A personalized learning path recommendation method based on graph convolutional neural network

Through the personalized learning path recommendation method based on graph convolution neural network, the problems of information overload and knowledge trekking in online learning are solved, and the personalized learning path is recommended, which improves learning efficiency and quality and enhances learners' trust in the recommended results. .

CN118656406BActive Publication Date: 2024-11-12辽宁向日葵数字技术股份有限公司
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411126094.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-16
Publication Date
2024-11-12
Estimated Expiration
2044-08-16

AI Technical Summary

Technical Problem

Due to information overload and knowledge trekking in online learning, it is difficult for learners to effectively acquire and utilize knowledge, resulting in poor learning results.

Method used

The personalized learning path recommendation method based on graph convolutional neural network is adopted. By constructing a knowledge graph for learners, courses and videos, the graph neural network method is used to obtain learners' interest scores in courses, and personalized learning paths are recommended.

Benefits of technology

Through the knowledge graph of deep semantic relationships, we can deeply explore learners' preferences, improve the accuracy of recommendations, provide comprehensive personalized services, improve learning efficiency and quality, and enhance learners' trust in recommended results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118656406B_ABST
    Figure CN118656406B_ABST
Patent Text Reader

Abstract

The present invention discloses a personalized learning path recommendation method based on graph convolutional neural network, which belongs to the field of personalized learning recommendation and includes the following steps: step 1, obtaining learner, course, video, knowledge point information, recording all learners, courses, videos, knowledge points as an entity set, establishing the relationship between entities, and constructing a knowledge graph according to the relationship between entities; step 2, obtaining all entity representation vectors and relationship representation vectors; step 3, performing learner feature representation and course feature representation, obtaining learner representation vectors and course representation vectors, designing a second loss function for iterative updating, and obtaining each learner's interest score for each course; step 4, recommending personalized learning paths for learners who need personalized learning path recommendation. The present invention uses the graph neural network method to deeply explore learners' preferences and improve the accuracy of recommendations.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of personalized learning recommendation, and specifically relates to a personalized learning path recommendation method based on graph convolutional neural network. Background Art

[0002] With the development of Massive Open Online Courses (MOOC), people's learning is no longer restricted by location and time, which greatly facilitates the acquisition of knowledge. However, the explosive growth of data on online learning platforms inevitably brings about problems such as information overload and knowledge wandering. Information overload refers to the difficulty for learners to effectively distinguish, process and utilize explosive data when conducting online learning; knowledge wandering refers to the fact that when learners learn in a complex network environment, they are often attracted by some irrelevant questions while collecting information, and lose the original learning goal direction, thereby increasing the difficulty of learning for learners, causing learning difficulties and slowing down the learning process.

[0003] In response to the problems of information overload and knowledge wandering, some researchers have proposed to construct learner learning characteristics based on learning history or statistical methods to achieve learning path resource recommendation, but such methods cannot fully display the knowledge architecture of the entire learning path. To address this problem, the present invention designs a personalized learning path recommendation method based on graph convolutional neural network. This method models the learning behavior between users and courses and videos as a knowledge graph, and then uses graph convolution and graph neural network methods to obtain the user's preference score for the course based on the user's historical records of learning courses and watching course videos. Summary of the invention

[0004] In order to solve the problem of poor learning effect caused by lack of professional guidance in online learning, the present invention proposes a personalized learning path recommendation method based on graph convolutional neural network. The learner's course information is preliminarily obtained through the graph convolutional neural network, and then the learner's course video viewing history is used as auxiliary information. The graph neural network method is used to recommend reasonable and efficient learning paths for learners.

[0005] The technical solution of the present invention is as follows:

[0006] A personalized learning path recommendation method based on graph convolutional neural network includes the following steps:

[0007] Step 1: Obtain information about learners, courses, videos, and knowledge points, record all learners, courses, videos, and knowledge points as an entity set, establish relationships between entities, and construct a knowledge graph based on the relationships between entities;

[0008] Step 2: Obtain all entity representation vectors and relationship representation vectors;

[0009] Step 3: Perform learner feature representation and course feature representation to obtain learner representation vector and course representation vector, design a second loss function for iterative update, and obtain each learner's interest score for each course;

[0010] Step 4: Recommend personalized learning paths for learners who need personalized learning path recommendations.

[0011] Furthermore, in step 1, a knowledge graph is constructed based on information about learners, courses, videos, and knowledge points. ,in is a collection of entities, is a set of relations, It is a triple A collection of and are respectively called the head entity vector and the tail entity vector, yes and relationship;

[0012] The relationship set includes: the relationship between a course and the video corresponding to the course, the relationship between a learner and the course he / she studies; the relationship between a learner and the video he / she watches; the relationship between a course and the prerequisite course corresponding to the course; and the relationship between a course and the knowledge points corresponding to the course.

[0013] Furthermore, the specific process of step 2 is as follows:

[0014] Step 2.1, vectorize entities and relationships, and normalize each representation vector;

[0015] For entity collection Each entity in the dimensional representation vector; let The representation vector of an entity , represents the transpose of a vector, For the Entity dimensional entity components; the representation vector of each entity is normalized:

[0016] (1);

[0017] in, For the The normalized representation vector of each entity; The calculation is performed using formula (2). , , ;

[0018] (2);

[0019] in, is a given vector; , For different parameters; For vector No. Quantity; is the total number of components;

[0020] Pair Relationship Set Each relationship in dimensional representation vector; let The representation vector of a relationship , For the Relationship The relationship component of the dimension; the representation vector of each relationship is normalized:

[0021] (3);

[0022] in, For the The normalized representation vector of the relationship;

[0023] Assume the set of normal vectors , is the number of relations; The relationship normal vector corresponding to the relationship , For the The relationship normal vector corresponding to the relationship Dimensional components; normalize the relationship normal vector corresponding to each relationship:

[0024] (4);

[0025] in, For the The normalized relationship normal vector corresponding to each relationship;

[0026] Step 2.2: For triples , calculate and update the representation vectors of the head entity and the tail entity:

[0027] (5);

[0028] (6);

[0029] in, and are the representation vectors of the updated head entity and tail entity respectively; and are the representation vectors of the head entity and the tail entity before updating respectively; For relationship The relationship normal vector of

[0030] Step 2.3: Iteratively update the entity representation vector and the relationship representation vector to obtain all the final entity representation vectors and relationship representation vectors.

[0031] Furthermore, the specific process of step 2.3 is as follows:

[0032] Step 2.3.1: Initialize the first set is an empty set;

[0033] Step 2.3.2, set is the average number of tail entities of all head entities, is the average number of head entities of all tail entities; for each triple , pick a random number in (0,1) ,if , then randomly select a new head entity vector Replace the head entity vector , so that ,Will Add to first set In; if , then randomly select a new tail entity vector Replace tail entity vector , so that ,Will Add to first set middle;

[0034] Step 2.3.3: For each triple , calculate the score function, the formula is as follows:

[0035] (7);

[0036] in, For triples The scoring function of For relationship The representation vector of The calculation is performed using the method in formula (2). , Take 2 for both;

[0037] Step 2.3.4: Calculate the first loss function value :

[0038] (8);

[0039] in, is a boundary hyperparameter;

[0040] Step 2.3.5, The normalized representation vector of the entity Updated to:

[0041] (9);

[0042] in, For the updated The normalized representation vector of each entity; is the gradient descent step size;

[0043] Step 2.3.6, The normalized representation vector of the relationship Updated to:

[0044] (10);

[0045] in, For the updated The normalized representation vector of the relationship;

[0046] Step 2.3.7, The normalized relationship normal vector corresponding to the relationship Updated to:

[0047] (11);

[0048] in, For the updated The normalized relationship normal vector corresponding to each relationship;

[0049] Step 2.3.8: Update the Assign to , the updated Assign to , the updated Assign to , and repeat steps 2.3.3 to 2.3.7 until Converge and obtain all the final entity representation vectors and relationship representation vectors.

[0050] Furthermore, the specific process of step 3 is as follows:

[0051] Step 3.1: Build a course knowledge graph based on the relationships between courses ;in is a collection of course entities, is a set of relations between courses, is a course triple A collection of is the head course entity, is the last course entity, It is a course relationship. and They are and The representation vector of for The representation vector of

[0052] Step 3.2: Take a learner , computational learner The representation vector of

[0053] Step 3.3: Get learners A course studied , computing courses The representation vector of

[0054] Step 3.4: Repeat steps 3.2 to 3.3 to obtain all learner representation vectors and course representation vectors;

[0055] Step 3.5: Calculate the second loss function ;

[0056] Step 3.6: Repeat steps 3.4 to 3.5 until Converge,get each learner's interest score for each course.

[0057] Furthermore, the specific process of step 3.2 is as follows:

[0058] Step 3.2.1: Build learners based on their course history Courses studied :

[0059] (12);

[0060] Step 3.2.2: , is a predetermined constant, which is the total number of hops; calculate the set of triples:

[0061] (13);

[0062] (14);

[0063] in, , In the course knowledge graph, The course starts with Jump, Jump to the collection of courses associated with it; for The corresponding triple set represents the user About Courses Order neighbor aggregation;

[0064] Step 3.2.3: For each course triplet , calculate the weight coefficient and representation vector as follows:

[0065] (15);

[0066] (16);

[0067] in, For the course triple exist The weight coefficient of is a natural constant; For learners exist The representation vector in ; It is a course relationship The representation vector of Is the head course entity The representation vector of

[0068] Step 3.2.4: Construct cross-feature matrix :

[0069] (17);

[0070] in, and Respectively and Middle The value of the element;

[0071] Step 3.2.5: Calculate the updated representation vector and :

[0072] (18);

[0073] in, For updated learners exist The representation vector in ; For the updated head course entity The representation vector of , and are three weight matrices; is the bias vector;

[0074] Step 3.2.6: and Assign values ​​to and ;

[0075] Step 3.2.7. Calculate the learner representation vector :

[0076] (19);

[0077] in, For learners exist The representation vector in ; Indicates user About Courses Order neighbor aggregation.

[0078] Furthermore, the specific process of step 3.3 is as follows:

[0079] Step 3.3.1. Take In the course knowledge graph Several first-order neighbor courses in In the collection;

[0080] Step 3.3.2: For each first-order neighbor course , let it be The triplet , computational learner For every first-order neighbor course and course Relationship preference:

[0081] (20);

[0082] in, For learners right degree of preference; For first-order neighbor courses and courses relationship; It's a relationship The representation vector of

[0083] Step 3.3.3: For learners Every first-order neighbor course and course The preference degree of the relationship is normalized:

[0084] (twenty one);

[0085] in, For the normalized learner right degree of preference; Be a learner right degree of preference, for For any first-order neighbor course in for and relationship;

[0086] Step 3.3.4, use The graph convolutional neural network composed of convolutional layers in series updates the course representation. The representation vector As a graph convolutional neural network The input of the layer is recorded as ,right ,calculate:

[0087] (twenty two);

[0088] (twenty three);

[0089] in, , Respectively The input of the convolutional layer and the updated intermediate results; For the normalized learner right degree of preference; It is The output of the convolutional layer is also the The input of the convolutional layer; and are the weight matrix and bias vector respectively; is a nonlinear transformation function;

[0090] The output of the last convolutional layer is the final representation vector of the course .

[0091] Furthermore, the specific process of step 3.5 is as follows:

[0092] Step 3.5.1. For a learner , a course , calculate the first interest score and the second interest score :

[0093] (twenty four);

[0094] (25);

[0095] in, For learners The representation vector of For Courses The representation vector of

[0096] Step 3.5.2: Calculate the learner ,course The loss function :

[0097] (26);

[0098] in, is the activation function;

[0099] Step 3.5.3, repeat steps 3.5.1 to 3.5.2, after obtaining the loss functions of all learners and courses, sum up the loss functions of all learners and courses to obtain the second loss function ;

[0100] Step 3.5.4: Update using stochastic gradient descent method , , , , and .

[0101] Furthermore, the specific process of step 4 is as follows:

[0102] Step 4.1, set For learners who need personalized learning path recommendations, A collection of history records of videos watched , express time The representation vector of the course videos watched, , is the total number of moments; extract hidden information from the learner's video viewing history to obtain the video representation vector, and concatenate all the video representation vectors into an initial video history matrix ;

[0103] Step 4.2: Model the watched video representation vector as a directed weighted graph ; Vertex Set The vertices in Videos watched; is the edge set, , are two different vertices, express Watching Then watched the corresponding video Corresponding video;

[0104] Step 4.3: Construct the input edge matrix , Middle The row corresponds to Vertices , No. Column corresponds to Vertex, Row, No. The elements of the column correspond to The weight of , then Row, No. The element value of the column is 0; construct the outgoing edge matrix , among which Row, No. The elements of the column are The weight of , then Row, No. The element value of the column is 0;

[0105] Step 4.4, take a watched video representation vector and calculate the splicing matrix:

[0106] (27);

[0107] in, for Moment splicing matrix; is the concatenation function; and Corresponding to and No. OK; and are different parameter matrices, and are different bias vectors;

[0108] Step 4.5, calculate the final representation vector of the video:

[0109] (28);

[0110] (29);

[0111] (30);

[0112] (31);

[0113] in, for Always update the door; for Reset the door at all times; for Course information that is always hidden; for time Representation vector of the watched course videos; is the tanh activation function; It is the Hadamard product operator, which means the product of the elements at corresponding positions of two matrices; , , , , , They are different weight matrices respectively; for The final representation vector of the moment video;

[0114] Step 4.6: Repeat steps 4.4 to 4.5 to obtain the final representation vectors of all viewed videos; concatenate the final representation vectors of all viewed videos into an updated global video matrix ;

[0115] Step 4.7: Calculate the function value :

[0116] (32);

[0117] in, is the softmax activation function; , , are different weight matrices;

[0118] Step 4.8: Calculate again to get the global video matrix after the second update :

[0119] (33);

[0120] in, , are different weight matrices;

[0121] Step 4.9: Repeat steps 4.7 to 4.8 to calculate the final global video matrix ;

[0122] Step 4.10, Calculation Long-term preference representation vector for videos :

[0123] (34);

[0124] in, For custom weights; for Last video watched The representation vector of

[0125] Step 4.11: Prediction Click on the candidate video Score :

[0126] (35);

[0127] calculate Click on the candidate video The initial score :

[0128] (36);

[0129] in, for The initial representation vector of ;

[0130] Step 4.12, calculate the third loss function :

[0131] (37);

[0132] in, is the boundary hyperparameter; is a random number in the interval (0,1);

[0133] Step 4.13, use stochastic gradient descent to update each weight matrix and bias vector;

[0134] Step 4.14: Repeat steps 4.4 to 4.13 until convergence;

[0135] Step 4.15: For each course, take the average score of all videos covered as the auxiliary score of the course; The auxiliary score is recorded as ,Will Mapped to a value in the (0,1) interval :

[0136] (38);

[0137] Step 4.16, Calculation About Courses Preference score :

[0138] (39);

[0139] in, for About Courses interest score; For custom weights;

[0140] Step 4.17: Based on the course knowledge graph ,from Starting from the currently studied course, several learning paths are obtained along the directed edge direction; for each path, the average of the preference scores of all courses on the path is taken as the preference score of the path;

[0141] Step 4.18: Recommend the learning path with the highest preference score as the personalized learning path to .

[0142] The beneficial technical effects brought by the present invention: the knowledge graph provides learners with deep semantic relationships between courses, deeply explores learners' preferences, and improves the accuracy of recommendations. The present invention personalizes the learning path map according to learner characteristics, provides learners with more comprehensive personalized services, and makes the learning path recommendation results more suitable for the learner from the perspective of learners and learning content, ensuring the learner's learning efficiency and quality. At the same time, the knowledge graph displays the project relationship between learners and course history interactions, as well as the relationship between courses and videos, schools, and teachers, which improves learners' acceptance and strengthens learners' trust in the recommendation results. The method is based on a graph neural network and combines a knowledge-aware water wave graph convolutional network and a sequence recommendation model to perform personalized learning path recommendations. While considering learners' learning goals, knowledge levels, and the relationship between knowledge points, it also considers the smoothness of the learning path and the learner's participation, and recommends efficient and reasonable learning paths for learners to solve the problem of poor learning results due to lack of effective guidance in online education. BRIEF DESCRIPTION OF THE DRAWINGS

[0143] Figure 1 This is a flowchart of the personalized learning path recommendation method based on graph convolutional neural network of the present invention. DETAILED DESCRIPTION

[0144] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0145] like Figure 1 As shown, a personalized learning path recommendation method based on graph convolutional neural network includes the following steps:

[0146] Step 1: Record all learners, courses, videos, and knowledge points as an entity set, establish the relationship between entities, and build a knowledge graph based on the relationship between entities;

[0147] Obtain learner, course, video, and knowledge point information to build a data set. The knowledge graph constructed based on course knowledge and other information is ,in is a collection of entities, is a set of relations, It is a triple A collection of and are respectively called the head entity vector and the tail entity vector, yes and relationship;

[0148] The relationship set includes: the relationship between a course and the video corresponding to the course, the relationship between a learner and the course he / she studies; the relationship between a learner and the video he / she watches; the relationship between a course and the prerequisite course corresponding to the course; and the relationship between a course and the knowledge points corresponding to the course.

[0149] Step 2: Get all entity representation vectors and relationship representation vectors; the specific steps are as follows:

[0150] Step 2.1, vectorize entities and relationships, and normalize each representation vector;

[0151] For entity collection Each entity in the dimensional representation vector. Let The representation vector of an entity , , ; Among them, the superscript Represents the transpose of a vector; For the Entity The entity component of the dimension, whose value is A random number within is the number of entities. The representation vector of each entity is normalized:

[0152] (1);

[0153] in, For the The normalized representation vector of an entity; for a given vector , The calculation function is as follows:

[0154] (2);

[0155] in, , For different parameters; For vector No. Quantity; is the total number of components;

[0156] In formula (1), The calculation is performed using formula (2). , , .

[0157] Pair Relationship Set Each relationship in dimensional representation vector. Let The representation vector of a relationship , , , For the Relationship The relationship component of the dimension, whose value is A random number within is the number of relations. The representation vector of each relation is normalized:

[0158] (3);

[0159] in, For the The normalized representation vector of the relationship;

[0160] Assume the set of normal vectors , among which The relationship normal vector corresponding to the relationship , For the The relationship normal vector corresponding to the relationship The component of the dimension is A random number within is used to normalize the relationship normal vector corresponding to each relationship:

[0161] (4);

[0162] in, For the The normalized relationship normal vector corresponding to each relationship;

[0163] Step 2.2: For triples , calculate and update the representation vectors of the head entity and the tail entity:

[0164] (5);

[0165] (6);

[0166] in, and are the representation vectors of the updated head entity and tail entity respectively; and are the representation vectors of the head entity and the tail entity before updating respectively; For relationship The relationship normal vector.

[0167] Step 2.3, iteratively update the entity representation vector and the relationship representation vector to obtain all the final entity representation vectors and relationship representation vectors; the specific process is:

[0168] Step 2.3.1: Initialize the first set Is an empty set.

[0169] Step 2.3.2, set is the average number of tail entities of all head entities, is the average number of head entities of all tail entities. For each triple , pick a random number in (0,1) ,if , then randomly select a new head entity vector Replace the head entity vector , so that , Add to first set In; if , then randomly select a new tail entity vector Replace tail entity vector , so that ,Will Add to first set middle.

[0170] Step 2.3.3: For each triple , calculate the score function, the formula is as follows:

[0171] (7);

[0172] in, For triples The scoring function of For relationship The representation vector of The calculation is performed using the method in formula (2). , Take 2 for both.

[0173] Step 2.3.4: Calculate the first loss function value :

[0174] (8);

[0175] in, is the boundary hyperparameter, .

[0176] Step 2.3.5, The normalized representation vector of the entity Updated to:

[0177] (9);

[0178] in, For the updated The normalized representation vector of each entity; is the gradient descent step size.

[0179] Step 2.3.6, The normalized representation vector of the relationship Updated to:

[0180] (10);

[0181] in, For the updated The normalized representation vector of the relationship;

[0182] Step 2.3.7, The normalized relationship normal vector corresponding to the relationship Updated to:

[0183] (11);

[0184] in, For the updated The normalized relationship normal vector corresponding to each relationship;

[0185] Step 2.3.8: Update the Assign to , the updated Assign to , the updated Assign to , and repeat steps 2.3.3 to 2.3.7 until Converge and obtain all the final entity representation vectors and relationship representation vectors.

[0186] Step 3: Perform learner feature representation and course feature representation, obtain learner representation vector and course representation vector, design a second loss function for iterative update, and obtain each learner's interest score for each course; the specific steps are as follows:

[0187] Step 3.1: Build a course knowledge graph based on the relationships between courses .in is a collection of course entities, is a set of relations between courses, is a course triple A collection of is the head course entity, is the last course entity, It is a course relationship. and They are and The representation vector of for The representation vector of .

[0188] Step 3.2: Take a learner , computational learner The representation vector; the specific process is as follows:

[0189] Step 3.2.1: Build learners based on their course history Courses studied :

[0190] (12);

[0191] Step 3.2.2: , is a predetermined constant, which is the total number of hops; calculate the set of triples:

[0192] (13);

[0193] (14);

[0194] in, , In the course knowledge graph, The course starts with Jump, Jump to the collection of courses associated with it; for The corresponding triple set represents the user About Courses Order neighbor aggregation.

[0195] Step 3.2.3: For each course triplet , calculate the weight coefficient and representation vector as follows:

[0196] (15);

[0197] (16);

[0198] in, For the course triple exist The weight coefficient of is a natural constant; For learners exist The representation vector in ; It is a course relationship The representation vector of Is the head course entity The representation vector of

[0199] Step 3.2.4: For each , construct the cross feature matrix :

[0200] (17);

[0201] in, and Respectively and Middle The value of the element, . and Respectively and Middle The value of an element.

[0202] Step 3.2.5: Calculate the updated representation vector and :

[0203] (18);

[0204] in, For updated learners exist The representation vector in ; For the updated head course entity The representation vector of , and are three weight matrices; is the bias vector; , , , Each element of takes a random value between 0 and 1. Each update is made by stochastic gradient descent , , , .

[0205] Step 3.2.6: and Assign values ​​to and ;

[0206] Step 3.2.7. Calculate the learner representation vector :

[0207] (19);

[0208] in, For learners exist The representation vector in ; for The corresponding triple set represents the user About Courses Order neighbor aggregation; In the course knowledge graph, The course starts with Jump to the collection of courses associated with it;

[0209] Step 3.3: Get learners A course studied , computing courses The specific steps are as follows:

[0210] Step 3.3.1. Take In the course knowledge graph Several first-order neighbor courses in In the collection;

[0211] Step 3.3.2: For each first-order neighbor course , let it be The triplet , computational learner For every first-order neighbor course and course Relationship preference:

[0212] (20);

[0213] in, For learners right degree of preference; For first-order neighbor courses and courses relationship; It's a relationship The representation vector of .

[0214] Step 3.3.3: For learners Every first-order neighbor course and course The preference degree of the relationship is normalized:

[0215] (twenty one);

[0216] in, For the normalized learner right degree of preference, Be a learner right The preference degree is obtained from step 3.3.2, for For any first-order neighbor course in for and relationship;

[0217] Step 3.3.4, use The graph convolutional neural network composed of convolutional layers in series updates the course representation. The representation vector As a graph convolutional neural network The input of the layer is recorded as ,right ,calculate:

[0218] (twenty two);

[0219] (twenty three);

[0220] in, , Respectively The input of the convolutional layer and the updated intermediate results; For the normalized learner right degree of preference; It is The output of the convolutional layer is also the The input of the convolutional layer; and are the weight matrix and bias vector respectively; is a nonlinear transformation function. and The initial value of is a random number. The output of the last convolutional layer is the final representation vector of the course. .

[0221] Step 3.4: Repeat steps 3.2 and 3.3 to obtain all learner representation vectors and course representation vectors.

[0222] Step 3.5: Calculate the second loss function , the specific steps are as follows:

[0223] Step 3.5.1. For a learner , a course , calculate the first interest score and the second interest score :

[0224] (twenty four);

[0225] (25);

[0226] in, For learners The representation vector of For Courses The representation vector of .

[0227] Step 3.5.2: Calculate the learner ,course The loss function :

[0228] (26);

[0229] in, is the activation function, and the calculation formula is: , For any variable.

[0230] Step 3.5.3, repeat steps 3.5.1 to 3.5.2, after obtaining the loss functions of all learners and courses, sum up the loss functions of all learners and courses to obtain the second loss function .

[0231] Step 3.5.4: Update using stochastic gradient descent method , , , , and .

[0232] Step 3.6: Repeat steps 3.4 to 3.5 until Converge,get each learner's interest score for each course.

[0233] Step 4: Set For learners who need personalized learning path recommendations, Recommend a personalized learning path; the specific steps are as follows:

[0234] Step 4.1, take A collection of history records of videos watched , express time The representation vector of the course videos watched, , is the total number of moments. Extract hidden information from the learner's video viewing history to obtain a video representation vector, and concatenate all video representation vectors into an initial video history matrix .

[0235] Step 4.2: Model the watched video representation vector as a directed weighted graph ; Vertex Set The vertices in Videos watched; is the edge set, , are two different vertices, express Watching Then watched the corresponding video For any vertex, set the weights of all its incoming edges to the same value and sum to 1, and set the weights of all its outgoing edges to the same value and sum to 1.

[0236] Step 4.3: Construct the input edge matrix , Middle The row corresponds to Vertices , No. Column corresponds to Vertex, Row, No. The elements of the column correspond to The weight of , then Row, No. The element value of the column is 0; construct the outgoing edge matrix , among which Row, No. The elements of the column are The weight of , then Row, No. The element value of the column is 0.

[0237] Step 4.4, take a watched video representation vector and calculate the splicing matrix:

[0238] (27);

[0239] in, for Moment splicing matrix; It is a concatenation function that concatenates matrices. and Corresponding to and No. OK; and are different parameter matrices, and are different bias vectors. , , , The initial value is a random number.

[0240] Step 4.5, calculate the final representation vector of the video:

[0241] (28);

[0242] (29);

[0243] (30);

[0244] (31);

[0245] in, for Always update the door; for Reset the door at all times; for Course information that is always hidden; for time Representation vector of the watched course videos; is the tanh activation function, and the calculation formula is , is any variable. It is the Hadamard product operator, which means the product of the elements at corresponding positions of two matrices; , , , , , They are different weight matrices, and their initial values ​​are all random numbers. for The final representation vector of the video at moment .

[0246] Step 4.6: Repeat steps 4.4 to 4.5 to obtain the final representation vectors of all watched videos. The final representation vectors of all watched videos are concatenated into an updated global video matrix .

[0247] Step 4.7: Calculate the function value :

[0248] (32);

[0249] in, is the softmax activation function; , , are different weight matrices, and their initial values ​​are random numbers.

[0250] Step 4.8: Calculate again to get the global video matrix after the second update :

[0251] (33);

[0252] in, , are different weight matrices, and their initial values ​​are all random numbers.

[0253] Step 4.9: Repeat steps 4.7 to 4.8 to calculate the final global video matrix .

[0254] Step 4.10, Calculation Long-term preference representation vector for videos :

[0255] (34);

[0256] in, For custom weights; for Last video watched The representation vector of

[0257] Step 4.11: Prediction Click on the candidate video Score :

[0258] (35);

[0259] calculate Click on the candidate video The initial score :

[0260] (36);

[0261] in, for The initial representation vector of ;

[0262] Step 4.12, calculate the third loss function :

[0263] (37);

[0264] in, is the boundary hyperparameter, ; is a random number in the interval (0,1);

[0265] Step 4.13, use stochastic gradient descent to update each weight matrix and bias vector;

[0266] Step 4.14: Repeat steps 4.4 to 4.13 until convergence.

[0267] Step 4.15: For each course, take the average of the scores of all videos covered as the auxiliary score of the course. The auxiliary score is recorded as ,Will Mapped to a value in the (0,1) interval :

[0268] (38);

[0269] Step 4.16, Calculation About Courses Preference score :

[0270] (39);

[0271] in, for About Courses The interest score is obtained by step 3.6; For custom weights;

[0272] Step 4.17: Based on the course knowledge graph ,from Starting from the current course, several learning paths are obtained along the directed edge direction. For each path, the average of the preference scores of all courses on the path is taken as the preference score of the path;

[0273] Step 4.18: Recommend the learning path with the highest preference score as the personalized learning path to , the algorithm ends.

[0274] In order to verify the feasibility of the method proposed in the present invention and the effectiveness of each step, a comparative experiment was conducted between the method of the present invention and the existing algorithm. The F1 score was used as the evaluation index. The higher the F1 score, the better the algorithm performance. The calculation process of the F1 score is formula (40), formula (41) and formula (42):

[0275] (40);

[0276] (41);

[0277] (42);

[0278] in, Indicates the accuracy value; Indicates the recall value; Indicates the F1 score value; It means to a learner The number of courses that made correct referrals, To a learner The number of courses incorrectly recommended, Represents a learner The number of courses taken but not recommended, Represented as a real learner in the set of learners.

[0279] The experiment was conducted using a real dataset, MOOCCube, which is an open source large-scale data warehouse serving MOOC-related research. In the dataset, 2,196,521 learners were used for training, and the training set included 424 courses and 22,909 course videos; 681,729 learners were used for testing, and the test set included 142 courses and 7,637 course videos.

[0280] On the MOOCCube dataset, the comparison results of the present invention with the four algorithms of CKE, KGAT, RippleNet, and KGCN are shown in Table 1.

[0281] Table 1 Comparison results with other algorithms

[0282] .

[0283] Among them, the CKE algorithm combines collaborative filtering and knowledge graph embedding recommendation algorithms, which aims to improve the accuracy and interpretability of recommendations by utilizing structured information in knowledge graphs; the KGAT algorithm is a recommendation model based on graph attention network, which explicitly integrates triple information in knowledge graphs into the recommendation process, adopts graph attention mechanism to learn the representation of users and items, and can capture high-order connectivity in knowledge graphs; RippleNet is a recommendation algorithm based on graph convolutional neural network, which propagates user preferences by performing graph convolution on knowledge graphs, which can effectively simulate the diffusion process of user interests; the KGCN algorithm is a convolutional neural network recommendation model based on knowledge graphs, which learns the feature representation of entities and relationships by applying convolution operations on knowledge graphs, can capture local structural information between entities, and fuse this information through convolution layers, thereby improving the accuracy of recommendations. As can be seen from the data in Table 1, the method of the present invention has the highest F1 score and the best performance.

[0284] In order to verify the effectiveness of steps 2, 4 and 5, ablation experiments were conducted on the MOOCCube dataset. The results are shown in Table 2.

[0285] Table 2 Ablation experiment results

[0286] .

[0287] It can be seen from Table 2 that any step of the present invention is indispensable, otherwise it will affect the final personalized learning path recommendation effect.

[0288] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A personalized learning path recommendation method based on graph convolutional neural network, characterized in that: The steps include: Step 1: Obtain information about learners, courses, videos, and knowledge points, record all learners, courses, videos, and knowledge points as an entity set, establish relationships between entities, and construct a knowledge graph based on the relationships between entities; Step 2: Obtain all entity representation vectors and relationship representation vectors; Step 3: Perform learner feature representation and course feature representation to obtain learner representation vector and course representation vector, design a second loss function for iterative update, and obtain each learner's interest score for each course; Step 4: Recommend personalized learning paths for learners who need personalized learning path recommendations; In the step 1, a knowledge graph G = (E, R, S) is constructed based on the learner, course, video, and knowledge point information, where E is an entity set, R is a relationship set, S is a set of triples (h, r, t), h∈E and t∈E are respectively called the head entity vector and the tail entity vector, and r∈R is the relationship between h and t; The relationship set includes: the relationship between a course and the video corresponding to the course, the relationship between a learner and the course he / she studies, the relationship between a learner and the video he / she watches, the relationship between a course and the prerequisite course corresponding to the course, and the relationship between a course and the knowledge point corresponding to the course. The specific process of step 2 is: Step 2.1, vectorize entities and relationships, and normalize each representation vector; Set an m-dimensional representation vector for each entity in the entity set E; let the representation vector E of the i-th entity be i =(e i,1 ,e i,2 ,…,e i,j ,…,e i,m ) T , T represents the transpose of the vector, e i,j is the entity component of the jth dimension of the i-th entity; the representation vector of each entity is normalized: in, is the normalized representation vector of the i-th entity; The calculation is performed using formula (2), where a=1, b=2, and x=E. i ; Among them, x is a given vector; a and b are different parameters; is the first Quantity; is the total number of components; Set an m-dimensional representation vector for each relation in the relation set R; let The representation vector of a relationship For the The relational component of the j-th dimension of each relation; the representation vector of each relation is normalized: in, For the The normalized representation vector of the relationship; Assume the set of normal vectors |R| is the number of relations; The relationship normal vector corresponding to the relationship For the The component of the j-th dimension of the relationship normal vector corresponding to each relationship; the relationship normal vector corresponding to each relationship is normalized: in, For the The normalized relationship normal vector corresponding to each relationship; Step 2.2: For the triple (h, r, t), calculate and update the representation vectors of the head entity and the tail entity: in, and are the representation vectors of the updated head entity and tail entity respectively; E h and E t are the representation vectors of the head entity and the tail entity before updating respectively; is the relation normal vector of relation r; Step 2.3, iteratively update the entity representation vector and the relationship representation vector to obtain all the final entity representation vectors and relationship representation vectors; The specific process of step 2.3 is as follows: Step 2.3.1, initialize the first set S' to an empty set; Step 2.3.2, let tph be the average number of tail entities of all head entities, and hpt be the average number of head entities of all tail entities; for each triple (h, r, t) ∈ S, pick a random number α in (0, 1), if Then randomly select a new head entity vector h′∈E to replace the head entity vector h, so that Put (h′, r, t) into the first set S′; if Then randomly select a new tail entity vector t′∈E to replace the tail entity vector t, so that Put (h, r, t′) into the first set S′; Step 2.3.

3. For each triple (h, r, t)∈S∪S′, calculate the score function, the formula is as follows: Among them, f(h,r,t) is the score function of the triple (h,r,t); is the representation vector of relation r; The calculation is performed using the method in formula (2), and a and b are both 2; Step 2.3.4: Calculate the first loss function value Among them, β is the boundary hyperparameter; Step 2.3.5: Normalized representation vector of the i-th entity Updated to: Among them, E i ′ is the normalized representation vector of the updated i-th entity; γ is the gradient descent step size; Step 2.3.6, The normalized representation vector of the relationship Updated to: in, For the updated The normalized representation vector of the relationship; Step 2.3.7, The normalized relationship normal vector corresponding to the relationship Updated to: in, For the updated The normalized relationship normal vector corresponding to each relationship; Step 2.3.8: Update the E i 'Assign to The updated Assign to The updated Assign to Repeat steps 2.3.3 to 2.3.7 until Converge, and finally obtain all entity representation vectors and relationship representation vectors; The specific process of step 3 is as follows: Step 3.1: Construct the course knowledge graph G based on the relationship between courses course =(C,R c ,S c );in is a collection of course entities, is a set of relations between courses, is a course triplet (h c ,r c ,t c ), h c ∈C is the head course entity, t c ∈C is the tail course entity, r c ∈R c It is a course relationship. and h c and t c The representation vector of For r c The representation vector of Step 3.2, take a learner u and calculate the representation vector of learner u; Step 3.3, take a course c that learner u has studied, and calculate the representation vector of course c; Step 3.4: Repeat steps 3.2 to 3.3 to obtain all learner representation vectors and course representation vectors; Step 3.5: Calculate the second loss function Step 3.6: Repeat steps 3.4 to 3.5 until Convergence, obtain each learner's interest score for each course; The specific process of step 3.2 is as follows: Step 3.2.1: Construct a set of courses that learner u has studied based on the learner’s course history Step 3.2.2, for k = 1, 2, ..., K, K is a predetermined constant, which is the total number of hops; calculate the set of triples: in, In the course knowledge graph, Starting from the course in, it is connected to the set of courses through k-1 hops and k hops respectively; for The corresponding triple set represents the aggregation of user u’s k-order neighbors for the course; Step 3.2.3: For each course triplet Calculate the weight coefficients and representation vectors as follows: Among them, p(h c ,r c ,t c ) is the course triplet (h c ,r c ,t c )exist The weight coefficient of ; e is a natural constant; For learners The representation vector in ; Is the course relationship c The representation vector of Is the head course entity h c The representation vector of Step 3.2.4, construct the cross feature matrix J: in, and Respectively and The value of the Mth element in ; Step 3.2.5: Calculate the updated representation vector and in, For updated learners u The representation vector in ; For the updated head course entity h c The representation vector of; ω1, ω2 and ω3 are three weight matrices; b c is the bias vector; Step 3.2.6: and Assign values ​​to and Step 3.2.

7. Calculate the learner representation vector in, For learners The representation vector in ; Represents the aggregation of user u’s K-order neighbors on a course; The specific process of step 3.3 is as follows: Step 3.3.

1. Take c in the course knowledge graph G course Put several first-order neighbor courses in N c In the collection; Step 3.3.2: For each first-order neighbor class c′∈N c , let it and c form a triple (c,r c,c′ ,c′)∈S c , calculate the preference of learner u for each first-order neighbor course and course c: in, For learners u to r c,c′ degree of preference; c,c′ is the relationship between the first-order neighbor course c′ and course c; It is a relationship c,c′ The representation vector of Step 3.3.3: Normalize the preference degree of each first-order neighbor course of learner u and the relationship between course c: in, is the normalized learner u versus r c,c′ degree of preference; Is the learner u right degree of preference, N c For any first-order neighbor course in for Relationship with c; Step 3.3.4: Update the course representation using a graph convolutional neural network consisting of L convolutional layers in series, transforming the representation vector E of course c into c As the input of the l=1th layer of the graph convolutional neural network, and recorded as For l = 1, 2, ..., L, calculate: in, They are the input of the l-1th convolutional layer and the updated intermediate result respectively; is the normalized learner u pair degree of preference; is the output of the l-1th convolutional layer and the input of the lth convolutional layer; W and b c ′ are weight matrix and bias vector respectively; ReLU(·) is nonlinear transformation function; The output of the last convolutional layer is the final representation vector of the course The specific process of step 3.5 is as follows: Step 3.5.

1. For a learner u and a course c, calculate the first interest score y u,c and the second interest score Among them, E u is the representation vector of learner u, E c is the representation vector of course c; Step 3.5.

2. Calculate the loss function of learner u and course c Among them, σ(·) is the activation function; Step 3.5.3: Repeat steps 3.5.1 to 3.5.2 to obtain the loss functions of all learners and courses. Then sum the loss functions of all learners and courses to obtain the second loss function. Step 3.5.4: Update ω1, ω2, ω3, b using stochastic gradient descent method c , W and b c ′.

2. The personalized learning path recommendation method based on graph convolutional neural network according to claim 1 is characterized in that: The specific process of step 4 is as follows: Step 4.1, set For learners who need personalized learning path recommendations, A collection of history records of videos watched express time The representation vector of the course videos watched, is the total number of moments; extract the hidden information of the video from the learner’s video viewing history to obtain the video representation vector, and concatenate all the video representation vectors into an initial video history matrix M Q ; Step 4.2: Model the watched video representation vector as a directed weighted graph G Q =(V Q ,E Q ); Vertex Set The vertices in Videos watched; is the edge set, are two different vertices, express Watching Then watched the corresponding video Corresponding video; Step 4.3: Construct the input edge matrix M I , M I Middle The row corresponds to Vertices No. Column corresponds to Vertex, Row, No. The elements of the column correspond to The weight of The first Row, No. The element value of the column is 0; construct the outgoing edge matrix M O , among which Row, No. The elements of the column are The weight of The first Row, No. The element value of the column is 0; Step 4.4, take a watched video representation vector and calculate the splicing matrix: in, for The moment concatenation matrix; contact(·) is the concatenation function; and Corresponding to M I and M O No. OK; and are different parameter matrices, b I and b O are different bias vectors; Step 4.5, calculate the final representation vector of the video: in, for Always update the door; for Reset the door at all times; for Course information that is always hidden; for time The representation vector of the watched course video; tanh(·) is the tanh activation function; ⊙ is the Hadamard product operator, which means the product of the corresponding position elements of two matrices; W z , W r , W h , P z , P r , P h They are different weight matrices respectively; for The final representation vector of the moment video; Step 4.6: Repeat steps 4.4 to 4.5 to obtain the final representation vectors of all viewed videos; concatenate the final representation vectors of all viewed videos into an updated global video matrix M Q ′; Step 4.7, calculate the function value F: Among them, softmax(·) is the softmax activation function; W Q , W K , W V are different weight matrices; Step 4.8: Calculate again to obtain the global video matrix M after the second update Q ″: M Q ″=ReLU(FW1)W2+F (33); Among them, W1 and W2 are different weight matrices; Step 4.9: Repeat steps 4.7 to 4.8 to calculate the final global video matrix Step 4.10, Calculation Long-term preference representation vector Seq for videos f : Among them, ω4 is a custom weight; for Last video watched The representation vector of Step 4.11: Prediction Click on the candidate video q c Score calculate Click on the candidate video q c The initial score in, for The initial representation vector of ; Step 4.12, calculate the third loss function Among them, λ is the boundary hyperparameter; θ is a random number in the interval (0,1); Step 4.13, use stochastic gradient descent to update each weight matrix and bias vector; Step 4.14: Repeat steps 4.4 to 4.13 until convergence; Step 4.15: For each course, take the average of the scores of all the videos it covers as the auxiliary score of the course; the auxiliary score of course c is recorded as Will Mapped to a value in the (0,1) interval Step 4.16, Calculation Preference score for course c in, for The interest score for course c; ρ is the custom weight; Step 4.17: Based on the course knowledge graph G course ,from Starting from the currently studied course, several learning paths are obtained along the directed edge direction; for each path, the average of the preference scores of all courses on the path is taken as the preference score of the path; Step 4.18: Recommend the learning path with the highest preference score as the personalized learning path to