Graph collaborative filtering course recommendation method for quantifying preference degree and level penalty

Abstract

The invention discloses a graph collaborative filtering curriculum recommendation method for quantifying preference degree and level penalty, which is used for improving the curriculum recommendation effect of an online learning platform and reducing the curriculum rate. According to the method, firstly, learning log data of a user is collected, and the preference degree of the user for registered courses is calculated by quantifying learning behaviors and learning duration of the user; thirdly, constructing a user-course bipartite graph, and applying the preference degree as an attention weight to a neighborhood feature aggregation process of the graph convolutional network; in order to relieve the problem of node over-smoothing, a new loss function item is introduced to measure and optimize the retention degree of the initial feature in the embedded representation, and different punishment intensities are applied to different levels of embedded representations, so that the over-smoothing phenomenon of the graph neural network in deep iteration is effectively relieved.

Description

Technical Field

[0001] The present invention relates to the technical field of recommendation algorithms, and in particular to a recommendation method in a course recommendation scenario, and specifically to a graph collaborative filtering course recommendation method with quantified preference and hierarchical penalty. Background Art

[0002] With the rapid development of Internet technology, more and more learners tend to obtain educational resources through online platforms. This model breaks the time and space limitations of traditional education, making high-quality educational content accessible to a wider audience. However, although online learning provides unprecedented convenience and flexibility, it also faces a significant challenge - high dropout rates. The completion rate of many online courses is much lower than that of courses in traditional classroom environments, which not only affects the personal development of learners, but also reduces the overall benefits of online education platforms.

[0003] In order to deal with the high dropout rate problem in online learning, existing research works have adopted graph neural networks to improve collaborative filtering models and thus optimize course recommendation effects. The graph collaborative filtering model constructs a user-item bipartite graph and uses graph neural networks to learn the embedded representations of users and items. However, the current graph collaborative filtering methods based on implicit feedback have some limitations. On the one hand, due to the lack of explicit ratings of registered courses by users, the existing models only rely on the behavior of users registering courses to construct the user-item graph. In this process, the weights of all interaction edges are uniformly set to the same value. This approach ignores the weight differences of edges and fails to fully reflect the learners' personalized preferences for different courses. Secondly, this homogenized edge weight processing method simplifies the complex interactions between users and courses, limits the model's ability to learn deep-level features of users and courses, and thus cannot learn better user and course representations, resulting in the accuracy of recommendation results being affected. On the other hand, during the training process of the graph neural network, as the number of layers increases, the feature representations of nodes tend to be consistent, resulting in the feature representations of nodes in the same connected component tending to be the same. This phenomenon reduces the model's ability to capture personalized information of different nodes, thereby affecting the performance and accuracy of the recommendation system. Summary of the invention

[0004] In view of the high dropout rate of online platform courses and the problems existing in the existing graph collaborative filtering recommendation model, the purpose of the present invention is to provide a graph collaborative filtering course recommendation method with quantified preference and hierarchical penalty.

[0005] The object of the present invention is achieved through the following technical solution: a graph collaborative filtering course recommendation method with quantified preference and hierarchical penalty, comprising the following steps:

[0006] 1. Collecting learning log data of users on the online course platform for registered courses, wherein the learning log data specifically includes what kind of interaction a user performed on a course and when;

[0007] 2. Quantify the user's preference for the registered courses based on the learning log data;

[0008] 3. Construct a bipartite graph of user-course nodes. If the user has registered for a course, an edge is established between the user node and the node of the course, and the edge is assigned an aggregated attention weight according to the preference obtained in step 2.

[0009] 4. Based on the existing simplified graph convolutional network model LightGCN, the weighted bipartite graph established in step 3 is constructed as the adjacency matrix of the aggregation update formula in the model, and the hierarchical penalty is added to the training loss function to learn the final embedding representation of each user node and course node;

[0010] 5. Based on the final embedding representation of the user node and the course node obtained in step 4, calculate the inner product of the user embedding and the course embedding as the prediction score; for each user, sort all courses in descending order according to the prediction score, and select the top 5% of courses to form a recommendation list.

[0011] Furthermore, the step 2 includes the following sub-steps:

[0012] 2.1、Define user u i In course c j During the learning process, different types of behavioral interaction action sets are performed on the learning modules of the registered course; wherein the learning modules include teaching videos, question-answering and problem-solving, and courseware resources; the different types of interaction action sets are as follows:

[0013]

[0014] Among them, |A(u i ,c j )| represents the user u i In course c j The total number of different interaction action types on ; Represents user u i In course c j The 1,2,p,|A(u i ,c j )|interaction types;

[0015] The user u i In course c j The total number of different types of behavioral interactions is:

[0016]

[0017] in, Respectively represent the user u i In course c j No.1,2,p,|A(u i ,c j )|The total number of interactive actions;

[0018] The learning time is expressed as user u i In course c j The total time spent on learning is calculated by summing the timestamp differences of adjacent actions (such as playing and pausing video actions). i In the registered course c j The learning time on is expressed as:

[0019]

[0020] 2.2 Definition For user u i For Course C j behavioral preference, For user u i For Course C j The time preference of X is defined as ij For user u i For Course C j preference of

[0021] Calculate the user u i In course c j The number of actions of the pth type on the course c is relative to the number of actions of all users on the course c. j The ratio of the average times of the first 20% of the p-th action times set

[0022]

[0023] in, Represents user u i In course c j The number of actions of the pth type on , Indicates that all users in course c j The number of the p-th action;

[0024] Calculate the user u i In course c j The ratio of the number of actions of the pth type in the course to the total number of actions in the course

[0025]

[0026] The vectorized representation is Finally, we get user u i For Course C j Behavioral preference

[0027]

[0028] Calculate according to the following formula MeanTop 20％ (t j ) for all users in the same course c j The average of the top 20% of study time;

[0029]

[0030] Among them, t ij Represents user u i In course c j The learning time on j Indicates that all users in course c j The learning duration set on ;

[0031] right and Weighted summation to obtain user u i For Course C j Preference X ij :

[0032]

[0033] Among them, β is a hyperparameter;

[0034] In the above way, the preference of all users and all courses is obtained, and finally the user-course preference matrix X is constructed.

[0035] Furthermore, the user-course preference matrix is ​​represented as X∈R |U|×|I| ; where |U| represents the total number of users, |i| represents the total number of courses, R represents a real number set, and X ij Represents user u i For Course C j preference.

[0036] Furthermore, the step 4 includes the following sub-steps:

[0037] 4.1. Use the softmax function to normalize the user-course preference matrix X:

[0038]

[0039] in, represents the user-course preference matrix after row normalization;

[0040] Construct the adjacency matrix A of the user-course node bipartite graph:

[0041]

[0042] The user embedding layer and the course embedding layer are randomly initialized using ID One Hot Encoding to represent the 0th layer embedding: E (0) ∈R (|U|+|I|)*T , T is the dimension size of the embedding layer; then the embedding representation of each layer is updated by simplifying the GCN aggregation:

[0043] E (k+1) =AE (k)

[0044] Among them, k represents the number of layers of the graph convolutional network; the embedding representation of the last layer node is directly used as the final embedding of the user-course and

[0045] 4.2. In network training, BPR loss is used as part of the total loss, and the objective function is expressed as:

[0046]

[0047] in, represents paired training data, X + represents the interaction pair between users and registered courses, X - represents the interaction pairs between the user and the remaining unregistered courses, σ is the sigmoid function, Θ represents all trainable model parameters, and λ controls the L2 regularization strength to prevent overfitting;

[0048] Define the user set U and the course set I, which contain the number of users |U| and the number of courses |I| respectively. e represents the initial feature embedding vector of the node. represents the final embedding vector after network learning, and the following formula is used to measure the degree of preservation of the initial features in the final embedding representation:

[0049]

[0050] in:

[0051]

[0052] From the above formula, we can see that Save avg The smaller the value, the better the initial features are preserved in the final embedding representation; represents the average degree of preservation of the initial features in the final embedding vector of the user set, represents the average degree of preservation of the initial features in the final embedding vector of the course set, e i 、e j Represent the initial feature vectors of user i and user j, or course i and course j, respectively. Respectively represent the final embedding vectors of user i and user j, or course i and course j after network learning, Diff cos (E i ,E j ) represents the cosine similarity difference between two sets of embedding vectors, which is used to quantify the degree of change between the final embedding representation and the initial embedding representation;

[0053] Therefore, the total loss function is expressed as:

[0054] L=L BPR +αL save

[0055] Among them, α is a hyperparameter, and L save The initial feature retention degree of each layer of user and course nodes calculated by GCN during network training is included in the loss, as shown below:

[0056]

[0057] in, Respectively represent the average retention degree of the initial features in the embedding vectors of the user set and the course set in the network layer l, Respectively represent the penalty coefficients of the user and course items corresponding to the lth layer, and All values ​​are It means that as the number of layers increases, the penalty coefficient decreases.

[0058] The beneficial effects of the present invention are: user preferences are quantified through user behavior and learning time in the learning log, and the aggregation weights in the user-course graph are optimized to enhance the model's ability to capture personalized preferences. At the same time, by introducing a measure of the degree of initial feature retention and adding it to the loss function, the problem of deep node over-smoothing is effectively alleviated, thereby improving the accuracy and personalization level of the recommendation system, which helps to reduce the dropout rate of online courses. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 It is a flowchart of the graph collaborative filtering course recommendation method with quantified preference and hierarchical penalty;

[0060] Figure 2 It is a graph convolutional network model diagram of preference weight and hierarchical penalty. DETAILED DESCRIPTION

[0061] The present invention is described in detail below in conjunction with the accompanying drawings. In the absence of conflict, the features of the following embodiments and implementations can be combined with each other.

[0062] like Figure 1 As shown, a graph collaborative filtering course recommendation method with quantified preference and hierarchical penalty provided by an embodiment of the present invention includes the following steps:

[0063] Step 1: Collect learning log data of users on the online course platform for their registered courses, such as when and what kind of interaction a user performs on a course.

[0064] Step 2: Based on the learning log data recorded in step 1, count and quantify the user's preference for the courses he or she has registered for.

[0065] (2.1) Online courses generally consist of multiple sections such as teaching videos, Q&A, and courseware resources. i In the registered course c j During the learning process, different behavioral interactions will be performed with the related learning modules.

[0066] Define user u i In course c j The following are different types of interactive action sets:

[0067]

[0068] Among them, |A(u i ,c j )| represents the user u i In course c j The total number of different interaction action types on ; Represents user u i In course c j The 1,2,p,|A(u i ,c j )|interaction types.

[0069] User i In course c j The total number of different types of interactive actions is expressed as:

[0070]

[0071] in, Respectively represent the user u i In course c j No.1,2,p,|A(u i ,c j )|The total number of interactions.

[0072] The learning time is expressed as user u i In course c j The total time spent on learning is calculated by summing the timestamp differences of adjacent actions (such as playing and pausing video actions), that is, user u i In the registered course c j The learning time on is expressed as:

[0073]

[0074] (2.2) Definition For user u i For Course C j behavioral preference, For user u i For Course C j The time preference of X ij For user u i For Course C j preference.

[0075] Calculate according to the following formula Represents user u i In course c j The ratio of the number of p-th actions on the same course to the average number of the top 20% of the number of p-th actions of all users on the same course, and the minimum value of 1 is taken to ensure that Between 0 and 1, and finally ensure It is also between 0 and 1.

[0076]

[0077] in, Represents user u i In course c j The number of actions of the pth type on , Indicates that all users in course c j The number of times the p-th action was performed.

[0078] Calculate user u i In course c j The ratio of the number of actions of the pth type in the course to the total number of actions in the course

[0079]

[0080] Vectorized Representation We can find user u i For Course C j Behavioral preference

[0081]

[0082] Calculate according to the following formula MeanTop 20％ (t j ) for all users in the same course c j The average learning time is in the top 20%.

[0083]

[0084] Last pair and The weighted summation is as follows, where β is a hyperparameter, and the user u can be obtained i For Course C j Preference X ij :

[0085]

[0086] Finally, the preference between all users and all courses is calculated and expressed as the user-course preference matrix X.

[0087] Step 3: Construct a bipartite graph of user-course nodes. That is, if a user has registered for a course, an edge is established between the user node and the course node. The preference degree required in step 2 is the aggregate attention weight given to the edge. The user-course preference matrix is ​​described as X∈R |U|×|I| , where |U| is the total number of users, |I| is the total number of courses, R is a set of real numbers, and X ij For user u i For Course C j preference.

[0088] Step 4: Use the simplified graph convolutional network model with preference weight and hierarchical penalty proposed in this invention to learn the final embedding representation of each user node and course node. The network model is as follows: Figure 2 As shown, it includes user and course embedding layers; aggregation update layer; prediction layer. The network training details are as follows:

[0089] (4.1) First, the user-course preference matrix X is row normalized using the softmax function:

[0090]

[0091] Thus, the adjacency matrix A of the user-course node bipartite graph is constructed:

[0092]

[0093] Depend on Figure 2The user embedding layer and course embedding layer represented in the figure are randomly initialized using ID One Hot Encoding. The embedding representation of the 0th layer is: (0) ∈R (|U|+|I|)*T , T is the dimension size of the embedding layer, and then the embedding of each layer is updated through 3 layers of simplified GCN aggregation without feature transformation matrix and nonlinear activation function. Different from the traditional simplified GCN, which usually adopts the average or summation method when aggregating the features of adjacent nodes, the model of the present invention weights the features of adjacent nodes according to the user's preference for the course. In other words, when updating its embedded representation, the node will pay more attention to those neighboring nodes with stronger connection strength, so that the information transmission is more accurate and personalized, and the expressiveness of the model is enhanced. The specific aggregation update formula is as follows:

[0094] E (k+1) =AE (k)

[0095] Where k represents the number of layers of the graph convolutional network; the embedding representation of the third layer node is used as the final embedding representation of the user-course, that is, and

[0096] (4.2) The BPR loss is used as part of the total loss in network training. Its objective function aims to maximize the difference between the predicted values ​​of positive samples (observed interactions) and negative samples (unobserved interactions). The objective function is:

[0097]

[0098] in represents paired training data, X + represents the interaction pair between users and registered courses, X - represents the interaction pairs between the user and the remaining unregistered courses, σ is the sigmoid function, Θ represents all trainable model parameters, and λ controls the L2 regularization strength to prevent overfitting.

[0099] In the research and practice of recommendation systems, although graph convolutional networks are widely used due to their advantages in processing graph-structured data, as the number of network layers increases, GCN may encounter a problem called "over-smoothing" when processing node embeddings. The over-smoothing phenomenon refers to the fact that as the number of network layers increases, the embedding representations of different nodes tend to be consistent, resulting in the inability to effectively distinguish the characteristics of different nodes, thereby damaging the accuracy of the recommendation system and the ability to make personalized recommendations. In order to solve this problem, the present invention proposes a new method to alleviate the over-smoothing phenomenon by introducing an indicator to measure the degree of retention of initial features in the final embedding. The core idea of ​​this method is to pay attention not only to the smoothness of node embeddings during network training, but also to retain the initial feature information of the nodes, so as to ensure that the nodes will not be over-smoothed by retaining the initial feature information. In addition, given that the distributions of the final embedding representations of users and items show significant differences, the user and item representation matrices are calculated separately, such as Figure 2 As shown, the retention loss is calculated for the user and item representation matrices of each layer. Specifically, the user set U and the course set I are defined to contain the number of users |U| and the number of courses |I|, respectively. Let e ​​represent the initial feature embedding vector of the node. represents the final embedding vector after network learning, and introduces the following formula to calculate the degree of preservation of the initial features in the final embedding representation:

[0100]

[0101] in:

[0102]

[0103]

[0104] From the above formula, we can see that Save avg The smaller the value, the better the initial features are preserved in the final embedding representation; represents the average degree of preservation of the initial features in the final embedding vector of the user set, represents the average degree of preservation of the initial features in the final embedding vector of the course set, e i 、e j Represent the initial feature vectors of user i and user j, or course i and course j, respectively. Respectively represent the final embedding vectors of user i and user j, or course i and course j after network learning, Diff cos (E i ,E j ) represents the cosine similarity difference between two sets of embedding vectors, which is used to quantify the degree of change between the final embedding representation and the initial embedding representation.

[0105] In order to alleviate the over-smoothness of user and item representations in the collaborative filtering model based on the simplified GCN with the feature transformation matrix and nonlinear activation function removed, the present invention directly incorporates the above-mentioned index for measuring the degree of retention into the loss function as part of the model training. By adjusting the loss function, the model not only pursues the smoothness of node embedding during training, but also retains the initial feature information of the node as much as possible. In this way, even in a deep network, the model can maintain sensitivity to different node characteristics, thereby improving the performance of the recommendation system. In addition, in the graph collaborative filtering model, the smoothness of the deep representation is a double-edged sword. On the one hand, smoothness can help the model capture the global relationship between nodes in the graph, thereby improving the overall performance of the recommendation system. On the other hand, excessive smoothness will lead to the homogenization of node features, making it difficult for the model to distinguish the unique attributes of different nodes, which is not conducive to personalized recommendations. Therefore, considering the contribution of embedding smoothness to the effectiveness of the graph collaborative filtering model, while limiting the over-smoothing of the final representation of the model, it is also necessary to ensure the appropriate smoothness of the deeper representation. The present invention introduces a hierarchical penalty mechanism in the retention loss function. The core idea is to assign different penalty coefficients to network layers of different depths to control the smoothness of each layer. Specifically, deeper network layers will receive lower penalty coefficients. In this way, the initial embedding is supplemented to each layer of embedding representation, thereby alleviating the over-smoothing problem. The total loss function can be expressed as:

[0106] L=L BPR +αL save

[0107] Among them, α is a hyperparameter, and L save The initial feature retention degree of each layer of user and course nodes calculated by GCN during network training is included in the loss, as shown below:

[0108]

[0109] in, Respectively represent the average retention degree of the initial features in the embedding vectors of the user set and the course set in the network layer l, Respectively represent the penalty coefficients of the user and course items corresponding to the lth layer, and All values ​​are It means that as the number of layers increases, the penalty coefficient decreases.

[0110] Step 5: Based on the user embedding and course embedding learned through the network in step 4, calculate the inner product of the user embedding and the course embedding as the prediction score. For each user, sort all courses in descending order according to the prediction score, and select the top 5% of courses to form a recommendation list.

[0111] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A graph collaborative filtering course recommendation method with quantified preference and hierarchical penalty, characterized by: The steps include: (1) Collecting learning log data of users on the online course platform for registered courses, wherein the learning log data specifically includes what kind of interaction a user performed on a course and when; (2) quantifying the user's preference for the registered courses based on the learning log data; (3) constructing a bipartite graph of user-course nodes. If the user has registered for a course, an edge is established between the user node and the node of the course, and an aggregated attention weight is assigned to the edge according to the preference obtained in step (2); (4) Based on the existing simplified graph convolutional network model LightGCN, the weighted bipartite graph established in step (3) is constructed as the adjacency matrix of the aggregation update formula in the model, and a hierarchical penalty is added to the training loss function to learn the final embedding representation of each user node and course node; (5) Based on the final embedding representation of the user node and the course node obtained in step (4), the inner product of the user embedding and the course embedding is calculated as the prediction score; for each user, all courses are sorted in descending order according to the prediction score, and the top 5% of courses are selected to form a recommendation list.

2. The graph collaborative filtering course recommendation method with quantified preference and hierarchical penalty according to claim 1 is characterized in that: The step (2) comprises the following sub-steps: (2.1) Define user u i In course c j During the learning process, different types of behavioral interaction action sets are performed on the learning modules of the registered course; wherein the learning modules include teaching videos, question-answering and problem-solving, and courseware resources; the different types of interaction action sets are as follows: Among them, |A(u i ,c j )| represents the user u i In course c j The total number of different interaction action types on ; Represents user u i In course c j The 1,2,p,|A(u i ,c j )|interaction types; The user u i In course c j The total number of different types of behavioral interactions is: in, Respectively represent the user u i In course c j No.1,2,p,|A(u i ,c j )|The total number of interactive actions; The user u i In the registered course c j The learning time on is expressed as: (2.2) Definition For user u i For Course C j behavioral preference, For user u i For Course C j The time preference of X is defined as ij For user u i For Course C j preference of Calculate the user u i In course c j The number of actions of the pth type on the course c is relative to the number of actions of all users on the course c. j The ratio of the average times of the first 20% of the p-th action times set in, Represents user u i In course c j The number of actions of the pth type on , Indicates that all users in course c j The number of the p-th action; Calculate the user u i In course c j The ratio of the number of actions of the pth type in the course to the total number of actions in the course The vectorized representation is Finally, we get user u i For Course C j Behavioral preference Calculate according to the following formula MeanTop 20％ (t j ) for all users in the same course c j The average of the top 20% of study time; Among them, t ij Represents user u i In course c j The learning time on j Indicates that all users in course c j The learning duration set on ; right and Weighted summation to obtain user u i For Course C j Preference X ij : Among them, β is a hyperparameter; In the above way, the preference of all users and all courses is obtained, and finally the user-course preference matrix X is constructed.

3. The graph collaborative filtering course recommendation method with quantified preference and hierarchical penalty according to claim 2 is characterized in that: The user-course preference matrix is ​​represented as X∈R |U|×|I| ; where |U| represents the total number of users, |I| represents the total number of courses, R represents a real number set, and X ij Represents user u i For Course C j preference.

4. The graph collaborative filtering course recommendation method with quantified preference and hierarchical penalty according to claim 1 is characterized in that: The step (4) comprises the following sub-steps: (4.1) Use the softmax function to normalize the user-course preference matrix X: in, represents the user-course preference matrix after row normalization; Construct the adjacency matrix A of the user-course node bipartite graph: The user embedding layer and the course embedding layer are randomly initialized using ID unique hot encoding to represent the 0th layer embedding: E (0) ∈R (|U|+|I|)*T , T is the dimension size of the embedding layer; then the embedding representation of each layer is updated by simplifying the GCN aggregation: AND (k+1) =AE (k) Among them, k represents the number of layers of the graph convolutional network; the embedding representation of the last layer node is used as the final embedding of the user-course and (4.2) In network training, BPR loss is used as part of the total loss, and the objective function is expressed as: in, represents paired training data, X + represents the interaction pair between users and registered courses, X - represents the interaction pairs between the user and the remaining unregistered courses, σ is the sigmoid function, Θ represents all trainable model parameters, and λ controls the L2 regularization strength to prevent overfitting; Define the user set U and the course set I, which contain the number of users |U| and the number of courses |I| respectively. e represents the initial feature embedding vector of the node. represents the final embedding vector after network learning, and the following formula is used to measure the degree of preservation of the initial features in the final embedding representation: in: From the above formula, we can see that Save avg The smaller the value, the better the initial features are preserved in the final embedding representation; represents the average degree of preservation of the initial features in the final embedding vector of the user set, represents the average degree of preservation of the initial features in the final embedding vector of the course set, e i 、e j Represent the initial feature vectors of user i and user j, or course i and course j, respectively. Respectively represent the final embedding vectors of user i and user j, or course i and course j after network learning, Diff cos (E i ,E j ) represents the cosine similarity difference between two sets of embedding vectors, which is used to quantify the degree of change between the final embedding representation and the initial embedding representation; Therefore, the total loss function is expressed as: L=L BPR +αL save Among them, α is a hyperparameter, and L save The initial feature retention degree of each layer of user and course nodes calculated by GCN during network training is included in the loss, as shown below: in, Respectively represent the average retention degree of the initial features in the embedding vectors of the user set and the course set in the network layer l, Respectively represent the penalty coefficients of the user and course items corresponding to the lth layer, and All values ​​are It means that as the number of layers increases, the penalty coefficient decreases.

Images