A personalized knowledge tracing method based on multi-concept representation

Abstract

The application discloses a kind of personalized knowledge tracking methods based on multi-concept representation, comprising: collecting student answer record data set and pre-processing, generate multi-concept answer record, form the concept coding matrix x for representing answer history record and the difficulty information matrix d of personalized history record based on multi-concept answer record, after splicing concept coding matrix x and difficulty information matrix d, it is sent into recurrent neural network, and student answer prediction is carried out;The application enriches input features by using multi-hot encoding to represent multiple concepts involved in the question, and considers the accuracy of each concept for different students as a personalized basis, realizes the personalized history record representation of each student, significantly improves the model prediction performance, and also realizes the personalized knowledge tracking of different students.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of knowledge tracking in smart education, and particularly relates to a personalized knowledge tracking method based on multi-concept representation. BACKGROUND

[0002] Knowledge tracking is a key technology in the field of smart education, aiming to predict students' future answering performance according to their historical answering records. This technology has strong practical significance for online education platforms, which can assess students' knowledge mastery level and plan the next stage of learning according to students' knowledge level to achieve personalized learning. In recent years, due to the powerful computing and predicting capabilities of deep learning algorithms, knowledge tracking models based on deep learning have emerged in large numbers.

[0003] In real life, each question involves multiple knowledge points to varying degrees, so the question should be a comprehensive representation of multiple knowledge points. However, existing conventional knowledge tracking methods often replace the question with a single concept for model training, which will cause a certain degree of input feature missing problem.

[0004] In real life, due to individual differences among students, even the same answering order should have different answering results. However, existing conventional knowledge tracking models treat students as a group to predict their knowledge mastery level, which deviates from the actual situation and fails to effectively track personalized learning.

[0005] In real life, since learning is a continuous and cumulative process, the knowledge points involved in the previous questions in the answering record will affect the knowledge mastery level of subsequent questions. However, existing conventional knowledge tracking models do not fully consider historical information when predicting students' knowledge level, which will result in poor prediction effect.

[0006] The patent document with application number 202110928810.3 discloses a personalized knowledge tracking method and system that fuses learning behavior features. It uses a convolutional neural network to extract effective features from a composite vector composed of learning behavior and answering result data. It also extracts question information features containing knowledge points through a denoising autoencoder. Finally, it combines learning behavior features and question information features, passes them through an LSTM network and a fully connected layer, and obtains the students' knowledge mastery level. In the modeling process, the present application fuses a series of learning behavior features of students in the learning process and rich information of the exercises and knowledge points, and more accurately predicts the knowledge mastery level of each student. However, its scheme also has the following problems: 1. It does not fully consider historical information when predicting students' knowledge level, which will result in poor prediction effect; 2. It does not make full use of personalized historical record representation.

[0007] To solve the above problems, a personalized knowledge tracking method based on multi-concept representation is needed to solve the existing problems. SUMMARY

[0008] To solve the above problems, the present application provides a personalized knowledge tracking method based on multi-concept representation, which solves the problems of rich input features of multiple concepts involved, considers different student individualization, combines historical records and attention mechanisms, and accurately predicts the knowledge mastery level of students.

[0009] The embodiment of the present application provides a personalized knowledge tracking method based on multi-concept representation.

[0010] The first aspect: a personalized knowledge tracking method based on multi-concept representation, comprising:

[0011] S1, preprocessing the data set of student answer records, obtaining each student's answer questions and all concepts involved in the questions, and generating multi-concept based answer records;

[0012] S2, representing each student's multi-concept answer record with multi-hot encoding to form a concept encoding matrix x for representing the answer history record;

[0013] S3, calculating the correct rate of each student for different concepts and representing it with multi-hot encoding to form a difficulty information matrix d for representing the individualized history record;

[0014] S4, concatenating the concept encoding matrix x and the difficulty information matrix d and inputting them into a recurrent neural network, introducing an attention mechanism in the recurrent neural network, calculating the correlation coefficient between the historical mastery hidden state and the current time hidden state, and updating the current time hidden state;

[0015] S5, according to the cross-entropy loss function, using the back propagation mechanism to train the recurrent neural network, updating the model parameters, obtaining the trained recurrent neural network, and predicting the student's answer.

[0016] Optionally, the S1 obtains each student's answer questions and all concepts involved in the questions, comprising the steps of:

[0017] S11, reading the data set file, deleting the rows with NA or NULL values in the data set, and the rows with correct column not being binary values;

[0018] S12, numbering the questions and concepts, finding all concept numbers corresponding to the question numbers, and generating a dictionary.

[0019] Optionally, the S1 generates multi-concept based answer records, comprising: converting each student's answer questions and all concepts involved in the questions into a three-row form according to the student ID based on the dictionary, wherein:

[0020] The first behavior {q1, q2,..., q i ,...,q n}∈Q, where Q is a set of single student's answers to questions, N q is the number of questions;

[0021] The second behavior {c1|c3, c2|c3,..., c i |c j ,}∈C, where C is a set of concepts involved in single student's answers to questions, N c is the number of concepts;

[0022] The third behavior {r1, r2,..., r i ,...,r n}, r i ∈{0,1}, where r i is whether the question q i is answered correctly.

[0023] Optionally, the concept encoding matrix x is represented as:

[0024]

[0025] where Batch is the batch size, Maxstep represents the maximum step, N c is the number of concepts, q∈[0, Batch-1], v∈[0, Maxstep-1], k∈[0, 2*Nc-1].

[0026] Optionally, the difficulty information matrix d is represented as:

[0027]

[0028] where c is the concept accuracy.

[0029] 1. Optionally, the concept accuracy c j is obtained by three rows, and the formula is:

[0030]

[0031] where n j is the total number of times that concept j appears in the second row, is the number of times that concept j is answered correctly in the third row.

[0032] Optionally, the updating of the current time hidden state in S4 is to update the current time hidden state by using the hidden state of the previous time, and the formula is:

[0033] m t = cat(x, d) (4)

[0034] h t ← LSTM(m t , h t-1 ) (5)

[0035] where c j is concatenation, m t is input after concatenation, h t is current time hidden state, and h t-1 is previous time hidden state.

[0036] Optionally, the previous time hidden state is a comprehensive hidden state The comprehensive hidden state is obtained by the following steps:

[0037] S41, obtaining hidden representation u f using tanh activation function, and the formula is:

[0038] u f = tanh(Wh f +b) (6)

[0039] where h f is hidden state at time f, 1≤f≤t, W and b are neural network model parameters;

[0040] S42, obtaining importance correlation coefficient w i using softmax function, and the formula is:

[0041] w f = softmax(u f ) (7)

[0042] where u f is hidden representation at time f, and w f is correlation coefficient at time f;

[0043] S43, calculating hidden state correlation of hidden state h f at time f, and the formula is:

[0044]

[0045] S44, summing hidden state correlation before time t-2, and then concatenating with hidden state at time t-1 as comprehensive hidden state at time t-1, and the formula is:

[0046]

[0047] where, The hidden state correlation at time f is used.

[0048] The second aspect: an electronic device, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method provided by the first aspect when executing the program.

[0049] The third aspect: a non-transitory computer readable storage medium, having stored thereon a computer program, which, when executed by a processor, implements the steps of the method provided by the first aspect.

[0050] Advantages of the present application:

[0051] 1、The present application enriches the input features by using multi-hot encoding to represent multiple concepts related to the problem, and considers the accuracy of each concept of different students as a personalized basis, to realize the personalized history record representation of each student, significantly improve the model prediction performance, and also realize the personalized knowledge tracking of different students.

[0052] 2、The present application designs a self-attention mechanism to calculate the correlation coefficient of historical pre-knowledge mastery, i.e. the importance coefficient, and the weighted sum is used to represent the final knowledge mastery, which combines the history record and the attention mechanism, simulates the accumulation of learning, and improves the prediction ability of the student's knowledge mastery. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1 The flowchart of the personalized knowledge tracking method of the present application;

[0054] Figure 2 The principle flowchart of the personalized knowledge tracking method of the present application;

[0055] Figure 3 The knowledge tracking task example diagram provided by the embodiment of the present application;

[0056] Figure 4 The visualization diagram of the present application for predicting the knowledge mastery of students on multiple concepts;

[0057] Figure 5 The structural schematic diagram of the electronic device of the present application. DETAILED DESCRIPTION

[0058] The embodiments of the present application are described in detail below, and the examples of the embodiments are shown in the drawings, wherein the same or similar symbols represent the same or similar elements or elements with the same or similar functions throughout. The embodiments described below by referring to the drawings are exemplary, only for explaining the present application, and cannot be understood as limiting the present application.

[0059] The existing conventional knowledge tracking method has the problems of: a single concept is often used instead of a problem for model training, which will cause a certain degree of input feature missing problem; individual differences among students cannot effectively track personalized learning; and the prediction of student knowledge level is not good due to insufficient consideration of historical information.

[0060] To solve the above problems, the present application provides a personalized knowledge tracking method based on multi-concept representation, Figure 1 The flowchart of the personalized knowledge tracking method based on multi-concept representation provided by the embodiment of the present application is shown.

[0061] For the convenience of understanding, the professional terms involved in the present application are first explained:

[0062] Batch Size (batch size): the number of samples used for each iteration training;

[0063] Multi-Hot (multi-hot encoding): used to represent the case of multiple categories being activated at the same time. Each category corresponds to an independent register bit, and multiple bits can be 1 at the same time, while other bits are 0;

[0064] Knowledge Hidden State (hidden state): the hidden state is the internal state used by RNN to store and transmit information when processing input sequences. It is updated at each time step to capture information and contextual relationships in the input sequence, and is used to represent the degree of mastery of students' knowledge.

[0065] As shown in Figure 2 The method of the present application comprises:

[0066] S1, pre-process the student answer record data set, obtain the answer question of each student and all the concepts involved in the question, and generate a multi-concept based answer record.

[0067] The data set of student answer record generally uses a csv file, reads the data set csv file, and pre-processes the obtained data set, including deleting the rows with NA or NULL values in the data set, and the rows with correct column not being binary values (1 or 0) and the like.

[0068] Then renumber the questions and concepts, for example, starting from 1, find all the concept numbers corresponding to the question numbers to generate a dictionary.

[0069] Then according to the dictionary, convert each student's answer question and all the concepts involved in the question into a three-row format according to the student ID, wherein:

[0070] The first row is the total number of questions answered by this student, represented as:

[0071]

[0072] where Q is the set of questions answered by a single student, N q is the number of questions.

[0073] The second line represents the multiple concepts corresponding to each question, denoted as:

[0074]

[0075] where C is the set of concepts involved in the answers to the questions by a single student, N c is the number of concepts.

[0076] The third line represents whether each question is answered correctly, denoted as:

[0077] {r1,r2,...,r i ,...,r n},r i ∈{0,1}

[0078] where r i is the answer to question q i .

[0079] S2, the multiple-concept answer records of each student are represented by multi-hot encoding to form a concept encoding matrix x representing the answer history records.

[0080] To distinguish between correct and incorrect answer records of each student, the concept encoding position can be used to represent the concept encoding matrix x, which can be represented as:

[0081]

[0082] where Batch is the batch size, Maxstep is the maximum step, N c is the number of concepts, i∈[0, Batch-1], j∈[0, Maxstep-1], k∈[0, 2*Nc-1].

[0083] For example, according to the Batch size of 32 and the Maxstep of 50, three rows are read, where Batch represents the Batch size, and Maxstep represents the maximum step. That is, each Maxstep is considered as a group of answer records, and a total of Batch size groups are trained.

[0084] In this concept encoding matrix x, if a student answers the first question involving concepts i and j correctly, the first i bit and the first j bit of the first row in the third dimension of x are both 1, i.e.:

[0085]

[0086] If the answer is wrong, then 1 is set on i+Ns and j+Ns bits, i.e.:

[0087]

[0088] For other positions not involved in the concept, the value is set to 0, and the remaining rows are also expressed in the same way. Finally, the concept encoding matrix x of the multi-concept answer history record is formed.

[0089] S3, calculate the correct rate of each student for different concepts, and express it with multi-hot encoding to form the difficulty information matrix d for representing personalized history records.

[0090] To realize the personalized knowledge tracking of students, the answer level of students is used as auxiliary information for tracking to form the difficulty information matrix d for representing personalized history records. The difficulty information matrix d can be expressed as:

[0091]

[0092] where Batch is the batch size, Maxstep represents the maximum step, N c is the concept number, i∈[0, Batch-1], j∈[0, Maxstep-1], k∈[0, 2*Nc-1], and c is the concept correct rate.

[0093] For example: according to the Batch size of 32 and the Maxstep of 50, three rows are read, where Batch represents the Batch size, and Maxstep represents the maximum step. That is, Maxstep is taken as a group of answer records, and a total of Batch size groups are trained.

[0094] c is the concept correct rate, which is obtained by calculating the correct rate of each student for each concept as the learning level auxiliary information according to the three rows. All concepts in the second row are traversed, and the number is counted. The historical correct rate is calculated according to the corresponding third row answer, i.e.:

[0095]

[0096] where n j is the total number of times that concept j appears in the second row, is the number of times that concept j is answered correctly in the third row.

[0097] In the difficulty information matrix d, if the student answers correctly on the first question involving concepts i and j, then the first i bit of the corresponding first row in the third dimension of x is the correct rate c i of concept i, and the j bit is the correct rate c j of concept j, i.e.:

[0098]

[0099] If the answer is wrong, the correct rate value is set on i+Ns and j+Ns bits, i.e.:

[0100]

[0101] The position value of other concepts not involved is set to 0, and the remaining rows are also expressed in the same way, and finally the difficulty information matrix d of the personalized history record representation is formed.

[0102] S4, the concept encoding matrix x and the difficulty information matrix d are spliced and sent to the recurrent neural network, the attention mechanism is introduced in the recurrent neural network, the correlation coefficient of the history mastering degree hidden state and the current time hidden state is calculated, and the current time hidden state is updated.

[0103] Fusion of multi-concept history record and personalized history record, splicing (cat) of multi-concept encoding matrix x and difficulty information matrix d, as the input of recurrent neural network (LSTM), using the hidden state of the previous time to update the hidden state coefficient h t of the current time, the formula is:

[0104] m t = cat(x, d) (4)

[0105] h t ← LSTM(m t , h t-1 ) (5) Wherein, m t is the spliced input, h t is the current time hidden state, h t-1 is the hidden state of the previous time;

[0106] The hidden state h is the intermediate value generated by the multi-layer iteration of the recurrent neural network, which is also considered as the mastering degree of students to each knowledge point in the knowledge tracking field.

[0107] The multi-concept history record concept encoding matrix x and the corresponding personalized history record difficulty information matrix d are spliced and input into the neural network, and the generated hidden state contains the information of multi-concept and personalized difficulty.

[0108] Considering the importance of history hidden state, the inheritance and continuity in the learning process of LSTM are maintained, and the hidden state coefficient h t of the current time is updated using the hidden state of the previous time, specifically:

[0109] In order to calculate the correlation coefficient of the hidden state at different times, the hidden state is input into a multi-layer perceptron, the complex linear mapping of model data is enhanced, and the comprehensive hidden state As the hidden state h in the previous moment t-1 Then, the hidden state h at the current time is obtained using formula (5). t Specifically:

[0110] The hidden representation u is generated using the tanh activation function. f This enhances the complex nonlinear mapping of the model data, and then the importance coefficients, i.e., the weights w, are generated through the softmax function. f The formula is:

[0111] u f =tanh(Wh) f +b) (6)

[0112] w f =softmax(u f (7)

[0113] Among them, h i Let be the hidden state at time f, 1 ≤ f ≤ t, and W and b be the LSTM model parameters. These are LSTM learnable weights. It is LSTM bias; u f Let w be the hidden representation at time f. f Let f be the correlation coefficient at time f.

[0114] Since the hidden states at all time points have varying degrees of influence on the knowledge state, their total influence is defined as the weighted sum of the hidden states at all time points. To emphasize the most recent knowledge state, i.e., the hidden state at time t-1, the hidden states before time t-1 and t-2 are concatenated, as shown in the formula:

[0115]

[0116]

[0117] in, The hidden state correlation at time f is the hidden state correlation. This represents the weighted sum at time t-2; To obtain the combined hidden state at time t-1 by concatenating the hidden states before time t-1 and t-2.

[0118] The final obtained comprehensive hidden state As input to the model prediction, the hidden state h at the current time is updated using formula (5). t， This approach takes into account both the impact of previous moments on the present and the importance of the previous t-1 moment.

[0119] According to the knowledge mastery of the student, that is, the hidden state h of the model t , the LSTM is used to predict the next time performance of the student, the comprehensive hidden state h is passed through a linear layer to convert it into a fixed output dimension, and then a softmax activation function is used to generate a prediction value The formula is:

[0120]

[0121] S5, according to the cross-entropy loss function, the recurrent neural network is trained by using the back propagation mechanism, the model parameters are updated, and a trained recurrent neural network is obtained, and the student answer prediction is performed.

[0122] The sum of the cross-entropy losses between the predicted student answer performance and the real label is calculated, the gradient of the loss with respect to the model parameters is calculated by using the back propagation mechanism, and the model parameters are updated in the opposite direction of the gradient, wherein the formula of the cross-entropy loss function is:

[0123]

[0124] As shown in Figure 3 , for example: three students have done four questions at t4 time, each question involves a certain number of concepts, for example, the first question of student L1 involves concepts c1 and c2, and the present application predicts the subsequent performance according to the previous answer record of the student, that is, the answer of the student at t5 time and the mastery of the concept c3 are predicted according to the question information, that is, the concept information of the student L1 at t4 time.

[0125] As shown in Figure 4 , Figure 4 The upper part of the dashed line is a personalized visualization result represented by multiple concepts, and the lower part of the dashed line is a visualization result based on a single concept. The darker the color, the better the student masters, and the green and blue circles represent two different concepts, and the red edge represents the correct answer, and the black edge represents the wrong answer.

[0126] The present application also discloses a personalized knowledge tracking system based on multiple concept representation for the above method, as shown in Figure 5 , the system comprises:

[0127] An input module is used to input a student answer history record data set;

[0128] A preprocessing module is used to preprocess the data set of the student answer record, and obtain a multiple-concept answer record;

[0129] A concept encoding matrix x module is used to obtain a concept encoding matrix x of the answer history record according to the multiple-concept answer record;

[0130] ​a difficulty information matrix d module, obtaining a difficulty information matrix d representing personalized historical records according to the multi-concept answer record;

[0131] a recurrent neural network module, performing student answer prediction according to the spliced concept encoding matrix x and the difficulty information matrix d;

[0132] an output module, outputting the student answer prediction result.

[0133] The system of the present application collects the data set of student historical answer records, pre-processes, obtains the answer question of each student and all the concepts involved in the question, generates a multi-concept answer record dictionary, then represents each student's multi-concept answer record with multi-hot encoding, forms a concept encoding matrix x representing the answer historical record, and the accuracy of each student on different concepts, forms a difficulty information matrix d representing personalized historical records.

[0134] The obtained concept encoding matrix x and difficulty information matrix d are spliced and sent to a recurrent neural network for student answer prediction, and the student answer prediction result is output; the multi-hot encoding of multiple concepts involved in the question enriches the input features, and the accuracy of different students on each concept is calculated as a personalized basis, realizing the personalized historical record representation of each student. In addition, the historical record and the attention mechanism are combined to simulate the accumulation of learning, improving the prediction ability of the student's knowledge mastery.

[0135] The present application also provides an electronic device, Figure 6 The structure diagram of the electronic device provided by the embodiment of the present application is shown in Figure 6 As shown, the electronic device can include a processor, a communications interface, a memory and a communications bus, wherein the processor, the communications interface and the memory complete mutual communication through the communications bus. The processor can invoke the logical instructions in the memory, for example, to execute the following method:

[0136] S1, pre-processing the data set of student answer records, obtaining the answer question of each student and all the concepts involved in the question, generating a multi-concept answer record;

[0137] S2, representing each student's multi-concept answer record with multi-hot encoding, forming a concept encoding matrix x representing the answer historical record;

[0138] S3, calculating the accuracy of each student on different concepts and representing it with multi-hot encoding, forming a difficulty information matrix d representing personalized historical records;

[0139] S4, the concept coding matrix x and the difficulty information matrix d are spliced and sent to a recurrent neural network, an attention mechanism is introduced in the recurrent neural network, a correlation coefficient of a history mastery hidden state and a current time hidden state is calculated, and the current time hidden state is updated;

[0140] S5, according to a cross-entropy loss function, a recurrent neural network is trained by using a back propagation mechanism, model parameters are updated, a trained recurrent neural network is obtained, and student answering prediction is performed.

[0141] In addition, the logic instructions in the above-mentioned memory can be realized in the form of a software function unit and sold or used as an independent product, and can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the parts that contribute to the prior art or parts of the technical solutions can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in various embodiments of the present application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various program code storage media.

[0142] The embodiments of the present application also provide a non-transitory computer readable storage medium having a computer program stored thereon, which is executed by a processor to implement the method provided by the above-mentioned embodiments, for example, including:

[0143] S1, pre-processing a data set of student answering records, obtaining each student's answering questions and all concepts involved in the questions, and generating a multi-concept based answering record;

[0144] S2, representing each student's multi-concept answering record by using multi-hot coding to form a concept coding matrix x for representing the answering history record;

[0145] S3, calculating the correct rate of each student for different concepts and representing it by using multi-hot coding to form a difficulty information matrix d for representing the personalized history record;

[0146] S4, the concept coding matrix x and the difficulty information matrix d are spliced and sent to a recurrent neural network, an attention mechanism is introduced in the recurrent neural network, a correlation coefficient of a history mastery hidden state and a current time hidden state is calculated, and the current time hidden state is updated;

[0147] S5, according to the cross-entropy loss function, using the back propagation mechanism, the recurrent neural network is trained, the model parameters are updated, the trained recurrent neural network is obtained, and student answering prediction is performed.

[0148] The system embodiments described above are only illustrative, wherein the units described as separate components can or can not be physically separated, and the components displayed as units can or can not be physical units, i.e., can be located in one place, or can be distributed on multiple network units. Part or all of the modules can be selected to achieve the purpose of the embodiment according to actual needs. Those skilled in the art can understand and implement without creative labor.

[0149] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be realized by means of software and the necessary general hardware platform, and of course, it can also be realized by hardware. Based on such understanding, the above technical solutions can be embodied in the form of software products, and the computer software products can be stored in a computer readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including a plurality of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute the method described in each embodiment or some parts of the embodiment.

[0150] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A personalized knowledge tracking method based on multi-concept representation, characterized in that, include: S1. Preprocess the dataset of student answer records to obtain the answer questions for each student and all the concepts involved in the questions, and generate answer records based on multiple concepts; S2. Represent each student's multi-concept answer record using multi-hot encoding to form a concept encoding matrix x for representing the answer history. S3. Calculate the accuracy rate of each student for different concepts and represent it using multi-hot encoding to form a difficulty information matrix d for representing personalized historical records; S4. After concatenating the concept encoding matrix x and the difficulty information matrix d, feed them into the recurrent neural network. Introduce an attention mechanism into the recurrent neural network, calculate the correlation coefficient between the hidden state of historical mastery and the hidden state at the current moment, and update the hidden state at the current moment. S5. Based on the cross-entropy loss function, the recurrent neural network is trained using the backpropagation mechanism, the model parameters are updated, the trained recurrent neural network is obtained, and student answer prediction is performed. The generation of multi-concept-based answer records in S1 includes: according to the dictionary and student ID, converting each student's answer question and all the concepts involved in the question into a three-line format, wherein: First line , Where Q is the set of questions answered by a single student. Number of questions; Second line , Where C is the set of concepts involved in a single student's answer to a question. For the number of concepts; Third line ,in For the question Did you answer correctly? The concept encoding matrix x is represented as: （1） Where Batch is the batch size, and Maxstep represents the maximum step size. For the number of concepts, , , ; The difficulty information matrix d is represented as: （2） Where c represents the concept accuracy rate.

2. The personalized knowledge tracking method based on multi-concept representation according to claim 1, characterized in that, S1 involves obtaining each student's answer question and all the concepts involved in the question, including the following steps: S11. Read the dataset file, delete rows with NA or NULL values, and rows where the correct column is not a binary value; S12. Number the questions and concepts, find all the concept numbers corresponding to the question numbers, and generate a dictionary.

3. The personalized knowledge tracking method based on multi-concept representation according to claim 1, characterized in that, The accuracy of the concept Obtained using a three-row formula, the formula is: （3） in, It is the total number of times concept j appears in the second line. It represents the number of times concept j answers correctly in the third line.

4. The personalized knowledge tracking method based on multi-concept representation according to claim 1, characterized in that, In step S4, updating the hidden state at the current moment involves using the hidden state from the previous moment to update the hidden state at the current moment. The formula is as follows: （4） （5） in, For splicing, m t For the concatenated input, h t The current state is hidden, h t-1 The state that was hidden in the previous moment.

5. The personalized knowledge tracking method based on multi-concept representation according to claim 4, characterized in that, The hidden state h at the previous moment t-1 For comprehensive hidden state The comprehensive hidden state The steps to obtain it are as follows: S41, Utilization Activation function retrieves hidden representation The formula is: （6） Among them, h f Let f be the hidden state at time f, 1≤f≤t, and W and b be the parameters of the neural network model. S42, Utilization Function to obtain importance and relevance coefficients The formula is: （7） in, The hidden representation at time f, Let f be the correlation coefficient at time f; S43. Calculate the hidden state h at time f. f The correlation of hidden states is given by the formula: （8） S44. Sum the correlations of the hidden states before time t-2, and then concatenate them with the hidden states at time t-1 to obtain the comprehensive hidden state at time t-1. The formula is as follows: （9） （10） in, The state correlation is hidden at time f.

6. An electronic 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, it implements the steps of a personalized knowledge tracking method based on multi-concept representation as described in any one of claims 1 to 5.

7. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of a personalized knowledge tracking method based on multi-concept representation as described in any one of claims 1 to 5.

Images