Personalized knowledge tracking method based on multi-concept representation

Abstract

The invention discloses a personalized knowledge tracking method based on multi-concept representation, and the method comprises the steps: collecting a student answer record data set, carrying out the preprocessing, generating a multi-concept answer record, forming a concept coding matrix x used for representing an answer historical record and a difficulty information matrix d of a personalized historical record based on the multi-concept answer record, the concept coding matrix x and the difficulty information matrix d are spliced and then sent to a recurrent neural network, and student answer prediction is carried out; according to the method, multiple concepts related to the problem are represented by multiple thermal codes to enrich input characteristics, and meanwhile, the correct rate of different students about each concept is considered as a personalized basis, so that personalized historical record representation of each student is realized, and personalized knowledge tracking of different students is also realized while the model prediction performance is remarkably improved.

Description

Technical Field

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

[0002] Knowledge tracking is a key technology in the field of smart education, which aims to predict students' future performance based on their historical answer records. This technology has great practical significance for online education platforms. The platform can evaluate students' knowledge mastery through knowledge tracking, make learning plans for the next stage based on their knowledge level, and realize personalized learning. In recent years, due to the powerful computing and inference capabilities of deep learning algorithms, knowledge tracking models based on deep learning have emerged one after another.

[0003] In real life, every problem involves multiple knowledge points to a greater or lesser extent, so the problem should be a comprehensive representation of multiple knowledge points. However, existing conventional knowledge tracking methods often use a single concept instead of a problem for model training, which will cause a certain degree of input feature missing problems.

[0004] In real life, due to individual differences among students, even if the order of answering questions is the same, there should be different answer results. However, existing conventional knowledge tracking models treat students as a group to predict their knowledge mastery level, which deviates from the actual reality and fails to effectively conduct personalized learning tracking.

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

[0006] The patent document with application number 202110928810.3 discloses a personalized knowledge tracking method and system that integrates learning behavior features, and uses a convolutional neural network to extract effective features in a composite vector composed of learning behavior and answer result data; and extracts question information features including knowledge points through a denoising autoencoder, and finally combines the learning behavior features with the question information features, and obtains the student's knowledge mastery status through an LSTM network and a fully connected layer. In the modeling process, the present invention integrates a series of learning behavior characteristics of students in the learning process and rich information about the exercises themselves and knowledge points to more accurately predict the knowledge mastery of each student. However, its solution also has the following problems: 1. Failure to fully consider historical information to predict students' knowledge level will result in poor prediction results; 2. Insufficient use of personalized historical records.

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

[0008] In view of the above-mentioned problems, the purpose of the present invention is to provide a personalized knowledge tracking method based on multi-concept representation, which enriches the input features of multiple concepts involved in the problem, considers the personalized basis of different students, combines historical records and attention mechanism, and accurately predicts the degree of students' knowledge mastery.

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

[0010] Aspect 1: A personalized knowledge tracking method based on multi-concept representation, comprising:

[0011] S1. Preprocess the data set of student answer records to obtain each student's answer question and all the concepts involved in the question, and generate answer records based on multiple concepts;

[0012] S2, represent 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, calculate the accuracy of each student for different concepts and represent them using multi-hot encoding to form a difficulty information matrix d for representing personalized historical records;

[0014] S4, concatenate the concept encoding matrix x and the difficulty information matrix d and send them to the recurrent neural network, introduce the attention mechanism into the recurrent neural network, calculate the correlation coefficient between the hidden state of the historical mastery degree and the hidden state at the current moment, and update the hidden state at the current moment;

[0015] S5. Based on the cross entropy loss function, use the back propagation mechanism to train the recurrent neural network, update the model parameters, obtain the trained recurrent neural network, and predict the students' answers.

[0016] Optionally, obtaining each student's answer question and all concepts involved in the question in S1 includes the following steps:

[0017] S11, read the data set file, delete the rows with NA or NULL values ​​in the data set, and the rows with non-binary values ​​in the correct column;

[0018] S12. Number the questions and concepts, find out all the concept numbers corresponding to the question numbers, and generate a dictionary.

[0019] Optionally, generating a multi-concept answer record in S1 includes: converting each student's answer question and all concepts involved in the question into three lines according to the dictionary and student ID, where:

[0020] The first line is {q1,q2,...,q i ,...,q n}∈Q, Where Q is the set of questions answered by a single student, N q is the number of questions;

[0021] The second line is {c1|c3,c2|c3,...,c i |c j ,,...,c i |c j ,}∈C, Where c is the set of concepts involved in answering questions by a single student, N s is the conceptual number;

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

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

[0024]

[0025] Among them, batch is the batch size, Maxstep is the maximum step size, N S is the number of concepts, i∈[0, Batch-1], j∈[0, Maxstep-1], k∈[0, 2*Ns-1].

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

[0027]

[0028] Among them, c is the concept accuracy rate.

[0029] Optionally, the concept accuracy c is obtained by a three-line formula, and the formula is:

[0030]

[0031] Among them, n j is the total number of times concept j appears in the second row, is the number of correct answers given to concept j in the third row.

[0032] Optionally, in S4, the current hidden state is updated by using the hidden state at the previous moment to update the current hidden state, and the formula is:

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

[0034] h t ←LSTM(m t ,h t-1 )(5)

[0035] Among them, m t is the input after splicing, h t is the hidden state at the current moment, h t-1 is the hidden state at the previous moment. Optionally, the hidden state at the previous moment is a comprehensive hidden state The combined hidden state The steps to obtain are:

[0036] S41. Use tanh activation function to obtain hidden representation u i , the formula is:

[0037] u i =tanh(Wh i +b) (6)

[0038] Among them, h i is the hidden state at time i, 1≤i≤t, W and b are the neural network model parameters;

[0039] S42. Use the softmax function to obtain the importance correlation coefficient w i , the formula is:

[0040] w i =softmax(u i ) (7)

[0041] Among them, u i is the hidden representation at time i, w i is the correlation coefficient at time i;

[0042] S43, calculate the hidden state h at time i i The hidden state correlation of is:

[0043]

[0044] S44, sum the hidden state correlation before time t-2, and then concatenate it with the hidden state at time t-1 to serve as the comprehensive hidden state at time t-1. The formula is:

[0045]

[0046] in, is the hidden state correlation at time i.

[0047] A second aspect: An electronic device comprises 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 in the first aspect when executing the program.

[0048] A third aspect: A non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method provided in the first aspect.

[0049] Beneficial effects of the present invention:

[0050] 1. The present invention enriches the input features by using multi-hot encoding to represent multiple concepts involved in the problem, and considers the accuracy of different students on each concept as a personalization basis, thereby realizing personalized historical record representation for each student, significantly improving the model prediction performance, and also realizing personalized knowledge tracking for different students.

[0051] 2. The present invention designs a self-attention mechanism to calculate the correlation coefficient, i.e., the importance coefficient, of the historical prior knowledge mastery level, and after weighted summation, comprehensively represents the final knowledge mastery level. It combines historical records with the attention mechanism to simulate the accumulation of learning and improve the ability to predict students' knowledge mastery level. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a flowchart of the personalized knowledge tracking method of the present invention;

[0053] Figure 2 It is a schematic diagram of the principle flow of the personalized knowledge tracking method of the present invention;

[0054] Figure 3 Provides an example diagram of a knowledge tracking task for an embodiment of the present invention;

[0055] Figure 4 A visualization diagram for predicting the students' mastery of multiple concepts and knowledge for the present invention;

[0056] Figure 5 It is a structural schematic diagram of the personalized knowledge tracking system of the present invention;

[0057] Figure 6 It is a schematic diagram of the structure of the electronic device of the present invention. DETAILED DESCRIPTION

[0058] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar symbols throughout represent the same or similar elements or elements with the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0059] The existing conventional knowledge tracking methods have the following problems: single concepts are often used instead of questions for model training, which will cause a certain degree of input feature missing problems; individual differences between students make it impossible to effectively track personalized learning; and historical information is not fully considered to predict students' knowledge level, which will result in poor prediction results.

[0060] In view of the above problems, the present invention provides a personalized knowledge tracking method based on multi-concept representation. Figure 1 A flowchart of a personalized knowledge tracking method based on multi-concept representation provided by an embodiment of the present invention.

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

[0062] Batch Size: The number of samples used in each training iteration;

[0063] Multi-Hot: used to represent the situation where multiple categories are 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: The hidden state is the internal state of the RNN used to store and transmit information when processing an input sequence. It is updated at each time step to capture the information and contextual relationships in the input sequence and is used to represent the student’s knowledge mastery.

[0065] like Figure 2 As shown, the method of the present invention comprises:

[0066] S1. Preprocess the data set of student answer records to obtain each student's answer question and all the concepts involved in the question, and generate answer records based on multiple concepts.

[0067] The data set of student answer records is generally in the form of a csv file. The data set csv file is read and the obtained data set is preprocessed, including deleting rows with NA or NULL values ​​in the data set, and rows whose correct columns are not binary values ​​(1 or 0).

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

[0069] Then, according to the dictionary and student ID, each student's answer question and all the concepts involved in the question are converted into three lines, where:

[0070] The first line shows how many questions this student answered, expressed 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 contains multiple concepts corresponding to each question, expressed as:

[0074]

[0075] Where c is the set of concepts involved in answering questions by a single student, N c It is a conceptual number.

[0076] The third line indicates whether each question is answered correctly, and the formula is:

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

[0078] where r i For question q i Is the answer correct?

[0079] S2. Represent each student's multi-concept answer records with multi-hot encoding to form a concept encoding matrix x for representing the answer history records.

[0080] In order to distinguish whether each student's answer record is correct or not, it can be represented by the concept encoding position. The concept encoding matrix x can be expressed as:

[0081]

[0082] Among them, batch is the batch size, Maxstep is the maximum step size, N S is the number of concepts, i∈[0, Batch-1], j∈[0, Maxstep-1], k∈[0, 2*Ns-1].

[0083] For example, if batch size is 32 and Maxstep is 50, read the three-line formula, where Batch represents the batch size and Maxstep represents the maximum step length. That is, each Maxstep is regarded as a group of answer records, and a total of batch size groups are taken for training.

[0084] In this concept encoding matrix x, if the student answers the first question involving concepts i and j correctly, the i-th and j-th bits of the corresponding first row on the third dimension of x are both 1, that is:

[0085]

[0086] If the answer is wrong, the i+Ns and j+Ns positions are 1, that is:

[0087]

[0088] The position values ​​of other concepts not involved are set to 0, and the remaining rows are represented in the same way, finally forming the concept encoding matrix x of the multi-concept answer history record.

[0089] S3. Calculate the accuracy of each student for different concepts and represent them using multi-hot encoding to form a difficulty information matrix d for representing personalized historical records.

[0090] In order to realize the personalized knowledge tracking of students, the students’ answering level is used as auxiliary information to track and form the difficulty information matrix d used to represent the personalized historical records. The difficulty information matrix d can be expressed as:

[0091]

[0092] Among them, batch is the batch size, Maxstep is the maximum step size, N S is the number of concepts, i∈[0, Batch-1], j∈[0, Maxstep-1], k∈[0, 2*Ns-1], and c is the concept accuracy.

[0093] For example, if batch size is 32 and Maxstep is 50, read the three-line formula, where Batch represents the batch size and Maxstep represents the maximum step length. That is, each Maxstep is regarded as a group of answer records, and a total of batch size groups are taken for training.

[0094] c is the concept accuracy, which is obtained by calculating the accuracy of each student for each concept according to the three-line formula as auxiliary information of their learning level, traversing all concepts in the second line, and counting the number, and calculating the historical accuracy through the corresponding answers in the third line, that is:

[0095]

[0096] Among them, n j is the total number of times concept j appears in the second row, is the number of correct answers given to concept j in the third row.

[0097] In the difficulty information matrix d, if the student answers the first question involving concepts i and j correctly, the i-th position of the first row on the third dimension of x is the correct rate c of concept i. i and the jth position is the accuracy rate c of concept j j ,Right now:

[0098]

[0099] If the answer is wrong, the correct rate values ​​at positions i+Ns and j+Ns are respectively:

[0100]

[0101] The position values ​​of other concepts not involved are set to 0, and the remaining rows are represented in the same way, finally forming the difficulty information matrix d representing the personalized historical records.

[0102] S4. Concatenate the concept encoding matrix x and the difficulty information matrix d and send them into the recurrent neural network. Introduce the attention mechanism into the recurrent neural network, calculate the correlation coefficient between the hidden state of the historical mastery level and the hidden state at the current moment, and update the hidden state at the current moment.

[0103] Integrate multi-concept historical records and personalized historical records, concatenate the multi-concept encoding matrix x and the difficulty information matrix d as the input of the recurrent neural network (LSTM), and use the hidden state of the previous moment to update the hidden state coefficient h at the current moment t , the formula is:

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

[0105] h t ←LSTM(m t ,h t-1 )(5)

[0106] Among them, m t is the input after splicing, h t is the hidden state at the current moment, h t-1 It is the hidden state at the previous moment;

[0107] The hidden state h is the intermediate value generated by the multi-layer iteration of the recurrent neural network. In the field of knowledge tracking, it is also considered to be the student's mastery of each knowledge point.

[0108] The multi-concept historical record concept encoding matrix x and the corresponding personalized historical record difficulty information matrix d are concatenated and input into the neural network, and the generated hidden state contains the information of multiple concepts and personalized difficulty.

[0109] Considering the importance of historical hidden states, the inheritance and continuity of the LSTM learning process are maintained, and the hidden state of the previous moment is used to update the hidden state coefficient h at the current moment. t , specifically:

[0110] In order to calculate the correlation coefficient of the hidden state at different times, the hidden state is input into a multi-layer perceptron to enhance the complex linear mapping of the model data and obtain the comprehensive hidden state As the hidden state h at the previous moment t-1 , and then use formula (5) to get the current hidden state h t , specifically:

[0111] Use tanh activation function to generate hidden representation u i , in order to enhance the complex nonlinear mapping of the model data, and then generate the importance coefficient, i.e., the weight w, through the softmax function i , the formula is:

[0112] u i =tanh(Wh i +b) (6)

[0113] w i =softmax(u i ) (7)

[0114] Among them, h i is the hidden state at time i, 1≤i≤t, W and b are LSTM model parameters, are the LSTM learnable weights, is the LSTM bias; u i is the hidden representation at time i, w i is the correlation coefficient at time i.

[0115] Since the hidden states at all times have different degrees of influence on the knowledge state, their total influence is defined as the weighted sum of the hidden states at all times. In order to highlight the most recent knowledge state, that is, the hidden state at time t-1, the hidden states before time t-1 and time t-2 are concatenated together. The formula is:

[0116]

[0117]

[0118] in, is the hidden state correlation of the hidden state at time i, is the weighted sum value at time t-2; The comprehensive hidden state at time t-1 is obtained by concatenating the hidden states before time t-1 and time t-2

[0119] The final comprehensive hidden state As the input of the model prediction, use formula (5) to update the current hidden state h t，This approach can not only consider the impact of previous moments on the current moment, but also take into account the importance of the previous t-1 moment.

[0120] According to the students' knowledge mastery, that is, the hidden state h of the model t , using LSTM to predict the student's next answer performance, by combining the comprehensive hidden state After a linear layer, it is converted into a fixed output dimension, and then the prediction value is generated through the softmax activation function The formula is:

[0121]

[0122] S5. Based on the cross entropy loss function, use the back propagation mechanism to train the recurrent neural network, update the model parameters, obtain the trained recurrent neural network, and predict the students' answers.

[0123] The sum of the cross entropy losses is calculated based on the predicted student answer performance and the true label. The back propagation mechanism is used to calculate the gradient of the loss to the model parameters, and the model parameters are updated in the opposite direction of the gradient. The cross entropy loss function formula is:

[0124]

[0125] like Figure 3 As shown, for example: three students did four questions at time t4 before, and each question involved a certain number of concepts. For example, the first question of student L1 involved concepts c1 and c2. The present invention predicts the subsequent performance of the students based on their previous answer records, that is, based on the questions or concept information of student L1 at time t4 before, the present invention predicts the student's answer situation at time t5 and the degree of mastery of concept c3.

[0126] like Figure 4 As shown, Figure 4 The above dotted line is a personalized visualization result represented by multiple concepts, while the below dotted line is a visualization result based on a single concept. The darker the color, the better the student's mastery. The green and blue circles represent two different concepts. The red edge indicates the correct answer, and the black edge indicates the wrong answer.

[0127] The present invention also discloses a personalized knowledge tracking system based on multi-concept representation for the above method, such as Figure 5 As shown, the system includes:

[0128] Input module, used to input student answer history data set;

[0129] The preprocessing module preprocesses the data set of students’ answer records to obtain answer records of multiple concepts;

[0130] The concept coding matrix x module obtains the concept coding matrix x of the answer history records based on the answer records of multiple concepts;

[0131] A difficulty information matrix d module obtains a difficulty information matrix d for representing personalized historical records based on the answer records of multiple concepts;

[0132] The recurrent neural network module predicts students’ answers based on the concatenated concept encoding matrix x and the difficulty information matrix d;

[0133] Output module, outputs the predicted results of students’ answers.

[0134] The system of the present invention collects a data set of students' historical answer records, pre-processes them, obtains each student's answer questions and all the concepts involved in the questions, generates a multi-concept-based answer record dictionary, and then represents each student's multi-concept answer records with multi-hot encoding to form a concept coding matrix x for representing the answer history records, as well as each student's accuracy rate for different concepts to form a difficulty information matrix d for representing personalized historical records.

[0135] The obtained concept encoding matrix x and difficulty information matrix d are concatenated and sent to the recurrent neural network to predict students' answers and output the predicted results. The input features are enriched by using multi-hot encoding to represent multiple concepts involved in the question, and the correct rate of each concept of different students is calculated as a personalized basis to achieve personalized historical record representation for each student. In addition, the historical record is combined with the attention mechanism to simulate the accumulation of learning and improve the ability to predict the degree of students' knowledge mastery.

[0136] The present invention also provides an electronic device, Figure 6 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention, such as Figure 6 As shown, the electronic device may include: a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other via the communication bus. The processor may call the logic instructions in the memory, for example, to execute the following method:

[0137] S1. Preprocess the data set of student answer records to obtain each student's answer question and all the concepts involved in the question, and generate answer records based on multiple concepts;

[0138] S2, represent each student's multi-concept answer record with multi-hot encoding to form a concept encoding matrix x for representing the answer history record;

[0139] S3, calculate the accuracy of each student for different concepts and represent them using multi-hot encoding to form a difficulty information matrix d for representing personalized historical records;

[0140] S4, concatenate the concept encoding matrix x and the difficulty information matrix d and send them to the recurrent neural network, introduce the attention mechanism into the recurrent neural network, calculate the correlation coefficient between the hidden state of the historical mastery degree and the hidden state at the current moment, and update the hidden state at the current moment;

[0141] S5. Based on the cross entropy loss function, use the back propagation mechanism to train the recurrent neural network, update the model parameters, obtain the trained recurrent neural network, and predict the students' answers.

[0142] In addition, the logic instructions in the above-mentioned memory can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0143] An embodiment of the present invention further provides a non-transitory computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the method provided in each of the above embodiments is implemented, for example, including:

[0144] S1. Preprocess the data set of student answer records to obtain each student's answer question and all the concepts involved in the question, and generate answer records based on multiple concepts;

[0145] S2, represent each student's multi-concept answer record with multi-hot encoding to form a concept encoding matrix x for representing the answer history record;

[0146] S3, calculate the accuracy of each student for different concepts and represent them using multi-hot encoding to form a difficulty information matrix d for representing personalized historical records;

[0147] S4, concatenate the concept encoding matrix x and the difficulty information matrix d and send them to the recurrent neural network, introduce the attention mechanism into the recurrent neural network, calculate the correlation coefficient between the hidden state of the historical mastery degree and the hidden state at the current moment, and update the hidden state at the current moment;

[0148] S5. Based on the cross entropy loss function, use the back propagation mechanism to train the recurrent neural network, update the model parameters, obtain the trained recurrent neural network, and predict the students' answers.

[0149] The system embodiment described above is merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art may understand and implement it without creative work.

[0150] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0151] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A personalized knowledge tracking method based on multi-concept representation, characterized in that: include: S1. Preprocess the data set of student answer records to obtain each student's answer question and all the concepts involved in the question, and generate answer records based on multiple concepts; S2, represent each student's multi-concept answer record with multi-hot encoding to form a concept encoding matrix x for representing the answer history record; S3, calculate the accuracy of each student for different concepts and represent them using multi-hot encoding to form a difficulty information matrix d for representing personalized historical records; S4, concatenate the concept encoding matrix x and the difficulty information matrix d and send them to the recurrent neural network, introduce the attention mechanism into the recurrent neural network, calculate the correlation coefficient between the hidden state of the historical mastery degree 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, use the back propagation mechanism to train the recurrent neural network, update the model parameters, obtain the trained recurrent neural network, and predict the students' answers.

2. According to claim 1, a personalized knowledge tracking method based on multi-concept representation is characterized in that: The step S1 of obtaining each student's answer question and all the concepts involved in the question includes the following steps: S11, read the data set file, delete the rows with NA or NULL values ​​in the data set, and the rows with non-binary values ​​in the correct column; S12. Number the questions and concepts, find out 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 2 is characterized in that: The generation of the answer record based on multiple concepts in S1 includes: converting each student's answer question and all the concepts involved in the question into three lines according to the dictionary and student ID, where: The first line is {q1,q2,…,q i ,…,q n }∈Q, Where Q is the set of questions answered by a single student, N q is the number of questions; The second line is {c1|c3,c2|c3,…,c i |c j ,,...,c i |c j ,}∈C, Where c is the set of concepts involved in answering questions by a single student, N s is the conceptual number; The third line is {r1,r2,...,r i ,...,r n },r i ∈{0,1}, where ri is whether the question qi is answered correctly.

4. The personalized knowledge tracking method based on multi-concept representation according to claim 3 is characterized in that: The concept encoding matrix x is expressed as: x∈R Batch×Maxstep×(2*Ns) ,x[i,j,k]∈{0,1} (1) where batch is the batch size, Maxstep is the maximum step size, and N S is the number of concepts, i∈[0, Batch-1], j∈[0, Maxstep-1], k∈[0, 2*Ns-1].

5. The personalized knowledge tracking method based on multi-concept representation according to claim 3 is characterized in that: The difficulty information matrix d is expressed as: d∈R Batch×Maxstep×(2*Ns) ,d[i,j,k]=c (2) Among them, c is the concept accuracy rate.

6. The personalized knowledge tracking method based on multi-concept representation according to claim 5 is characterized in that: The concept accuracy c is obtained by a three-line formula, and the formula is: Among them, n j is the total number of times concept j appears in the second row, is the number of correct answers given to concept j in the third row.

7. The personalized knowledge tracking method based on multi-concept representation according to claim 1 is characterized in that: In S4, the current hidden state is updated by using the hidden state at the previous moment to update the current hidden state. The formula is: m t =cat(x t ,d t ) (4) h t ←LSTM(m t ,h t-1 ) (5) Where m t is the input after splicing, h t is the hidden state at the current moment, h t-1 It is the hidden state at the previous moment.

8. The personalized knowledge tracking method based on multi-concept representation according to claim 7 is characterized in that: The previous hidden state h t-1 is the comprehensive hidden state The combined hidden state The steps to obtain are: S41. Use tanh activation function to obtain hidden representation u i , the formula is: u i =tanh(Wh i +b) (6) Among them, h i is the hidden state at time i, 1≤i≤t, W and b are the neural network model parameters; S42. Use the softmax function to obtain the importance correlation coefficient w i , the formula is: w i =softmax(u i ) (7) Among them, u i is the hidden representation at time i, w i is the correlation coefficient at time i; S43, calculate the hidden state h at time i i The hidden state correlation of is: S44, sum the hidden state correlation before time t-2, and then concatenate it with the hidden state at time t-1 to serve as the comprehensive hidden state at time t-1. The formula is: in, is the hidden state correlation at time i.

9. 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, the steps of a personalized knowledge tracking method based on multi-concept representation as described in any one of claims 1 to 8 are implemented.

10. 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, the steps of a personalized knowledge tracking method based on multi-concept representation as described in any one of claims 1 to 8 are implemented.

Images