Learning recommendation method and system based on cluster model, terminal and storage medium

Abstract

The invention provides a learning recommendation method and system based on a cluster model, a terminal and a storage medium, and the method comprises the steps: obtaining an evaluation data set of a target student, the evaluation data set comprising learning behavior data, score data and character data; according to data types existing in the evaluation data set, feature extraction is conducted on corresponding data in the evaluation data set through corresponding feature extraction models, corresponding feature vectors are obtained, and the data types comprise a video data type, a text data type and a time series data type; performing feature fusion on each feature vector to obtain a fused feature vector; and inputting the fusion feature vector into a preset hybrid model, so that the hybrid model evaluates the learning ability of the target student according to the fusion feature vector to obtain an evaluation result, and generating a learning path corresponding to the target student according to the evaluation result, thereby improving the accuracy and flexibility of personalized learning recommendation.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a learning recommendation method, system, terminal and storage medium based on a cluster model. Background Art

[0002] In terms of modern educational needs and personalized learning, with the rapid development of information technology and artificial intelligence, the modern education system is increasingly developing towards digitalization and personalization. The advancement of educational technology makes personalized learning possible, that is, adjusting the teaching content and teaching speed according to each student's learning ability, knowledge base, interest preferences and learning progress. This teaching method can greatly improve learning efficiency, increase learning motivation, and help students achieve better learning results. Among them, the personalized learning recommendation system is one of the key technologies to achieve personalized education. It collects, analyzes and predicts learners' learning behaviors and preferences to provide them with accurate learning resource recommendations. According to students' learning characteristics, the most suitable resources for students are selected from a large number of learning resources for recommendation. This technology can make more effective use of learning resources, thereby greatly improving learners' learning effects and efficiency.

[0003] However, existing recommendation systems are difficult to fully meet the personalized needs of learners because they rely on a single model or algorithm. Existing recommendation systems mainly include rule-based recommendation systems, model-based recommendation systems, and hybrid recommendation systems. Early personalized learning recommendation systems were mostly based on fixed rules and recommended learning content through preset logic. For example, if a student scores below a certain standard in a test, the system will recommend basic learning materials. The main disadvantage of this system is that it has poor flexibility, cannot respond to students' specific performance and changes in real time, and has difficulty handling complex learning situations. With the application of machine learning technology, model-based recommendation systems have begun to be developed and used. Such systems analyze students' behavioral data through data mining and machine learning techniques to predict their learning preferences and needs. However, most systems rely on a single model, such as decision trees, support vector machines, or neural networks. These models may perform well in specific scenarios, but their applicability and accuracy are still limited when dealing with a wider range of learning data. In order to overcome the limitations of a single model, a hybrid recommendation system has been proposed. This system combines multiple recommendation technologies, such as collaborative filtering and content recommendation. Although hybrid systems perform better in some aspects, they still face the following problems in practical applications: First, data integration is difficult: data from different sources have different formats and standards, making data integration a major challenge. Second, high processing complexity: when the system needs to process large-scale and diverse data, its performance is often limited by computing resources and algorithm efficiency. Third, lack of adaptability: although hybrid systems are improved over single model systems, they are still insufficient in dynamically adapting to learners' real-time changing needs. Summary of the invention

[0004] In response to the above problems, the present application provides a learning recommendation method, system, terminal and storage medium based on a cluster model to improve the accuracy and flexibility of personalized learning recommendations.

[0005] In a first aspect, the present application provides a learning recommendation method based on a cluster model, comprising:

[0006] Acquire an evaluation data set of a target student, wherein the evaluation data set includes learning behavior data, performance data, and personality data;

[0007] According to the data types existing in the evaluation data set, using respective corresponding feature extraction models to respectively extract features of the corresponding data in the evaluation data set to obtain respective corresponding feature vectors, the data types including video data class, text data class and time series data class;

[0008] Performing feature fusion on each of the feature vectors to obtain a fused feature vector;

[0009] The fused feature vector is input into a preset hybrid model so that the hybrid model evaluates the target student's learning ability according to the fused feature vector to obtain an evaluation result, and then generates a learning path corresponding to the target student according to the evaluation result, wherein the hybrid model is constructed by several machine learning models and trained based on evaluation data sets of several students, and the learning path includes several learning resources and learning sequence.

[0010] The embodiment of the present application provides a learning recommendation method based on a cluster model. First, the evaluation data set of the target student is obtained, and different feature extraction models are used to extract features according to the data type, and then the features of each feature vector are fused to achieve integrated processing of multiple types of data, so that the hybrid model in the embodiment of the present application can efficiently process diverse data, thereby improving the adaptability of learning recommendations. At the same time, the embodiment of the present application utilizes the collaborative drive of the cluster model, integrates the model advantages of multiple fields, and evaluates the learning ability of students based on the fused fusion feature vector through collective wisdom and collaborative learning methods, and further generates corresponding learning paths based on the evaluation results, thereby achieving personalized learning recommendations. Through multi-model clustering, the accuracy of learning recommendations is significantly improved.

[0011] Furthermore, obtaining the evaluation data set of the target student includes:

[0012] Obtain the target student's learning behavior data through the learning records of the online learning platform and the video records of the offline classroom, wherein the learning behavior data includes classroom interaction data, homework and task completion status, and learning habits and resource usage data;

[0013] Obtaining the target student's performance data through test and quiz performance records, the performance data including course performance, assessment and quiz performance, and comprehensive evaluation data;

[0014] The personality data of the target student is obtained through the questionnaire survey results, and the personality data includes personality assessment data, learning style assessment data, and mental health status data.

[0015] In an embodiment of the present application, the learning behavior data, performance data, and personality data of the target students are obtained respectively through different data sources, thereby improving the diversity and richness of the evaluation data set, so that the hybrid model can comprehensively evaluate the learning ability of the target students from multiple angles, thereby improving the accuracy and rationality of the model evaluation, making the subsequent generation of learning paths more accurate, and improving the accuracy of personalized learning recommendations.

[0016] In a possible implementation, the method further includes, after obtaining the evaluation data set of the target student, performing data preprocessing on the evaluation data set, including data cleaning, standardization, data synchronization, data integration, and data verification;

[0017] Wherein, the data cleaning includes removing missing values, removing duplicate values ​​and correcting erroneous values;

[0018] The standardization process includes unifying data format, standardizing data range, classifying coding and text data preprocessing;

[0019] The data synchronization includes aligning timestamps, dividing time windows, and aggregating data within the time windows;

[0020] The data integration includes building a database and fusing data;

[0021] The data verification includes checking data consistency and assessing data quality.

[0022] In an embodiment of the present application, the obtained evaluation data set is further subjected to data preprocessing. Since the evaluation data set is collected from different data sources, and the data types in the data set also present diversified forms, different data sources may adopt different data formats and standards, making data integration work complicated and error-prone. In addition, traditional systems are often limited by computing resources and algorithm efficiency when processing large-scale data from different sources. Therefore, in order to ensure the consistency and comparability of each data in the data set and improve the processing efficiency of the model to data, the embodiment of the present application needs to perform data preprocessing on the evaluation data set. By combining the means of data cleaning, standardization, data synchronization, data integration and data verification, the embodiment of the present application can automatically identify and convert the data formats of different data sources, ensure the consistency and comparability of the data, and then enable the subsequent hybrid model to quickly process large-scale multi-source data, improve the operating efficiency and response speed of the model.

[0023] In a possible implementation, according to the data types in the evaluation data set, using respective corresponding feature extraction models to extract features from the corresponding data in the evaluation data set to obtain respective corresponding feature vectors includes:

[0024] Using a preset computer vision model to extract features from the video data class data in the evaluation data set to obtain a corresponding video feature vector;

[0025] Using a preset natural language processing model to perform feature extraction on the text data class data in the evaluation data set to obtain a corresponding text feature vector;

[0026] A preset time series data model is used to perform feature extraction on the time series data class data in the evaluation data set to obtain a corresponding time series feature vector.

[0027] An embodiment of the present application provides a feature extraction method. Since there are multiple data types in the evaluation data set, including video data, text data, and time series data, the present application uses different feature extraction models to extract feature vectors of different data types respectively, thereby improving the efficiency and specificity of data processing and providing data support for subsequent model input.

[0028] Furthermore, the step of fusing the feature vectors to obtain a fused feature vector includes:

[0029] Calculate the mean and standard deviation of each of the eigenvectors;

[0030] According to the mean and standard deviation of each of the eigenvectors, each of the eigenvectors is subjected to z-score standardization processing to obtain each corresponding standardized eigenvector;

[0031] A Transformer encoder based on a self-attention mechanism is used to perform feature fusion on each of the standardized feature vectors to obtain the fused feature vector.

[0032] The embodiment of the present application provides a feature vector fusion method. Since each feature vector is extracted by different feature extraction models, each feature vector needs to be standardized before feature fusion, and then the Transformer encoder based on the self-attention mechanism is used to perform feature fusion on each of the standardized feature vectors. The self-attention mechanism can capture the relationship between each part of the input sequence and combine different features by weighted averaging. This method can not only deal with the problem of inconsistent feature dimensions, but also establish more complex dependencies between features. Improve the accuracy of subsequent personalized learning recommendations.

[0033] In a possible implementation, the hybrid model evaluates the learning ability of the target student according to the fused feature vector to obtain an evaluation result, and then generates a learning path corresponding to the target student according to the evaluation result, including:

[0034] Each machine learning model in the hybrid model evaluates the learning ability of the target student according to the fused feature vector, and generates each corresponding initial evaluation result;

[0035] The hybrid model integrates the initial evaluation results according to the preset weights of the machine learning models to obtain the evaluation result;

[0036] Each machine learning model recommends a number of learning resources based on the evaluation results and the preset knowledge graph;

[0037] The hybrid model integrates and sorts a number of learning resources recommended by each machine learning model according to the preset weights of each machine learning model, and generates a learning path corresponding to the target student.

[0038] In an embodiment of the present application, each machine learning model is clustered into a hybrid model by means of preset weights, thereby realizing the collaborative work of the clustered large model. Through the collaboration of the clustered large models and the advantages of each model in different fields, a more comprehensive learning data analysis is achieved. This method not only improves the generalization ability of the model, but also enhances the accuracy and personalization level of the recommendation. Before performing the learning ability assessment and learning path generation process, different weights can be set for each machine learning model according to the learning field and model performance, effectively integrating the wisdom of large models from different fields and functions to form a comprehensive and powerful hybrid model. This model collaboration method not only improves the accuracy and adaptability of the system, but also increases the system's ability to handle complex learning scenarios.

[0039] In a possible implementation, the method further includes, after generating the learning path, dynamically adjusting the learning path according to the learning resource preference and real-time learning data of the target student, specifically:

[0040] Acquire the learning resource preference and real-time learning data of the target student, wherein the real-time learning data includes learning progress, real-time learning behavior data and real-time achievement data;

[0041] Inputting the real-time learning data into the hybrid model so that the hybrid model predicts the future learning performance of the target student;

[0042] Scoring each learning resource in the learning resource library and the learning path according to the learning resource preference and future learning performance;

[0043] The learning path is dynamically adjusted according to the scores of the various learning resources and the future learning performance, and the dynamic adjustment includes replacing or adding or deleting the learning resources in the learning path and adjusting the learning sequence in the learning path.

[0044] The embodiment of the present application provides a method for dynamically adjusting a learning path. In the prior art, although some hybrid recommendation systems attempt to improve the flexibility and adaptability of the system by combining multiple recommendation technologies, in actual operation, these systems are often difficult to respond to changes in learner behavior in real time and dynamically adjust the learning path and recommended content. In the real learning process, affected by various subjective and objective reasons, learners often find it difficult to strictly implement the preset learning path or learning plan, resulting in the actual learning progress deviating from the preset learning path, and thus unable to achieve the learning effect in theory. Therefore, it is necessary to flexibly adjust the learning path according to the actual learning effect during the learning process. The embodiment of the present application introduces a dynamic learning path optimization mechanism, which can adjust the learning path in real time according to the learner's immediate learning performance, feedback and progress. By predicting students' future academic performance through real-time learning data and dynamically adjusting the learning path according to the predicted results, the flexibility of personalized learning recommendations is greatly improved, thereby achieving a truly personalized learning experience.

[0045] In a second aspect, accordingly, the present application provides a learning recommendation system based on a cluster model, including an acquisition module, a feature extraction module, a feature fusion module and a path generation module;

[0046] Wherein, the acquisition module is used to acquire an evaluation data set of a target student, wherein the evaluation data set includes learning behavior data, performance data and personality data;

[0047] The feature extraction module is used to extract features from the corresponding data in the evaluation data set using corresponding feature extraction models according to the data types in the evaluation data set, so as to obtain corresponding feature vectors, wherein the data types include video data, text data and time series data;

[0048] The feature fusion module is used to fuse the features of each feature vector to obtain a fused feature vector;

[0049] The path generation module is used to input the fused feature vector into a preset hybrid model so that the hybrid model evaluates the target student's learning ability based on the fused feature vector to obtain an evaluation result, and then generates a learning path corresponding to the target student based on the evaluation result, wherein the hybrid model is constructed by several machine learning models and trained based on evaluation data sets of several students, and the learning path includes several learning resources and a learning sequence.

[0050] In a third aspect, the present application provides a terminal, comprising a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, the cluster model-based learning recommendation method described in the present application is implemented.

[0051] In a fourth aspect, the present application provides a computer-readable storage medium, which includes a stored computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute the cluster model-based learning recommendation method described in the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 : A flow chart of a learning recommendation method based on a cluster model provided in an embodiment of the present application.

[0053] Figure 2 : A schematic diagram of a process for dynamically adjusting a learning path in a learning recommendation method based on a cluster model provided in an embodiment of the present application.

[0054] Figure 3 : A flow chart of another learning recommendation method based on a cluster model provided in an embodiment of the present application.

[0055] Figure 4 : A structural diagram of a learning recommendation system based on a cluster model provided in an embodiment of the present application. DETAILED DESCRIPTION

[0056] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0057] It should be noted that the step numbers in the text are only for the convenience of explanation of the specific embodiments and do not serve to limit the order in which the steps are executed. In the description of this application, the terms "first" and "second" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features.

[0058] Embodiment 1:

[0059] like Figure 1 As shown, the first embodiment provides a learning recommendation method based on a cluster model, including steps S1-S4:

[0060] Step S1, obtaining an evaluation data set of a target student, wherein the evaluation data set includes learning behavior data, performance data, and personality data;

[0061] Step S2, according to the data types existing in the evaluation data set, using each corresponding feature extraction model to respectively extract features of the corresponding data in the evaluation data set to obtain each corresponding feature vector, the data types including video data class, text data class and time series data class;

[0062] Step S3, performing feature fusion on each of the feature vectors to obtain a fused feature vector;

[0063] Step S4: input the fused feature vector into a preset hybrid model, so that the hybrid model evaluates the target student's learning ability according to the fused feature vector to obtain an evaluation result, and then generates a learning path corresponding to the target student according to the evaluation result, wherein the hybrid model is constructed by several machine learning models and trained based on evaluation data sets of several students, and the learning path includes several learning resources and a learning sequence.

[0064] The embodiment of the present application provides a learning recommendation method based on a cluster model. First, the evaluation data set of the target student is obtained, and different feature extraction models are used to extract features according to the data type, and then the features of each feature vector are fused to achieve integrated processing of multiple types of data, so that the hybrid model in the embodiment of the present application can efficiently process diverse data, thereby improving the adaptability of learning recommendations. At the same time, the embodiment of the present application utilizes the collaborative drive of the cluster model, integrates the model advantages of multiple fields, and evaluates the learning ability of students based on the fused fusion feature vector through collective wisdom and collaborative learning methods, and further generates corresponding learning paths based on the evaluation results, thereby achieving personalized learning recommendations. Through multi-model clustering, the accuracy of learning recommendations is significantly improved.

[0065] Furthermore, in step S1, obtaining the evaluation data set of the target student includes:

[0066] Obtain the target student's learning behavior data through the learning records of the online learning platform and the video records of the offline classroom, wherein the learning behavior data includes classroom interaction data, homework and task completion status, and learning habits and resource usage data;

[0067] Obtaining the target student's performance data through test and quiz performance records, the performance data including course performance, assessment and quiz performance, and comprehensive evaluation data;

[0068] The personality data of the target student is obtained through the questionnaire survey results, and the personality data includes personality assessment data, learning style assessment data, and mental health status data.

[0069] In a preferred embodiment, the classroom interaction data includes: Attendance record: student attendance at each class. Participation: the frequency and quality of students asking questions, answering questions and participating in discussions in class. Classroom concentration: student behaviors related to learning concentration in class, including listening, standing to answer, discussing with classmates, drinking water, sleeping, playing with mobile phones, writing, recording with electronic devices, playing with mobile phones, using notebooks, etc. Use of interactive tools: such as click-through rate, browsing time, discussion area speeches, online test participation, etc.

[0070] The completion of the assignments and tasks includes: Assignment submission time: whether the assignment was submitted on time, and the specific time of advance or delay. Assignment completion quality: the score of the assignment, teacher feedback, and whether it needs to be redone. Task participation: participation and completion of tasks such as project work and lab reports.

[0071] The learning habits and resource usage data include: Online learning duration: the total time spent on the online learning platform every day and every week. Resource usage frequency: the frequency of use of learning resources such as e-books, video lectures, and simulation tests. Learning rhythm: the frequency and time interval of learning, whether there is a concentrated intensive learning or a continuous and stable learning habit.

[0072] The course grades include: mid-term and final exam scores: the scores of each major exam and their proportion to the total score. Regular grades: the overall score consisting of quizzes, class performance, homework scores, etc.

[0073] The assessment and test scores include: Periodic test scores: the scores of unit tests or periodic assessments conducted during the learning process. Skill assessments: Programming ability tests, experimental operation ability assessments and other skill test scores related to the course.

[0074] The comprehensive evaluation data includes: GPA (Grade Point Average): the average grade point average calculated by combining the grades of various courses. Academic year ranking: the student's ranking in the class or grade.

[0075] The personality assessment data includes linking the behavioral data of students completing homework with the Big Five Personality Test (Big Five Personality Test) for analysis, inferring certain personality traits through behavioral patterns, and helping us to more comprehensively understand students' personality traits and their impact on learning behavior. The detailed correlation analysis between behavioral data and Big Five personality is as follows:

[0076] Openness

[0077] Behavioral data: whether students like to explore new problem-solving methods and try different problem-solving strategies.

[0078] Correlation analysis: High openness: May use multiple methods to solve homework problems, showing creativity and flexibility. Low openness: Tend to use familiar methods and are reluctant to try new problem-solving strategies.

[0079] Conscientiousness

[0080] Behavioral data: whether the assignment was submitted on time, whether it was completed in multiple sessions, and the completeness and detail of the assignment.

[0081] Correlation analysis: High sense of responsibility: assignments are submitted on time, usually completed in one go, and the assignments are completed very carefully and in detail. Low sense of responsibility: often procrastinates, assignments are completed in multiple times, and omissions and errors may occur.

[0082] Extraversion

[0083] Behavioral data: participation in group work, frequency and quality of collaboration with classmates.

[0084] Correlation analysis: High extraversion: actively participates in group work, frequently interacts and cooperates with classmates. Low extraversion: prefers to complete homework independently and participates less in group activities.

[0085] Agreeableness

[0086] Behavioral data: cooperative attitude in group work, whether willing to help others.

[0087] Correlation analysis: High agreeableness: Willing to help classmates, positive attitude towards cooperation, easy to cooperate with others. Low agreeableness: More focused on their own tasks, relatively cold attitude towards cooperation, may be difficult to cooperate with others.

[0088] Emotional stability (Neuroticism)

[0089] Behavioral data: performance when faced with difficult questions and emotional fluctuations during the completion of the assignment.

[0090] Correlation analysis: High emotional stability: Stay calm when facing difficulties and manage time and stress effectively. Low emotional stability: Get anxious easily when encountering difficult questions and have large mood swings during the homework process.

[0091] The learning style assessment data includes: Learning style questionnaire: assessing students' learning preferences, such as visual, auditory, kinesthetic, etc. Motivation and attitude survey: understanding students' motivation, interest and attitude towards learning through questionnaires.

[0092] The mental health status data include: Stress and anxiety assessment: assessing students' stress level and anxiety through questionnaires or professional assessment tools. Social skills and emotional regulation: assessing students' interpersonal skills and emotional regulation abilities.

[0093] The above data can be collected through the following methods:

[0094] 1) Online platform data collection: Use the learning management system (LMS) to record students’ online learning behavior and homework completion behavior, such as online teaching platforms, YuKeTang, PTA (Program Online Assessment Platform) and other platforms.

[0095] 2) Questionnaire survey: Collect students’ personality assessment and learning style information through online questionnaires (such as WJX, Tencent Online Documents, WPS Online Documents, etc.) or paper questionnaires.

[0096] 3) Examinations and tests: Obtain students’ performance data through classroom tests, homework, regular examinations and unit tests.

[0097] 4) Recording and analysis data of smart classrooms: Teachers can use smart classrooms (intelligent analysis of recorded class videos) to record students’ classroom performance and participation in daily teaching, and provide feedback to students’ process learning files.

[0098] In an embodiment of the present application, the learning behavior data, performance data, and personality data of the target students are obtained respectively through different data sources, thereby improving the diversity and richness of the evaluation data set, so that the hybrid model can comprehensively evaluate the learning ability of the target students from multiple angles, thereby improving the accuracy and rationality of the model evaluation, making the subsequent generation of learning paths more accurate, and improving the accuracy of personalized learning recommendations.

[0099] In a possible implementation, the method further includes, after obtaining the evaluation data set of the target student, performing data preprocessing on the evaluation data set, including data cleaning, standardization, data synchronization, data integration, and data verification;

[0100] Wherein, the data cleaning includes removing missing values, removing duplicate values ​​and correcting erroneous values;

[0101] The standardization process includes unifying data format, standardizing data range, classifying coding and text data preprocessing;

[0102] The data synchronization includes aligning timestamps, dividing time windows, and aggregating data within the time windows;

[0103] The data integration includes building a database and fusing data;

[0104] The data verification includes checking data consistency and assessing data quality.

[0105] In a preferred embodiment, the removing of missing values, removing of duplicate values ​​and correction of erroneous values ​​are specifically as follows:

[0106] Remove missing values: Use deletion, filling, or interpolation methods to handle missing data.

[0107] De-duplicate: Delete redundant duplicate data records.

[0108] Correcting Bad Values: Checking and correcting obviously erroneous data (e.g., negative scores, outliers).

[0109] The unified data format, standardized data range, classification coding and text data preprocessing are specifically as follows:

[0110] Unify data formats: Unify date and time formats: Convert all timestamps to a unified format (such as ISO8601). Unify numerical units: For example, unify all scores to a percentage system and unify time units to hours or minutes.

[0111] Standardize the data range: Numerical standardization: convert data of different dimensions into dimensionless data so that they can be compared on the same scale. Common methods include: Min-Max Normalization: linearly convert data to the interval [0,1]; Percentile Normalization: convert data into percentiles. Unify data types: The data types of the same attributes of the same student collected by different data sources should be unified, such as student IDs unified into integer data, etc.

[0112] Categorical Encoding: One-Hot Encoding: Converts categorical data into binary vector form. Label Encoding: Converts categorical data into integer values.

[0113] Text data preprocessing: including word segmentation, stop word removal, stemming and encoding.

[0114] The aligning of timestamps, dividing time windows, and aggregating data within the time windows are specifically as follows:

[0115] Align timestamps: Unify time intervals: Sample all data to the same time interval (such as daily, weekly). Interpolation processing: For data with inconsistent time intervals, use interpolation methods to fill in missing time point data.

[0116] Divide the time window: Fixed window: Divide the data into time windows of fixed length, such as weekly or monthly. Sliding window: Use sliding window technology to slide on the time axis with a certain step size to generate overlapping data windows.

[0117] Aggregate the data in the time window: Aggregation function: Aggregate the data in the time window using aggregation functions such as mean, median, and maximum. Feature engineering: Extract features from the aggregated data based on demand, such as calculating the change trend and fluctuation range within the time period.

[0118] The said construction of database and fusion of data are specifically as follows:

[0119] Build a database: Relational database: Design a standardized relational database structure to ensure efficient data storage and query. Data warehouse: Build a data warehouse to centrally store and analyze historical data and real-time data.

[0120] Fusion of data: Key-value matching: Use unique identifiers (such as student ID) to match and merge data from different sources. Data linking: Use foreign key relationships to connect related tables to form a complete data view.

[0121] The checking of data consistency and evaluation of data quality are specifically as follows:

[0122] Check data consistency: Data range check: Verify that the values ​​are within the expected range. Data integrity check: Ensure that all required data fields have valid values.

[0123] Assess data quality: Statistical analysis: Use descriptive statistical analysis methods (such as mean, variance, frequency distribution) to assess data quality. Anomaly detection: Use anomaly detection algorithms to identify and handle potential data anomalies.

[0124] The implementation tools and technologies used in the above data processing process include but are not limited to: Programming languages: Python (pandas, numpy), R, etc. Database management systems: MySQL, PostgreSQL, etc. Data processing platforms: Apache Spark, Hadoop, etc. Visualization tools: Tableau, Power BI, Matplotlib, etc.

[0125] In an embodiment of the present application, the obtained evaluation data set is further subjected to data preprocessing. Since the evaluation data set is collected from different data sources, and the data types in the data set also present diversified forms, different data sources may adopt different data formats and standards, making data integration work complicated and error-prone. In addition, traditional systems are often limited by computing resources and algorithm efficiency when processing large-scale data from different sources. Therefore, in order to ensure the consistency and comparability of each data in the data set and improve the processing efficiency of the model to data, the embodiment of the present application needs to perform data preprocessing on the evaluation data set. By combining the means of data cleaning, standardization, data synchronization, data integration and data verification, the embodiment of the present application can automatically identify and convert the data formats of different data sources, ensure the consistency and comparability of the data, and then enable the subsequent hybrid model to quickly process large-scale multi-source data, improve the operating efficiency and response speed of the model.

[0126] In a possible implementation, in step S2, according to the data types existing in the evaluation data set, using respective corresponding feature extraction models to respectively extract features from the corresponding data in the evaluation data set to obtain respective corresponding feature vectors, including:

[0127] Using a preset computer vision model to extract features from the video data class data in the evaluation data set to obtain a corresponding video feature vector;

[0128] Using a preset natural language processing model to perform feature extraction on the text data class data in the evaluation data set to obtain a corresponding text feature vector;

[0129] A preset time series data model is used to perform feature extraction on the time series data class data in the evaluation data set to obtain a corresponding time series feature vector.

[0130] In a preferred embodiment, the preset computer vision model is used to extract features from the video data in the evaluation data set to obtain corresponding video feature vectors. Taking the viewing behavior data of students as an example, the specific extraction is as follows:

[0131] In order to extract features from students’ viewing behavior data (such as viewing time, pause frequency, repeated playback, etc.) through video models and evaluate students’ understanding and attention concentration on different video contents, we can use advanced video analysis models such as 3D convolutional neural networks (3D-CNN) or long short-term memory networks (LSTM) combined with visual feature extraction models (such as ResNet).

[0132] Specifically, we first use the online teaching platform provided by the school to obtain basic data such as students' viewing time, pause frequency, and repeated playback (the behavior records of logging into the platform to learn videos can be exported); synchronize the above data with the video content. We obtained the following data: Viewing time: the total viewing time of each video clip. Pause frequency: the number of pauses during video playback and the duration of each pause. Repeated playback: the number of repeated views of a video clip. Video content: video frame sequence and its corresponding timestamp.

[0133] Here we use three models: 3D-CNN combined with ResNet for video frame feature extraction, and LSTM model for temporal feature analysis. 3D-CNN is used to extract spatiotemporal features between video frames, ResNet is used to extract visual features of each frame, and the LSTM model is used to analyze temporal features and combine viewing behavior data to evaluate students' understanding of video content and their concentration. From the analysis results, on the one hand, the video clip comprehension score predicted by the LSTM model is used to evaluate students' understanding of different video contents; on the other hand, combined with viewing time, pause frequency, number of repeated playbacks and other data, the students' concentration is evaluated. The detailed mathematical description is as follows:

[0134] Defining and quantifying characteristics

[0135] 1) Watch Time

[0136] Suppose the total viewing time of student i on video j is T ij , in seconds.

[0137] 2) Pause Frequency

[0138] Let the number of pauses of student i on video j be P ij .

[0139] 3) Repeat Count

[0140] Let R be the number of times student i repeatedly plays video j. ij .

[0141] 4) Video Features

[0142] Let the feature vector of video j be V j , which consists of video frame features (such as features extracted by ResNet) and spatiotemporal features (such as features extracted by 3D-CNN).

[0143] 5) Understanding Level

[0144] Assume that student i’s understanding of video j is U ij , ranging from 0 to 1.

[0145] 6) Attention Level

[0146] Assume that the concentration level of student i when watching video j is A ij , ranging from 0 to 1.

[0147] Then the process of extracting features using 3D-CNN is as follows:

[0148] We first extract the spatiotemporal features of the video. Assume that video j consists of n frames, and the feature vector of each frame is extracted through ResNet as F jk (k represents the kth frame), the dimension of the feature vector of each frame is d.

[0149] The spatiotemporal features of the video are extracted through 3D-CNN to obtain the overall feature vector V of the video. j :

[0150] V j =3D-CNN(F j1 , F j2 , F j3 , …, F jn )

[0151] For each student's behavior data in video j, we define the feature vector B ij :

[0152] B ij =[T ij , P ij , R ij ]

[0153] Combined with video content features V j , we get the comprehensive feature vector C of student i on video j ij :

[0154] C ij =[B ij ,V j ]

[0155] Use the LSTM model to analyze the time series features and predict the student i’s understanding of video j U ij :

[0156] U ij =LSTM(C ij )

[0157] We quantify students' attention span through viewing behavior data ij:

[0158]

[0159] Among them, T max , P max and R max The maximum values ​​of viewing time, pause frequency and repeated playback times are set to ensure A ij The range is between 0 and 1. Finally, combined with U ij and A ij Evaluate students’ overall learning outcomes:

[0160] S ij =αU ij +(1-α)A ij

[0161] Among them, α is the weight coefficient, which reflects the relative importance of understanding and concentration in the comprehensive score.

[0162] Example Calculation

[0163] Assume that the student's video viewing behavior data is as follows: Viewing time T ij =300 seconds, pause frequency P ij =2, repeat playback times R ij =1, video content feature vector V j It is extracted by the 3D-CNN model. Then the feature vector B ij =[300,2,1], C ij =[300,2,1,V j ] Use LSTM model to predict the level of understanding U ij =LSTM(C ij ), assuming that the calculated result U ij =0.75.

[0164] Assume T max = 600 seconds, P max =5, R max =3, then:

[0165]

[0166] Assume α = 0.7, then:

[0167] S ij =0.7×0.75+0.3×0.2=0.525+0.06=0.585

[0168] The method of using a preset natural language processing model to extract features from the text data in the evaluation data set to obtain a corresponding text feature vector is as follows:

[0169] If you use a large model you built yourself, such as BERT, a bidirectional encoding model based on the Transformer architecture, you can extract context-related features from the input text. The process of feature extraction using BERT is divided into three steps: text preprocessing, that is, text segmentation, and converting the words into the input format of the BERT model (including Token IDs, Segment IDs, and Attention Masks). The second step is model input. Input the preprocessed text into the BERT model. The third step is feature extraction: Get the feature vector of the text from the last layer or the second to last layer of BERT. These vectors can represent the semantic information of the entire sentence, or the context-related features of each word.

[0170] If you use a public large model to extract text features, you can do so by calling the model API. For example, through the APIs of large models such as "GPT" and "Tongyi Qianwen", the text can be sent to the corresponding large models for processing.

[0171] Finally, the desired features are extracted from the responses of these large models, such as the generated embedding vectors or the generated text snippets.

[0172] The method of using a preset time series data model to extract features from the time series data in the evaluation data set to obtain a corresponding time series feature vector is as follows:

[0173] Use time series data models such as LSTM and Transformer to extract time series data features. For example, students' classroom answer data for multiple times in a semester, answer data on online platforms (multi-dimensional / multi-attribute learning data exported from the system or platform), feature extraction is as follows:

[0174] The data of students' answers in a semester. Each data includes multiple attributes such as timestamp, question number, answer correctness, answering time, etc., and uses the Transformer model (any transformer type price model is acceptable) to extract features from time series data.

[0175] An embodiment of the present application provides a feature extraction method. Since there are multiple data types in the evaluation data set, including video data, text data, and time series data, the present application uses different feature extraction models to extract feature vectors of different data types respectively, thereby improving the efficiency and specificity of data processing and providing data support for subsequent model input.

[0176] Furthermore, in step S3, the feature fusion of each of the feature vectors to obtain a fused feature vector includes:

[0177] Calculate the mean and standard deviation of each of the eigenvectors;

[0178] According to the mean and standard deviation of each of the eigenvectors, each of the eigenvectors is subjected to z-score standardization processing to obtain each corresponding standardized eigenvector;

[0179] A Transformer encoder based on a self-attention mechanism is used to perform feature fusion on each of the standardized feature vectors to obtain the fused feature vector.

[0180] In a preferred embodiment, suppose we have a student named Zhang San, whose learning data in one semester is as follows:

[0181] Video learning data: 10 videos were watched, and the frame features of each video have been extracted and normalized into a 2048-dimensional feature vector.

[0182] Text data: 5 study notes were written, and the text features of each note have been extracted and standardized into a 768-dimensional feature vector.

[0183] Time series data: 20 online test questions were completed, and the time series features of each test question have been extracted and standardized into a 128-dimensional feature vector.

[0184] In order to integrate features from different sources, feature standardization and fusion are required. Standardization steps:

[0185] 1. Calculate the mean and standard deviation: Calculate the mean and standard deviation for each feature.

[0186] 2. Standardized features: Standardize each feature by z-score

[0187] After that, feature fusion is performed to concatenate the standardized feature vectors together to form a comprehensive feature vector. Feature fusion is achieved using a Transformer encoder based on the self-attention mechanism. The self-attention mechanism can capture the relationship between the various parts of the input sequence and combine different features by weighted averaging. This method can not only handle the problem of inconsistent feature dimensions, but also establish more complex dependencies between features.

[0188] The embodiment of the present application provides a feature vector fusion method. Since each feature vector is extracted by different feature extraction models, each feature vector needs to be standardized before feature fusion, and then the Transformer encoder based on the self-attention mechanism is used to perform feature fusion on each of the standardized feature vectors. The self-attention mechanism can capture the relationship between each part of the input sequence and combine different features by weighted averaging. This method can not only deal with the problem of inconsistent feature dimensions, but also establish more complex dependencies between features. Improve the accuracy of subsequent personalized learning recommendations.

[0189] In a possible implementation, in step S4, the hybrid model evaluates the learning ability of the target student according to the fused feature vector to obtain an evaluation result, and then generates a learning path corresponding to the target student according to the evaluation result, including:

[0190] Each machine learning model in the hybrid model evaluates the learning ability of the target student according to the fused feature vector, and generates each corresponding initial evaluation result;

[0191] The hybrid model integrates the initial evaluation results according to the preset weights of the machine learning models to obtain the evaluation result;

[0192] Each machine learning model recommends a number of learning resources based on the evaluation results and the preset knowledge graph;

[0193] The hybrid model integrates and sorts a number of learning resources recommended by each machine learning model according to the preset weights of each machine learning model, and generates a learning path corresponding to the target student.

[0194] At the intersection of educational technology and artificial intelligence, the multidimensionality and complexity of data often make it difficult for a single model to provide comprehensive and accurate recommendations. This application adopts a strategy of clustered large models working together, that is, multiple specially trained large models are used to jointly analyze learners' data, and each model conducts in-depth analysis on a specific learning dimension (such as mathematical ability, language comprehension, and learning style). The outputs of these models are then integrated together to form a comprehensive, multidimensional learning feature vector, which provides data support for subsequent recommendation algorithms. For example, a mathematical model may focus on analyzing students' problem-solving ability and logical reasoning speed, while a language model analyzes grammatical comprehension and vocabulary usage frequency. After analyzing the data independently, these models will send their feature outputs to a central processing unit, which is responsible for coordinating and integrating data from different models. Through weighted algorithms, such as weighted average or weighted logistic regression, the output of each model is weighted according to its performance and reliability in the prediction task to optimize the accuracy and applicability of the overall recommendation results.

[0195] In a preferred embodiment, a formula based on a weighted ensemble learning method is designed, which can effectively integrate the advantages of multiple large models to form a comprehensive learning feature vector, as shown below:

[0196]

[0197] Among them, suppose we have n large models, each model M i Output a feature vector v i These feature vectors are integrated to form a comprehensive learning feature vector v. Among them, w i are the weights of the i-th model, which can be adjusted based on the model’s performance or expertise in a specific task. i Usually satisfied To keep the size of the feature vector appropriate.

[0198] Weight w i The determination can be based on a variety of strategies:

[0199] Performance weight: Each model is assigned a weight based on its performance on the validation set, with higher performing models receiving higher weights.

[0200] Inversely proportional to the error rate: The weight is inversely proportional to the model error rate, and the model with the lower error rate gets a higher weight.

[0201] Specialized fitness: Assign weights based on the model’s fitness for a specific learning task. For example, if a model performs well in optimizing mathematical learning paths, it will be given a higher weight for such tasks.

[0202] For example, suppose the hybrid model ensembles three large models:

[0203] Model 1 (mathematical ability analysis): feature vector V1 = [0.9, 0.1], performance score 85%;

[0204] Model 2 (language understanding analysis): feature vector V2 = [0.2, 0.8], performance score 75%;

[0205] Model 3 (learning style identification): feature vector V3 = [0.5, 0.5], performance score 90%;

[0206] If we assign weights by performance, the weights can be calculated as the ratio of performance scores:

[0207] W1=0.85 / (0.85+0.75+0.90)≈0.34

[0208] W2=0.75 / (0.85+0.75+0.90)≈0.30

[0209] W3=0.90 / (0.85+0.75+0.90)~0.36

[0210] The combined feature vector will be:

[0211] V=0.34·[0.9,0.1]+0.30·[0.2,0.8]+0.36·[0.5,0.5]=[0.53,0.47]

[0212] This comprehensive vector v reflects the learner's overall performance in mathematical ability, language comprehension, and learning style, and can be used for further personalized recommendations and path optimization. In this way, we can effectively integrate the wisdom of large models from different fields and functions to form a comprehensive and powerful recommendation system. This model collaboration approach not only improves the accuracy and adaptability of the system, but also increases the system's ability to handle complex learning scenarios.

[0213] In an embodiment of the present application, each machine learning model is clustered into a hybrid model by means of preset weights, thereby realizing the collaborative work of the clustered large model. Through the collaboration of the clustered large models and the advantages of each model in different fields, a more comprehensive learning data analysis is achieved. This method not only improves the generalization ability of the model, but also enhances the accuracy and personalization level of the recommendation. Before performing the learning ability assessment and learning path generation process, different weights can be set for each machine learning model according to the learning field and model performance, effectively integrating the wisdom of large models from different fields and functions to form a comprehensive and powerful hybrid model. This model collaboration method not only improves the accuracy and adaptability of the system, but also increases the system's ability to handle complex learning scenarios.

[0214] In a possible implementation, the method further includes, after generating the learning path, dynamically adjusting the learning path according to the learning resource preferences and real-time learning data of the target student, such as Figure 2 As shown, steps S501-S504 are included:

[0215] Step S501, obtaining the learning resource preference and real-time learning data of the target student, wherein the real-time learning data includes learning progress, real-time learning behavior data and real-time achievement data;

[0216] Step S502, inputting the real-time learning data into the hybrid model so that the hybrid model predicts the future learning performance of the target student;

[0217] Step S503: scoring each learning resource in the learning resource library and the learning path according to the learning resource preference and future learning performance;

[0218] Step S504: dynamically adjusting the learning path according to the scores of the various learning resources and the future learning performance, wherein the dynamic adjustment includes replacing or adding or deleting the learning resources in the learning path and adjusting the learning sequence in the learning path.

[0219] In general, the learner's real-time performance data at time t is defined as Pt, where Pt can be calculated based on test scores, exercise completion, and other quantifiable learning activity data. The goal of learning path adjustment is to optimize future learning outcomes, expressed as L(t+1,+)L(t+1,t+n), where n is the predicted time range. The decision function D for learning path adjustment can be expressed as:

[0220] L(t+1,t+n)=D(P t ,H t , C t ,F t ,I t , Q t )

[0221] Among them, P t is the learner performance at time t, H t is the learner’s historical data up to time t, C t is the course content and resources at time t, D is the decision function that calculates the optimal learning path based on the given performance, history and current course content, and F t Learning concentration includes learners’ concentration level in classroom and online learning. t Classroom interaction, including the frequency and quality of interaction between students and teachers and classmates. t The question-and-answer situation of classroom exercises reflects students' performance and interaction in classroom exercises.

[0222] The algorithm uses the comprehensive learning feature vector obtained from the cluster model, combined with the learner's immediate feedback and real-time learning data, to adjust the recommended learning resources and learning sequence in real time. This process not only takes into account the learner's current needs, but also predicts their future learning development direction, thereby providing a personalized and adaptable learning path. Specifically, the system first presets a learning path based on the learner's initial learning performance. As the learner progresses and provides feedback, the system collects these dynamic data and re-evaluates the effectiveness of the original path. If the system detects that the learner's learning effect in a certain area is not good, it will automatically adjust the learning resources, such as recommending more basic or intensive practice materials, or changing the teaching methods and the presentation of learning content to meet the learner's actual needs. In terms of data collection,

[0223] 1. Study concentration F t :Not only the data collected through the online learning platform (such as mouse click frequency, page dwell time, etc.), but also focus on behavioral observations in the physical classroom (information collected by digital devices under the deep learning model, such as eye tracking and facial expression analysis detected by smart devices and classroom behavior analysis, such as listening to lectures, sleeping, and other different classroom behaviors).

[0224] 2. Classroom Interaction I t : The number of students speaking, the quality of questions asked and answered, recorded by the classroom management system, and the frequency of electronic interactions recorded by the intelligent teaching system.

[0225] 3. Classroom practice Q&A t : Based on students’ performance in automated classroom tests or exercises, including analysis of accuracy, answer speed, and error types.

[0226] Suppose a student in a digital logic design course shows a decrease in learning concentration through behavioral analysis on the online platform, and performs averagely in classroom interaction and question-and-answer exercises. The system will automatically identify these problems and formulate a series of specific measures, such as adding video teaching content, providing more real-time online tests, or arranging additional one-on-one tutoring, to enhance interaction and improve learning motivation.

[0227] The embodiment of the present application provides a method for dynamically adjusting a learning path. In the prior art, although some hybrid recommendation systems attempt to improve the flexibility and adaptability of the system by combining multiple recommendation technologies, in actual operation, these systems are often difficult to respond to changes in learner behavior in real time and dynamically adjust the learning path and recommended content. In the real learning process, affected by various subjective and objective reasons, learners often find it difficult to strictly implement the preset learning path or learning plan, resulting in the actual learning progress deviating from the preset learning path, and thus unable to achieve the learning effect in theory. Therefore, it is necessary to flexibly adjust the learning path according to the actual learning effect during the learning process. The embodiment of the present application introduces a dynamic learning path optimization mechanism, which can adjust the learning path in real time according to the learner's immediate learning performance, feedback and progress. By predicting students' future academic performance through real-time learning data and dynamically adjusting the learning path according to the predicted results, the flexibility of personalized learning recommendations is greatly improved, thereby achieving a truly personalized learning experience.

[0228] In a preferred embodiment, the specific process of steps S501-S504 is:

[0229] 1) Predict learners’ future learning performance, continuously collect and monitor learners’ behavioral data, learning performance, interaction logs, etc. These data sources can be online learning platforms, smart classroom equipment, learning management systems, etc. Key data points may include: learners’ access to and use of different learning resources; learners’ performance in online tests or assessments; learners’ participation in discussion forums or interactive sessions. Use the aforementioned hybrid model to predict results and predict learners’ future learning performance based on their real-time data. For example, the model can predict the difficulties that learners may encounter in a specific course or concept, or predict learners’ interest and acceptance of specific learning materials.

[0230] 2) Implement effective feedback mechanisms to allow learners to give feedback on recommended resources and learning paths. This can be achieved through a direct evaluation system, such as learners can rate and comment on the course content, teaching methods, or the usefulness of resources.

[0231] 3) Based on the prediction results and learners' feedback, the system dynamically adjusts the learning path and resource configuration. This includes: Resource replacement and adjustment: If the prediction system finds that certain learning materials fail to achieve the expected teaching effect, or the learners show low interest in certain content, the system can recommend other more suitable learning materials. Path optimization: Adjust the difficulty or advancement speed of the course based on the learner's learning progress and mastery. For example, for learners who progress faster and have good mastery, the system can recommend more advanced learning content or additional challenges. Specifically, a new algorithm can also be designed to implement: Dynamic Learning Path Adjustment and Resource Configuration Algorithm (DLPRCA). The goal of the DLPRCA algorithm is to dynamically adjust the learning path and configure resources to maximize the learning effect. The algorithm needs to respond to the following inputs:

[0232] Learners’ real-time behavior data B, learning effectiveness feedback F, learning resource library R, ​​and preset learning path L. The DLPRCA algorithm includes the following steps:

[0233] 1. Data Collection:

[0234] Collect learners' interaction data, such as click-through rate, video viewing time, question answering, etc.

[0235] Collect learners' evaluation and feedback on the resources.

[0236] 2. Feature extraction and analysis:

[0237] Data mining technology is used to extract learner behavior characteristics and learning performance characteristics.

[0238] Analyze learners’ learning styles and preferences.

[0239] 3. Build a prediction model:

[0240] Use machine learning methods (e.g., random forests, neural networks) to train a model based on historical data to predict learners’ resource preferences and likely learning performance.

[0241] 4. Resource scoring and ranking:

[0242] Rating each available learning resource is performed based on the output of the predictive model and the learner's preferences.

[0243] Sort resources by rating, and recommend resources with high ratings first.

[0244] 5. Path adjustment decision:

[0245] Analyze learners’ current learning progress and future performance predicted by the model.

[0246] If the predicted result is lower than the expected threshold, suggestions are made to replace learning resources or modify the learning path.

[0247] 6. Dynamic adjustment:

[0248] Dynamically adjust learning paths and resource configurations based on learners' real-time feedback and prediction results.

[0249] Apply the tuning rules in feedback control theory, such as PID control strategy, to optimize the tuning process.

[0250] 7. Continuous learning and optimization:

[0251] The model continuously learns from new data and improves its prediction accuracy.

[0252] The system regularly updates the learning resource library and path design to adapt to new learning needs and technological advances.

[0253] Example: Personalized English learning path adjustment

[0254] Suppose a learner is using a large model to learn English. The DLPRCA algorithm is as follows:

[0255] The learner completed a vocabulary test, and the system collected the answer data by calling the big model API. By analyzing the test results and the learner's past learning records, the system predicted the learner's preference for different types of English learning resources (such as video courses, reading materials, and interactive exercises). The algorithm scores and recommends the resources, giving priority to video courses because the model predicts that this type of resource is most effective for learners. After the learner uses the recommended resources, the system collects feedback and finds that the learner's interest in advanced reading materials has increased. The system dynamically adjusts the recommendation strategy to increase the recommendation ratio of advanced reading materials and reduce the recommendation of basic video courses. In this way, the DLPRCA algorithm is able to achieve personalized adjustment of the learning path, making the configuration of learning resources more flexible and adaptable, thereby improving learning efficiency and learner satisfaction.

[0256] In a preferred embodiment, Figure 3 As shown, after adjusting the learning path, the embodiment of the present application will further collect learners' feedback data, and use the analysis results to optimize the data integration and feature extraction stages. The process of optimizing data integration and feature extraction based on learners' feedback data mainly involves gaining insights from feedback, and then applying these insights to improve data processing and model training. This optimization can enable the system to more accurately understand learners' needs and behaviors, thereby improving the effectiveness of personalized recommendations. Specifically:

[0257] 1) Collect feedback data: including direct feedback (such as ratings, comments) and indirect feedback (such as learning behavior patterns, speed of completing tasks, and frequently encountered difficult course parts). Collect through online surveys, built-in feedback mechanisms, learning analysis tools, etc.

[0258] 2) Data analysis and insight extraction: Statistical analysis, quantitative analysis of collected feedback, such as calculating average ratings and identifying common problems in feedback. Text analysis: sentiment analysis and keyword extraction of text feedback to understand learners’ feelings and needs. Feature extraction optimization: For example, adjusting feature weights, adjusting the weights of these features based on which features revealed in the feedback are most relevant to learner satisfaction and learning outcomes. For example, adding new features: introducing new features, such as new behavioral patterns discovered based on feedback data, such as learners spending an unusually large amount of time on a certain type of resource.

[0259] 3) Data integration optimization: In data cleaning, adjust the data cleaning rules based on feedback, such as correcting the identification rules of erroneous data. In data fusion methods, improve the integration methods between different data sources, such as optimizing data matching algorithms to reduce data loss and redundancy. In iterative testing and deployment, retrain the model using optimized features and data integration methods, and verify the improvement effect on the test set. In deployment and monitoring, deploy the optimized model to the production environment, continue to monitor new learner feedback, and form a continuous optimization cycle.

[0260] For example, in the process of language learning, we analyze the feedback of the recommended resources. Many learners give low satisfaction scores after completing the intermediate grammar course, or the frequency of using our recommended resources decreases. We analyze the collected feedback data, such as the feedback of the intermediate grammar course, and find that learners generally reflect that the difficulty of the course suddenly increases, resulting in a decline in learning experience. Feature extraction can be optimized: First, add a new feature "course difficulty change rate" to calculate the increase ratio of each course difficulty relative to the previous course. Second, adjust the feature weight: increase the weight of "learner course interruption rate" (the frequency of course unfinished exit) because it is highly correlated with learning satisfaction. Data fusion optimization: Improve the course content and learner ability matching algorithm to ensure that each learner's ability matches the course difficulty growth curve. Cleaning rule adjustment: Adjust the error data detection rules based on feedback. For example, conduct special audits on course completion records with abnormally short learning time to eliminate possible data entry errors. Therefore, long-term tracking and timely adjustment based on feedback to complete feature extraction optimization can more accurately adjust the learning path of each learner, prevent learning interruptions caused by excessive difficulty, and thus improve learners' overall satisfaction and learning effects. For example, if learners progress quickly, there is no need to practice the knowledge they have mastered step by step. This feedback-based optimization process helps to build a more humane and efficient learning environment.

[0261] Embodiment 2:

[0262] like Figure 4 As shown, accordingly, the second embodiment provides a learning recommendation system based on a cluster model, including an acquisition module 10, a feature extraction module 20, a feature fusion module 30 and a path generation module 40;

[0263] Wherein, the acquisition module 10 is used to acquire an evaluation data set of a target student, wherein the evaluation data set includes learning behavior data, performance data and personality data;

[0264] The feature extraction module 20 is used to extract features from the corresponding data in the evaluation data set using corresponding feature extraction models according to the data types in the evaluation data set, and obtain corresponding feature vectors, wherein the data types include video data, text data, and time series data.

[0265] The feature fusion module 30 is used to fuse the features of each feature vector to obtain a fused feature vector;

[0266] The path generation module 40 is used to input the fused feature vector into a preset hybrid model so that the hybrid model evaluates the target student's learning ability based on the fused feature vector to obtain an evaluation result, and then generates a learning path corresponding to the target student based on the evaluation result, wherein the hybrid model is constructed by several machine learning models and is trained based on evaluation data sets of several students, and the learning path includes several learning resources and a learning sequence.

[0267] The embodiment of the present application provides a learning recommendation method based on a cluster model. First, the evaluation data set of the target student is obtained, and different feature extraction models are used to extract features according to the data type, and then the features of each feature vector are fused to achieve integrated processing of multiple types of data, so that the hybrid model in the embodiment of the present application can efficiently process diverse data, thereby improving the adaptability of learning recommendations. At the same time, the embodiment of the present application utilizes the collaborative drive of the cluster model, integrates the model advantages of multiple fields, and evaluates the learning ability of students based on the fused fusion feature vector through collective wisdom and collaborative learning methods, and further generates corresponding learning paths based on the evaluation results, thereby achieving personalized learning recommendations. Through multi-model clustering, the accuracy of learning recommendations is significantly improved.

[0268] The more detailed working principle and step flow of this embodiment can refer to, but are not limited to, the relevant records of Embodiment 1.

[0269] Embodiment three:

[0270] Embodiment 3 provides a terminal, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor implements the cluster model-based learning recommendation method described in the present application when executing the computer program.

[0271] The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc. The processor is the control center of the terminal, and various interfaces and lines are used to connect various parts of the entire terminal.

[0272] The memory can be used to store the computer program, and the processor realizes various functions of the terminal by running or executing the computer program stored in the memory and calling the data stored in the memory. The memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, an application required for at least one function (such as a sound playback function, an image playback function, etc.), etc.; the data storage area can store data created according to the use of the mobile phone (such as audio data, a phone book, etc.), etc. In addition, the memory can include a high-speed random access memory, and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (SmartMedia Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0273] Embodiment 4:

[0274] Embodiment 4 provides a computer-readable storage medium, wherein the computer-readable storage medium includes a stored computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located is controlled to execute the cluster model-based learning recommendation method described in the present application.

[0275] Wherein, if the module integrated by the learning recommendation method based on the cluster model is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the present invention implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned various method embodiments when executed by the processor. Wherein, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form, etc. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electrical carrier signal, telecommunication signal and software distribution medium, etc.

[0276] The specific embodiments described above further describe the purpose, technical solutions and beneficial effects of the present application in detail. It should be understood that the above description is only a specific embodiment of the present application and is not intended to limit the scope of protection of the present application. It is particularly pointed out that for those skilled in the art, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of protection of the present application.

Claims

1. A learning recommendation method based on a cluster model, characterized in that: include: Acquire an evaluation data set of a target student, wherein the evaluation data set includes learning behavior data, performance data, and personality data; According to the data types existing in the evaluation data set, using respective corresponding feature extraction models to respectively extract features of the corresponding data in the evaluation data set to obtain respective corresponding feature vectors, the data types including video data class, text data class and time series data class; Performing feature fusion on each of the feature vectors to obtain a fused feature vector; The fused feature vector is input into a preset hybrid model so that the hybrid model evaluates the target student's learning ability according to the fused feature vector to obtain an evaluation result, and then generates a learning path corresponding to the target student according to the evaluation result, wherein the hybrid model is constructed by several machine learning models and trained based on evaluation data sets of several students, and the learning path includes several learning resources and learning sequence.

2. A learning recommendation method based on a cluster model as claimed in claim 1, characterized in that: The step of obtaining the target student's evaluation data set includes: Obtain the target student's learning behavior data through the learning records of the online learning platform and the video records of the offline classroom, wherein the learning behavior data includes classroom interaction data, homework and task completion status, and learning habits and resource usage data; Obtaining the target student's performance data through test and quiz performance records, the performance data including course performance, assessment and quiz performance, and comprehensive evaluation data; The personality data of the target student is obtained through the questionnaire survey results, and the personality data includes personality assessment data, learning style assessment data, and mental health status data.

3. The learning recommendation method based on cluster model as claimed in claim 1, characterized in that: The method further includes, after obtaining the evaluation data set of the target student, performing data preprocessing on the evaluation data set, including data cleaning, standardization, data synchronization, data integration, and data verification; Wherein, the data cleaning includes removing missing values, removing duplicate values ​​and correcting erroneous values; The standardization process includes unifying data format, standardizing data range, classifying coding and text data preprocessing; The data synchronization includes aligning timestamps, dividing time windows, and aggregating data within the time windows; The data integration includes building a database and fusing data; The data verification includes checking data consistency and assessing data quality.

4. The learning recommendation method based on cluster model as claimed in claim 1, characterized in that: According to the data types existing in the evaluation data set, using each corresponding feature extraction model to extract features of the corresponding data in the evaluation data set to obtain each corresponding feature vector, including: Using a preset computer vision model to extract features from the video data class data in the evaluation data set to obtain a corresponding video feature vector; Using a preset natural language processing model to perform feature extraction on the text data class data in the evaluation data set to obtain a corresponding text feature vector; A preset time series data model is used to perform feature extraction on the time series data class data in the evaluation data set to obtain a corresponding time series feature vector.

5. The learning recommendation method based on cluster model as claimed in claim 1, characterized in that: The step of fusing the feature vectors to obtain a fused feature vector includes: Calculate the mean and standard deviation of each of the eigenvectors; Performing z-score normalization processing on each of the feature vectors according to the mean and standard deviation of each of the feature vectors to obtain each corresponding normalized feature vector; A Transformer encoder based on a self-attention mechanism is used to perform feature fusion on each of the standardized feature vectors to obtain the fused feature vector.

6. A learning recommendation method based on a cluster model as claimed in claim 1, characterized in that: The hybrid model evaluates the learning ability of the target student according to the fused feature vector to obtain an evaluation result, and then generates a learning path corresponding to the target student according to the evaluation result, including: Each machine learning model in the hybrid model evaluates the learning ability of the target student according to the fused feature vector, and generates each corresponding initial evaluation result; The hybrid model integrates the initial evaluation results according to the preset weights of the machine learning models to obtain the evaluation result; Each machine learning model recommends a number of learning resources based on the evaluation results and the preset knowledge graph; The hybrid model integrates and sorts a number of learning resources recommended by each machine learning model according to the preset weights of each machine learning model, and generates a learning path corresponding to the target student.

7. A learning recommendation method based on a cluster model as described in any one of claims 1 to 6, characterized in that: The method further includes dynamically adjusting the learning path according to the learning resource preference and real-time learning data of the target student after generating the learning path, specifically: Acquire the learning resource preference and real-time learning data of the target student, wherein the real-time learning data includes learning progress, real-time learning behavior data and real-time achievement data; Inputting the real-time learning data into the hybrid model so that the hybrid model predicts the future learning performance of the target student; Scoring each learning resource in the learning resource library and the learning path according to the learning resource preferences and future learning performance; The learning path is dynamically adjusted according to the scores of the various learning resources and the future learning performance, and the dynamic adjustment includes replacing or adding or deleting the learning resources in the learning path and adjusting the learning sequence in the learning path.

8. A learning recommendation system based on a cluster model, characterized in that: It includes an acquisition module, a feature extraction module, a feature fusion module and a path generation module; Wherein, the acquisition module is used to acquire an evaluation data set of a target student, wherein the evaluation data set includes learning behavior data, performance data and personality data; The feature extraction module is used to extract features from the corresponding data in the evaluation data set using corresponding feature extraction models according to the data types in the evaluation data set, so as to obtain corresponding feature vectors, wherein the data types include video data, text data and time series data; The feature fusion module is used to fuse the features of each feature vector to obtain a fused feature vector; The path generation module is used to input the fused feature vector into a preset hybrid model so that the hybrid model evaluates the target student's learning ability based on the fused feature vector to obtain an evaluation result, and then generates a learning path corresponding to the target student based on the evaluation result, wherein the hybrid model is constructed by several machine learning models and is trained based on evaluation data sets of several students, and the learning path includes several learning resources and a learning sequence.

9. A terminal, characterized in that: The invention comprises a processor, a memory and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, a learning recommendation method based on a cluster model as described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located is controlled to execute a learning recommendation method based on a cluster model as described in any one of claims 1 to 7.

Images