A course recommendation method, system and medium based on cross-graph enhancement

Abstract

The present invention discloses a course recommendation method, system and medium based on cross-graph enhancement, the method comprising the following steps: generating a student-course graph, a student-course feature graph and a student group graph; constructing a comparative learning model, and generating each embedded node in an embedding layer; obtaining a course node embedding representation and a student node embedding representation through a main view channel, and selecting a course recommendation list; obtaining a student node embedding representation of a student-course feature graph through a static feature channel; obtaining a final representation of a student node based on a student node embedding representation and a student node embedding representation of a student-course feature graph; obtaining a student node embedding representation of a student group graph through a student group channel, and aggregating the student node embedding representation with the final representation to obtain a student global representation, and performing comparative learning between the student global representation and the student node embedding representation; constructing a loss function training comparative learning model, and obtaining a course recommendation result. The present invention provides students with more accurate and efficient personalized course recommendations.

Description

Technical Field

[0001] The present invention relates to the technical field of course recommendation, and in particular to a course recommendation method, system and medium based on cross-graph enhancement. Background Art

[0002] At present, online education has become an important part of daily educational activities, and a large number of course resources have poured into online education platforms. However, this has also brought about problems such as uneven course quality and difficulty in effectively helping students grow. The explosive growth of course information makes it difficult for students to quickly find the learning content they are really interested in among the massive resources; at the same time, the disorder, randomness and complexity of course resources have not only consumed students' time and energy, but also buried valuable online course resources.

[0003] In this context, personalized course recommendations came into being, aiming to help students find course resources that suit their interests and needs more quickly and accurately. There are many personalized recommendation methods currently applied in the field of courses. For example, in order to ensure the accuracy and timeliness of course recommendations, a hierarchical and phased attention network model is used to model the interaction sequence between users and courses according to the dynamic changes in user interests to recommend course resources; for example, personalized recommendations based on learner knowledge and personality, which adaptively integrates the learner's knowledge level, sequence behavior and learner's personality to model the learner's portrait, and a personalized recommendation plan for student courses; for example, by constructing the similarity between users, the degree of match between users and courses, and the characteristics of user groups of similar courses, a course recommendation plan is analyzed from multiple angles to achieve a course recommendation plan that is closer to the actual needs of users; for example, by obtaining the user's course learning status information, predicting the user's theoretical course learning purpose, and recommending to the user an online course recommendation method for learning courses to achieve the theoretical course learning purpose; for example, a deep learning convolutional neural network model is built through course resource evaluation text data, the convolutional neural network model is iteratively trained for multiple rounds using a training set, the convolutional neural network model is tested using a test set, the prediction effect of the convolutional neural network model is tested based on the MSE mean square error, and the model parameters are adjusted until the model converges to obtain the desired convolutional neural network model.

[0004] However, current personalized recommendation methods often rely too much on users' historical behavior data or attribute characteristics, without taking into account the limited number of interactions between students and courses in online education and the scarcity of interaction data. For example, freshmen who have just entered college may not understand the course content, have no course selection records, or have not determined their interests. Different from common music recommendations and news recommendations, there is an obvious user cold start problem in course resource recommendations, which affects the construction effect of the recommendation model. Secondly, existing recommendation algorithms usually only consider information at the individual user level, but ignore the social relationships and mutual influences between users. In real life, students' course selection decisions are often influenced by classmates, teachers or other people on social media. Moreover, existing recommendation systems usually lack effective evaluation of course quality. Even if they can accurately predict users' interests, such recommendations are meaningless if the recommended courses are of low quality or do not meet teaching standards.

[0005] Therefore, existing personalized course recommendation methods face challenges such as data scarcity, neglected social network influence, and over-reliance on multimodal data. These problems limit the accuracy and universality of recommendation models. In the current online education context, how to help students quickly and accurately find resources that meet their needs from massive courses and reduce dependence on multimodal data to achieve smarter and more personalized course recommendations is a technical problem that needs to be solved urgently. Summary of the invention

[0006] In order to overcome the defects and shortcomings of the prior art, the present invention provides a course recommendation method based on cross-graph enhancement. The present invention can fully capture students' interest preferences through the powerful processing ability of contrastive learning on sparse interactive data and the advantages of deep learning in feature extraction and similarity matching, and overcome the data sparsity and cold start problems in personalized recommendation of educational resources, so as to provide students with more accurate and efficient personalized recommendations in course recommendation scenarios.

[0007] The second object of the present invention is to provide a course recommendation system based on cross-graph enhancement.

[0008] A third object of the present invention is to provide a storage medium.

[0009] In order to achieve the above object, the present invention adopts the following technical solutions:

[0010] The present invention provides a course recommendation method based on cross-graph enhancement, comprising the following steps:

[0011] Generate a student-course graph G0 based on the interaction history between students and courses, and generate a student-course feature graph G based on the connection between students and course features. L , based on the students’ descriptive information, generate the student group graph Gs ;

[0012] Constructing a contrastive learning model, including an embedding layer, a main view channel, a static feature channel, and a student group channel; the embedding layer generates student nodes, course nodes, and course static feature nodes based on random initialization;

[0013] The main view channel performs graph convolution and aggregation operations on the student-course graph G0 to obtain the course node embedding representation and student node embedding representation , and perform inner product operation to generate similarity scores, and select the top k k The courses are listed as course recommendations;

[0014] The static feature channel is used for each student-course feature graph G L Perform graph convolution and aggregation operations to obtain the student node embedding representation of each subgraph in the student-course feature graph ;

[0015] Embedding the student node into And the student node embedding representation of each subgraph in the student-course feature graph Input the fully connected feature extraction layer to get the final representation of the student node ;

[0016] The student group channel to the student group graph G s Perform graph convolution and aggregation operations to obtain the student group graph G s The student node embedding representation in ;

[0017] The final representation of the student node and student node embedding representation Aggregate to get the student's global representation , embedding the student node into Global representation with students Conduct comparative learning;

[0018] Construct a loss function to train a contrastive learning model, and obtain course recommendation results based on the trained contrastive learning model.

[0019] As a preferred technical solution, the main view channel performs graph convolution and aggregation operations on the student-course graph G0 to obtain the course node embedding representation and student node embedding representation , the specific steps include:

[0020] ;

[0021] ;

[0022] in, 、 Respectively expressed in r In the +1-layer student-course graph G0, the student node u i and course nodes c j The result after aggregating and propagating the information of adjacent nodes, , Respectively expressed in r In the student-course graph G0, the student node u i and course nodes c j The result after aggregating and propagating the information of adjacent nodes, and For student nodes u i and course nodes c j The degree in the student-course graph G0, and For student nodes u i and course nodes c j Aggregation of adjacent nodes in the student-course graph G0;

[0023] Aggregate the node embedding representations of all layers in the student-course graph G0 to generate student nodes u i and course nodes c j Embedding representation of course nodes in the student-course graph G0 and student node embedding representation , as the student local representation in contrastive learning.

[0024] As a preferred technical solution, a graph convolution encoder is used for graph convolution, and the feature transformation matrix, nonlinear activation method and initial layer embedding are removed;

[0025] The course node embedding representation and student node embedding representation Specifically expressed as:

[0026] ;

[0027] ;

[0028] in, , , Represents the corresponding weight value, , , Respectively represent the student nodes in the 1st, 2nd, and 3rd layer student-course graph G0. u i The result after aggregating and propagating the information of adjacent nodes, , , Respectively represent the course nodes in the 1st, 2nd, and 3rd layer student-course graph G0. c j The result after aggregating and propagating neighboring node information.

[0029] As a preferred technical solution, the similarity score is used to predict the student node in the student-course graph G0 u i and course nodes c j The weight of the edge between , specifically expressed as:

[0030] .

[0031] As a preferred technical solution, the final representation of the student node The specific calculation process is expressed as:

[0032] ;

[0033] ;

[0034] in, represents the nonlinear activation function tanh, , , …, is a trainable matrix for adaptively extracting personal preferences, Represents a trainable matrix used to reduce the dimensionality of the connected node representation.

[0035] As a preferred technical solution, students globally expressed The specific calculation process is expressed as:

[0036] ;

[0037] in, Represents an aggregation operation.

[0038] As a preferred technical solution, the loss function is constructed to train the contrastive learning model. The course recommendation adopts Bayesian personalized ranking loss, and the contrastive learning adopts information noise contrast estimation loss. The total loss function is expressed as:

[0039] ;

[0040] ;

[0041] ;

[0042] in, , Represents the corresponding weight value, Used to control the strength of contrastive learning, To control total loss L The regularization strength of represents the set of model parameters, including the trainable parameter matrix of the fully connected feature extraction layer and the initialization embedding of all nodes, represents the Bayesian personalized ranking loss, Indicates students u Interactive courses i The positive sample of Indicates students u and j Negative samples where no interaction occurs in the course, Indicates students u The set of all courses that have been interacted with in the current data, Represents the similarity score of the positive sample pair, Represents the similarity score of the negative sample pair, represents the Sigmoid activation function, represents the information noise contrast estimation loss, and Represent the same node i The global and local representations of represents a negative sample node, Represents the parameters of the Softmax function.

[0043] In order to achieve the above second purpose, the present invention adopts the following technical solutions:

[0044] A course recommendation system based on cross-graph enhancement, used to implement the above-mentioned course recommendation method based on cross-graph enhancement, the system comprises: a data graph construction module, a contrastive learning model construction module, a contrastive learning model training module, and a course recommendation result output module;

[0045] The data graph construction module is used to generate a student-course graph G0 based on the interaction history between students and courses, and to generate a student-course feature graph G0 based on the connection between students and course features. L , based on the students’ descriptive information, generate the student group graph G s ;

[0046] The contrastive learning model construction module is used to construct a contrastive learning model, including an embedding layer, a main view channel, a static feature channel and a student group channel; the embedding layer generates student nodes, course nodes and course static feature nodes based on random initialization;

[0047] The main view channel performs graph convolution and aggregation operations on the student-course graph G0 to obtain the course node embedding representation and student node embedding representation , and perform inner product operation to generate similarity scores, and select the top k k The courses are listed as course recommendations;

[0048] The static feature channel is used for each student-course feature graph G L Perform graph convolution and aggregation operations to obtain the student node embedding representation of each subgraph in the student-course feature graph ;

[0049] Embedding the student node into and student node embedding representation Input the fully connected feature extraction layer to get the final representation of the student node ;

[0050] The student group channel to the student group graph G s Perform graph convolution and aggregation operations to obtain the student group graph G s The student node embedding representation in ;

[0051] The final representation of the student node and student node embedding representation Aggregate to get the student's global representation , embedding the student node into Global representation with students Conduct comparative learning;

[0052] The contrastive learning model training module is used to construct a loss function to train a contrastive learning model;

[0053] The course recommendation result output module is used to obtain course recommendation results based on the trained contrastive learning model.

[0054] In order to achieve the third objective, the present invention adopts the following technical solutions:

[0055] A computer-readable storage medium stores a program, which, when executed by a processor, implements the above-mentioned course recommendation method based on cross-graph enhancement.

[0056] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0057] (1) In view of the difficulty of modeling comprehensive interest preference portraits of users with information overload and cold data, the present invention can fully capture students' interest preferences through the powerful processing ability of contrastive learning on sparse interactive data, and the advantages of deep learning in feature extraction and similarity matching, and overcome the data sparsity and cold start problems in personalized recommendation of educational resources, thus providing students with more accurate and efficient personalized recommendations in course recommendation scenarios.

[0058] (2) The contrastive learning model EduLGCL of the present invention uses the clustering relationship of student groups and the static characteristics of courses to expand the relationship between student nodes, so that the model has better performance in the user cold start problem than models such as LightGCN that use traditional collaborative filtering methods;

[0059] The contrastive learning model EduLGCL of the present invention adopts a joint learning mode that combines recommendation tasks with contrastive learning. Compared with the model that does not adopt the contrastive learning mode, the contrastive learning model EduLGCL performs better in processing long-tail problems, thereby alleviating the long-tail problems that commonly occur in recommendation scenarios.

[0060] The contrastive learning model EduLGCL of the present invention adopts a bidirectional interaction graph such as a student-course interaction graph for data representation to save data space, extracts effective information from the global feature representation to enhance the local feature representation, and effectively alleviates the data sparsity problem.

[0061] (3) The present invention uses the EduLGCL model to generate a recommended list of courses for students to choose based on their personal interests and the interaction between them and the courses, and displays it to the users. This can alleviate the current problems in colleges and universities, such as the difficulty for students to choose courses and the burial of some high-quality course resources.

[0062] (4) The present invention introduces non-traditional student user characteristics, such as college entrance examination subject selection, professional background, interest survey, etc., combined with limited historical behavior data to jointly construct a user portrait; at the same time, the graph convolution operation recommends courses for new users based on behaviors similar to those of student users, which can effectively deal with the student user cold start problem and provide reasonable recommendations even for new students or users with limited interaction data.

[0063] (5) The present invention integrates social network analysis technology into the course recommendation method, adding information such as majors and classes, such as a social network propagation model based on graph theory, to simulate and predict the information flow and decision-making influence between users; at the same time, it uses user interaction data on social media and online learning platforms to quantify the strength of social relationships between users, which can fully consider the social relationships and mutual influences between users, making the recommendation results more in line with the users' real needs and preferences. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 It is a flowchart of the course recommendation method based on cross-graph enhancement of the present invention;

[0065] Figure 2 Schematic diagram of the overall architecture of the comparative learning model EduLGCL of the present invention. DETAILED DESCRIPTION

[0066] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0067] Example 1

[0068] like Figure 1 As shown, this embodiment provides a course recommendation method based on cross-graph enhancement, comprising the following steps:

[0069] S1: Generate a student-course graph G0 to represent the interaction history between students and courses. y ui Represents the interaction between students and courses, y ui >0 means that there is an interaction between students and courses. y ui = 0 means no interaction was observed;

[0070] Generate course-course static feature graph G c , used to record course features and the relationship between courses. All courses are divided into several groups with different indirect nodes according to their static feature types. Using these indirect nodes, a student-course feature graph G that reflects the relationship between students and course features can be obtained. L ;

[0071] Generate a student group graph G based on the student's descriptive information (such as grade, gender, etc.) s (i.e., student-student graph G s );

[0072] In this embodiment, three graphs are constructed using multiple types of data from different sources (self-collected, public data sets, etc.). The multi-source data are cleaned, integrated, and standardized. The data from different sources include user behavior data, attribute characteristics, social data, course quality evaluation, etc., to build a unified data view. Then, these data are modeled and analyzed to generate personalized recommendation results and provide more comprehensive and accurate recommendations.

[0073] S2: Construct the contrastive learning model EduLGCL, such as Figure 2 As shown in Figure 2, the overall structure of the model includes: embedding layer, main view channel, static feature channel and student group channel;

[0074] The embedding layer provides the embedding representation of students and each course feature shared by all views and can be used for training based on the random initialization method.

[0075] The embedding layer vectorizes the data, maps students, courses, and feature labels into a vector using a unique identifier, and initializes the embedding lookup table to the corresponding embedding representation. These embedding representations are , and , respectively represent student nodes u i , Course Node c j and course static feature nodes , all student nodes and course nodes are respectively and It indicates that, n and m Represent the number of students and the number of courses respectively, d Indicates the dimensionality of the embedding representation.

[0076] Local representation learning is performed in the main view channel: three graph convolution operations are performed on the student-course graph G0 to propagate adjacency information, and the first-order and second-order adjacent node information of the node is learned to obtain the local representation of the student node and the course feature node. The model is designed using a graph convolution encoder, and the feature transformation matrix, nonlinear activation method and initial layer embedding are removed.

[0077] In this embodiment, non-traditional student features are introduced, such as college entrance examination subjects, professional background, interest surveys, etc., combined with limited historical behavior data to jointly construct a student user portrait; at the same time, the graph convolution operation recommends courses for new student users through behaviors similar to student users, which can effectively deal with the student user cold start problem and provide reasonable recommendations even for new students or users with limited interaction data.

[0078] In this embodiment, the graph convolution process can be expressed in the following form:

[0079] ;

[0080] ;

[0081] in, 、 Respectively expressed inr In the +1-layer student-course graph G0, the student node u i and course nodes c j The result after aggregating and propagating the information of adjacent nodes; and For student nodes u i and course nodes c j degree in the student-course graph G0; and For student nodes u i and course nodes c j Aggregation of adjacent nodes in the student-course graph G0. Aggregate the node embedding representations of all layers in the student-course graph G0 to generate student nodes in the main view channel. u i and course nodes c j Embedding representation of course nodes in the student-course graph G0 and student node embedding representation , as the student local representation in contrastive learning, the calculation result is expressed in the following form:

[0082] ;

[0083] ;

[0084] To simplify the calculation process, the weights in the model are set as: , for student nodes u i and course nodes c j After the inner product operation is performed on the final representation of , a similarity score is generated to predict the student node in the student-course graph G0. u i and course nodes c j The weight of the edge between , indicating the predicted preference of students for the course, and weighting them according to the generated similarity scores Sort the values ​​and select the first k Courses as a recommended list of courses ( k The value of can be set freely), the calculation result is expressed in the following form:

[0085] ;

[0086] Global representation learning: to reduce the common phenomenon of feature blurring and transition smoothing in graph convolution calculations, including:

[0087] By calculating the mutual information value between student and course features, we can obtain the appropriate feature combination. The statistical data is plotted as a mutual information value heat map. The results show that the feature combinations with the largest and smallest mutual information values ​​are least affected by the feature ambiguity problem, and the calculation effect of selecting these combinations is also the best.

[0088] In the static feature channel, each student-course feature graph uses a three-layer graph convolution encoder to learn and aggregate the high-order adjacent node relationships of each node, that is, each student-course feature graph performs three graph convolution operations. The model refers to SocialLGN and designs an information extraction mechanism. The final representation of the student node and the course feature node is aggregated using the fully connected feature extraction layer to generate the final global representation of the student node. The calculation results are as follows:

[0089] ;

[0090] ;

[0091] in, represents the nonlinear activation function tanh, , , …, is a trainable matrix for adaptively extracting personal preferences, is a trainable matrix used to reduce the dimensionality of the connected node representation. For student nodes u i In the student-course feature subgraph G Cp The embedding representation in is, C p represents the number of student-course feature subgraphs, and Input the fully connected feature extraction layer together to get , For student nodes u i The final representation of Excessive phenomenon.

[0092] In addition, the connections between students, especially the interactions between freshmen groups, can also effectively reflect the students' interests. This embodiment integrates social network analysis technology and adds information such as majors and classes, such as a social network propagation model based on graph theory, to simulate and predict the information flow and decision-making influence between students. At the same time, it uses student interaction data on social media and online learning platforms to quantify the strength of social relationships between students, which can fully consider the social relationships and mutual influence between students, so that the recommendation results are more in line with the students' real needs and preferences.

[0093] In the student group channel, for the student group graph G s Three rounds of graph convolution operations are also performed to reveal the first-order direct connections and second-order connections between students, and then these connection information is combined with the final global representation of student nodes. Specifically, and Aggregate to obtain the student global representation for contrastive learning , the calculation process is expressed as:

[0094] ;

[0095] in, Represents the student group graph G s Obtained student nodes u i Embedded representation of u i and information from its first- and second-order neighbors, represents an aggregation operation. To simplify calculation, the aggregation operation in this embodiment takes the average value.

[0096] S3: Construct a loss function to train a contrastive learning model, and obtain course recommendation results based on the trained contrastive learning model;

[0097] In this embodiment, the contrastive learning model adopts a joint learning method to construct a joint learning mode that combines the recommendation task with the contrastive learning task, extracts effective information from the global feature representation to enhance the local feature representation, and alleviates the data sparsity problem. Therefore, in different views, the application of the loss function used to optimize the model parameters is also different. The main view performs the recommendation task, using Bayesian Personalized Ranking Loss (BPR Loss) as the loss function, and using BPR Loss to optimize student nodes and course nodes; the contrast view performs the contrastive learning task, using Infonoise contrastive estimation Loss (InfoNCE Loss) as the loss function.

[0098] The total loss of the joint learning model is expressed as:

[0099] ;

[0100] in, Control the intensity of contrastive learning, control L The regularization strength of Represents a collection of model parameters, which include the trainable parameter matrix of the fully connected layer and the initialized embeddings of all nodes;

[0101] The formula for BPR Loss is as follows:

[0102] ;

[0103] In the training of mini-batch data, the observed student-course interactions are regarded as positive samples, and the unobserved interactions are regarded as negative samples. Indicates students u Interactive courses i The sample is the positive sample; Indicates students u and j The samples where no interaction occurs in the course are called negative samples. Indicates students u The set of all classes that are interacted with in the current mini-batch, Represents the similarity score of the positive sample pair, Represents the similarity score of the negative sample pair, Represents the Sigmoid activation function;

[0104] After the introduction of InfoNCE Loss, the distance between the local representation and the global representation of the same node in the embedding space is reduced, while the distance between different nodes is increased, so that the local representation used for the recommendation task can extract useful information from the global representation and alleviate the data sparsity problem. The formula of InfoNCE Loss is as follows:

[0105] ;

[0106] in, and Represent the same node i The global and local representations of Represents a negative sample node, that is, other nodes of the same type that do not include the node itself in the current mini-batch data. Represents the parameters of the Softmax function, which is used to control the degree of participation of negative samples in the training process.

[0107] Example 2

[0108] This embodiment provides a course recommendation system based on cross-graph enhancement, which is used to implement the course recommendation method based on cross-graph enhancement in the above embodiment 1. The system includes: a data graph construction module, a contrastive learning model construction module, a contrastive learning model training module, and a course recommendation result output module;

[0109] In this embodiment, the data graph construction module is used to generate a student-course graph G0 based on the interaction history between students and courses, and to generate a student-course feature graph G0 based on the connection between students and course features. L , based on the students’ descriptive information, generate the student group graph G s ;

[0110] In this embodiment, the contrastive learning model construction module is used to construct a contrastive learning model, including an embedding layer, a main view channel, a static feature channel, and a student group channel;

[0111] The embedding layer generates student nodes, course nodes, and course static feature nodes based on random initialization;

[0112] The main view channel performs graph convolution and aggregation operations on the student-course graph G0 to obtain the course node embedding representation and student node embedding representation , and perform inner product operation to generate similarity scores, and select the top k k The courses are listed as course recommendations;

[0113] Static feature channel for each student-course feature graph G L Perform graph convolution and aggregation operations to obtain the student node embedding representation of each subgraph in the student-course feature graph ;

[0114] Embedding the student node into And the student node embedding representation of each subgraph in the student-course feature graph Input the fully connected feature extraction layer to get the final representation of the student node ;

[0115] Student Group Channel to Student Group Graph G s Perform graph convolution and aggregation operations to obtain the student group graph G s The student node embedding representation in ;

[0116] The final representation of the student node and student node embedding representation Aggregate to get the student's global representation , embedding the student node into Global representation with students Conduct comparative learning;

[0117] In this embodiment, the contrastive learning model training module is used to construct a loss function to train the contrastive learning model;

[0118] In this embodiment, the course recommendation result output module is used to obtain the course recommendation result based on the trained contrastive learning model.

[0119] Example 3

[0120] This embodiment provides a storage medium, which may be a storage medium such as ROM, RAM, disk, or CD. The storage medium stores one or more programs. When the program is executed by a processor, the course recommendation method based on cross-graph enhancement of Embodiment 1 is implemented.

[0121] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.

Claims

1. A course recommendation method based on cross-graph enhancement, characterized in that: The steps include: Generate a student-course graph G0 based on the interaction history between students and courses, and generate a student-course feature graph G based on the connection between students and course features. L , based on the students’ descriptive information, generate the student group graph G s ; Constructing a contrastive learning model, including an embedding layer, a main view channel, a static feature channel, and a student group channel; the embedding layer generates student nodes, course nodes, and course static feature nodes based on random initialization; The main view channel performs graph convolution and aggregation operations on the student-course graph G0 to obtain the course node embedding representation and student node embedding representation , j Indicates j course nodes, c represents the course, i Indicates i student nodes, u represents students, and performs inner product operation to generate similarity scores, and selects the top students based on top-k k The courses are listed as course recommendations; The static feature channel is used for each student-course feature graph G L Perform graph convolution and aggregation operations to obtain the student node embedding representation of each subgraph in the student-course feature graph , , C p Represents the number of student-course feature subgraphs; Embedding the student node into And the student node embedding representation of each subgraph in the student-course feature graph Input the fully connected feature extraction layer to get the final representation of the student node ; The student group channel to the student group graph G s Perform graph convolution and aggregation operations to obtain the student group graph G s The student node embedding representation in ; The final representation of the student node and student node embedding representation Aggregate to get the student's global representation , embedding the student node into Global representation with students Conduct comparative learning; Construct a loss function to train a contrastive learning model, and obtain course recommendation results based on the trained contrastive learning model.

2. The course recommendation method based on cross-graph enhancement according to claim 1 is characterized in that: A graph convolution encoder is used for graph convolution, and the feature transformation matrix, nonlinear activation method and initial layer embedding are removed; The course node embedding representation and student node embedding representation Specifically expressed as: ； ； in, , , Represents the corresponding weight value, , , Respectively represent the student nodes in the 1st, 2nd, and 3rd layer student-course graph G0. u i The result after aggregating and propagating the information of adjacent nodes, , , Respectively represent the course nodes in the 1st, 2nd, and 3rd layer student-course graph G0. c j The result after aggregating and propagating neighboring node information.

3. The course recommendation method based on cross-graph enhancement according to claim 1 is characterized in that: The similarity score is used to predict the student nodes in the student-course graph G0 u i and course nodes c j The weight of the edge between , specifically expressed as: 。 4. The course recommendation method based on cross-graph enhancement according to claim 1 is characterized in that: Final representation of student nodes The specific calculation process is expressed as: ； ； in, represents the nonlinear activation function tanh, , , …, is a trainable matrix for adaptively extracting personal preferences, Represents a trainable matrix used to reduce the dimensionality of the connected node representation.

5. The course recommendation method based on cross-graph enhancement according to claim 1 is characterized in that: Student Global Representation The specific calculation process is expressed as: ； in, Represents an aggregation operation.

6. A course recommendation system based on cross-graph enhancement, characterized in that: The system is used to implement the course recommendation method based on cross-graph enhancement as described in any one of claims 1 to 5, comprising: a data graph construction module, a contrastive learning model construction module, a contrastive learning model training module, and a course recommendation result output module; The data graph construction module is used to generate a student-course graph G0 based on the interaction history between students and courses, and to generate a student-course feature graph G0 based on the connection between students and course features. L , based on the students’ descriptive information, generate the student group graph G s ; The contrastive learning model construction module is used to construct a contrastive learning model, including an embedding layer, a main view channel, a static feature channel and a student group channel; the embedding layer generates student nodes, course nodes and course static feature nodes based on random initialization; The main view channel performs graph convolution and aggregation operations on the student-course graph G0 to obtain the course node embedding representation and student node embedding representation , j Indicates j course nodes, c represents the course, i Indicates i student nodes, u represents students, and performs inner product operation to generate similarity scores, and selects the top students based on top-k k The courses are listed as course recommendations; The static feature channel is used for each student-course feature graph G L Perform graph convolution and aggregation operations to obtain the student node embedding representation of each subgraph in the student-course feature graph , , C p Represents the number of student-course feature subgraphs; Embedding the student node into and student node embedding representation Input the fully connected feature extraction layer to get the final representation of the student node ; The student group channel to the student group graph G s Perform graph convolution and aggregation operations to obtain the student group graph G s The student node embedding representation in ; The final representation of the student node and student node embedding representation Aggregate to get the student's global representation , embedding the student node into Global representation with students Conduct comparative learning; The contrastive learning model training module is used to construct a loss function to train a contrastive learning model; The course recommendation result output module is used to obtain course recommendation results based on the trained contrastive learning model.

7. A computer-readable storage medium storing a program, characterized in that: When the program is executed by a processor, the course recommendation method based on cross-graph enhancement as described in any one of claims 1 to 5 is implemented.

Images