Accurate test question pushing system based on learning behavior analysis

Abstract

The invention discloses an accurate test question pushing system based on learning behavior analysis, and particularly relates to the technical field of accurate test question recommendation, which comprises a learning ability dynamic evaluation module, a multi-task model refining module, a speculation and complementation module and a self-adaptive recommendation module, the learning ability dynamic evaluation module constructs a time sequence model according to the learning behavior data of the user to predict the future learning ability of the user for primary classification, and the multi-task model refining module refines the primary classification result; the speculation and completion module generates learning behavior characteristics by combining known learning behavior data based on a graph neural network, and the self-adaptive recommendation module performs personalized recommendation on different types of users according to an output result of the speculation and completion module, evaluates the learning ability of the users by using a time sequence model, and performs refined classification by combining a multi-task learning model, so that the learning ability of the users is improved. Personalized test question recommendation is achieved by speculating potential learning behaviors through the graph neural network, the problem of data sparsity is effectively solved, and recommendation accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of Internet of Things, and more specifically, to a test question accurate push system based on learning behavior analysis. Background Art

[0002] The potential subjects for accurate test question push are usually arranged in a normal distribution. Students with extremely strong and extremely weak learning abilities account for only a very small number. The needs of a very small number of users have the problem of data sparsity, making it difficult to perform accurate and personalized push.

[0003] To address the problem of data sparsity, the current common method is to adopt a cold start solution and conduct trial processing based on the personal experience and recommended rules of industry experts. This method is highly subjective and has high requirements on the breadth of rules. It is still difficult to adapt to a small number of extreme users. Another common method is usually to extract and transfer existing data from other fields to the target domain based on transfer learning. However, there are differences between the source domain and the target domain. Smoothing out the differences and mastering the behavioral logic of learning data is a daunting challenge in practical applications.

[0004] In order to solve the above defects, a technical solution is now proposed. Summary of the invention

[0005] The purpose of the present invention is to provide a test question accurate push system based on learning behavior analysis to solve the deficiencies in the background technology.

[0006] In order to achieve the above object, the present invention provides the following technical solutions: a test question accurate push system based on learning behavior analysis, including a learning ability dynamic evaluation module, a multi-task model refinement module, a speculation completion module, and an adaptive recommendation module;

[0007] The learning ability dynamic evaluation module is used to build a time series model based on the user's learning behavior data to predict the user's future learning ability, and classify the user's learning status based on the user's future learning ability and the user's current learning performance;

[0008] The multi-task model refinement module is used to refine the results of a classification, and to establish a multi-task learning model after labeling different questions. The multi-task learning model includes a first output layer for predicting the question required by the user, a second output layer for classifying the question type, and a third output layer for evaluating the difficulty of the question. The target users with extremely strong and extremely weak learning abilities are marked according to the output results of the multi-task shared underlying network.

[0009] The inference completion module is used to infer the potential learning behavior of the target user through a graph neural network-based method, and generate learning behavior features based on known learning behavior data;

[0010] The adaptive recommendation module is used to make personalized recommendations for different types of users based on the output results of the inference completion module.

[0011] Preferably, a feature vector representing the change in user learning ability is constructed based on time series extraction features, and the feature vector is constructed by:

[0012] The number of answers within the calibration time window t is N t , the correct answer rate in the time window t is A t , A t The calculation expression is In the formula, C t is the number of correct answers in time window t, and the average answering time is T t , the calculation expression is Among them, t i is the answering time for each question, and the review frequency of wrong questions is F t , the calculation expression is Learning progress rate In the formula, δ and ∈ are both extremely small positive numbers, A t-1 is the correct answer rate in the previous time window, and the learning stability is S t , the calculation expression is In the formula, n is the amount of learning performance data, x i is the i-th learning achievement, and the feature vector v represents the change of user learning ability t The form is v t =[N t , A t , T t , F t , P t , S t ].

[0013] Preferably, the feature vector v t After inputting the time series model, the prediction of the user's future learning ability is output. The logistic regression model is used to classify the users according to the prediction of different users' future learning abilities. The specific method is as follows:

[0014] Building a logistic regression model In the formula, P(y=1|X) is the classification bias value of the user, y is a binary dependent variable, which represents the judgment category of the user's learning ability, and x is an independent variable vector, including the output prediction of the time series model, the user's learning performance and various features, β0, β1, ...β m is the coefficient of the logistic regression model, m is the number of independent variable vectors, and m is a positive integer;

[0015] Set the first classification threshold Pf and the first classification second threshold P s , and the first classification threshold P f Less than the second threshold P of the first classification s , when the calculated user classification deviation value P is less than or equal to the first classification threshold P f , the user is marked as a Class A target user. When the calculated user classification deviation value P is greater than or equal to the first classification second threshold P s , the user is marked as a Class B target user. When the calculated user classification deviation value P is greater than the first classification threshold P f And less than the second threshold P of the first classification s , mark the user as a non-target user.

[0016] Preferably, the logic of establishing a multi-task learning model is:

[0017] Set the multi-task learning model as Z = f(X u , X q ), where Z is the shared eigenvector, X u is a user feature, X q is the question feature, f is the underlying shared network, the first output layer of the multi-task learning model is used to predict the user's required question, and the loss function is the cross entropy function, expressed as In the formula, y j For target classification, p j is the predicted category probability, B is the number of questions, and the second output layer of the multi-task learning model is used to classify questions according to the question labels and determine the question type. The loss function is the cross entropy function, which is expressed as In the formula, E k is the true label of the question type, q k is the predicted question type probability, V is the number of question types, the third output layer of the multi-task learning model is used to predict the difficulty of the question and output the difficulty value of the question. The loss function is the mean square error loss, expressed as Where D h is the actual difficulty value of the question, is the predicted difficulty value, and the overall loss function is the weighted sum of L1, L2 and L3.

[0018] Preferably, the prediction result of the user's learning ability by the multi-task learning model is classified and refined based on the first classification as follows:

[0019] Define the explicit capability function S(X u )The underlying network output based on the multi-task learning model is represented as S(X u ) = f cl (Z u ), where Zu is the user feature extracted from the shared network, f cl is the classification function, and the explicit ability classification threshold is set to E T , then the output type in Among them, T y =1 means the user has a strong learning ability, T y =0 indicates that the user's learning ability is extremely weak. For target users of type A, when the output type is T y = 0, the user is marked as a user with extremely weak explicit ability. When the output type is T y =1, the user is marked as an invisible floating user;

[0020] For Class B users, when the output type is T y =1, marking the user as a user with strong explicit ability. When the user's output type is T y =0, the user is marked as an invisible floating user.

[0021] Preferably, the logic for generating learning behavior features based on graph neural network is:

[0022] The nodes of the graph include user nodes, question nodes, and knowledge point nodes. User nodes are used to represent individual users, question nodes are used to represent questions, and knowledge point nodes are used to represent knowledge points involved in questions. The edges of the graph include user-question edges and question-knowledge point edges. User-question edges are used to represent the relationship between questions that users have done, and the weights of user-question edges are assigned according to the accuracy of the user's answers and the answering time. Question-knowledge point edges are used to represent the knowledge points associated with the questions, and the weights of question-knowledge point edges are assigned according to the difficulty of the questions and the importance of the knowledge points. The graph is represented by G=(W, A), where W is the node set of the graph, which includes users, questions, and knowledge points, and A is the edge set of the graph, which includes the association between users and questions, and the association between questions and knowledge points. Information is propagated through the connections between nodes, the feature representation of each node is updated, and node features are generated using known learning behavior data.

[0023] Preferably, the logic for recommending test questions based on the output of the graph neural network is:

[0024] The convolution of the graph neural network integrates the user's learning behavior data and the relationship between questions and knowledge points, and trains the graph neural network to obtain the user's feature representation as h u , the expression is h u =GNN(X u ), where X u is the user’s feature input, GNN represents graph neural network, h uIt is the user feature representation after graph convolution. For users with extremely strong explicit abilities, a regression model is used to predict the user's future learning ability performance and recommend test questions. For users with extremely weak explicit abilities, a classification model is used to predict the user's learned knowledge and recommend test questions.

[0025] Preferably, the method of using a regression model to predict the user's future learning ability performance and recommend test questions is:

[0026] Define the regression model as To predict users with strong explicit abilities, the difficulty of the recommended questions is dynamically adjusted based on the user's learning ability performance predicted by the regression model. The adjustment logic is: In the formula, To recommend the difficulty of the test questions, A function that adjusts the difficulty of questions based on the user's predicted learning performance.

[0027] Preferably, the method of using the classification model to predict the user's learned knowledge and recommend test questions is:

[0028] Define the classification model as P(y|h u )=softmax(W2·σ(W1·h u +b1)+b2), where W1 and W2 are the weight matrices of the classification model, σ is the activation function, and the output is passed through the softmax function to obtain the category distribution that the user has learned. u and the characteristics of the topic q Match and calculate the matching similarity to recommend test questions. The matching similarity is calculated as follows: The higher the matching similarity, the higher the recommendation priority.

[0029] In the above technical solution, the technical effects and advantages provided by the present invention are:

[0030] This application solves the data sparsity problem of extreme users by dual classification of user properties. It includes dynamic evaluation of learning ability, refinement of multi-task model, inference completion module and adaptive recommendation module. It builds a time series model through learning behavior data to predict users' future learning ability, and refines user classification in combination with multi-task learning model. It uses graph neural network to infer potential learning behavior, generates personalized recommendations, evaluates learning ability in real time, provides accurate test question recommendations, uses multi-task learning to improve evaluation accuracy and avoid repeated training, and uses graph neural network to capture complex learning behavior relationships and improve the quality of test question recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0032] Figure 1 It is a system module diagram of the present invention. DETAILED DESCRIPTION

[0033] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0034] Example 1: Please refer to Figure 1 As shown, the present invention is a test question accurate push system based on learning behavior analysis, including a learning ability dynamic evaluation module, a multi-task model refinement module, a speculation completion module, and an adaptive recommendation module;

[0035] The learning ability dynamic evaluation module is used to build a time series model based on the user's learning behavior data to predict the user's future learning ability, and classify the user's learning status based on the user's future learning ability and the user's current learning performance;

[0036] The multi-task model refinement module is used to refine the results of a classification, and to establish a multi-task learning model after labeling different questions. The multi-task learning model includes a first output layer for predicting the question required by the user, a second output layer for classifying the question type, and a third output layer for evaluating the difficulty of the question. The target users with extremely strong and extremely weak learning abilities are marked according to the output results of the multi-task shared underlying network.

[0037] The inference completion module is used to infer the potential learning behavior of the target user through a graph neural network-based method, and generate learning behavior features based on known learning behavior data;

[0038] The adaptive recommendation module is used to make personalized recommendations for different types of users based on the output results of the inference completion module.

[0039] The learning ability dynamic evaluation module is used to build a time series model based on the user's learning behavior data to predict the user's future learning ability, and classify the user's learning status based on the user's future learning ability and the user's current learning performance;

[0040] The user's learning behavior data includes the number of answers, accuracy, answering time, and frequency of reviewing wrong questions. A timestamp is added to each learning behavior of the user, and the learning behavior data is converted into time series data. A time window is defined, and the time window is a period of learning time. In each window, the change trend of the learning behavior data is analyzed, and a feature vector representing the change of the user's learning ability is constructed based on the time series extraction features. The method for constructing the feature vector is as follows:

[0041] The number of answers within the calibration time window t is N t , the correct answer rate in the time window t is A t , A t The calculation expression is In the formula, C t is the number of correct answers in time window t, and the average answering time is T t , the calculation expression is Among them, t i is the answering time for each question, and the review frequency of wrong questions is F t , the calculation expression is Learning progress rate In the formula, δ and ∈ are both extremely small positive numbers, A t-1 is the correct answer rate in the previous time window, and the learning stability is S t , the calculation expression is In the formula, n is the amount of learning performance data, x i is the i-th learning achievement, and the feature vector v represents the change of user learning ability t The form is v t =[N t , A t , T t , F t , P t , S t ].

[0042] The feature vector v t After inputting the time series model, the prediction of the user's future learning ability is output. The logistic regression model is used to classify the users according to the prediction of different users' future learning abilities. The specific method is as follows:

[0043] Building a logistic regression model In the formula, P(y=1|X) is the classification bias value of the user, y is a binary dependent variable, which represents the judgment category of the user's learning ability, and x is an independent variable vector, including the output prediction of the time series model, the user's learning performance and various features, β0, β1, ...β m is the coefficient of the logistic regression model, m is the number of independent variable vectors, and m is a positive integer;

[0044] Set the first classification threshold P f and the first classification second threshold P s , and the first classification threshold P f Less than the second threshold P of the first classification s , when the calculated user classification deviation value P is less than or equal to the first classification threshold P f , the user is marked as a Class A target user. When the calculated user classification deviation value P is greater than or equal to the first classification second threshold P s , the user is marked as a Class B target user. When the calculated user classification deviation value P is greater than the first classification threshold P f And less than the second threshold P of the first classification s , mark the user as a non-target user.

[0045] The learning ability of target users of type A is low and their learning status is poor, while the learning ability of target users of type B is strong and their learning status is good. Both types of target users are extreme user types with data sparsity problems. The existing data may not be able to meet the recommendation needs of extreme users. The fundamental reason is that the extreme user types with very few existing data are limited by data sparsity.

[0046] The traditional recommendation algorithm layout logic is solved by cold start processing or transfer learning. However, in question recommendation, the interaction of question recommendation and recycling evaluation always exists. Users who do not have the need for question feedback will not continuously take questions. The demand does exist, but the stock of extreme users is too scarce to be analyzed through sufficient learning behavior data. The transfer learning solution has high requirements for the source domain and target domain of the transfer learning application, and it is difficult to implement in actual application.

[0047] Time series analysis can capture the trends and periodicity of learning behaviors over time, reveal the inherent laws in the learning process, and generate personalized learning paths and test question recommendations based on the results of time series analysis. Real-time evaluation of changes in learning ability helps to adjust the content of test question recommendations in a timely manner. On the other hand, the learning abilities of different users may not remain constant. There is a possibility that the learning ability performance of some users may change from a lower level to a higher level or from a higher level to a lower level. Therefore, real-time evaluation of users is necessary, and therefore users must be doubly classified.

[0048] The multi-task model refinement module is used to refine the results of a classification, and to establish a multi-task learning model after labeling different questions. The multi-task learning model includes a first output layer for predicting the question required by the user, a second output layer for classifying the question type, and a third output layer for evaluating the difficulty of the question. The target users with extremely strong and extremely weak learning abilities are marked according to the output results of the multi-task shared underlying network.

[0049] The multi-task learning model refines the evaluation of different users' learning abilities and labels the questions. The model should include three tasks: predicting user-required questions, classifying the question types, and evaluating the difficulty of the questions. Each task shares the feature representation of the underlying network and is trained together through the multi-task learning framework. The underlying network of the multi-task learning model includes three input layers:

[0050] User feature input: user learning behavior data, including accuracy, answering speed, and knowledge points learned;

[0051] Question feature input: question label information, including question type, knowledge point, and question difficulty;

[0052] Historical learning path: The user’s learning path sequence is input as a time series;

[0053] The logic of building a multi-task learning model is:

[0054] Set the multi-task learning model as Z = f(X u , X q ), where Z is the shared eigenvector, X u is a user feature, X q is the question feature, f is the underlying shared network, the first output layer of the multi-task learning model is used to predict the user's required question, and the loss function is the cross entropy function, expressed as In the formula, y j For target classification, p j is the predicted category probability, B is the number of questions, and the second output layer of the multi-task learning model is used to classify questions according to the question labels and determine the question type. The loss function is the cross entropy function, which is expressed as In the formula, E k is the true label of the question type, q k is the predicted question type probability, V is the number of question types, the third output layer of the multi-task learning model is used to predict the difficulty of the question and output the difficulty value of the question. The loss function is the mean square error loss, expressed as Where D h is the actual difficulty value of the question, is the predicted difficulty value, and the overall loss function is the weighted sum of L1, L2 and L3;

[0055] Based on the prediction results of user learning ability by the multi-task learning model, the target users of category A and category B obtained in one classification are screened, and the users with extremely strong and extremely weak learning abilities are classified and refined. The logic of classification refinement is as follows:

[0056] Define the explicit capability function S(X u ) is the underlying network output representation based on the multi-task learning model, S(X u ) = f cl (Z u ), where Z u is the user feature extracted from the shared network, f cl is the classification function, and the explicit ability classification threshold is set to E T , then the output type in, Among them, T y =1 means the user has a strong learning ability, T y =0 indicates that the user's learning ability is extremely weak. For target users of type A, when the output type is T y = 0, the user is marked as a user with extremely weak explicit ability. When the output type is T y =1, the user is marked as an invisible floating user;

[0057] For Class B users, when the output type is T y =1, marking the user as a user with strong explicit ability. When the user's output type is T y =0, the user is marked as an invisible floating user.

[0058] It should be noted that the weights of each loss function are determined by professionals in this field through hyperparameter tuning.

[0059] The multi-task learning model conducts targeted training based on different question types and difficulty levels to make learning more accurate and efficient. Compared with the traditional single-task learning model, multi-task learning can avoid repeated training, save time and resources, and evaluate the user's learning ability from multiple dimensions, including the mastery of different question types, problem-solving speed, accuracy, etc. The multi-dimensional evaluation can more comprehensively understand the user's learning status. Multi-task learning enables the model to learn more general and robust feature representations by sharing the representation layer between different tasks, and the auxiliary tasks have a regularization effect on the main task, reducing overfitting and improving the stability and reliability of the model.

[0060] The inference completion module is used to infer the potential learning behavior of the target user through a graph neural network-based method, and generate learning behavior features based on known learning behavior data;

[0061] Based on the graph neural network, a learning behavior graph is constructed. Through the learning behavior graph, users with strong explicit ability and users with weak explicit ability are connected with the question and knowledge point entities, so that the learning behavior is abstracted into a graph model. The construction logic of the graph neural network is as follows:

[0062] The nodes of the graph include user nodes, question nodes, and knowledge point nodes. User nodes are used to represent individual users, question nodes are used to represent questions, and knowledge point nodes are used to represent knowledge points involved in questions. The edges of the graph include user-question edges and question-knowledge point edges. User-question edges are used to represent the relationship between questions that users have done, and the weights of user-question edges are assigned according to the accuracy of the user's answers and the answering time. Question-knowledge point edges are used to represent the knowledge points associated with the questions, and the weights of question-knowledge point edges are assigned according to the difficulty of the questions and the importance of the knowledge points. The graph is represented by G=(W, A), where W is the node set of the graph, which includes users, questions, and knowledge points, and A is the edge set of the graph, which includes the association between users and questions, and the association between questions and knowledge points. Information is propagated through the connections between nodes, the feature representation of each node is updated, and node features are generated using known learning behavior data.

[0063] The adaptive recommendation module is used to make personalized recommendations for different types of users based on the output results of the inference completion module.

[0064] The convolution of the graph neural network integrates the user's learning behavior data and the relationship between questions and knowledge points, and trains the graph neural network to obtain the user's feature representation as h u , the expression is h u =GNN(X u ), where X u is the user’s feature input, GNN represents graph neural network, h u It is the user feature representation after graph convolution. For users with strong explicit abilities, a regression model is used to predict the user's future learning ability performance and recommend test questions. For users with weak explicit abilities, a classification model is used to predict the user's learned knowledge and recommend test questions.

[0065] The method of using regression model to predict the user's future learning ability performance and recommend test questions is:

[0066] Define the regression model as To predict users with strong explicit abilities, the difficulty of the recommended questions is dynamically adjusted based on the user's learning ability performance predicted by the regression model. The adjustment logic is: In the formula, To recommend the difficulty of the test questions, A function that adjusts the difficulty of questions based on the user's predicted learning performance;

[0067] The method of using the classification model to predict the user's learned knowledge and recommend test questions is:

[0068] Define the classification model as P(y|h u )=softmax(W2·σ(W1·h u +b1)+b2), where W1 and W2 are the weight matrices of the classification model, σ is the activation function, and the output is passed through the softmax function to obtain the category distribution that the user has learned. u and the characteristics of the topic q Match and calculate the matching similarity to recommend test questions. The matching similarity is calculated as follows: The higher the matching similarity, the higher the recommendation priority.

[0069] This application solves the data sparsity problem of extreme users by dual classification of user properties. It includes dynamic evaluation of learning ability, refinement of multi-task model, inference completion module and adaptive recommendation module. It builds a time series model through learning behavior data to predict users' future learning ability, and refines user classification in combination with multi-task learning model. It uses graph neural network to infer potential learning behavior, generates personalized recommendations, evaluates learning ability in real time, provides accurate test question recommendations, uses multi-task learning to improve evaluation accuracy and avoid repeated training, and uses graph neural network to capture complex learning behavior relationships and improve the quality of test question recommendations.

[0070] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0071] The above embodiments can be implemented in whole or in part by software, hardware, firmware or any other combination. When implemented by software, the above embodiments can be implemented in whole or in part in the form of computer program goods. The computer program goods include one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from one website site, computer, server or data center to another website site, computer, server or data center by wired or wireless (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that contains one or more available media sets. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state hard disk.

[0072] It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0073] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0074] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the system described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0075] If the functions are implemented in the form of software functional units and sold or used as independent goods, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application can essentially or in other words, the part that contributes to the prior art or the part of the technical solution can be embodied in the form of software goods, and the computer software goods are stored in a storage medium, including several instructions for a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, and other media that can store program codes.

[0076] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. The test question accurate push system based on learning behavior analysis is characterized by: It includes a learning ability dynamic evaluation module, a multi-task model refinement module, a speculation completion module, and an adaptive recommendation module; The learning ability dynamic evaluation module is used to build a time series model based on the user's learning behavior data to predict the user's future learning ability, and classify the user's learning status based on the user's future learning ability and the user's current learning performance; The multi-task model refinement module is used to refine the results of a classification, and to establish a multi-task learning model after labeling different questions. The multi-task learning model includes a first output layer for predicting the question required by the user, a second output layer for classifying the question type, and a third output layer for evaluating the difficulty of the question. The target users with extremely strong and extremely weak learning abilities are marked according to the output results of the multi-task shared underlying network. The inference completion module is used to infer the potential learning behavior of the target user through a graph neural network-based method, and generate learning behavior features based on known learning behavior data; The adaptive recommendation module is used to make personalized recommendations for different types of users based on the output results of the inference completion module.

2. The test question accurate push system based on learning behavior analysis according to claim 1 is characterized in that: Based on the time series extraction features, a feature vector representing the change in user learning ability is constructed. The method for constructing the feature vector is: The number of answers within the calibration time window t is N t , the correct answer rate in the time window t is A t , A t The calculation expression is In the formula, C t is the number of correct answers in time window t, and the average answering time is T t , the calculation expression is Among them, t i is the answering time for each question, and the review frequency of wrong questions is F t , the calculation expression is Learning progress rate In the formula, δ and ∈ are both extremely small positive numbers, A t-1 is the correct answer rate in the previous time window, and the learning stability is S t , the calculation expression is In the formula, n is the amount of learning performance data, x i is the i-th learning achievement, then the feature vector v representing the change in the user's learning ability t The form is v t =[N t ,A t ,T t ,F t ,P t ,S t ].

3. The test question accurate push system based on learning behavior analysis according to claim 2 is characterized in that: The feature vector v t After inputting the time series model, the prediction of the user's future learning ability is output. The logistic regression model is used to classify the users according to the prediction of different users' future learning abilities. The specific method is as follows: Building a logistic regression model In the formula, P(y=1|X) is the classification bias value of the user, y is a binary dependent variable, which represents the judgment category of the user's learning ability, and x is an independent variable vector, including the output prediction of the time series model, the user's learning performance and various features, β0,β1,…β m is the coefficient of the logistic regression model, m is the number of independent variable vectors, and m is a positive integer; Set the first classification threshold P f and the first classification second threshold P s , and the first classification threshold P f Less than the second threshold P of the first classification s , when the calculated user classification deviation value P is less than or equal to the first classification threshold P f , the user is marked as a Class A target user. When the calculated user classification deviation value P is greater than or equal to the first classification second threshold P s , the user is marked as a Class B target user. When the calculated user classification deviation value P is greater than the first classification threshold P f And less than the second threshold P of the first classification s , mark the user as a non-target user.

4. The test question accurate push system based on learning behavior analysis according to claim 1 is characterized in that: The logic of building a multi-task learning model is: Set the multi-task learning model as Z = f(X u ,X q ), where Z is the shared eigenvector, X u is a user feature, X q is the question feature, f is the underlying shared network, the first output layer of the multi-task learning model is used to predict the user's required question, and the loss function is the cross entropy function, expressed as In the formula, y j For target classification, p j is the predicted category probability, B is the number of questions, and the second output layer of the multi-task learning model is used to classify questions according to the question labels and determine the question type. The loss function is the cross entropy function, which is expressed as In the formula, E k is the real label of the question type, q k is the predicted question type probability, V is the number of question types, the third output layer of the multi-task learning model is used to predict the difficulty of the question and output the difficulty value of the question. The loss function is the mean square error loss, expressed as Where D h is the actual difficulty value of the question, is the predicted difficulty value, and the overall loss function is the weighted sum of L1, L2 and L3.

5. The test question accurate push system based on learning behavior analysis according to claim 4 is characterized in that: The prediction results of user learning ability by multi-task learning model are classified into the following logic based on the first classification: Define the explicit capability function S(X u )The underlying network output based on the multi-task learning model is represented as S(X u ) = f cl (Z u ), where Z u is the user feature extracted from the shared network, f cl is the classification function, and the explicit ability classification threshold is set to E T , then the output type in Among them, T y =1 means the user has a strong learning ability, T y =0 indicates that the user's learning ability is extremely weak. For target users of type A, when the output type is T y = 0, the user is marked as a user with extremely weak explicit ability. When the output type is T y =1, the user is marked as an invisible floating user; For Class B users, when the output type is T y =1, marking the user as a user with strong explicit ability. When the user's output type is T y =0, the user is marked as an invisible floating user.

6. The test question accurate push system based on learning behavior analysis according to claim 1 is characterized in that: The logic of generating learning behavior features based on graph neural networks is: The nodes of the graph include user nodes, question nodes, and knowledge point nodes. User nodes are used to represent individual users, question nodes are used to represent questions, and knowledge point nodes are used to represent knowledge points involved in questions. The edges of the graph include user-question edges and question-knowledge point edges. User-question edges are used to represent the relationship between questions that users have done. The weight of the user-question edge is assigned according to the user's answer accuracy and answering time. Question-knowledge point edges are used to represent the knowledge points associated with the questions. The weight of the question-knowledge point edge is assigned according to the difficulty of the question and the importance of the knowledge point. The graph is represented by G=(W,A), where W is the node set of the graph, which includes users, questions and knowledge points, and A is the edge set of the graph, which includes the association between users and questions, as well as the association between questions and knowledge points. Information is propagated through the connections between nodes, the feature representation of each node is updated, and node features are generated using known learning behavior data.

7. The test question accurate push system based on learning behavior analysis according to claim 6 is characterized in that: The logic of recommending test questions based on the output of the graph neural network is: The convolution of the graph neural network integrates the user's learning behavior data and the relationship between questions and knowledge points, and trains the graph neural network to obtain the user's feature representation as h u , the expression is h u =GNN(X u ), where X u is the user’s feature input, GNN represents graph neural network, h u It is the user feature representation after graph convolution. For users with extremely strong explicit abilities, a regression model is used to predict the user's future learning ability performance and recommend test questions. For users with extremely weak explicit abilities, a classification model is used to predict the user's learned knowledge and recommend test questions.

8. The test question accurate push system based on learning behavior analysis according to claim 7 is characterized in that: The method of using regression model to predict the user's future learning ability performance and recommend test questions is: Define the regression model as To predict users with strong explicit abilities, the difficulty of the recommended questions is dynamically adjusted based on the user's learning ability performance predicted by the regression model. The adjustment logic is: In the formula, To recommend the difficulty of the test questions, A function that adjusts the difficulty of questions based on the user's predicted learning performance.

9. The test question accurate push system based on learning behavior analysis according to claim 7 is characterized in that: The method of using the classification model to predict the user's learned knowledge and recommend test questions is: Define the classification model as P(y|h u )=softmax(W2·σ(W1·h u +b1)+b2), where W1 and W2 are the weight matrices of the classification model, σ is the activation function, and the output is passed through the softmax function to obtain the category distribution that the user has learned. u and the characteristics of the topic q Match and calculate the matching similarity to recommend test questions. The matching similarity is calculated as follows: The higher the matching similarity, the higher the recommendation priority.

Images