Personalized education recommendation method and system, medium and equipment

Abstract

The invention discloses a personalized education recommendation method, system, medium and device, the system comprises a data acquisition unit, a knowledge graph construction unit, a propagation unit, a course scoring unit and a personalized education recommendation scheme generation unit, the data acquisition unit is in communication connection with the knowledge graph construction unit; the knowledge graph construction unit is in communication connection with the propagation unit, the propagation unit is in communication connection with the course scoring unit, and the course scoring unit is in communication connection with the personalized education recommendation scheme generation unit. According to the method and the system, knowledge graph information such as first modification and second modification among courses is fused, and the influence of social relations of students on course selection is introduced, so that the relevance between recommendation results and course targets is clearer, and meanwhile, the recommendation results are more personalized and better fit the characteristics and requirements of the students.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of education. In particular, it is a personalized education recommendation method, system, medium and device. BACKGROUND

[0002] In the field of education recommendation system, with the popularity of online courses, MOOCs (Massive Open Online Courses) and smart education platforms, "recommended courses" have become an important research direction to solve the burden of course selection for students and improve the personalization of learning paths. Traditional recommendation systems are mainly based on the "student-course" binary interaction relationship, and on this basis, enhanced models integrating semantics, graph structure, social relationships, etc. are continuously developed.

[0003] The existing course recommendation scheme improves the semantic level and recommendation accuracy to some extent, but still has the following main defects:

[0004] 1. Ignoring the prerequisite / successor structure between courses

[0005] Although some schemes construct a course-knowledge point or course-theme graph, many systems do not explicitly include the "prerequisite course" / "successor course" knowledge path structure as an edge into the model. This leads to a recommendation that is rich in semantics, but difficult to reflect the logic of "knowledge progression path", such as the prerequisite relationship that students should first take A and then take B;

[0006] 2. Ignoring the influence of student social relationships on course selection

[0007] Existing schemes focus on user-course interaction and course semantic relationship, but often lack modeling of social relationships between student nodes, such as friends, study groups, and learning interaction frequency. Social influence can lead to group characteristics and peer effects in student course selection behavior, and ignoring this can result in serious homogenization and lack of personalization in the recommendation results. Literature also points out that: "... most online education recommendation research has not deeply considered the social network or group influence of learners. SUMMARY

[0008] To this end, the technical problem to be solved by the present application is to provide a personalized education recommendation method, system, medium and device, which integrates knowledge graph information such as prerequisite and successor between courses and introduces the influence of student social relationships on course selection, making the relevance between the recommendation results and the course goals clearer, and making the recommendation results more personalized and more suitable for the students' own characteristics and needs.

[0009] To solve the above technical problems, the present application provides the following technical solutions:

[0010] A personalized education recommendation system, comprising:

[0011] The data acquisition unit is used to collect course nodes. Course data and student nodes Its own data and social data; among them, course nodes The course data includes the knowledge domain category to which the course belongs, the course difficulty level, the number of students who have enrolled in the course in the past, and the course milestones. The number of subsequent courses and prerequisite courses; the course difficulty level indicates the ease or difficulty of passing the course; the number of subsequent courses represents the course milestones. The number of courses that can be taken after passing the courses represented, and the number of prerequisite courses is the elective course node. The number of courses required before the represented course, student node Our proprietary data includes statistics on completed course categories, historical grade distribution, learning objectives and interest tags, and student nodes. Social data includes the number of friends and the number of interactions;

[0012] The knowledge graph construction unit is used to generate knowledge graphs using the data collected by the data acquisition unit.

[0013] The propagation unit is used to run a graph neural network model on the knowledge graph constructed by the knowledge graph construction unit. The propagation unit includes a first adaptive gate control module and a second adaptive gate control module. The first adaptive gate control module is used to adjust the path based on the course nodes. Adjust course level and course nodes Propagation weight, course degree is a metric related to the number of students enrolled in the course, the number of subsequent courses, and the number of prerequisite courses; the second adaptive gate control module is used to determine the student node... With student nodes The social interaction weight between course nodes is adjusted based on their correlation with student nodes. Number of friends, frequency of interaction, and interactions with students The similarity of course selection paths and the metrics related to social influence show that course nodes with a higher correlation to social interaction are those with student nodes. The nodes to which the courses selected by students with high social engagement belong;

[0014] The course scoring unit is used to calculate the matching score between student nodes and course nodes based on the final representation of student nodes and the final representation of course nodes output by the propagation unit, and to sort all courses according to the matching score.

[0015] The personalized education recommendation scheme generation unit is used to generate personalized learning paths based on the sorting structure of the course scoring unit, the prerequisite / successor relationships between courses, and the knowledge graph constructed by the knowledge graph construction unit.

[0016] The data acquisition unit is in communication connection with the knowledge graph construction unit, the knowledge graph construction unit is in communication connection with the propagation unit, the propagation unit is in communication connection with the course scoring unit, and the course scoring unit is in communication connection with the personalized education recommendation scheme generation unit.

[0017] The personalized education recommendation system further comprises an annotation unit for explaining the personalized learning path, and the annotation unit is in communication connection with the data acquisition unit, the course scoring unit and the personalized education recommendation scheme generation unit respectively.

[0018] A method for personalized education recommendation by using the personalized education recommendation system comprises the following steps:

[0019] S1) collecting course data, student self data and student social data;

[0020] S2) the knowledge graph construction unit constructs a course-student-social knowledge graph by using the course data, the student self data and the student social data collected in step S1), and sets a course node set V (c) and a student node set V (s) , sets a prerequisite / successor relationship edge E (cc) between the course nodes for representing the knowledge dependency path between courses, establishes a student-course edge E (sc) between the student nodes and the course nodes for representing the elective / finished / planned to finish relationship, establishes a social relationship edge E (ss) between the student nodes for representing the social influence between students, defines an initial feature vector for each course node, each student node, each social relationship edge E (ss) and each student-course edge E (sc) .

[0021] S3) setting a graph neural network model for relationship modeling of the student nodes and the course nodes on the course-student-social knowledge graph constructed in step S2), and generating vector representations of the student nodes and the course nodes in a low-dimensional space by using the graph neural network model;

[0022] S4) performing L rounds of propagation on the student nodes and the course nodes by using the graph neural network model, and then calculating a matching score of the student nodes and the course nodes, wherein L is an integer greater than or equal to 2;

[0023] S5) sorting the courses by using the matching score, and selecting the top N courses as the recommendation result;

[0024] S6) generating a personalized learning path along the path in the course-student-social knowledge graph from the course recommended in step S5) E (cc) .

[0025] The method, in step S4), when propagating the student nodes and the course nodes by using the graph neural network model, in each round of propagation, adjusts the propagation weight of the course nodes by using the first adaptive gate control module The propagation weight, and adjusts the propagation weight of the course nodes with a higher association with the social degree by using the second adaptive gate control module, so as to automatically adjust the propagation intensity of different courses.

[0026] The method, the first adaptive gate control module calculates the propagation weight of the course nodes by the following formula:

[0027]

[0028] In the formula, denotes the degree or normalized connection strength of the course node ; denotes all course neighbors of the student node .

[0029] The method, the second adaptive gate control module calculates the propagation weight of the course nodes with a higher association with the social degree by the following formula:

[0030]

[0031] In the formula, denotes the social feature score of the student node and the student node ; denotes the social neighbor set of the student node .

[0032] The method, in step S4), when propagating the student nodes and the course nodes by using the graph neural network embedding propagation module, the propagation formula used is:

[0033]

[0034] In the formula, denotes the initial representation of the student node when propagating in the first layer; denotes a nonlinear activation function; and both denote the trainable parameters in the first layer; denotes the relationship weight of the student node in the first layer to the student node .

[0035] ​​​The above method, in step S4), calculates the matching score by the following formula:

[0036]

[0037] In the formula, f() represents an inner product, a bilinear function, or a probability or score output by a multi-layer perception; represents the final representation of the student node . represents the final representation of the course node .

[0038] A computer readable storage medium having stored thereon a computer program, which, when executed by a processor, implements the above method.

[0039] A computer device comprising a readable storage medium, a processor, and a computer program stored on the readable storage medium and executable on the processor, which, when executed by the processor, implements the above method.

[0040] The technical solution of the present application achieves the following beneficial technical effects:

[0041] 1. The output of the present application is not only a discrete course list, but also a structured learning sequence. The system automatically plans a feasible path for the student from the current state to the learning goal using the inherent inter-course relationship (prerequisite / successor) in the constructed knowledge graph.

[0042] 2. The present application can generate a complex explanation such as "recommend you to learn course B because you have mastered its prerequisite course A, and your learning friends also highly evaluate this course", which greatly enhances the transparency and credibility of the recommendation.

[0043] 3. The present application introduces an "adaptive gating mechanism" in the propagation layer to automatically generate course-student relationship feature weights and student social feature weights, thereby improving the flexibility and accuracy of the model. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 is a working principle diagram of the personalized education recommendation system of the present application;

[0045] Figure 2 is a flowchart of personalized education recommendation in the present application;

[0046] Figure 3 is a principle diagram of a computer device that can perform personalized education recommendation. DETAILED DESCRIPTION

[0047] The present application will be further described below in conjunction with examples.

[0048] As Figure 1As shown, the personalized education recommendation system in the application includes a data acquisition unit, a knowledge graph construction unit, a propagation unit, a course scoring unit, a personalized education recommendation scheme generation unit and an annotation unit, the data acquisition unit is in communication connection with the knowledge graph construction unit, the knowledge graph construction unit is in communication connection with the propagation unit, the propagation unit is in communication connection with the course scoring unit, the course scoring unit is in communication connection with the personalized education recommendation scheme generation unit, and the annotation unit is in communication connection with the data acquisition unit, the course scoring unit and the personalized education recommendation scheme generation unit respectively.

[0049] In the application, the data acquisition unit is used to acquire course data, student self-owned data and student social data, and the above three kinds of data are directly or indirectly related to the course selection of students, that is, they directly or indirectly affect the decision of the students to select courses. Among them, the course data of a course Course includes the knowledge field category to which the course Course belongs, the difficulty level of the course Course, the historical number of students selecting the course Course, and the number of successor courses and prerequisite courses of the course Course, the successor course is a course that can be taken after the course Course, and the prerequisite course is a course that should be taken before the course Course. The student self-owned data includes student completed course category statistics, historical score distribution, learning goal and interest label. The student social data includes the number of friends and the number of activities of the student and his friends.

[0050] The knowledge graph construction unit is used to generate a knowledge graph by using the data acquired by the data acquisition unit. In the application, the knowledge graph constructed by the knowledge graph construction unit is a course-student-social knowledge graph, which has student nodes, course nodes and social relationship edges. In the course-student-social knowledge graph, the "course (Course)" node V (c) and the "student (Student)" node set V (s) are set, and the "prerequisite / succession (Prerequisite / Succession)" relationship edge E (cc) for representing the knowledge dependency path between courses is set between the course nodes, the student-course edge E (sc) for reflecting the interaction or planning state of the student and the course is established between the student nodes and the course nodes, and the social relationship edge E (ss) for representing the social influence between students (such as friend relationship, learning group relationship and interaction frequency relationship, etc.) is established between the student nodes. Based on the above setting, in the application, the course-student-social knowledge graph can be represented as G=(V,E), wherein V=V (s) ∪V (c) , E=E (sc) ∪E(cc) ∪E (ss) In constructing the course-student-social knowledge graph, initial feature vectors can be defined for each course node, each student node, and each edge. The initial feature vectors for course nodes include the knowledge domain category to which the course belongs, the difficulty level of the course, the number of students who have previously enrolled in the course, and the number of subsequent courses and prerequisite courses. The initial features for student nodes include statistics on the types of courses the student has taken, the distribution of historical grades, learning objectives, and interest tags. Social relationship edges and student-course edges are accompanied by weights or interaction features (such as the number of interactions, course selection frequency, and peer success rate).

[0051] The propagation unit is used to run a graph neural network model on the knowledge graph constructed by the knowledge graph construction unit. The graph neural network model is used to model the relationship between student nodes and course nodes, and generate vector representations of student nodes and course nodes in a low-dimensional space. The propagation unit includes a first adaptive gate control module and a second adaptive gate control module. The first adaptive gate control module is used to adjust the parameters based on the course nodes. Adjust course level and course nodes Propagation weight, course degree is a metric related to the number of students enrolled in the course, the number of subsequent courses, and the number of prerequisite courses; the second adaptive gate control module is used to determine the student node... With student nodes The social interaction weight between course nodes is adjusted based on their correlation with student nodes. Number of friends, frequency of interaction, and interactions with students The similarity of course selection paths and the metrics related to social influence show that course nodes with a higher correlation to social interaction are those with student nodes. The nodes to which the courses selected by students with high social engagement belong.

[0052] In this invention, the first adaptive gate control module, based on the course nodes... Adjust course level and course nodes When propagating weights, for student nodes With course nodes The edge between (belonging to E) (sc) The generation of the gating weights (\alpha_{i,k}^{(l)}) is based on the course nodes. The "degree" or connection strength of a course (e.g., how many students have enrolled in it, how many are its subsequent courses, how many are its prerequisites, etc.). That is, nodes with higher course degrees (more connections, wider influence) have correspondingly higher course propagation weights, thus reflecting the "key" or "core" position of the course in the knowledge graph. Specifically, in this invention, course nodes can be calculated using the following formula. Propagation weight:

[0053]

[0054] In the formula, Represented as course nodes The degree or normalized connection strength; Represented as student node All courses in the neighborhood.

[0055] The second adaptive gate control module, based on the student node... With student nodes When adjusting the propagation weight of course nodes with high social correlation, for student nodes... With its social neighbors The edge between (belonging to E) (ss) The weights (\alpha_{i,j}^{(l)}) are generated based on students' social characteristics (such as social degree, interaction frequency, and relationship with students). (Course selection path similarity, social influence indicators, etc.). Specifically, in this invention, the propagation weight of course nodes with high social correlation can be calculated using the following formula:

[0056]

[0057] In the formula, Represented as student node and student nodes Social characteristic scores; Represented as student node A collection of social neighbors.

[0058] In this invention, utilizing For student nodes The relevant social relationships are ranked, and those ranked higher are compared with those ranked lower. Relevant students (non-students) The selected courses are marked as those with a high degree of social relevance, and the dissemination intensity of the course nodes representing the courses is increased during the dissemination process. This reflects the impact of students' social relationships on course selection and is conducive to students forming mutual learning assistance during the course selection process.

[0059] The first and second adaptive gate control modules enable the propagation intensity of different courses to be automatically adjusted according to their structural characteristics, thereby allowing the graph neural network model to give higher weight to "popular or core courses" and "key social peer influence", thus improving personalization and interpretability.

[0060] A course scoring unit is configured to calculate a matching score of the student node and the course node according to the final representation of the student node and the final representation of the course node output by the propagation unit, and sort all courses according to the matching score. Specifically, the matching score is calculated by the following formula:

[0061]

[0062] wherein f() represents an inner product, a bilinear function or a probability or score output by a multi-layer perception; represents the final representation of the student node ; represents the final representation of the course node .

[0063] When the course scoring unit gives the matching score of the student node and the course node, a personalized education recommendation scheme generation unit generates a personalized learning path according to the sorting structure of the course scoring unit, the prerequisite / successor relationship between courses and the knowledge graph constructed by the knowledge graph construction unit.

[0064] An annotation unit is configured to annotate the personalized learning path. Specifically, the annotation unit provides explanation information for the personalized learning path generated by the personalized education recommendation scheme generation unit, and explains the reason for the recommendation, for example: “Recommend course A because the course is a high-degree core course in the knowledge graph, and you have already taken its prerequisite course B; in addition, your social friend X has chosen A, and the influence enhances the recommendation adaptation degree”. In this way, the transparency and student trust of the recommendation result are improved.

[0065] As shown in Figure 2 , when the personalized education recommendation system in the application is used to perform personalized education recommendation for students, the personalized education recommendation is completed through the following steps:

[0066] S1) Collecting course data, student self data and student social data; wherein the course data can be provided by the school, and the student self data and the student social data are provided by the student and input into the personalized education recommendation system through different ports;

[0067] S2) The knowledge graph construction unit constructs a course-student-social knowledge graph using the course data, student self data and student social data collected in step S1), and sets a course node set V (c) and a student node set V (s) , sets a prerequisite / successor relationship edge E (cc) representing the knowledge dependency path between courses between the course nodes, and establishes a student-course edge E (sc)establishing a social relationship edge E representing social influence between students between student nodes (ss) for each course node, each student node, each social relationship edge E (ss) and each student-course edge E (sc) defining an initial feature vector;

[0068] S3) setting a graph neural network model for modeling the relationship between student nodes and course nodes on the course-student-social knowledge graph constructed in step S2), and generating vector representations of student nodes and course nodes in a low-dimensional space using the graph neural network model;

[0069] S4) propagating student nodes and course nodes using the graph neural network model for L rounds, and then calculating matching scores of student nodes and course nodes, wherein L is an integer greater than or equal to 2, preferably 3 or 4, to ensure that the node representation can integrate the information of its direct neighbors and two-hop neighbors. However, as the number of rounds increases, the propagation range that the node can receive expands, but in order to avoid over-smoothing or information redundancy, L should not be set too large;

[0070] S5) ranking courses using the matching scores, and selecting the top N courses as the recommendation results;

[0071] S6) using the prerequisite / successor relationship edge E (cc) from the courses recommended in step S5) to generate a personalized learning path along the path in the course-student-social knowledge graph.

[0072] In step S4), when propagating student nodes and course nodes using the graph neural network embedding propagation module, the propagation formula used is:

[0073]

[0074] wherein, represents the initial representation of the student node in the first layer propagation; represents a nonlinear activation function; and both represent trainable parameters of the first layer; represents the relationship weight of the student node in the first layer to the student node .

[0075] In step S6), when generating the personalized learning path, a graph traversal algorithm (such as depth-first, breadth-first, limit path length, weight threshold screening, etc.) can be used to ensure the rationality and executability of the personalized learning path.

[0076] In the personalized education recommendation, the personalized education recommendation system in the application constructs an integrated heterogeneous knowledge graph, a course-student-social knowledge graph, and fuses three types of information, i.e., a student-course interaction relationship, a course-course knowledge structure relationship and a student-student social relationship, which are traditionally independent, in a unified graph data structure. This breaks the limitation of traditional recommendation systems that only focus on user-item binary relationship, and lays a foundation for modeling a student learning environment comprehensively and stereoscopically. Specifically, the application explicitly models a core relationship edge between courses as a prerequisite / successor, thereby introducing an internal logical path of knowledge into the system.

[0077] A computer readable storage medium, having stored thereon a computer program, the computer program being executed by a processor to implement the personalized education recommendation method in the application.

[0078] As shown in Figure 3 A computer device, including a readable storage medium, a processor and a computer program stored on the readable storage medium and executable on the processor, the computer program being executed by the processor to implement the personalized education recommendation method in the application.

[0079] Obviously, the above embodiments are only examples for clearly illustrating but not limiting the implementation. Based on the above description, other different forms of changes or variations can be made by those skilled in the art. Here, all the implementation is not required or can not be exhausted. The obvious changes or variations derived therefrom are still within the protection scope of the claims of the present application.

Claims

1. A personalized education recommendation system, characterized in that, include: The data acquisition unit is used to collect course node data. Course data and student nodes Its own data and social data; among them, course nodes The course data includes the knowledge domain category to which the course belongs, the course difficulty level, the number of students who have enrolled in the course in the past, and the course milestones. The number of subsequent courses and prerequisite courses; the course difficulty level indicates the ease or difficulty of passing the course; the number of subsequent courses represents the course milestones. The number of courses that can be taken after passing the courses represented, and the number of prerequisite courses is the elective course node. The number of courses required before the represented course, student node The proprietary data includes statistics on the types of courses taken, historical grade distribution, learning objectives and interest tags, and student nodes. Social data includes the number of friends and the number of interactions; The knowledge graph construction unit is used to generate knowledge graphs using the data collected by the data acquisition unit. The propagation unit is used to run a graph neural network model on the knowledge graph constructed by the knowledge graph construction unit. The propagation unit includes a first adaptive gate control module and a second adaptive gate control module. The first adaptive gate control module is used to adjust the path based on the course nodes. Adjust course level and course nodes Propagation weight, course degree is a metric related to the number of students enrolled in the course, the number of subsequent courses, and the number of prerequisite courses; the second adaptive gate control module is used to determine the student node... With student nodes The social interaction weight between course nodes is adjusted based on their correlation with student nodes. Number of friends, frequency of interaction, and interactions with students The similarity of course selection paths and the metrics related to social influence show that course nodes with a higher correlation to social interaction are those with student nodes. The nodes to which the courses selected by students with high social engagement belong; The course scoring unit is used to calculate the matching score between student nodes and course nodes based on the final representation of student nodes and the final representation of course nodes output by the propagation unit, and to sort all courses according to the matching score. The personalized education recommendation scheme generation unit is used to generate personalized learning paths based on the sorting structure of the course scoring unit, the prerequisite / successor relationships between courses, and the knowledge graph constructed by the knowledge graph construction unit. The data acquisition unit communicates with the knowledge graph construction unit, the knowledge graph construction unit communicates with the dissemination unit, the dissemination unit communicates with the course scoring unit, and the course scoring unit communicates with the personalized education recommendation scheme generation unit.

2. The personalized education recommendation system according to claim 1, characterized in that, It also includes annotation units for describing personalized learning paths, which are communicatively connected to the data acquisition unit, the course scoring unit, and the personalized education recommendation generation unit.

3. A method for making personalized education recommendations using the personalized education recommendation system described in claim 1, characterized in that, Includes the following steps: S1) Collect course data, student personal data, and student social data; S2) The knowledge graph construction unit uses the course data, student personal data, and student social data collected in step S1) to construct a course-student-social knowledge graph, and simultaneously sets a course node set V. (c) and student node set V (s) Define prerequisite / successor relationship edges E between course nodes to represent knowledge dependency paths between courses. (cc) Establish a student-course edge E between student nodes and course nodes to represent the relationship between selected / completed / planned courses. (sc) Establish social relationship edges E between student nodes to represent the social influence between students. (ss) For each course node, each student node, and each social relationship edge E (ss) And each student - course side E (sc) Define the initial feature vector; S3) On the course-student-social knowledge graph constructed in step S2), set up a graph neural network model for modeling the relationship between student nodes and course nodes, and use the graph neural network model to generate vector representations of student nodes and course nodes in low-dimensional space respectively; S4) Use a graph neural network model to propagate through student nodes and course nodes in L rounds, and then calculate the matching score between student nodes and course nodes, where L is an integer greater than or equal to 2. S5) Rank the courses using matching scores and select the top N courses as recommendations. S6) Utilizing the prerequisite / successor relationship edges E in the course-student-social knowledge graph (cc) Starting from the courses recommended in step S5), a personalized learning path is generated along the path in the course-student-social knowledge graph.

4. The method according to claim 3, characterized in that, In step S4), when propagating the information between student nodes and course nodes using the graph neural network model, the course nodes are adjusted using the first adaptive gate control module in each round of propagation. The propagation weight is adjusted using the second adaptive gate control module to adjust the propagation weight of course nodes that are highly correlated with social interaction, thereby automatically adjusting the propagation intensity of different courses.

5. The method according to claim 4, characterized in that, The first adaptive gate control module calculates the course nodes using the following formula. Propagation weight: ; In the formula, Represented as course nodes The degree or normalized connection strength; Represented as student node All courses in the neighborhood.

6. The method according to claim 4, characterized in that, The second adaptive gate control module calculates the propagation weight of course nodes with high social relevance using the following formula: ; In the formula, Represented as student node and student nodes Social characteristic scores; Represented as student node A collection of social neighbors.

7. The method according to claim 3, characterized in that, In step S4), when using the graph neural network embedding propagation module to propagate information between student nodes and course nodes, the propagation formula used is: ; In the formula, Represented as student node In the Initial representation during layer propagation; Represents a non-linear activation function; and All are represented as the first Layer trainable parameters; Indicates the first Layer student node To student node Relationship weights.

8. The method according to claim 3, characterized in that, In step S4), the matching score is calculated using the following formula: ; In the formula, f() represents the probability or score of the inner product, bilinear function, or multilayer perceptron output; Represented as student node The final representation; Represented as course nodes The final representation.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 3 to 8.

10. A computer device, comprising a readable storage medium, a processor, and a computer program stored on the readable storage medium and executable on the processor, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 3 to 8.

Images