Deep learning-based personalized learning scheme generation and optimization method for college entrance examination

Abstract

The invention discloses a method for generating and optimizing a personalized learning scheme for college entrance examination based on deep learning, and aims to solve the problems of large learning difference of students, difficulty in personalized tutoring and low utilization rate of learning resources. Comprising the following steps: S1, collecting multi-source learning data of students through a multi-platform data interface and an embedded tracking technology; s2, preprocessing the multi-source learning data set to obtain a standardized learning data set; s3, constructing a deep learning model to extract student learning feature parameter vectors; s4, generating a student personalized learning scheme through a personalized recommendation algorithm; s5, collecting student learning process data in real time; s6, dynamically adjusting and optimizing the personalized learning scheme in combination with a feedback optimization mechanism; and S7, performing quantitative evaluation and generating a learning effect report. According to the method, intelligent and adaptive personalized learning recommendation and dynamic optimization can be realized, and the scientificity and learning efficiency of college entrance examination preparation are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of educational informatization and artificial intelligence technology, in particular to a personalized learning scheme generation and optimization method based on deep learning for middle and high school examinations, which is suitable for application scenarios such as personalized resource recommendation, dynamic optimization of learning process, and learning effectiveness evaluation in the process of middle and high school examination preparation. BACKGROUND

[0002] With the continuous development of educational informatization and artificial intelligence technology, data-driven personalized learning has become an important research direction in the field of current middle and high school examination preparation. Existing middle and high school learning schemes still rely mainly on traditional teaching experience and static score analysis, which is difficult to fully tap the multi-dimensional behavior data and individual differences of students in the learning process. Traditional learning management systems and online education platforms mainly focus on exercise question pushing, error question sorting and stage score feedback, but they lack in systematic collection, dynamic analysis and personalized modeling of students' overall learning data, resulting in homogenization of learning resource recommendation and lack of strong targeting, which makes it difficult to fully meet the actual needs of different students.

[0003] In the prior art, although some platforms can achieve a certain degree of personalized learning resource pushing by using big data analysis, they usually only rely on students' historical scores or single behavior sequences, ignoring the multi-dimensional characteristics such as operation details, knowledge point mastery level and interest preferences of students in different learning scenarios. At the same time, the data processing process mainly uses manual rules or simple algorithms, which lack efficient data cleaning, anomaly detection and standardization means, resulting in a lot of data noise and affecting the accuracy of subsequent analysis. In addition, the existing recommendation algorithm mainly focuses on collaborative filtering or content recommendation, and the model structure is relatively simple, which makes it difficult to effectively capture the dynamic changes of students' knowledge mastery, interest shifts and evolution of weak links, thereby limiting the scientificity and effectiveness of personalized learning schemes.

[0004] In terms of dynamic optimization, traditional methods mainly rely on stage test results to adjust the learning scheme, lack real-time tracking and feedback of students' learning process, and are difficult to dynamically optimize the learning path and resource recommendation according to the actual learning performance and feedback of students, resulting in a lag between the learning scheme and the actual needs of students, which affects the learning efficiency and effectiveness. At the same time, the existing learning effect evaluation often only stays in single indicators such as accuracy statistics and subjective evaluation, lacking multi-dimensional quantitative analysis of knowledge point mastery improvement and learning efficiency, which is difficult to fully reflect the actual implementation effect of personalized learning scheme and provide strong teaching decision support for teachers and parents.

[0005] In summary, the existing personalized learning scheme generation and optimization methods for middle and high school examinations have deficiencies in deep feature modeling, dynamic scheme adjustment, and multi-dimensional effectiveness evaluation, making it difficult to achieve efficient, accurate, and dynamic personalized learning support under the whole-process data-driven mode. Therefore, a new method that can integrate deep learning modeling, dynamic recommendation optimization, and multi-dimensional effectiveness evaluation is urgently needed to better meet the personalized preparation needs of middle and high school students and comprehensively improve learning efficiency and effectiveness. SUMMARY

[0006] One object of the present application is to provide a deep learning-based personalized learning scheme generation and optimization method for middle and high school examinations. The present application makes full use of deep learning modeling, data mining, and personalized recommendation technology, systematically collects and processes students' multi-source learning data, automatically extracts learning features, dynamically generates and optimizes personalized learning schemes, and evaluates learning effectiveness in multiple dimensions. It can realize real-time tracking and adaptive recommendation of students' learning process, has the advantages of high personalization, fast response, and significant optimization effect, and can effectively improve the learning efficiency and preparation effectiveness of middle and high school students.

[0007] According to the deep learning-based personalized learning scheme generation and optimization method for middle and high school examinations, the following steps are included:

[0008] S1, collecting students' multi-source learning data through multi-platform data interfaces and embedded tracking technology to form a multi-source learning data set;

[0009] S2, data cleaning, outlier detection, and standardization processing of the multi-source learning data set to obtain a standardized learning data set;

[0010] S3, based on the middle and high school knowledge system and students' learning goals, constructing and training a student feature extraction deep learning model, and inputting the standardized learning data set into the model to extract student learning feature parameter vectors;

[0011] S4, generating a student personalized learning scheme based on the student learning feature parameter vectors and the middle and high school knowledge system through a personalized recommendation algorithm;

[0012] S5, collecting students' learning process data in real time during the execution of the personalized scheme to form a learning process data sequence;

[0013] S6, inputting the learning process data sequence into the student feature extraction deep learning model, and dynamically adjusting and optimizing the personalized learning scheme in combination with a feedback optimization mechanism;

[0014] S7, quantifying and evaluating the execution effect of the student personalized learning scheme, and generating a learning effect report.

[0015] Optionally, the S1 further includes:

[0016] S11. Automatically obtain students' learning behavior data, performance data, answering process data, and learning resource usage data on various platforms by connecting to the data interfaces of various educational platforms used by students;

[0017] S12. Embed a data collection script in the student's learning terminal using embedded tracking technology to record the student's login time, learning time, operation sequence, answer sequence, and interactive behavior data in real time;

[0018] S13. Aggregate the above data in a unified data format to form the original learning data set D raw , where D raw ={d1,d2,...,d n}, d i represents the i-th original learning data, and n is the total amount of data;

[0019] S14, collect the original learning data set D raw Perform timestamp unification, student identity verification, and data deduplication to form a multi-source learning dataset D orig .

[0020] Optionally, the S2 further includes:

[0021] S21, multi-source learning dataset D orig Perform missing value detection and use mean filling, interpolation or similar sample filling methods to fill in the missing items to obtain the completed data set D fill ;

[0022] S22, D fill The data in the dataset are detected by using the 3σ principle or box plot method, and the abnormal data are eliminated or corrected to obtain the dataset D clean ;

[0023] S23, D clean The categorical data in the dataset are encoded using one-hot encoding (One-Hot Encoding), and the text variables are converted into numerical vectors to obtain the encoded dataset D enc ;

[0024] S24, to D enc The numerical features in are standardized using the Z-score standardization method to obtain the standardized learning dataset D std Specifically, the standardized value z is calculated for each feature x using the following formula:

[0025]

[0026] Wherein, x is the original feature value, μ is the mean value of the feature, σ is the standard deviation of the feature;

[0027] Optionally, the S3 further comprises:

[0028] S31, based on the high school knowledge system K and the student learning goal G, a student feature extraction deep learning model M is constructed and trained, the model M combines knowledge tracking and interest modeling mechanism, and can process learning behavior sequence and multi-source learning data;

[0029] S32, the standardized learning data set D std Is input into the model M, and the hidden state h t Of the fusion knowledge point mastery and behavior performance at each time step is obtained through multi-dimensional feature extraction and time sequence modeling, and the calculation formula is:

[0030] h t =f att (W x x t +W h h t-1 +b)+f kn (K t ),

[0031] Wherein, f att Is a nonlinear activation function focusing on student behavior patterns, f kn Is a knowledge point mastery extraction function based on the knowledge system, x t Is the input feature at the t time, h t-1 Is the hidden state of the previous time step, W x And W h Are input weight matrix and hidden state weight matrix respectively, b is the bias term, K t Is the knowledge point feature vector of the current time step;

[0032] S33, the hidden state of the last time step or the output layer transformation result thereof is taken as the student learning feature parameter vector F=[f1,f2,...,f m ], wherein f j Is the jth learning feature parameter, and m is the total number of parameters, the learning feature parameters include basic behavior features, knowledge point mastery, weak points and interest preferences;

[0033] Optionally, the S4 further comprises:

[0034] S41, the student learning feature parameter vector F=[f1,f2,...,f m ] and the high school knowledge system K are taken as inputs, and combined with the current learning goal G of the student, an individualized recommendation input set S={F,K,G} is constructed;

[0035] S42. Input the input set S to a personalized recommendation algorithm R. The personalized recommendation algorithm R may adopt a collaborative filtering algorithm, a content recommendation algorithm, a deep interest network (DIN), DeepFM, or a recommendation model based on knowledge tracking;

[0036] S43. Calculate each optional learning resource r through the recommendation algorithm R i The matching score s i , the formula is:

[0037] s i =R(F,K,G,r i ),

[0038] Among them, s i Represents learning resources i Match score with students' individual needs;

[0039] S44. According to the matching score s i Sort all candidate learning resources, select the top N with the highest scores, and generate personalized learning plans for students in, represents the jth recommended learning resource, and N is the number of resources in the solution;

[0040] S45. The student's personalized learning plan includes learning goals, learning paths, and learning resource recommendations.

[0041] Optionally, the S5 further includes:

[0042] S51. During the student's implementation of the personalized learning plan, real-time collection of student learning process data is performed through system logs, the online answering platform, and the learning terminal. The learning process data includes learning time, learning frequency, answering order, answering time, accuracy rate, error rate, and feedback information.

[0043] S52, organize the above learning process data into a learning process data sequence E={e1, e2, ..., e T}, where e t represents the learning process data collected at time t, where T is the total duration of the learning process or the number of acquisition time steps;

[0044] S53. Each learning process data is recorded in a structured manner according to the learning time, learning frequency, answering order, answering time, accuracy rate, error rate, and feedback information fields;

[0045] S54: Store the collected learning process data sequence E in a learning process database.

[0046] Optionally, the S6 further includes:

[0047] S61, the learning process data sequence E={e1,e2,...,e T} Input to the student feature extraction deep learning model M, and dynamically update the student learning feature parameter vector. The process is:

[0048] F′=M(D std ,E),

[0049] Among them, F′ is the updated student learning feature parameter vector, D std is the standardized learning data set, M is the model, and E is the latest learning process data sequence;

[0050] S62: Input the updated student learning feature parameter vector F′, the high school entrance examination knowledge system K, and the learning goal G into the personalized recommendation algorithm R again, and recalculate the candidate learning resources r i The matching score s′ i :

[0051] s′ i =R(F′,K,G,r i );

[0052] S63, according to the latest matching score s′ i Re-sort learning resources, select the top N with the highest scores, and dynamically adjust and optimize personalized learning plans in The jth learning resource is recommended;

[0053] S64. Push the optimized personalized learning plan P′ to students to achieve continuous adaptive optimization.

[0054] Optionally, the S7 further includes:

[0055] S71. After the student completes the personalized learning plan or the optimized learning plan, the student's learning result data R = {r1, r2, ..., r L}, where r l represents the result data of the lth learning task, where L is the total number of learning tasks;

[0056] S72, for each result data r l Extracting key evaluation indicators, including correct answer rate, improvement in knowledge point mastery, learning efficiency, and students' subjective feedback;

[0057] S73. Use the weighted comprehensive scoring method to calculate the implementation effect score of the student's personalized learning plan S eval , the calculation formula is:

[0058] S eval = alpha * ACC + beta * Delta K + gamma * EFF + delta * SAT,

[0059] Wherein, ACC is the average correct rate of answering, Delta K is the average amplitude of knowledge point mastery improvement, EFF is the learning efficiency, SAT is the satisfaction score, alpha, beta, gamma, delta are index weights, and alpha + beta + gamma + delta = 1 is satisfied;

[0060] S74, generate a learning effect report REP, including various evaluation indexes and comprehensive scores S eval And REP is pushed to the student and the related teacher or parent.

[0061] Optionally, the student feature extraction deep learning model can select a recurrent neural network, a long short-term memory network, a Transformer, a deep knowledge tracking network or other applicable structures according to data characteristics and multi-source fusion requirements, and has a continuous self-learning ability, and can use new data to perform incremental training or online updating on model parameters.

[0062] Optionally, the method further comprises: automatically generating a teaching analysis report for teachers based on the student learning feature parameter vector and the learning process data, and providing personalized teaching suggestions and intervention measures to the teachers in combination with artificial intelligence analysis results.

[0063] Beneficial effects

[0064] The application provides a personalized learning scheme generation and optimization method based on deep learning for middle and high school examinations, which fully integrates artificial intelligence and education data analysis technology, and can significantly improve the learning efficiency and examination preparation effect of middle and high school students.

[0065] Firstly, the application realizes comprehensive collection and integration of multi-source learning data of students through a multi-platform data interface and embedded tracking technology, and can restore the behavior characteristics and knowledge mastery of students in different learning scenarios in detail. Secondly, the feature extraction and modeling method based on deep learning effectively overcomes the limitation that traditional rules or simple algorithms cannot deeply mine multi-dimensional features of students, can dynamically capture the behavior patterns, knowledge mastery and interest changes in the learning process of students, and realize accurate identification and analysis of individual differences of students. In addition, the application combines personalized recommendation algorithm and dynamic feedback mechanism, can continuously optimize the learning scheme according to the real-time learning performance and feedback information of students, provide timely, accurate and high adaptation degree learning resources for students, improve the flexibility and self-adaptation ability of the learning path. At the same time, through multi-dimensional effect evaluation and intelligent report pushing, the application provides scientific decision support for teachers and parents, facilitates timely discovery of weak links and growth changes of students, and promotes personalized teaching and school-family cooperation in educating people.

[0066] In summary, the present application effectively makes up for the deficiencies of the prior art in feature modeling, dynamic optimization and effectiveness evaluation, significantly improves the scientificity, effectiveness and intelligent level of the personalized learning scheme for the middle and high school entrance examination, and has important significance for promoting students' efficient preparation for the examination and innovation of the education and teaching mode. BRIEF DESCRIPTION OF DRAWINGS

[0067] The accompanying drawings are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and are included to provide a further understanding of the application, and are made a part of the specification and are used to explain the application, but do not constitute a limitation on the application. In the drawings:

[0068] Fig. 1 is the overall flowchart of the personalized learning scheme generation and optimization method based on deep learning for the middle and high school entrance examination proposed by the present application;

[0069] Fig. 2 is the implementation flowchart of the personalized learning method based on deep learning for the middle and high school entrance examination in the present application;

[0070] Fig. 3 is the functional structure diagram of the personalized learning scheme generation and optimization for the middle and high school entrance examination in the present application. DETAILED DESCRIPTION

[0071] The present application will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, and only illustrate the basic structure of the present application in a schematic manner, and therefore only show the components related to the present application.

[0072] Reference Figs. 1-3 A personalized learning scheme generation and optimization method based on deep learning for the middle and high school entrance examination, comprising the following steps:

[0073] S1, collecting multi-source learning data of students through a multi-platform data interface and embedded tracking technology to form a multi-source learning data set;

[0074] S2, performing data cleaning, outlier detection and standardization processing on the multi-source learning data set to obtain a standardized learning data set;

[0075] S3, based on the middle and high school entrance examination knowledge system and the student learning goal, constructing and training a student feature extraction deep learning model, and inputting the standardized learning data set into the model to extract a student learning feature parameter vector;

[0076] S4, generating a student personalized learning scheme based on the student learning feature parameter vector and the middle and high school entrance examination knowledge system through a personalized recommendation algorithm;

[0077] S5, collecting student learning process data in real time during the execution of the personalized scheme by the student to form a learning process data sequence;

[0078] S6. Input the learning process data sequence into the student feature extraction deep learning model, and dynamically adjust and optimize the personalized learning plan in combination with the feedback optimization mechanism;

[0079] S7. Quantify and evaluate the effectiveness of students’ personalized learning plans and generate learning effectiveness reports.

[0080] In this embodiment, the S1 further includes:

[0081] S11. Automatically obtain students' learning behavior data, performance data, answering process data, and learning resource usage data on various platforms by connecting to the data interfaces of various educational platforms used by students;

[0082] S12. Embed a data collection script in the student's learning terminal using embedded tracking technology to record the student's login time, learning time, operation sequence, answer sequence, and interactive behavior data in real time;

[0083] S13. Aggregate the above data in a unified data format to form the original learning data set D raw , where D raw ={d1,d2,...,d n}, d i represents the i-th original learning data, and n is the total amount of data;

[0084] S14, collect the original learning data set D raw Perform timestamp unification, student identity verification, and data deduplication to form a multi-source learning dataset D orig .

[0085] In this embodiment, the S2 further includes:

[0086] S21, multi-source learning dataset D orig Perform missing value detection and use mean filling, interpolation or similar sample filling methods to fill in the missing items to obtain the completed data set D fill ;

[0087] S22, D fill The data in the dataset are detected by using the 3σ principle or box plot method, and the abnormal data are eliminated or corrected to obtain the dataset D clean ;

[0088] S23, D clean The categorical data in the dataset are encoded using one-hot encoding (One-Hot Encoding), and the text variables are converted into numerical vectors to obtain the encoded dataset D enc ;

[0089] S24, Denc The numerical features in the above table are standardized by using the Z-score standardization method to obtain a standardized learning data set D std , which is specifically calculated by the following formula:

[0090]

[0091] wherein x is the original feature value, μ is the mean of the feature, and σ is the standard deviation of the feature;

[0092] In this embodiment, the S3 further comprises:

[0093] S31, based on the high school knowledge system K and the student learning goal G, a student feature extraction deep learning model M is constructed and trained, the model M combines knowledge tracking and interest modeling mechanism, and can process learning behavior sequence and multi-source learning data;

[0094] S32, the standardized learning data set D std is input into the model M, and the hidden state h t of each time step is obtained by multi-dimensional feature extraction and time sequence modeling, and the calculation formula is:

[0095] h t =f att (W x x t +W h h t-1 +b)+f kn (K t ),

[0096] wherein f att is a nonlinear activation function focusing on student behavior patterns, f kn is a knowledge point mastery extraction function based on the knowledge system, x t is the input feature at the t time, h t-1 is the hidden state of the previous time step, W x and W h are input weight matrix and hidden state weight matrix respectively, b is the bias term, K t is the knowledge point feature vector of the current time step;

[0097] S33, the hidden state of the last time step or the output layer transformation result thereof is taken as a student learning feature parameter vector F = [f1, f2,..., f m ], wherein f j is the jth learning feature parameter, and m is the total number of parameters, and the learning feature parameters include basic behavior features, knowledge point mastery, weak points, and interest preferences;

[0098] In this embodiment, the S4 further comprises:

[0099] S41, the student learning feature parameter vector F=[f1, f2,..., f m ] and the high school knowledge system K as input, and combining the current learning goal G of the student, an individualized recommendation input set S={F, K, G} is constructed;

[0100] S42, input the input set S into the individualized recommendation algorithm R, which can adopt collaborative filtering algorithm, content recommendation algorithm, deep interest network (DIN), DeepFM or knowledge tracking-based recommendation model;

[0101] S43, the matching score s i of each optional learning resource r i is calculated by the recommendation algorithm R, and the formula is:

[0102] s i =R(F, K, G, r i ),

[0103] Wherein, s i indicates the matching score of the learning resource r i and the individualized needs of the student;

[0104] S44, all candidate learning resources are sorted according to the matching score s i , and the top N items with the highest score are selected to generate the student individualized learning scheme Wherein, Indicates the jth recommended learning resource, and N is the number of resources in the scheme;

[0105] S45, the student individualized learning scheme includes learning goal, learning path and learning resource recommendation.

[0106] In this embodiment, the S5 further comprises:

[0107] S51, in the process of the student executing the individualized learning scheme, the learning process data of the student is collected in real time through the system log, the online answering platform and the learning terminal, and the learning process data includes learning time, learning frequency, answering order, answering time, accuracy, error rate and feedback information;

[0108] S52, the above learning process data is organized into a learning process data sequence E={e1, e2,..., e T} in time sequence, wherein e t Indicates the learning process data collected at the tth moment, and T is the total time length or the collection time step number of the learning process;

[0109] S53. Each learning process data is recorded in a structured manner according to the learning time, learning frequency, answering order, answering time, accuracy rate, error rate, and feedback information fields;

[0110] S54: Store the collected learning process data sequence E in the learning process database.

[0111] In this embodiment, the S6 further includes:

[0112] S61, the learning process data sequence E={e1,e2,...,e T} Input to the student feature extraction deep learning model M, and dynamically update the student learning feature parameter vector. The process is:

[0113] F′=M(D std ,E),

[0114] Among them, F′ is the updated student learning feature parameter vector, D std is the standardized learning data set, M is the model, and E is the latest learning process data sequence;

[0115] S62: Input the updated student learning feature parameter vector F′, the high school entrance examination knowledge system K, and the learning goal G into the personalized recommendation algorithm R again, and recalculate the candidate learning resources r i The matching score s′ i :

[0116] s′ i =R(F′,K,G,r i );

[0117] S63, based on the latest matching score s′ i Re-sort learning resources, select the top N with the highest scores, and dynamically adjust and optimize personalized learning plans in The jth learning resource is recommended;

[0118] S64. Push the optimized personalized learning plan P′ to students to achieve continuous adaptive optimization.

[0119] In this embodiment, the S7 further includes:

[0120] S71. After the student completes the personalized learning plan or the optimized learning plan, the student's learning result data R = {r1, r2, ..., r L}, where r l represents the result data of the lth learning task, where L is the total number of learning tasks;

[0121] S72, for each result data rl extracting key evaluation indexes, including correct answer rate, knowledge point mastery improvement range, learning efficiency and student subjective feedback;

[0122] S73, using a weighted comprehensive scoring method to calculate the execution effect score S of the student individualized learning scheme eval , and the calculation formula is:

[0123] S eval = a ACC + b AK + g EFF + d SAT,

[0124] Wherein, ACC is the average correct answer rate, AK is the average improvement range of knowledge point mastery, EFF is the learning efficiency, SAT is the satisfaction score, a, b, g, d are index weights, and a+b+g+d=1;

[0125] S74, generating a learning effect report REP, including various evaluation indexes and comprehensive score S eval , and pushing REP to students and related teachers or parents.

[0126] In the embodiment, the student feature extraction deep learning model can select a recurrent neural network, a long short-term memory network, a Transformer, a deep knowledge tracking network or other applicable structures according to data characteristics and multi-source fusion requirements, and has a continuous self-learning ability, which can use new data to incrementally train or online update model parameters.

[0127] In the embodiment, the method further comprises: automatically generating a teacher-oriented teaching analysis report based on the student learning feature parameter vector and the learning process data, and providing personalized teaching suggestions and intervention measures to the teacher in combination with artificial intelligence analysis results.

[0128] Example 1

[0129] To verify the application effect of the individualized learning scheme generation and optimization method based on deep learning in actual teaching scenarios, the application is applied to the mathematics review process of the third grade of high school in a certain city in Jiangsu Province. In the high school entrance examination preparation, the school generally faces problems such as uneven student foundation level, difficulty in accurately positioning weak knowledge points, low utilization rate of learning resources, and difficulty in effectively carrying out individualized tutoring. The traditional review mode mainly adopts large-class teaching and unified assignment of exercises, which is difficult to carry out targeted teaching according to individual differences of students, resulting in low efficiency of some students in repeated review, repeated appearance of weak knowledge points, and slow improvement of scores.

[0130] To ensure the scientificity of the experimental design and the effectiveness of the comparative data, the school's academic affairs department selected 8 classes of grade 3 to participate in the pilot, a total of 384 people, fully considering the students' basic academic level, gender ratio, class size, teacher allocation and other key factors when classifying, to ensure the balance of the grouping. Before the start of the pilot, all students participating in the experiment took a unified high school entrance examination mathematics level test, the results showed that the average score of the experimental group and the control group was 85.6 and 85.2 respectively, the difference was very small, effectively ensuring the comparability and scientificity of the experimental data. Subsequently, the 8 classes were randomly divided into 4 classes of the experimental group, a total of 196 people, and 4 classes of the control group, a total of 188 people, the pilot period was from February to May 2025.

[0131] During the pilot, the school connected with the mainstream online homework platform and classroom interaction system to automatically collect students' multi-source learning behavior data, including online homework completion, daily test scores, knowledge point mistake records, daily learning time, answer order and time, learning resource usage frequency, and classroom interaction. At the same time, learning behavior tracking scripts were embedded in the learning tools commonly used by students on tablets and mobile phones to record students' daily learning period, knowledge point learning path, mistake repeated practice times, and self-learning feedback in real time. After all the collected data were cleaned, de-duplicated, missing value and outlier value processed and standardized, they were input into the student feature extraction deep learning model proposed in the invention.

[0132] The method of the invention is based on the high school entrance examination mathematics knowledge system, combined with students' high school entrance examination goals and learning characteristic parameters, including knowledge point mastery, weak points, interest preferences, and answering habits, etc. Through personalized recommendation algorithms, the most suitable learning resources such as detailed explanation videos, targeted training questions, mistake micro-lessons, and hierarchical reinforcement questions are dynamically pushed. For example, for students with high mistake rates on function questions, the system will preferentially recommend explanation videos and special training for related knowledge points; for students with low learning enthusiasm, interesting and moderately difficult questions are recommended to improve learning motivation. At the same time, the software platform developed by applying the method of the invention also analyzes students' answering speed, learning time changes, mistake repetition, and other behavior data to automatically identify learning bottlenecks and optimize learning content and review pace in a timely manner. The platform dynamically optimizes the personalized learning plan every week based on the latest learning behavior data and stage test results to ensure continuous and efficient learning for students. In addition, the platform also automatically generates a stage analysis report to push the class's overall knowledge point mastery, common weak points, individual progress, and learning warnings to teachers to assist teachers in carrying out targeted hierarchical teaching and individual tutoring, and improve overall teaching efficiency.

[0133] In the pilot period, the data of the experimental group and the control group were statistically analyzed and compared to comprehensively evaluate the effect of the application. The key indicators such as the scores of the periodic mock exam, the improvement of knowledge points, the change of wrong answer rate, the utilization rate of learning resources, the subjective satisfaction of students and teachers, etc. were investigated. The specific results are shown in Table 1.

[0134] Table 1 Comparison table of the effect of mathematics review for high school graduates

[0135]

[0136] As shown in the above table, the pre-test scores of the experimental group and the control group were basically the same before the pilot. After four months of application, the daily average learning time of the students in the experimental group was significantly improved, reaching 172.4 minutes, which was 23.1 minutes higher than that of the control group. The average score of the mock exam was improved by 7.9%, the excellent rate was improved by 48.2%, and the passing rate was improved by 7.4%. The improvement of knowledge point mastery in the experimental group reached 20.2%, which was significantly higher than that of the control group of 12.1%. The average decline of the wrong answer rate reached 14.2%, while that of the control group was only 8.3%. In terms of learning resource utilization rate, the experimental group reached 78.5%, which was much higher than that of the control group of 59.2%. In terms of subjective satisfaction, the scores of students and teachers were improved by 14.7% and 14.2%, respectively. The above data fully show that under the premise of balanced basic conditions and scientific grouping, the application can significantly improve the overall learning quality, knowledge point mastery and self-learning enthusiasm of high school graduates, and has wide practical application and promotion value.

[0137] The above is only the preferred specific embodiment of the application, but the protection scope of the application is not limited thereto. Any person skilled in the art can make equivalent replacements or changes to the technical range disclosed in the application according to the technical solution and inventive concept of the application, which should be covered within the protection scope of the application.

Claims

1. A method for generating and optimizing personalized learning plans for the college entrance examination and high school entrance examination based on deep learning, characterized in that: The steps include: S1. Collect students’ multi-source learning data through multi-platform data interfaces and embedded tracking technology to form a multi-source learning dataset; S2. Perform data cleaning, outlier detection, and standardization on the multi-source learning dataset to obtain a standardized learning dataset; S3. Based on the knowledge system of the National College Entrance Examination and the students' learning goals, a deep learning model for student feature extraction is constructed and trained, and the standardized learning dataset is input into the model to extract the student learning feature parameter vector; S4. Based on the student's learning characteristic parameter vector and the high school entrance examination and college entrance examination knowledge system, generate a personalized learning plan for the student through a personalized recommendation algorithm; S5. Collect students’ learning process data in real time while they are executing their personalized plans, forming a learning process data sequence; S6. Input the learning process data sequence into the student feature extraction deep learning model, and dynamically adjust and optimize the personalized learning plan in combination with the feedback optimization mechanism; S7. Quantify and evaluate the effectiveness of students’ personalized learning plans and generate learning effectiveness reports.

2. The method for generating and optimizing personalized learning plans for the high school entrance exam and college entrance exam based on deep learning according to claim 1 is characterized in that: Said S1 further comprises: S11. Automatically obtain students' learning behavior data, performance data, answering process data, and learning resource usage data on various platforms by connecting to the data interfaces of various educational platforms used by students; S12. Embed a data collection script in the student's learning terminal using embedded tracking technology to record the student's login time, learning time, operation sequence, answer sequence, and interactive behavior data in real time; S13. Aggregate the above data in a unified data format to form the original learning data set D raw , where D raw ={d1,d2,...,d n }, d i represents the i-th original learning data, and n is the total amount of data; S14, collect the original learning data set D raw Perform timestamp unification, student identity verification, and data deduplication to form a multi-source learning dataset D orig .

3. The method for generating and optimizing personalized learning plans for the high school entrance exam and college entrance exam based on deep learning according to claim 1 is characterized in that: Said S2 further comprises: S21, multi-source learning dataset D orig Perform missing value detection and use mean filling, interpolation or similar sample filling methods to fill in the missing items to obtain the completed data set D fill ; S22, D fill The data in the dataset are detected by using the 3σ principle or box plot method, and the abnormal data are eliminated or corrected to obtain the dataset D clean ; S23, D clean The categorical data in the dataset are encoded using one-hot encoding (One-Hot Encoding), and the text variables are converted into numerical vectors to obtain the encoded dataset D enc ; S24, D enc The numerical features in are standardized using the Z-score standardization method to obtain the standardized learning dataset D std .

4. The method for generating and optimizing personalized learning plans for the high school entrance exam and college entrance exam based on deep learning according to claim 1 is characterized in that: Said S3 further comprises: S31. Based on the high school entrance exam knowledge system K and the student learning goals G, a student feature extraction deep learning model M is constructed and trained. The model M combines knowledge tracking and interest modeling mechanisms and can process learning behavior sequences and multi-source learning data. S32, standardize the learning data set D std Input into the model M, and obtain the hidden state h of the fusion knowledge point mastery and behavior performance at each time step through multi-dimensional feature extraction and time series modeling t , the calculation formula is: h t =f att (W x x t +W h h t-1 +b)+f kn (K t ), Among them, f att is a nonlinear activation function that focuses on students’ behavior patterns, f kn is the knowledge point mastery extraction function based on the knowledge system, x t is the input feature at time t, h t-1 is the hidden state of the previous time step, W x and W h are the input weight matrix and the hidden state weight matrix respectively, b is the bias term, K t is the feature vector of the knowledge point at the current time step; S33, the hidden state of the last time step or its output layer transformation result is used as the student learning feature parameter vector F = [f1, f2, ..., f m ], where f j is the jth learning feature parameter, m is the total number of parameters, and the learning feature parameters include basic behavior characteristics, knowledge point mastery, weaknesses, and interest preferences.

5. The method for generating and optimizing personalized learning plans for the high school entrance exam and college entrance exam based on deep learning according to claim 1 is characterized in that: Said S4 further comprises: S41, the student learning characteristic parameter vector F=[f1,f2,...,f m ] and the high school entrance exam knowledge system K as input, and combined with the student’s current learning goal G, a personalized recommendation input set S = {F, K, G} is constructed; S42. Input the input set S to a personalized recommendation algorithm R. The personalized recommendation algorithm R may adopt a collaborative filtering algorithm, a content recommendation algorithm, a deep interest network (DIN), DeepFM, or a recommendation model based on knowledge tracking; S43. Calculate each optional learning resource r through the recommendation algorithm R i The matching score s i , the formula is: s i =R(F,K,G,r i ), Among them, s i Represents learning resources i Match score with students' individual needs; S44. According to the matching score s i Sort all candidate learning resources, select the top N with the highest scores, and generate personalized learning plans for students in, represents the jth recommended learning resource, and N is the number of resources in the solution; S45. The student's personalized learning plan includes learning goals, learning paths, and learning resource recommendations.

6. The method for generating and optimizing personalized learning plans for the high school entrance exam and college entrance exam based on deep learning according to claim 1 is characterized in that: Said S5 further comprises: S51. During the student's implementation of the personalized learning plan, real-time collection of student learning process data is performed through system logs, the online answering platform, and the learning terminal. The learning process data includes learning time, learning frequency, answering order, answering time, accuracy rate, error rate, and feedback information. S52, organize the above learning process data into a learning process data sequence E={e1, e2, ..., e T }, where e t represents the learning process data collected at time t, where T is the total duration of the learning process or the number of acquisition time steps; S53. Each learning process data is recorded in a structured manner according to the learning time, learning frequency, answering order, answering time, accuracy rate, error rate, and feedback information fields; S54: Store the collected learning process data sequence E in the learning process database.

7. The method for generating and optimizing personalized learning plans for the high school entrance exam and college entrance exam based on deep learning according to claim 1 is characterized in that: Said S6 further comprises: S61, the learning process data sequence E={e1,e2,...,e T } Input to the student feature extraction deep learning model M, and dynamically update the student learning feature parameter vector. The process is: F′=M(D std ,E), Among them, F′ is the updated student learning feature parameter vector, D std is the standardized learning data set, M is the model, and E is the latest learning process data sequence; S62: Input the updated student learning characteristic parameter vector F′, the high school entrance examination knowledge system K, and the learning goal G into the personalized recommendation algorithm R again, and recalculate the candidate learning resources r i The matching score s i ′: s′ i =R(F′,K,G,r i ); S63. According to the latest matching score s i Re-sort learning resources, select the top N with the highest scores, and dynamically adjust and optimize personalized learning plans in The jth learning resource is recommended; S64. Push the optimized personalized learning plan P′ to students to achieve continuous adaptive optimization.

8. The method for generating and optimizing personalized learning plans for the high school entrance exam and college entrance exam based on deep learning according to claim 1 is characterized in that: The S7 further includes: S71. After students complete their personalized learning plans or optimized learning plans, collect their learning result data R = {r1, r2, ..., r L }, where r l represents the result data of the lth learning task, where L is the total number of learning tasks; S72, for each result data r l Extracting key evaluation indicators, including correct answer rate, improvement in knowledge point mastery, learning efficiency, and students' subjective feedback; S73. Use the weighted comprehensive scoring method to calculate the implementation effect score of the student's personalized learning plan S eval , the calculation formula is: S eval =α·ACC+β·ΔK+γ·EFF+δ·SAT, Where ACC is the average correct rate of answering questions, ΔK is the average improvement in knowledge point mastery, EFF is the learning efficiency, SAT is the satisfaction score, α, β, γ, δ are indicator weights, and α+β+γ+δ=1; S74. Generate a learning effect report REP, including various evaluation indicators and comprehensive scores S eval and push REP to students and relevant teachers or parents.

9. The method for generating and optimizing personalized learning plans for the high school entrance exam and college entrance exam based on deep learning according to claim 1 is characterized in that: The student feature extraction deep learning model can select recurrent neural networks, long short-term memory networks, transformers, deep knowledge tracking networks or other applicable structures according to data characteristics and multi-source fusion requirements, and has continuous self-learning capabilities, and can use new data to perform incremental training or online updates on model parameters.

10. The method for generating and optimizing personalized learning plans for the high school entrance exam and college entrance exam based on deep learning according to claim 1, characterized in that: The method also includes: automatically generating a teaching analysis report for teachers based on student learning feature parameter vectors and learning process data, and providing personalized teaching suggestions and intervention measures to teachers in combination with artificial intelligence analysis results.

Images