A dropout prediction method for online courses based on association and video interaction

Abstract

The present invention discloses an online education dropout prediction method based on association relationship and video interaction. First, a heterogeneous information graph of a user's learning course video and a video interaction time series matrix are obtained. Then, on the one hand, the heterogeneous information graph is input into a deep association relationship feature extraction module based on node triples to obtain the association feature representation of the user and the course. Then, the deep association relationship feature is obtained through feature fusion. On the other hand, the video interaction time series matrix is ​​input into an interaction feature extraction module to obtain the interaction feature. Finally, the dropout probability of the user in his elective course is calculated by combining the deep association relationship feature and the interaction feature. The present invention models the user's online learning situation as a heterogeneous information graph, constructs a node representation with stronger representation ability by mining the complex and diverse node association relationships in the graph, and extracts the interaction features by combining the video interaction information with the largest amount of information in the real situation, thereby greatly enhancing the prediction ability and generalization ability of the model.

Description

Technical Field

[0001] The present invention belongs to the field of online education dropout behavior prediction, and in particular relates to an online course dropout prediction method based on association relationship and video interaction. Background Art

[0002] In recent years, Massive open online courses (MOOCs) have used the Internet as a carrier to deliver high-quality course resources from top universities around the world to the terminal devices of learners around the world, promoting educational equity. In addition, through online education, learners can control their own learning progress and choose their own learning environment according to their own time and ability, which undoubtedly breaks the time and place restrictions of traditional education.

[0003] However, unlike traditional education, online education has few interactions between teachers and students, fragmented class time, and lack of supervision. It is difficult for learners to complete the entire course. Relevant research shows that more than 90% of learners in online education will drop out. The extremely high dropout rate has become a major obstacle to the development of online education and the transformation of educational technology. On the other hand, due to the large number of online education groups, teachers or platforms cannot track the learning behavior of each learner and determine the possibility of dropout through human resources. Therefore, it is particularly important to design an automated prediction method to accurately and timely predict learners' dropouts.

[0004] The patent application with publication number CN110059716A discloses a CNN-LSTM-SVM network model construction and MOOC dropout prediction method. The method processes the original online learning behavior records of known students to obtain multiple time slice matrices, performs network training according to the time slice matrices, and obtains a trained CNN-LSTM network; finally, a trained dropout prediction network is obtained.

[0005] The patent application with publication number CN109558983A discloses a method and device for predicting the dropout rate of online courses. The method constructs a variety of behavioral attribute information into a behavior matrix, and calculates the weight vector corresponding to the behavior matrix based on the behavior matrix. The dropout prediction model is obtained according to the dropout rate of sample data and the behavioral data corresponding to the user's multiple behavioral attribute information and the weight vector.

[0006] The above two methods extract user behavior characteristics from the user's behavior matrix. However, there are many factors that affect user dropouts in the online education environment. The behavior matrix cannot fully represent the deep reasons for user dropouts. Therefore, the performance of dropout prediction is relatively general.

[0007] The patent application with publication number CN112734105A discloses a method for preventing dropout behavior in online education. The method first performs deep data mining on the original data obtained from the online education platform to obtain the user's activity time series matrix, context information and "friend" relationship, and then inputs the activity time series matrix, context information and course selection feature similarity relationship into their respective constructed and trained feature extraction modules to obtain the user's deep sequence features, attention features and classmate features. Finally, the three features are fused and the fused features are input into the behavior prediction model to obtain the user's dropout behavior prediction results.

[0008] This method not only extracts the user's temporal features but also introduces contextual information and a single node relationship such as "friends". However, it still cannot fully represent the real situation of users' online learning. In real situations, users mainly learn knowledge by watching course videos, so video interaction information often has the largest amount of information, but this method does not fully exploit this feature. In addition, there are rich deep correlations between nodes in the learning scenario, and this method does not fully exploit this feature, so the effect of dropout prediction still needs to be improved. Summary of the invention

[0009] The purpose of this application is to provide an online course dropout prediction method based on association relationships and video interactions to characterize the real context of users' online learning, enrich the semantic representation of users and courses using node triples, and improve the accuracy of prediction.

[0010] In order to achieve the above purpose, the technical solution of this application is as follows:

[0011] A method for predicting dropout of online courses based on association relationship and video interaction, including:

[0012] According to the user's course selection record, the user's video viewing record and the course video resource record, a heterogeneous information graph G of the user's course video learning is constructed. After the node embedding module, the node embedding corresponding to each node in the graph G is obtained.

[0013] The nodes of each node are embedded in a predefined user triplet set t u Input into the deep correlation feature extraction module to obtain the user's correlation feature representation f u ;

[0014] The nodes of each node are embedded in a predefined course triple set t c Input into the deep correlation feature extraction module to obtain the course correlation feature representation f c ;

[0015] Based on the relationship between users and course selections, the user's associated features f u and the associated features of the course c Perform feature fusion to obtain the deep correlation feature F of the user and course selection relationship pair R ;

[0016] According to the user's course selection record and the user's video viewing record, the video interaction timing matrix A is constructed, and the video interaction timing matrix A is input into the interaction feature extraction module to obtain the interaction feature F T ;

[0017] Based on the deep correlation feature F R and the interaction feature F T , calculate the probability of a user dropping out of his or her elective courses.

[0018] Furthermore, the heterogeneous information graph G of the user's course video learning is constructed according to the user's course selection record, the user's video viewing record and the course video resource record, and the node embedding corresponding to each node in the graph G is obtained through the node embedding module, including:

[0019] Based on the user's course selection records, the user's video viewing records and the course's video resource records, a heterogeneous information graph G(U,C,V) is formed, where U represents the set of all user nodes, C represents the set of all course nodes, V represents the set of all video nodes, the edge between the user and the course represents the user's selection relationship with the course, the edge between the user and the video represents the user's viewing relationship with the video, and the edge between the course and the video represents the subordinate relationship between the video and the course;

[0020] Based on the heterogeneous information graph G, one-hot encoding is used to uniquely identify all user nodes, course nodes, and video nodes in the graph G;

[0021] The one-hot encoding of all nodes obtained in the previous step is input into the node embedding module, which includes a fully connected layer. The parameter matrix of the fully connected layer is obtained, and the node embedding corresponding to each node is obtained by searching the parameter matrix of the fully connected layer.

[0022] Furthermore, the nodes of each node are embedded in a predefined user triplet set t u Input into the deep correlation feature extraction module to obtain the user's correlation feature representation f u ,include:

[0023] Step 2.1: Construct the user triple set t according to the association relationship between the user node and any node u ;

[0024] Step 2.2: Based on any triple η∈t u , sample neighbor nodes of different user nodes to obtain the first-order neighbor node set of all users, and then sample neighbor nodes of the first-order neighbor nodes of all user nodes in turn to obtain the second-order neighbor node set of the user;

[0025] Step 2.3: Based on any triple η∈t u ,Using the correlation measurement method, we get the similarity matrix between all user nodes and their second-order neighbor nodes;

[0026] Step 2.4: Based on any triple η∈t u , multiply the similarity value between the user node and its second-order neighbor node and the node embedding of the second-order neighbor node itself, and obtain the embedding representation of the second-order neighbor nodes of all user nodes under the triple η based on similarity guidance;

[0027] Step 2.5: Based on any triple η∈t u , apply the soft attention mechanism to obtain the associated feature representation of the user under the triple η, and traverse the triple set t u , get the associated feature representation of all users under different triples;

[0028] Step 2.6: Perform a self-attention mechanism operation on the associated feature representations of all users under different triples to obtain the associated feature representations of the users.

[0029] Furthermore, the nodes of each node are embedded in a predefined course triple set t c Input into the deep correlation feature extraction module to obtain the course correlation feature representation f c ,include:

[0030] Step 3.1: Construct a course triplet set t according to the relationship between the course node and any node c ;

[0031] Step 3.2: Based on any triple ξ∈t c , sample neighbor nodes for different course nodes to obtain the first-order neighbor node set of all courses, and then sample neighbor nodes for the first-order neighbor nodes of all course nodes in turn to obtain the second-order neighbor node set of the course;

[0032] Step 3.3: Based on any triple ξ∈t c ,Using the correlation measurement method, we get the similarity matrix between all course nodes and their second-order neighbor nodes;

[0033] Step 3.4: Based on any triple ξ∈t c, multiply the similarity value between the course node and its second-order neighbor node and the node embedding of the second-order neighbor node itself, and obtain the embedding representation of the second-order neighbor nodes of all course nodes under the triple ξ based on similarity guidance;

[0034] Step 3.5: Based on any triple ξ∈t c , apply the soft attention mechanism, get the associated feature representation of the course under the triple ξ, and traverse the triple set t c , get the associated feature representation of all courses under different triples;

[0035] Step 3.6: Perform self-attention mechanism operation on the associated feature representations of all courses under different triples to obtain the associated feature representation of the course f c .

[0036] Furthermore, based on the user and course selection relationship pair, the user's associated feature f u and the associated features of the course f c Perform feature fusion to obtain the deep correlation feature F of the user and course selection relationship pair R ,include:

[0037] Step 4.1, according to the course selection record of the user, obtain the set E of the relationship pairs between the user and the course selection;

[0038] Step 4.2: Based on the set E of user and course selection relationship pairs, the user's associated features f u and the associated characteristics of the courses they choose c Concatenate and flatten, then send to the fully connected layer and activate with the sigmoid function to get the final deep correlation feature F R .

[0039] Furthermore, the video interaction timing matrix A is constructed based on the user's course selection record and the user's video viewing record, and the video interaction timing matrix A is input into the interaction feature extraction module to obtain the interaction feature F T ,include:

[0040] Step 5.1, based on the user's course selection record and the video viewing record d days before the course, obtain the video interaction time series matrix set A;

[0041] Step 5.2: Input the video interaction timing matrix A into the interaction feature extraction module, which is a bidirectional LSTM model, to obtain the interaction feature F T .

[0042] Furthermore, the deep correlation feature F R and the interaction feature F T, calculate the probability of a user dropping out of his or her elective courses, including:

[0043] Step 6.1: Deep correlation feature F R and the interaction feature F T After concatenation, the data is sent to a multi-layer perceptron and the sigmoid function is used to constrain the output between 0 and 1 to obtain the probability of a user dropping out of their elective course:

[0044]

[0045] Where y represents the probability of a user dropping out of his / her elective course, and MLP(·) represents a multi-layer perceptron;

[0046] Step 6.2: All parameters in the above model are optimized by minimizing the following objective function using stochastic gradient descent:

[0047]

[0048] where Θ is the parameter set, y k Corresponding to the elective relationship E k User u i Whether from course c j is the true value of dropout, λ is the regularization coefficient, and K represents the number of elective records.

[0049] The present application proposes a method for predicting dropouts of online courses based on association relationships and video interactions. First, the original data is modeled to construct a heterogeneous information graph of the user's course video and a video interaction time series matrix. Then, on the one hand, the heterogeneous information graph is input into a deep association relationship feature extraction module based on node triples to obtain the association feature representations of the user and the course respectively. Then, the deep association relationship features are obtained through feature fusion. On the other hand, the video interaction time series matrix is ​​input into an interaction feature extraction module to obtain interaction features. Finally, the dropout probability of the user in his elective course is calculated by combining the deep association relationship features and the interaction features. The present application models the user's online learning situation well as a heterogeneous information graph. By mining the complex and diverse node association relationships in the graph, a node representation with stronger representation ability is constructed. The interaction features extracted are combined with the video interaction information with the largest amount of information in the real situation, which effectively improves the dropout prediction effect of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 is a flowchart of the online course dropout prediction method based on association relationship and video interaction of the present application;

[0051] Figure 2 It is a schematic diagram of a heterogeneous information graph of a course video for a user of this application;

[0052] Figure 3 It is a schematic diagram of the node embedding module of this application;

[0053] Figure 4 It is a schematic diagram of the deep correlation feature extraction module of the present application;

[0054] Figure 5 It is a schematic diagram of the interactive feature extraction module of this application. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application 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 application and are not used to limit the present application.

[0056] In one embodiment, Figure 1 As shown, a method for predicting dropout of online courses based on association relationship and video interaction includes:

[0057] Step 1: Based on the user's course selection record, the user's video viewing record and the course video resource record, a heterogeneous information graph G of the user's course video learning is constructed, and the node embedding corresponding to each node in the graph G is obtained through the node embedding module;

[0058] Specifically, they include:

[0059] Step 1.1: Based on the user's course selection record, the user's video viewing record and the course's video resource record, a heterogeneous information graph G(U,C,V) is formed, where U represents the set of all user nodes, C represents the set of all course nodes, V represents the set of all video nodes, the edge between the user and the course represents the user's selection relationship with the course, the edge between the user and the video represents the user's viewing relationship with the video, and the edge between the course and the video represents the subordinate relationship between the video and the course;

[0060] Step 1.2: Based on the heterogeneous information graph G, use one-hot encoding to uniquely identify all user nodes, course nodes, and video nodes in the graph G;

[0061] Step 1.3: Input the unique hot encoding of all nodes obtained in the previous step into the node embedding module, wherein the node embedding module includes a fully connected layer, obtain the parameter matrix of the fully connected layer, and obtain the node embedding corresponding to each node by searching the parameter matrix of the fully connected layer.

[0062] In this embodiment, the heterogeneous information graph G(U, C, V) of the user learning course video is as follows: Figure 2As shown, the edge between the user node and the course node represents the elective relationship, the edge between the user node and the video node represents the viewing relationship, and the edge between the course node and the video node represents the subordinate relationship.

[0063] In this embodiment, the node embedding module is as follows: Figure 3 As shown, the unique hot encodings of the five nodes are input into the fully connected layer with 3 neurons to obtain the weight matrix of the fully connected layer. The weight matrix of the fully connected layer is the low-dimensional dense and continuous vector representation of all nodes in the graph G. For the node x3 = [0, 0, 1, 0, 0], the low-dimensional dense and continuous embedding representation of the node [0.3, 0.2, 0.5] is obtained by table lookup (looking up the parameter matrix of the fully connected layer). It is worth noting that through the back-propagation learning of the entire model, the embedding representation of each node will contain the interaction relationship between nodes.

[0064] Step 2: embed the nodes of each node into a predefined user triplet set t u Input into the deep correlation feature extraction module to obtain the user's correlation feature representation f u ;

[0065] Specifically, Figure 4 As shown, including:

[0066] Step 2.1: Construct the user triple set t according to the association relationship between the user node and any node u .

[0067] The constructed user triple set is expressed as Where m represents m types of user triples. For any triple η∈t u , η can be expressed as <User,X,Y>, where User represents the user node type, X and Y represent any node type, such as user node type User, course node type Course and video node type Video.

[0068] Step 2.2: Based on any triple η∈t u , sample neighbor nodes for different user nodes to obtain the first-order neighbor node set of all users, and then sample neighbor nodes for the first-order neighbor nodes of all user nodes in turn to obtain the second-order neighbor node set of the user.

[0069] Based on any triple η∈t u , sample neighbor nodes for different user nodes, and get the first-order neighbor node set of all users expressed as There are n u user nodes, the first-order neighbor node set of user node u1 is expressed as And so on. Then sample neighbor nodes for the first-order neighbor nodes to obtain the user's second-order neighbor node set Same reason Represents the set of second-order neighbor nodes of user node u1, and so on.

[0070] Step 2.3: Based on any triple η∈t u ,Using the correlation measurement method, the similarity matrix between all user nodes and their second-order neighbor nodes is obtained.

[0071] Based on any triple η∈t u , using the HeteSim correlation measurement method, the similarity matrix between all user nodes and their second-order neighbor nodes is expressed as:

[0072]

[0073] Where α1(u i |η) represents user node u i The similarity vector with its second-order neighbor nodes based on triple η, is the vector α1(u i Any element in |η) represents the user node u i Any second-order neighbor node based on the triple η The similarity value of .

[0074] Step 2.4: Based on any triple η∈t u , multiply the similarity value between the user node and its second-order neighbor node and the node embedding of the second-order neighbor node itself, and obtain the similarity-guided embedding representation of the second-order neighbor nodes of all user nodes under the triple η.

[0075] The following generates second-order neighbor nodes Similarity-guided embedding representation For example:

[0076]

[0077] in Represents the second-order neighbor node The node itself is embedded.

[0078] Step 2.5: Based on any triple η∈t u , apply the soft attention mechanism to obtain the associated feature representation of the user under the triple η, and traverse the triple set t u , and obtain the associated feature representations of all users under different triples.

[0079] Based on any triple η∈t uAccording to the soft attention mechanism shown in formulas (2) and (3), we can get user u i The associated feature representation f under the triple η η (u i ), similarly, we get the associated feature representation of all users under triple η:

[0080]

[0081]

[0082] in Represents user u i An arbitrary second-order neighbor node under the triple η, is its similarity-guided embedding representation, v1, w1, b1 are all learnable parameter vectors, and tanh() is the activation function.

[0083] Traverse the triple set t u Repeat steps 2.2 to 2.5 to obtain the associated feature representations of all users under different triples in Represents a concatenation operation.

[0084] Step 2.6: Perform a self-attention mechanism operation on the associated feature representations of all users under different triples to obtain the associated feature representations of the users.

[0085] According to the self-attention mechanism shown in formulas (4), (5), (6) and (7), the user's associated feature representation f is obtained u , as follows:

[0086]

[0087] in are all learnable parameter matrices, σ is the sigmoid activation function,

[0088] This embodiment constructs multiple triples on the user node type to describe the possible association relationship between the user node and any node.

[0089] For example, the triple of user node type and user node type It has a triple instance <u1,c1,u2>, which indicates that two different user nodes u1 and u2 have both taken the course c1, thus establishing a classmate relationship between users who have taken the same course. It has a triple instance 〈u1,v1,c1〉, which indicates that user u1 watched video v1 belonging to course c1, thereby constructing the video watching relationship of the user in the course.

[0090] In this embodiment, the deep correlation feature extraction module is as follows: Figure 4 As shown. For user node u i First, based on a single triplet η, sample the user node u i The set of second-order neighbor nodes Then use HeteSim as the correlation measurement method to get the user node u i and any of its second-order neighbor nodes The similarity value of And thus generate the second-order neighbor nodes Similarity-guided embedding representation Then, the soft attention mechanism is used to obtain the attention weight of the second-order neighbor node to the user node. And thus get the user node u i Relation feature representation based on a single association relationship η (u i ), traverse user node u i Repeat the above steps and use the self-attention mechanism to finally get the user node u i The association feature representation f that integrates multiple association relationships u (u i ).

[0091] Step 3: embed the nodes of each node into a predefined course triple set t c Input into the deep correlation feature extraction module to obtain the course correlation feature representation f c .

[0092] Specifically, they include:

[0093] Step 3.1: Construct a course triplet set t according to the relationship between the course node and any node c .

[0094] The constructed course triple set is expressed as Where n represents n types of course triples. For any triple ξ∈t c , ξ can be expressed as <Course,X,Y>, where Course represents the course node type, and X and Y represent any node type, such as user node type User, course node type Course, and video node type Video.

[0095] Step 3.2: Based on any triple ξ∈tc , sample neighbor nodes for different course nodes to obtain the first-order neighbor node set of all courses, and then sample neighbor nodes for the first-order neighbor nodes of all course nodes in turn to obtain the second-order neighbor node set of the course.

[0096] Based on any triple ξ∈t c , sample neighbor nodes for different course nodes, and get the first-order neighbor node set of all courses as There are n c course nodes, the first-order neighbor node set of course node c1 is expressed as And so on. Then sample the neighbor nodes of the first-order neighbor nodes of the course node to obtain the second-order neighbor node set of the course Same reason Represents the set of second-order neighbor nodes of course node c1, and so on.

[0097] Step 3.3: Based on any triple ξ∈t c ,The correlation measurement method is used to obtain the similarity matrix between all course nodes and their second-order neighbor nodes.

[0098] Based on any triple ξ∈t c , using the HeteSim correlation measurement method, the similarity matrix of all course nodes and their second-order neighbor nodes is expressed as:

[0099]

[0100] Where α1(c j |ξ) represents the course node c j The similarity vector with its second-order neighbor nodes based on triple ξ, is the vector α1(c j |ξ) represents any element in the course node c j Any second-order neighbor node based on the triple ξ The similarity value of .

[0101] Step 3.4: Based on any triple ξ∈t c , multiply the similarity value between the course node and its second-order neighbor node and the node embedding of the second-order neighbor node itself, and obtain the similarity-guided embedding representation of the second-order neighbor nodes of all course nodes under the triple ξ.

[0102] The following generates second-order neighbor nodes Similarity-guided embedding representation For example:

[0103]

[0104] in Represents the second-order neighbor node The node itself is embedded.

[0105] Step 3.5: Based on any triple ξ∈t c , apply the soft attention mechanism, get the associated feature representation of the course under the triple ξ, and traverse the triple set t c , and obtain the associated feature representations of all courses under different triples.

[0106] Based on any triple ξ∈t c According to the soft attention mechanism shown in formulas (9) and (10), we get course c j The associated feature representation f under the triple ξ ξ (c j ), and similarly, we get the associated feature representation of all courses under the triple ξ:

[0107]

[0108] in Indicates course c j For any second-order neighbor node under the triple ξ, is its similarity-guided embedding representation, v2, w2, b2 are all learnable parameter vectors, and tanh() is the activation function.

[0109] Traverse the triple set t c Repeat steps 3.2 to 3.5 to get the associated feature representation of all courses under different triples in Represents a concatenation operation.

[0110] Step 3.6: Perform self-attention mechanism operation on the associated feature representations of all courses under different triples to obtain the associated feature representation of the course f c .

[0111] According to the self-attention mechanism shown in formulas (11), (12), (13) and (14), the associated feature representation f of the course is obtained c , as follows:

[0112]

[0113] in are all learnable parameter matrices, σ is the sigmoid activation function,

[0114] This embodiment constructs multiple triples on the course node type to describe the possible association relationship between the course node and any node.

[0115] For example, the triple of course node type and course node type It has a triple instance <c1,u2,c3>, which indicates that two different course nodes c1 and c3 are selected by the same user u2, thereby constructing a similarity relationship between courses selected by the same user.

[0116] This embodiment designs different triplets for user node types and course node types as shown in the following table. In other embodiments, triplets can be freely designed and combined according to actual needs.

[0117]

[0118]

[0119] Table 1- Triple corresponding node association relationship description table

[0120] Step 4: Based on the user and course selection relationship pair, the user's associated features f u and the associated features of the course f c Perform feature fusion to obtain deep correlation feature F R .

[0121] Specifically, they include:

[0122] Step 4.1: According to the course selection record of the user, obtain the set E of the relationship pairs between the user and the course selection.

[0123] The set of user and course selection relationship pairs is represented as E = [E1,…,E k ,…,E K ], where K means there are K elective records. For the kth elective record E k , we can express it as (u i ,c j ), that is, user u i ∈U took course c j ∈C.

[0124] Step 4.2: Based on the set E of user and course selection relationship pairs, the user's associated features f u and the associated characteristics of the courses they choose c Concatenate and flatten, then send to the fully connected layer and activate with the sigmoid function to get the final deep correlation feature F R .

[0125] The final deep correlation feature is expressed as The kth user and course selection relationship pair (u i ,cj )'s deep correlation characteristics For example:

[0126]

[0127] Where δ is a flattening function that flattens the matrix into a row vector, and w3 and b3 are both trainable parameter vectors.

[0128] Step 5: Construct a video interaction timing matrix A based on the user's course selection record and the user's video viewing record, and input the video interaction timing matrix A into the interaction feature extraction module to obtain the interaction feature F T .

[0129] Specifically, they include:

[0130] Step 5.1: Based on the user's course selection record and his / her video viewing record d days before taking the course, obtain the video interaction time series matrix set A.

[0131] Video interaction timing matrix set A = [A (1) ,…,A (k) ,…,A (K) ], where the kth user-course selection relationship pair (u i ,c j )’s video interaction timing matrix is ​​A (k) =[a1,…,a e ,…,a d ], a e Represents user u i Study Coursec j Statistics of course videos watched on day e. d is the preset number of days.

[0132] Step 5.2: Input the video interaction timing matrix A into the interaction feature extraction module to obtain the interaction feature F T .

[0133] In this embodiment, in order to predict the probability of a user dropping out of the course in the early stage of learning, the data of the first 7 days of the user's learning in the original data is used as the data set, that is, d = 7. The data set records a total of 16 types of user activity records of watching course videos, including the user's video viewing time, the user's video viewing speed, the number of fast forwards, etc., so a e The dimension is 1×16.

[0134] The interactive feature extraction module used in this embodiment is a bidirectional LSTM model. Figure 5 As shown in the figure, each day's video viewing data will be input as a time step and input into the bidirectional LSTM, thereby extracting the interactive features.

[0135] Step 6: Based on the deep correlation feature F R and the interaction feature F T , calculate the probability of a user dropping out of his or her elective courses.

[0136] Specifically, they include:

[0137] Step 6.1: Deep correlation feature F R and the interaction feature F T After concatenation, the data is sent to a multi-layer perceptron and the sigmoid function is used to constrain the output between 0 and 1 to obtain the probability of a user dropping out of their elective course:

[0138]

[0139] Where y represents the probability of a user dropping out of his / her elective course, and MLP(·) represents a multi-layer perceptron;

[0140] Step 6.2: All parameters in the above model are optimized by minimizing the following objective function using stochastic gradient descent:

[0141]

[0142] where Θ is the parameter set, y k Corresponding to the elective relationship E k User u i Whether from course c j is the true value of dropout, λ is the regularization coefficient, and K means there are K elective records.

[0143] After completing the above steps, the dropout prediction of users in online education can be achieved.

[0144] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. A method for predicting dropout of online courses based on association relationship and video interaction, characterized in that: The online course dropout prediction method based on association relationship and video interaction includes: According to the user's course selection record, the user's video viewing record and the course video resource record, a heterogeneous information graph G of the user's course video learning is constructed. After the node embedding module, the node embedding corresponding to each node in the graph G is obtained. The nodes of each node are embedded in a predefined user triplet set t u Input into the deep correlation feature extraction module to obtain the user's correlation feature representation f u ; The nodes of each node are embedded in a predefined course triple set t c Input into the deep correlation feature extraction module to obtain the course correlation feature representation f c ; Based on the relationship between users and course selections, the user's associated features f u and the associated features of the course f c Perform feature fusion to obtain the deep correlation feature F of the user and course selection relationship pair R ; According to the user's course selection record and the user's video viewing record, the video interaction timing matrix A is constructed, and the video interaction timing matrix A is input into the interaction feature extraction module to obtain the interaction feature F T ; Based on the deep correlation feature F R and the interaction feature F T , calculate the probability of users dropping out of their elective courses; Wherein, the nodes of each node are embedded in a predefined user triplet set t u Input into the deep correlation feature extraction module to obtain the user's correlation feature representation f u ,include: Step 2.1: Construct the user triple set t according to the association relationship between the user node and any node u ; Step 2.2: Based on any triple η∈t u , sample neighbor nodes of different user nodes to obtain the first-order neighbor node set of all users, and then sample neighbor nodes of the first-order neighbor nodes of all user nodes in turn to obtain the second-order neighbor node set of the user; Step 2.3: Based on any triple η∈t u ,Using the correlation measurement method, we get the similarity matrix between all user nodes and their second-order neighbor nodes; Step 2.4: Based on any triple η∈t u , multiply the similarity value between the user node and its second-order neighbor node and the node embedding of the second-order neighbor node itself, and obtain the embedding representation of the second-order neighbor nodes of all user nodes under the triple η based on similarity guidance; Step 2.5: Based on any triple η∈t u , apply the soft attention mechanism to obtain the associated feature representation of the user under the triple η, and traverse the triple set t u , get the associated feature representation of all users under different triples; Step 2.6: Perform a self-attention mechanism operation on the associated feature representations of all users under different triples to obtain the associated feature representations of the users; The nodes of each node are embedded in a predefined course triple set t c Input into the deep correlation feature extraction module to obtain the course correlation feature representation f c ,include: Step 3.1: Construct a course triplet set t according to the relationship between the course node and any node c ; Step 3.2: Based on any triple ξ∈t c , sample neighbor nodes for different course nodes to obtain the first-order neighbor node set of all courses, and then sample neighbor nodes for the first-order neighbor nodes of all course nodes in turn to obtain the second-order neighbor node set of the course; Step 3.3: Based on any triple ξ∈t c ,Using the correlation measurement method, we get the similarity matrix between all course nodes and their second-order neighbor nodes; Step 3.4: Based on any triple ξ∈t c , multiply the similarity value between the course node and its second-order neighbor node and the node embedding of the second-order neighbor node itself, and obtain the embedding representation of the second-order neighbor nodes of all course nodes under the triple ξ based on similarity guidance; Step 3.5: Based on any triple ξ∈t c , apply the soft attention mechanism, get the associated feature representation of the course under the triple ξ, and traverse the triple set t c , get the associated feature representation of all courses under different triples; Step 3.6: Perform self-attention mechanism operation on the associated feature representations of all courses under different triples to obtain the associated feature representation of the course f c ; The video interaction timing matrix A is constructed according to the user's course selection record and the user's video viewing record, and the video interaction timing matrix A is input into the interaction feature extraction module to obtain the interaction feature F T ,include: Step 5.1, based on the user's course selection record and the video viewing record d days before the course, obtain the video interaction time series matrix set A; Step 5.2: Input the video interaction timing matrix A into the interaction feature extraction module, which is a bidirectional LSTM model, to obtain the interaction feature F T .

2. The online course dropout prediction method based on association relationship and video interaction according to claim 1 is characterized in that: According to the course selection record of the user, the video viewing record of the user and the video resource record of the course, a heterogeneous information graph G of the user's course video learning is constructed, and the node embedding corresponding to each node in the graph G is obtained through the node embedding module, including: Based on the user's course selection record, the user's video viewing record and the course's video resource record, a heterogeneous information graph G(U, C, V) is formed, where U represents the set of all user nodes, C represents the set of all course nodes, V represents the set of all video nodes, the edge between the user and the course represents the user's selection relationship with the course, the edge between the user and the video represents the user's viewing relationship with the video, and the edge between the course and the video represents the subordinate relationship between the video and the course; Based on the heterogeneous information graph G, one-hot encoding is used to uniquely identify all user nodes, course nodes, and video nodes in the graph G; The one-hot encoding of all nodes obtained in the previous step is input into the node embedding module, which includes a fully connected layer. The parameter matrix of the fully connected layer is obtained, and the node embedding corresponding to each node is obtained by searching the parameter matrix of the fully connected layer.

3. The online education dropout prediction method based on association relationship and video interaction according to claim 1 is characterized in that: Based on the user and course selection relationship pair, the user's associated feature f u and the associated features of the course f c Perform feature fusion to obtain the deep correlation feature F of the user and course selection relationship pair R ,include: Step 4.1, according to the course selection record of the user, obtain the set E of the relationship pairs between the user and the course selection; Step 4.2: Based on the set E of user and course selection relationship pairs, the user's associated features f u and the associated characteristics of the courses they choose c Concatenate and flatten, then send to the fully connected layer and activate with the sigmoid function to get the final deep correlation feature F R .

4. The online education dropout prediction method based on association relationship and video interaction according to claim 1 is characterized in that: Based on the deep correlation feature F R and the interaction feature F T , calculate the probability of a user dropping out of his or her elective courses, including: Step 6.1: Deep correlation feature F R and the interaction feature F T After concatenation, the data is sent to a multi-layer perceptron and the sigmoid function is used to constrain the output between 0 and 1 to obtain the probability of a user dropping out of their elective course: Where y represents the probability of a user dropping out of his / her elective course, and MLP(·) represents a multi-layer perceptron; Step 6.2: All parameters in the above model are optimized by minimizing the following objective function using stochastic gradient descent: where Θ is the parameter set, y k Corresponding to the elective relationship E k User u i Whether from course c j is the true value of dropout, λ is the regularization coefficient, and K represents the number of elective records.

Images