An improved online education resource recommendation method and system for graph neural networks

Abstract

The embodiment of the application provides an online education resource recommendation method and system for improving a graph neural network. In the method, a multi-path heterogeneous graph is generated based on education resources, learning ability of students and knowledge points of the resources. A recommendation model for predicting preferences of students is constructed by introducing the multi-path heterogeneous graph into the graph neural network. The recommendation model includes a time sequence feature module and a graph neural network module. In the graph neural network module, a path perception convolution layer constructs a type conversion matrix for different types of nodes and edges in the multi-path heterogeneous graph, respectively performs convolution operations on different paths, and aggregates features of adjacent nodes. A feature fusion layer converts processed heterogeneous node features to a unified space through a cross-modal mapping matrix to obtain node representation. The recommendation model is used to recommend education resources to students. The application is more suitable for the characteristics of knowledge reserve decay, dynamic change of learning ability and multi-dimensional correlation in the online education scene, and improves the pertinence and effectiveness of the recommendation result.

Description

TECHNICAL FIELD

[0001] The embodiment of the application belongs to the technical field of online education, and particularly relates to an online education resource recommendation method and system for improving a graph neural network. BACKGROUND

[0002] With the rapid development of Internet technology, the online education industry is showing explosive growth, and a large amount of education resources is emerging. In order to help learners quickly find suitable content from numerous resources, online education resource recommendation systems have emerged.

[0003] Early online education resource recommendation is mainly based on simple keyword matching technology. This technology finds resources with high matching degree by analyzing the keywords input by students and the title, abstract and other text information of education resources. This method is relatively simple to implement and can meet the needs of students to find resources based on clear needs to some extent. However, keyword matching technology has some limitations, which are reflected in that the traditional recommendation method regards the learning ability of students as a static feature, does not consider the natural decay of knowledge reserve over time (such as the forgetting curve law), and does not dynamically update the learning speed according to the recent learning task completion, which makes it difficult to accurately capture the real-time ability state of students, and the recommended results often appear to be mismatched with the current ability of students and lag behind the learning progress. SUMMARY

[0004] The purpose of the embodiment of the application is to provide an online education resource recommendation method and system for improving a graph neural network, which has at least solved the technical problem that the recommendation method of the related art is difficult to accurately capture the real-time ability state of students, and the recommended results often appear to be mismatched with the current ability of students and lag behind the learning progress.

[0005] To achieve the above-mentioned purpose, the embodiment of the application provides the following technical solution.

[0006] According to one embodiment of the application, an online education resource recommendation method for improving a graph neural network is provided.

[0007] The method comprises the following steps:

[0008] A multi-path heterogeneous graph is generated based on education resources, student learning ability and resource knowledge points, a multi-path heterogeneous graph is introduced into a graph neural network, a recommendation model for predicting student preferences is constructed, the recommendation model comprises a time sequence feature module and a graph neural network module, and the edges between nodes in the heterogeneous graph are used to represent the interaction relationship between education resources, student learning ability and resource knowledge points.

[0009] In the time sequence feature module, the time sequence learning data of the student is processed, key information is extracted, a decay coefficient is introduced, the decay amount of the knowledge reserve is calculated according to the time interval, and the learning speed is dynamically updated according to the completion of the recent learning task of the student, so as to obtain the knowledge reserve feature and the learning speed feature of the student; the multi-path heterogeneous graph is reconstructed based on the knowledge reserve feature and the learning speed feature output by the time sequence feature module;

[0010] In the graph neural network module, the path perception convolution layer constructs a type conversion matrix for different types of nodes and edges in the multi-path heterogeneous graph, performs convolution operation on different paths respectively, aggregates the features of adjacent nodes, and filters the edges that do not match the current ability of the student through the edge type mask matrix; the time sequence-structure fusion attention layer introduces a time sequence decay factor to assign dynamic attention weights to the edges of different paths and time intervals; the feature fusion layer converts the processed heterogeneous node features to a unified space through a cross-modal mapping matrix to obtain node representations;

[0011] The trained recommendation model is used to predict the student's preference by taking the node representation as the input of the recommendation model, and to recommend educational resources to the student based on the student's preference.

[0012] Preferably, the learning ability of the student includes the knowledge reserve and the learning speed, and an evaluation model for measuring the learning ability of the student is constructed by obtaining the learning data of the student in a target time period, wherein in the construction process of the evaluation model:

[0013] The knowledge reserve of the student is calculated through the learning performance of the student in the target time period; wherein the learning performance is quantified by completion, correct answer rate and video watching completion rate; the knowledge reserve is represented as: ; in the formula, S represents the knowledge reserve of the student, t represents the time point, i represents the i-th learning task, n represents the total number of learning tasks, w represents the weight of i, and P represents the learning performance of the student on task i;

[0014] The learning speed of the student is calculated by comparing the knowledge reserves of the same student at different time points; wherein the learning speed is represented as: ; V represents the learning speed of the student, 、 respectively represent the knowledge reserves of the student at time and time ;

[0015] Based on the historical knowledge reserve and the learning speed of the student, a linear regression is used to calculate the learning ability of the student at a future time point, and a first model for measuring the learning ability of the student is constructed, represented as: ; wherein, denote the learning ability model; a, b denote the parameters of the linear regression model; denote the future time point, which is the prediction time point.

[0016] Preferably, the step of generating a multi-path heterogeneous graph based on the educational resources, the cognitive ability of the students and the knowledge points of the resources comprises:

[0017] Obtain multi-source heterogeneous data, which includes student behavior logs, educational resource metadata and knowledge point hierarchical structure;

[0018] Construct a heterogeneous graph structure including a student node layer, a resource node layer and a knowledge point node layer; in the student node layer, each node corresponds to a unique student ID and carries a learning ability feature vector output by the first model; in the resource node layer, subcategories are divided according to types, and node features are fused with resource text vectors and metadata; in the knowledge point node layer, based on the hierarchical relationship of the subject knowledge graph, directed edges are established between nodes;

[0019] Construct edges between nodes, wherein student-resource edges are established through student behavior logs; resource-knowledge point edges are established through the matching degree of resource content and knowledge points; student-knowledge point edges are established through the mastery of knowledge points by students;

[0020] Based on the constructed heterogeneous graph, form a weighted path network based on multi-path types to obtain a multi-path heterogeneous graph, wherein the multi-path types include a learning path, a demand path and an association path, the learning path is used to reflect the trajectory of students obtaining knowledge through resources, the path weight adopts edge weight continuous multiplication and adds a time attenuation factor; the demand path is used to depict the matching relationship between the ability of students and the requirements of knowledge points, and a knowledge point importance weight is introduced in the path; the association path is used to mine the potential connection between knowledge points, and the path weight is calculated based on the cross-knowledge point association degree matrix of resources.

[0021] Preferably, in the path perception convolution layer, the features of different paths are aggregated differently, and the aggregated features of a student node s are represented as:

[0022]

[0023] In the formula, denote the set of resource nodes associated with the student s; σ denotes an activation function; denote the attention weight of the student-resource edge, which is represented as:

[0024]

[0025] In the formula, denote the conversion matrix of the node s, which is used to convert the features of the node s; a transformation matrix representing node r, used to transform the features of node r; a transformation matrix representing node k, used to transform the features of node k; an activation function used to introduce nonlinearity, a learnable weight vector; a feature vector representing node r; a feature vector representing node s; a set of neighbor nodes of node s; a neighbor node; a feature vector of a neighbor node; a concatenation operation on vectors.

[0026] Preferably, in the time-structure fusion attention layer, a knowledge reserve amount decay coefficient of the time sequence module output is introduced to dynamically adjust the path weight, represented as:

[0027]

[0028]

[0029] wherein, a convolutional feature corresponding to path p, a path importance weight; a time sequence decay factor of path p, a timestamp of path p, and δ represents a knowledge reserve amount decay coefficient.

[0030] Preferably, in the feature fusion layer:

[0031] A mapping matrix is constructed for different node types, including student mapping, resource mapping, and knowledge point mapping;

[0032] An adversarial learning constraint is introduced, the minimum mapping type discrimination accuracy is taken as the generator target, the ability to distinguish different types of nodes is taken as the discriminator target, so that the mapped features meet the semantic consistency and type distinguishability;

[0033] In the mapping process, the residual connection mechanism is used to retain key information in the original features;

[0034] An attention mechanism is used to assign different weights to features from different sources, and interactive features between nodes are constructed;

[0035] A gating mechanism is introduced to filter effective information, and the fused features after gating are converted to a unified space through a cross-modal mapping matrix, and finally the node representation is output.

[0036] Preferably, in the step of introducing a gating mechanism to screen effective information, a multi-path attention gating mechanism is used to separately construct gating parameters for learning paths, demand paths and association paths in the multi-path heterogeneous graph, and the gating parameters are represented as:

[0037]

[0038]

[0039] wherein p∈{learning path, demand path, association path}, represents a path weight, represents a knowledge reserve change rate; represents a learning speed; represents a sigmoid activation function for mapping input values between 0 and 1; represents element-by-element multiplication; and [;] represents a vector concatenation operation. represents original mapping features; represents features after gating processing and fusion; represents a gating weight matrix; represents a gating signal.

[0040] Preferably, the step of separately constructing gating parameters for each path includes:

[0041] The gating signal constructed for each path is generated by nonlinear transformation of the mapping features, the knowledge reserve change rate and the learning speed corresponding to the path, and is used to dynamically balance the static mapping features and dynamic learning features of a single path.

[0042] The features of each path after gating processing are assigned a path weight, wherein the path weight is dynamically set based on the importance of the path to the student preference prediction; the correlation between the path features and the student preference label is calculated through an attention mechanism that fuses a time decay factor, and the path weight is adjusted in real time; at the same time, in the training phase of the recommendation model, the path weight is iteratively adjusted from a global perspective through back propagation of the prediction error.

[0043] The weighted features of all paths are aggregated to obtain the final fusion features.

[0044] According to another embodiment of the present application, an online education resource recommendation system for improving a graph neural network is provided.

[0045] The system comprises the following modules:

[0046] The recommendation model construction module is configured to generate a multi-path heterogeneous graph based on the educational resources, the learning ability of the student and the knowledge points of the resources, introduce the multi-path heterogeneous graph into a graph neural network, and construct a recommendation model for predicting the preference of the student, the recommendation model comprising a time sequence feature module and a graph neural network module, and edges between nodes in the heterogeneous graph being configured to represent the interaction relationship between the educational resources, the learning ability of the student and the knowledge points of the resources;

[0047] The heterogeneous graph reconstruction module is configured to process the time sequence learning data of the student in the time sequence feature module, extract key information, introduce a decay coefficient, calculate the decay amount of the knowledge reserve amount according to the time interval, and dynamically update the learning speed according to the completion of the recent learning task of the student, to obtain the knowledge reserve amount feature and the learning speed feature of the student; and the multi-path heterogeneous graph is reconstructed based on the knowledge reserve amount feature and the learning speed feature output by the time sequence feature module.

[0048] The node representation module is configured to, in the graph neural network module, the path perception convolutional layer constructs a type conversion matrix for different types of nodes and edges in the multi-path heterogeneous graph, performs convolutional operation on different paths respectively, aggregates the features of adjacent nodes, and filters the edges that do not match the current ability of the student through an edge type mask matrix; the time sequence-structure fusion attention layer introduces a time sequence decay factor to assign dynamic attention weights to edges of different paths and time intervals; and the feature fusion layer converts the processed heterogeneous node features to a unified space through a cross-modal mapping matrix to obtain the node representation.

[0049] The resource recommendation module is configured to use the trained recommendation model to take the node representation as the input of the recommendation model, predict the preference of the student, and recommend the educational resources to the student based on the preference of the student.

[0050] According to still another embodiment of the present application, a computer readable storage medium is provided, and the computer readable storage medium stores a computer program, wherein the computer program is executed by a processor to implement the steps of the online education resource recommendation method of the improved graph neural network provided in the above embodiments.

[0051] Compared with the prior art, the technical advantages of the online education resource recommendation method and system of the improved graph neural network provided in the embodiments of the present application are as follows:

[0052] First, the knowledge reserve amount decay coefficient and the learning speed dynamic updating mechanism are introduced by the time sequence feature module, the change of the learning ability of the student (such as knowledge forgetting and recent learning efficiency fluctuation) is tracked in real time, the limitation that the learning ability is statically processed in the traditional method is solved, the recommendation result can closely match the current knowledge level and learning rhythm of the student, the recommended content is avoided to be inconsistent with the actual ability, the dynamic learning state of the student can be accurately captured, and the recommendation adaptability is improved.

[0053] Secondly, the application is based on educational resources, student learning ability and resource knowledge points to construct a multi-path heterogeneous graph, and captures multi-dimensional interaction relationships in a targeted manner, thereby breaking through the problem of insufficient description of complex associations in an education scene by a single graph structure. Meanwhile, the path perception convolution layer and the edge type mask matrix can accurately filter irrelevant information, so that the model focuses on effective associations that match the student's ability, thereby improving the pertinence of feature extraction.

[0054] Thirdly, the time sequence-structure fusion attention layer of the application introduces a time sequence attenuation factor to assign dynamic weights to interactions of different time intervals, thereby reflecting the strong influence of recent learning behaviors and taking into account the reference value of historical associations. The feature fusion layer realizes unified representation of heterogeneous features through cross-modal mapping, solves the fusion problem caused by differences in multi-source feature dimensions, makes the node representation more comprehensively reflect the student's preferences, and finally improves the prediction accuracy.

[0055] In summary, through the cooperation of dynamic ability modeling, multi-path graph structure design, time sequence-structure fusion and precise gate screening, the application significantly improves the accuracy and educational practicality of resource recommendation. BRIEF DESCRIPTION OF DRAWINGS

[0056] The drawings accompanying the specification of this application form a part thereof, serve to provide further understanding of the application, and together with the description of the exemplary embodiments of the application and the explanation thereof serve to explain the application, and do not constitute an improper limitation of the application.

[0057] In the drawings:

[0058] Figure 1 An implementation flowchart of an online education resource recommendation method of an improved graph neural network provided for the embodiments of the application;

[0059] Figure 2 A sub-flowchart of an online education resource recommendation method of an improved graph neural network provided for the embodiments of the application;

[0060] Figure 3 A structure block diagram of a recommendation model provided for the embodiments of the application;

[0061] Figure 4 Another sub-flowchart of an online education resource recommendation method of an improved graph neural network provided for the embodiments of the application;

[0062] Figure 5 Still another sub-flowchart of an online education resource recommendation method of an improved graph neural network provided for the embodiments of the application;

[0063] Figure 6 A structure block diagram of an online education resource recommendation system of an improved graph neural network provided for the embodiments of the application. DETAILED DESCRIPTION

[0064] In order for those skilled in the art to better understand the scheme of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should be within the scope of protection of the present application.

[0065] It should be noted that the user data (including but not limited to data for display, analyzed data, etc.) involved in the present application is information and data authorized by the user or authorized by all parties, and the collection, storage, use, processing, transmission, provision, disclosure and application of related data comply with relevant laws, regulations and standards, necessary security measures are taken, do not violate public order and good customs, and provide corresponding operation portal for user to choose authorization or refusal.

[0066] According to an embodiment of the present application, an online education resource recommendation method for improving graph neural network is provided, please refer to Figure 1 and Figure 3 The online education resource recommendation method for improving graph neural network provided by the embodiments of the present application includes the following steps:

[0067] S101, based on the education resources, the learning ability of students and the knowledge points of resources, a multi-path heterogeneous graph is generated, a multi-path heterogeneous graph is introduced into the graph neural network, a recommendation model for predicting student preferences is constructed, the recommendation model includes a time sequence feature module and a graph neural network module, and the edges between nodes in the heterogeneous graph are used to represent the interaction relationship between the education resources, the learning ability of students and the knowledge points of resources;

[0068] The multi-path heterogeneous graph of the embodiments of the present application simultaneously includes three types of nodes of education resources, learning ability of students and knowledge points of resources, and distinguishes the interaction relationship through the type of edge, such as learning behavior edge of student-resource, matching degree edge of student ability-knowledge point, and coverage degree edge of resource-knowledge point, accurately captures the multi-dimensional association of behavior-ability-knowledge in online education, and solves the problem of one-sidedness of traditional models in describing complex relationships in education scene.

[0069] The time sequence feature module built in the recommendation model provided by the present application is used to process dynamic learning ability, and the graph neural network module for processing the structure relationship of the heterogeneous graph and the time sequence feature module form a complement: the time sequence module injects time dynamics (such as knowledge forgetting and learning speed change) into the graph structure, and the graph neural network module provides structural association (such as the influence of the dependency relationship between knowledge points on the learning path) for the time sequence feature, and the combination of the two makes the model understand the real-time state of the students.

[0070] Specifically, the weight of the learning behavior edge is the learning duration of the learner on the resource; the weight of the ability matching edge is the matching degree of the current ability of the learner and the difficulty of the knowledge point; the weight of the knowledge coverage edge is the coverage depth of the resource on the knowledge point; and the weight of the knowledge association edge is the logical dependency degree between knowledge points.

[0071] Exemplarily, the multiple paths can be represented as a learning behavior path, an ability improvement path, and a knowledge association path.

[0072] Further, in step S101, the learning ability of the learner provided by the embodiment of the present application includes a knowledge reserve amount and a learning speed, and an evaluation model for measuring the learning ability of the learner is constructed by acquiring the learning data of the learner in a target time period;

[0073] Specifically, in the embodiment of the present application, in the construction process of the evaluation model:

[0074] The knowledge reserve amount of the learner is calculated through the learning performance of the learner in the target time period; wherein the learning performance is quantified by completion degree, correct answer rate, and video watching completion rate; and the knowledge reserve amount is represented as:

[0075]

[0076] In the formula, S represents the knowledge reserve amount of the learner, t represents a time point, i represents the i-th learning task, n represents the total number of learning tasks, w represents the weight of i, and P represents the learning performance of the learner on the task i;

[0077] The learning speed of the learner is calculated by comparing the knowledge reserve amounts of the same learner at different time points; wherein the learning speed is represented as: ; V represents the learning speed of the learner, 、 respectively represent the knowledge reserve amounts of the learner at time and time ;

[0078] Based on the historical knowledge reserve amount and the learning speed of the learner, a linear regression is used to calculate the learning ability of the learner at a future time point, and a first model for measuring the learning ability of the learner is constructed, represented as: ; wherein, represents the learning ability model; a and b represent the parameters of the linear regression model; represents a future time point, which is a predicted time point.

[0079] Further, in step S101, as shown in Figure 2 , the step of generating a multiple-path heterogeneous graph based on the educational resources, the cognitive ability of the learner, and the knowledge points of the resources includes:

[0080] S201, acquire multi-source heterogeneous data, the multi-source heterogeneous data including a student behavior log, education resource metadata and a knowledge point hierarchical structure;

[0081] In step S201, in the step of acquiring the student behavior log, the interaction track of the student and the resource (such as clicking, learning duration, exercise submission record, wrong question marking, etc.) is recorded, which is used to reflect the actual learning preference and behavior habit of the student; the education resource metadata includes resource type (video, document, exercise), title, content abstract, difficulty label, release time, etc., which is used to extract resource features (such as text vector, type attribute) to support the differentiated representation of the resource node; in the knowledge point hierarchical structure, based on the subject knowledge graph, the dependency relationship between knowledge points is clear, which provides a logical basis for the construction of the directed edge of the knowledge point node layer.

[0082] S202, construct a heterogeneous graph structure including a student node layer, a resource node layer and a knowledge point node layer; in the student node layer, each node corresponds to a unique student ID and carries a learning ability feature vector output by the first model; in the resource node layer, subcategories are divided according to types, and node features are fused with resource text vectors and metadata; in the knowledge point node layer, based on the hierarchical relationship of the subject knowledge graph, directed edges are established between nodes;

[0083] In the student node layer of step S202, each node is bound to a unique student ID, and the core feature is the learning ability feature vector output by the evaluation, such as knowledge reserve amount, learning speed, weak knowledge point distribution and the like; in the resource node layer, subcategories are divided according to resource types, and node features adopt a text vector and metadata fusion mode to ensure that the resource node can reflect the dual features of content attribute and interaction attribute; in the knowledge point node layer, based on the hierarchical relationship of the subject knowledge graph, directed edges are established between nodes, so that the knowledge point node layer itself has a structured knowledge association logic.

[0084] S203, construct edges between nodes, wherein the student-resource edge is established through the student behavior log; the resource-knowledge point edge is established through the matching degree of resource content and knowledge point; the student-knowledge point edge is established through the mastery of the knowledge point by the student;

[0085] In step S203, the association link of behavior-ability-knowledge is established, the three-layer nodes are connected through the definition of edges, the interaction intensity of student-resource-knowledge point is quantified, and the association skeleton of the heterogeneous graph is formed; wherein, in the student-resource edge, the weight of the edge quantifies the interaction depth based on the construction of the student behavior log. For example, student A completely watches the educational resource video (time length accounts for 90%), and the edge weight of the student and the resource is higher than that of student B who only watches 10%, which intuitively reflects the attraction or importance of the resource to the student; in the resource-knowledge point edge, the weight is calculated through the matching degree of resource content and knowledge point, which reflects the coverage degree of the resource to the knowledge point; in the student-knowledge point edge, the weight is quantified based on the mastery of the student to the knowledge point.

[0086] S204, based on the constructed heterogeneous graph, a weighted path network is formed based on multiple path types, and a multi-path heterogeneous graph is obtained.

[0087] In step S204, the differentiated representation of the multi-path is realized, wherein, based on the heterogeneous graph, the weight is assigned according to the path type (such as learning behavior path, ability improvement path, knowledge association path), and the network structure of “multi-path parallel and weighted” is formed, so that the graph neural network can learn the features of different paths.

[0088] Preferably, in step S204 of the embodiment, the multi-path type provided by the embodiment of the application includes a learning path, a demand path and an association path; wherein the learning path is used to reflect the trajectory of the student obtaining knowledge through the resource, the path weight is multiplied by the edge weight and a time decay factor is added; the demand path is used to depict the matching relationship between the student's ability and the knowledge point requirement, and the knowledge point importance weight is introduced in the path; the association path is used to mine the potential connection between knowledge points, and the path weight is calculated based on the cross-knowledge point association matrix of the resource.

[0089] Further, please continue to refer to Figure 1 In the embodiment of the application, the online education resource recommendation method for improving the graph neural network provided by the embodiment of the application further includes the following steps:

[0090] S102, in the time sequence feature module, the student time sequence learning data is processed, the key information is extracted, the decay coefficient is introduced, the decay amount of the knowledge reserve is calculated according to the time interval, and the learning speed is dynamically updated according to the completion of the recent learning task of the student, and the knowledge reserve feature and the learning speed feature of the student are obtained; the multi-path heterogeneous graph is reconstructed based on the knowledge reserve feature and the learning speed feature output by the time sequence feature module;

[0091] The embodiment introduces a decay coefficient and calculates the knowledge reserve decay based on a time interval, conforms to the cognitive law of the knowledge forgetting curve (such as the learned knowledge points become fuzzy over time), and breaks through the limitation of the traditional model that regards the knowledge reserve as a static value; and dynamically updates the learning speed according to the recent learning task completion situation (such as the correct rate of exercises and the learning time length), so that the model can capture the state fluctuation of the learner (such as the learning speed is improved due to high concentration in a certain stage), and the dynamic nature makes the learner's ability characteristics more in line with the actual situation, solving the problem of lagging behind in ability evaluation caused by the fixed learning speed of the traditional model.

[0092] In the step of processing the learner time sequence learning data in the time sequence feature module, key information is extracted from the learner time sequence learning data (such as historical learning records, task completion logs, and timestamp information), including knowledge point learning timestamp, learning task completion quality, and knowledge point initial mastery degree; and original data support is provided for subsequent decay coefficient calculation and learning speed updating, ensuring the pertinence of time sequence features.

[0093] In an exemplary embodiment, for the calculation of the knowledge reserve decay amount, an exponential decay function is adopted, the longer the interval, the smaller the decay coefficient, and the more obvious the knowledge reserve decay; the decay amount calculation is represented as: knowledge reserve decay amount = initial mastery degree x (1-decay coefficient). For example, the learner's initial mastery of a certain knowledge point is 80%, and the decay coefficient is 0.6 after an interval of 7 days, then the current knowledge reserve amount = 80% x 0.6 = 48%, the decay amount is 32%, and finally the real-time knowledge reserve amount feature is obtained, which is used to reflect the learner's actual mastery level of each knowledge point;

[0094] In an embodiment, based on the completion of the recent learning task, the learning speed parameter of the learner is adjusted, such as adopting a sliding window mechanism (such as taking the data of the last 7 days as the window), to avoid the interference of single task anomaly on the speed, and to ensure the stability and timeliness of the learning speed feature;

[0095] The updated knowledge reserve feature and learning speed feature are fed back to the multi-path heterogeneous graph to adjust the node features and edge weights.

[0096] It also includes the steps of:

[0097] In the graph neural network module, the path perception convolutional layer constructs a type conversion matrix for different types of nodes and edges in the multi-path heterogeneous graph, performs convolution operation on different paths respectively, aggregates the features of adjacent nodes, and filters the edges that do not match the current ability of the learner through the edge type mask matrix; the time sequence-structure fusion attention layer introduces a time sequence decay factor to assign dynamic attention weights to the edges of different paths and time intervals; and the feature fusion layer converts the processed heterogeneous node features to a unified space through a cross-modal mapping matrix to obtain node representation.

[0098] S104, using the trained recommendation model, inputting the node representation into the recommendation model, predicting the learner preference, and recommending the educational resources to the learner based on the learner preference.

[0099] In an implementation manner of the present application, as shown in the figure, Figure 3 The recommendation model provided by the embodiment includes a time sequence feature module 301, a heterogeneous graph reconstruction module 302, and a graph neural network module 303, wherein the graph neural network module 303 includes a path-aware convolution layer 3031, a time sequence-structure fusion attention layer 3032, and a feature fusion layer 3033.

[0100] The time sequence feature module 301 is configured to process the learner time sequence learning data, extract key information, introduce a decay coefficient, calculate the decay amount of the knowledge reserve according to the time interval, and dynamically update the learning speed according to the completion of the learner's recent learning task, so as to obtain the knowledge reserve feature and the learning speed feature of the learner; the heterogeneous graph reconstruction module 302 is configured to reconstruct the multi-path heterogeneous graph based on the knowledge reserve feature and the learning speed feature output by the time sequence feature module; the path-aware convolution layer 3031 constructs a type conversion matrix for different types of nodes and edges in the multi-path heterogeneous graph, performs convolution operation on different paths respectively, aggregates the features of adjacent nodes, and filters the edges that do not match the current ability of the learner through an edge type mask matrix.

[0101] Further, in the embodiment, the time sequence-structure fusion attention layer 3032 introduces a time sequence decay factor to assign dynamic attention weights to the edges of different paths and time intervals; and the feature fusion layer 3033 converts the processed heterogeneous node features to a unified space through a cross-modal mapping matrix to obtain the node representation.

[0102] In the step S103 of the embodiment of the present application, through the processing logic of path-aware convolution-time sequence-structure attention fusion-cross-modal feature unification, the limitations of the traditional graph neural network in aspects such as heterogeneous feature adaptation, time sequence and structure correlation split, and cross-modal feature conflict are solved.

[0103] Specifically, the traditional graph neural network adopts a single conversion matrix for convolution operation on the heterogeneous graph, ignoring the characteristic differences of different types of edges such as learner-resource and resource-knowledge point (for example, the former reflects behavioral interaction and the latter reflects content association), while the path-aware convolution layer designs exclusive processing logic for nodes / edges of different paths through a type conversion matrix + an edge type mask matrix, which can accurately aggregate features of the same association and filter invalid edges that do not match the current ability of the learner (for example, when the knowledge reserve is low, the associated edges of high-difficulty knowledge points are automatically shielded), avoiding irrelevant information interference with feature learning.

[0104] In addition, the time-structure fusion attention layer introduces a time decay factor, taking the time interval and path type as the basis for allocating attention weights, so that the model can focus on the paths with strong recent correlations and distinguish the long-term value of different paths.

[0105] The feature fusion layer converts heterogeneous features to a unified space through a cross-modal mapping matrix, which not only preserves the core information of each modality (such as the text semantics of resources and the quantitative values of students' abilities), but also eliminates the differences in dimensions and distributions, so that the node representation can be directly used for subsequent preference prediction, improving the stability of model reasoning.

[0106] In one exemplary embodiment, a separate conversion matrix is constructed for different types of node pairs, and the matrix parameters are learned through training to adapt to the feature dimensions and correlation rules of different node types. In the path convolution operation, convolution is performed separately for each path: taking the path as an example, the resource features are first aggregated with the student node features by converting the resource features; then the knowledge point features are aggregated with the aggregation results of the previous step by converting the knowledge point features, and finally the local feature representation of the path is obtained. In the edge type mask matrix filtering, based on the matching degree of the student's current knowledge reserve and the difficulty of the resource / knowledge point, a mask matrix is generated, and during convolution, the edge features with a mask of 0 are not involved in aggregation, avoiding interference from high-difficulty resource features. In the time-structure fusion attention layer, dynamic attention weights are assigned to the convolved features based on the path type and time interval, strengthening the influence of important correlations.

[0107] In one preferred embodiment,

[0108] Time decay factor design: the same logic as the decay coefficient in step S102 is adopted, and the longer the interval, the smaller the factor value.

[0109] In the path-time attention weight calculation, the weight is determined by the path importance and the time decay factor, where the path importance is obtained based on prior knowledge or training.

[0110] Finally, in the step of weighted aggregation, the convolution features of each path are weighted and summed according to the attention weights to obtain the global feature that integrates the structure correlation strength and the time influence degree. The spatial differences of heterogeneous features are eliminated through cross-modal mapping, and the node representation that can be directly used for prediction is output.

[0111] Further, in the embodiments of the present application, in the path-aware convolution layer, the features of different paths are aggregated differently, and the aggregated feature representation of the student node s is:

[0112]

[0113] wherein, represents the set of resource nodes associated with the student s; and σ represents an activation function. denotes the attention weight of the student-resource edge, denoted as:

[0114]

[0115] wherein, denotes the transformation matrix of the node s, used for transforming the features of the node s; denotes the transformation matrix of the node r, used for transforming the features of the node r; denotes the transformation matrix of the node k, used for transforming the features of the node k; denotes an activation function used for introducing nonlinearity, denotes a learnable weight vector; denotes the feature vector of the node r; denotes the feature vector of the node s; denotes the set of neighbor nodes of the node s; denotes a neighbor node; denotes the feature vector of the neighbor node; denotes the concatenation operation of vectors.

[0116] Further, in the embodiment of the present application, in the time-structure fusion attention layer, a knowledge reserve amount decay coefficient of the time sequence module output is introduced to dynamically adjust the path weight, denoted as: , wherein, denotes the convolution feature corresponding to the path p, denotes the path importance weight; denotes the time sequence decay factor of the path p, , denotes the timestamp of the path p, and δ denotes the knowledge reserve amount decay coefficient.

[0117] Further, in the embodiment of the present application, as shown in Figure 4 , the feature fusion layer converts the processed heterogeneous node features to a unified space through a cross-modal mapping matrix to obtain the node representation, and the steps include:

[0118] S401, constructing a mapping matrix for different node types, including student mapping, resource mapping and knowledge point mapping;

[0119] In step S401, in the student mapping matrix, the input is the dynamic characteristics of the student (knowledge reserve, learning speed, etc. Numerical vector), and the matrix parameter is converted into a feature space suitable for interaction with other modalities through learning; In the resource mapping matrix, the input is the multi-source characteristics of the resources (text vector, difficulty label, type attribute, etc.), and the matrix needs to focus on preserving the semantics of the resources (such as the knowledge point direction of the exercises and the teaching target of the video); For example, the 384-dimensional text vector is mapped to 256 dimensions, while embedding the difficulty coefficient and other metadata characteristics; In the knowledge point mapping matrix, the input is the structured characteristics of the knowledge points (hierarchical coding, correlation strength, etc.), and the matrix needs to strengthen the logical relationship between knowledge points. Map 64-dimensional hierarchical features to 256 dimensions to ensure that the structure of the knowledge system can still be reflected after mapping; The above three types of matrices are independently optimized to avoid interference of feature distribution of different node types, providing the same dimension and interactive basic features for subsequent cross-modal fusion.

[0120] S402, introduce the constraint of adversarial learning, minimize the type discrimination accuracy after mapping as the generator target, maximize the ability to distinguish different types of nodes as the discriminator target, so that the mapped features meet the semantic consistency and type distinguishability;

[0121] In step S402, through the adversarial training of the generator and the discriminator, the contradiction between semantic distortion and type confusion in cross-modal mapping is solved, so that the mapped features not only retain the original semantics, but also can clearly distinguish the node types;

[0122] Generator target: minimize the type discrimination accuracy as the optimization direction, that is, make the generated mapping features as "ambiguous" as possible to the node type difference (such as making the student features and resource features interactive in space), while ensuring that the feature semantics are consistent with the original input (such as the mathematical meaning of the derivative knowledge point feature after mapping);

[0123] Discriminator target: maximize the type discrimination ability as the optimization direction, try to accurately identify the node type from the generator output by learning the discriminant features of the node type (such as the time sequence fluctuation pattern in the student feature and the text semantic identifier in the resource feature);

[0124] When the adversarial training reaches equilibrium, the mapped features will neither lose key semantics due to excessive pursuit of type uniformity (such as dilution of student learning ability features), nor hinder cross-modal interaction due to type isolation (such as the inability of resource and knowledge point features to associate), ultimately forming intermediate features with semantic fidelity and type distinguishability.

[0125] S403, in the mapping process, based on the residual connection mechanism to retain the key information in the original features;

[0126] Further, in step S403, key information in the original features is reserved based on the residual connection mechanism in the mapping process, avoiding information loss caused by feature mapping; specifically, by fusing the residuals of the original features and the mapped features, the problem of forgetting key information (such as losing core weak knowledge points of students after mapping) that may occur in deep mapping is solved; wherein, the original features (such as the knowledge reserve amount vector of students) of the nodes and the features converted by the mapping matrix are added element by element, which realizes cross-modal adaptation through the mapping matrix and ensures the retention of core information in the original features through the residual connection, avoiding feature distortion caused by deviation in optimization of the mapping matrix parameters;

[0127] S404, differentiating weights are given to features of different sources through an attention mechanism, and interaction features between nodes are constructed;

[0128] In step S404, differentiating weights are given to features of different sources through an attention mechanism, and interaction features between nodes are constructed to strengthen the influence weight of important features; based on the contribution of features to preference prediction, weights are dynamically allocated, and the correlation pattern between nodes is captured to generate more discriminative features; for multi-source features of nodes, weights are calculated through an attention mechanism.

[0129] Further, in the construction of interaction features, cross features are generated in combination with the correlation strength between nodes;

[0130] S405, a gating mechanism is introduced to filter effective information, and the fused features processed by the gating mechanism are converted to a unified space through a cross-modal mapping matrix, and finally the node representation is output.

[0131] Specifically, in step S405, in the step of introducing a gating mechanism to filter effective information, the embodiment of the present application adopts a multi-path attention gating mechanism to separately construct gating parameters for learning paths, demand paths and correlation paths in the multi-path heterogeneous graph, and the gating parameters are represented as:

[0132]

[0133]

[0134] In the formula, p∈{learning path, demand path, correlation path}, represents the path weight, represents the knowledge reserve amount change rate; represents the learning speed; represents a sigmoid activation function for mapping input values to between 0 and 1; represents element-wise multiplication; [;] represents a vector splicing operation; represents the original mapping feature; representing the fused features after the gating processing; representing a gating weight matrix; representing a gating signal;

[0135] As Figure 5 shown, in the embodiment of the present application, the step of constructing gating parameters for each path includes:

[0136] S501, the gating signal constructed for each path is generated by nonlinear transformation on the mapping features corresponding to the path, the knowledge reserve change rate and the learning speed, and is used for dynamically balancing the static mapping features and dynamic learning features of a single path;

[0137] In step S501, the proportion of static mapping features and dynamic learning features in a single path is adjusted by the gating signal, so that the path features can not only retain stable properties, but also respond to changes in the real-time learning state of the learner; wherein the mapping features corresponding to the path (static features) refer to the path basic properties after conversion by the type mapping matrix; the knowledge reserve change rate reflects the fluctuation of the learner's mastery of the path-related knowledge points, and embodies the dynamic decay trend; the learning speed is used to represent the efficiency of the learner in processing the path resources in the near future, and reflects the real-time learning state; the gating signal enables the feature representation of a single path to adapt to the learner's state, for example, when the learner has just finished learning a certain knowledge point, the gating signal tends to retain the static knowledge point mapping features; two weeks later, the knowledge reserve decays, and the gating signal automatically increases the proportion of dynamic features, so that the path features are more consistent with the current ability.

[0138] S502, the features of each path after the gating processing are given path weights, wherein the path weights are dynamically set based on the importance of the path to the preference prediction of the learner; including calculating the correlation between the path features and the learner preference label through the attention mechanism of fusing the timing decay factor, and adjusting the path weight in real time; at the same time, in the training stage of the recommendation model, the path weight is iteratively adjusted from a global perspective through the back propagation of the prediction error;

[0139] S503, aggregate the weighted features of all paths to obtain the final fused features.

[0140] In step S502, through the dual mechanism of real-time dynamic adjustment and global iterative optimization, a weight matching the contribution of each path to the preference prediction is allocated to the path, which not only ensures the real-time adaptability, but also realizes the global optimization;

[0141] In the real-time dynamic adjustment mechanism, the correlation between path features and student preference tags is calculated by combining the time-series decay factor, and weights are dynamically allocated. In the time-series decay factor, historical interaction behavior of the path is given higher weight, and distant interaction is given lower weight (the decay factor decreases as the time interval increases), ensuring that the importance of recent paths is given priority. In the correlation calculation, the matching degree between path features and student preference tags is quantified by attention score. The higher the score, the greater the path weight. In each round of forward propagation, the weights are adjusted in real time according to the current student's preference feedback.

[0142] In addition, in the global iterative optimization mechanism, during the recommendation model training phase, the global adjustment of path weights is driven by prediction error: the error between the model's predicted student preferences and the true labels is traced back to the weight parameters of each path, and the contribution of the weights to the error is calculated; the weights of high-contribution paths that continuously reduce prediction error are increased through gradient descent; and the weights of low-contribution paths that aggravate the error are decreased.

[0143] In step S503, the weighted features of each path are integrated into a unified fusion feature, which not only preserves the unique value of each path, but also reflects the difference in their contribution through weights, providing high-quality input for subsequent node representation or preference prediction.

[0144] like Figure 6 As shown, according to another embodiment of this application, an improved graph neural network-based online educational resource recommendation system is provided, comprising the following modules:

[0145] The recommendation model building module 601 is used to generate a multi-path heterogeneous graph based on educational resources, learners' learning abilities, and resource knowledge points. The multi-path heterogeneous graph is introduced into the graph neural network to build a recommendation model for predicting learners' preferences. The recommendation model includes a temporal feature module and a graph neural network module. The edges between nodes in the heterogeneous graph are used to represent the interaction relationship between educational resources, learners' learning abilities, and resource knowledge points.

[0146] The heterogeneous graph reconstruction module 602 is used to process the student's time-series learning data in the time-series feature module, extract key information, introduce attenuation coefficients, calculate the attenuation of knowledge reserves based on time intervals, and dynamically update the learning speed based on the student's recent learning task completion status to obtain the student's knowledge reserve characteristics and learning speed characteristics; and reconstruct the multi-path heterogeneous graph based on the knowledge reserve characteristics and learning speed characteristics output by the time-series feature module.

[0147] The node representation module 603 is configured to, in the graph neural network module, the path-aware convolutional layer constructs a type conversion matrix for different types of nodes and edges in the multi-path heterogeneous graph, performs convolutional operation on different paths respectively, aggregates features of adjacent nodes, and filters edges that do not match the current ability of the student through an edge type mask matrix; a time-structure fusion attention layer introduces a time decay factor to assign dynamic attention weights to edges of different paths and time intervals; and a feature fusion layer converts the processed heterogeneous node features to a unified space through a cross-modal mapping matrix to obtain node representations.

[0148] The resource recommendation module 604 is configured to use the trained recommendation model to take the node representation as an input of the recommendation model, predict a student preference, and recommend an educational resource to the student based on the student preference.

[0149] In an embodiment of the present application, an electronic device is also provided, which includes a processor, a memory, a network interface, an input device and a display screen connected through a system bus. The memory includes a non-volatile storage medium and an internal memory. The non-volatile storage medium of the computer device stores an operating system, and can also store a computer program. When the computer program is executed by the processor, the processor can implement the online education resource recommendation method of improving the graph neural network.

[0150] The internal memory can also store a computer program. When the computer program is executed by the processor, the processor can execute the online education resource recommendation method of improving the graph neural network.

[0151] Another embodiment of the present application provides a computer readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the processor executes the steps of the online education resource recommendation method of improving the graph neural network as described above.

[0152] It should be understood that although each step in the flowchart of each embodiment of the present application is displayed in sequence according to the arrow, these steps are not necessarily executed in sequence according to the arrow. Unless otherwise specified herein, the execution of these steps is not strictly limited in sequence, and these steps can be executed in other sequences. Moreover, at least a part of the steps in each embodiment can include multiple sub-steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution sequence of these sub-steps or stages is not necessarily sequential, but can be executed alternately or alternately with at least a part of other steps or sub-steps or stages of other steps.

[0153] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The program can be stored in a non-volatile computer readable storage medium, and when the program is executed, it can include the processes of the above-mentioned embodiments of the online education resource recommendation method for improving the graph neural network.

[0154] Any reference to memory, storage, database, or other medium used in the embodiments provided by the present application can include non-volatile and volatile memory. Non-volatile memory can include read-only memory, programmable ROM, electrically programmable ROM, electrically erasable programmable ROM, or flash memory. Volatile memory can include random access memory or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM, dynamic RAM, synchronous DRAM, double data rate SDRAM, enhanced SDRAM, synchronous link DRAM, memory bus direct RAM, direct memory bus dynamic RAM, and memory bus dynamic RAM.

[0155] Any combination of the technical features of the above-mentioned embodiments can be combined. In order to make the description concise, all possible combinations of the technical features in the above-mentioned embodiments are not described, but as long as the combination of the technical features does not exist, it should be considered as the scope of the present application.

[0156] The above-mentioned embodiments only express several embodiments of the present application, and the description is more specific and detailed, but it should not be understood as limiting the scope of the present application. It should be pointed out that for ordinary skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are within the scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

[0157] The above-mentioned is only the preferred embodiment of the present application, and does not limit the present application. Any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. An improved graph neural network-based method for recommending online educational resources, characterized in that, Includes the following steps: Based on educational resources, learners' learning abilities, and resource knowledge points, a multi-path heterogeneous graph is generated. This graph is then introduced into a graph neural network to construct a recommendation model for predicting learners' preferences. The recommendation model includes a temporal feature module and a graph neural network module. The edges between nodes in the heterogeneous graph represent the interaction relationships between educational resources, learners' learning abilities, and resource knowledge points. In the time series feature module, the time series learning data of students is processed, key information is extracted, a decay coefficient is introduced, the decay of knowledge reserves is calculated according to the time interval, and the learning speed is dynamically updated according to the completion of recent learning tasks of students, so as to obtain the knowledge reserve characteristics and learning speed characteristics of students. The multi-path heterogeneous graph is reconstructed based on the knowledge reserve feature and learning speed feature output by the time series feature module; In the graph neural network module, the path-aware convolutional layer constructs a type conversion matrix for different types of nodes and edges in a multi-path heterogeneous graph, performs convolution operations on different paths, aggregates the features of adjacent nodes, and filters edges that do not match the trainee's current ability through the edge type mask matrix; in the temporal-structure fusion attention layer, a temporal decay factor is introduced to assign dynamic attention weights to edges of different paths and time intervals. The feature fusion layer transforms the processed heterogeneous node features into a unified space through a cross-modal mapping matrix to obtain node representations; Using a trained recommendation model, node representations are used as input to predict student preferences, and educational resources are recommended to students based on these preferences.

2. The online educational resource recommendation method based on the improved graph neural network according to claim 1, characterized in that, Learner learning ability includes knowledge reserves and learning speed. By acquiring learner learning data within a target time period, an assessment model is constructed to measure learner learning ability. The process of constructing this assessment model includes: The knowledge reserve of students is calculated based on their learning performance within the target time period; learning performance is quantified by completion rate, quiz accuracy, and video viewing completion rate; the knowledge reserve is expressed as: In the formula, S represents the student's knowledge reserves, t represents the time point, i represents the i-th learning task, n represents the total number of learning tasks, w represents the weight of task i, and P represents the student's learning performance on task i. The learning speed of a student is calculated by comparing their knowledge reserves at different points in time; where learning speed is expressed as: ； In the formula, V represents the student's learning speed. , These represent the time periods of the trainees. and time The amount of knowledge reserves; Based on students' historical knowledge reserves and learning speed, linear regression is used to calculate students' learning ability at a future point in time, constructing a first model to measure students' learning ability, expressed as: ;in, represents the learning ability model; a and b represent the parameters of the linear regression model; This indicates a future point in time, and is a predicted time point.

3. The improved graph neural network-based online education resource recommendation method according to claim 2, characterized in that, The steps for generating a multi-path heterogeneous graph based on educational resources, learners' cognitive abilities, and resource knowledge points include: Acquire multi-source heterogeneous data, including student behavior logs, educational resource metadata, and knowledge point hierarchical structure; A heterogeneous graph structure is constructed, comprising a student node layer, a resource node layer, and a knowledge point node layer. In the student node layer, each node corresponds to a unique student ID and carries a learning ability feature vector output by the first model. In the resource node layer, subcategories are divided by type, and node features are fused with resource text vectors and metadata. In the knowledge point node layer, directed edges are established between nodes based on the hierarchical relationship of the subject knowledge graph. Construct edges between nodes, including student-resource edges based on student behavior logs; resource-knowledge point edges based on the matching degree between resource content and knowledge points; and student-knowledge point edges based on students' mastery of knowledge points. Based on the constructed heterogeneous graph, a weighted path network is formed based on multiple path types to obtain a multi-path heterogeneous graph. The multiple path types include learning paths, demand paths, and association paths. The learning path is used to reflect the trajectory of learners acquiring knowledge through resources, and the path weight is calculated by multiplying the edge weights and adding a time decay factor. The demand path is used to characterize the matching relationship between learners' abilities and knowledge point requirements, and the importance weight of knowledge points is introduced into the path. The association path is used to explore the potential connections between knowledge points, and the path weight is calculated based on the cross-knowledge point association degree matrix of resources.

4. The improved graph neural network-based online educational resource recommendation method according to claim 3, characterized in that, In the path-aware convolutional layer, features from different paths are aggregated differentially. The aggregated features of student node s are represented as follows: ； In the formula, σ represents the set of resource nodes associated with node s; σ represents the activation function. The attention weight of the student-resource edge is represented as: ； In the formula, This represents the transformation matrix of node s, used to transform the features of node s; This represents the transformation matrix of node r, used to transform the features of node r; This represents the transformation matrix of node k, used to transform the features of node k; This represents the activation function used to introduce nonlinearity. This represents a learnable weight vector; The feature vector of node r; The feature vector of node s; Indicate neighboring nodes; Represents the feature vector of a neighboring node; This represents the vector concatenation operation.

5. The improved graph neural network-based online education resource recommendation method according to claim 4, characterized in that, In the temporal-structural fusion attention layer, a knowledge reserve decay coefficient from the temporal module output is introduced to dynamically adjust the path weights, expressed as: ； ； In the formula, This represents the convolutional feature corresponding to path p. Indicates the importance weight of the path; This represents the time decay factor of path p. δ represents the timestamp of path p, and δ represents the knowledge reserve decay coefficient.

6. The improved graph neural network-based online educational resource recommendation method according to claim 5, characterized in that, In the feature fusion layer: Construct mapping matrices for different node types, including student mapping, resource mapping, and knowledge point mapping; By introducing adversarial learning constraints, the generator aims to minimize the accuracy of type discrimination after mapping, while the discriminator aims to maximize the ability to distinguish different types of nodes, so that the mapped features satisfy semantic consistency and type distinguishability. During the mapping process, key information in the original features is preserved based on the residual connection mechanism; Differentiated weights are assigned to features from different sources through an attention mechanism, and interaction features between nodes are constructed. A gating mechanism is introduced to filter valid information. The fused features processed by gating are transformed into a unified space through a cross-modal mapping matrix, and the final output is the node representation.

7. The improved graph neural network-based online education resource recommendation method according to claim 6, characterized in that, In the step of introducing a gating mechanism to filter effective information, a multi-path attention gating mechanism is used to construct gating parameters separately for the learning path, demand path, and association path in the multi-path heterogeneous graph. The gating parameters are expressed as follows: ； ； In the formula, p∈{learning path, demand path, association path}, Indicates path weight, This represents the rate of change in the amount of knowledge stored. Indicates learning speed; This represents the sigmoid activation function, used to map input values ​​to the range of 0 to 1; [;] indicates element-wise multiplication; [;] indicates vector concatenation. Represents the original mapping features; This indicates the features after gating and fusion. Represents the gate weight matrix; This indicates a gating signal.

8. The improved graph neural network-based online education resource recommendation method according to claim 7, characterized in that, The step of constructing gating parameters for each path individually includes: The gate signal constructed for each path is generated by nonlinear transformation of the mapping features, the rate of change of knowledge reserves, and the learning speed corresponding to the path, and is used to dynamically balance the static mapping features and dynamic learning features of a single path. Each path is assigned a path weight after gating, and the path weight is dynamically set based on the importance of the path to the prediction of student preferences. This includes calculating the correlation between path features and student preference labels through an attention mechanism that incorporates a time decay factor, and adjusting the path weight in real time. At the same time, during the training phase of the recommendation model, the path weight is iteratively adjusted from a global perspective through backpropagation of the prediction error. The weighted features of all paths are aggregated to obtain the final fused features.

9. A recommendation system for implementing an improved graph neural network-based online educational resource recommendation method as described in any one of claims 1 to 8, characterized in that, The recommendation system includes the following modules: The recommendation model building module is used to generate a multi-path heterogeneous graph based on educational resources, learners' learning abilities, and resource knowledge points. The multi-path heterogeneous graph is introduced into the graph neural network to build a recommendation model for predicting learners' preferences. The recommendation model includes a temporal feature module and a graph neural network module. The edges between nodes in the heterogeneous graph are used to represent the interaction relationships between educational resources, learners' learning abilities, and resource knowledge points. The heterogeneous graph reconstruction module is used in the temporal feature module to process the student's temporal learning data, extract key information, introduce attenuation coefficient, calculate the attenuation of knowledge reserve based on time interval, and dynamically update the learning speed based on the student's recent learning task completion status, so as to obtain the student's knowledge reserve characteristics and learning speed characteristics. The multi-path heterogeneous graph is reconstructed based on the knowledge reserve feature and learning speed feature output by the time series feature module; The node representation module is used in the graph neural network module. The path-aware convolutional layer constructs a type conversion matrix for different types of nodes and edges in the multi-path heterogeneous graph, performs convolution operations on different paths, aggregates the features of adjacent nodes, and filters edges that do not match the trainee's current ability through the edge type mask matrix. The temporal-structure fusion attention layer introduces a temporal decay factor to assign dynamic attention weights to edges of different paths and time intervals. The feature fusion layer transforms the processed heterogeneous node features into a unified space through a cross-modal mapping matrix to obtain node representations; The resource recommendation module uses a trained recommendation model, taking node representations as input, to predict student preferences and recommend educational resources to students based on those preferences.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, wherein the computer program, when executed by a processor, implements the steps of the online educational resource recommendation method of the improved graph neural network as described in any one of claims 1 to 8.

Images