CN115688863A - Depth knowledge tracking method based on residual connection and student near-condition feature fusion - Google Patents

Abstract

The invention relates to a deep knowledge tracking method based on fusion of residual connection and current situation features of students, and belongs to the technical field of knowledge tracking. The present invention improves the problems existing in the Deep Knowledge Tracking (DKT) model. The Deep Knowledge Tracking is to use the time-sequential knowledge point answering data of the learner and the relevant data of the learner's answer to the knowledge point correctly or not, and use the recurrent neural network (RNN), to predict the students' future answers. The invention adds the residual connection to the DKT model to solve the problem of network information degradation, and enhances the recent learning situation of the students through the recent learning data of the students, effectively improving the accuracy of the model prediction, and is a great contribution to the field of knowledge tracking. Development offers feasible solutions.

Description

A Deep Knowledge Tracking Method Based on Fusion of Residual Connections and Students' Current Situation Features

technical field

The invention relates to a deep knowledge tracking method based on fusion of residual connection and current situation characteristics of students, and belongs to the technical field of knowledge tracking.

Background technique

Knowledge tracking is one of the important practices of artificial intelligence in education. Its main research task is to establish a model to predict students' future performance based on the students' past learning data. In recent years, deep knowledge tracing (DKT) based on recurrent neural network (RNN) creation has been proposed that outperforms previously proposed knowledge tracing methods and traditional methods. The cyclic neural network has unexpected advantages and effects in predicting students' performance in real time, but after observation and experimentation, it is found that DKT still has a lot of room for improvement.

With the advent of the era of big data, the application of artificial intelligence in education is becoming more and more popular. The data generated by students in the learning process is stored in large quantities, and the ability of computers to process data is also greatly enhanced. Advances in educational data mining and educational data analysis have provided impetus for the development of learning prediction. In-depth knowledge tracking is to build an RNN-based model based on students' past learning data to predict students' future performance. RNN is very effective for data with sequential characteristics. It can mine the timing information and semantic information in the data. Using this ability of RNN, it can predict the future situation of the students from the past learning data. It provides assistance for teachers' personalized teaching, and at the same time, it also improves the learning efficiency of students.

Contents of the invention

The technical problem to be solved by the present invention is to provide a deep knowledge tracking method based on the fusion of residual connection and students' current situation features, which is used to solve the problem of ignoring the rationality of other knowledge points' predicted values in the existing IQ tracking method, thus aggravating the After understanding the error problem when predicting other knowledge points.

The technical solution of the present invention is: a deep knowledge tracking method based on the fusion of residual connection and current situation characteristics of students, one-hot encoding is performed on the required data as the input of the cyclic neural network, and each piece of data input to the cyclic neural network generates a For the prediction list, set a CUT_STEP value. Every time each student calculates CUT_STEP data in the cyclic neural network, multiply the previous CUT_STEP original data by weight and add them together, put them into the fully connected layer as input, and then put the fully connected layer The output of is added to the hidden layer as supplementary information to continue the operation until the initial prediction matrix of the input data is obtained. Each piece of data in the initial prediction matrix is spliced with the sum of the students' recent learning data, that is, the recent learning situation of the students is enhanced, and finally the prediction matrix is obtained through the linear layer. Read the required prediction value from the prediction matrix, then calculate the loss between the prediction value and the real value, and optimize the model according to the loss until the model reaches a relative optimum.

The specific steps are:

Step1: Data preprocessing, process each piece of raw data into one-hot data form, each piece of raw data is combined to form a one-hot matrix, the specific steps are as follows:

Step1.1: Data reading and cleaning, read the required features from the student's test record, and form a new data form data. The required features include time series id, student id, knowledge point id, and correctness of the question.

Step1.2: Calculate the types of knowledge points, and form a list of knowledge point ids K={k ₁ , k ₂ ,..., k _l }.

Step1.3: Form a dictionary according to the knowledge point list K, the dictionary form is {knowledge point id: the position of the knowledge point in the knowledge point list formed in Step1.2}, and the mathematical symbol is expressed as dict={k:e}, where k∈K, K is the list of knowledge points formed in Step1.2, e∈{0, 1, 2, ..., l-1}.

Step1.4: Extract the list of knowledge points of each student and the corresponding list of correct answers from the data to form a sequence of sequences.

Step1.5: Convert the sequences into one-hot encoded form, and process the data into a unified form, that is, each student has MAX_STEP pieces of data, and each piece of one-hot data corresponds to a question that the student has done.

Step2: Read the training data in batches from the data processed in Step1.

Step3: Use the cyclic neural network to predict the correctness of the student's next question, that is, put the data read in Step 2 into the cyclic neural network for calculation, and generate the original prediction data. Its formula is:

h _t ＝tanh(W _x x _t +b _x +W _h h _(t-1) +b _h ) (1)

In the formula, h _t is the hidden state at time t, x _t is the input at time t, and h _(t-1) is the hidden state at time t-1.

Step4: The original prediction data generated by the cyclic neural network is used to enhance the recent learning situation of the students through the recent learning data of the students, and make the final prediction. The specific steps are as follows.

If the student performed well in the previous problem-solving situations, but the recent problem-solving situation performed poorly, it is very likely that the student's recent learning situation has problems. And it will affect the current problem-making situation. Judge the student's recent learning situation through the student's past learning data. Through experiments, the effect is the best when splicing the sum of the students' recent learning data.

Step4.1: Splicing the original prediction data generated in Step3 and the sum of the student’s recent test data to form the experience data o _t , the formula is:

In the formula, o _t is the data after splicing the original forecast data and the sum of the student’s recent test data, h _t is the original forecast data, x _i is the one-hot test data of the student at time i, and n is the required The number of spliced students' recent test data.

Step4.2: Put the spliced data o _t of Step4.1 into the fully connected layer fc ₁ , and adjust the predicted dimension. The expression of fc ₁ is:

和

为线性层fc₁的输入和输出，

为权重，

为偏置。In the formula,

and

is the input and output of the linear layer fc ₁ ,

is the weight,

for the bias.

Step4.3: Put the output vector of Step4.2 into the activation function, control each element of the output between 0 and 1, and get the final prediction result. The numerical conversion is expressed as:

Step5: Add residual connections on the basis of the cyclic neural network described in Step3 to improve the prediction accuracy. The specific steps are:

On the basis of the Step3 cyclic neural network, the residual connection is added to improve the prediction accuracy. It is found that the prediction results in the cyclic neural network will be too pursuit of the results of the next knowledge point answering questions, while ignoring the rationality of the prediction values of other knowledge points, thus exacerbating the error when predicting other knowledge points in the future, so set a CUT_STEP( 1<CUT_STEP<MAX_STEP) value is used to segment the MAX_STEP of a student's one-hot matrix, and CUT_STEP is the segmentation value of segmented data aggregation, adding the segmented CUT_STEP data and adding them to the implicit Layer, enter the next cycle, the above is the residual connection, the specific steps to realize the residual connection are as follows:

Step5.1: Every time each student calculates CUT_STEP data in the cyclic neural network, multiply the previous CUT_STEP data by the weight W and add them together, and use the residual idea to add the previous CUT_STEP data according to their importance (weight) Get a vector containing the previous CUT_STEP data information, where the weight W={w ₁ , W ₂ ,...,W _C }, and C is the abbreviation of CUT_STEP.

Step5.2: Put the vector generated by Step5.1 containing the previous CUT_STEP data information into the fully connected layer fc ₂ as input, and convert the result of Step5.1 into a form that can be directly added to the hidden layer, fc ₂ The expression is:

和

为线性层fc₂的输入和输出，

为权重，

为偏置。In the formula,

and

is the input and output of the linear layer fc ₂ ,

is the weight,

for the bias.

的表达式为：Step5.3: Then add the output of fc ₂ to the hidden layer to continue the operation, then calculate CUT_STEP data in RNN, add the previous CUT_STEp data, repeat the above steps, after the hidden layer changes at time t

The expression is:

即{s_d(i-C+1)，...s_di}乘以W再求和。In the formula, C is the abbreviation of CUT_STEP, s _hd is the one-hot data of the dth student, W refers to the weight assigned to s _hd participating in the calculation,

That is {s _d(i-C+1) ,...s _di } multiplied by W and summed.

Step6: Read the true value of the student's correct answer and the predicted value of the model's predicted student's correct or incorrect answer. The specific steps are:

Since there is no clear label given in the data set, the scheme adopted is: the i+1th piece of data of the student is the label of the prediction list generated by the i-th piece of data of the student, because the prediction list generated by the i-th piece of data of the student is The next step for the student is to predict all knowledge points, and the student's i+1th piece of data is the one-hot code of whether the student is correct or not when doing a certain knowledge point, so it can be seen from here that the student's i+ 1 piece of data can be found from the prediction list generated by the i-th piece of data of the student. The specific steps are as follows:

Step6.1: From the input one-hot matrix, find the positions of all the knowledge points made by the student except the first one (that is, the zero position) to form a position list.

Step6.2: Find the corresponding predicted value from the predicted list according to the found position list.

Step6.3: From the input one-hot matrix, find the true value list consisting of all knowledge points except the first one (that is, the zero position) that the student has done.

Step7: Calculate the prediction error and optimize the network to obtain a model that can predict whether the student will do the next question correctly or not. The specific steps are:

Step7.1: According to the real value and predicted value read in Step6, use the loss function to calculate the loss value.

Step7.2: According to the loss value, use the optimizer to optimize the network.

Step7.3: Because optimization is not optimized once, if you want to define the size of Epoch and determine the number of optimizations according to your needs, you can repeat the above steps several times.

Step7.4: Obtain a trained model that can predict whether the student will do the next question correctly or not.

In the existing technology, for the knowledge point answers with time series and the student learning data corresponding to the correct answers, the recurrent neural network is used to predict the correctness of the students' next answer, because each time it is the result of all the knowledge points of the students in the next step. Forecasting, and students will only do one question in the next step, so it is found that the prediction result will be too pursuit of the result of answering the next knowledge point, while ignoring the rationality of the predicted value of other knowledge points, thus exacerbating the error when predicting other knowledge points in the future , in order to solve this problem, the present invention adds a residual connection on the basis of the recurrent neural network, and together completes the prediction of the student's next answer. At the same time, it was found that if the students performed well in the previous problems, but performed poorly in the recent problems, it is very likely that the students have problems in their recent learning. And it will affect the current problem-making situation. Therefore, the present invention enhances the recent learning situation of the students through the recent learning data of the students, and achieves good results. The present invention uses the residual connection to effectively solve the problem that the prediction results are too pursuit of the answer to the next knowledge point, while ignoring the rationality of the prediction values of other knowledge points, thereby exacerbating the problem of errors when predicting other knowledge points in the future.

The beneficial effects of the present invention are: the present invention improves the problems existing in the Deep Knowledge Tracking (DKT) model, and the Deep Knowledge Tracking is to use the relevant data of the learner's time-sequenced knowledge point answer questions and the learner's information on whether the knowledge point answer is correct or not. Relevant data, using recurrent neural network (RNN), to predict students' future answers. The invention adds the residual connection to the DKT model to solve the problem of network information degradation, and enhances the recent learning situation of the students through the recent learning data of the students, effectively improving the accuracy of the model prediction, and is a great contribution to the field of knowledge tracking. Development offers feasible solutions.

Description of drawings

Fig. 1 is a flow chart of steps of the present invention;

Fig. 2 is a schematic diagram of a model of the present invention.

Detailed ways

The present invention will be further described below in combination with the accompanying drawings and specific embodiments.

Example 1: As shown in Figure 2, this example takes the public dataset 2009-2010ASSISTment dataset as an example, processes the data of each student, and then puts it in the built model to train the model. The specific process includes: processing the data into a data form that can be input into the model, reading the training data in batches from the processed data, using the recurrent neural network to make a preliminary prediction of the correctness of the students' next question, and using the students' recent learning data to Enhance students' recent learning situation, make final predictions, add residual connections on the basis of recurrent neural networks to improve prediction accuracy, read real values and predicted values, calculate prediction errors and optimize the network, and get predictable students' next questions correct or not model.

As shown in Figure 1, a deep knowledge tracking method based on the fusion of residual connections and students' current situation features, the specific steps are:

Step1: Data preprocessing.

To process the raw data into data that can be directly input into the model, the specific steps are as follows:

Step1: Data preprocessing, processing each piece of raw data into one-hot data form, the specific steps are as follows:

Step1.1: Data reading and cleaning:

Read the required features from the student's record of doing questions to form a new data form data. The required features include the number in non-chronological order, the number of students, the number of knowledge points, and the four characteristics of correctness. Among them, the knowledge points are the knowledge points that students need to master. Each knowledge point may correspond to several questions to assist students to master This knowledge point, that is, each knowledge point may appear multiple times in a student's data.

Step1.2: Calculate the types of knowledge points, and form a list K={k ₁ , k ₂ ,...,k _l } composed of knowledge point ids, the length of the list is 123. The main function of the list K is to provide each The position of the knowledge point is determined, so that in the following one-hot matrix, which knowledge point can be determined only according to the 1 position in the one-hot matrix.

Step1.3: Form a dictionary according to the knowledge point list K formed in Step1.2, the dictionary form is {knowledge point number: the position of the knowledge point in the knowledge point list formed in Step1.2}, and the mathematical symbol is expressed as dict={ k:e}, where k∈K, e∈{0, 1, 2, ... 122}, K is the list of knowledge points generated in Step1.2, 122 is the length of the list of knowledge points in Step1.2 -1, Prepare for the formation of one-hot encoding in the future.

Step1.4: Extract the list of knowledge points of each student and the corresponding list of correct answers from the data to form a sequence sequence. The specific steps are as follows:

Step1.4.1: Extract all the data of each student from data S _r = {s _r1 , s _r2 ,..., s _rt }, S _r represents the unprocessed data of all students, which is the same as in Step1.6.4 The difference between Sh _h , where s _rt is a two-dimensional array of unequal length (the unequal length is because each student has a different number of questions to do), containing all the data of the tth student.

Step1.4.2: Extract the student's knowledge point list S _Ki = {q _ij |0≤i<b, j∈N ^* } and the corresponding list of correct answers S _Ai = from all the data of each student {a _ij |0≤i<b, j∈N ^* } (1 means correct, 0 means wrong), where S _K refers to the student's knowledge point list, i refers to the i-th student, that is, S _Ki refers to the i-th student The list of knowledge points of each student, q _ij represents the jth question done by the i-th student in chronological order, b is the length of the S list in Step1.4.1, that is, the number of students, and there are 4163 students in this data set. The range of j depends on the number of questions that the corresponding students do (because the number of questions for each student is different, so the range of j for each student is also different), S _A refers to the student's list of correct and incorrect answers, and i refers to The i-th student, that is, S _Ai refers to the correct and incorrect answer list of the i-th student.

Step1.4.3: The list of knowledge points of all students and the corresponding list of correct answers form a sequence sequences.

Step1.5: Convert the sequences into one-hot coded form, each one-hot data corresponds to a question for the students, the specific steps are as follows:

Step1.5.1: Read the list of knowledge points of each student from the sequences accordingly (the length of the list of knowledge points of the students is recorded as M) and the list of whether the answers are correct or not.

Step1.5.2: M is the length of the student's knowledge point list (each student has a different number of questions, so the value of M is also different), L is twice the number of all knowledge points, that is, L=246, set a MAX_STEP as needed is 50, if a student's data number M is an integer multiple of 50, then the student forms a M*246 zero matrix, otherwise:

If M%50=P

Then the student forms a zero matrix of M _ch *246, and the expression of M _ch is as follows:

M _ch =M+(50-P) (1)

The role of MAX_STEP is to unify the data shape of students. Since each student has a different number of questions, the length M of the list of students' knowledge points is also different, resulting in different one-hot matrices formed by students. The input model The shape of the data needs to be the same, so a MAX_STEP is set to correct the one-hot matrix formed by each student, and the length of each student's one-hot matrix is changed to an integer multiple of MAX_STEP, and the insufficient one-hot matrix is filled with zeros , in order to further unify the data into the input style required by the model.

Step1.5.3: Use the student knowledge point list S _Ki generated by Step1.4.2 q _ij in {q _ij |0≤i<b, j∈N ^* } (that is, the jth question done by the i-th student), Find the position s where _qij forms the knowledge point list in Step1.2 according to the dictionary dict formed in Step1.3,

If the knowledge point is correct in the corresponding Step1.4.2 answer list S _Ai , the sth position of the jth row of the i-th student’s zero matrix is recorded as 1, otherwise the student’s zero-matrix The s+l position of the jth row becomes 1, l is the length of the knowledge point list formed in Step1.2, and the above steps are repeated until the zero matrix of all students becomes the corresponding one-hot matrix.

Step1.5.4: Next, change each student's one-hot matrix M*246 (or M _ch *246) to X*50*246, and X is automatically generated according to MAX_STEP and M, that is, a student's one-hot data is based on MAX_STEP is divided into multiple blocks, which are regarded as one-hot data of multiple students, and the above steps are repeated. The data of all students according to MAX_STEP segmentation forms the final train_data. The data in train_data can be expressed as Sh ₌ {s _h1 , s _h2 , ..., s _hi }, the dimension of S _h is Y*50*246, where Y is the sum of all X in X*50*246, 1≤i≤Y, s _hi ={s _i1 , s _i2 ,. ..s _im }, where s _hi is the data of the jth one-hot student, and its dimension is 50*246, 1≤m≤50.

The above steps convert the data data into a one-hot matrix of Y*50*246.

Step2: Read the training data in batches from the data processed in Step1, set the batch size to 64, and decide how many one-hot student data to load at a time, and its dimension is 64*50*246.

Step3: Use the cyclic neural network to predict the correctness of the students' next question, and generate the original prediction data. The specific steps are:

Put the data read in Step 2 into the recurrent neural network for calculation. Since the data of the students' questions is time-sequential, that is, the questions the students did in the previous step will affect the correct rate of the students' next questions, so the choice has memory Build a model based on the RNN cyclic neural network with data sharing characteristics, set the input_size to 246, the hidden layer size (hidden_size) to 10, the number of cyclic layers to 1, select the activation function 'tanh', and the initial hidden layer parameters are random Generated, h _t (hidden state at time t) will be used as the output of the recurrent neural network at time t, and will enter the next cycle as the hidden state at time t, then the hidden state at time t is:

h _t ＝tanh(W _x x _t +b _x +W _h h _(t-1) +b _h ) (2)

In the formula, h _t is the hidden state at time t, x _t is the input at time t, and h _(t-1) is the hidden state at time t-1.

Step4: The original prediction data generated by the cyclic neural network is used to enhance the student's recent learning data through the student's recent learning data, and make the final prediction. The specific steps are:

If the student performed well in the previous problem-solving situations, but the recent problem-solving situation performed poorly, it is very likely that the student's recent learning situation has problems. And it will affect the current problem-making situation. Judge the student's recent learning situation through the student's past learning data. Through experiments, the effect is the best when splicing the sum of the students' recent learning data.

Step4.1: Splicing the original prediction data generated in Step3 and the sum of the student’s recent test data to form the experience data o _t , the formula is:

In the formula, o _t is the data after splicing the original forecast data and the sum of the student’s recent test data, h _t is the original forecast data, x _i is the one-hot test data of the student at time i, and n is the required The number of spliced students' recent test data.

Step4.2: Put the spliced data o _t of Step4.1 into the fully connected layer fc ₁ , and adjust the predicted dimension. The expression of fc ₁ is:

和

为线性层fc₁的输入和输出，

为权重，

为偏置。In the formula,

and

is the input and output of the linear layer fc ₁ ,

is the weight,

for the bias.

Step4.3: Put the output vector of Step4.2 into the activation function, control each element of the output between 0 and 1, and get the final prediction result. The numerical conversion is expressed as:

Step5: Add residual connection on the basis of Step3 cyclic neural network to improve prediction accuracy. The specific steps are:

On the basis of the Step3 cyclic neural network, the residual connection is added to improve the prediction accuracy. It is found that the prediction results in the cyclic neural network will pursue the results of the next knowledge point answer too much, while ignoring the rationality of the prediction values of other knowledge points, which will aggravate the error when predicting other knowledge points in the future, so set a CUT_STEP as 10. It is used to segment the MAX_STEP of a student's one-hot matrix. CUT_STEP is the segmentation value of segmented data aggregation. Add the 10 segmented data and add them to the hidden layer after transformation, and enter the next time cycle, the above is the residual connection, and the specific steps to realize the residual connection are as follows:

Step5.1: Every time each student calculates 10 data in the cyclic neural network, multiply the previous 10 data by the weight W and add them together (because the importance of each data is different), and use the residual idea to divide the previous 10 According to their importance (weight), add up a vector containing the information of the previous 10 data, where the weight W={w ₁ , W ₂ ,...,W _C }, the weight is given to the network to learn by itself, and C is Shorthand for CUT_STEP.

Step5.2: Put the vector generated by Step5.1 containing the previous 10 data information as input into the fully connected layer fc ₂ , the input size of fc ₂ is 246, and the output is the hidden layer size 10. The main function of fc ₂ is Convert the result of Step5.1 into a form that can be directly added to the hidden layer, the expression of fc ₂ is:

和

为线性层fc₂的输入和输出，

为权重，

为偏置。In the formula,

and

is the input and output of the linear layer fc ₂ ,

is the weight,

for the bias.

的表达式为：Step5.3: Then add the output of fc ₂ to the hidden layer to continue the operation, then calculate 10 data in the RNN, add the previous 10 data, repeat the above steps, after the hidden layer changes at time t

The expression is:

即{s_d(i-C+1)，...s_di}乘以W再求和。In the formula, C is the abbreviation of CUT_STEP, _Shd is the one-hot data of the dth student, W refers to the weight assigned to the _shd involved in the calculation,

That is {s _d(i-C+1) ,...s _di } multiplied by W and summed.

Step6: Read the true value of the student's correct answer and the predicted value of the model's predicted student's correct or incorrect answer. The specific steps are:

Since there is no clear label given in the data set, the scheme adopted is: the i+1th piece of data of the student is the label of the prediction list generated by the i-th piece of data of the student, because the prediction list generated by the i-th piece of data of the student is The next step for the student is to predict all knowledge points, and the student's i+1th piece of data is the one-hot code of whether the student is correct or not when doing a certain knowledge point, so it can be seen from here that the student's i+ 1 piece of data can be found from the prediction list generated by the i-th piece of data of the student. The specific steps are as follows:

Step6.1: From the input one-hot matrix, find the positions of all knowledge points made by the student except the first one (that is, the zero position) to form a position list. The reason for removing the first one is: from the first piece of data Generate a prediction for the second item, the second item generates a prediction for the third item, and so on, obviously, there is no prediction for the first item of data, that is, there is no way to compare its real value and predicted value, and calculate the error Loss, so remove it.

Step6.2: According to the location list found in Step6.1, find out the corresponding prediction value from the prediction list (because each prediction is the prediction of the correct probability of the student doing all the knowledge points next time, but the next time the student What is done is a knowledge point, so only the predicted value of this knowledge point is needed, and the others are discarded) to form a prediction list.

Step6.3: From the input one-hot matrix, find the true value list consisting of all knowledge points except the first one (that is, the zero position) that the student has done.

Step7: Calculate the prediction error and optimize the network to obtain a model that can predict whether the student will do the next question correctly or not. The specific steps are:

Step7.1: According to the real value and predicted value read in Step6, use the loss function to calculate the loss value.

Step7.2: According to the loss value, use the optimizer to optimize the network.

Step7.3: Define the size of Epoch as 70 according to the needs, decide how many rounds to train, and repeat the above steps several times.

Step7.4: Obtain a trained model that can predict whether the student will do the next question correctly or not.

The specific embodiments of the present invention have been described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the above embodiments. Variations.

Claims (7)

1. A deep knowledge tracking method based on residual connection and student current situation feature fusion, characterized in that: Step1: Data preprocessing, processing each piece of raw data into one-hot data form; Step2: Read the training data in batches from the data processed in Step1; Step3: Use the cyclic neural network to predict the correctness of the students' next question, and generate the original prediction data; Step4: Circulate the original prediction data, use the students' recent learning data to enhance the students' recent learning situation, and make the final prediction; Step5: Add residual connections on the basis of the cyclic neural network described in Step3; Step6: Read the true value of the student's correct answer and the predicted value of the model's predicted student's correct answer; Step7. Calculate the prediction error and optimize the network to obtain a model that can predict whether students will be correct or not next time. 2. the depth knowledge tracking method based on residual connection and student current situation feature fusion according to claim 1, characterized in that, said Step1 is specifically: Step1.1: Data reading and cleaning, read the required features from the student's record of doing questions, and form a new data form data. The required features include time series id, student id, knowledge point id, and correctness and error of the question; Step1.2: Calculate the types of knowledge points and form a list of knowledge point ids K={k ₁ , k ₂ ,...,k _l }; Step1.3: Form a dictionary according to the knowledge point list K, the dictionary form is {knowledge point id: the position of the knowledge point in the knowledge point list formed in Step1.2}, and the mathematical symbol is expressed as dict={k:e}, where k∈K, K is the list of knowledge points formed in Step1.2, e∈{0, 1, 2, ..., l-1}; Step1.4: Extract the list of knowledge points of each student and the corresponding list of correct answers from the data to form a sequence of sequences; Step1.5: Convert the sequences into one-hot encoded form, and process the data into a unified form, that is, each student has MAX_STEP pieces of data, and each piece of one-hot data corresponds to a question that the student has done. 3. the depth knowledge tracking method based on residual connection and student's current situation feature fusion according to claim 1, characterized in that, said Step3 is specifically: Put the data read in Step2 into the cyclic neural network for calculation to generate the original forecast data, specifically: h _t ＝tanh(W _x x _t +b _x +W _h h _(t-1) +b _h ) (1) In the formula, h _t is the hidden state at time t, x _t is the input at time t, and h _(t-1) is the hidden state at time t-1. 4. the depth knowledge tracking method based on the fusion of residual connection and student's recent situation features according to claim 1, wherein, said Step4 is specifically: Step4.1: Splicing the original prediction data generated in Step3 and the sum of the student’s recent test data to form the experience data o _t , the formula is:

In the formula, o _t is the data after splicing the original forecast data and the sum of the student’s recent test data, h _t is the original forecast data, x _i is the one-hot test data of the student at time i, and n is the required The number of spliced students' recent test data; Step4.2: Put the spliced data o _t of Step4.1 into the fully connected layer fc ₁ , and adjust the predicted dimension. The expression of fc ₁ is:

In the formula, o _fc1 and y _fc1 are the input and output of the linear layer _fc1 , A _fc1 is the weight, b _fc1 is the bias; Step4.3: Put the output vector of Step4.2 into the activation function, control each element of the output between 0 and 1, and get the final prediction result. The numerical conversion is expressed as:

5. the depth knowledge tracking method based on residual connection and student's current situation feature fusion according to claim 1, is characterized in that, described Step5 is specifically: Step5.1: Every time each student calculates CUT_STEP data in the cyclic neural network, multiply the previous CUT_STEP data by the weight W and add them together, and use the residual idea to add the previous CUT_STEP data according to their importance to get a A vector containing the previous CUT_STEP data information, where the weight W={w ₁ , w ₂ ,...,w _C }, C is the abbreviation of CUT_STEP; Step5.2: Put the vector generated by Step5.1 containing the previous CUT_STEP data information into the fully connected layer fc ₂ as input, and convert the result of Step5.1 into a form that can be directly added to the hidden layer, fc ₂ The expression is:

In the formula, x _fc2 and y _fc2 are the input and output of the linear layer _fc2 , A _fc2 is the weight, b _fc2 is the bias; Step5.3: Then add the output of fc ₂ to the hidden layer to continue the operation, then calculate CUT_STEP data in RNN, add the previous CUT_STEP data, repeat the above steps, after the hidden layer changes at time t

The expression is:

In the formula, C is the abbreviation of CUT_STEP, s _hd is the one-hot data of the dth student, W refers to the weight assigned to s _hd participating in the calculation,

That is {s _d(i-C+1) ,...s _di } multiplied by W and summed. 6. the depth knowledge tracking method based on residual connection and student current situation feature fusion according to claim 1, characterized in that, said Step6 is specifically: Step6.1: From the input one-hot matrix, find the positions of all knowledge points made by the student except the first one to form a list of positions; Step6.2: Find the corresponding predicted value from the predicted list according to the found position list; Step6.3: From the input one-hot matrix, find the true value list consisting of all knowledge points except the first one that the student has done correctly or not. 7. the depth knowledge tracking method based on residual connection and student's current situation feature fusion according to claim 1, wherein, said Step7 is specifically: Step7.1: According to the actual value and predicted value read in Step6, use the loss function to calculate the loss value; Step7.2: According to the loss value, use the optimizer to optimize the network; Step7.3: Define the size of Epoch according to the needs, determine the number of optimizations, and repeat the above steps several times; Step7.4: Obtain a trained model that can predict whether the student will do the next question correctly or not.

Images