An academic early warning method based on meta-learning and support vector description

Abstract

The application discloses a learning achievement early warning method based on meta learning and support vector description, and relates to the technical field of learning achievement early warning. The method comprises the following steps: a learning achievement prediction main model is established through meta learning; a learning achievement prediction auxiliary model is established based on support vector description; current learning behavior data of a student in a first stage is acquired and input into the learning achievement prediction main model to obtain a first learning achievement result; current learning behavior data of the student in a second stage is acquired and input into the learning achievement prediction auxiliary model to obtain a second learning achievement result; a learning achievement early warning condition is set, and whether the student meets the learning achievement early warning condition is judged according to the first learning achievement result and the second learning achievement result; and when the student meets the learning achievement early warning condition, corresponding early warning measures are taken for the student. The application judges whether the student has a learning achievement early warning condition through the learning achievement prediction main model and the learning achievement prediction auxiliary model, the learning achievement prediction auxiliary model compensates for errors of the learning achievement prediction main model, the accuracy of learning achievement early warning is improved, and the one-sidedness of a learning achievement early warning report is reduced.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence, and in particular to an academic early warning method based on meta-learning and support vector description. Background Technology

[0002] Currently, universities are vigorously promoting credit-based teaching, which, while giving students more choices and autonomy, also places higher demands on students' learning abilities and self-management skills. Universities generally lack systematic learning guidance systems, and some students' insufficient learning abilities and poor self-discipline lead to failing grades and GPAs that do not meet graduation requirements, resulting in delayed graduation. Furthermore, in recent years, the number of students in universities who have delayed graduation due to academic problems has been steadily increasing. Therefore, it is urgent to help students improve their learning abilities to complete their studies.

[0003] To address the aforementioned issues, some universities have begun exploring academic early warning systems. These systems provide timely academic warnings when students' credits or GPAs are abnormal, aiming to help them successfully complete their studies. The types of academic early warning methods are limited. The most common method utilizes information technology to statistically analyze students' course data from previous academic years, such as GPA, to inform and warn them of potential academic difficulties. This involves communication and cooperation among schools, parents, and students to help students complete their studies smoothly. However, this method, which relies on a single data type, has limitations. It cannot truly reflect a student's learning situation in the current academic year. For example, some students may have failed courses in previous years but have a positive learning attitude in the current year. If an academic warning report is issued solely because such students' previous GPAs meet the warning requirements, the report may contain significant errors. This method has low accuracy and high simplification, failing to comprehensively assess a student's learning situation in the current academic year. Summary of the Invention

[0004] The purpose of this invention is to provide an academic early warning method based on meta-learning and support vector description, in order to solve one or more technical problems existing in the prior art, or at least provide a beneficial option or create conditions.

[0005] The solution to the technical problem of this invention is to provide an academic early warning method based on meta-learning and support vector description, the method comprising the following steps:

[0006] S100: Based on deep neural networks, an initial academic prediction model is established. Historical course data of students in different subjects in the historical academic year is obtained, the historical course data is preprocessed, a cross-task dataset is established, and the initial academic prediction model is pre-trained based on the cross-task dataset to obtain the academic prediction base model.

[0007] S200: Obtain historical learning behavior data corresponding to historical course data, establish a single-task dataset, and use the single-task dataset to fine-tune and train the academic prediction base model to obtain the academic prediction master model.

[0008] S300. Obtain the student's current learning behavior data in the first stage of the current academic year, input the learning behavior data into the main academic prediction model, and obtain the student's first academic result.

[0009] The first academic result includes future course data corresponding to the second time point, which is after the first stage;

[0010] S400: An academic prediction auxiliary model is established based on support vector description. The student's current learning behavior data in the second stage of the current academic year is obtained and input into the academic prediction auxiliary model to obtain the student's second academic result.

[0011] The second academic result is the probability of future academic warning corresponding to the second stage, which is after the first stage and before the second moment.

[0012] S500 sets academic warning conditions, and judges whether the student meets the academic warning conditions based on the first academic result and the second academic result. When the student meets the academic warning conditions, an academic warning is issued for the student.

[0013] As a further improvement to the above technical solution, the step S100 of pre-training the initial academic prediction model based on the cross-task dataset to obtain the academic prediction base model is as follows:

[0014] S110, Initialize the network parameters of the initial academic prediction model. Given N training task classifiers, set the network parameters The network structure θ assigned to these N training task classifiers N ,Right now

[0015] S120, divide the cross-task dataset into N groups of training data. Each group of training data has its corresponding training task classifier. Each group of training data is divided into a query set and a support set.

[0016] S130, each training task classifier is trained on its corresponding support set, and the total loss function of all training task classifiers is calculated based on the query set corresponding to each training task classifier. The network parameters of the initial academic prediction model are updated using gradient descent and the total loss function.

[0017] S140, Set the expected value range of the network parameters of the initial academic prediction model, and determine whether the network parameters of the initial academic prediction model are within the expected value range. If so, proceed to step S141.

[0018] S141, Output the academic prediction base model;

[0019] If the network parameters of the initial academic prediction model are outside the predicted value range, proceed to step S142;

[0020] S142, repeat steps S130 to S140.

[0021] As a further improvement to the above technical solution, step S130 specifically includes:

[0022] S131, select n training task classifiers from N training task classifiers through random sampling. Train the n selected training task classifiers on their respective support sets, and update their network parameters θ once using gradient descent. n For θ n * n = 1, 2, 3…;

[0023] S132, based on the network parameters θ obtained in step S131 n * These n training task classifiers are tested on their corresponding query sets, and the first sub-loss function L corresponding to each training task classifier is calculated. n (θ n * );

[0024] S133, the total loss function is calculated based on all the first sub-loss functions; wherein the total loss function satisfies the following formula:

[0025]

[0026] in, Represents the total loss function;

[0027] S134, Update the network parameters of the initial academic prediction model based on the total loss function. The network parameters of the updated academic prediction initial model satisfy the following formula:

[0028]

[0029]

[0030] Where β represents the learning rate of the initial academic prediction model. Represents the total loss function Network parameters of the initial model for academic prediction The derivative of .

[0031] As a further improvement to the above technical solution, step S200 specifically includes:

[0032] S210, Obtain historical learning behavior data corresponding to historical course data. The format of the historical course data is the same as that of the historical course data during preprocessing. The historical learning behavior data constitutes a single-task dataset.

[0033] S220: The single-task dataset is randomly sampled and divided into m training sets and j test sets. A gradient descent operation is then performed on the academic prediction base model using the data from these m training sets to update the network parameters of the academic prediction base model. Updated to

[0034] S230, calculate the second sub-loss function for each training set based on these j test sets, and calculate the overall loss function based on all the second sub-loss functions; wherein the overall loss function satisfies the following formula:

[0035]

[0036] Among them, L m This represents the second sub-loss function corresponding to the m-th training set. Represents the overall loss function;

[0037] S240, Set the expected range of the overall loss function of the academic prediction base model, and determine whether the overall loss function is within its expected range. If the overall loss function is within its expected range, proceed to step S241; if the overall loss function is outside its expected range, proceed to step S242.

[0038] S241, Output the main model for academic prediction;

[0039] S242, repeat steps S220 to S240.

[0040] As a further improvement to the above technical solution, the step of establishing the academic prediction auxiliary model in step S400 is as follows:

[0041] S410: Obtain sample behavioral data of students with academic warnings in the second phase of the historical academic year, perform PCA analysis and preprocessing on the sample behavioral data to obtain an auxiliary sample set;

[0042] The preprocessing includes: the sample behavior data is of multiple categories, each with different attributes, and a normalization method is used to uniformly measure the sample behavior data of all categories.

[0043] S420, the auxiliary sample set is divided into a training sample set X = {x} by random sampling. i {i = 1, 2, 3, ..., l} and test sample set X′ = {x i ′ For each x, i = 1, 2, 3, ..., n, a prediction model is built based on a support vector descriptor classifier. The training sample set is used as the input to the prediction model, and the training sample set X = {x, i = 1, 2, 3, ..., n} is mapped using a mapping function F. i The region {i,i = 1, 2, 3, ..., l} is mapped to a high-dimensional feature space; wherein the mapping function F is a Gaussian kernel function, which satisfies the following formula:

[0044]

[0045] Where y represents the center of the Gaussian kernel function, and σ represents the width parameter of the Gaussian kernel function;

[0046] S430, Set the slack variable ε of the support vector to describe the classifier. i With the penalty parameter C, the support vectors describe how the classifier gradually forms a hypersphere capable of describing the auxiliary sample set during training; wherein, the hypersphere satisfies the following formula:

[0047]

[0048]

[0049] Where R represents the radius of the hypersphere, μ represents the center of the hypersphere, C represents the penalty parameter of the support vector classifier, and ε i Let x represent a slack variable. i Let represent the course data of the i-th sample in the auxiliary sample set, and st denotes constraints;

[0050] The hypersphere is centered at μ and contains the training sample set X = {x}. i The data field is i = 1, 2, 3, ...;

[0051] S440, introducing the Lagrange operator, constructing the Lagrange function, and combining the Lagrange function and the KKT conditions, obtain the optimal hypersphere and its corresponding center and radius; wherein, the optimal hypersphere satisfies the following formula:

[0052]

[0053]

[0054] Where L represents the optimal hypersphere, α i α j All are Lagrange operators;

[0055] The center of the optimal hypersphere satisfies the following formula:

[0056]

[0057] The radius of the optimal hypersphere is calculated from its center, and its radius satisfies the following formula:

[0058]

[0059] Where, x k Represents any data point in the training sample set;

[0060] S450: Evaluate the performance of the academic prediction auxiliary model based on the test sample set, obtain the model accuracy of the academic prediction auxiliary model, set the expected value of the model accuracy, and determine whether the model accuracy is less than its expected value.

[0061] If the model accuracy is less than the expected value, then repeat steps S420 to S450.

[0062] If the model accuracy is not less than the expected value, then output the academic prediction auxiliary model.

[0063] As a further improvement to the above technical solution, step S400, which involves obtaining the student's current learning behavior data for the second phase of the current academic year and inputting it into the academic prediction auxiliary model, includes:

[0064] The current behavioral data is input into the academic prediction auxiliary model to obtain the support vectors of the current behavioral data in the hypersphere state of the academic prediction auxiliary model; wherein, the support vectors satisfy the following formula:

[0065]

[0066] Where z represents the current behavior data;

[0067] Determine whether the support vector is located within the hypersphere of the academic prediction auxiliary model, calculate the academic warning probability, and output the second academic result.

[0068] As a further improvement to the above technical solution, the academic warning probability satisfies the following formula:

[0069]

[0070] Where P(z) represents the academic warning probability, when f(z)-R 2 A value ≥ 0 indicates that the support vectors are located outside the hypersphere; when f(z) - R 2 A value less than 0 indicates that the support vector is located within the hypersphere.

[0071] As a further improvement to the above technical solution, the step of determining whether a student meets the academic warning conditions based on the first and second academic results, and issuing an academic warning to the student when the student meets the academic warning conditions, includes:

[0072] S510, the academic warning conditions include the course data pass range, the academic warning threshold probability and the compensation value. Determine whether the first academic result is within the course data pass range. If not, proceed to step S520; if yes, do not output the academic warning report.

[0073] S520, determine whether the probability of the second academic result is lower than the academic warning threshold; if so, proceed to step S521;

[0074] S521, Generate an academic warning report for the student based on the first and second academic results. The academic warning report includes the student's current learning behavior data in the first and second phases of the current academic year, future course data, future academic warning probability, and the student's personal information.

[0075] If the second academic result is not lower than the academic warning threshold probability, proceed to step S522;

[0076] S522, the absolute values ​​of the first academic result and the second academic result are multiplied and the compensation result is output. When the compensation result is lower than the compensation value, an academic warning report for the student is generated.

[0077] A computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the steps of the academic early warning method based on meta-learning and support vector description.

[0078] An academic early warning system based on meta-learning and deep learning, the system comprising:

[0079] At least one processor;

[0080] At least one memory for storing at least one program;

[0081] When the at least one program is executed by the at least one processor, the at least one processor implements the academic early warning method based on meta-learning and support vector description.

[0082] The beneficial effects of this invention are: it provides an academic early warning method based on meta-learning and support vector description, comprising: establishing an academic prediction master model through meta-learning; establishing an academic prediction auxiliary model based on support vector description; acquiring the student's current learning behavior data in the first stage of the current academic year and inputting it into the academic prediction master model to obtain a first academic result; acquiring the student's current learning behavior data in the second stage of the current academic year and inputting it into the academic prediction auxiliary model to obtain a second academic result; setting academic early warning conditions; judging whether the student meets the academic early warning conditions based on the first and second academic results; and taking corresponding early warning measures for the student when the student meets the academic early warning conditions. This application determines whether a student has an academic early warning situation through the academic prediction master model and the academic prediction auxiliary model. Meta-learning is used to improve the accuracy of the model when constructing the academic prediction master model, while support vector description is used to reduce the computational load when constructing the academic prediction auxiliary model. The academic prediction auxiliary model compensates for errors in the academic prediction master model, improving the accuracy of academic early warning, reducing the bias of academic early warning reports, and enabling more accurate academic early warning. Attached Figure Description

[0083] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly explained below. Obviously, the described drawings are only a part of the embodiments of the present invention, and not all of them. Those skilled in the art can obtain other design schemes and drawings based on these drawings without creative effort.

[0084] Figure 1 This is a flowchart of an academic early warning method based on meta-learning and support vector description;

[0085] Figure 2 This is a flowchart of an academic early warning method based on meta-learning and support vector description, which pre-trains an initial academic prediction model based on a cross-task dataset.

[0086] Figure 3 This is a flowchart illustrating a method for updating the network parameters of the initial academic prediction model based on meta-learning and support vector description in academic early warning.

[0087] Figure 4 This is a flowchart illustrating a method for establishing a master model for academic prediction based on meta-learning and support vector description for academic early warning.

[0088] Figure 5 This is a flowchart illustrating a method for establishing an academic prediction auxiliary model based on meta-learning and support vector description for academic early warning.

[0089] Figure 6This is a flowchart illustrating a method for determining the existence of academic warning situations based on meta-learning and support vector description. Detailed Implementation

[0090] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0091] It should be noted that although functional modules are divided in the system diagram and the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the system or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0092] Please refer to Figures 1 to 6 This application provides an academic early warning method based on meta-learning and support vector description, which includes the following steps:

[0093] S100: Based on deep neural networks, an initial academic prediction model is established. Historical course data of students in different subjects in the historical academic year is obtained, the historical course data is preprocessed, a cross-task dataset is established, and the initial academic prediction model is pre-trained based on the cross-task dataset to obtain the academic prediction base model.

[0094] The steps for preprocessing historical course data include:

[0095] Find and discard missing values ​​in historical course data;

[0096] Standardize the history course data and convert the data format of the history course data into an input format that conforms to deep neural networks;

[0097] S200: Obtain historical learning behavior data corresponding to historical course data, establish a single-task dataset, and use the single-task dataset to fine-tune and train the academic prediction base model to obtain the academic prediction master model.

[0098] S300. Obtain the student's current learning behavior data in the first stage of the current academic year, input the learning behavior data into the main academic prediction model, and obtain the student's first academic result.

[0099] The first academic result includes future course data corresponding to the second time point, which is after the first stage;

[0100] S400: An academic prediction auxiliary model is established based on support vector description. The student's current learning behavior data in the second stage of the current academic year is obtained and input into the academic prediction auxiliary model to obtain the student's second academic result.

[0101] The second academic result is the probability of future academic warning corresponding to the second stage, which is after the first stage and before the second moment.

[0102] S500 sets academic warning conditions, and judges whether the student meets the academic warning conditions based on the first academic result and the second academic result. When the student meets the academic warning conditions, an academic warning is issued for the student.

[0103] In steps S100 and S200, the amount of course data collected for students in this application is small. Due to the limited course data, if the network model is directly subjected to machine learning based on this data, the prediction accuracy of the output model will be low, which is very detrimental to predicting students' future course data. Therefore, this application adopts the concept of meta-learning to update the network parameters of the initial academic prediction model. Meta-learning, also known as "learning to learn," enables the network model to acquire the ability to learn and tune parameters, allowing the network model to quickly learn new tasks based on existing knowledge. Meta-learning is suitable for situations with few training samples and the existence of another new training task, and it is also suitable for finding the initial parameters of a network model.

[0104] In machine learning, the specific steps for training a neural network are: preprocessing the dataset, selecting a network model, setting hyperparameters, initializing parameters, selecting an optimizer, defining a loss function, and updating parameters using gradient descent. The parameters involved are called network parameters, which need to be set and defined manually. Meta-learning, on the other hand, involves first training a better network with the parameters of another task, and then training it on a specific task.

[0105] Furthermore, in step S100, an initial academic prediction model is established based on a deep neural network. Historical course data is used as the model input, and the predicted future course data is used as the model output. The course data represents grade points. Please refer to [link / reference]. Figure 2 The specific steps for pre-training the initial academic prediction model based on a cross-task dataset to obtain the base academic prediction model are as follows:

[0106] S110, Initialize the network parameters of the initial academic prediction model. Given N training task classifiers, set the network parameters The network structure θ assigned to these N training task classifiers N ,Right now

[0107] S120, divide the cross-task dataset into N groups of training data. Each group of training data has its corresponding training task classifier. Each group of training data is divided into a query set and a support set.

[0108] S130, each training task classifier is trained on its corresponding support set, and the total loss function of all training task classifiers is calculated based on the query set corresponding to each training task classifier. The network parameters of the initial academic prediction model are updated using gradient descent and the total loss function.

[0109] S140, Set the expected value range of the network parameters of the initial academic prediction model, and determine whether the network parameters of the initial academic prediction model are within the expected value range. If so, proceed to step S141.

[0110] S141, Output the academic prediction base model;

[0111] If the network parameters of the initial academic prediction model are outside the predicted value range, proceed to step S142;

[0112] S142, repeat steps S130 to S140.

[0113] Meta-learning is divided into two stages: training task training and testing task training. In this application, step S100, which describes pre-training the initial academic prediction model based on a cross-task dataset, belongs to the first stage of meta-learning, namely training task training. Its purpose is to find the optimal network parameters of the initial academic prediction model and assign these optimal network parameters to the initial academic prediction model to improve its network performance.

[0114] In the training task, the initial parameters of the academic prediction initial model are first initialized. Given N training task classifiers, each training task's dataset is divided into a query set and a support set. The network parameters of the initial learning prediction model are trained using the support sets of these N training task classifiers, and the parameters of each training task classifier are trained separately. The performance of the corresponding training task classifier is tested using the query sets of these N training task classifiers, and the total loss function of all training task classifiers is calculated. Finally, the gradient descent method is used for calculation. To update the network parameters of the initial academic prediction model.

[0115] Please refer to Figure 3 Step S130 specifically includes:

[0116] S131, select n training task classifiers from N training task classifiers through random sampling. Train the n selected training task classifiers on their respective support sets, and update their network parameters θ once using gradient descent. n For θ n * n = 1, 2, 3…;

[0117] S132, based on the network parameters θ obtained in step S131 n * These n training task classifiers are tested on their corresponding query sets, and the first sub-loss function L corresponding to each training task classifier is calculated. n (θ n * );

[0118] S133, the total loss function is calculated based on all the first sub-loss functions; wherein the total loss function satisfies the following formula:

[0119]

[0120] in, Represents the total loss function;

[0121] S134, Update the network parameters of the initial academic prediction model based on the total loss function. The network parameters of the updated academic prediction initial model satisfy the following formula:

[0122]

[0123]

[0124] Where β represents the learning rate of the initial academic prediction model. Represents the total loss function Network parameters of the initial model for academic prediction The derivative of .

[0125] This application achieves the second stage of meta-learning through step S200, namely, test task training, which can be regarded as a normal machine learning process. Please refer to... Figure 4 The specific steps of step S200 are as follows:

[0126] S210, Obtain historical learning behavior data corresponding to historical course data. The format of the historical course data is the same as that of the historical course data during preprocessing. The historical learning behavior data constitutes a single-task dataset.

[0127] S220: The single-task dataset is randomly sampled and divided into m training sets and j test sets. A gradient descent operation is then performed on the academic prediction base model using the data from these m training sets to update the network parameters of the academic prediction base model. Updated to

[0128] S230, calculate the second sub-loss function for each training set based on these j test sets, and calculate the overall loss function based on all the second sub-loss functions; wherein the overall loss function satisfies the following formula:

[0129]

[0130] Among them, L m This represents the second sub-loss function corresponding to the m-th training set. Represents the overall loss function;

[0131] S240, Set the expected range of the overall loss function of the academic prediction base model, and determine whether the overall loss function is within its expected range. If the overall loss function is within its expected range, proceed to step S241; if the overall loss function is outside its expected range, proceed to step S242.

[0132] S241, Output the main model for academic prediction;

[0133] S242, repeat steps S220 to S240.

[0134] The purpose of step S200 is to fine-tune the academic prediction base model. First, historical learning behavior data corresponding to historical course data is obtained and a single-task dataset is constructed, the format of which must be consistent with the format of historical course data. Then, the single-task dataset is divided into a training set and a test set through random sampling. The academic prediction base model is trained on the training set and a gradient descent is performed to update the network parameters of the academic prediction base model. The overall loss function of the training set is calculated based on the test set to obtain the academic prediction master model.

[0135] It should be noted that steps S100 and S200 are training for the training task and testing task of meta-learning, respectively. The first stage obtains the optimal initialization parameters of the network model based on the cross-task dataset, while the second stage uses the optimal initialization parameters obtained from the first stage to obtain the final academic prediction master model based on the single-task training set.

[0136] Step S300 involves obtaining the student's current learning behavior data for the first phase of the current academic year, inputting the learning behavior data into the academic prediction master model, and obtaining the student's first academic result. The first academic result includes future course data corresponding to the second time point; that is, based on the academic prediction master model, the predicted future course data can be obtained through the current learning behavior data.

[0137] In this embodiment, step S400 is to establish an academic prediction auxiliary model based on Support Vector Data Description (SVDD) and obtain the student's second academic result through the academic prediction auxiliary model, wherein the second academic result is the future academic warning probability corresponding to the second time point.

[0138] Please refer to the following: Figure 5 The steps for establishing the academic prediction auxiliary model are as follows:

[0139] S410: Obtain sample behavioral data of students with academic warnings in the second phase of the historical academic year, perform PCA analysis and preprocessing on the sample behavioral data to obtain an auxiliary sample set;

[0140] The preprocessing includes: the sample behavior data is of multiple categories, each with different attributes, and a normalization method is used to uniformly measure the sample behavior data of all categories.

[0141] S421, the auxiliary sample set is divided into a training sample set X = {x} by random sampling. i {i = 1, 2, 3, ..., l} and test sample set X′ = {x′} i For each x, i = 1, 2, 3, ..., n, a prediction model is built based on a support vector descriptor classifier. The training sample set is used as the input to the prediction model, and the training sample set X = {x, i = 1, 2, 3, ..., n} is mapped using a mapping function F. i The region {i,i = 1, 2, 3, ..., l} is mapped to a high-dimensional feature space; wherein the mapping function F is a Gaussian kernel function, which satisfies the following formula:

[0142]

[0143] Where y represents the center of the Gaussian kernel function, and σ represents the width parameter of the Gaussian kernel function;

[0144] S430, Set the slack variable ε of the support vector to describe the classifier. i With the penalty parameter C, the support vectors describe how the classifier gradually forms a hypersphere capable of describing the auxiliary sample set during training; wherein, the hypersphere satisfies the following formula:

[0145]

[0146]

[0147] Where R represents the radius of the hypersphere, μ represents the center of the hypersphere, C represents the penalty parameter of the support vector classifier, and ε i Let x represent a slack variable. i Let represent the course data of the i-th sample in the auxiliary sample set, and st denotes constraints;

[0148] The hypersphere is centered at μ and contains the training sample set X = {x}. i The data field is i = 1, 2, 3, ...;

[0149] S440, introducing the Lagrange operator, constructing the Lagrange function, and combining the Lagrange function and the KKT conditions, obtain the optimal hypersphere and its corresponding center and radius; wherein, the optimal hypersphere satisfies the following formula:

[0150]

[0151]

[0152] Where L represents the optimal hypersphere, α i α j All are Lagrange operators;

[0153] The center of the optimal hypersphere satisfies the following formula:

[0154]

[0155] The radius of the optimal hypersphere is calculated from its center, and its radius satisfies the following formula:

[0156]

[0157] Where, x k Represents any data point in the training sample set;

[0158] S450: Evaluate the performance of the academic prediction auxiliary model based on the test sample set, obtain the model accuracy of the academic prediction auxiliary model, set the expected value of the model accuracy, and determine whether the model accuracy is less than its expected value.

[0159] If the model accuracy is lower than expected, then repeat steps S420 to S450.

[0160] If the model accuracy is not less than the expected value, then output the academic prediction auxiliary model.

[0161] In this technical field, Support Vector Description (SVM) is a single classifier commonly used for detecting outliers or remote values ​​in data. Its basic idea is to find a hypersphere that describes a given target dataset, containing all the data points in the dataset with the smallest possible radius. The data points located on the hypersphere are called support vectors. The most crucial issue during the training of an SVM classifier is obtaining the minimum radius of the hypersphere.

[0162] In this embodiment, the training sample set is first mapped to a high-dimensional feature space using a Gaussian kernel function. Preferably, the mapping function can be other kernel functions. Then, a hypersphere is constructed in the high-dimensional feature space to contain a high-probability region, that is, the data domain of the training sample set is contained with the center of the hypersphere as the center. The optimal hypersphere is obtained through the Lagrangian function and KKT conditions.

[0163] In step S430 of this invention, a slack variable and a penalty parameter are set for the support vector classifier. The slack variable is used to allow some outliers in the training sample set to be located outside the hypersphere, while the penalty parameter is used to set the degree of penalty for outliers. The penalty parameter is the ratio of the number of support vectors mapped to the high-dimensional feature space of the training sample set to the total number of data in the training sample set, which satisfies the following formula:

[0164]

[0165] Where SVs is the number of support vectors, and l is the total number of data in the training sample set.

[0166] In this embodiment, the step of obtaining the student's current learning behavior data in the second phase of the current academic year and inputting it into the academic prediction auxiliary model to obtain the student's second academic result includes:

[0167] The current learning behavior data is input into the academic prediction auxiliary model to obtain the support vectors of the current behavior data in the hypersphere state of the academic prediction auxiliary model; wherein, the support vectors satisfy the following formula:

[0168]

[0169] Where z represents the current behavior data;

[0170] Determine whether the support vector lies within the hypersphere of the academic prediction auxiliary model, calculate the academic warning probability, and output the second academic result; wherein, the academic warning probability satisfies the following formula:

[0171]

[0172] Where P(z) represents the academic warning probability, when f(z)-R 2A value ≥ 0 indicates that the support vectors are located outside the hypersphere; when f(z) - R 2 A value less than 0 indicates that the support vector is located within the hypersphere.

[0173] Specifically, the current learning behavior data is input into the academic prediction auxiliary model to obtain the support vector of the current behavior data in the hypersphere state. If the support vector falls inside the hypersphere, P(z) is negative; if the support vector falls outside the hypersphere, P(z) is positive.

[0174] Please refer to Figure 6 In this embodiment, step S500 involves determining whether the student meets the academic warning conditions based on the first and second academic results. When the student meets the academic warning conditions, an academic warning is issued to the student. Specifically, this involves:

[0175] S510, the academic warning conditions include the course data pass range, the academic warning threshold probability and the compensation value. Determine whether the first academic result is within the course data pass range. If not, proceed to step S520; if yes, do not output the academic warning report.

[0176] S520, determine whether the probability of the second academic result is lower than the academic warning threshold; if so, proceed to step S521;

[0177] S521, Generate an academic warning report for the student based on the first and second academic results. The academic warning report includes the student's current learning behavior data in the first and second phases of the current academic year, future course data, future academic warning probability, and the student's personal information.

[0178] If the second academic result is not lower than the academic warning threshold probability, proceed to step S522;

[0179] S522, the absolute values ​​of the first academic result and the second academic result are multiplied and the compensation result is output. When the compensation result is lower than the compensation value, an academic warning report for the student is generated.

[0180] This invention measures whether a student meets the academic warning criteria from two dimensions. The first dimension is the future course data obtained through the main academic prediction model (i.e., the first academic outcome), and the second dimension is the future academic warning probability obtained through the auxiliary academic prediction model (i.e., the second academic outcome). The second dimension is used to compensate for the errors in the first dimension and reduce the one-sidedness of the academic warning results.

[0181] In step S500, the academic warning conditions are first set as the course data pass range, the academic warning threshold probability, and the compensation value. The course data pass range can be determined according to the school's relevant regulations on grade points, and it is used to determine whether future course data is a pass value. The academic warning threshold probability is used to determine whether the future academic warning probability is lower than the normal value, and the compensation value is used to compensate for the judgment result.

[0182] After setting the academic warning conditions, the system judges whether a student has an academic warning based on the first and second academic results. The student is considered to have an academic warning only when the first academic result exceeds the course data pass range and the second academic result is lower than the academic warning threshold probability, and an academic warning report is output.

[0183] In another implementation, if the first academic result exceeds the qualified range of the course data but the second academic result is not lower than the academic warning threshold probability, the next step is required: the absolute values ​​of the first and second academic results are divided into grades and a compensation result is output, and the student is judged to meet the academic warning based on the compensation value.

[0184] In this embodiment, the numerical type of the second academic result is a percentage. If the current learning behavior data is located within the hypersphere, the second academic result is negative. Therefore, when performing the product operation, the absolute value of the second academic result needs to be taken. The purpose of multiplying the absolute value of the second academic result with the first academic result is to compensate for any errors in the first academic result using the second academic result.

[0185] The beneficial effects of the above technical solution are: by analyzing whether students have academic warning problems through two dimensions, the second dimension plays a role in compensating for the error of the first dimension, thereby improving the accuracy of academic warnings and reducing the one-sidedness of academic warning reports.

[0186] This application also discloses a computer-readable storage medium, characterized in that a computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, it implements the steps of the academic early warning method based on meta-learning and support vector description.

[0187] This application also discloses an academic early warning system based on meta-learning and deep learning, characterized in that the system includes:

[0188] At least one processor;

[0189] At least one memory for storing at least one program;

[0190] When the at least one program is executed by the at least one processor, the at least one processor implements the academic early warning method based on meta-learning and support vector description.

[0191] The preferred embodiments of the present invention have been described in detail above, but the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention, and these equivalent modifications or substitutions are all included within the scope defined by the claims of this application.

Claims

1. An academic early warning method based on meta-learning and support vector description, characterized in that, The method includes the following steps: S100: Based on deep neural networks, an initial academic prediction model is established. Historical course data of students in different subjects in the historical academic year is obtained, the historical course data is preprocessed, a cross-task dataset is established, and the initial academic prediction model is pre-trained based on the cross-task dataset to obtain the academic prediction base model. S200: Obtain historical learning behavior data corresponding to historical course data, establish a single-task dataset, and use the single-task dataset to fine-tune and train the academic prediction base model to obtain the academic prediction master model. S300. Obtain the student's current learning behavior data in the first stage of the current academic year, input the learning behavior data into the main academic prediction model, and obtain the student's first academic result. The first academic result includes future course data corresponding to the second time point, which is after the first stage; S400: An academic prediction auxiliary model is established based on support vector description. The student's current learning behavior data in the second stage of the current academic year is obtained and input into the academic prediction auxiliary model to obtain the student's second academic result. The second academic result is the probability of future academic warning corresponding to the second stage, which is after the first stage and before the second moment. S500 sets academic warning conditions, and judges whether the student meets the academic warning conditions based on the first academic result and the second academic result. When the student meets the academic warning conditions, an academic warning is issued to the student. Specifically, step S200 includes: S210, Obtain historical learning behavior data corresponding to historical course data. The format of the historical course data is the same as that of the historical course data during preprocessing. The historical learning behavior data constitutes a single-task dataset. S220: The single-task dataset is randomly sampled and divided into m training sets and j test sets. A gradient descent operation is then performed on the academic prediction base model using the data from these m training sets to update the network parameters of the academic prediction base model. Updated to ; S230, calculate the second sub-loss function for each training set based on these j test sets, and calculate the overall loss function based on all the second sub-loss functions; wherein the overall loss function satisfies the following formula: in, This represents the second sub-loss function corresponding to the m-th training set. Represents the overall loss function; S240, Set the expected range of the overall loss function of the academic prediction base model, and determine whether the overall loss function is within its expected range. If the overall loss function is within its expected range, proceed to step S241; if the overall loss function is outside its expected range, proceed to step S242. S241, Output the main model for academic prediction; S242, repeat steps S220 to S240 in a loop; The step of determining whether a student meets the academic warning criteria based on the first and second academic results, and issuing an academic warning to the student when the student meets the criteria, includes: S510, the academic warning conditions include the course data pass range, the academic warning threshold probability, and the compensation value. Determine whether the first academic result is within the course data pass range. If not, proceed to step S520; if yes, do not output an academic warning report. S520, determine whether the probability of the second academic result is lower than the academic warning threshold; if so, proceed to step S521; S521, Generate an academic warning report for the student based on the first and second academic results. The academic warning report includes the student's current learning behavior data in the first and second phases of the current academic year, future course data, future academic warning probability, and the student's personal information. If the second academic result is not lower than the academic warning threshold probability, proceed to step S522; S522, the absolute values ​​of the first academic result and the second academic result are multiplied and the compensation result is output. When the compensation result is lower than the compensation value, an academic warning report for the student is generated.

2. The academic early warning method based on meta-learning and support vector description according to claim 1, characterized in that, The specific steps in step S100, which involve pre-training the initial academic prediction model based on the cross-task dataset to obtain the base academic prediction model, are as follows: S110, Initialize the network parameters of the initial academic prediction model. Given N training task classifiers, set the network parameters... The network structure assigned to these N training task classifiers ,Right now ; S120, divide the cross-task dataset into N groups of training data. Each group of training data has its corresponding training task classifier. Each group of training data is divided into a query set and a support set. S130, each training task classifier is trained on its corresponding support set, and the total loss function of all training task classifiers is calculated based on the query set corresponding to each training task classifier. The network parameters of the initial academic prediction model are updated using gradient descent and the total loss function. S140, Set the expected value range of the network parameters of the initial academic prediction model, and determine whether the network parameters of the initial academic prediction model are within the expected value range. If so, proceed to step S141. S141, Output the academic prediction base model; If the network parameters of the initial academic prediction model are outside the predicted value range, proceed to step S142; S142, repeat steps S130 to S140.

3. The academic early warning method based on meta-learning and support vector description according to claim 2, characterized in that, Step S130 specifically includes: S131: Select n training task classifiers from N training task classifiers through random sampling. Train each of the n sampled training task classifiers on its corresponding support set and update their network parameters once using gradient descent. for ; S132, based on the network parameters obtained in step S131 These n training task classifiers are tested on their corresponding query sets, and the first sub-loss function corresponding to each training task classifier is calculated. ; S133, the total loss function is calculated based on all the first sub-loss functions; wherein the total loss function satisfies the following formula: in, Represents the total loss function; S134, Update the network parameters of the initial academic prediction model based on the total loss function. The network parameters of the updated academic prediction initial model satisfy the following formula: in, This represents the learning rate of the initial model for academic prediction. Represents the total loss function Network parameters of the initial model for academic prediction The derivative of .

4. The academic early warning method based on meta-learning and support vector description according to claim 1, characterized in that, The steps in step S400 to establish the academic prediction auxiliary model are as follows: S410: Obtain sample behavioral data of students with academic warnings in the second phase of the historical academic year, perform PCA analysis and preprocessing on the sample behavioral data to obtain an auxiliary sample set; The preprocessing includes: the sample behavior data is of multiple categories, each with different attributes, and a normalization method is used to uniformly measure the sample behavior data of all categories. S420, the auxiliary sample set is divided into a training sample set through random sampling. and test sample set A prediction model is built based on a support vector classifier, using the training sample set as input to the prediction model, and then through a mapping function. training sample set Mapping to a high-dimensional feature space; wherein, the mapping function F is a Gaussian kernel function, which satisfies the following formula: Where y represents the center of the Gaussian kernel function, The parameter representing the width of the Gaussian kernel function; S430, Set the slack variables for the support vector description classifier With the penalty parameter C, the support vectors describe how the classifier gradually forms a hypersphere capable of describing the auxiliary sample set during training; wherein, the hypersphere satisfies the following formula: Where R represents the radius of the hypersphere. Let C represent the center of the hypersphere, and let C represent the penalty parameters of the support vectors describing the classifier. Represents slack variables. Let represent the course data of the i-th sample in the auxiliary sample set, and st denotes constraints; The hypersphere is Centered on the training sample set Data field; S440, introducing the Lagrange operator, constructing the Lagrange function, and combining the Lagrange function and the KKT conditions, obtain the optimal hypersphere and its corresponding center and radius; wherein, the optimal hypersphere satisfies the following formula: Where L represents the optimal hypersphere. , All are Lagrange operators; The center of the optimal hypersphere satisfies the following formula: The radius of the optimal hypersphere is calculated from its center, and its radius satisfies the following formula: in, Represents any data point in the training sample set; S450: Evaluate the performance of the academic prediction auxiliary model based on the test sample set, obtain the model accuracy of the academic prediction auxiliary model, set the expected value of the model accuracy, and determine whether the model accuracy is less than its expected value. If the model accuracy is less than the expected value, then repeat steps S420 to S450. If the model accuracy is not less than the expected value, then output the academic prediction auxiliary model.

5. The academic early warning method based on meta-learning and support vector description according to claim 4, characterized in that, Step S400 involves obtaining students' current learning behavior data for the second phase of the current academic year and inputting it into the academic prediction auxiliary model, including: The current behavioral data is input into the academic prediction auxiliary model to obtain the support vectors of the current behavioral data in the hypersphere state of the academic prediction auxiliary model; wherein, the support vectors satisfy the following formula: in, This represents current behavioral data; Determine whether the support vector is located within the hypersphere of the academic prediction auxiliary model, calculate the academic warning probability, and output the second academic result.

6. The academic early warning method based on meta-learning and support vector description according to claim 5, characterized in that, The academic warning probability satisfies the following formula: in, Indicates the probability of academic warning, when When it means that the support vector is located outside the hypersphere; when The time indicates that the support vector is located within the hypersphere.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of an academic early warning method based on meta-learning and support vector description as described in any one of claims 1-6.

8. An academic early warning system based on meta-learning and deep learning, characterized in that, The system includes: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements an academic early warning method based on meta-learning and support vector description as described in any one of claims 1-6.

Images