrete learning progress state, which corresponds to the learning progress feature vector. One-hot encoding is used to represent one of the K states.

[0221] Map the knowledge mastery vector to the knowledge state:

[0222] The user state representation includes a knowledge mastery feature vector, each element of which corresponds to the mastery probability of a knowledge point.

[0223] The knowledge mastery vectors are clustered to obtain M cluster centers. Traditional clustering algorithms such as K-Means can be used. Each cluster center represents a knowledge mastery mode (strong, weak, etc.).

[0224] The knowledge mastery vector is classified into the nearest neighbor cluster center, and the knowledge state is represented by one-hot encoding of the cluster center index.

[0225] Determine the candidate resource subset:

[0226] In the resource concept bipartite graph, each resource node corresponds to a resource. Calculating the relevance score of each resource with the current user status will map the user status (a combination of learning progress and knowledge status) into a vector, and calculate the similarity between the vector and each resource feature vector (such as cosine similarity). The higher the similarity score, the more the resource fits the current user status. Finally, the top N resources with the highest scores are selected as the candidate resource set.

[0227] Build status-resource response collection:

[0228] Pair the learning progress status, knowledge status and candidate resource set by Cartesian product;

[0229] A triple of the form (progress state i, knowledge state j, resource k).

[0230] All such paired triplets constitute a state-resource response set, which represents which resources can be recommended in response to a specific state, providing support for the state transfer process of the subsequent MDP model.

[0231] Through the above steps, we discretize the continuous user state vector space, and filter the resource space according to resource similarity, and finally form a limited set of state-resource responses. This mapping enables the subsequent MDP-based resource recommendation problem to be efficiently modeled and solved. In terms of implementation details, the state division granularity, similarity calculation method, etc. can be adjusted according to specific needs.

[0232] When building an educational resource recommendation model, we first need to define a quality evaluation function to extract features from user feedback data:

[0233] Instant feedback features: number of likes, rating scores, etc.

[0234] Long-term feedback characteristics: post-class test scores, final exam grades, etc.

[0235] Then build a machine learning model to map the features into a quality score Q for a single recommendation:

[0236] Construct a Markov Decision Process (MDP) model:

[0237] The state set S of the MDP needs to use the discrete state space determined previously, and each state s is represented as (learning progress state i, knowledge state j).

[0238] Define the behavior set A of MDP, behavior a is to recommend a candidate resource r to the user, and the candidate resource set is given by the state-resource response set;

[0239] Define an immediate reward function R. When executing action a (recommending resource r) in state s, the immediate reward obtained is set to the score corresponding to the quality evaluation function Q.

[0240] Define the state transition probability P, which is the probability of executing action a in state s and transitioning to state s'.

[0241] Solve the MDP and obtain the optimal strategy:

[0242] The goal is to maximize the sum of all recommended paths, that is, the overall quality assessment score brought by the recommended sequence.

[0243] Use temporal difference technology to solve, such as Q-Learning, Sarsa and other reinforcement learning algorithms, and converge to the optimal Q function through continuous exploration and updating.

[0244] The optimal Q function corresponds to the optimal strategy. In each state s, the Q value corresponding to the recommended behavior a is the largest.

[0245] Through the above steps, the resource recommendation problem is modeled as an MDP, where:

[0246] Status is the user's current state;

[0247] The behavior is to recommend a resource;

[0248] The reward function is based on the quality assessment score;

[0249] The goal is to maximize the cumulative quality score of the entire recommendation path;

[0250] By solving the problem with reinforcement learning algorithm, we can get an optimal resource recommendation strategy. This strategy dynamically considers the changes in user status and takes long-term education quality as the optimization goal. Compared with traditional recommendations based only on feature scoring and sorting, this sequential decision modeling method is more forward-looking and global.

[0251] Execute resource recommendation process:

[0252] Initial time t=0, based on the current user state s0;

[0253] Output the corresponding recommended resource set R0 through the MDP model;

[0254] Select recommended resource r from R0 according to the strategy (such as based on the resource concept association strength), push r to the user, and record the user feedback. The user feedback updates the user state to s1, iterate t=1, 2, 3... Perform the above steps until the termination condition (such as the upper limit of the number of iterations, convergence, etc.) is reached, and the recommended resource sequence is recorded at each moment to form a recommendation path.

[0255] Determine the best recommended resource, calculate the cumulative score of each recommended path based on the quality evaluation function, consider the immediate feedback of each step and the final educational effect feedback, select the recommended path with the highest cumulative score, and the resource of the last step on the path as the best recommended resource output.

[0256] Figure 2 FIG. 1 is a schematic diagram of a structure of an educational resource recommendation system based on user personalization according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0257] The first unit is used to summarize all educational resources, extract keywords for each educational resource, and establish a match with a pre-constructed educational knowledge graph, construct a resource concept mapping set, organize it into a bipartite graph structure, determine the edge weight by calculating the co-occurrence frequency, construct a co-occurrence matrix, use a non-negative matrix, combine the structural constraints of the educational knowledge graph, decompose the co-occurrence matrix to obtain resource implicit representation and concept implicit representation, calculate the inner product between the resource implicit representation and the concept implicit representation, determine the resource concept association strength, and obtain a resource concept bipartite graph;

[0258] The second unit is used to collect user historical behavior data, construct a behavior sequence, match the behavior sequence with the educational knowledge graph, identify the weight of each knowledge point, calculate the completion degree of each knowledge point, and generate a learning progress feature vector; use the knowledge tracking model to calculate the user's mastery probability of each knowledge point and generate a knowledge mastery feature vector; splice the learning progress feature vector and the knowledge mastery feature vector, combine the user's basic attributes, and generate a user state representation through feature fusion;

[0259] The third unit is used to determine the state space based on the user state representation, and select a subset of candidate resources from the resource concept bipartite graph to determine the resource space; based on the state space and the resource space, an educational resource recommendation model is constructed, and the educational resource recommendation model calculates the cumulative score through a quality evaluation function constructed based on feedback data, and determines the best recommended resource by selecting the maximum value of the cumulative score.

[0260] According to a third aspect of the embodiments of the present invention,

[0261] An electronic device is provided, comprising:

[0262] processor;

[0263] a memory for storing processor-executable instructions;

[0264] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0265] A fourth aspect of the embodiments of the present invention is:

[0266] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0267] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0268] 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 it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. Educational resource recommendation method based on user personalization, It is characterized in that include: Summarize all educational resources, extract keywords for each educational resource, and match them with the pre-built educational knowledge graph, build a resource concept mapping set, organize it into a bipartite graph structure, determine the edge weights by calculating the co-occurrence frequency, build a co-occurrence matrix, use a non-negative matrix, combine the structural constraints of the educational knowledge graph, decompose the co-occurrence matrix to obtain implicit representation of resources and implicit representation of concepts, calculate the inner product between the implicit representation of resources and the implicit representation of concepts, determine the resource concept association strength, and obtain a bipartite graph of resource concepts; Collect user historical behavior data, construct a behavior sequence, match the behavior sequence with the educational knowledge graph, identify the weight of each knowledge point, calculate the completion degree of each knowledge point, and generate a learning progress feature vector; Using the knowledge tracking model, the probability of the user mastering each knowledge point is calculated to generate a knowledge mastery feature vector; the learning progress feature vector and the knowledge mastery feature vector are concatenated, combined with the user's basic attributes, and the user state representation is generated through feature fusion; Determine a state space based on the user state representation, and select a subset of candidate resources from the resource concept bipartite graph to determine a resource space; Based on the state space and the resource space, construct an educational resource recommendation model, wherein the educational resource recommendation model calculates a cumulative score through a quality evaluation function constructed based on feedback data, and determines the best recommended resource by selecting the maximum value of the cumulative score; Extract keywords from each educational resource and match them with the pre-built educational knowledge graph to build a resource concept mapping set, including: Extract keywords from each of the educational resources, find literal matching concepts corresponding to the keywords in the educational knowledge graph, and establish a primary mapping set; Based on the literal matching concept, the neighbor concept of the literal matching concept is found in the educational knowledge graph, and the semantic similarity value between the educational resource and the neighbor concept is calculated. When the semantic similarity value is higher than a preset adoption threshold, the mapping relationship between the educational resource and the neighbor concept is added to the primary mapping set to construct a resource concept mapping set, wherein the neighbor concept includes a superordinate concept and a subordinate concept.

2. The method according to claim 1, It is characterized in that The resource concept mapping set is organized into a bipartite graph structure. The edge weight is determined by calculating the co-occurrence frequency, and a co-occurrence matrix is ​​constructed. The co-occurrence matrix is ​​decomposed by using a non-negative matrix and combining the structural constraints of the educational knowledge graph to obtain implicit resource representation and implicit concept representation. The resource concept association strength is determined by calculating the inner product between the implicit resource representation and the implicit concept representation. The resource concept bipartite graph includes: The resource concept mapping set is split and combined into a bipartite graph structure according to the classification of resource nodes and concept nodes, the edge weights of resources and concepts are calculated by a word frequency weighting method, and all edge weights are organized into a co-occurrence matrix; Decomposing the co-occurrence matrix into two non-negative matrices, adding regularization terms based on the structural constraints of the educational knowledge graph, constructing a decomposition error function, minimizing the decomposition error, using the alternating least squares method to obtain the optimal solution of the two non-negative matrices, and determining the implicit representation of resources and the implicit representation of concepts; Based on each pair of the implicit representation of the resource and the implicit representation of the concept, the inner product is calculated to determine the resource concept association strength, and the resources are arranged in descending order. According to the preset preferred index, the optimal association result is determined to obtain a bipartite graph of resource concepts.

3. The method according to claim 2, It is characterized in that The edge weights of resources and concepts are calculated using the word frequency weighting method. The formula is as follows: ; in, r Represents resources, c Represents the concept, ω rc represents the edge weights of resources and concepts, N rc Representation Resources r With concept c The number of co-occurrences, N r Representation Resources r Number of occurrences, R Represents a resource collection, | R | indicates the total number of resources, R c Representation and Concept c The associated resource collection, | R c |Representation and Concept c The number of associated resources; Combined with the regularization term, the decomposition error is minimized, and the formula is as follows: ; Among them, U represents the first non-negative matrix after decomposition, V represents the second non-negative matrix after decomposition, W represents the co-occurrence matrix, and F represents the Frobenius norm. λ 1 represents the first regularization coefficient, u r Represents the corresponding resources in matrix U r The characteristic vector of M r Representation and Resources r The set of all concepts associated with j express M r The first in the collection j The feature vector of a resource, λ 2 represents the second regularization coefficient, C represents a concept set, v c Represents the corresponding concept in the matrix V c The characteristic vector of M c Representation and Concept c The collection of all associated resources, v i express M c The first in the collection i The feature vector of a resource.

4. The method according to claim 1, It is characterized in that Collecting user historical behavior data, constructing a behavior sequence, matching the behavior sequence with the educational knowledge graph, identifying the weight of each knowledge point, calculating the completion degree of each knowledge point, generating a learning progress feature vector, and using the knowledge tracking model to calculate the user's mastery probability of each knowledge point. Generating a knowledge mastery feature vector includes: Arrange the user's historical behavior data in chronological order to construct a behavior sequence, use a multi-layer long short-term memory network to model the behavior sequence, determine the temporal dependency, and determine the user's historical behavior state characteristics based on the last hidden state of the multi-layer long short-term memory network and the temporal dependency; Based on the educational knowledge graph, the weight of the knowledge point is determined, the behavior sequence is matched with the educational knowledge graph, the knowledge point corresponding to each behavior is determined, and the completion degree of each knowledge point is determined in combination with the weight of the knowledge point. Based on the knowledge point association structure in the educational knowledge graph, the completion degree of the associated knowledge points is aggregated, and the overall completion degree corresponding to the behavior sequence is recursively calculated to generate a learning progress feature vector; Based on each knowledge point, using the knowledge tracking model, determine the transition process of the user's mastery of the knowledge point, wherein the mastery state is determined by the initial mastery probability, the transition probability, the error rate and the guessing rate; Based on the user's answer sequence on each knowledge point, the user's knowledge mastery rate is determined by the error rate and the guessing rate. Combined with the mastery probability at the previous moment, the mastery probability of the user mastering the knowledge point is recursively calculated. The expectation maximization algorithm is used to calculate the user's mastery probability sequence for the knowledge point, and the mastery probability corresponding to the last step of the mastery probability sequence is extracted to determine the mastery feature. The above steps are repeated to calculate the user's mastery features for all knowledge points and generate a knowledge mastery feature vector.

5. The method according to claim 4, It is characterized in that Calculating knowledge to master the feature vector includes: Master probability calculation and recursive calculation, the formula is as follows: ; in, k a t Indicates that at time step t Whether users have mastered the knowledge points a status, k a t-1 Indicates that at time step t- 1. Whether the user has mastered the knowledge points a status, o 1 t From time step 1 to t The user answer sequence, o 1 t-1 From time step 1 to t- 1 user answer sequence, o t Indicates that at time step t The answer result of Indicates the answer sequence o 1 t In the case of user at time step t About knowledge points a The probability of mastering Indicates that if the user is at time step t Master the knowledge points a The probability of giving the correct answer, Indicates the answer sequence o 1 t-1 In the case of user at time step t About knowledge points a The probability of mastering Indicates the answer sequence o 1 t-1 In the case of user at time step t-1 About knowledge points a The probability of mastering Represents the probability that a user moves from not mastering to mastering after one practice.

6. The method according to claim 1, It is characterized in that Based on the user state representation, a state space is determined, and a subset of candidate resources is selected from the resource concept bipartite graph to determine a resource space; based on the state space and the resource space, an educational resource recommendation model is constructed, and the educational resource recommendation model calculates a cumulative score through a quality evaluation function constructed based on feedback data, and determines the best recommended resource by selecting the maximum value of the cumulative score, including: Based on the user state representation, analyzing the user's learning progress and knowledge mastery, and determining the state space; Selecting a subset of candidate resources matching the user state representation from the resource concept bipartite graph to determine a resource space; Pairing the state space with the resource space to determine a state resource response set, and determining a quality evaluation function based on user feedback to construct an educational resource recommendation model, wherein the user feedback includes user instant feedback and long-term feedback on educational effects; At the initial moment, based on the current user state representation, a first state is determined, and a first recommended resource set corresponding to the first state is generated through the educational resource recommendation model; recommended resources are selected from the first recommended resource set according to a preset selection quantity, and based on the user feedback, the first state is updated to a second state, and the next moment is entered; Iterate in sequence until the preset termination state is reached; Based on the recommended resources selected at each moment, a recommended path is constructed, based on the quality evaluation function, the cumulative score of each recommended path is calculated respectively, the recommended path with the highest cumulative score is selected, and the best recommended resource is determined.

7. An educational resource recommendation system based on user personalization, used to implement the method described in any one of claims 1 to 6. It is characterized in that include: The first unit is used to summarize all educational resources, extract keywords for each educational resource, and establish a match with a pre-constructed educational knowledge graph, construct a resource concept mapping set, organize it into a bipartite graph structure, determine the edge weight by calculating the co-occurrence frequency, construct a co-occurrence matrix, use a non-negative matrix, combine the structural constraints of the educational knowledge graph, decompose the co-occurrence matrix to obtain resource implicit representation and concept implicit representation, calculate the inner product between the resource implicit representation and the concept implicit representation, determine the resource concept association strength, and obtain a resource concept bipartite graph; The second unit is used to collect user historical behavior data, construct a behavior sequence, match the behavior sequence with the educational knowledge graph, identify the weight of each knowledge point, calculate the completion degree of each knowledge point, and generate a learning progress feature vector; Using the knowledge tracking model, the probability of the user mastering each knowledge point is calculated to generate a knowledge mastery feature vector; the learning progress feature vector and the knowledge mastery feature vector are concatenated, combined with the user's basic attributes, and the user state representation is generated through feature fusion; A third unit is used to determine a state space based on the user state representation, and select a candidate resource subset from the resource concept bipartite graph to determine a resource space; Based on the state space and the resource space, an educational resource recommendation model is constructed. The educational resource recommendation model calculates the cumulative score through a quality evaluation function constructed based on feedback data, and determines the best recommended resource by selecting the maximum value of the cumulative score.

8. An electronic device, It is characterized in that include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, It is characterized in that When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.