A MOOC recommendation method integrating content and graph structure information

Abstract

The invention discloses a method for recommending MOOCs that integrates content and graph structure information, comprising the following steps: step S1, data preprocessing, constructing input data of a GEMRec model and a CNNMRec model; step S2, inputting an adjacency matrix set A into the GEMRec model to obtain a user graph embedding vector and a course graph embedding vector; step S3, inputting the user graph embedding vector, the course graph embedding vector, the user word embedding vector and the course word embedding vector into the CNNMRec model, and outputting the probability that the user studies the course; step S4, during training, using a binary cross entropy loss function, reducing the loss by a gradient descent method, and during recommendation, calculating the cosine similarity between the user and all candidate courses, and recommending the top K courses with the greatest similarity to the user. The invention enables the model to adapt to the dynamic interactive environment between the user and the course, thereby improving the recommendation effect.

Description

Technical Field

[0001] The present invention relates to a method for recommending MOOCs that integrates content and graph structure information, and in particular to a method for recommending MOOCs that integrates content and graph structure information. Background Art

[0002] Massive Open Online Courses (MOOCs) have gradually become a mature form of online and distance education. For example, the three major MOOC platforms, Coursera, edX, and Udacity, have provided numerous courses from internationally renowned universities to millions of users. In China, XuetangX has provided more than 1,000 courses and has more than 6 million users worldwide, making it one of the largest MOOC platforms in China. Online education is an inevitable trend in the future development of education, which further highlights the significance and importance of online learning. There are many types of courses on the MOOC platform. Course recommendations are the main way to systematically guide users' learning. How to enable users with different learning backgrounds to find courses of interest and capture users' dynamic interests that change over time to achieve accurate personalized recommendations is currently a challenging problem;

[0003] Existing course recommendation methods are mainly based on content or collaborative filtering. These methods directly recommend courses to users based on the historical interaction information between users and courses or the content information between users and courses themselves, which ignores the professional levels of different users and is not very interpretable. In addition, using only single-level feature information often cannot fully explore the potential relationship between users and courses, which has a great impact on the accuracy of recommendations. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a MOOC recommendation method that integrates content and graph structure information. In response to the problems mentioned in the background technology, the complex nonlinear relationship between users and courses is mined from two levels of text content and heterogeneous graph structure information, thereby enhancing the interpretability of recommendations and improving the accuracy of recommendations.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: a method for recommending MOOCs that integrates content and graph structure information, comprising the following steps:

[0006] Step S1, data preprocessing, constructing input data for GEMRec model and CNNMRec model;

[0007] Step S2: Input the adjacency matrix set A into the GEMRec model to obtain the user graph embedding vector and the course graph embedding vector;

[0008] Step S3: Input the user graph embedding vector, the course graph embedding vector, the user word embedding vector and the course word embedding vector into the CNNMRec model, and output the interaction probability between user learning and the course;

[0009] Step S4: During training, a binary cross entropy loss function is used, and the loss is reduced by the gradient descent method. During recommendation, the cosine similarity between the user and all candidate courses is calculated, and the top K courses with the greatest similarity are recommended to the user.

[0010] A further improvement of the technical solution of the present invention is that the specific steps of step S1 are as follows:

[0011] Step S11, using the concepts of users, courses and knowledge to construct a heterogeneous information network represented by an adjacency matrix set A including an identity matrix I;

[0012] Step S12: remove HTML tags and escape characters at the beginning and end of the course introduction text, replace the escape characters in the middle of the text with [UNK] marks, and truncate the text after the 510th character if the length exceeds 512 characters;

[0013] Step S13: Input the processed course introduction into the pre-trained model BERT, and take the feature vector corresponding to the [CLS] tag in the output layer as the word embedding vector of the course;

[0014] Step S14: The word embedding vector of each course in the course sequence studied by the user is weighted and summed to obtain the user word embedding vector. The weight is assigned to the course using the exponential decay method. The formula is as follows:

[0015]

[0016] in, Represents user u i The word embedding vector of Represents user u i The number of courses studied, t represents the tth course studied by the user, represents the word embedding vector of the t-th course, Indicates the latest course that the user has studied.

[0017] A further improvement of the technical solution of the present invention is that the specific steps of step S2 are as follows:

[0018] Step S21, the adjacency matrix set A is input into the MG sublayer including two parallel multi-channel 1×1 convolutional neural networks to generate a meta-path adjacency matrix;

[0019] Step S22: Apply the graph convolutional network to the meta-path adjacency matrix A (l) ∈R N×N×C Each channel of , and concatenate the representations of different channels of each node, the formula is as follows:

[0020]

[0021] Among them, || is the concatenation operator, C represents the number of channels, σ is the Sigmoid activation function, is the element path adjacency matrix with self-connection of the ith channel, yes The degree matrix, W∈R d×d is a trainable weight matrix shared across channels, X∈R N×d is the feature matrix, Z contains node representations from C different meta-path graphs, and the length of each meta-path adjacency matrix is ​​different;

[0022] Step S23: Input the user representation in Z as the state into the reinforcement learning model Policy Gradient, update the heterogeneous information network by maximizing the reward, and thus update the node representation in the graph, and finally obtain the graph embedding vector of the user and the course.

[0023] A further improvement of the technical solution of the present invention is that the specific steps of step S21 are as follows:

[0024] Step S211, using multi-channel 1×1 convolution to softly select two adjacency matrices Q1 and Q2 from the adjacency matrix set A, the formula is as follows:

[0025]

[0026] Among them, Q is the soft-selected adjacency matrix, φ is the convolutional layer, and W φ ∈R 1×1×K is the parameter of φ, T e represents a set of edge types, is edge type t i The weight at the lth MG sublayer, is edge type t i The corresponding adjacency matrix;

[0027] Step S212, calculating the element path adjacency matrix by matrix multiplication Q1*Q2 according to the two soft-selected adjacency matrices;

[0028] Step S213: The adjacency matrix of a meta-path of any length l can be realized by stacking MG sub-layers. The specific calculation formula is as follows:

[0029]

[0030] Among them, A P Represents the meta-path adjacency matrix.

[0031] A further improvement of the technical solution of the present invention is that the specific steps of step S23 are as follows:

[0032] Step S231: Given a specific user, the user graph embedding vector is input as the state into the Policy Gradient, and the probability of learning each course is output. The course with the highest probability is taken as the predicted value. If the predicted course is correct, the reward is set to 1, and the model will update the heterogeneous information network by adding an edge between the user and the correctly recommended course, thereby updating the state to obtain a new user representation, which is then input into Policy Gradient, and the round continues;

[0033] Step S232: If the predicted course is wrong, the reward is set to -1, then the round ends and the model parameters are updated. The specific process is as follows:

[0034] Step S2321: The reward obtained after a round is the accumulation of rewards obtained in a single time step. The formula is as follows:

[0035]

[0036] Where γ represents the discount factor, t and k represent the time step, and r t+k+1 represents the reward obtained at the t+k+1 time step.

[0037] Step S2322: The calculation formula for the cumulative reward of all rounds is as follows:

[0038]

[0039] in, represents the probability of adopting the τ strategy, and N is the number of sampling τ;

[0040] Step S2323: The goal is to maximize the cumulative reward If it is used as a loss function, we need to find the minimum value of the loss function, that is, minimize - By solving its derivative, we can find the extreme point. The formula of the derivative process is as follows:

[0041]

[0042] in, They represent the state and the selected action at the tth time step in the nth round respectively;

[0043] Step S2324: After each round, the parameter update formula is as follows:

[0044]

[0045] in, They represent the gradients of the n+1th and nth rounds respectively, α represents the learning rate, and si 、a i Respectively represent the state and selected action of the i-th time step;

[0046] Step S2325: In order to alleviate the special problem of the trade-off between exploration and utilization in reinforcement learning, the entropy regularization method is adopted, and the formula is as follows:

[0047]

[0048] Among them, s t 、a t They represent the state and the selected action at the tth time step, respectively, and C represents the set of all courses;

[0049] Step S2326: The final objective function formula is as follows:

[0050]

[0051] A further improvement of the technical solution of the present invention is that the specific steps of step S3 are as follows:

[0052] Step S31: Input the obtained graph embedding vector and word embedding vector into the CNNMRec model, and output the interaction probability between the user and the course. The formula of the whole process is as follows:

[0053]

[0054] in, Represents the predicted value of the model output, f corresponds to the convolutional layer and the fully connected layer in the CNNMRec model, U g , U w , C g and C w They represent the user graph embedding vector, user word embedding vector, course graph embedding vector and course word embedding vector respectively, and θ is the model parameter to be solved, which includes the weight matrix in the neural network and the bias of the neuron.

[0055] A further improvement of the technical solution of the present invention is that the specific steps of step S31 are as follows:

[0056] Step S311, inputting the graph embedding vector and the word embedding vector into two parallel convolutional neural networks respectively;

[0057] Step S312: pass the output of the convolutional neural network into the fully connected layer, output two feature vectors, namely user embedding and course embedding, and calculate their cosine similarity as the interaction probability between the user and the course.

[0058] A further improvement of the technical solution of the present invention is that the specific steps of step S4 are as follows:

[0059] Step S41, calculate the binary cross entropy loss function, the formula is as follows:

[0060]

[0061] Among them, y (i) The value of is 0 or 1, indicating whether the user has studied the current course;

[0062] Step S42: Measure the recommendation effect by HR@K and NDCG@K:

[0063] The HR@K formula is as follows:

[0064]

[0065] Among them, #Hits@K represents the correct number of Top-K predicted by the model, and |GT| represents the size of the test set;

[0066] The NDCG@K formula is as follows:

[0067]

[0068] Where |Q| represents the total number of users participating in the evaluation, r(j) represents the relevant score of each item, and Z kq is the normalization factor.

[0069] Due to the adoption of the above technical solution, the technical progress achieved by the present invention is:

[0070] 1. The present invention models MOOCs as heterogeneous information networks and combines them with reinforcement learning methods to make the model adaptive to the dynamic interaction environment between users and courses, and obtains graph embedding vectors of users and courses with richer semantic information. In addition, feature fusion of multiple feature information improves the recommendation effect.

[0071] 2. The present invention realizes the innovative supply of course resources through the active recommendation of the platform, helping learners to quickly locate the target courses and grasp the essence of the courses. Introducing knowledge concepts into the course recommendation task, building a heterogeneous information network, and exploring the user's interest preferences at a more fine-grained knowledge concept level enhances the interpretability of the recommendation. In addition, considering the text content and heterogeneous graph structure information at the same time, the method of integrating multi-level feature information reduces the impact of the inherent defects of using only single-level features, realizes feature complementarity, and improves the accuracy of recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 is a flow chart of the method of the present invention;

[0073] Figure 2 It is the structural diagram of MG sublayer;

[0074] Figure 3 It is the overall structure diagram of GEMRec model;

[0075] Figure 4 This is the overall structure diagram of the CNNMRec model. DETAILED DESCRIPTION

[0076] The present invention is further described in detail below in conjunction with embodiments:

[0077] like Figure 1 As shown, a method for recommending MOOCs that integrates content and graph structure information includes the following steps:

[0078] Step S1: Data preprocessing, constructing the input data of the GEMRec model and the CNNMRec model; the specific steps of step S1 are as follows:

[0079] Step S11, using the concepts of users, courses and knowledge to construct a heterogeneous information network represented by an adjacency matrix set A including an identity matrix I;

[0080] Step S12: remove HTML tags and escape characters at the beginning and end of the course introduction text, replace the escape characters in the middle of the text with [UNK] marks, and truncate the text after the 510th character if the length exceeds 512 characters;

[0081] Step S13: Input the processed course introduction into the pre-trained model BERT, and take the feature vector corresponding to the [CLS] tag in the output layer as the word embedding vector of the course. After BERT is pre-trained with large-scale corpus, its feature extraction capability is stronger. In addition, it can simultaneously utilize the text context information. The vector obtained after encoding can better express the semantic information of the sentence.

[0082] Step S14: The word embedding vector of each course in the course sequence studied by the user is weighted and summed to obtain the user word embedding vector. Generally, the course that the user has studied most recently can better represent the user's current interest preference and should be given a greater weight. In order to consider the influence of time factors, the exponential decay method is used to assign weights to the courses. The formula is as follows:

[0083]

[0084] in, Represents user u i The word embedding vector of Represents user u i The number of courses learned, t represents the user's first learning record, represents the word embedding vector of the t-th course, Indicates the latest course that the user has studied.

[0085] Step S2: Input the adjacency matrix set A into Figure 3 The GEMRec model shown in the figure obtains the user graph embedding vector and the course graph embedding vector. Heterogeneous graph embedding can not only represent the nodes in the heterogeneous graph as low-dimensional dense vectors, but also expand the semantic information represented by the original nodes in the heterogeneous graph, and has a good application in heterogeneous graph representation learning. The specific steps are as follows:

[0086] Step S21, the adjacency matrix set A is input to the MG sublayer containing two parallel multi-channel 1×1 convolutional neural networks to generate a metapath adjacency matrix. The structure diagram of the MG sublayer is as follows: Figure 2 shown.

[0087] The specific steps of step S21 are as follows:

[0088] Step S211, using multi-channel 1×1 convolution to softly select two adjacency matrices Q1 and Q2 from the adjacency matrix set A, the formula is as follows:

[0089]

[0090] Among them, Q is the soft-selected adjacency matrix, φ is the convolutional layer, and W φ ∈R 1×1×K is the parameter of φ, T e represents a set of edge types, is edge type t i The weight at the lth MG sublayer, is edge type t i The corresponding adjacency matrix;

[0091] Step S212, calculating the element path adjacency matrix by matrix multiplication Q1*Q2 according to the two soft-selected adjacency matrices;

[0092] Step S213: The adjacency matrix of a meta-path of any length l can be realized by stacking MG sub-layers. The specific calculation formula is as follows:

[0093]

[0094] Among them, A P Represents the meta-path adjacency matrix.

[0095] Step S22: Apply the graph convolutional network (GCN) to the meta-path adjacency matrix A (l) ∈R N×N×C Each channel of , and concatenate the representations of different channels of each node, the formula is as follows:

[0096]

[0097] Among them, || is the concatenation operator, C represents the number of channels, σ is the Sigmoid activation function, is the element path adjacency matrix with self-connection of the ith channel, yes The degree matrix, W∈R d×d is a trainable weight matrix shared across channels, X∈R N×d is the feature matrix, Z contains node representations from C different meta-path graphs, and the length of each meta-path adjacency matrix is ​​different;

[0098] Step S23, reinforcement learning is a mechanism for learning how to map states to actions to obtain maximum rewards. Therefore, the recommendation strategy based on reinforcement learning is more conducive to considering the influence of users' long-term interest preferences and dynamic embedding. The user representation in Z, that is, the user graph embedding vector, is input as the state into the reinforcement learning model Policy Gradient. The heterogeneous information network is updated by maximizing the reward, thereby updating the node representation in the graph, and finally obtaining the graph embedding vector of users and courses. The specific steps are as follows:

[0099] Step S231: Given a specific user, the user graph embedding vector is input as the state into the Policy Gradient, and the probability of learning each course is output. The course with the highest probability is taken as the predicted value. If the predicted course is correct, the reward is set to 1, and the model will update the heterogeneous information network by adding an edge between the user and the correctly recommended course, thereby updating the state to obtain a new user representation, which is then input into Policy Gradient, and the round continues;

[0100] Step S232: If the predicted course is wrong, the reward is set to -1, then the round ends and the model parameters are updated. The specific process is as follows:

[0101] Step S2321: The reward obtained after a round is the accumulation of rewards obtained in a single time step. The formula is as follows:

[0102]

[0103] Here, γ represents the discount factor, and t and k represent the time steps.

[0104] Step S2322: The calculation formula for the cumulative reward of all rounds is as follows:

[0105]

[0106] in, represents the probability of adopting the τ strategy, and N is the number of sampling τ;

[0107] Step S2323: The goal is to maximize the cumulative reward If it is used as a loss function, we need to find the minimum value of the loss function, that is, minimize - By solving its derivative, we can find the extreme point. The formula of the derivative process is as follows:

[0108]

[0109] in, They represent the state and the selected action at the tth time step in the nth round respectively;

[0110] Step S2324: After each round, the parameter update formula is as follows:

[0111]

[0112] in, They represent the gradients of the n+1th and nth rounds respectively, α represents the learning rate, and s i 、a i Respectively represent the state and selected action of the i-th time step;

[0113] Step S2325: In order to alleviate the special problem of the trade-off between exploration and utilization in reinforcement learning, the entropy regularization method is adopted, and the formula is as follows:

[0114]

[0115] Among them, s t 、a t They represent the state and the selected action at the tth time step, respectively, and C represents the set of all courses;

[0116] Step S2326: The final objective function formula is as follows:

[0117]

[0118] Step S3: Input the user graph embedding vector, the course graph embedding vector, the user word embedding vector, and the course word embedding vector into the CNNMRec model, and output the probability of the user learning the course; the specific steps are as follows:

[0119] Step S31: Input the obtained graph embedding vector and word embedding vector into Figure 4 In the CNNMRec model shown, the learning probability of the course is output, and the formula of the whole process is expressed as follows:

[0120]

[0121] in, Represents the predicted value of the model output, f corresponds to the convolutional layer and the fully connected layer in the CNNMRec model, U g , U w , C g and C w They represent the user graph embedding vector, user word embedding vector, course graph embedding vector and course word embedding vector respectively, and θ is the model parameter to be solved, which includes the weight matrix in the neural network and the bias of the neuron.

[0122] The specific steps of step S31 are as follows:

[0123] Step S311: Input the graph embedding vector and the word embedding vector into two parallel convolutional neural networks respectively. Different from the network structure of a general convolutional neural network, it only has a convolution layer but no pooling layer, so as to retain more features;

[0124] Step S312: pass the output of the convolutional neural network into the fully connected layer, output two feature vectors, namely user embedding and course embedding, and calculate their cosine similarity as the interaction probability between the user and the course.

[0125] Step S4: During training, use the binary cross entropy loss function and reduce the loss through the gradient descent method. During recommendation, calculate the cosine similarity between the user and all candidate courses, and recommend the top K courses with the greatest similarity to the user. The specific steps are as follows:

[0126] Step S41, calculate the binary cross entropy loss function, the formula is as follows:

[0127]

[0128] Among them, y (i) The value of is 0 or 1, indicating whether the user has studied the current course;

[0129] Step S42: Measure the recommendation effect by HR@K and NDCG@K:

[0130] HR@K is used to measure the accuracy of recommendations, and its formula is as follows:

[0131]

[0132] Among them, #Hits@K represents the correct number of Top-K predicted by the model, and |GT| represents the size of the test set;

[0133] NDCG@K is used to measure and evaluate the accuracy of the ranking, which assigns greater weights to the top-ranked positions. The formula is as follows:

[0134]

[0135] Where |Q| represents the number of users participating in the evaluation, r(j) represents the relevant score of each item, and Z kq is the normalization factor.

Claims

1. A method for recommending MOOCs that integrates content and graph structure information, characterized in that: The steps include: Step S1: Data preprocessing to construct input data for the GEMRec model and the CNNMRec model; the specific steps are as follows: Step S11, using the concepts of users, courses and knowledge to construct a heterogeneous information network represented by an adjacency matrix set A including an identity matrix I; Step S12: remove HTML tags and escape characters at the beginning and end of the course introduction text, replace the escape characters in the middle of the text with [UNK] marks, and truncate the text after the 510th character if the length exceeds 512 characters; Step S13: Input the processed course introduction into the pre-trained model BERT, and take the feature vector corresponding to the [CLS] tag in the output layer as the word embedding vector of the course; Step S14: The word embedding vector of each course in the course sequence studied by the user is weighted and summed to obtain the user word embedding vector. The weight is assigned to the course using the exponential decay method. The formula is as follows: Among them, e ui Represents user u i The word embedding vector, N ui Represents user u i The number of courses studied, t represents the tth course studied by the user, represents the word embedding vector of the t-th course, Indicates the latest course studied by the user; Step S2, input the adjacency matrix set A into the GEMRec model to obtain the user graph embedding vector and the course graph embedding vector; the specific steps are as follows: Step S21, the adjacency matrix set A is input into the MG sublayer including two parallel multi-channel 1×1 convolutional neural networks to generate a meta-path adjacency matrix; Step S22: Apply the graph convolutional network to the meta-path adjacency matrix A (l) ∈R N×N×C Each channel of , and concatenate the representations of different channels of each node, the formula is as follows: Among them, || is the concatenation operator, C represents the number of channels, σ is the Sigmoid activation function, is the element path adjacency matrix of the ith channel with self-connection, yes The degree matrix, W∈R d×d is a trainable weight matrix shared across channels, X∈R N×d is a feature matrix, Z contains node representations from C different meta-path graphs, and the length of each meta-path adjacency matrix is ​​different; Step S23: Input the user representation in Z as the state into the reinforcement learning model Policy Gradient, update the heterogeneous information network by maximizing the reward, and thus update the node representation in the graph, and finally obtain the graph embedding vector of the user and the course; Step S3: Input the user graph embedding vector, the course graph embedding vector, the user word embedding vector and the course word embedding vector into the CNNMRec model, and output the interaction probability between the user learning and the course; Step S4: During training, a binary cross entropy loss function is used, and the loss is reduced by the gradient descent method. During recommendation, the cosine similarity between the user and all candidate courses is calculated, and the top K courses with the greatest similarity are recommended to the user.

2. A method for recommending MOOCs that integrates content and graph structure information according to claim 1, characterized in that: The specific steps of step S21 are as follows: Step S211, using multi-channel 1×1 convolution to softly select two adjacency matrices Q1 and Q2 from the adjacency matrix set A, the formula is as follows: Among them, φ is the convolutional layer, W φ ∈R 1×1×K is the parameter of φ, T e represents a set of edge types, is edge type t i The weight of the lth MG sublayer, A ti is edge type t i The corresponding adjacency matrix; Step S212, calculating the element path adjacency matrix by matrix multiplication Q1*Q2 according to the two soft-selected adjacency matrices; Step S213: The adjacency matrix of a meta-path of any length l can be realized by stacking MG sub-layers. The specific calculation formula is as follows: Among them, A P Represents the meta-path adjacency matrix.

3. A method for recommending MOOCs that integrates content and graph structure information according to claim 2, characterized in that: The specific steps of step S23 are as follows: Step S231: Given a specific user, the user graph embedding vector is input as the state into the Policy Gradient, and the probability of learning each course is output. The course with the highest probability is taken as the predicted value. If the predicted course is correct, the reward is set to 1, and the model will update the heterogeneous information network by adding an edge between the user and the correctly recommended course, thereby updating the state to obtain a new user representation, which is then input into Policy Gradient, and the round continues; Step S232: If the predicted course is wrong, the reward is set to -1, then the round ends and the model parameters are updated. The specific process is as follows: Step S2321: The reward obtained after a round is the accumulation of rewards obtained in a single time step. The formula is as follows: Where γ represents the discount factor, t and k represent the time step; Step S2322: The calculation formula for the cumulative reward of all rounds is as follows: Among them, π θ (τ) represents the probability of adopting the τ strategy; Step S2323, the goal is to maximize the cumulative reward L(θ). If it is used as a loss function, it is necessary to solve the minimum value of the loss function, that is, minimize -L(θ). By solving its derivative, the extreme point is obtained. The formula for the derivation process is as follows: Step S2324: After each round, the parameter update formula is as follows: Step S2325: In order to alleviate the special problem of the trade-off between exploration and utilization in reinforcement learning, the entropy regularization method is adopted, and the formula is as follows: Step S2326: The final objective function formula is as follows: -L(θ)+λH[π θ (a t |s t )]。 4. A method for recommending MOOCs that integrates content and graph structure information according to claim 3, characterized in that: The specific steps of step S3 are as follows: Step S31: Input the obtained graph embedding vector and word embedding vector into the CNNMRec model, and output the interaction probability between the user and the course. The formula of the whole process is as follows: in, Represents the predicted value of the model output, f corresponds to the convolutional layer and the fully connected layer in the CNNMRec model, U g , U w , C g and C w They represent the user graph embedding vector, user word embedding vector, course graph embedding vector and course word embedding vector respectively, and θ is the model parameter to be solved, which includes the weight matrix in the neural network and the bias of the neuron.

5. A method for recommending MOOCs that integrates content and graph structure information according to claim 4, characterized in that: The specific steps of step S31 are as follows: Step S311, inputting the graph embedding vector and the word embedding vector into two parallel convolutional neural networks respectively; Step S312: pass the output of the convolutional neural network into the fully connected layer, output two feature vectors, namely user embedding and course embedding, and calculate their cosine similarity as the interaction probability between the user and the course.

6. A method for recommending MOOCs that integrates content and graph structure information according to claim 5, characterized in that: The specific steps of step S4 are as follows: Step S41, calculate the binary cross entropy loss function, the formula is as follows: Among them, y (i) The value of is 0 or 1, indicating whether the user has studied the current course; Step S42: Measure the recommendation effect by HR@K and NDCG@K: The HR@K formula is as follows: Among them, |GT| represents the size of the test set; The NDCG@K formula is as follows: Among them, Q' represents the number of users participating in the evaluation, r(j) represents the relevant score of each item, and Z kq is the normalization factor.

Images