CN-119397088-B - Accurate pushing system of examination questions based on learning behavior analysis

CN119397088BCN 119397088 BCN119397088 BCN 119397088BCN-119397088-B

Abstract

The invention discloses a test question accurate pushing system based on learning behavior analysis, in particular to the technical field of test question accurate recommendation, which comprises a learning ability dynamic evaluation module, a multitask model refinement module, a speculation completion module and an adaptive recommendation module, wherein the learning ability dynamic evaluation module constructs a time sequence model according to learning behavior data of a user to predict future learning ability of the user for primary classification, the multitask model refinement module refines the result of the primary classification, the self-adaptive recommendation module performs personalized recommendation on different types of users according to the output result of the prediction completion module by generating learning behavior characteristics based on the graph neural network and combining known learning behavior data, evaluates the learning ability of the users by using a time sequence model, refines and classifies the users by combining a multi-task learning model, and predicts potential learning behaviors through the graph neural network to realize personalized test question recommendation, thereby effectively solving the problem of data sparsity and improving recommendation accuracy.

Inventors

GU XIAOQING
HUANG WENHUI
SHI HAIFENG
XU YUTING

Assignees

江苏优利信科技有限公司

Dates

Publication Date: 20260508
Application Date: 20240914

Claims (6)

1. The test question accurate pushing system based on learning behavior analysis is characterized by comprising a learning ability dynamic evaluation module, a multitask model refinement module, a speculation completion module and a self-adaptive recommendation module; The learning ability dynamic evaluation module is used for constructing a time sequence model according to the learning behavior data of the user so as to predict the future learning ability of the user, and classifying the learning state of the user once according to the future learning ability of the user and the current learning score of the user; The multi-task model refinement module is used for refining the results of primary classification, labeling different topics and then establishing a multi-task learning model, wherein the multi-task learning model comprises a first output layer for predicting the topics required by a user, a second output layer for classifying the topics, and a third output layer for evaluating the topic difficulty, and the target users with extremely strong learning ability and extremely weak learning ability are marked according to the output results of the multi-task sharing bottom network; The presumption completion module is used for presuming the potential learning behavior of the target user through a method based on a graph neural network, and generating learning behavior characteristics by combining known learning behavior data; The self-adaptive recommendation module is used for conducting personalized recommendation on different types of users according to the output result of the presumption completion module; the prediction result of the multi-task learning model on the learning ability of the user is classified and refined on the basis of one classification as follows: Definition of the display Capacity function The underlying network output based on the multitasking learning model is expressed as In which, in the process, For the user characteristics extracted from the shared network, Setting the display capability classification threshold as the classification function Output type Wherein , wherein, Indicating that the learning ability of the user is extremely strong, Indicating that the learning ability of the user is extremely weak, for the class A target user, when the output type is When the output type is the dominant weak user When the user is marked as a hidden floating user; for class B target users, when the output type is Marking the user as a user with extremely strong display capability, when the output type is When the user is marked as a hidden floating user; The learning ability of the class A target user is lower, the learning state is continuously poor, and the learning ability of the class B target user is stronger, and the learning state is continuously good; Logic for generating learning behavior features based on a graph neural network is as follows: The nodes of the graph comprise user nodes, question nodes and knowledge point nodes, wherein the user nodes are used for representing individual users, the question nodes are used for representing questions, the knowledge point nodes are used for representing knowledge points related to the questions, the edges of the graph comprise user-question edges and question-knowledge point edges, the user-question edges are used for representing the relation of the user doing the questions, the weights of the user-question edges are assigned according to the accuracy and the answering time of the answers of the user, the question-knowledge point edges are used for representing the knowledge points related to the questions, the weights of the question-knowledge point edges are assigned according to the question difficulty and the importance of the knowledge points, and the graph is represented as follows , wherein, For the node set of the graph, the node set includes users, topics and knowledge points, For the edge set of the graph, the edge set comprises the association between a user and a question and the association between the question and a knowledge point, information transmission is carried out through connection between nodes, the characteristic representation of each node is updated, and node characteristics are generated by using known learning behavior data; The logic for recommending test questions according to the output of the graph neural network is as follows: Integrating the relation between learning behavior data and topics and knowledge points of a user through convolution of a graph neural network, and training the graph neural network to obtain the characteristic representation of the user as The expression is In which, in the process, For the feature input by the user, A neural network of the graph is shown, For the user characteristic representation after the graph is rolled, for the user with extremely strong display capability, predicting the future learning capability representation of the user by adopting a regression model and recommending test questions, and for the user with extremely weak display capability, predicting the learned knowledge of the user by adopting a classification model and recommending the test questions.
2. The accurate pushing system for test questions based on learning behavior analysis according to claim 1, wherein feature vectors representing the change of learning ability of users are constructed based on time series extracted features, and the construction method of the feature vectors is as follows: The answer number in the calibrated time window t is The answer accuracy rate in the time window t is , The calculated expression of (2) is In which, in the process, The average answer time is that the correct answer number in the time window t is The expression is calculated as , wherein, For each question answering time, the wrong question review frequency is The expression is calculated as Learning progress rate In which, in the process, And All are extremely small positive numbers, and the number of the positive numbers is very small, For the answer accuracy of the previous time window, the learning stability is that The expression is calculated as Wherein n is the data amount of the learning result, For the i-th learning result, a feature vector representing a change in learning ability of the user Is in the form of 。
3. The test question accurate pushing system based on learning behavior analysis according to claim 2, wherein feature vectors are used for pushing Outputting prediction of future learning ability of the user after inputting the time sequence model, and classifying the user once by using a logistic regression model according to the prediction of future learning ability of different users, wherein the specific method comprises the following steps: establishing a logistic regression model In which, in the process, The classification bias value of the user is y is a binary dependent variable, represents the judgment category of the learning ability of the user, x is an independent variable vector, comprises output prediction of a time sequence model, the learning performance of the user and various characteristics, The m is the number of independent variable vectors and is a positive integer; Setting a first threshold value of primary classification And a first classification second threshold And classify the first threshold value once Less than a first classification second threshold When calculating the classification deviation value of the obtained user Less than or equal to the primary classification first threshold When the classification deviation value of the user is calculated, the user is marked as the class A target user Greater than or equal to the primary classification second threshold When the user is marked as the class B target user, the classification deviation value of the obtained user is calculated More than one classification first threshold And less than a first classification second threshold When the user is marked as a non-target user.
4. The accurate pushing system of test questions based on learning behavior analysis of claim 1, wherein the logic for establishing the multi-task learning model is: Setting the multi-task learning model as Where Z is the shared feature vector, Is a feature of the user and is, Is a characteristic of the title and is characterized by, Is a bottom shared network, a first output layer of the multi-task learning model is used for predicting the problem of user requirements, the loss function is a cross entropy function, and the expression is In which, in the process, For the purpose of classifying the objects, For the predicted class probability, B is the number of topics, the second output layer of the multi-task learning model is used for classifying the topics according to the topic labels, judging the topic types of the topics, the loss function is a cross entropy function, and the expression is In which, in the process, Is a true label for the question type, For the predicted question probability, V is the number of questions, the third output layer of the multi-task learning model is used for predicting the question difficulty, outputting the question difficulty value, the loss function is mean square error loss, and the expression is In which, in the process, As the true difficulty value of the subject, For the predicted difficulty value, the overall loss function is 、 And Is a weighted sum of (c).
5. The accurate pushing system for test questions based on learning behavior analysis according to claim 1, wherein the method for predicting future learning ability performance of a user and recommending test questions by using a regression model is as follows: Defining a regression model as , In order to predict the users with extremely strong dominant ability, the difficulty of recommending questions is dynamically adjusted according to the learning ability performance of the users predicted by the regression model, and the adjustment logic is as follows In which, in the process, In order to recommend the difficulty of the test questions, The question difficulty function is adjusted according to the predicted learning performance of the user.
6. The accurate pushing system for test questions based on learning behavior analysis according to claim 5, wherein the method for predicting the learned knowledge of the user and recommending the test questions by using the classification model comprises the following steps: defining a classification model as In which, in the process, 、 For the weight matrix of the classification model, To activate the function, the output is The function obtains the learned category distribution of the user and the characteristics of the user Characteristics of the title Matching is carried out, matching similarity is calculated to recommend test questions, and the matching similarity is calculated by the following method The higher the matching similarity, the stronger the priority of the recommendation.

Description

Accurate pushing system of examination questions based on learning behavior analysis Technical Field The invention relates to the technical field of the Internet of things, in particular to a test question accurate pushing system based on learning behavior analysis. Background Potential objects accurately pushed by test questions are generally arranged in normal distribution, students with extremely strong learning ability and extremely weak learning ability only occupy very few students, the requirement of very few users is that the problem of data sparsity exists, and precise personalized pushing of the cutting is difficult to carry out; Aiming at the problem of data sparsity, the current common means is to adopt a cold start scheme to carry out heuristic processing based on personal experience and recommended rules of an industry expert, the method has strong subjectivity, meanwhile, has high requirements on rule breadth, is still difficult to adapt to a small number of users in extreme types, the other common method is to transfer the data extraction existing in other fields to the target field usually based on transfer learning, but differences exist between a source field and the target field, and the smooth processing of the differences and the grasping of learning data behavior logic are difficult challenges in practical application. In order to solve the above-mentioned defect, a technical scheme is proposed. Disclosure of Invention The invention aims to provide a test question accurate pushing system based on learning behavior analysis so as to solve the defects in the background technology. In order to achieve the aim, the invention provides the technical scheme that the test question accurate pushing system based on learning behavior analysis comprises a learning ability dynamic evaluation module, a multi-task model refinement module, a speculation completion module and an adaptive recommendation module; The learning ability dynamic evaluation module is used for constructing a time sequence model according to the learning behavior data of the user so as to predict the future learning ability of the user, and classifying the learning state of the user once according to the future learning ability of the user and the current learning score of the user; The multi-task model refinement module is used for refining the results of primary classification, labeling different topics and then establishing a multi-task learning model, wherein the multi-task learning model comprises a first output layer for predicting the topics required by a user, a second output layer for classifying the topics, and a third output layer for evaluating the topic difficulty, and the target users with extremely strong learning ability and extremely weak learning ability are marked according to the output results of the multi-task sharing bottom network; The presumption completion module is used for presuming the potential learning behavior of the target user through a method based on a graph neural network, and generating learning behavior characteristics by combining known learning behavior data; the self-adaptive recommendation module is used for conducting personalized recommendation on different types of users according to the output result of the presumption completion module. Preferably, a feature vector representing the variation of learning ability of a user is constructed based on the time series extracted features, and the construction method of the feature vector is as follows: Calibrating the answer number in the time window t to be N t, and calibrating the answer accuracy rate in the time window t to be A t,At to be the calculation expression Wherein C t is the correct answer number in the time window T, the average answer time is T t, and the calculation expression isWherein t i is the answering time of each question, the wrong question review frequency is F t, and the calculation expression isRate of progress of learningWherein, delta and epsilon are extremely small positive numbers, A t-1 is the answer accuracy of the previous time window, the learning stability is S t, and the calculation expression isWhere n is the data amount of the learning result, x i is the i-th learning result, and the feature vector v t representing the change in learning ability of the user is in the form of v t＝[Nt,At,Tt,Ft,Pt,St. Preferably, the feature vector v t is input into a time sequence model and then output to predict the future learning ability of the user, and the user is classified once by using a logistic regression model according to the prediction of the future learning ability of different users, and the specific method is as follows: establishing a logistic regression model Wherein P (y= 1|X) is a classification bias value of the user, y is a binary dependent variable, and represents a judgment type of learning ability of the user, x is an independent variable vector, including output prediction of a time series mode