A MOOCs learning resource recommendation method

By constructing the student-learning resource scoring matrix and co-occurrence relationship matrix, combining the topic word characteristics, an additive low-rank sparse matrix decomposition model was constructed, and resource recommendation problems caused by data sparseness in the MOOCs platform were solved, and resources that are interested in students were realized accurate prediction and recommendation.

CN118377972BActive Publication Date: 2024-11-12CHANGSHU INSTITUTE OF TECHNOLOGY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410815994.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-06-24
Publication Date
2024-11-12
Estimated Expiration
2044-06-24

AI Technical Summary

Technical Problem

In the MOOCs platform, there is a problem of sparse data when recommending resource recommendations based on the similarity of student users or the similarity of learning resource content, and it is impossible to accurately predict the learning resources students are interested in next time.

Method used

Construct the student-learning resource scoring matrix, student co-occurrence relationship matrix and learning resource co-occurrence relationship matrix, combine the topical word characteristics of the students' learning resources and the topical word characteristics of the learning resources, and construct the additive low-rank sparse matrix decomposition model, through Iterative optimization solves the potential factor matrix of students and learning resources, calculates the recommended scores and sorts the recommended learning resources.

Benefits of technology

It effectively overcomes the problem of data sparseness, realizes accurate prediction of learning resources that are of interest to students, and improves the accuracy and utilization of recommendations.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118377972B_ABST
    Figure CN118377972B_ABST
Patent Text Reader

Abstract

The invention discloses a MOOCs learning resource recommendation method, which comprises constructing a student-learning resource scoring matrix, constructing a student co-occurrence relationship matrix based on the co-learning situation of the learning resources learned by the students, the registered courses and the selected teachers, constructing a cosine similarity matrix based on the content keywords of the learning resources learned by the students, constructing a learning resource co-occurrence relationship matrix based on the situation of the learning resources being learned by the students, the situation of the creating teachers and the names of the courses to which they belong, constructing a cosine similarity matrix based on the content keywords of the learning resources, constructing an additive low-rank sparse matrix decomposition model, iteratively optimizing the student potential matrix and the learning resource potential matrix of the model, calculating the low-rank approximation of the student potential factor matrix and the sparse approximation of the learning resource potential factor matrix, and then calculating the recommendation score of the student learning resource, and recommending the corresponding learning resource according to the recommendation score.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to a learning resource recommendation method, in particular to a MOOCs learning resource recommendation method. Background Art

[0002] MOOCs platforms provide a large number of learning resources. Students face the problem of information overload and need to spend a lot of time and energy to browse, retrieve and locate the required learning resources. Therefore, MOOCs platforms should be able to filter and select the most relevant and valuable learning resources according to learners' interest preferences, backgrounds and learning goals, helping them to more accurately select learning resources that suit them and improve the utilization rate of learning resources. However, the current MOOCs platforms have data sparsity problems when recommending resources based on the similarity of student users or the similarity of learning resource content. They cannot predict and recommend the learning resources that student users may be interested in next time based on the learning resources selected by student users. To this end, we propose a MOOCs learning resource recommendation method. Summary of the invention

[0003] Purpose of the invention: The technical problem to be solved by the present invention is the problem of data sparsity when recommending resources based on the similarity of student users or the similarity of learning resource contents in MOOCs platforms, and the inability to predict the learning resources that student users will be interested in next time based on the learning resources selected by student users each time.

[0004] In order to solve the above technical problems, the present invention provides the following technical solutions: a MOOCs learning resource recommendation method, comprising the following steps:

[0005] Step 1: Construct a student-learning resource scoring matrix for the MOOCs platform , assuming there is students , Learning Resources , the constructed student-learning resource scoring matrix is If the student Learned learning resources ,but ,otherwise ;

[0006] Step 2: Construct the co-occurrence relationship matrix of students and the cosine similarity matrix of the subject word features of the learning resources learned by students , the student co-occurrence relationship matrix includes the student co-learning resource number matrix , the matrix of the number of courses selected by students , and the matrix of the number of teachers shared by students ;

[0007] Step 3: Construct the co-occurrence relationship matrix of learning resources and the cosine similarity matrix of learning resource keyword features The learning resource co-occurrence relationship matrix includes the learning resource total student number matrix , learning resource course co-occurrence matrix , and the learning resource teacher co-occurrence matrix ;

[0008] Step 4: Based on the student-learning resource scoring matrix , student co-occurrence relationship matrix, cosine similarity matrix of the subject word features of the learning resources learned by students , the co-occurrence relationship matrix of learning resources, and the cosine similarity matrix of learning resource keyword features , construct an additive low-rank sparse matrix decomposition model for MOOCs learning resources:

[0009] ,

[0010] in , , and Represents the student-learning resource scoring matrix Decomposing the error term , the student co-occurrence matrix and the cosine similarity matrix of the subject word features of the learning resources studied by the students and The loss error term , learning resource co-occurrence relationship matrix and cosine similarity matrix of learning resource keyword features and The loss error term and regular penalty to prevent overfitting The weight of , , L is the student latent factor matrix, which is used to represent the correlation between students' interest in learning resources, S is the learning resource latent factor matrix, which is used to represent the learning resource characteristics and potential predictions of each student's preference, T is the transposition operation, and the matrix and is a balanced matrix, and the constraint is that the rank of the student latent factor matrix L is less than or equal to a constant , the number of non-zero elements of the learning resource latent factor matrix S, the balance matrix N, and the balance matrix M are less than or equal to a constant , and , is the matrix Frobenius norm;

[0011] Step 5, perform iterative optimization based on the additive low-rank sparse matrix decomposition model to solve the student latent factor matrix L and the learning resource latent factor matrix S of the additive low-rank sparse matrix decomposition model:

[0012] Step 5-1, initialize matrix variables: L=X, S=0, M=0, N=0;

[0013] Step 5-2, fix the learning resource latent factor matrix S, the balance matrix M and the balance matrix N, and calculate the update expression of the student latent factor matrix L:

[0014] According to the additive low-rank sparse matrix decomposition model, the partial derivative of the student latent factor matrix L is equal to 0, and the expression of the student latent factor matrix L is calculated as:

[0015] ,

[0016] in is the identity matrix;

[0017] Step 5-3, fix the student latent factor matrix L, balance matrix M and balance matrix N, and calculate the updated expression of the learning resource latent factor matrix S:

[0018] According to the additive low-rank sparse matrix decomposition model, the partial derivative of the learning resource potential factor matrix S is equal to 0, and the expression of the learning resource potential factor matrix S is calculated as follows:

[0019] ;

[0020] Step 5-4, fix the learning resource latent factor matrix S, the student latent factor matrix L and the balance matrix N, and calculate the update expression of the balance matrix M:

[0021] According to the additive low-rank sparse matrix decomposition model, the partial derivative of the balanced matrix M is equal to 0, and the expression of the balanced matrix M is calculated as follows:

[0022] ;

[0023] Step 5-5, fix the learning resource latent factor matrix S, the student latent factor matrix L and the balance matrix M, and calculate the update expression of the balance matrix N:

[0024] According to the additive low-rank sparse matrix decomposition model, the partial derivative of the balanced matrix N is equal to 0, and the expression for calculating the balanced matrix N is:

[0025] ;

[0026] Step 5-6, repeat iterative steps 5-2 to 5-5 until the number of iterations reaches the maximum or the additive low-rank sparse matrix decomposition model converges, and output the final learning resource latent factor matrix S and student latent factor matrix L;

[0027] Step 6: Calculate the low-rank approximation of the student latent factor matrix L to obtain the student latent factor low-rank matrix , calculate the sparse approximation of the learning resource potential factor matrix S, and obtain the learning resource potential factor sparse matrix ;

[0028] Step 7: Based on the student latent factor low-rank matrix and the learning resource latent factor sparse matrix Calculate the recommended scores of all learning resources for students :

[0029] Recommended score is the low-rank matrix of student latent factors and the learning resource latent factor sparse matrix The sum of , for a given student ,in , retrieve the given student The recommendation scores of the corresponding learning resources are sorted in descending order, and the learning resources corresponding to the top K largest recommendation scores are recommended to the student.

[0030] As a further preference of the present invention, step 2 specifically comprises the following steps:

[0031] Step 2-1: Construct a matrix of the number of students’ shared learning resources based on the shared learning situation of the learning resources that students have learned ;

[0032] Assumption Student The resource set learned is ,student The resource set learned is ,statistics and Resource sets learned , construct the matrix of the number of common learning resources between students as follows:

[0033] ,in ;

[0034] Step 2-2: Construct a matrix of the number of courses selected by students based on the courses of the learning resources that students have registered for. ;

[0035] Assumption Student The course set registered is ,student The course set registered is ,statistics and Common courses that everyone has registered for , construct the matrix of the number of courses selected by students as follows:

[0036] ,in ;

[0037] Step 2-3: Construct a matrix of the number of teachers shared by students based on the teachers selected by the students ;

[0038] Assumption Student The set of all teachers is ,student The set of all teachers is ,statistics and Total teachers set , construct the matrix of the total number of teachers among students:

[0039] ,in ;

[0040] Step 2-4: Based on the content keywords of the learning resources learned by students, a cosine similarity matrix of the keyword characteristics of the learning resources learned by students reflecting their learning interests is constructed. ;

[0041] Step 2-4-1, extract the keyword features of the learning resources learned by students based on the Transformer model:

[0042] Assumption Student The subject word set of the learning resources is , For Students The number of subject words in the learning resources studied, students The subject word set of the learning resources is , For Students The number of keywords in the learning resources is learned. The Transformer model is used to learn and extract the features of each keyword. The learning resources The subject word features are , ,student The learning resources The subject word features are , ;

[0043] Step 2-4-2, construct the cosine similarity matrix of the subject word features of the learning resources learned by students that reflects their learning interests:

[0044] Connect with students Learning resources Keyword Features As input, it is processed through a fully connected layer and a normalization layer and then outputs the student The characteristics of the learning resources studied, i.e. ; Connect students Learning resources Keyword Features As input, it is processed through a fully connected layer and a normalization layer and then outputs the student The characteristics of the learning resources studied, i.e. , construct the cosine similarity matrix of the subject word features of the learning resources learned by students ,in .

[0045] As a further preference of the present invention, step 3 specifically comprises the following steps:

[0046] Step 3-1: Construct a matrix of the total number of students studying the learning resources based on how the learning resources are studied by students. ;

[0047] Assuming you have studied the learning resources The student set is , studied learning resources The student set is , then you have learned and The student set is , construct the total number of students in learning resources matrix as:

[0048] ,in ;

[0049] Step 3-2: Construct a learning resource course co-occurrence matrix based on the course name to which the learning resource belongs ;

[0050] Hypothesis Learning Resources The course name is , learning resources The course name is , then the co-occurrence matrix of learning resource courses is constructed as:

[0051] ,in ,

[0052] if , then the matrix elements , otherwise ;

[0053] Step 3-3, construct the learning resource teacher co-occurrence matrix based on the teacher who created the learning resource ;

[0054] Hypothesis Learning Resources The teacher who created , learning resources The teacher who created , then the learning resource teacher co-occurrence matrix is ​​constructed as:

[0055] ,in ,

[0056] if , then the matrix elements , otherwise ;

[0057] Step 3-4: Construct a cosine similarity matrix of learning resource keyword features based on the content keywords of the learning resource ;

[0058] Step 3-4-1, extract the topic features of resources based on Transformer model learning;

[0059] Hypothesis Learning Resources The subject word set is , Resources for learning The number of keywords, learning resources The subject word set is , Resources for learning The number of keywords, using the Transformer model to learn and extract the features of each keyword, learning resources No. The subject word features are , , Learning Resources No. The subject word features are , ;

[0060] Step 3-4-2, construct the cosine similarity matrix of learning resource keyword features:

[0061] Connect to learning resources of Keyword Features As the input layer, it is processed through a fully connected layer and a normalization layer to output the learning resource The characteristics of , connect to learning resources of Keyword Features , processed through a fully connected layer and a normalization layer, outputting the learning resource The characteristics of , construct the cosine similarity matrix of learning resource theme word features

[0062] ,in .

[0063] As a further preference of the present invention, step 6 specifically comprises the following steps:

[0064] Step 6-1, calculate the low-rank approximation of the student latent factor matrix L:

[0065] If the rank of the student latent factor matrix L is Greater than , using singular value decomposition for low-rank approximation:

[0066] First, the singular value decomposition of the student latent factor matrix L is , where matrices A and B are unitary matrices, and the diagonal matrix And the size of the eigenvalue satisfies ;

[0067] Secondly, keep the previous The rest are set to 0, that is, , through matrices A, B and Multiply them together to perform low-rank approximation and obtain the low-rank matrix of student potential factors ,Right now ;

[0068] If the rank of the student latent factor matrix L is Less than or equal to ,but ;

[0069] Step 6-2, calculate the sparse approximation of the learning resource latent factor matrix S:

[0070] If the number of non-zero elements in the learning resource latent factor matrix S is greater than , then by retaining the previous The largest matrix elements and the rest are set to 0 to perform matrix sparse approximation and obtain the sparse matrix of learning resource potential factors ;

[0071] If the number of non-zero elements in the learning resource latent factor matrix S is less than or equal to ,but ;

[0072] If the number of non-zero elements in the balance matrix N is greater than , then by retaining the previous The largest matrix elements and the rest are set to 0 to perform matrix sparse approximation to obtain a sparse balanced matrix ;

[0073] If the number of non-zero elements in the balanced matrix N is less than or equal to ,but ;

[0074] If the number of non-zero elements in the balance matrix M is greater than , then by retaining the previous The largest matrix elements and the rest are set to 0 to perform matrix sparse approximation to obtain a sparse balanced matrix ;

[0075] If the number of non-zero elements in the balance matrix M is less than or equal to ,but .

[0076] The beneficial effects of the present invention are as follows: the present invention effectively overcomes the problem of sparse data in the student-learning resource scoring matrix by constructing a student co-occurrence relationship matrix, a learning resource co-occurrence relationship matrix, a cosine similarity matrix of the subject word features of the learning resources learned by students, and a cosine similarity matrix of the subject word features of the learning resources; an additive low-rank sparse matrix decomposition model of MOOCs learning resources is constructed based on the student-learning resource scoring matrix, the student co-occurrence relationship matrix, the cosine similarity matrix of the subject word features of the learning resources learned by students, the learning resource co-occurrence relationship matrix, and the cosine similarity matrix of the subject word features of the learning resources, and the model is iteratively optimized to solve the student latent factor matrix and the learning resource latent factor matrix; the student latent factor matrix represents the common patterns and potential structures among students who are interested in learning resources, There is a high correlation between students' interests and there are relatively few patterns. Therefore, the rank of the student latent factor matrix should be small and needs to be low-rank approximated to obtain the low-rank matrix of student latent factors. The learning resource latent factor matrix represents the characteristics and potential predictions of learning resources preferred by each student. Individual students will only be interested in a small number of learning resources. Therefore, sparse approximation is required to obtain a sparse matrix of learning resource latent factors. The recommendation score is obtained by adding the sparse matrix of learning resource latent factors and the low-rank matrix of student latent factors. For a given student, the recommendation score of the learning resource corresponding to the given student is retrieved, and the recommendation scores are sorted in order from large to small. The learning resources corresponding to the top K largest recommendation scores are recommended to the student, which can accurately predict the learning resources that students are interested in. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0078] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0079] The specific working process of the present invention is as follows Figure 1 shown.

[0080] Step 1: Construct a student-learning resource scoring matrix for the MOOCs platform , assuming there is students , Learning Resources , the constructed student-learning resource scoring matrix is If the student Learned learning resources ,but ,otherwise ;

[0081] Construct a student-learning resource scoring matrix for the MOOCs platform based on which learning resources students have studied Due to the limited time and energy of students, it is impossible to learn all the learning resources completely. They can only learn a small part. If we want to predict which of the remaining learning resources students are interested in based on the small part of learning resources they have learned, we will face the problem of data sparsity. In other words, there is too little data on the learning resources that students are interested in, which makes the accuracy of learning resource recommendations insufficient and requires more data.

[0082] Step 2: Construct the co-occurrence relationship matrix of students and the cosine similarity matrix of the subject word features of the learning resources learned by students , the student co-occurrence relationship matrix includes the student co-learning resource number matrix , the matrix of the number of courses selected by students , and the matrix of the number of teachers shared by students ;

[0083] Preferably, step 2 specifically includes the following steps:

[0084] Step 2-1: Construct a matrix of the number of students’ shared learning resources based on the shared learning situation of the learning resources that students have learned ;

[0085] Assumption Student The resource set learned is ,student The resource set learned is ,statistics and Resource sets learned , construct the matrix of the number of common learning resources between students as follows:

[0086] ,in ;

[0087] If students learn one or more of the same learning resources together, they can get the connection that they may be in the same class or have the same interests or majors. This is a relationship constructed through the common learning resources as a medium.

[0088] Step 2-2: Construct a matrix of the number of courses selected by students based on the courses of the learning resources that students have registered for. ;

[0089] Assumption Student The course set registered is ,student The course set registered is ,statistics and Common courses that everyone has registered for , construct the matrix of the number of courses selected by students as follows:

[0090] ,in ;

[0091] If students register for the same course, then the learning resources of the course they study are the same. This relationship of studying the same course further clarifies the relevance of learning resources between students.

[0092] Step 2-3: Construct a matrix of the number of teachers shared by students based on the teachers selected by the students ;

[0093] Assumption Student The set of all teachers is ,student The set of all teachers is ,statistics and Total teachers set , construct the matrix of the total number of teachers among students:

[0094] ,in ;

[0095] If students choose the same teacher, then the students of the same teacher are also a kind of relationship, and the relevance of learning resources between students can be further clarified based on this relationship;

[0096] Step 2-4: Based on the content keywords of the learning resources learned by students, a cosine similarity matrix of the keyword characteristics of the learning resources learned by students reflecting their learning interests is constructed. ;

[0097] Step 2-4-1, extract the keyword features of the learning resources learned by students based on the Transformer model:

[0098] Assumption Student The subject word set of the learning resources is , For Students The number of subject words in the learning resources studied, students The subject word set of the learning resources is , For Students The number of keywords in the learning resources is learned. The Transformer model is used to learn and extract the features of each keyword. The learning resources The subject word features are , ,student The learning resources The subject word features are , ;

[0099] Step 2-4-2, construct the cosine similarity matrix of the subject word features of the learning resources learned by students that reflects their learning interests:

[0100] Connect with students Learning resources Keyword Features As input, it is processed through a fully connected layer and a normalization layer and then outputs the student The characteristics of the learning resources studied, i.e. ; Connect students Learning resources Keyword Features As input, it is processed through a fully connected layer and a normalization layer and then outputs the student The characteristics of the learning resources studied, i.e. , construct the cosine similarity matrix of the subject word features of the learning resources learned by students ,in ;

[0101] According to the general recommendation algorithm, only based on the student-learning resource scoring matrix It is possible to predict the data of learning resources that students are interested in, but the student-learning resource scoring matrix There is too little data and data sparsity, so we need more information about the characteristics of the learning resources themselves, such as what common features are there in the content of the learning resources that students learn.

[0102] Step 3: Construct the co-occurrence relationship matrix of learning resources and the cosine similarity matrix of learning resource keyword features The learning resource co-occurrence relationship matrix includes the learning resource total student number matrix , learning resource course co-occurrence matrix , and the learning resource teacher co-occurrence matrix ;

[0103] Preferably, step 3 specifically includes the following steps:

[0104] Step 3-1: Construct a matrix of the total number of students studying the learning resources based on how the learning resources are studied by students. ;

[0105] Assuming you have studied the learning resources The student set is , studied learning resources The student set is , then you have learned and The student set is , construct the total number of students in learning resources matrix as:

[0106] ,in ;

[0107] If different learning resources are learned by the same group of students, then there will be some correlation between these same group of students, and a connection relationship can be established through these same group of students;

[0108] Step 3-2: Construct a learning resource course co-occurrence matrix based on the course name to which the learning resource belongs ;

[0109] Hypothesis Learning Resources The course name is , learning resources The course name is , then the co-occurrence matrix of learning resource courses is constructed as:

[0110] ,in ,

[0111] if , then the matrix elements , otherwise ;

[0112] Different learning resources belong to a certain course name, so the relationship between learning resources can be established through this course name;

[0113] Step 3-3, construct the learning resource teacher co-occurrence matrix based on the teacher who created the learning resource ;

[0114] Hypothesis Learning Resources The teacher who created , learning resources The teacher who created , then the learning resource teacher co-occurrence matrix is ​​constructed as:

[0115] ,in ,

[0116] if , then the matrix elements , otherwise ;

[0117] If different learning resources are created by the same teacher, the relationship between the learning resources can be established through this teacher;

[0118] Step 3-4: Construct a cosine similarity matrix of learning resource keyword features based on the content keywords of the learning resource ;

[0119] Step 3-4-1, extract the topic features of resources based on Transformer model learning;

[0120] Hypothesis Learning Resources The subject word set is , Resources for learning The number of keywords, learning resources The subject word set is , Resources for learning The number of keywords, using the Transformer model to learn and extract the features of each keyword, learning resources No. The subject word features are , , Learning Resources No. The subject word features are , ;

[0121] Step 3-4-2, construct the cosine similarity matrix of learning resource keyword features:

[0122] Connect to learning resources of Keyword Features As the input layer, it is processed through a fully connected layer and a normalization layer to output the learning resource The characteristics of , connect to learning resources of Keyword Features , processed through a fully connected layer and a normalization layer, outputting the learning resource The characteristics of , construct the cosine similarity matrix of learning resource theme word features

[0123] ,in ;

[0124] Through all the learning resources, the characteristics of the learning resource keywords are extracted, and then the similarities in the content of the learning resources are constructed, and then the connection relationship is established.

[0125] Step 4: Based on the student-learning resource scoring matrix , student co-occurrence relationship matrix, cosine similarity matrix of the subject word features of the learning resources learned by students , the co-occurrence relationship matrix of learning resources, and the cosine similarity matrix of learning resource keyword features , construct an additive low-rank sparse matrix decomposition model for MOOCs learning resources:

[0126] ,

[0127] in , , and Represents the student-learning resource scoring matrix Decomposing the error term , the student co-occurrence matrix and the cosine similarity matrix of the subject word features of the learning resources studied by the students and The loss error term , learning resource co-occurrence relationship matrix and cosine similarity matrix of learning resource keyword features and The loss error term and regular penalty to prevent overfitting The weight of , , L is the student latent factor matrix, which is used to represent the correlation between students' interest in learning resources, S is the learning resource latent factor matrix, which is used to represent the learning resource characteristics and potential predictions of each student's preference, T is the transposition operation, and the matrix and is a balanced matrix, and the constraint is that the rank of the student latent factor matrix L is less than or equal to a constant , the number of non-zero elements of the learning resource latent factor matrix S, the balance matrix N, and the balance matrix M are less than or equal to a constant , and , is the matrix Frobenius norm;

[0128] constant , , and Can be any constant. When using this model, the user can set the constant , , and The specific value of .

[0129] Step 5, perform iterative optimization based on the additive low-rank sparse matrix decomposition model to solve the student latent factor matrix L and the learning resource latent factor matrix S of the additive low-rank sparse matrix decomposition model:

[0130] When calculating the iterative update expression of a variable along the direction of the objective function's gradient descent, the remaining variables are treated as constants, and then other variables are calculated alternately until the objective function is optimal;

[0131] Step 5 specifically includes the following steps:

[0132] Step 5-1, initialize matrix variables: L=X, S=0, M=0, N=0;

[0133] Step 5-2, fix the learning resource latent factor matrix S, the balance matrix M and the balance matrix N, and calculate the update expression of the student latent factor matrix L:

[0134] According to the additive low-rank sparse matrix decomposition model, the partial derivative of the student latent factor matrix L is equal to 0, and the expression of the student latent factor matrix L is calculated as:

[0135] ,

[0136] in is the identity matrix;

[0137] Step 5-3, fix the student latent factor matrix L, balance matrix M and balance matrix N, and calculate the updated expression of the learning resource latent factor matrix S:

[0138] According to the additive low-rank sparse matrix decomposition model, the partial derivative of the learning resource potential factor matrix S is equal to 0, and the expression of the learning resource potential factor matrix S is calculated as follows:

[0139] ;

[0140] Step 5-4, fix the learning resource latent factor matrix S, the student latent factor matrix L and the balance matrix N, and calculate the update expression of the balance matrix M:

[0141] According to the additive low-rank sparse matrix decomposition model, the partial derivative of the balanced matrix M is equal to 0, and the expression of the balanced matrix M is calculated as follows:

[0142] ;

[0143] Step 5-5, fix the learning resource latent factor matrix S, the student latent factor matrix L and the balance matrix M, and calculate the update expression of the balance matrix N:

[0144] According to the additive low-rank sparse matrix decomposition model, the partial derivative of the balanced matrix N is equal to 0, and the expression for calculating the balanced matrix N is:

[0145] ;

[0146] Step 5-6, repeat iterative steps 5-2 to 5-5 until the number of iterations reaches the maximum or the additive low-rank sparse matrix decomposition model converges, and output the final learning resource latent factor matrix S and student latent factor matrix L.

[0147] Step 6: Calculate the low-rank approximation of the student latent factor matrix L to obtain the student latent factor low-rank matrix , calculate the sparse approximation of the learning resource potential factor matrix S, and obtain the learning resource potential factor sparse matrix ;

[0148] Preferably, step 6 specifically includes the following steps:

[0149] Step 6-1, calculate the low-rank approximation of the student latent factor matrix L:

[0150] The student latent factor matrix L describes the common patterns and latent structures among students in terms of their interest in learning resources. The interests among students are highly correlated, so the rank of the student latent factor matrix is It cannot be too high and needs to satisfy the constraints of the additive low-rank sparse matrix decomposition model: that is, the rank of the student's latent factor matrix Less than or equal to the set constant , so the rank of the student latent factor matrix L needs to be Make judgments and categorize calculations;

[0151] If the rank of the student latent factor matrix L is Greater than , using singular value decomposition for low-rank approximation:

[0152] First, the singular value decomposition of the student latent factor matrix L is , where matrices A and B are unitary matrices, and the diagonal matrix And the size of the eigenvalue satisfies ;

[0153] Secondly, keep the previous The rest are set to 0, that is, , through matrices A, B and Multiply them together to perform low-rank approximation and obtain the low-rank matrix of student potential factors ,Right now ;

[0154] If the rank of the student latent factor matrix L is Less than or equal to ,but ;

[0155] By the rank of the student latent factor matrix L and The size relationship between them is used to determine whether the student potential factor matrix L needs to be low-rank processed. If the rank of the student potential factor matrix L is Less than or equal to , then the student latent factor matrix L defaults to be equal to the student latent factor low-rank matrix ;

[0156] Step 6-2, calculate the sparse approximation of the learning resource latent factor matrix S:

[0157] The learning resource latent factor matrix S describes the learning resource characteristics and potential predictions of each student's preference. Therefore, the learning resource latent factor matrix S needs to be sparsely processed to meet the constraints of the additive low-rank sparse matrix decomposition model: that is, the number of non-zero elements of the learning resource latent factor matrix S, the balance matrix N, and the balance matrix M are less than or equal to the constant , and , so it is necessary to judge and calculate the learning resource potential factor matrix S, the balance matrix N, and the balance matrix M;

[0158] If the number of non-zero elements in the learning resource latent factor matrix S is greater than , then by retaining the previous The largest matrix elements and the rest are set to 0 to perform matrix sparse approximation and obtain the sparse matrix of learning resource potential factors ;

[0159] If the number of non-zero elements in the learning resource latent factor matrix S is less than or equal to ,but ;

[0160] If the number of non-zero elements in the balance matrix N is greater than , then by retaining the previous The largest matrix elements and the rest are set to 0 to perform matrix sparse approximation to obtain a sparse balanced matrix ;

[0161] If the number of non-zero elements in the balanced matrix N is less than or equal to ,but ;

[0162] If the number of non-zero elements in the balance matrix M is greater than , then by retaining the previous The largest matrix elements and the rest are set to 0 to perform matrix sparse approximation to obtain a sparse balanced matrix ;

[0163] If the number of non-zero elements in the balance matrix M is less than or equal to ,but .

[0164] Step 7: Based on the student latent factor low-rank matrix and the learning resource latent factor sparse matrix Calculate the recommended scores of all learning resources for students :

[0165] Recommended score is the low-rank matrix of student latent factors and the learning resource latent factor sparse matrix The sum of , for a given student ,in , retrieve the given student The recommendation scores of the corresponding learning resources are sorted in descending order, and the learning resources corresponding to the top K largest recommendation scores are recommended to the student.

[0166] The existing learning resource recommendation algorithm only predicts which learning resources students are interested in through the interaction data between students and learning resources, but the data is too little and has data sparsity. The present invention decomposes the characteristics of the learning resources learned by students and the characteristics of the learning resources, increases the interaction data between students and learning resources, and then uses the student-learning resource scoring matrix to predict which learning resources students are interested in. , student co-occurrence relationship matrix, cosine similarity matrix of the subject word features of the learning resources learned by students , the co-occurrence relationship matrix of learning resources, and the cosine similarity matrix of learning resource keyword features The additive low-rank sparse matrix decomposition model of MOOCs learning resources is constructed. The student latent factor matrix L and the learning resource latent factor matrix S are obtained by iterative optimization of the model. The student latent factor matrix L describes the common patterns and potential structures of students' interest in learning resources. The interests of students are highly correlated, so the rank of the student latent factor matrix is It cannot be too high and needs to satisfy the constraints of the additive low-rank sparse matrix decomposition model: that is, the rank of the student's latent factor matrix Less than or equal to the set constant , so the rank of the student latent factor matrix L needs to be Make judgments and classify calculations: If the rank of the student's potential factor matrix L is Greater than , singular value decomposition is used for low-rank approximation to obtain the low-rank matrix of student potential factors , if the rank of the student latent factor matrix L is Less than or equal to ,but ; The learning resource latent factor matrix S describes the learning resource characteristics and potential predictions of each student's preference. Therefore, the learning resource latent factor matrix S needs to be sparsely processed to meet the constraints of the additive low-rank sparse matrix decomposition model: that is, the number of non-zero elements of the learning resource latent factor matrix S, the balance matrix N, and the balance matrix M are less than or equal to the constant , and , so it is necessary to judge and calculate the learning resource potential factor matrix S, the balance matrix N, and the balance matrix M: If the number of non-zero elements in the learning resource potential factor matrix S is greater than , then by retaining the previous The largest matrix elements and the rest are set to 0 to perform matrix sparse approximation and obtain the sparse matrix of learning resource potential factors ; If the number of non-zero elements in the learning resource potential factor matrix S is less than or equal to ,but ; If the number of non-zero elements in the balanced matrix N is greater than , then by retaining the previous The largest matrix elements and the rest are set to 0 to perform matrix sparse approximation to obtain a sparse balanced matrix ; If the number of non-zero elements in the balanced matrix N is less than or equal to ,but ; If the number of non-zero elements in the equilibrium matrix M is greater than , then by retaining the previous The largest matrix elements and the rest are set to 0 to perform matrix sparse approximation to obtain a sparse balanced matrix ; If the number of non-zero elements in the balanced matrix M is less than or equal to ,but ; Finally, by learning the resource potential factor sparse matrix and the student latent factor low-rank matrix Add up to get the recommended score , for a given student ,in , retrieve the recommendation scores of the learning resources corresponding to a given student, sort the recommendation scores in descending order, and recommend the learning resources corresponding to the top K largest recommendation scores to the student, so as to accurately predict the learning resources that the student is interested in.

[0167] 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 preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A MOOCs learning resource recommendation method, characterized in that: The steps include: Step 1: Construct a student-learning resource scoring matrix for the MOOCs platform , assuming there is students , Learning Resources , the constructed student-learning resource scoring matrix is If the student Learned learning resources ,but ,otherwise ; Step 2: Construct the co-occurrence matrix of students and the cosine similarity matrix of the subject word features of the learning resources learned by students , the student co-occurrence relationship matrix includes the student co-learning resource number matrix , the matrix of the number of courses selected by students , and the matrix of the number of teachers shared by students ; Step 3: Construct the co-occurrence relationship matrix of learning resources and the cosine similarity matrix of learning resource keyword features The learning resource co-occurrence relationship matrix includes the learning resource total student number matrix , learning resource course co-occurrence matrix , and the learning resource teacher co-occurrence matrix ; Step 4: Based on the student-learning resource scoring matrix , student co-occurrence relationship matrix, cosine similarity matrix of the subject word features of the learning resources learned by students , the co-occurrence relationship matrix of learning resources, and the cosine similarity matrix of learning resource keyword features , construct an additive low-rank sparse matrix decomposition model for MOOCs learning resources: , in , , and Represents the student-learning resource scoring matrix Decomposing the error term , the student co-occurrence matrix and the cosine similarity matrix of the subject word features of the learning resources studied by the students and The loss error term , learning resource co-occurrence relationship matrix and cosine similarity matrix of learning resource keyword features and The loss error term and regular penalty to prevent overfitting The weight of , , L is the student latent factor matrix, which is used to represent the correlation between students' interest in learning resources, S is the learning resource latent factor matrix, which is used to represent the learning resource characteristics and potential predictions of each student's preference, T is the transposition operation, and the matrix and is a balanced matrix, and the constraint is that the rank of the student latent factor matrix L is less than or equal to a constant , the number of non-zero elements of the learning resource latent factor matrix S, the balance matrix N, and the balance matrix M are less than or equal to a constant , and , is the matrix Frobenius norm; Step 5, perform iterative optimization based on the additive low-rank sparse matrix decomposition model to solve the student latent factor matrix L and the learning resource latent factor matrix S of the additive low-rank sparse matrix decomposition model: Step 5-1, initialize matrix variables: L=X, S=0, M=0, N=0; Step 5-2, fix the learning resource latent factor matrix S, the balance matrix M and the balance matrix N, and calculate the update expression of the student latent factor matrix L: According to the additive low-rank sparse matrix decomposition model, the partial derivative of the student latent factor matrix L is equal to 0, and the expression of the student latent factor matrix L is calculated as: , in is the identity matrix; Step 5-3, fix the student latent factor matrix L, balance matrix M and balance matrix N, and calculate the updated expression of the learning resource latent factor matrix S: According to the additive low-rank sparse matrix decomposition model, the partial derivative of the learning resource potential factor matrix S is equal to 0, and the expression of the learning resource potential factor matrix S is calculated as follows: ; Step 5-4, fix the learning resource latent factor matrix S, the student latent factor matrix L and the balance matrix N, and calculate the update expression of the balance matrix M: According to the additive low-rank sparse matrix decomposition model, the partial derivative of the balanced matrix M is equal to 0, and the expression of the balanced matrix M is calculated as follows: ; Step 5-5, fix the learning resource latent factor matrix S, the student latent factor matrix L and the balance matrix M, and calculate the update expression of the balance matrix N: According to the additive low-rank sparse matrix decomposition model, the partial derivative of the balanced matrix N is equal to 0, and the expression for calculating the balanced matrix N is: ; Step 5-6, repeat iterative steps 5-2 to 5-5 until the number of iterations reaches the maximum or the additive low-rank sparse matrix decomposition model converges, and output the final learning resource latent factor matrix S and student latent factor matrix L; Step 6: Calculate the low-rank approximation of the student latent factor matrix L to obtain the student latent factor low-rank matrix , calculate the sparse approximation of the learning resource potential factor matrix S, and obtain the learning resource potential factor sparse matrix ; Step 7: Based on the student latent factor low-rank matrix and the learning resource latent factor sparse matrix Calculate the recommended scores of all learning resources for students : Recommended score is the low-rank matrix of student latent factors and the learning resource latent factor sparse matrix The sum of , for a given student ,in , retrieve the given student The recommendation scores of the corresponding learning resources are sorted in descending order, and the learning resources corresponding to the top K largest recommendation scores are recommended to the student.

2. A MOOCs learning resource recommendation method as claimed in claim 1, characterized in that: Step 2 specifically includes the following steps: Step 2-1: Construct a matrix of the number of students’ shared learning resources based on the shared learning situation of the learning resources that students have learned ; Assumption Student The resource set learned is ,student The resource set learned is ,statistics and Resource sets learned , construct the matrix of the number of common learning resources between students as follows: ,in ; Step 2-2: Construct the matrix of students’ co-selected courses based on the courses of the learning resources registered by the students Array ; Assumption Student The course set registered is ,student The course set registered is ,statistics and Common courses that everyone has registered for , construct the matrix of the number of courses selected by students as follows: ,in ; Step 2-3: Construct a matrix of the number of teachers shared by students based on the teachers selected by the students ; Assumption Student The set of all teachers is ,student The set of all teachers is ,statistics and Total teachers set , construct the matrix of the total number of teachers among students: ,in ; Step 2-4: Based on the content keywords of the learning resources learned by students, a cosine similarity matrix of the keyword characteristics of the learning resources learned by students reflecting their learning interests is constructed. ; Step 2-4-1, extract the keyword features of the learning resources learned by students based on the Transformer model: Assumption Student The subject word set of the learning resources is , For Students The number of subject words in the learning resources studied, students The subject word set of the learning resources is , For Students The number of keywords in the learning resources is learned. The Transformer model is used to learn and extract the features of each keyword. The learning resources The subject word features are , ,student The learning resources The subject word features are , ; Step 2-4-2, construct the cosine similarity matrix of the subject word features of the learning resources learned by students that reflects their learning interests: Connect with students Learning resources Keyword Features As input, it is processed through a fully connected layer and a normalization layer and then outputs the student The characteristics of the learning resources studied, i.e. ; Connect students Learning resources Keyword Features As input, it is processed through a fully connected layer and a normalization layer and then outputs the student The characteristics of the learning resources studied, i.e. , construct the cosine similarity matrix of the subject word features of the learning resources learned by students ,in .

3. A MOOCs learning resource recommendation method as claimed in claim 1, characterized in that: Step 3 specifically includes the following steps: Step 3-1: Construct a matrix of the total number of students studying the learning resources based on how the learning resources are studied by students. ; Assuming you have studied the learning resources The student set is , studied learning resources The student set is , then you have learned and The student set is , construct the total number of students in the learning resources matrix as: ,in ; Step 3-2: Construct a learning resource course co-occurrence matrix based on the course name to which the learning resource belongs ; Hypothesis Learning Resources The course name is , learning resources The course name is , then the co-occurrence matrix of learning resource courses is constructed as: ,in , if , then the matrix elements , otherwise ; Step 3-3, construct the learning resource teacher co-occurrence matrix based on the teacher who created the learning resource ; Hypothesis Learning Resources The teacher who created , learning resources The teacher who created , then the learning resource teacher co-occurrence matrix is ​​constructed as: ,in , if , then the matrix elements , otherwise ; Step 3-4: Construct a cosine similarity matrix of learning resource keyword features based on the content keywords of the learning resource ; Step 3-4-1, extract the topic features of resources based on Transformer model learning; Hypothesis Learning Resources The subject word set is , Resources for learning The number of keywords, learning resources The subject word set is , Resources for learning The number of keywords, using the Transformer model to learn and extract the features of each keyword, learning resources No. The subject word features are , , Learning Resources No. The subject word features are , ; Step 3-4-2, construct the cosine similarity matrix of learning resource keyword features: Connect to learning resources of Keyword Features As the input layer, it is processed through a fully connected layer and a normalization layer to output the learning resource The characteristics of , connect to learning resources of Keyword Features , processed through a fully connected layer and a normalization layer, outputting the learning resource The characteristics of , construct the cosine similarity matrix of learning resource theme word features ,in .

4. A MOOCs learning resource recommendation method as claimed in claim 1, characterized in that: Step 6 specifically includes the following steps: Step 6-1, calculate the low-rank approximation of the student latent factor matrix L: If the rank of the student latent factor matrix L is Greater than , using singular value decomposition for low-rank approximation: First, the singular value decomposition of the student latent factor matrix L is , where matrices A and B are unitary matrices, and the diagonal matrix And the size of the eigenvalue satisfies ; Secondly, keep the previous The rest are set to 0, that is, , through matrices A, B and Multiply them together to perform low-rank approximation and obtain the low-rank matrix of student potential factors ,Right now ; If the rank of the student latent factor matrix L is Less than or equal to ,but ; Step 6-2, calculate the sparse approximation of the learning resource latent factor matrix S: If the number of non-zero elements in the learning resource latent factor matrix S is greater than , then by retaining the previous The largest matrix elements and the rest are set to 0 to perform matrix sparse approximation and obtain the sparse matrix of learning resource potential factors ; If the number of non-zero elements in the learning resource latent factor matrix S is less than or equal to ,but ; If the number of non-zero elements in the balance matrix N is greater than , then by retaining the previous The largest matrix elements and the rest are set to 0 to perform matrix sparse approximation to obtain a sparse balanced matrix ; If the number of non-zero elements in the balanced matrix N is less than or equal to ,but ; If the number of non-zero elements in the balance matrix M is greater than , then by retaining the previous The largest matrix elements and the rest are set to 0 to perform matrix sparse approximation to obtain a sparse balanced matrix ; If the number of non-zero elements in the balance matrix M is less than or equal to ,but .