Question recommendation method based on multi-factor modeling learning score prediction

Abstract

The invention discloses a question recommendation method based on multi-factor modeling learning score prediction, and relates to the technical field of education, and the method comprises the steps: carrying out the modeling of student features, and obtaining an overall feature vector; analyzing a corresponding relation between each question and the knowledge points, obtaining weight vectors of the knowledge points, performing normalization processing, constructing a knowledge point Embedding matrix, and combining the weight vectors of the knowledge points associated with the questions to calculate question feature vectors of the current question; constructing a score prediction model based on the historical question scores, the corresponding overall feature vectors and the question feature vectors, optimizing model internal parameters through a loss function of the score prediction model, and obtaining a trained score prediction model; and based on the trained score prediction model, predicting scores of the students on the candidate questions, constructing learning course recommendation rules, and recommending learning courses to the students. According to the invention, personalized training question recommendation for trainees is realized.

Description

Technical Field

[0001] The present invention relates to the field of educational technology, and more particularly to a topic recommendation method based on multi-factor modeling and learning score prediction. Background Art

[0002] In the current mainstream education model, teachers conduct teaching activities according to a pre-determined curriculum. Drawing on their accumulated teaching experience, teachers meticulously develop a unified learning plan for all students. During daily instruction, teachers assess students' performance based on multiple metrics, including attendance, class performance, and homework completion. Periodic exams are also used to further comprehensively assess students' mastery of knowledge. If a student's weaknesses in a particular area are identified, remedial measures, whether initiated by the teacher or discovered by the student, are typically implemented, including retraining. Throughout this process, curriculum planning is centrally controlled by professional educational planners, who treat every student equally, from the sequence of courses, class time allocation, to the choice of teaching methods. The greatest drawback of this model is that it completely ignores the significant differences among students in their speed of absorption, comprehension, and initiative. Students with excellent academic performance and strong receptive abilities may find the slow pace of instruction boring, limiting their potential for development. Students with weaker foundations and less advanced learning abilities may struggle to keep up, resulting in increasing knowledge gaps and a decline in their confidence.

[0003] Under the elective course system, students do have a certain degree of autonomy in their choices, and can independently select subjects and courses within the range of optional courses defined by the school. Most students tend to be guided by their own interests when making choices, and tend to choose areas that they are already interested in and are relatively proficient in. Although the school took into account the diversity and comprehensiveness of the disciplines when setting the range of optional courses, due to the limitations of the overall planning, students' choices are relatively limited. Moreover, this selection method based on interests and areas of expertise limits students' room for improvement in the knowledge points corresponding to the selected courses. In the long run, students may only linger in areas they are familiar with, finding it difficult to expand the boundaries of their knowledge, which is not conducive to the cultivation of comprehensive qualities for all-round development. For some learning tasks that require the integration of interdisciplinary knowledge and self-challenge, students may feel overwhelmed due to the lack of knowledge reserves in relevant elective courses.

[0004] With the rapid development of educational informatization, data-based analytical methods are gradually gaining prominence in the field. For example, by analyzing students' past academic records, data analysis tools can be used to pinpoint each student's weaknesses. Targeted training plans can then be tailored to address these weaknesses, aiming to strengthen them. This approach is undoubtedly a significant improvement over traditional teaching, fully accounting for individual differences in students' mastery of knowledge. However, it also has significant limitations. When developing study plans, it focuses solely on knowledge-based data, completely ignoring students' overall physical fitness and the crucial subjective willingness to learn. Every student's physical condition varies, as do their learning stamina and energy levels. If students are forced to study difficult and unfamiliar subjects for a long time, even with targeted training, repeated setbacks can easily lead to a loss of interest and aversion to learning. This can even negatively impact their physical and mental health, hindering their long-term development and the cultivation of lifelong learning skills.

[0005] Therefore, there is an urgent need for a personalized training question recommendation method that can comprehensively consider multi-dimensional factors such as students' knowledge mastery, physical fitness, and subjective learning willingness. Summary of the Invention

[0006] In view of this, the present invention provides a question recommendation method based on multi-factor modeling learning score prediction, which can comprehensively consider multi-dimensional factors such as students' knowledge mastery, physical fitness, and subjective learning willingness to recommend personalized training questions.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions:

[0008] A topic recommendation method based on multi-factor modeling and learning score prediction, including:

[0009] Model the student characteristics of each student and obtain the overall feature vector of the student;

[0010] Analyze the correspondence between each question and knowledge point, obtain the weight vector of the knowledge point and perform normalization processing, build the knowledge point embedding matrix, combine the weight vector of the question-related knowledge point, and calculate the question feature vector of the current question;

[0011] Constructing a score prediction model based on the historical question scores, the corresponding overall feature vectors, and the question feature vectors, calculating the loss function of the score prediction model, optimizing the internal parameters of the score prediction model, and obtaining a trained score prediction model;

[0012] Based on the trained score prediction model, the student's scores on the candidate questions are predicted, and a learning course recommendation rule is constructed to recommend learning courses to the students.

[0013] Preferably, the student characteristics include the student's ID and personal characteristics.

[0014] Preferably, the step of obtaining the overall feature vector of the student specifically includes:

[0015] Perform Embedding modeling for each student, and set the Embedding layer of the student ID to E id , mapping the student ID to a low-dimensional vector The low-dimensional vector of the student ID and attribute feature vector Splice and get the overall feature vector of the current student

[0016] Preferably, the calculation of the current question's feature vector Specifically include:

[0017]

[0018] Among them, l is the question q j The number of associated knowledge points, represents the normalized weight of the u-th knowledge point, Indicates question q j The associated u-th knowledge point The embedding vector of .

[0019] Preferably, the constructing of the score prediction model includes: using the overall feature vector and the question feature vector, combined with the student's past question scores as model input, to predict the student's score for the current question, wherein the calculation process of the Transformer's multi-head self-attention mechanism is as follows:

[0020]

[0021] X=(x i ,x2,…,x p );

[0022] Among them, X is the input sequence, which represents the ordered feature set of the joint features of the overall feature vector, the question feature vector, and the historical question score in the time dimension, x t is the overall eigenvector F Si , question feature vector and the history question score y it After the concatenation, first calculate the query matrix Q = XW Q , key matrix K = XW K Sum matrix V = XW V , where W Q 、W K 、W VIs a learnable weight matrix, then Q, K and V are divided into h heads, and for each head i, the attention score is calculated:

[0023]

[0024] Among them, d k is the dimension of the key vector;

[0025] Finally, the outputs of all heads are concatenated and passed through a linear layer to obtain the output of the multi-head self-attention mechanism:

[0026] MultiHead(X)=Concat(Attention1,Attention2,…,Attention h )W O ;

[0027] Among them, W O is a learnable weight matrix.

[0028] Preferably, the Transformer-based neural network model also includes a feedforward neural network, which consists of two linear layers and an activation function:

[0029] FFN(x)=ReLU(xW1+b1)W2+b2;

[0030] Where W1 and W2 are weight matrices, b1 and b2 are bias vectors, ReLU is the activation function, and the input x of the feedforward neural network is the output MultiHead(X) of the multi-head self-attention mechanism;

[0031] After being processed by the Transformer-based neural network model, the output is linearly transformed through a fully connected layer to obtain the predicted score y pred,ij :

[0032]

[0033] in, is the question feature vector, is the student feature vector, y i For students i For historical topics Constructed scoring sequence FC is a fully connected layer, which means linear transformation of the input.

[0034] Preferably, the loss function specifically includes:

[0035]

[0036] Among them, y pred.nis the nth sample question The prediction score, N is the number of samples, y true,n is the nth sample question The real score;

[0037] Based on the learning and test data of several students, the Adam optimizer is used to update the learnable parameters in the Transformer-based neural network model to minimize the loss function L. After 200 rounds of iterative training, or when the loss of the learning and test data decreases by less than 5% in 10 rounds of iterations, the training is stopped early to produce a score prediction model based on student and question characteristics.

[0038] Through the above technical solutions, it can be seen that compared with the existing technology, the present invention discloses a method for recommending questions based on multi-factor modeling learning score prediction. By accurately predicting scores, it provides students with personalized training questions, enabling them to learn more efficiently, concentrate on overcoming weak knowledge points, and reduce ineffective learning time; recommends questions that meet the students' subjective learning intentions and ability levels, avoiding students losing interest in learning due to questions that are too difficult or too easy; comprehensively considers the students' physical fitness and personal characteristics, formulates a reasonable study plan, and avoids the negative impact on students' physical and mental health due to excessive learning or learning inappropriate content. In the long run, it helps to cultivate students' lifelong learning ability and improve their comprehensive quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0040] Figure 1 A schematic structural diagram of the present invention;

[0041] Figure 2 This is a diagram of topic recommendations provided by the present invention. DETAILED DESCRIPTION

[0042] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0043] The embodiment of the present invention discloses a method for recommending topics based on multi-factor modeling learning score prediction, such as Figure 1 As shown, including:

[0044] Model the student characteristics of each student and obtain the overall feature vector of the student;

[0045] Analyze the correspondence between each question and knowledge point, obtain the weight vector of the knowledge point and perform normalization processing, build the knowledge point embedding matrix, combine the weight vector of the question-related knowledge point, and calculate the question feature vector of the current question;

[0046] Constructing a score prediction model based on the historical question scores, the corresponding overall feature vectors, and the question feature vectors, calculating the loss function of the score prediction model, optimizing the internal parameters of the score prediction model, and obtaining a trained score prediction model;

[0047] Based on the trained score prediction model, the student's scores on the candidate questions are predicted, and a learning course recommendation rule is constructed to recommend learning courses to the students.

[0048] In a specific embodiment, the student characteristics include the student's ID and personal characteristics, and modeling is performed on the student's ID and multi-dimensional personal characteristics.

[0049] Students' personal characteristic data covers multiple aspects, including basic information such as age and gender, learning-related information such as study habits and study time, physical fitness indicators (such as fatigue recovery ability and concentration duration), and subjective learning intention data (which can be obtained through questionnaires and learning behavior analysis, such as the proportion of time spent actively learning and interest in different subjects). Using feature engineering methods in machine learning, this data is preprocessed, including data cleaning and normalization, to convert it into feature vectors suitable for model input. These features are then transformed through neural network layers to obtain student attribute characteristics.

[0050] In a specific embodiment, obtaining the overall feature vector of the student specifically includes:

[0051] Perform Embedding modeling for each student, and set the Embedding layer of the student ID to E id , mapping the student ID to a low-dimensional vector The low-dimensional vector of the student ID and attribute feature vector Splice and get the overall feature vector of the current student The attribute feature vector is transformed from the personal characteristics of the students.

[0052] In a specific embodiment, the correspondence between each question and the knowledge point is analyzed, the weight vector is obtained and normalized, the knowledge point embedding matrix is ​​constructed, the knowledge point is modeled and embedded, and the weight vector of the knowledge point associated with the question is combined. Calculate the eigenvector of the current question Specifically include:

[0053]

[0054] Among them, l is the question q j The number of associated knowledge points, represents the normalized weight of the u-th knowledge point, Indicates question q j The associated u-th knowledge point The embedding vector of .

[0055] In a specific embodiment, a score prediction model is constructed, the input of which includes the student ID, the student's personal characteristics, the weight vector of the question-related knowledge points, and the score of the historical questions, and the output is the student's score or correct probability for the current question.

[0056] The student ID and personal characteristics are converted into student feature vector F through the student model Si , the weight of the knowledge point of the question is converted into the feature vector of the question For the history question score, let the student s i For historical topics The score sequence is

[0057] In a specific embodiment, constructing a score prediction model includes: using the overall feature vector and the question feature vector, combined with the student's historical question scores as model input, to predict the student's score or correct probability for the current question.

[0058] The overall model process is as follows: the student, current question, and past question score data enter their respective feature extraction models and are converted into the student's overall features, current question features, and past question score features. The features of the three different dimensions are converted into joint features through splicing operations and enter the score prediction model.

[0059] The overall model structure is: the feature transformation embedding model of three different dimensions is connected to the score prediction model.

[0060] The score prediction model is a Transformer model, which mainly consists of a multi-head self-attention mechanism and a feedforward neural network. The calculation process of the multi-head self-attention mechanism is as follows:

[0061]

[0062] X=(x i ,x2,…,x p );

[0063] Among them, X is the input sequence, which represents the ordered feature set of the joint features of the overall feature vector, the question feature vector, and the historical question score in the time dimension, x t is the overall eigenvector F Si , question feature vector and history question scores it After concatenating the vector, first calculate the query matrix Q, key matrix K, and value matrix V.

[0064] Q=XW Q

[0065] K=XW K

[0066] V=XW V

[0067] Where W Q 、W K 、W V Is a learnable weight matrix, then Q, K and V are divided into h heads, and for each head i, the attention score is calculated:

[0068]

[0069] Among them, d k is the dimension of the key vector. Q, K, and V are all obtained by linearly transforming the input sequence X through their respective weight matrices W, representing different ways in which the model processes and utilizes input information in the self-attention calculation.

[0070] Finally, the outputs of all heads are concatenated and passed through a linear layer to obtain the output of the multi-head self-attention mechanism:

[0071] MultiHead(X)=Concat(Attention1,Attention2,…,Attention h )W O ;

[0072] Among them, W O is a learnable weight matrix.

[0073] In a specific embodiment, the Transformer-based neural network model further includes a feedforward neural network, which consists of two linear layers and an activation function:

[0074] FFN(x)=ReLU(xW1+b1)W2+b2;

[0075] Where W1 and W2 are weight matrices, b1 and b2 are bias vectors, and ReLU (Rectified Linear Unit) is an activation function that performs a nonlinear transformation on the input value, setting negative values ​​to 0 and leaving positive values ​​unchanged. FFN is a feedforward neural network, which represents two linear transformations and one nonlinear activation operation on the input. The input x of the feedforward neural network is the output of the multi-head self-attention function MultiHead(X);

[0076] After being processed by the Transformer-based neural network model, the output is linearly transformed through a fully connected layer to obtain the predicted score y pred,ij :

[0077]

[0078] in, is the question feature vector, is the student feature vector, y i For students i For historical topics Constructed scoring sequence FC is a fully connected layer (Fully Connected Layer), which means that the input is linearly transformed.

[0079] In a specific embodiment, the binary cross entropy loss function (BCELoss) is used to train the model. Assume that there are N samples in the training data set. For the nth sample, the student Topic The true score (correct is 1, wrong is 0) is y true,n , the predicted score (probability value) is y pred.n .

[0080] The calculation formula of the binary cross entropy loss function is:

[0081]

[0082] Among them, y pred.n is the nth sample question The prediction score, N is the number of samples, y true,n is the nth sample question The real score;

[0083] Based on the learning and test data of several students, the Adam optimizer is used to update the learnable parameters in the Transformer-based neural network model to minimize the loss function L. After multiple iterative training, a score prediction model based on student and question characteristics is produced.

[0084] Through the score prediction model reasoning, the student's score or accuracy rate for a large number of candidate questions is predicted, combined with the student's predicted score for the question, such as Figure 2 As shown in the figure, we construct a course recommendation rule. For example, we prioritize history questions with low predicted scores and a low number of history points studied. We reduce the recommendation probability for questions with high predicted scores to ensure student learning enthusiasm. We periodically recommend questions with very high predicted scores to consolidate knowledge mastery. We also limit recommendations to questions with rapidly increasing predicted scores to improve learning efficiency. Based on the score prediction model and course recommendation rules, we provide each student with a personalized library of recommended questions for selection.

[0085] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0086] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A topic recommendation method based on multi-factor modeling learning score prediction, characterized in that: include: Model the student characteristics of each student and obtain the overall feature vector of the student; Analyze the correspondence between each question and knowledge point, obtain the weight vector of the knowledge point and perform normalization processing, build the knowledge point embedding matrix, combine the weight vector of the question-related knowledge point, and calculate the question feature vector of the current question; Constructing a score prediction model based on the historical question scores, the corresponding overall feature vectors, and the question feature vectors, calculating the loss function of the score prediction model, optimizing the internal parameters of the score prediction model, and obtaining a trained score prediction model; Based on the trained score prediction model, the student's scores on the candidate questions are predicted, and a learning course recommendation rule is constructed to recommend learning courses to the students.

2. The method for recommending topics based on multi-factor modeling and learning score prediction according to claim 1, characterized in that: The student characteristics include the student's ID and personal characteristics.

3. The method for recommending topics based on multi-factor modeling and learning score prediction according to claim 2, characterized in that: The obtaining of the overall feature vector of the student specifically includes: Perform Embedding modeling for each student, and set the Embedding layer of the student ID to E id , mapping the student ID to a low-dimensional vector The low-dimensional vector of the student ID and attribute feature vector Splice and get the overall feature vector of the current student 4. The method for recommending topics based on multi-factor modeling and learning score prediction according to claim 1, characterized in that: The calculation of the current question's feature vector Specifically include: Among them, l is the question q j The number of associated knowledge points, represents the normalized weight of the u-th knowledge point, Indicates question q j The associated u-th knowledge point The embedding vector of .

5. The method for recommending topics based on multi-factor modeling and learning score prediction according to claim 1, characterized in that: The score prediction model is constructed by taking the overall feature vector and the question feature vector, and combining them with the student's previous question scores as model inputs to predict the student's score for the current question. The calculation process of the Transformer's multi-head self-attention mechanism is as follows: X=(x i ,x2,…,x p ); Among them, X is the input sequence, which represents the ordered feature set of the joint features of the overall feature vector, the question feature vector, and the historical question score in the time dimension, x t is the overall eigenvector F Si , question feature vector and history question scores it After the concatenation, first calculate the query matrix Q = XW Q , key matrix K = XW K Sum matrix V = XW V , where W Q 、W K 、W V Is a learnable weight matrix, then Q, K and V are divided into h heads, and for each head i, the attention score is calculated: Among them, d k is the dimension of the key vector; Finally, the outputs of all heads are concatenated and passed through a linear layer to obtain the output of the multi-head self-attention mechanism: MultiHead(X)=Concat(Attention1,Attention2,…,Attention h )W O ; Among them, W O is a learnable weight matrix.

6. The method for recommending topics based on multi-factor modeling and learning score prediction according to claim 5, characterized in that: The Transformer-based neural network model also includes a feedforward neural network, which consists of two linear layers and an activation function: FFN(x)=ReLU(xW1+b1)W2+b2 Where W1 and W2 are weight matrices, b1 and b2 are bias vectors, ReLU is the activation function, and the input x of the feedforward neural network is the output MultiHead(X) of the multi-head self-attention mechanism; After being processed by the Transformer-based neural network model, the output is linearly transformed through a fully connected layer to obtain the predicted score y pred,ij : in, is the question feature vector, is the student feature vector, y i For students i For historical topics Constructed scoring sequence FC is a fully connected layer.

7. The method for recommending questions based on multi-factor modeling and learning score prediction according to claim 1, characterized in that: The loss function specifically includes: Among them, y pred.n is the nth sample question The prediction score, N is the number of samples, y true,n is the nth sample question The real score; Based on the learning and test data of several students, the Adam optimizer is used to update the learnable parameters in the Transformer-based neural network model to minimize the loss function L. After 200 rounds of iterative training, or when the loss of the learning and test data decreases by less than 5% in 10 rounds of iterations, the training is stopped early to produce a score prediction model based on student and question characteristics.

Images