An intelligent interdisciplinary educational resource integration and recommendation system

Abstract

The present invention belongs to the technical field of learning recommendation, and specifically refers to an intelligent interdisciplinary educational resource integration and recommendation system, comprising a resource integration module, a resource recommendation module and a feedback information module; an aggregation and filtering method is used to aggregate and filter search results related to the same knowledge point, so as to extract the key content of the subject education resources; the subject education resources are represented as a graph structure through a graph isomorphic network, the neighborhood features of each node and edge in the graph isomorphic network are extracted, and aggregation and conversion are performed through a graph convolution layer and a perturbation adversarial function, so as to deeply explore the potential features and relationships of the knowledge points, and provide an accurate and effective basis for the recommended educational resources; students' learning progress and feedback information are collected, and the graph convolution layer and the perturbation adversarial function are optimized, so as to improve the timeliness and adaptability of the recommendation list.

Description

Technical Field

[0001] The present invention belongs to the technical field of learning recommendation, and specifically refers to an intelligent interdisciplinary educational resource integration and recommendation system. Background Art

[0002] Interdisciplinary education is an educational model that breaks the boundaries of traditional disciplines and integrates the knowledge, methods and ways of thinking of two or more different disciplines. It aims to cultivate students' comprehensive qualities and the ability to solve complex problems. It emphasizes the interconnection and integration between different disciplines and encourages students to think, analyze and solve problems from a multidisciplinary perspective.

[0003] The current education platform recommendation system recommends learning based on a single subject, which has the following shortcomings:

[0004] Educational resources are usually simply classified by subject, and the data representation is relatively simple. There is a lack of effective representation of the complex relationship between educational resources. Moreover, it is limited to the knowledge system within the subject, and it is difficult to reflect the connection and integration of interdisciplinary knowledge. It is impossible to effectively guide students to conduct interdisciplinary learning, making it difficult for students to discover the connection and specific application of knowledge points between different subjects.

[0005] Usually, students’ learning progress is not tracked in a timely enough manner, and recommendation lists cannot be generated based on students’ latest learning progress and feedback information. This leads to a lack of targeted recommendations, failure to meet the learning needs and interests of different students, difficulty in achieving personalized resource recommendations, and poor timeliness and adaptability of recommendations. Summary of the invention

[0006] In view of the above situation, the present invention provides an intelligent interdisciplinary educational resource integration and recommendation system. The resource integration module adopts an aggregation and filtering method to aggregate and filter the search results related to the same knowledge point, and extract the key content of the subject education resources; the resource recommendation module represents the subject education resources as a graph structure through a graph isomorphic network, extracts the neighborhood features of each node and edge in the graph isomorphic network, and aggregates and transforms them through a graph convolution layer and a perturbation adversarial function, which can deeply explore the potential features and relationships of knowledge points and provide an accurate and effective basis for recommending educational resources; collect students' learning progress and feedback information, optimize the graph convolution layer and the perturbation adversarial function, and enhance the timeliness and adaptability of the recommendation list.

[0007] An intelligent interdisciplinary educational resource integration and recommendation system includes: a resource integration module, a resource recommendation module and a feedback information module;

[0008] The resource integration module is connected to the resource recommendation module, the resource recommendation module is connected to the feedback information module, and the feedback information module is connected to the resource integration module and the resource recommendation module;

[0009] The resource integration module uses an aggregation and filtering method to search and integrate subject education resources, collect subject education resources from learning websites, and form search results;

[0010] The resource recommendation module uses a graph isomorphic network to integrate and summarize subject education resources in the form of nodes and edges, and establishes associations between students, teachers, subjects, courses and resources, extracts neighborhood features of nodes and edges in the graph isomorphic network, aggregates and transforms the neighborhood features to obtain available features, generates a recommendation list of subject education resources based on the available features, and continuously optimizes and improves the recommendation list;

[0011] The feedback information module collects students' learning progress and feedback information on the recommendation list, and transmits it back to the resource recommendation module for optimizing and improving the recommendation list.

[0012] The resource integration module uses an aggregate filtering method to search and integrate subject education resources. The method includes the following steps:

[0013] Step G1: Obtain search results: Write a crawler program to simulate a browser to automatically access the learning website, collect subject education resources from the learning website, and form search results;

[0014] Step G2: Aggregate search results: Aggregate and filter search results related to the same knowledge point to obtain integrated results;

[0015] Step G3: Extract key content: Preset keywords, subject topics and subject knowledge, and use lexical analysis, syntactic analysis and named entity recognition methods based on the keywords, subject topics and subject knowledge to extract relevant key content involving the same knowledge point from the integrated results.

[0016] Furthermore, step G2 specifically includes the following steps:

[0017] Step G21: Knowledge point marking: Preset the user's learning goal, determine the definition and scope of the knowledge point based on the learning goal, and use the automatic classification algorithm to mark the knowledge point involved in each search result;

[0018] Step G22: Similarity calculation: Calculate the similarity between the search results, use the cosine similarity algorithm to calculate the similarity of the knowledge points involved in the search results, obtain a similarity score, and group the search results according to the similarity score;

[0019] Step G23: Aggregation operation: select search results in descending order according to user evaluations, perform aggregation operation by grouping, and obtain aggregated results;

[0020] Step G24: Filtering and screening: Set filtering conditions, filter and screen the aggregation results according to the filtering conditions, remove the aggregation results that do not meet the filtering conditions, and obtain the integrated results.

[0021] The resource recommendation module adopts a graph isomorphic network method to make personalized recommendations for subject education resources. The method specifically includes the following steps:

[0022] Step S1: Resource representation: Use a graph isomorphic network to represent the key content of subject education resources as a graph structure, which includes nodes, edges, neighborhood features, and an adjacency matrix. Initialize the graph isomorphic network, where nodes include student nodes, teacher nodes, subject nodes, course nodes, and resource nodes, and edges represent the relationship between nodes.

[0023] Step S2: Feature extraction: extract the neighborhood features of each node and edge of the graph isomorphic network, design the graph convolution layer and perturbation adversarial function, aggregate and transform the neighborhood features of the nodes, and obtain usable features;

[0024] Step S3: Resource recommendation: Generate a recommended list of subject education resources based on available features, and collect students’ learning progress and feedback information on the recommended list from the feedback information module;

[0025] Step S4: Dynamic update: According to the learning progress and feedback information, the available features of the graph isomorphic network nodes are updated in real time to optimize the recommended list of subject education resources.

[0026] Furthermore, step S2 specifically includes the following steps:

[0027] Step S21: Initialize nodes and edges: initialize node features of student nodes, teacher nodes, subject nodes, course nodes, and resource nodes, and initialize edge features;

[0028] Step S22: define the node aggregation method: use the five-layer graph isomorphism network convolution layer to define the aggregation method between nodes, select a student node as the initial node, and the nodes connected to the initial node by edges are the neighboring nodes. Use the adjacency matrix to obtain the neighborhood features of the initial node, and the neighborhood features are the node features of the neighboring nodes;

[0029] The aggregation method between nodes is defined as follows:

[0030] First convolutional layer: For each node, use the sum function to aggregate the node features of neighboring nodes, combine the aggregated node features with the node's own node features, and obtain the feature representation of the node after the first convolution layer, thus obtaining the first-layer node features;

[0031] The second convolutional layer: collects the node features of the neighboring nodes of each node again, and also uses the aggregation function to aggregate the node features of the neighboring nodes, and combines them with the node features of the first layer to obtain the node features of the second layer;

[0032] The third to fifth convolutional layers: Repeat the above steps, perform the third, fourth, and fifth convolution operations in sequence, and continuously update the representation of node features;

[0033] Step S23: assigning weights to nodes: assigning weights to the neighborhood nodes of the initial node, constructing a perturbation adversarial function, and generating a perturbation graph of the graph isomorphic network. In the perturbation graph, the importance of interaction is evaluated according to the weights of the neighborhood nodes;

[0034] Assign weights to the neighborhood nodes of the initial node using the following formula:

[0035] ;

[0036] Among them, i represents the initial node, j represents the neighboring node, represents the weight of the neighboring node j for the initial node i, represents the strength of the learning progress between the initial node and the neighboring nodes, represents the correlation of learning time between the initial node and the neighboring nodes, and is the adjustment parameter;

[0037] Step S24: Aggregate neighborhood features: Aggregate the neighborhood features of the initial node, obtain related subject nodes based on the neighborhood features and interaction importance of the initial node, aggregate the neighborhood features of the subject nodes, and obtain available features;

[0038] Aggregate the neighborhood features of the initial node, and use the following formula:

[0039] ;

[0040] Among them, k represents the level of the graph isomorphic network, v represents the node, and u represents the initial node currently traversed. represents the neighborhood features of the k-th layer node, represents the neighborhood features of the k+1th layer node, represents the set of neighboring nodes of node v, represents the learnable parameters, represents the multilayer perceptron of the kth layer, and the update node represents the further processing of the neighborhood features before. Represents the neighborhood characteristics of traversing the initial node u.

[0041] Furthermore, step S21 specifically includes the following steps:

[0042] Step S211: setting node characteristics of the student node, including learning history, hobbies, learning goals and ability level;

[0043] Learning history includes courses studied, course grades, and study time. Interests and hobbies are obtained through students' course records and browsing records. Learning goals are set to improve grades in a certain subject and expand knowledge in a specific field. Competence levels assess students' knowledge mastery based on learning grades and test performance;

[0044] Step S212: setting node features of the teacher node, where the teacher node features include teaching subjects and teaching styles;

[0045] Step S213: setting node features of the subject node, the subject node features including subject knowledge system, difficulty level, related subjects and application fields;

[0046] The subject knowledge system includes knowledge points and concepts, related subjects include the correlation and knowledge intersection with other subjects, and application fields include the actual application scope of the subject;

[0047] Step S214: the node features of the course node include knowledge points, teaching outline and course evaluation;

[0048] Step S215: The node characteristics of the resource node include resource type, subject content, difficulty level, usage frequency and usage evaluation.

[0049] Resource types include documents, videos, exercises, and experiments, and subject content includes specific topics and knowledge points;

[0050] Step S216: The features of the edges include: student subject edge: including students' study time, academic performance and interests in the subject, student course edge: including students' study progress and academic performance in the course, teacher course edge: including teachers' teaching effectiveness, course resource edge: including a list of resources used in the course, and the importance of resources in the course, subject edge: including the correlation between subjects and the degree of knowledge overlap.

[0051] Furthermore, step S3 specifically includes the following steps:

[0052] Step S31: Use a five-layer graph isomorphism network convolutional layer to convert the available features to obtain an embedded representation of the student node. The embedded representation is a representation that maps the neighborhood features of the student node into a low-dimensional vector space. According to the embedded representation of the student node, the appropriate probabilities of the student node, the course node, and the resource node are calculated. The subject education resources are selected in descending order of appropriate probabilities for recommendation, and a recommendation list is generated.

[0053] The aggregated neighborhood features are transformed using the following formula:

[0054] ;

[0055] Where A represents the adjacency matrix, Represents the convolution operation of the graph isomorphic network, performs linear transformation on the neighborhood features, and ReLU represents the nonlinear activation function;

[0056] Step S32: Collect user feedback information on the recommendation list, the feedback information includes learning frequency, learning time and learning evaluation, and filter and extract the feedback information.

[0057] Furthermore, step S4 specifically includes the following steps:

[0058] Step S41: Divide the training set and the validation set, randomly sample positive samples and negative samples, train the perturbation adversarial function, and use epoch to control the training process. The positive sample is the observed interaction importance, and the negative sample is the unobserved interaction importance.

[0059] Step S42: Screening negative samples based on semantic analysis, selecting negative samples containing neighborhood features of the subject knowledge system, and labeling the negative samples to learn the relationship between interaction importance and appropriate probability;

[0060] Step S43: Repeatedly sample positive samples and negative samples, and update the perturbation graph in each epoch, so that the similarity between the perturbation graph and the graph isomorphic network gradually decreases;

[0061] Step S44: using a personalized loss function to measure the degree of adaptation of the recommendation list according to the feedback information;

[0062] The formula used by the personalized loss function is as follows:

[0063] ;

[0064] ;

[0065] Where L represents the loss function, D represents the set of graph isomorphic networks, N represents the total number of pairs of student nodes and resource nodes, s represents the student node, and z represents the resource node. represents the true label, represents the importance of the interaction between the student node and the resource node, represents the prediction score, represents the interaction score between the predicted student node and the resource node, Represents the sigmoid function, which converts the score into a suitable probability. It indicates the predicted probability of interaction between student nodes and resource nodes. represents the constraint condition, argmax represents the appropriate probability parameter value that returns the maximum value of the expression, and Q represents the perturbation adversarial function;

[0066] Step S45: Calculate the perturbation of the embedding representation of the student node using the projection matrix according to the degree of adaptation of the recommendation list;

[0067] Step S46: Construct an embedding perturbation function, add perturbations to the embedding representation of the first layer of the graph isomorphic network, and add perturbations layer by layer. The formula used is as follows:

[0068] ;

[0069] ;

[0070] ;

[0071] ;

[0072] Among them, l represents the lth layer of the graph isomorphism network, represents the d×d projection matrix, represents the embedding representation of the student node after projection, represents the perturbation added by the student node in layer l, represents the perturbed embedding representation of the student node at layer l, represents the embedding representation of the student node, represents the embedding representation of the student node at layer l, represents the embedding representation of the student node at layer l+1, represents the amplitude of added disturbance, A represents the adjacency matrix of the student node, D represents the degree matrix of A, T represents the transpose operation of the matrix, f represents the encoder, represents the learnable embedding representation, Q represents the perturbation adversarial function, P represents the embedding perturbation function, represents the perturbation graph, L represents the loss function;

[0073] Step S47: Update the perturbation graph, generate a comparative view based on the calculated embedding representations of student nodes, course nodes, and resource nodes, and update the perturbation adversarial function and the embedding perturbation function.

[0074] The beneficial effects achieved by the present invention are as follows:

[0075] (1) By representing disciplinary education resources as a graph structure through a graph isomorphic network, we can comprehensively and systematically integrate elements such as students, teachers, disciplines, courses, and resources, break down barriers between disciplines, and achieve an organic combination of interdisciplinary resources. By extracting the neighborhood features of each node and edge in the graph isomorphic network and aggregating and transforming them through graph convolutional layers and perturbation adversarial functions, we can deeply explore the potential features and relationships of resources and provide a more accurate and effective basis for recommending disciplinary education resources.

[0076] (2) Generate a recommendation list based on available features and collect students' learning progress and feedback information. This can realize personalized resource recommendations, meet the learning needs and interests of different students, and improve learning outcomes. Based on learning progress and feedback information, the available features of graph isomorphic network nodes are updated in real time to optimize the recommendation list. The recommendation strategy can be adjusted in a timely manner to provide resources that are more in line with students' current learning status, thereby enhancing the timeliness and adaptability of recommendations.

[0077] (3) Through the accurate recommendation list of subject education resources, students can obtain the required resources more efficiently, avoid waste of resources, and improve the utilization rate of education resources. It also provides teachers and education administrators with information on student learning and resource usage, which helps them make more scientific teaching decisions and optimize teaching arrangements and resource allocation. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] Figure 1 A module diagram of an intelligent interdisciplinary educational resource integration and recommendation system proposed by the present invention;

[0079] Figure 2 This is a flow chart of the polymerization filtering method proposed by the present invention;

[0080] Figure 3 This is a flow chart of the graph isomorphic network method proposed in the present invention. DETAILED DESCRIPTION

[0081] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0082] Example 1: Reference Figure 1 ,This embodiment provides an intelligent interdisciplinary educational resource integration and recommendation system, including a resource integration module, a resource recommendation module and a feedback information module;

[0083] The resource integration module is connected to the resource recommendation module, the resource recommendation module is connected to the feedback information module, and the feedback information module is connected to the resource integration module and the resource recommendation module;

[0084] The resource integration module uses an aggregation and filtering method to search and integrate subject education resources, collect subject education resources from learning websites, and form search results;

[0085] The resource recommendation module uses a graph isomorphic network to integrate and summarize subject education resources in the form of nodes and edges, and establishes associations between students, teachers, subjects, courses and resources, extracts neighborhood features of nodes and edges in the graph isomorphic network, aggregates and transforms the neighborhood features to obtain available features, generates a recommendation list of subject education resources based on the available features, and continuously optimizes and improves the recommendation list;

[0086] The feedback information module collects students' learning progress and feedback information on the recommendation list, and transmits it back to the resource recommendation module for optimizing and improving the recommendation list.

[0087] Example 2: Reference Figure 2 This embodiment is based on the above embodiment. The resource integration module uses an aggregate filtering method to search and integrate subject education resources. The method includes the following steps:

[0088] Step G1: Obtain search results: Write a crawler program to simulate a browser to automatically access the learning website, collect subject education resources from the learning website, and form search results;

[0089] Subject education resources include:

[0090] (1) Text resources

[0091] Textbooks, supplementary teaching materials, academic papers, and study notes;

[0092] (2) Video resources

[0093] Online courses, teaching lectures, experimental demonstration videos, knowledge popularization videos;

[0094] (3) Audio resources

[0095] Listening practice materials, audiobooks, and lecture audios;

[0096] (4) Image resources

[0097] Diagrams, illustrations, mind maps;

[0098] Step G2: Aggregate search results: Aggregate and filter search results related to the same knowledge point to obtain integrated results;

[0099] Step G3: Extract key content: Preset keywords, subject topics and subject knowledge, and use lexical analysis, syntactic analysis and named entity recognition methods based on the keywords, subject topics and subject knowledge to extract key content related to the same knowledge point from the integrated results.

[0100] Embodiment 3: This embodiment is based on the above embodiment, and step G2 specifically includes the following steps:

[0101] Step G21: Knowledge point marking: Preset the user's learning goal, determine the definition and scope of the knowledge point based on the learning goal, and use the automatic classification algorithm to mark the knowledge point involved in each search result;

[0102] Step G22: Similarity calculation: Calculate the similarity between the search results, use the cosine similarity algorithm to calculate the similarity of the knowledge points involved in the search results, obtain a similarity score, and group the search results according to the similarity score.

[0103] Step G23: Aggregation operation: select search results in descending order according to user evaluations, perform aggregation operation by grouping, and obtain aggregated results;

[0104] Step G24: Filtering and screening: Set filtering conditions, filter and screen the aggregation results according to the filtering conditions, remove the aggregation results that do not meet the filtering conditions, and obtain the integrated results.

[0105] Example 4: Reference Figure 3 This embodiment is based on the above embodiment. The resource recommendation module adopts a graph isomorphic network method to make personalized recommendations for subject education resources. The method specifically includes the following steps:

[0106] Step S1: Resource integration: Use a graph isomorphic network to represent subject education resources as a graph structure, which includes nodes, edges, neighborhood features, and an adjacency matrix. Initialize the graph isomorphic network, where nodes include student nodes, teacher nodes, subject nodes, course nodes, and resource nodes, and edges represent the relationship between nodes.

[0107] Step S2: Feature extraction: extract the neighborhood features of each node and edge of the graph isomorphic network, design the graph convolution layer and perturbation adversarial function, aggregate and transform the neighborhood features of the nodes, and obtain usable features;

[0108] Step S3: Resource recommendation: Generate a recommendation list of subject education resources based on available features, and the feedback information module collects students’ learning progress and feedback information on the recommendation list;

[0109] Step S4: Dynamic update: According to the learning progress and feedback information, the available features of the graph isomorphic network nodes are updated in real time to optimize the recommended list of subject education resources.

[0110] Embodiment 5: This embodiment is based on the above embodiment, and step S2 specifically includes the following steps:

[0111] Step S21: Initialize nodes and edges: initialize node features of student nodes, teacher nodes, subject nodes, course nodes, and resource nodes, and initialize edge features;

[0112] Step S22: define the node aggregation method: use the five-layer graph isomorphism network convolution layer to define the aggregation method between nodes, select a student node as the initial node, and the nodes connected to the initial node by edges are the neighboring nodes. Use the adjacency matrix to obtain the neighborhood features of the initial node, and the neighborhood features are the node features of the neighboring nodes;

[0113] Step S23: assigning weights to nodes: assigning weights to the neighborhood nodes of the initial node, constructing a perturbation adversarial function, and generating a perturbation graph of the graph isomorphic network. In the perturbation graph, the importance of interaction is evaluated according to the weights of the neighborhood nodes;

[0114] Assign weights to the neighborhood nodes of the initial node using the following formula:

[0115] ;

[0116] Among them, i represents the initial node, j represents the neighboring node, represents the weight of the neighboring node j for the initial node i, represents the strength of the learning progress between the initial node and the neighboring nodes, represents the correlation of learning time between the initial node and the neighboring nodes, and is the adjustment parameter;

[0117] Step S24: Aggregate neighborhood features: Aggregate the neighborhood features of the initial node, obtain related subject nodes based on the neighborhood features and interaction importance of the initial node, aggregate the neighborhood features of the subject nodes, and obtain available features;

[0118] Aggregate the neighborhood features of the initial node, and use the following formula:

[0119] ;

[0120] Among them, k represents the level of the graph isomorphic network, v represents the node, and u represents the initial node currently traversed. represents the neighborhood features of the k-th layer node, represents the neighborhood features of the k+1th layer node, represents the set of neighboring nodes of node v, represents the learnable parameters, represents the multilayer perceptron of the kth layer, and the update node represents the further processing of the neighborhood features before. Represents the neighborhood characteristics of traversing the initial node u;

[0121] In this embodiment, student Li Hua has studied mathematics course A, and the neighborhood features of mathematics knowledge points related to A and knowledge points of physics and chemistry related to A are aggregated.

[0122] Embodiment 6: This embodiment is based on the above embodiment, and step S21 specifically includes the following steps:

[0123] Step S211: setting node characteristics of the student node, including learning history, hobbies, learning goals and ability level;

[0124] Learning history includes courses studied, course grades, and study time. Interests and hobbies are obtained through students' course records and browsing records. Learning goals are set to improve grades in a certain subject and expand knowledge in a specific field. Competence levels assess students' knowledge mastery based on learning grades and test performance;

[0125] This embodiment provides specific settings for student Li Hua:

[0126] 1. Learn History

[0127] Courses studied:

[0128] Chinese, mathematics, English, physics, chemistry, biology, history, geography, politics;

[0129] Course Grade:

[0130] Chinese: 80 points; fair performance in reading comprehension, but unstable scores in composition;

[0131] Mathematics: 85 points; good grasp of the algebra part, but some errors in solving the geometry part;

[0132] English: 88 points; relatively good at listening and reading comprehension, but still has room for improvement in grammar and writing;

[0133] Physics: 75 points; I have a good understanding of the mechanics part, but I have a hard time with the acoustics part;

[0134] Chemistry: 78 points; performed better in experimental operation, but the memory and understanding of chemical equations were not deep enough.

[0135] Biology: 82 points; good understanding of ecosystems and genetics;

[0136] History: 80 points; interested in ancient history and well-versed in it, but not fully memorized modern history;

[0137] Geography: 78 points; good understanding of physical geography, but easily confused by knowledge points in human geography;

[0138] Politics: 80 points; I have a certain ability to think about the philosophy part, but I don’t have a thorough understanding of some concepts in the economic part;

[0139] Study time:

[0140] On average, I spend 3 hours a day studying school courses, including 40 minutes each for Chinese and English, 50 minutes for mathematics, and about 20 minutes for each of the other subjects. I spend an additional 2 hours on review and preview on weekends;

[0141] 2. Hobbies

[0142] Li Hua showed high enthusiasm and creativity in art class, often actively participating in class discussions and creating works. In music class, he also showed a strong interest in different types of music and took the initiative to learn about the stories and culture behind the music;

[0143] Li Hua often browses web pages about painting techniques, art exhibitions, music recommendations and musicians' introduction pages. He also pays attention to videos about scientific experiments and technological innovations.

[0144] 3. Learning Objectives

[0145] Improve math scores to 90 points;

[0146] Expand knowledge in the field of computer science and plan to learn the basics of programming;

[0147] Step S212: setting node features of the teacher node, where the teacher node features include teaching subjects and teaching styles;

[0148] Step S213: setting node features of the subject node, the subject node features including subject knowledge system, difficulty level, related subjects and application fields;

[0149] The subject knowledge system includes knowledge points and concepts, related subjects include the correlation and knowledge intersection with other subjects, and application fields include the actual application scope of the subject;

[0150] Step S214: the node features of the course node include knowledge points, teaching outline and course evaluation;

[0151] Step S215: The node characteristics of the resource node include resource type, subject content, difficulty level, usage frequency and usage evaluation.

[0152] Resource types include documents, videos, exercises, and experiments, and subject content includes specific topics and knowledge points;

[0153] Step S216: The features of the edges include: student subject edge: including students' study time, academic performance and interests in the subject, student course edge: including students' study progress and academic performance in the course, teacher course edge: including teachers' teaching effectiveness, course resource edge: including a list of resources used in the course, and the importance of resources in the course, subject edge: including the correlation between subjects and the degree of knowledge overlap.

[0154] Embodiment 7: This embodiment is based on the above embodiment, and step S3 specifically includes the following steps:

[0155] Step S31: Use a five-layer graph isomorphism network convolutional layer to convert the available features to obtain an embedded representation of the student node. The embedded representation is a representation that maps the neighborhood features of the student node into a low-dimensional vector space. According to the embedded representation of the student node, the appropriate probabilities of the student node, the course node, and the resource node are calculated. The subject education resources are selected in descending order of appropriate probabilities for recommendation, and a recommendation list is generated.

[0156] The aggregated neighborhood features are transformed using the following formula:

[0157] ;

[0158] Where A represents the adjacency matrix, Represents the convolution operation of the graph isomorphic network, performs linear transformation on the neighborhood features, and ReLU represents the nonlinear activation function;

[0159] The recommended list is as follows:

[0160] Course resources: "The Beauty of Mathematics and Art Design" online course;

[0161] This course introduces the application of geometric figures in mathematics in art design, such as the embodiment of the golden ratio in painting and architecture, helping Li Hua to combine mathematical knowledge with artistic interests;

[0162] Course resources: "The Physical Principles of Sound" popular science article and related videos;

[0163] This course explains the relationship between physical knowledge such as the generation and propagation of sound and music, deepening his understanding of the two subjects while satisfying his interests and hobbies;

[0164] Course resources: Lecture on "Ancient Medicine and Biological Development";

[0165] This course introduces the development of biomedicine in ancient history, combining Li Hua's knowledge of biology and interest in history;

[0166] Course resources: "Introduction to Geographic Information Systems" online tutorial;

[0167] This course explains how to use computer technology to process geographic information. Li Hua aims to expand his knowledge in computer science. He also has a certain foundation in geography. This course allows him to get in touch with interdisciplinary practical applications.

[0168] Step S32: Collect user feedback information on the recommendation list, the feedback information includes learning frequency, learning time and learning evaluation, and filter and extract the feedback information.

[0169] Embodiment 8: This embodiment is based on the above embodiment, and step S4 specifically includes the following steps:

[0170] Step S41: Divide the training set and the validation set, randomly sample positive samples and negative samples, train the perturbation adversarial function, and use epoch to control the training process. The positive sample is the observed interaction importance, and the negative sample is the unobserved interaction importance.

[0171] Step S42: Screening negative samples based on semantic analysis, selecting negative samples containing neighborhood features of the subject knowledge system, and labeling the negative samples to learn the relationship between interaction importance and appropriate probability;

[0172] Step S43: Repeatedly sample positive samples and negative samples, and update the perturbation graph in each epoch, so that the similarity between the perturbation graph and the graph isomorphic network gradually decreases;

[0173] Step S44: using a personalized loss function to measure the degree of adaptation of the recommendation list according to the feedback information;

[0174] The formula used by the personalized loss function is as follows:

[0175] ;

[0176] ;

[0177] Where L represents the loss function, D represents the set of graph isomorphic networks, N represents the total number of pairs of student nodes and resource nodes, s represents the student node, and z represents the resource node. represents the true label, represents the importance of the interaction between the student node and the resource node, represents the prediction score, represents the interaction score between the predicted student node and the resource node, Represents the sigmoid function, which converts the score into a suitable probability. It indicates the predicted probability of interaction between student nodes and resource nodes. represents the constraint condition, argmax represents the appropriate probability parameter value that returns the maximum value of the expression, and Q represents the perturbation adversarial function;

[0178] Step S45: Calculate the perturbation of the embedding representation of the student node using the projection matrix according to the degree of adaptation of the recommendation list;

[0179] Step S46: Construct an embedding perturbation function, add perturbations to the embedding representation of the first layer of the graph isomorphic network, and add perturbations layer by layer. The formula used is as follows:

[0180] ;

[0181] ;

[0182] ;

[0183] ;

[0184] Among them, l represents the lth layer of the graph isomorphism network, represents the d×d projection matrix, represents the embedding representation of the student node after projection, represents the perturbation added by the student node in layer l, represents the perturbed embedding representation of the student node at layer l, represents the embedding representation of the student node, represents the embedding representation of the student node at layer l, represents the embedding representation of the student node at layer l+1, represents the amplitude of added disturbance, A represents the adjacency matrix of the student node, D represents the degree matrix of A, T represents the transpose operation of the matrix, f represents the encoder, represents the learnable embedding representation, Q represents the perturbation adversarial function, P represents the embedding perturbation function, represents the perturbation graph, L represents the loss function;

[0185] Step S47: Update the perturbation graph, generate a comparative view based on the calculated embedding representations of student nodes, course nodes, and resource nodes, and update the perturbation adversarial function and the embedding perturbation function.

[0186] The present invention and its implementation methods are described above. Such description is not restrictive. If a person skilled in the art is inspired by it and creatively designs structures and implementation methods similar to the technical solution without departing from the purpose of the present invention, they should all fall within the protection scope of the present invention.

Claims

1. An intelligent interdisciplinary educational resource integration and recommendation system, characterized by: It includes resource integration module, resource recommendation module and feedback information module; The resource integration module uses an aggregation and filtering method to search and integrate subject education resources, collect subject education resources from learning websites, and form search results; The resource recommendation module adopts a graph isomorphic network method to integrate and summarize subject education resources in the form of nodes and edges, extracts neighborhood features of nodes in the graph isomorphic network, aggregates and transforms the neighborhood features to obtain available features, generates a recommendation list of subject education resources based on the available features, and continuously optimizes and improves the recommendation list; The feedback information module collects students' learning progress and feedback information on the recommendation list and transmits it back to the resource recommendation module for optimizing and improving the recommendation list; The resource recommendation module adopts a graph isomorphic network method to make personalized recommendations for subject education resources. The method specifically includes the following steps: Step S1: Resource integration: Use a graph isomorphic network to represent subject education resources as a graph structure, which includes nodes, edges, neighborhood features, and an adjacency matrix. Initialize the graph isomorphic network, where nodes include student nodes, teacher nodes, subject nodes, course nodes, and resource nodes, and edges represent the relationship between nodes. Step S2: Feature extraction: extract the neighborhood features of the graph isomorphic network nodes, design the graph convolution layer and perturbation adversarial function, aggregate and transform the neighborhood features of the nodes, and obtain usable features; Step S3: Resource recommendation: Generate a recommendation list of subject education resources based on available features, and receive students’ learning progress and feedback information on the recommendation list collected by the feedback information module; Step S4: Dynamic update: According to the learning progress and feedback information, the available features of the graph isomorphic network nodes are updated in real time to optimize the recommended list of subject education resources; Step S2 specifically includes the following steps: Step S21: Initialize the node features of student nodes, teacher nodes, subject nodes, course nodes and resource nodes, and initialize the features of edges; Step S22: using a five-layer graph isomorphism network convolutional layer, defining the aggregation method between nodes, selecting a student node as the initial node, and the nodes connected to the initial node by edges are the neighboring nodes, and using the adjacency matrix to obtain the neighborhood features of the initial node, the neighborhood features are the node features of the neighboring nodes; Step S23: assigning weights to the neighborhood nodes of the initial node, constructing a perturbation adversarial function, generating a perturbation graph of the graph isomorphic network, and evaluating the interactive importance of the neighborhood nodes to the initial node according to the weights of the neighborhood nodes in the perturbation graph; Step S24: Aggregate the neighborhood features of the initial node, obtain related subject nodes based on the neighborhood features and interaction importance of the initial node, aggregate the neighborhood features of the subject nodes, and obtain available features; Step S3 specifically includes the following steps: Step S31: Use a five-layer graph isomorphism network convolution layer to convert the available features to obtain an embedded representation of the student node. The embedded representation maps the neighborhood features of the student node to a representation in a low-dimensional vector space. According to the embedded representation of the student node, the appropriate probabilities of the student node, the course node, and the resource node are calculated. The subject education resources are selected in descending order of appropriate probabilities for recommendation, and a recommendation list is generated. Step S32: Collect user feedback information on the recommendation list, the feedback information includes learning frequency, learning time and learning evaluation, and filter and extract the feedback information.

2. The intelligent interdisciplinary educational resource integration and recommendation system according to claim 1, characterized in that: Step S4 The specific steps include: Step S41: Divide the training set and the validation set, randomly sample positive samples and negative samples, train the perturbation adversarial function, and use epoch to control the training process. The positive sample is the observed interaction importance, and the negative sample is the unobserved interaction importance. Step S42: Screening negative samples based on semantic analysis, selecting negative samples containing neighborhood features of the subject knowledge system, and labeling the negative samples to learn the relationship between interaction importance and appropriate probability; Step S43: Repeatedly sample positive samples and negative samples, and update the perturbation graph in each epoch, so that the similarity between the perturbation graph and the graph isomorphic network gradually decreases; Step S44: using a personalized loss function to measure the degree of adaptation of the recommendation list according to the feedback information; Step S45: Calculate the perturbation of the embedding representation of the student node using the projection matrix according to the degree of adaptation of the recommendation list; Step S46: construct an embedding perturbation function, add perturbations to the embedding representation of the first layer of the graph isomorphic network, and add perturbations layer by layer; Step S47: Update the perturbation graph, generate a comparative view based on the calculated embedding representations of student nodes, course nodes, and resource nodes, and update the perturbation adversarial function and the embedding perturbation function.

3. The intelligent interdisciplinary educational resource integration and recommendation system according to claim 2, characterized in that: The resource integration module uses an aggregate filtering method to search and integrate subject education resources. The method includes the following steps: Step G1: Obtain search results: Write a crawler program to simulate a browser to automatically access the learning website, collect subject education resources from the learning website, and form search results; Step G2: Aggregate search results: Aggregate and filter search results related to the same knowledge point to obtain integrated results; Step G3: Extract key content: Preset keywords, subject topics and subject knowledge, and use lexical analysis, syntactic analysis and named entity recognition methods based on the keywords, subject topics and subject knowledge to extract key content related to the same knowledge point from the integrated results.

4. The intelligent interdisciplinary educational resource integration and recommendation system according to claim 3 is characterized by: Step G2 specifically includes the following steps: Step G21: Knowledge point marking: Preset the user's learning goal, determine the definition and scope of the knowledge point based on the learning goal, and use the automatic classification algorithm to mark the knowledge point involved in each search result; Step G22: Similarity calculation: Calculate the similarity between the search results, use the cosine similarity algorithm to calculate the similarity of the knowledge points involved in the search results, obtain a similarity score, and group the search results according to the similarity score; Step G23: Aggregation operation: select search results in descending order according to user evaluations, perform aggregation operation by grouping, and obtain aggregated results; Step G24: Filtering and screening: Set filtering conditions, filter and screen the aggregation results according to the filtering conditions, remove the aggregation results that do not meet the filtering conditions, and obtain the integrated results.

Images