Performance prediction method, processor, readable storage medium and computer device based on PCANet-BiGRU
Through the PCANet-BiGRU-combined grade prediction method, the contradiction between efficiency and accuracy of the statistical school and deep learning methods in the existing technology is solved, and efficient and accurate grade prediction is achieved, which can promptly warn students with low grades. Improve academic performance.
Patent Information
- Application Number
- CN202110800902.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-07-15
- Publication Date
- 2024-11-01
- Estimated Expiration
- 2041-07-15
AI Technical Summary
Among the existing achievement prediction methods, the statistical school method has low calculation consumption but low accuracy. Although the deep learning method has high accuracy, the calculation complexity is high, and it is difficult to find a balance between efficiency and accuracy.
Using the PCANet-BiGRU-based grade prediction method, students' basic statistical data and grade data are collected from the online learning platform, data cleaning and normalization are carried out, and features are extracted by using principal component analysis network PCANet, combined with a bidirectional gated loop unit The neural network Bi-GRU performs performance predictions, reducing computational complexity and improving accuracy.
It has achieved the improvement of the accuracy of grade prediction while ensuring computing efficiency, and can promptly warn students with low grades and improve their academic performance, which has important teaching guidance significance.
Smart Images

Figure CN113962424B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to a performance prediction method, in particular to a student performance prediction method based on PCANet-BiGRU, belonging to the field of educational technology. Background Art
[0002] With the widespread use of online teaching platforms such as MOOC and Rain Classroom in Chinese universities, education is becoming more efficient, convenient, autonomous and humane. Strengthening the management of the entire process of students' online learning has become the key to ensuring and improving the quality of distance education. As we all know, online learning is a long-term mechanism, and the effort and cognition of daily learning basically determine the quality of learning.
[0003] However, due to the sparsity of data and the scientific nature of the methods, there are currently few effective results on how to conduct scientific research on the learning process, learning intervention, and learning outcomes in a quantitative manner.
[0004] The problems existing in the prior art are:
[0005] Currently, most methods of performance prediction include statistics, traditional machine learning, and deep learning. The statistical methods have low computational cost but most of them cannot achieve high accuracy. Although machine learning and deep learning have higher accuracy, their computational complexity is too high. Summary of the invention
[0006] The purpose of the present invention is to provide a performance prediction method based on PCANet-BiGRU to solve the problem that statistical methods have low computational cost but most of them cannot achieve high accuracy; to solve the problem that machine learning and deep learning have high accuracy but too high computational complexity.
[0007] The object of the present invention is achieved in this way: a performance prediction method based on PCANet-BiGRU, comprising the following steps:
[0008] Step 1) using the basic statistical data and performance data of students collected from the online learning platform, specifically including: obtaining the basic statistical data (such as student number, course name, school year) and performance data (such as the number of homework assignments, the number of completions, and the performance of each homework) of the corresponding students from a specific online learning platform (such as Chinese University MOOC);
[0009] Step 2) Divide the original score data collected in step 1) into three independent parts: training set, validation set and test set, and perform data cleaning and normalization on the training set (since the total score of each regular score is different, normalization is required to eliminate the impact of the dimension on data prediction), and construct a data matrix. Data cleaning specifically includes: based on the training set data obtained in the previous stage, if there are missing values in a column of data needed for training, delete the row of data, that is, delete the raw data; if the column of data is not used in training, even if there are missing values in the column, the row of data is still retained (based on the training set score data obtained in the previous stage, there may be a situation where the score of a regular homework is missing due to missing homework. This data anomaly will cause a great deviation in model training, so these missing data must be cleaned);
[0010] Step 3) Inputting the matrix data into the principal component analysis network PCANet to extract the features of the performance data, wherein the principal component analysis network PCANet is composed of a PCA convolution layer, a nonlinear processing layer, and a feature pooling layer;
[0011] Step 4) Input the data processed by PCANet into the Bi-GRU layer of the bidirectional gated recurrent unit neural network to predict the students' regular grades.
[0012] As a further limitation of the present invention, in step 2), the training set is used to train the model, find the initial parameters of the model through fitting, establish the model, that is, determine the parameters such as the weight and bias of the model; the validation set is used to determine the network structure or control the complexity of the model and other parameters to optimize and adjust the model; and the test set is used to test the performance of the optimal model finally selected. The common ratio of training set, validation set and test set is 6:2:2; when there are not many data sets, 20% of the data is usually randomly selected as the test set, and then the remaining data is subjected to the cross-validation algorithm; the specific steps of the cross-validation algorithm are as follows:
[0013] a. Randomly divide the training data into k equal parts;
[0014] b. Select k-1 copies for training in turn and the remaining copy for verification, and calculate the sum of squares of the prediction errors;
[0015] c. Finally, the sum of squares of the k prediction errors is averaged as the basis for selecting the optimal model structure.
[0016] As a further limitation of the present invention, the bidirectional gated recurrent unit neural network Bi-GRU described in step 4) has both forward and backward propagation added to its hidden layer based on the structure of the gated recurrent unit GRU, and captures the long-term dependency of daily learning performance in different periods by running forward and backward in both directions, thereby obtaining more accurate performance prediction.
[0017] As a further limitation of the present invention, the PCA convolution layer, for each data j in the input layer l, has a convolution kernel P around it. j The window is sampled, and then the convolution kernel is slid to concatenate all the sampling blocks as the representation X of the sample. i =[x i,1 ,x i,2 ,...,x i,n ] and take the average; perform the above operation on N data sets to obtain a new feature matrix X; then perform principal component analysis (PCA) on the matrix; PCA is a commonly used method for data analysis and modeling, which mainly retains the most important features of high-dimensional data, removes noise and unimportant features, and performs dimensionality reduction, greatly reducing data processing costs and speed; its specific algorithm steps are:
[0018] a. Let matrix X have n rows and m columns;
[0019] b. Standardize each row of X;
[0020] c. Find the covariance matrix C of X:
[0021]
[0022] d. Find the eigenvalue E and eigenvector D corresponding to C:
[0023] [E,D]=eig(C)
[0024] Among them, eig is the function of finding eigenvalues and eigenvectors;
[0025] e. Arrange the eigenvectors D according to their corresponding eigenvalues, and select the first k columns to form a new matrix, which is the characteristic component of the data after dimensionality reduction;
[0026] f. Take these k groups of feature vectors as PCA filters and use them as the convolution kernel K to convolve the data set with these k groups of data to complete the convolution operation of extracting data.
[0027] As a further limitation of the present invention, the nonlinear part is used to enhance the characteristic expression of the data, specifically: nonlinear processing is performed on the data obtained by the two-layer PCA convolution: each convolution result is binarized, and the Heaviside step function is selected. Binarization is performed; then, the binarization result is weighted to obtain the "integer map" T of the i-th value on the output feature of the l-th layer i l :
[0028] As a further limitation of the present invention, the feature pooling layer uses the local histogram to perform the feature pooling operation of PCANet (because the output range of the nonlinear processing layer is [0,2 l-1 ], so PCANet is not suitable for the maximum pooling and average pooling commonly used in CNN), integer map T i l The local histogram is divided into blocks, and each histogram is counted and vectorized, denoted as Bhist(T i l ), the feature vector generated by cascading k integer graphs is expressed as: i =[Bhist(T i 1 ),...,Bhist(T i k )] T .
[0029] As a further limitation of the present invention, the GRU in the Bi-GRU neural network has two gate functions: a “reset gate” and an “update gate”; wherein the “reset gate” has two gate functions: t Used to control the state of the previous time period h t-1 For the current pending status The degree of influence of t It is used to determine h t-1 The information is passed to h t The amount of information; the GRU neural network unit update method is
[0030] Among them, tanh is the hyperbolic tangent function, that is Its output value is always in the interval (-1, 1); f is the sigmoid function, and its output value is always in the interval (0, 1), which is used to determine the importance of information and is more helpful in deciding whether to update or abandon data; i t As the input at time t, y t As the output at time t; W ir , W iz , W ih are the weight matrices input to the “reset gate”, “update gate” and state respectively; W hr , W hz , W hh , W o are the weight matrices from the state at a certain moment to the “reset gate”, “update gate”, current state, and initial state; b r 、b z 、b h They are the "reset gate", "update gate" and the state bias.
[0031] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the steps of the above method are implemented when the processor executes the program.
[0032] A computer-readable storage medium stores a computer program, which implements the steps of the above method when executed by a processor.
[0033] A processor is used to run a program, wherein the program executes the above method when it is run.
[0034] Compared with the prior art, the present invention adopts the above technical solution and has the following technical effects:
[0035] The PCANet-BiGRU score prediction model proposed in the present invention takes into account the low consumption of the statistical PCA classification model and the accurate prediction results of deep learning, so as to predict the student's grades and give timely warnings to students who may fail the exam, so that students who are lagging behind in learning can check for omissions and make up for them in time, thereby improving their academic performance, which has important teaching guidance and practical significance. BRIEF DESCRIPTION OF THE DRAWINGS
[0036] Figure 1 This is the network structure of the performance warning model based on PCANet and Bi-GRU.
[0037] Figure 2 This is the network structure of PCANet.
[0038] Figure 3 Schematic diagram of the internal structure of GRU DETAILED DESCRIPTION
[0039] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings:
[0040] Learning is a long-term sequential behavior, with good and bad performances during the process. Different sequential performances will lead to different outputs. In view of the series of defects such as the long parameter training time and the need for special parameter adjustment techniques, the following are proposed: Figure 1 A hybrid prediction model based on principal component analysis network (PCANet) and bidirectional gated recurrent unit (Bi-GRU) neural network is proposed and applied to C language, C++, Python and other courses on the online learning platform; based on the structure of GRU, Bi-GRU captures the long-term dependency of daily learning performance in different periods through forward and backward bidirectional operation in its hidden layer, and obtains more accurate performance prediction; experimental results show that the PCANet-BiGRU performance prediction model effectively predicts performance and improves the accuracy and efficiency of performance prediction.
[0041] The PCANet-BiGRU performance prediction model specifically includes: PCANet and BiGRU;
[0042] PCANet Figure 2 As shown in the figure, PCANet mainly follows the structure and ideas of CNN, but its convolution kernel uses the principal component analysis (PCA) kernel, the nonlinear layer uses the hash algorithm, and the features are generated using histogram statistics. The principal component analysis network PCANet consists of a PCA convolution layer, a nonlinear processing layer, and a feature pooling layer to extract the spatial features of the performance data.
[0043] PCA Convolutional Layer
[0044] For each data in the input layer, a convolution kernel P is constructed around it. j The window is sampled, and then the convolution kernel is slid to concatenate all the sampling blocks as the representation X of the sample. i =[x i,1 ,x i,2 ,...,x i,n ] and take the average; perform the above operation on N data sets to obtain a new feature matrix X; then perform principal component analysis (PCA) on the matrix; PCA is a commonly used method for data analysis and modeling, which mainly retains the most important features of high-dimensional data, removes noise and unimportant features, and performs dimensionality reduction, greatly reducing data processing costs and speed; its specific algorithm steps are:
[0045] aRemember that the matrix X has n rows and m columns;
[0046] b Standardize each row of X;
[0047] cFind the covariance matrix C of X:
[0048] d. Find the eigenvalue E and eigenvector D corresponding to C: [E, D] = eig(C), where eig is the function for finding the eigenvalue and eigenvector;
[0049] e. Arrange the eigenvectors D according to their corresponding eigenvalues, and select the first k columns to form a new matrix, which is the characteristic component of the data after dimensionality reduction;
[0050] f takes these k groups of feature vectors as PCA filters and uses them as the convolution kernel K to convolve the data set with these k groups of data to complete the convolution operation of extracting data.
[0051] PCANet uses two layers of PCA convolutional layers for convolution; if only one layer of convolution is performed, the feature extraction effect is not satisfactory, and if three or more layers of convolution are performed, the amount of calculation will increase sharply due to the increased dimension.
[0052] Non-linear processing layer
[0053] In order to enhance the feature expression of the data, the data from the two-layer PCA convolution is processed nonlinearly: each convolution result is binarized; the Heaviside step function is selected Binarization: In H(x), if the data convolved by PCA is greater than 0, the function is set to 1, otherwise it is set to 0; then, the binarized result is weighted to obtain the "integer map" T of the i-th value on the output feature of the l-th layer i l :
[0054] Feature pooling layer
[0055] Since the output range of the nonlinear processing layer is [0,2 l-1 ], so PCANet is not suitable for the maximum pooling and average pooling commonly used in the second part of CNN, but uses local histograms for PCANet feature pooling operations; integer map T i l The local histogram is divided into blocks, and each histogram is counted and vectorized, denoted as Bhist(T i l );The feature vector obtained by concatenating the vectors generated by k integer graphs is expressed as: f i =[Bhist(T i 1 ),...,Bhist(T i k )] T .
[0056] The GRU neural network in the Bi-GRU network is a type of recurrent neural network RNN, which can fully reflect time series data; but when facing long-term dependencies, it is difficult for RNN to transmit earlier information to the later ones, and important information may be missed; in addition, RNN may also have problems such as gradient disappearance and explosion. The long short-term memory artificial neural network LSTM neural network introduces "input gate", "forget gate" and "output gate" to regulate the information flow to solve the above problems.
[0057] GRU is a variant of the Long Short-Term Memory (LSTM) network. The "input gate" and "forget gate" in LSTM are merged into an "update gate" to construct GRU. GRU not only retains the effect of the Long Short-Term Memory (LSTM) network, but also has fewer parameters, a simpler structure, is easier to implement in calculation, and is less prone to problems such as overfitting.
[0058] like Figure 3As shown, each GRU neural network unit has two gate functions: "reset gate" and "update gate"; the "reset gate" r t Used to control the state of the previous time period h t-1 For the current pending status The degree of influence of "update gate" t It is used to determine h t-1 The information is passed to h t The amount of information.
[0059] The GRU neural network unit is updated as follows:
[0060] Among them, tanh is the hyperbolic tangent function, that is Its output value is always in the interval (-1, 1); f is the sigmoid function, and its output value is always in the interval (0, 1), which is used to determine the importance of information and is more helpful in deciding whether to update or abandon data; i t As the input at time t, y t As the output at time t; W ir , W iz , W ih are the weight matrices input to the “reset gate”, “update gate” and state respectively; W hr , W hz , W hh , W o are the weight matrices from the state at a certain moment to the “reset gate”, “update gate”, current state, and initial state; b r 、b z 、b h They are the "reset gate", "update gate" and the state bias.
[0061] In the GRU structure of the above model, information is transmitted unidirectionally from front to back. However, we believe that the performance data is not only affected by the data of the previous period, but also by the data of the next period. Therefore, a bidirectional gated recurrent neural network Bi-GRU is used. Based on the structure of GRU, Bi-GRU adds both forward and backward propagation to its hidden layer. By running forward and backward in both directions, it can capture the long-term dependency of the usual learning performance in different periods and obtain more accurate performance prediction.
[0062] PCANet-BiGRU
[0063] PCANet has the characteristics of local perception and weight sharing, so the present invention uses PCANet to extract spatial features related to the position of the performance data. PCANet and BiGRU are combined to mine the temporal and spatial features of the performance data through the spatial perception characteristics of PCANet and the bidirectional memory characteristics of BiGRU to achieve performance prediction. The PCANet-BiGRU performance prediction process is as follows:
[0064] Step 1 Preprocess the homework score data and construct a matrix
[0065] Step 2 Input emergency data into PCANet to extract data spatial features
[0066] Step 3: Input the data processed by PCANet into the Bi-GRU layer. The model structure for predicting regular performance data is shown in the figure.
[0067] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the steps of the above method are implemented when the processor executes the program.
[0068] A computer-readable storage medium stores a computer program, which implements the steps of the above method when executed by a processor.
[0069] A processor is used to run a program, wherein the method is executed when the program is run. A specific application embodiment of the present invention is given below:
[0070] Data Acquisition
[0071] The experimental data of the present invention comes from the information of each course stored in the MOOC online learning system. These courses were opened in the system in 2019, and the student performance information was stored in the form of Excel tables. The student performance evaluation criteria in this data set are based on the performance of each regular homework, which is basically the regular homework submitted on a weekly basis. The system preprocesses the assessed data into vectors as input to the neural network based on internal standards for these regular grades.
[0072] Dataset Partitioning
[0073] Randomly extract 20% of the data set as a test set, and then use the remaining data using the cross-validation algorithm:
[0074] a Randomly divide the training data into k equal parts;
[0075] b. Select k-1 copies for training in turn and the remaining copy for verification, and calculate the sum of squares of the prediction errors;
[0076] cFinally, the sum of squares of the k prediction errors is averaged as the basis for selecting the optimal model structure.
[0077] Data cleaning
[0078] Data cleaning for the training set data: Based on the training set score data obtained in the previous stage, there may be a situation where the score of a homework is missing due to missing a regular homework. This data anomaly will cause a huge deviation in model training, so these missing data must be cleaned. Based on the training set data obtained in the previous stage, if there are missing values in a column of data needed for training, delete the row of data, that is, delete the raw data; if the column of data is not used in training, even if there are missing values in the column, the row of data will still be retained.
[0079] Data Normalization
[0080] Since the total scores of each regular exam are different, we normalize them to eliminate the impact of the dimension on data prediction.
[0081] Build a model using the training set
[0082] Construct the PCANet-BiGRU network structure, and use the training set to build the network model, and determine the model's weights and bias parameters through fitting;
[0083] To prevent overfitting, the DropOut technique is used during the training process to enhance the generalization ability of deep neural networks by perturbing the feature graph.
[0084] Use the validation set to determine the hyperparameters for optimizing the model
[0085] Since the learning rate has a significant impact on the performance of the model, too large or too small a learning rate will cause the network to oscillate and fail to converge to the optimal solution. Therefore, it is necessary to use the overall information in the gradient descent to update the learning rate optimization method to improve the model training speed. If the number of PCA layers is too small, the extracted features will be incomplete, while too many will bring huge computational complexity. Therefore, according to the optimization of the validation set, the number of PCA convolution layers is determined to be 2, the number of Bi-GRU layers is 4, the number of neurons in the Bi-GRU layer is 30, the optimized learning algorithm is Adam, and the learning rate is 0.1.
[0086] Evaluate the model using the test set
[0087] Evaluation indicators
[0088] In order to measure the performance of the prediction model, the present invention uses the root mean square error (RMSE) and the mean absolute percentage error (MAPE) as performance evaluation indicators to evaluate the model.
[0089]
[0090]
[0091] Where N represents the number of samples, y i Represents a student's actual grade. Represents the predicted score of the student. MAPE reflects the overall deviation of the predicted score value and can measure the prediction accuracy of the model; RMSE reflects the error between the predicted value and the true value and can measure the accuracy of the predicted value.
[0092] Experimental results analysis
[0093] The experiment predicts the total score based on the regular homework scores of C language, C++ and Python on the MOOC platform. The results are shown in the following table
[0094] Table 1: Results prediction table
[0095] Regular grades C C++ Python RMSE 0.0252 0.0256 0.0263 MAPE / % 2.6139 2.6732 2.7141
[0096] As shown in the table above, the RMSE and MAE values of the prediction results are both below 0.03, and the MAPE of the prediction results is below 2.8%, indicating that this model can effectively predict the grades of different MOOC courses, and the prediction results are relatively accurate, which can meet the requirements of actual MOOC grade prediction. Compared with the previous models that use GM (1,1) and PSO-SVM models for grade prediction, this model has also significantly improved the prediction accuracy.
[0097] The PCANet-BiGRU model established in this paper also has obvious advantages in running time.
[0098] Table 2 Training time comparison table
[0099]
[0100] According to the table above, PCANet-BiGRU has a clear advantage in computational efficiency over CNN-GRU. The more times you train, the greater the advantage.
[0101] Grade prediction can make teachers better aware of students' online learning situation and promote the effective improvement of students' grades. The present invention proposes a deep neural network prediction model combining PCANet and Bi-GRU, which uses PCANet to extract data hidden features to reduce data size, and uses Bi-GRU to extract its internal dynamic laws to achieve early warning of grades. Compared with previous models, this model can integrate data in a timely manner, give students grade warnings, and has significantly improved data utilization and time, which can ultimately promote the development of educational informatization and has high practical value.
[0102] The above description is only a specific implementation of the present invention, but the protection scope of the present invention is not limited thereto. Any person familiar with the technology can understand and think of any changes or substitutions within the technical scope disclosed by the present invention, which should be included in the scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.
Claims
1. A performance prediction method based on PCANet-BiGRU, characterized in that: The following steps are involved: Step 1) using the basic statistical data and performance data of students collected from the online learning platform, the basic statistical data includes student number, course name, and school year, and the performance data includes the number of homework assignments, the number of completions, and the performance of each homework; Step 2) Divide the original score data collected in step 1) into three independent parts: training set, validation set and test set, and perform data cleaning and normalization on the training set to construct a data matrix. The data cleaning specifically includes: based on the training set data obtained in the previous stage, if there are missing values in a column of data needed for training, delete the column of data, that is, delete the student data; if the column of data is not used in training, even if there are missing values in the column, the row of data is still retained; Step 3) Input the matrix data into the principal component analysis network PCANet to extract the features of the performance data. The principal component analysis network PCANet consists of a PCA convolution layer, a nonlinear processing layer, and a feature pooling layer. The PCA convolution layer, for each data j in the input layer l, convolution kernel P is added around it. j The window is sampled, and then the convolution kernel is slid to concatenate all the sampling blocks as the representation X of the sample. i =[x i,1 ,x i,2 ,...,x i,n ] and take the average; perform the above operation on N data sets to obtain a new feature matrix X; then perform principal component analysis PCA on the matrix; PCA is a commonly used method for data analysis and modeling, which mainly retains the most important features of high-dimensional data, removes noise and unimportant features, and performs dimensionality reduction, greatly reducing data processing costs and speed; its specific algorithm steps are: a. Let matrix X have n rows and m columns; b. Standardize each row of X; c. Find the covariance matrix C of X: d. Find the eigenvalue E and eigenvector D corresponding to C: [E,D]=eig(C) Among them, eig is the function of finding eigenvalues and eigenvectors; e. Arrange the eigenvectors D according to their corresponding eigenvalues, and select the first k columns to form a new matrix, which is the characteristic component of the data after dimensionality reduction; f. Take these k groups of feature vectors as PCA filters and use them as convolution kernel K to convolve the data set with these k groups of data to complete the convolution operation of extracting data; The nonlinear processing is used to enhance the feature expression of the data, specifically: nonlinear processing is performed on the data obtained by the two-layer PCA convolution: each convolution result is binarized, and the Heaviside step function is selected Binarization is performed; then, the binarization result is weighted to obtain the "integer map" T of the i-th value on the output feature of the l-th layer i l : The feature pooling layer uses the local histogram to perform the feature pooling operation of PCANet. i l The local histogram is divided into blocks, and each histogram is counted and vectorized, denoted as Bhist(T i l ), the feature vector generated by cascading k integer graphs is expressed as: f i =[Bhist(T i 1 ),...,Bhist(T i k )] T ; Step 4) Input the data processed by PCANet into the Bi-GRU layer of the bidirectional gated recurrent unit neural network to predict the students' regular grades. The Bi-GRU neural network adds forward and backward propagation to its hidden layer based on the structure of the gated recurrent unit GRU. It captures the long-term dependency of regular learning grades in different periods by running forward and backward in both directions, and obtains more accurate grade prediction.
2. The PCANet-BiGRU-based performance prediction method according to claim 1, characterized in that: In step 2), the training set is used to train the model, find the initial parameters of the model through fitting, and establish the model, that is, determine the parameters such as the weight and bias of the model; the validation set is used to determine the network structure or control the complexity of the model to optimize and adjust the parameters of the model; and the test set is used to test the performance of the optimal model finally selected; the common ratio of training set, validation set and test set is 6:2:2; when there are not many data sets, first randomly select 20% of the data as the test set, and then use the cross-validation algorithm for the remaining data; the specific steps of the cross-validation algorithm are as follows: a. Randomly divide the training data into k equal parts; b. Select k-1 copies for training in turn and the remaining copy for verification, and calculate the sum of squares of prediction errors; c. Finally, the sum of squares of the k prediction errors is averaged as the basis for selecting the optimal model structure.
3. The PCANet-BiGRU-based performance prediction method according to claim 1, characterized in that: The GRU in the Bi-GRU neural network has two gate functions: "reset gate" and "update gate"; t Used to control the state of the previous time period h t-1 For the current pending status The degree of influence; and the "update gate" z t It is used to determine h t-1 The information is passed to h t The amount of information; the GRU neural network unit update method is Among them, tanh is the hyperbolic tangent function, that is, Its output value is always in the interval (-1, 1); f is the sigmoid function, and its output value is always in the interval (0, 1), which is used to determine the importance of information and is more helpful in deciding whether to update or abandon data; i t As the input at time t, y t As the output at time t; W ir , W iz , W ih are the weight matrices input to the "reset gate", "update gate" and state respectively; W hr , W hz , W hh W0 and W1 are the weight matrices from the state at a certain moment to the "reset gate", "update gate", current state and initial state respectively; b r 、b z 、b h They are the "reset gate", "update gate" and state bias.
4. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of the method according to any one of claims 1 to 3 are implemented.
5. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 3 are implemented.
6. A processor, characterized in that: The processor is used to run a program, wherein the program executes the method according to any one of claims 1 to 3 when running.
