Personalized online test question recommendation method based on student emotion perception

Abstract

Disclosed in the present invention is a personalized online test question recommendation method based on student emotion perception. The method comprises the following steps: S1, collecting answering behaviors of students for online learning and corresponding emotion records, and carrying out data cleaning to extract effective information; S2, on the basis of the emotion data of the students, using PCA to classify the complex emotions of the students, adjusting the impacts of the emotions on the students by means of the weight of a self-attention module, and performing modeling on the external cognition of the students; S3, integrating the external cognition of the students and the inherent cognition of the students, introducing an NCB-IRM framework, and comprehensively predicting the personalized response of the students; S4, using PMF to decompose the answering generality of the students, and jointly predicting the response of the students to test questions on the basis of the predicted personalized response of the students; and S5, on the basis of the predicted response of the students, providing test questions with appropriate difficulty for the students, and forming a final recommendation list. In the present invention, on the basis of PMF and NCB-IRM in combination with the personality and answering generality of students, the accuracy and reasonability of test question recommendation are realized in accordance with the test question mastering condition of the students.

Description

Personalized online test question recommendation method based on students' emotion perception Technical Field

[0001] The present invention relates to the technical field of online learning big data mining, and in particular to a personalized online test question recommendation method based on student emotion perception. Background Art

[0002] The development of online learning platforms has provided students with a convenient way to practice exams. Students can access a rich library of question banks through online learning platforms and learn anytime, anywhere. Online learning platforms offer a broad, convenient, and intelligent learning method for both teachers and students. In recent years, numerous online learning platforms have emerged, such as ASSISTment, Coursera, and EdX, offering students opportunities to practice exams. However, due to the sheer volume of available exam resources, students often struggle to effectively find the right questions, which impacts their learning interest and efficiency. Therefore, it is crucial to leverage educational big data mining technology to accurately recommend appropriate exam resources to students based on their individual characteristics.

[0003] Recommender systems in education can provide students with a variety of learning and test resources, such as video courses and practice tests. Recommender systems are a key branch of information retrieval within educational big data mining. Their goal is to provide students with learning materials that meet their needs and assist them in practicing test questions. Traditional online resource recommendation methods are primarily based on collaborative filtering algorithms, which construct a user-resource matrix and then recommend learning resources based on user or resource similarities. However, in online learning scenarios, these methods have several drawbacks: they fail to fully consider students' individual characteristics; the recommended content lacks diversity and rationality, resulting in the test resources provided by the recommendation system failing to meet students' individual needs and preventing targeted training.

[0004] To uncover students' test practice patterns, most existing online test resource recommendation solutions utilize artificial intelligence algorithms such as neural networks, deep learning, and reinforcement learning. However, these methods often overemphasize the accuracy of the recommendation model and overlook the individual characteristics exhibited by students during practice. They fail to tailor recommendation solutions to students' individual learning characteristics, resulting in a lack of targeted and informative test questions. Cognitive diagnostics analyze students' test logs to assess their knowledge and measure their learning status, providing a theoretical basis for personalized recommendation solutions for different students. Using cognitive diagnostic methods (CDM) for test question recommendation can not only diagnose students' knowledge mastery but also infer test question characteristics such as difficulty and discrimination, significantly accounting for individual student characteristics. Because the effectiveness of CDM diagnostics directly impacts the accuracy of test question recommendations, inaccurate CDM diagnostics can lead to reduced recommendation performance.

[0005] In cognitive diagnosis, increasing the dimensionality of data input can effectively improve its effectiveness. Furthermore, leveraging multi-source information to enhance the semantic fusion of data in recommendation systems can also enrich the representation of users and items. Therefore, by adding the emotional dimension, emotion-aware test question recommendations can both improve diagnostic accuracy and leverage more personalized student knowledge. However, CDM fails to consider the commonalities among students' answers, resulting in a narrow coverage of recommended test questions. This leads to sparse data in the subsequent recommendation process, resulting in limitations in the recommended test questions.

[0006] Summary of the Invention

[0007] Purpose of the invention: In response to the above problems, the purpose of the present invention is to provide a personalized online test question recommendation method based on student emotional perception, introduce student emotional information, design an online test question recommendation scheme based on the individuality and commonality of students' test questions, recommend test questions suitable for students, and improve students' practice effects.

[0008] In order to solve at least one of the above technical problems, according to one aspect of the present invention, a method for recommending personalized online test questions based on student emotion perception is provided, comprising the following steps:

[0009] S1. Collect students' online learning answering behaviors and corresponding emotional records, perform data cleaning, and extract effective information;

[0010] S2. Based on students’ emotional data, PCA is used to classify students’ complex emotions, and the influence of emotions on students is adjusted through the weight of the self-attention module to model students’ external cognition;

[0011] S3. Integrate students’ external cognition and inherent cognition, introduce the NCB-IRM framework, and comprehensively predict students’ personalized responses;

[0012] S4. Use PMF to decompose the common characteristics of students' answers and combine it with students' personalized predicted responses to jointly predict students' responses to test questions;

[0013] S5. Based on the predicted student responses, provide students with test questions of appropriate difficulty and form a final recommendation list.

[0014] 2. The method according to claim 1, characterized in that step S1 collects the online learning answering behavior records including the student's emotions and extracts effective information; specifically:

[0015] S11, collect student ID, test question ID, knowledge point ID, student's emotional data E on the test question, and student's response R on the test question;

[0016] W includes six emotional dimensions: bored, concentrating, confused, frustrated, offtask, and gaming; the emotional element A in each dimension ij It belongs to [0,1]. The closer it is to 1, the stronger the student's performance in this emotion is.

[0017] Student i's response to question j is expressed as r ij Indicates; if r ij =1 means student i answered question j correctly; if r ij =0, it means that student i answered question j incorrectly;

[0018] S12. Clean the data, delete empty entries and redundant answer records, and extract valid information as follows:

[0019] Extract the potential knowledge mastery of student i at knowledge point k from student behavior, and use α ik Indicates that α ik ∈[0,1];

[0020] The test matrix Q contains the relationship between all test questions and knowledge points, and the element q jk Indicates the situation of each test question q testing knowledge points; if q jk =1, indicating that question j does not test knowledge point k; if q jk =0, indicating that question j does not test knowledge point k.

[0021] 3. The method according to claim 1, characterized in that step S2 models students' external cognition based on students' emotions, students' answering behaviors and the test matrix through a PCA algorithm module and a hierarchical self-attention module, specifically:

[0022] S21. Using the PCA algorithm, the students’ complex emotions are clustered into two categories, namely positive emotions and negative emotions, expressed as:

[0023] Where E is a matrix containing six emotion dimensions, represents the dot product operation of the vector, It is the projection value of the emotional sample in the direction of the first two principal components, V1 and V2 represent the eigenvectors respectively;

[0024] S22, using two self-attention modules in the interaction layer to simulate the clustered emotions. For example, the potential mastery of students' knowledge points α ik And the knowledge points tested in the test questions jk Interaction;

[0025] Input positive emotions When the self-attention module takes the sequence p of knowledge points tested in the test j As the key in the attention mechanism, the sentiment representation Sequences of students' potential mastery of knowledge points i As the query and value in the attention mechanism respectively;

[0026] First, the student's answer vector p is calculated by cosine similarity j With positive emotion vector The mutual influence weight of interaction ω p ; Then, after considering the feature interaction, by ω p and students’ potential mastery of knowledge points i The weighted sum of the positive emotion vector o after interaction is obtained p , the expression is:

[0027] where p j =(q j1 ,q j2 ,…q jk ), s i =(α i1 ,α i2 ,…α ik );

[0028] Similarly, inputting negative sentiment When , the interaction layer simulates the interaction between students’ negative emotions, students’ knowledge mastery and test answering through self-attention, and obtains the negative emotion vector o after interaction. n ;

[0029] S23, the aggregation layer uses another self-attention module to calculate the different interactive emotions o p 、o n polymerization;

[0030] The interactive emotion results of different categories o p and o n As the query and key in the attention mechanism respectively, the student's knowledge mastery is used as the value in the attention mechanism, and different interactive emotions are aggregated to obtain the student's external cognition θ that integrates emotional features. o , the expression is:

[0031] 4. The method according to claim 1, characterized in that step S3 integrates external cognition and students' inherent cognition, introduces the NCB-IRM framework, and comprehensively predicts students' personalized responses; given that students' responses r ij , considering two extreme cases where students’ emotional characteristics affect their responses; specifically:

[0032] S31. Influenced only by the emotional characteristics, student i correctly answers question j. The student learning response function is modeled as:

[0033] Among them, θ o , α j 、b j They are students’ external cognitive level, test item discrimination and test item difficulty;

[0034] S32. Unaffected by the emotional characteristics, student i correctly answered question j. The student learning response function is modeled as:

[0035] Among them, r ij A represents the response of student i to question j, ij represents the emotion expressed by student i when answering question j, θ i 、a j d j They are students’ intrinsic cognitive level, test item discrimination and test item difficulty;

[0036] S33. Assuming that the response to each question is statistically independent of the student's cognition, use the Bernoulli distribution to model the responses of all students:

[0037] where η ij and ζ ij represent the probability that student i answers question j correctly based on the practice and emotional characteristics of the test questions, respectively;

[0038] S34, the student's response matrix is ​​the comprehensive response to the test question prior and the student prior; given the response r ij , students' cognitive θ i ,θ o , test variable a j , α j、b j d j , the item response function of the student response is as follows:

[0039] S35, the item response function that maximizes the student response is obtained by a Markov chain Monte Carlo method based on Metropolis-Hastings. i 、a j , α j 、b j d j The optimal solution of

[0040] S36. Predicting Personalized Student Responses

[0041] 5. The method according to claim 1, characterized in that step S4 is based on the actual responses of students, uses PMF to decompose the common characteristics of students' answers, combines students' personalized predicted responses, and jointly predicts students' responses to the test questions; specifically:

[0042] S41. Perform probability matrix decomposition on the student's true response through PMF to obtain the characteristic matrix U of the student and the test question. i 、V j ;

[0043] S42. Predict responses that include common characteristics among students

[0044] S43. Calculate the weighted sum of the student's personalized prediction response and the student's common prediction response, as follows:

[0045] Among them, R p is the predicted student response, μ is the mean score of all students, r′ ij is the predicted student personalized response, R′ is the student common response predicted by PMF, and the ratio between student individuality and commonality is adjusted by parameter ρ.

[0046] 6. The method for recommending personalized online test questions based on student emotion perception according to claim 1, wherein step S5 provides test questions of appropriate difficulty to the student based on the predicted student response to form a final recommendation list; specifically, the method comprises:

[0047] S51, based on the predicted student response R p Design test difficulty [d1, d2];

[0048] The difficulty of the test item is calculated as the mean of the students' responses:

[0049] Among them, r ijis the response of student i to test item j, and n is the number of records of answering test item j;

[0050] S52, according to the difficulty of the test questions d j , filter out the list of recommended test questions R p ∈[1-d2,1-d1].

[0051] The present invention uses PMF to explore the common characteristics of students' answers and combines them with students' personalized prediction responses, which not only improves the prediction ability of the model but also ensures the comprehensiveness of the recommendation plan.

[0052] The present invention designs the difficulty of test questions according to the students' mastery of the test questions after fully considering the students' individual characteristics and the common characteristics of their answers, thereby improving the rationality of the recommended test questions.

[0053] According to another aspect of the present invention, a computer-readable storage medium is provided, on which a computer program is stored. When the program is executed by a processor, the steps of the student emotion-aware personalized online test question recommendation method of the present invention are implemented.

[0054] According to another aspect of the present invention, a computer device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps of the student emotion-aware personalized online test question recommendation method of the present invention are implemented.

[0055] Compared with the prior art, the present invention has at least the following beneficial effects:

[0056] 1. Based on PMF, this paper introduces NCB-IRM and proposes a personalized online test question recommendation method based on student emotional perception. First, NCB-IRM combines students' external cognition of emotional perception with their internal cognition to more comprehensively consider students' individual needs. Second, PMF is used to explore the common characteristics of students' answers, ensuring the comprehensiveness of the recommendation plan and improving the rationality of the recommended test questions.

[0057] 2. In NCB-IRM, an external cognitive modeling method based on students' emotional perception is designed. On the one hand, PCA is used to reduce the complexity of sentiment analysis. On the other hand, a hierarchical self-attention method is used to adjust the impact on students according to different sentiment weights. Based on students' emotions, it provides an effective method for exploring students' personalized needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] FIG1 is a flow chart of the method of the present invention;

[0059] Figure 2 is a schematic diagram of the external cognitive modeling structure of students’ emotional perception;

[0060] FIG3 shows the performance of the accuracy of students' answers to questions of different difficulty levels according to the present invention;

[0061] Figure 4 shows examples of two students’ cognitive results and recommended test questions;

[0062] 4(a) shows the enhanced cognition of diagnosed students A and B, and 4(b) shows the test questions recommended to students A and B and the relationship between the test questions and knowledge points. DETAILED DESCRIPTION

[0063] The dataset used in this example comes from the online learning platform ASSISTment. Data from 2009 and 2017 were collected from the platform. ASSIST2009 contains 401,756 student practice records, while ASSIST2017, which incorporates student emotion data, contains 942,816 student practice records. Each student's response and emotion data are recorded as they answer the questions.

[0064] As shown in FIG1 , the method for recommending personalized online test questions based on student emotion perception described in this embodiment includes the following steps:

[0065] S1. Collect students’ learning behavior records when watching educational videos, clean the learning behavior records, and extract effective information.

[0066] Student ID, test question ID, knowledge point ID, student's emotion E on the test question, student's response R on the test question;

[0067] E includes six emotional dimensions: bored, concentrating, confused, frustrated, off-task, and gaming. Each dimension belongs to [0,1]. The closer the value of a certain dimension is to 1, the stronger the student's emotion is.

[0068] Student i's response to question j is expressed as r ij Indicates; if r ij =1 means student i answered question j correctly; if r ij =0, it means that student i answered question j incorrectly;

[0069] The data was cleaned, empty entries and redundant answer records were deleted, and the valid information was extracted as follows:

[0070] Student i’s potential knowledge mastery of knowledge point k is expressed as α ik Indicates that α is randomly initialized ik , α ik ∈[0,1];

[0071] The test matrix Q contains the relationship between all test questions and knowledge points. Test question q examines the situation of knowledge point k, and q jk Indicates; if qjk =1, indicating that question j does not test knowledge point k; if q jk =0, indicating that question j does not test knowledge point k.

[0072] Data cleaning refers to deleting empty entries and redundant answer records. In the two data sets of this embodiment, the cleaned data sets are divided into training set, test set, and validation set in a ratio of 7:2:1. The statistical information is shown in Table 1.

[0073] Table 1 Statistical description of the dataset

[0074] S2. Based on students’ emotions, answering behaviors and test matrix, the PCA algorithm module and the hierarchical self-attention module are used to model students’ external cognition.

[0075] S21 uses the PCA algorithm to cluster students’ complex emotions into two categories, namely positive emotions and negative emotions, expressed as:

[0076] Where E = E bored ,E concentrate ,E confused ,E frustrated ,E offtast ,E gaming , represents the dot product operation of the vector, n_component after dimensionality reduction = 2, V1, V2 are the eigenvectors, It is the projection value of the sentiment sample in the direction of the first two principal components.

[0077] S22 uses two self-attention modules in the interaction layer to simulate the clustered emotions (with For example) and students' potential mastery of knowledge points α ik And the knowledge points tested in the test questions jk Interaction;

[0078] Input positive emotions When the self-attention module takes the test question knowledge point sequence p j As the key in the attention mechanism, the sentiment representation Sequences of students' potential mastery of knowledge points i They serve as the query and value in the attention mechanism respectively;

[0079] First, the student's answer vector p is calculated by cosine similarity j With positive emotion vector The mutual influence weight of interaction ω p, assign weights to students’ positive emotions based on the characteristics of the test questions; then, after considering the interaction of features, p and students’ potential mastery of knowledge points i The weighted sum of the positive emotion vector o after interaction is obtained p , adjust the impact of positive emotions on students' knowledge mastery, and the expression is:

[0080] where p j =(q j1 ,q j2 ,…q jk ), s i =(α i1 ,α i2 ,…α ik );

[0081] Similarly, inputting negative sentiment When , the interaction layer uses self-attention to understand the interactive relationship between students’ negative emotions, students’ knowledge mastery and test answering, and obtains the negative emotion vector o after interaction. n ;

[0082] The S23 aggregation layer uses another self-attention module to calculate the different interaction emotions o p 、o n polymerization;

[0083] Specifically, the interactive emotion results of different categories are o p and o n They are respectively used as query and key in the attention mechanism, and the influence of emotions on students is adjusted according to the attention weight; the students' knowledge mastery is used as the value in the attention mechanism, different interactive emotions are aggregated, and the students' external cognition θ with integrated emotional features based on the students' characteristics is calculated. o , the expression is:

[0084] When modeling the external cognition of emotional influence, students’ potential knowledge mastery α ik Take it as 0.001. Step S2 can solve the problem that the students’ external cognitive modeling mechanism is not clear.

[0085] S3 integrates external cognition and students’ inherent cognition, introduces the NCB-IRM framework, and comprehensively predicts students’ personalized responses. ij ,considering two extreme cases where students’ emotional characteristics affect students’ responses;

[0086] S31 is only affected by the emotional characteristics. Student i correctly answers question j. The student learning response function is modeled as:

[0087] in, θo , α j 、b j They are students’ external cognitive level, test item discrimination and test item difficulty;

[0088] S32 is not affected by the emotional characteristics. Student i correctly answers question j. The student learning response function is modeled as:

[0089] Among them, r ij A represents the response of student i to question j, ij represents the emotion expressed by student i when answering question j, θ i 、a j d j They are students’ intrinsic cognitive level, test item discrimination and test item difficulty;

[0090] S33 assumes that the response to each question is statistically independent of the student's cognition and uses a Bernoulli distribution to model all student responses:

[0091] where η ij and ζ ij represent the probability that student i answers question j correctly based on the practice and emotional characteristics of the test questions, respectively;

[0092] S34 The student's response matrix is ​​a comprehensive response to the test question prior and the student's emotional prior; given a response r ij , students' cognitive θ i ,θ o , test variable a j , α j 、b j d j , the item response function of the student response is as follows:

[0093] S35 maximizes the item response function of student responses and obtains θ through a Markov chain Monte Carlo method based on Metropolis-Hastings. i ,a j ,α j ,d j ,b j The optimal solution:

[0094] S351 first randomizes all parameters to initial values;

[0095] S352 starts from the first number and traverses the samples of T moments in sequence;

[0096] S353 from θ i(t)~U(θ i(t-1) -δ θi ,θ i(t-1) +δ θi ) and collect a sample with probability Accept θ i ;

[0097] S354 from a j(t) ~U(a j(t-1) -δ aj ,a j(t-1) +δ aj ) and collect a sample with probability Accept a j ;

[0098] S355 from α j(t) ~U(α j(t-1) -δ αj ,α j(t-1) +δ αj ) and collect a sample with probability Accept α j ;

[0099] S356 from d j(t) ~U(d j(t-1) -δ dj ,d j(t-1) +δ dj ) and collect a sample with probability Accept j ;

[0100] S357 from b j(t) ~U(b j(t-1) -δ bj ,b j(t-1) +δ bj ) and collect a sample with probability Accept b j ;

[0101] S358 traverses all T moments, and if the convergence condition is met, it outputs θ i ,a j ,α j ,d j ,b j estimates;

[0102] S36 Predicting Personalized Student Responses

[0103] Combined with students' emotional data, step S3 can solve the problem of inaccurate personalized cognitive modeling of students.

[0104] Based on students’ real responses, S4 uses PMF to decompose the common characteristics of students’ answers and combines them with students’ personalized predicted responses to jointly predict students’ responses to test questions.

[0105] S41 performs probability matrix decomposition based on the student response R through PMF to obtain the characteristic matrix U of the student and the test question i 、V j ;

[0106] S411 assumes that the known response data and the distribution of student and test item feature vectors satisfy Gaussian distribution, and defines the conditional distribution on the observed response:

[0107] Among them U i represents the potential characteristics of students, V j Represents the potential characteristics of the test questions, N is the mean μ, and the variance is σ 2 Gaussian distribution probability density function.

[0108] S412 calculates the posterior probability of the potential characteristics of students and test questions as follows:

[0109] S413 To maximize the posterior probability of the potential features of students and test questions containing hyperparameters, it is necessary to minimize the sum of squared errors containing quadratic regularization terms in the objective function:

[0110] Among them, R is the true value, is the regularization parameter, U i and V j Represent the latent feature vectors of students and test questions respectively, represents the matrix norm, I ij is an indicator function (if student i answers question j, the value is 1; otherwise, it is 0);

[0111] S414 respectively for U i ,V j Derivative:

[0112] S415 Update U using stochastic gradient descent i ,V j :

[0113] Until convergence or reaching the maximum number of iterations, get U i ,V j Estimates.

[0114] S42 Predict student responses that include student commonalities

[0115] S43 calculates the weighted sum of the student's personalized prediction response and the student's common prediction response, as follows:

[0116] Among them, R p is the predicted student response, μ is the mean score of all students, r′ ij is the predicted student personalized response, R′ is the student common response predicted by PMF, and the ratio between the student’s individuality and commonality is adjusted by the parameter ρ;

[0117] S431 To make the predicted value close to the true value, the optimization goal is to minimize the function:

[0118] Among them, R p is the predicted student response, R is the true value, is the regularization parameter, U i and V j Represent the D-dimensional student and test question latent feature vectors, represents the matrix norm, I ij is the indicator function. When student i answers question j, I ij =1, otherwise 0;

[0119] S432 Estimation of r′ by Markov Chain Monte Carlo Method Based on Metropolis-Hastings ij , estimate the latent factor vector U by maximizing the posterior probability i 、V j , output the trained predicted response R p .

[0120] By integrating students' individuality and commonalities, step S4 can solve the problems of weak prediction ability and incomplete recommendation schemes in traditional recommendation schemes.

[0121] Based on the predicted student responses, S5 provides students with test questions of appropriate difficulty and forms a final recommendation list.

[0122] S51 Based on the predicted student response R p Design test difficulty [d1, d2];

[0123] The difficulty of the test item is calculated as the mean of the students' responses:

[0124] Among them, r ij is the response of student i to test item j, and n is the number of records of answering test item j;

[0125] S52 According to the difficulty of the test questions d j , filter out the list of recommended test questions R p∈[1-d2,1-d1]. Step S5 makes recommendations based on the students’ mastery of the test questions, which can solve the problem of unreasonable test question recommendation schemes.

[0126] The experimental method of this embodiment is further described below.

[0127] During training, Xavier is used to initialize the parameters. Specifically, these parameters are initialized from N(0, std 2 ) is filled with random values ​​sampled from n i Represents the input dimension of the neural network, n i+1 represents the output dimension of the neural network. In addition, epoch = 5 and lr (learning rate) = 0.001 were set. All models were implemented in PyTorch in Python. The experimental hardware configuration was a 2.3GHz Dual-Core Intel Core i5 with 8GB of RAM and the operating system was macOS Big Sur 11.3.1.

[0128] In this embodiment, the performance indicators used to evaluate the recommendation scheme proposed by the present invention are divided into two categories: precision indicators and non-precision indicators. Precision indicators include accuracy (ACC), area under the ROC curve (AUC), root mean square error (RMSE), and mean absolute error (MAE). The purpose is to measure whether the cognition of emotional perception is more in line with the students' actual mastery. Non-precision indicators include the rationality of the recommended test questions, that is, whether the recommended test questions meet the students' needs and are consistent with their cognitive status. The specific meanings of these five indicators are as follows:

[0129] Since it is impossible to accurately obtain the true value of students' cognition, it is difficult to directly evaluate the performance of students' emotional perception and cognitive diagnosis. Therefore, based on the existing work, this embodiment uses student responses to supervise the performance of the model. Specifically, depending on whether the predicted result is a score rate or an answer result (0 or 1), this embodiment can indirectly evaluate the performance of emotional perception and cognitive diagnosis from the perspectives of regression and classification. When the problem is treated as a regression task, the root mean square error (RMSE) and the mean absolute error (MAE) are used to quantify the distance between the predicted score and the actual score.

[0130] Where y' represents the observed value of the i-th observation, y represents the actual value of the i-th observation, and n represents the total number of observations. The smaller the RMSE and MAE values, the better the prediction effect of the model.

[0131] When the problem is treated as a classification task, the predicted result (1, 0) represents positive and negative instances, and the evaluation indicators often use the area under the ROC curve (AUC) and the prediction accuracy (ACC).

[0132] Where M represents the number of positive samples, N represents the number of negative samples, and CorrectPair represents the number of correctly ranked pairs of samples. TP (True Positives) represents the number of samples correctly predicted as positive by the model, TN (True Negatives) represents the number of samples correctly predicted as negative by the model, FP (False Positives) represents the number of samples incorrectly predicted as positive by the model, and FN (False Negatives) represents the number of samples incorrectly predicted as negative by the model. AUC and ACC values ​​range from 0 to 1, with values ​​closer to 1 indicating better predictions.

[0133] The test question recommendation in this embodiment does not consider recommending the most difficult / easiest test questions to students, but attempts to recommend test questions that are in line with their cognitive level. Therefore, the evaluation indicator is set as the correct answer rate of the recommended test questions, that is, the ratio of the number of correctly answered test questions to the number of recommended test questions:

[0134] The numerator represents the number of recommended questions that the student answered correctly, and the denominator represents the number of recommended questions. The larger the RACC, the closer the recommended questions are to the student's cognitive level.

[0135] In this example, three frameworks in CDM, IRT, MIRT, and DINA, are used for comparison. E-CDM represents CDM using the attention layer. Their prediction results in ASSIST2009 are shown in Table 2.

[0136] IRT: A logistic function is used to describe students' knowledge mastery. It can provide interpretable parameters (such as students' latent characteristics, problem discrimination and difficulty) to analyze students' performance. It is a continuous one-dimensional CDM.

[0137] MIRT: It has a logistic item response function and is an extension of the unidimensional IRT. It is a continuous multidimensional CDM.

[0138] DINA: It introduces the test knowledge point matrix Q, also focuses on the impact of students' sliding and guessing behaviors, and uses discrete binary vectors to represent students' knowledge mastery. It is a discrete multidimensional CDM.

[0139] In CDM, the discrimination and difficulty of the test items in IRT are respectively a j =4,d j =0; a in MIRTj =0,d j＝0 ;DINA's sliding, guessing, and step size are set to max slip =0.4, max guess =0.4, max step =1000. The input structures of CDM are: IRT(maximum student id+1, maximum test question id+1), MIRT(maximum student id+1, maximum test question id+1, latent_number), and DINA(maximum student id+1, maximum test question id+1, knowledge_number). In this embodiment, let latent_number=knowledge_number. In the ASSIST2009 dataset, the input structure of IRT is (4164, 17747), the input structure of MIRT is (4164, 17747, 102), and the input structure of DINA is (4164, 17747, 102); similarly, in the ASSIST2017 dataset, the input structure of IRT is (7784, 3164), the input structure of MIRT is (7784, 3164, 123), and the input structure of DINA is (7784, 3164, 123).

[0140] The experimental results are analyzed below with reference to the accompanying figures and tables.

[0141] Since ASSIST2009 does not record emotion-related data, we first verify the effect of the self-attention layer in this dataset. j At the same time, s is used as the query and key in self-attention. i As the value in self-attention, adjust the weight of the test question's influence on students' mastery.

[0142] Table 2 ASSIST2009 cognitive prediction results for middle school students

[0143] (1) After the self-attention module was added to the CDM in the ASSIST2009 dataset, the ACC of all E-CDMs increased compared with CDM, and the MAE of E-CDM decreased compared with CDM, which indicates that the self-attention module can more effectively capture the internal correlation between students and test questions.

[0144] (2) From the results of the models, E-IRT (ACC=0.659, MAE=0.375) has the best effect among these models, indicating that IRT has good adaptability to the ASSIST2009 dataset.

[0145] To explore whether emotional input further enhances model performance, this example conducts experiments on the ASSIST2017 dataset, which contains emotional data. Based on ASSIST2009, we add A-CDM and AE-CDM. A-CDM adds student emotional characteristics as model inputs to CDM, and experiments are conducted within the NCB-IRM framework. AE-CDM is the emotion-aware cognitive diagnostic model proposed in this example. Building on NCB-IRM, it uses PCA and hierarchical self-attention to model student emotions as external cognition. The prediction results for ASSIST2017 middle school students are shown in Table 3.

[0146] Table 3 ASSIST2017 cognitive prediction results for middle school students

[0147] (1) Compared with the models that only focus on students' answer logs (CDM, E-CDM), the CDMs that integrate emotional features (A-CDM, AE-CDM) have significantly improved in performance. When only emotional features are added as input, IRT (AUC = 0.535, RMSE = 0.575), MIRT (AUC = 0.551, RMSE = 0.633), and DINA (AUC = 0.508, RMSE = 0.572), while A-IRT (AUC = 0.701, RMSE = 0.465), A-MIRT (AUC = 0.608, RMSE = 0.533), and A-DINA (AUC = 0.543, RMSE = 0.554) have significantly improved in performance. This shows that the addition of emotional features has a significant impact on the predictive ability of the model.

[0148] (2) AE-CDM also has a certain degree of improvement over A-CDM. According to the evaluation indicators, compared with A-IRT, the AUC value of AE-IRT increased by 0.040 and the MAE value decreased by 0.019; compared with A-MIRT, the AUC value of AE-MIRT increased by 0.034 and the MAE value decreased by 0.034; compared with A-DINA, the AUC value of AE-DINA increased by 0.023 and the MAE value decreased by 0.017. This shows that the self-attention module plays a regulatory role in the emotional characteristics of students' cognition.

[0149] (3) Among the three CDMs, IRT has the best prediction results and EA-IRT has the most significant improvement effect. The AUC of EA-IRT is 38.64% higher than that of IRT, and the ACC of EA-IRT is 45.12% higher than that of the baseline IRT.

[0150] As shown in FIG3 , this embodiment explores the relationship between students' mastery of test questions and the difficulty of recommendations.

[0151] In the ASSIST2017 dataset, as the difficulty of the recommended questions increases, students' RACC for the recommended questions decreases, indicating that these models can recommend questions of appropriate difficulty levels based on students' needs. According to experimental results, the RACC of the AE-PMFDINA proposed in this paper (applying this invention to the DINA framework) exceeds that of PMF-DINA. By introducing student emotional characteristics, this paper can more accurately model students' cognition, more accurately predict their scores, and recommend questions within the set recommended difficulty range to each student.

[0152] This example also demonstrates the recommended test questions for two students through case analysis. Figure 4 shows the mastery of six knowledge points by Student A and Student B on ASSISIT 2017. It can be seen that Student A has a good grasp of knowledge points such as square-root and pattern-finding, but a poor grasp of knowledge points such as area and probability. Student B has a good grasp of knowledge points such as area and pattern-finding, but a poor grasp of knowledge points such as square-root, probability, and equation-solving.

[0153] In the test data set, the correct answer rates of student A and student B are both 50%. When test questions with a difficulty between 0.3 and 0.5 are selected for recommendation, the present invention recommends test questions 4558, 2115, and 1624 to student A, and recommends test questions 894, 2401, and 1597 to student B.

[0154] Figure 4(a) shows that Student A's grasp of the knowledge points "area" and "probability" is weak. Figure 4(b) shows the correspondence between some test questions and knowledge points. Recommended question 4558 tests the knowledge point "area," and question 2115 tests the knowledge point "probability." Student B's grasp of the knowledge points "square-root" and "probability" is weak. The recommended questions for Student B are question 894, which tests the knowledge point "square-root," and question 2401, which tests the knowledge point "probability." Because Student A has a strong grasp of "square-root," no related test questions were recommended.

[0155] From this case analysis, it can be seen that the personalized test question recommendation method proposed in the present invention can recommend corresponding test questions to each student based on their personalized learning status when recommending test questions with higher difficulty, and the test question recommendation results are highly interpretable.

[0156] Example 2:

[0157] The computer-readable storage medium of this embodiment stores a computer program thereon, which, when executed by a processor, implements the steps of the student emotion-aware personalized online test question recommendation method of Example 1.

[0158] The computer-readable storage medium of this embodiment may be an internal storage unit of the terminal, such as a hard disk or memory of the terminal; the computer-readable storage medium of this embodiment may also be an external storage device of the terminal, such as a plug-in hard disk, a smart memory card, a secure digital card, a flash memory card, etc. equipped on the terminal; further, the computer-readable storage medium may also include both an internal storage unit of the terminal and an external storage device.

[0159] The computer-readable storage medium of this embodiment is used to store computer programs and other programs and data required by the terminal. The computer-readable storage medium can also be used to temporarily store data that has been output or is to be output.

[0160] Example 3:

[0161] The computer device of this embodiment includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps of the student emotion-aware personalized online test question recommendation method of Example 1 are implemented.

[0162] In this embodiment, the processor can be a central processing unit, or other general-purpose processors, digital signal processors, application-specific integrated circuits, off-the-shelf programmable gate arrays or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc. The memory can include read-only memory and random access memory, and provide instructions and data to the processor. A part of the memory can also include non-volatile random access memory. For example, the memory can also store information about the device type.

[0163] Those skilled in the art will appreciate that the disclosed contents of the embodiments may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a hardware embodiment, a software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage and optical storage, etc.) containing computer-usable program code.

[0164] The present solution is described with reference to the method according to the embodiment of the present solution, and the flowchart and / or block diagram of the computer program product. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions; these computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate a device for implementing the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.

[0165] These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce a product including an instruction device that implements the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.

[0166] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.

[0167] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing related hardware through a computer program. The program can be stored in a computer-readable storage medium, and when executed, the program can include the processes in the above-described method embodiments. The storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0168] The examples described in the present invention are merely descriptions of the preferred embodiments of the present invention and are not intended to limit the concept and scope of the present invention. Without departing from the design concept of the present invention, various modifications and improvements made to the technical solutions of the present invention by engineers and technicians in this field should fall within the scope of protection of the present invention.

Claims

1. A personalized online test question recommendation method based on student emotion perception, characterized in that: The steps include: S1. Collect students’ online learning answering behaviors and corresponding emotional records, clean the data, and extract effective information; S2. Based on the students’ emotional data, PCA is used to classify students’ complex emotions, and the influence of emotions on students is adjusted through the weight of the self-attention module to model students’ external cognition; S3, integrating students’ external cognition and students’ inherent cognition, introducing the NCB-IRM framework, and comprehensively predicting students’ personalized responses; S4, using PMF to decompose the commonality of students’ answers, combined with students’ personalized predicted responses, to jointly predict students’ responses to test questions; S5. Based on the predicted student responses, provide students with test questions of appropriate difficulty to form a final recommendation list.

2. The method according to claim 1, characterized in that Step S1 collects the online learning answering behavior records including students' emotions and extracts effective information; specifically: S11, collect student ID, test question ID, knowledge point ID, student's emotional data E on the test question, and student's response R on the test question; E includes six emotional dimensions: bored, concentrated, confused, frustrated, offtask, and gaming; the emotional element A in each dimension ij It belongs to [0,1]. The closer it is to 1, the stronger the student's performance in this emotion is. Student i's response to question j is expressed as r ij Indicates; if r ij =1 means student i answered question j correctly; if r ij =0, it means that student i answered question j incorrectly; S12. Clean the data, delete empty entries and redundant answer records, and extract valid information as follows: Extract the potential knowledge mastery of student i at knowledge point k from student behavior and use α ik Indicates that α ik ∈[0,1]; The test matrix Q contains the relationship between all test questions and knowledge points. The element q jk Indicates the knowledge point tested by each test question q; if q jk =1, indicating that question j does not test knowledge point k; if q jk =0, indicating that question j does not test knowledge point k.

3. The method according to claim 1, characterized in that Step S2 models students’ external cognition based on students’ emotions, students’ answering behaviors and test matrix through PCA algorithm module and hierarchical self-attention module, specifically: S21. Using the PCA algorithm, the students’ complex emotions are clustered into two categories, namely positive emotions and negative emotions, expressed as: Where E is a matrix containing six emotion dimensions, · represents the dot product operation of the vector, It is the projection value of the sentiment sample in the direction of the first two principal components, V1 and V2 represent the feature vectors respectively; S22, two self-attention modules are used in the interaction layer to simulate the clustered emotions. For example, with Students' potential mastery of knowledge points α ik And the knowledge points tested in the test questions jk Interaction; Enter positive emotions When the self-attention module takes the sequence p of knowledge points tested in the test j As the key in the attention mechanism, the sentiment representation Sequences of students' potential mastery of knowledge points j As the query and value in the attention mechanism respectively; First, the student answer vector p is calculated by cosine similarity j With positive emotion vector The mutual influence weight of interaction ω p ; Then, after considering the feature interactions, by ω p and students’ potential mastery of knowledge points i The weighted sum of the positive sentiment vector o after interaction is obtained. p , the expression is: among them p j =(q j1 ,q j2 ,…q jk ),s i =(a i1 ,a i2 ,…a ik ); Similarly, inputting negative sentiment When , the interaction layer simulates the interaction between students’ negative emotions, students’ knowledge mastery and answering questions through self-attention, and obtains the negative emotion vector o after interaction. n ; S23, the aggregation layer uses another self-attention module to calculate different interactive emotions o p , o n polymerization; The interactive emotion results of different categories are o p and n As the query and key in the attention mechanism respectively, the student's knowledge mastery is used as the value in the attention mechanism, and different interactive emotions are aggregated to obtain the student's external cognition θ with integrated emotional features. o , the expression is:

4. The method according to claim 1, characterized in that: Step S3 integrates external cognition and students’ inherent cognition, introduces the NCB-IRM framework, and comprehensively predicts students’ personalized responses; given that students’ responses are ij , consider two extreme cases where students’ emotional characteristics affect their responses; specifically: S31. Only affected by the emotional characteristics, student i correctly answered question j, and the student learning response function was modeled as: Among them, θ o , α j , b j They are students’ external cognitive level, test item discrimination and test item difficulty; S32. Unaffected by the emotional characteristics, student i correctly answered question j. The student learning response function is modeled as: Among them, r ij represents the response of student i on question j, A ij represents the emotion expressed by student i when answering question j, θ i 、a j ,d j They are students’ intrinsic cognitive level, test item discrimination and test item difficulty; S33. Assuming that the response to each question is statistically independent of the students’ cognition, the responses of all students are modeled using a Bernoulli distribution: where η ij and ij represent the probability that student i correctly answers question j based on the test questions practiced and the sentiment characteristics, respectively; S34, the student's response matrix is ​​the comprehensive response to the test question prior and the student prior; given the response r ij , students’ cognition θ i ,θ o , test variable a j , α j 、b j ,d j , the item response function of the student response is as follows: S35, the item response function that maximizes the student response is obtained by a Markov chain Monte Carlo method based on Metropolis-Hastings. i 、a j , α j , b j d j The optimal solution of S36. Predicting Personalized Student Responses 5. The method according to claim 1, characterized in that Step S4 is based on the students' real responses, uses PMF to decompose the commonality of students' answers, combines students' personalized predicted responses, and jointly predicts students' responses to test questions; specifically: S41. Perform probability matrix decomposition on the student's true response through PMF to obtain the characteristic matrix U of the student and the test question. i 、V j ; S42. Predict responses that include students’ commonalities S43. Calculate the weighted sum of the student's personalized prediction response and the student's common prediction response, as follows: Among them, R p is the predicted student response, μ is the mean score of all students, r′ ij Personalize for predicted students Response, R′ is the student commonality response predicted by PMF, and the ratio between students’ individuality and commonality is adjusted by parameter ρ.

6. The method for recommending personalized online test questions based on student emotion perception according to claim 1, characterized in that: Step S5 provides students with test questions of appropriate difficulty based on the predicted student responses to form a final recommendation list; specifically: S51, based on the predicted student response R p Design test difficulty [d1, d2]; The difficulty of the test question is calculated by the mean of the students' responses: Among them, r ij is the response of student i to test item j, and n is the number of records of answering test item j; S52, according to the difficulty of the test questions d j , filter out the list of recommended test questions R p ∈[1-d2,1-d1].

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps in the student emotion-aware personalized online test question recommendation method as described in any one of claims 1 to 6 are implemented.

8. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps in the student emotion-aware personalized online test question recommendation method according to any one of claims 1 to 6 are implemented.

Images