CN112734105B - Method for preventing breaking behavior in online education - Google Patents

Method for preventing breaking behavior in online education Download PDF

Info

Publication number
CN112734105B
CN112734105B CN202110022690.0A CN202110022690A CN112734105B CN 112734105 B CN112734105 B CN 112734105B CN 202110022690 A CN202110022690 A CN 202110022690A CN 112734105 B CN112734105 B CN 112734105B
Authority
CN
China
Prior art keywords
user
course
feature
behavior
attention
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202110022690.0A
Other languages
Chinese (zh)
Other versions
CN112734105A (en
Inventor
郝鹏翼
吴冬恩
郑宇祥
韩天星
白琮
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Zhejiang University of Technology ZJUT
Original Assignee
Zhejiang University of Technology ZJUT
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Zhejiang University of Technology ZJUT filed Critical Zhejiang University of Technology ZJUT
Priority to CN202110022690.0A priority Critical patent/CN112734105B/en
Publication of CN112734105A publication Critical patent/CN112734105A/en
Application granted granted Critical
Publication of CN112734105B publication Critical patent/CN112734105B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10—Complex mathematical operations
    • G06F17/16—Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/047—Probabilistic or stochastic networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
    • G06Q50/10—Services
    • G06Q50/20—Education
    • G06Q50/205—Education administration or guidance

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Business, Economics & Management (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Mathematical Physics (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Strategic Management (AREA)
  • Human Resources & Organizations (AREA)
  • Tourism & Hospitality (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Economics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Probability & Statistics with Applications (AREA)
  • General Business, Economics & Management (AREA)
  • Marketing (AREA)
  • Educational Technology (AREA)
  • Mathematical Optimization (AREA)
  • Educational Administration (AREA)
  • Mathematical Analysis (AREA)
  • Computational Mathematics (AREA)
  • Pure & Applied Mathematics (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
  • Development Economics (AREA)
  • Quality & Reliability (AREA)
  • Algebra (AREA)

Abstract

The invention discloses a method for preventing a learning-dropping behavior in online education, which comprises the steps of firstly carrying out data depth mining on original data acquired from an online education platform to obtain an activity time sequence matrix, context information and a friend relation of a user, then respectively inputting the activity time sequence matrix, the context information and a course selection characteristic similarity relation into respective constructed and trained characteristic extraction modules to obtain a deep sequence characteristic, an attention characteristic and a classmatic characteristic of the user, finally carrying out characteristic fusion on the three characteristics, and inputting the fused characteristics into a behavior prediction model to obtain a learning-dropping behavior prediction result of the user. The invention greatly enhances the prediction capability of the network by means of a plurality of factors influencing the behavior of the conjugate, and has stronger generalization capability.

Description

Method for preventing breaking behavior in online education
Technical Field
The invention belongs to the field of prediction of a learning dropping behavior of online education, and particularly relates to a method for preventing a learning dropping behavior in online education.
Background
In recent years, the development of large-scale online open courses (MOOCs) has been rapidly progressed, resulting in a number of excellent online education platforms (such as courera, Udacity, edX, and schoolhouse online, etc.), which provide efficient and convenient learning resources and learning opportunities for users by providing thousands of carefully designed online courses, so that education is not limited by time and place, users can learn for their life, and thus a large number of learning users are attracted.
Although online education is continuously developed and users participating in learning are continuously increased, the effectiveness of online education is questioned. In fact, the problem of high leakage rate of online education is widely existed in online education platforms due to the problems of poor interactivity, uneven education level of users and the like of online education. Statistics show that the average completion rate of courses on the world famous online education platform edX is only 5%, and that the average completion rate of courses is only 4.5% after the statistics of 1000 courses on the online education platform "class online" established by Qinghua university. Low completion rates have therefore become a major obstacle to further development of online education.
The Chinese patent application with publication number CN110059716A discloses a CNN-LSTM-SVM network model construction and MOOC (learning object model) learning-by-learning prediction method, which comprises the steps of processing the original online learning behavior records of known students to obtain a plurality of time slice matrixes, and performing network training according to the time slice matrixes to obtain a trained CNN-LSTM network; finally, a trained network of the prediction of the conjugate is obtained. However, the method extracts features according to the time slice matrix of the user, and factors influencing the behavior of the user after learning are numerous, so the extracted features are not comprehensive enough, and the prediction performance is relatively general.
The Chinese patent application with the publication number of CN109558983 discloses a method and a device for predicting the rate of dropped learning of network courses. However, the method only considers the characteristics of the behavior information, and actually, the differences of the users participating in online education are large, and the influence of the same behavior of different types of users on the behavior of the user to the study is different, so that the method does not fully utilize the data, and the prediction performance is general.
The problems that extracted features are not comprehensive enough, prediction accuracy and precision are poor and the like exist generally in the prior art, and due to the fact that prediction is not accurate enough, users are difficult to be reminded in a targeted mode to prevent the behavior of breaking.
Disclosure of Invention
The application aims to provide a method for preventing a learning-breaking behavior in online education, which is used for solving the problems of incomplete extracted features, poor prediction accuracy, poor precision and the like in the prior art.
In order to achieve the purpose, the technical scheme of the application is as follows:
a method of preventing a dropped behavior in online education comprising:
obtaining an activity time sequence matrix R of the user from an online education platform, inputting the activity time sequence matrix R to a time sequence feature extraction module which is constructed and trained to obtain deep sequence features F of the userR；
Obtaining context information N of a user from an online education platform, and combining the context information N and deep sequence characteristics FRConnecting, inputting to the constructed and trained attention feature generation module to obtain the attention feature F of the userN；
Obtaining course selection characteristic similarity relation G of the user from the online education platform, inputting the course selection characteristic similarity relation G into a classmate feature extraction module which is constructed and trained to obtain classmate features F of the userG；
Will be characterized by FR、FNAnd FGDepth feature fusion is carried out to obtain the dropping behavior feature F of the userfision；
Characterizing a dropped behavior of the user FfusionInputting a constructed and trained behavior prediction model, obtaining a prediction result of the user's behavior of dropping learning, and sending reminding information to the user who has the behavior of dropping learning.
Further, the activity time sequence matrix R of the user is obtained from the online education platform and is input to the constructed and trained time sequence feature extraction module, and the deep sequence feature F of the user is obtainedRThe method comprises the following steps:
step 1.1, constructing an activity time sequence matrix R of the user through the records of n activities of the user within m days after the start of the course:
Figure BDA0002889215800000031
wherein, when i is less than or equal to m and j is less than or equal to n, ri,jRepresenting the number of occurrences of activity j by the user on the ith day after the start of the course(ii) a When i is not more than m, j is n +1, ri,n+1Representing the total number of times that the user's activity occurred during the ith day after the start of the course; when i is not more than m, j is n +2, ri,n+2Representing the total activity duration of the user on the ith day after the beginning of the course; when i is m +1, j is less than or equal to n, rm+1,jRepresenting the total number of the ith activity of the user within m days after the beginning of the course; when i is m +1 and j is n +1, rm+1,n+1Representing the total number of times that all activities of the user occur within m days after the beginning of the course; when i is m +1 and j is n +2, rm+1,n+2Representing the total activity duration of the user within m days after the beginning of the course;
step 1.2, inputting the activity time sequence matrix R of the user into a time sequence feature extraction module to obtain a deep sequence feature FR。
Further, the context information N of the user is obtained from the online education platform, and the context information N and the deep sequence characteristics F are combinedRConnecting, inputting to the constructed and trained attention feature generation module to obtain the attention feature F of the userNThe method comprises the following steps:
step 2.1, constructing a behavior matrix R' of the user according to the activity record of the user:
R′＝[r′1……r′i…r′n]
wherein n represents a total of n activities, r'iRepresenting the total number of times of the ith activity of the user;
step 2.2, classifying the users by using a clustering algorithm through the behavior matrix R' of the users to obtain user types, and combining the user types with the ages, sexes, academic calendars and course types and difficulties of the users to form the context information N of the users:
N＝[n1 n2 n3 n4 n5 n6]
wherein n is1To n6Sequentially representing the user type, the user age, the user gender, the user study history, the course type and the course missing rate;
step 2.3, the context information N and the deep sequence characteristics F of the userRIs connected to obtainThe deep sequence characteristics of the context information are fused;
step 2.4, inputting the deep sequence features fused with the context information into an attention feature generation module to obtain an attention feature FN。
Further, the course selection characteristic similarity relation G of the user is obtained from the online education platform, the course selection characteristic similarity relation G is input to a classmate characteristic extraction module which is constructed and trained, and the classmate characteristic F of the user is obtainedGThe method comprises the following steps:
step 3.1, according to the Course selection records SR (User, Course) of the users, regarding each User as a User node, regarding each Course as a Course node, and forming a Course selection record bipartite graph BG(User, Course, Edge), where User represents the set of all User nodes: useriBelongs to the User; course represents the set of all Course nodes: coursejE.g. Course; edge represents the set of edges connecting two nodes: edgek＝(Useri,Coursej) E, the Edge is equal to SR, which indicates that the user node i selects and repairs the course node j;
step 3.2, recording the course selection into bipartite graph BGInputting the data into an unsupervised graph neural network to obtain the characteristics F of all user nodesu：
Figure BDA0002889215800000041
Where t represents the total number of users, FuEach row u iniThe course selection characteristics of the user i are represented;
step 3.3 course selection characteristics u for different usersi、ujAnd performing similarity calculation, determining that the course selection characteristics are similar when the defined similarity is greater than a certain value, and obtaining a course selection characteristic similarity relation G of the user:
Figure BDA0002889215800000042
wherein each row in G represents a pair of course selection feature phasesSimilar relation, (User)i,Userj) Representing that the user i and the user j are similar to each other in course selection characteristics;
step 3.4, forming a classmate relation graph G according to the course selection characteristic similarity relation G of the userclassmate，GclassmateEach node in the system represents a Course selection record SR (User, Course), the edge connecting the two nodes represents that the users represented by the two nodes are in the Course selection characteristic similarity relation, and the selected courses are consistent;
step 3.5, drawing the relationship graph G of the classmatesclassmateInputting the data into a classmate feature extraction module to generate a classmate feature FG。
Further, the feature FR、FNAnd FGDepth feature fusion is carried out to obtain the dropping behavior feature F of the userfusionThe method comprises the following steps:
step 4.1, attention feature FNSoftmax activation was performed according to the following formula, and the attention weight W of each day after the course was started was obtainedattention＝(w1,w2,…,wk)：
Figure BDA0002889215800000051
Wherein, wiIndicating the attention weight on day i after the start of the course, FN,iExpress attention feature FNK represents the total number of days recorded in the raw data;
step 4.2, the attention weight W obtained in the last stepattentionAccording to the following formula and deep sequence characteristics FRMultiplying to obtain the deep sequence features combined with attention
Figure BDA0002889215800000053
Figure BDA0002889215800000052
Wherein k represents FRNumber of lines of，FR,iIs represented by FRThe value of row i;
step 4.3, deep sequence features to be combined with attention
Figure BDA0002889215800000054
Characteristic of classmate FGAdding and splicing attention feature FNObtaining a feature fusion vector Fffusion：
Figure BDA0002889215800000055
Further, the characterizing of the user's dropped behavior as FfusionInputting the user behavior prediction model which is built and trained to obtain a prediction result of the user behavior of the science, wherein the prediction result comprises the following steps:
step 5.1, fusing the feature into a vector FfusionInputting the data into a constructed and trained behavior prediction model to obtain a classification result PoutputIf P isoutputIf 1 indicates that the user is predicted as a conjugate, if Poutput0 indicates that the user is predicted as not likely to be learned.
The method for preventing the behavior of the missed learning in the online education comprises the steps of firstly carrying out data deep mining on original data to obtain the similar relation among an activity time sequence matrix, context information and course selection characteristics of a user, then inputting the similar relation among the activity time sequence matrix, the context information and the course selection characteristics of the user to a constructed and trained characteristic extraction module to obtain deep sequence characteristics, attention characteristics and classmatic characteristics of the user, and then sending the characteristics to a behavior prediction model to predict the behavior of the missed learning. According to the method and the device, data are deeply mined firstly, the mined data are utilized to extract features, different attention features are generated for different types of users by utilizing user information and course information, the problem that the users participating in online education are small in difference and diverse is solved, a classmate relation graph is ingeniously constructed by utilizing a graph mode, the model can be assisted to accurately predict the behavior of the dropped students, and the prediction effect of the model is improved. Therefore, the user is reminded through various modes, online learning is continued, and the behavior of learning dropping is reduced.
Drawings
FIG. 1 is a flow chart of a method of the present application for preventing a dropped school behavior in online education;
FIG. 2 is a schematic diagram of a bipartite graph of course selection records of the present application;
FIG. 3 is a schematic diagram of a forward LSTM module and a reverse LSTM module of the present application;
FIG. 4 is a schematic representation of the relationship between the present application;
FIG. 5 is an attention feature generation module;
FIG. 6 is a behavior prediction model.
Detailed Description
In order to make the objects, technical solutions and advantages of the present application more apparent, the present application is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present application and are not intended to limit the present application.
A method for preventing a dropped school behavior in online education, as shown in fig. 1, includes:
step S1, obtaining the activity time sequence matrix R of the user from the online education platform, inputting the activity time sequence matrix R to the constructed and trained time sequence feature extraction module, and obtaining the deep sequence feature F of the userR。
The online education platform accumulates a large amount of raw data related to the user during operation, such as user activity logs, user conversations, user personal information (age, sex, calendar), course information (type of course, start time of course), and so on. According to the method and the system, data deep mining is carried out on original data provided by an online education platform to obtain data such as an activity time sequence matrix R of a user, context information N of the user, class selection characteristic similarity relation G of the user and the like.
The activity time sequence matrix R of the user is obtained from the online education platform and is input to the time sequence feature extraction module which is constructed and trained to obtain the deep sequence feature F of the userRThe method specifically comprises the following steps:
step 1.1, constructing an activity time sequence matrix R of the user through the records of n activities of the user within m days after the start of the course:
Figure BDA0002889215800000071
wherein, when i is less than or equal to m and j is less than or equal to n, ri,jRepresenting the number of occurrences of activity j of the user on the ith day after the start of the course; when i is not more than m, j is n +1, ri,n+1Representing the total number of times that the user's activity occurred during the ith day after the start of the course; when i is not more than m, j is n +2, ri,n+2Representing the total activity duration of the user on the ith day after the beginning of the course; when i is m +1, j is less than or equal to n, rm+1,jRepresenting the total number of times of the ith activity of the user within m days after the beginning of the course; when i is m +1 and j is n +1, rm+1,n+1Representing the total number of times of all activities of the user within m days after the beginning of the course; when i is m +1 and j is n +2, rm+1,n+2Representing the total duration of the user's activity within m days after the start of the lesson.
Step 1.2, inputting the activity time sequence matrix R of the user into a time sequence feature extraction module to obtain a deep sequence feature FRWith dimension of heightR×widthR。
In the present embodiment, the data set used records 21 activities of the user within 36 days from the beginning of the lesson, i.e., m is 36, and n is 21, so the size of the activity time series matrix R of the user is 37 × 23, and the activity time series matrix R of the user can be expanded or reduced according to actual needs.
The time sequence feature extraction module adopted in this embodiment is a bidirectional LSTM model, as shown in fig. 3, the time sequence feature extraction module includes an LSTM module with a depth of 2, the number of neurons is 128, the number of nodes in the hidden layer is 30, and the forward direction and the reverse direction LSTM refer to an input receiving sequence, and are input in a forward direction according to a time sequence and input in a reverse direction according to a time reverse sequence.
Step S2, obtaining the context information N of the user from the online education platform, and combining the context information N and the deep sequence characteristics FRConnected to, input toThe attention feature generation module is built and trained to obtain the attention feature F of the userN。
The embodiment obtains the context information N of the user from the online education platform, and combines the context information N and the deep sequence characteristics FRConnecting, inputting to the constructed and trained attention feature generation module to obtain the attention feature F of the userNThe method comprises the following steps:
step 2.1, constructing a behavior matrix R' of the user according to the activity record of the user:
R′＝[r′1……r′i…r′n]
wherein n represents a total of n activities, r'iRepresenting the total number of times the user has taken place in the ith activity.
Step 2.2, classifying the users by using a clustering algorithm through the behavior matrix R' of the users to obtain user types, and combining the user types with the ages, sexes, academic calendars and course types and difficulties of the users to form the context information N of the users:
N＝[n1 n2 n3 n4 n5 n6]
wherein n is1To n6The user type, the user age, the user gender, the user study history, the course type and the course dropped rate are sequentially represented.
Step 2.3, the context information N and the deep sequence characteristics F of the userRConnecting to obtain deep sequence characteristics F fused with context informationn。
Deep sequence feature F fused with context informationn:
Figure BDA0002889215800000081
Step 2.4, inputting the deep sequence features fused with the context information into an attention feature generation module to obtain an attention feature FNWith dimension of heightN×widthN。
In the present embodiment, it is preferred that,the data set used has a total of 21 records of activity types of user activity, i.e. n-21. In this embodiment, a K-means clustering algorithm is adopted to classify users into 5 classes, so n1There are 5 kinds of values; this example defines males as 1 and females as 0; in this embodiment, different calendars are sorted from low to high in sequence, i.e. n4(ii) a In this embodiment, all the course types are numbered, i.e. n5；
The attention feature generation module adopted in the present embodiment is shown in fig. 5, and includes an embedding layer, a full connection layer, and softmax to calculate the attention weight.
Step S3, obtaining course selection characteristic similarity relation G of the user from the online education platform, inputting the course selection characteristic similarity relation G into a classmate characteristic extraction module which is constructed and trained to obtain classmate characteristics F of the userG。
Specifically, the course selection characteristic similarity relation G of the user is obtained from the online education platform, the course selection characteristic similarity relation G is input to a classmate characteristic extraction module which is constructed and trained, and the classmate characteristic F of the user is obtainedGThe method comprises the following steps:
step 3.1, according to the Course selection records SR (User, Course) of the users, regarding each User as a User node, regarding each Course as a Course node, and forming a Course selection record bipartite graph BG(User, Course, Edge), where User represents the set of all User nodes: useriBelongs to the User; course represents the set of all Course nodes: coursejE.g. Course; edge represents the set of edges connecting two nodes: edgek＝(Useri,Coursej) And e, Edge is SR, which indicates that the user node i has selected the course node j.
Step 3.2, recording the course selection into bipartite graph BGInputting the data into an unsupervised graph neural network to obtain the characteristics F of all user nodesuWith dimension of t × widthR：
Figure BDA0002889215800000091
Wherein, t represents the total number of users,Fueach row u iniIndicating the course selection characteristics of user i.
Step 3.3 course selection characteristics u for different usersi、ujAnd (3) performing similarity calculation, and determining that the similarity is greater than a certain value as a course selection characteristic similarity relation G:
Figure BDA0002889215800000092
wherein each row in G represents a pair of similar relationship of course selection characteristics, e.g. (User)i,Userj) Indicating that user i and user j are similar in course selection characteristics.
Step 3.4, forming a classmate relation graph G according to the course selection characteristic similarity relation G of the userclassmate，GclassmateEach node in the list represents a Course selection record SR (User, Course), the edge connecting the two nodes represents that the users represented by the two nodes are in the similar relationship of Course selection characteristics, and the selected courses are consistent.
Step 3.5, drawing the relationship graph G of the classmatesclassmateInputting the data into a classmate feature extraction module to generate a classmate feature FGWith dimension of 1 xwidthG。
In this embodiment, the class selection record bipartite graph BG(User, Course, Edge) As shown in figure 2, the unsupervised graph neural network adopted for the method is unsupervised graph SAGE, for each User node u, firstly, 10 nodes are randomly sampled from all nodes connected with edges of the User node u, the node characteristics of the 10 nodes are multiplied by weights one by one and then added, then, the average value is taken to obtain the neighbor node characteristics, and then, the neighbor node characteristics and the User node u are added to obtain the node characteristics of the User node u. And finally, using a Markov chain Monte Carlo Negative Sampling (MCNS) strategy to obtain a node which is most dissimilar to the characteristics of the user node u through negative sampling, calculating a difference value Loss between the two nodes, and updating the value of the weight through a back propagation algorithm. After the features of each user u are obtained, the similarity relationship of the course selection features is defined in a manner that the cosine similarity is greater than 0.85. In practical applicationIn the course selection method, a proper unsupervised graph neural network can be adopted according to different online education platforms, and a proper course selection characteristic similarity relation definition mode can be adopted to define course selection characteristic similarity relations.
Classmatic relation graph G in this exampleclassmateAs shown in FIG. 4, SR2、SR3、SR4Are all combined with SR1For the similar relationship of course selection characteristics, the relationship graph G of classmatesclassmateEach node SR in (1) represents a course selection record, and the edge connecting the two nodes represents that the two users are in a course selection characteristic similarity relationship, and the selected courses are consistent.
In this embodiment, the classmate feature extraction module performs random sampling with 10 numbers on all nodes connected with edges of the SR node, multiplies the node features of the 10 nodes by weights one by one, and then adds the node features, and then averages the node features to obtain the classmate features.
Step S4, feature FR、FNAnd FGDepth feature fusion is carried out to obtain the dropping behavior feature F of the userfusion。
Specifically, feature FR、FNAnd FGDepth feature fusion is carried out to obtain the dropping behavior feature F of the userfusionThe method comprises the following steps:
step 4.1, attention feature FNSoftmax activation was performed according to the following formula, obtaining the attention weight W for each day after the start of the courseattention＝(w1,w2,…,wk)：
Figure BDA0002889215800000101
Wherein, wiIndicating the attention weight on day i after the start of the course, FN,iFeature of attention FNK represents the total number of days recorded in the raw data;
step 4.2, the attention weight W obtained in the last stepattentionAccording to the following formula and deep sequence characteristics FRMultiply to obtain and combineDeep sequence characterization of attention
Figure BDA0002889215800000103
Figure BDA0002889215800000102
Wherein k represents FRNumber of lines of (F)R,iIs represented by FRThe value of row i;
step 4.3, deep sequence characteristics combined with attention
Figure BDA0002889215800000111
Characteristic of classmate FGAdding and splicing attention feature FNObtaining a feature fusion vector FfusionWith a dimension of 1 × widthfusionWherein widthfusion＝widthR+widthN：
Figure BDA0002889215800000112
Note that the data set used in the present embodiment records all user data within 36 days from the beginning of the lesson, that is, k is 36.
Step S5, characterizing the user' S dropped behavior FfusionInputting a constructed and trained behavior prediction model, obtaining a prediction result of the user's behavior of dropping learning, and sending reminding information to the user who has the behavior of dropping learning.
Specifically, the method comprises the following steps:
step 5.1, fusing the feature into a vector FfusionInputting the data into a constructed and trained behavior prediction model to obtain a classification result PoutputIf P isoutputIf 1 indicates that the user is predicted as a conjugate, if Poutput0 indicates that the user is predicted as not likely to be learned.
In this embodiment, the behavior prediction model includes three fully connected layers with dimensions of 1 × 128, 1 × 32, and 1 × 16, respectively.
It should be noted that the activation functions used in this embodiment are all relus, which avoids the problems of gradient disappearance and gradient explosion of the prediction model.
The application directly learns the deep sequence characteristics of the user through the forward LSTM module and the backward LSTM module, and provides direct support for the prediction of the conjugate. For the difference and diversity of users in online education, the application generates specific attention characteristics according to the personal information of the users and the selected course information, and the prediction accuracy of the application is improved.
After the user's behavior of breaking out is predicted, the user can be contacted by telephone, mail, etc. to provide various welfare or incentive policies to reduce the user's behavior of breaking out and to prompt the user to return to online learning.
The above-mentioned embodiments only express several implementation methods of the present application, and the description thereof is more specific and detailed, but not construed as limiting the scope of the invention. It should be noted that, for a person skilled in the art, several variations and modifications can be made without departing from the concept of the present application, which falls within the scope of protection of the present application. Therefore, the protection scope of the present patent application shall be subject to the appended claims.

Claims (5)

1. A method of preventing a dropped behavior in online education, the method comprising:
obtaining an activity time sequence matrix R of the user from an online education platform, inputting the activity time sequence matrix R to a time sequence feature extraction module which is constructed and trained to obtain deep sequence features F of the userR；
Obtaining context information N of a user from an online education platform, and combining the context information N and deep sequence characteristics FRConnecting, inputting to the constructed and trained attention feature generation module to obtain the attention feature F of the userN；
Obtaining the class selection characteristic similarity relation G of the user from the online education platform, and selecting according to the userThe class characteristic similarity relation G forms a classmate relation graph GclassmateA graph G of the relationship between studentsclassmateInputting the data to a classmate feature extraction module which is constructed and trained to obtain the classmate feature F of the userG；
Will be characteristic FR、FNAnd FGDepth feature fusion is carried out to obtain the dropping behavior feature F of the userfusion；
Characterizing a dropped behavior of the user FfusionInputting a constructed and trained behavior prediction model to obtain a prediction result of a user behavior of breaking down, and sending reminding information to the user with the behavior of breaking down;
wherein the obtaining of the activity time series matrix R of the user from the online education platform comprises:
constructing an activity time sequence matrix R of the user through the records of n activities of the user within m days after the beginning of the course:
Figure FDA0003553497770000011
wherein, when i is less than or equal to m and j is less than or equal to n, ri，jRepresenting the number of occurrences of activity j of the user on the ith day after the start of the course; when i is not more than m, j is n +1, ri，n+1Representing the total number of times that the user's activity occurred during the ith day after the start of the course; when i is not more than m, j is n +2, ri，n+2Representing the total activity duration of the user on the ith day after the course is started; when i is m +1, j is less than or equal to n, rm+1，jRepresenting the total number of times of the ith activity of the user within m days after the beginning of the course; when i is m +1 and j is n +1, rm+1，n+1Representing the total number of times of all activities of the user within m days after the beginning of the course; when i is m +1 and j is n +2, rm+1，n+2Representing the total duration of the user's activity within m days after the start of the lesson.
2. The method of preventing dropped learning behavior in online education of claim 1 wherein the obtaining contextual information N for the user from an online education platform,the context information N and the deep sequence characteristics FRConnecting, inputting to the constructed and trained attention feature generation module to obtain the attention feature F of the userNThe method comprises the following steps:
step 2.1, constructing a behavior matrix R' of the user according to the activity record of the user:
R′＝[r′1……r′i…r′n]
wherein n represents a total of n activities, r'iRepresenting the total number of times of the ith activity of the user;
step 2.2, classifying the users by using a clustering algorithm through the behavior matrix R' of the users to obtain user types, and combining the user types with the ages, the sexes, the academic histories and the types of courses and the learning missing rate of the users to form context information N of the users:
N＝[n1 n2 n3 n4 n5 n6]
wherein n is1To n6Sequentially representing the user type, the user age, the user gender, the user academic history, the course type and the course dropping rate;
step 2.3, the context information N and the deep sequence characteristics F of the userRConnecting to obtain deep sequence characteristics fused with context information;
step 2.4, inputting the deep sequence features fused with the context information into an attention feature generation module to obtain an attention feature FN。
3. The method of claim 1, wherein said obtaining a class selection feature similarity G for a user from an online education platform is performed by inputting the class selection feature similarity G to a class feature extraction module configured and trained to obtain a class feature F for the userGThe method comprises the following steps:
step 3.1, according to the Course selection records SR (Uset, Course) of the users, regarding each user as a user node, regarding each Course as a Course node, and forming a Course selection record bipartite graph BG(User，Course, Edge), where User represents the set of all User nodes: useriBelongs to the User; course represents the set of all Course nodes: coursejE.g. Course; edge represents the set of edges connecting two nodes: edgek＝(Useri，Coursej) E, the Edge is equal to SR, which indicates that the user node i selects and repairs the course node j;
step 3.2, recording the course selection into bipartite graph BGInputting the data into an unsupervised graph neural network to obtain the characteristics F of all user nodesu：
Figure FDA0003553497770000031
Where t represents the total number of users, FuEach row u iniThe course selection characteristics of the user i are represented;
step 3.3 course selection characteristics u for different usersi、ujAnd performing similarity calculation, determining that the course selection characteristics are similar when the defined similarity is greater than a certain value, and obtaining a course selection characteristic similarity relation G of the user:
Figure FDA0003553497770000032
wherein, each line in G represents a pair of similar relations of course selection characteristics, (User)i，Userj) Representing that the user i and the user j are similar to each other in course selection characteristics;
step 3.4, forming a classmate relation graph G according to the class selection characteristic similarity relation G of the userclassmate，GclassmateEach node in the system represents a Course selection record SR (User, Course), the edge connecting the two nodes represents that the users represented by the two nodes are in the Course selection characteristic similarity relation, and the selected courses are consistent;
step 3.5, drawing the relationship graph G of the classmatesclassmateInputting the data into a classmate feature extraction module to generate a classmate feature FG。
4. The method of preventing dropped learning behavior in online education of claim 1 wherein feature F is usedR、FNAnd FGDepth feature fusion is carried out to obtain the dropping behavior feature F of the userfusionThe method comprises the following steps:
step 4.1, attention feature FNSoftmax activation was performed according to the following formula, obtaining the attention weight W for each day after the start of the courseattention＝(w1，w2，...，wk)：
Figure FDA0003553497770000033
Wherein, wiIndicating the attention weight on day i after the start of the course, FN，iExpress attention feature FNThe ith value of (a), k represents the total number of days recorded in the raw data;
step 4.2, the attention weight W obtained in the last stepattentionThe following formula is used to determine the deep sequence characteristics FRMultiplying to obtain the deep sequence features combined with attention
Figure FDA0003553497770000041
Figure FDA0003553497770000042
Wherein k represents FRNumber of lines of (F)R，iIs represented by FRThe value of row i;
step 4.3, deep sequence features to be combined with attention
Figure FDA0003553497770000043
Characteristic of classmate FGAdding and splicing attention feature FNObtaining a feature fusion vector Ffusion：
Figure FDA0003553497770000044
5. A method of preventing dropped learning behavior in online education as recited in claim 1, where said characterizing dropped learning behavior of said user as FfusionInputting the user behavior prediction model which is built and trained to obtain a prediction result of the user behavior of the science, wherein the prediction result comprises the following steps:
fusing the features into a vector FfusionInputting the data into a constructed and trained behavior prediction model to obtain a classification result PoutputIf P isoutputIf 1 indicates that the user is predicted as a conjugate, if Poutput0 indicates that the user is predicted as not likely to be learned.
CN202110022690.0A 2021-01-08 2021-01-08 Method for preventing breaking behavior in online education Active CN112734105B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110022690.0A CN112734105B (en) 2021-01-08 2021-01-08 Method for preventing breaking behavior in online education

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110022690.0A CN112734105B (en) 2021-01-08 2021-01-08 Method for preventing breaking behavior in online education

Publications (2)

Publication Number Publication Date
CN112734105A CN112734105A (en) 2021-04-30
CN112734105B true CN112734105B (en) 2022-06-28

Family

ID=75591266

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110022690.0A Active CN112734105B (en) 2021-01-08 2021-01-08 Method for preventing breaking behavior in online education

Country Status (1)

Country Link
CN (1) CN112734105B (en)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114358135B (en) * 2021-12-10 2024-02-09 西北大学 MOOC (motion on-the-fly) conjugate prediction method realized by utilizing data enhancement and feature weighting
CN114880525A (en) * 2022-05-25 2022-08-09 浙江工业大学 Online course breakout prediction method based on incidence relation and video interaction
CN118569451A (en) * 2024-08-01 2024-08-30 北京长亭科技有限公司 Discontinue one's studies prediction method and device based on MOOC comprehensive feature fusion

Family Cites Families (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20110047224A1 (en) * 2009-08-21 2011-02-24 Seymour Ian Fox Apparatus and method for processing and/or for providing education information and/or education related information
CN105930875A (en) * 2016-05-05 2016-09-07 清华大学 User course dropping behavior prediction method and apparatus
US11551570B2 (en) * 2018-02-15 2023-01-10 Smarthink Srl Systems and methods for assessing and improving student competencies
CN109272164B (en) * 2018-09-29 2021-09-28 清华大学深圳研究生院 Learning behavior dynamic prediction method, device, equipment and storage medium
KR102134339B1 (en) * 2018-10-11 2020-07-15 연세대학교 산학협력단 Method and Apparatus for Detecting Fault of Multi-Core in Multi-Layer Perceptron Structure with Dropout
CN109558983A (en) * 2018-12-03 2019-04-02 华中师范大学 Network courses dropping rate prediction technique and device
CN110059716B (en) * 2019-03-12 2023-06-02 西北大学 Construction of CNN-LSTM-SVM network model and MOOC conjugate prediction method
CN110866162B (en) * 2019-10-10 2021-11-19 西安交通大学 Causal relationship mining method based on conjugate behaviors in MOOC data

Also Published As

Publication number Publication date
CN112734105A (en) 2021-04-30

Similar Documents

Publication Publication Date Title
CN112734105B (en) Method for preventing breaking behavior in online education
US10115167B2 (en) System and method for identifying key targets in a social network by heuristically approximating influence
Freyne et al. Increasing engagement through early recommender intervention
CN112199608A (en) Social media rumor detection method based on network information propagation graph modeling
CN111428127B (en) Personalized event recommendation method and system integrating theme matching and bidirectional preference
CN111737427B (en) Method for recommending lesson forum posts by combining forum interaction behaviors and user reading preference
Furini et al. TSentiment: On gamifying Twitter sentiment analysis
Macdonald et al. Simulating impacts of rapid forest loss on population size, connectivity and genetic diversity of Sunda clouded leopards (Neofelis diardi) in Borneo
CN110473123A (en) A kind of multi-element intelligent educational method and system
Schumann et al. Group fairness in bandit arm selection
CN109934681A (en) The recommended method of user's commodity interested
Reilly et al. The role of place meanings in opposition to water-related infrastructure projects: the case of the Mactaquac Dam, New Brunswick, Canada.
Li et al. Knowledge Graph‐Enhanced Intelligent Tutoring System Based on Exercise Representativeness and Informativeness
Wilson et al. Costly information and the evolution of self-organization in a small, complex economy
Yitshaki et al. Becoming a social entrepreneur: understanding motivations using life story analysis
Liu et al. Toward personalized activity level prediction in community question answering websites
Ho et al. Multi-objective parallel test-sheet composition using enhanced particle swarm optimization
CN104516873A (en) Method and device for building emotion model
Losee Information and Knowledge: Combining Justification, Truth, and Belief.
CN106096653A (en) Ascribed characteristics of population estimating method based on cross-platform user social contact multimedia behavior
CN115564013B (en) Method for improving learning representation capability of network representation, model training method and system
Jiang et al. Learning analytics in a blended computer education course
CN114861766A (en) Dynamic link prediction method and system based on multi-granularity evolution
Li et al. An effective deep learning approach for personalized advertisement service recommend
CN113268583A (en) Knowledge recommendation method based on community content

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant