Intelligent course arrangement method fusing learning characteristics of students and teaching nodes of teachers

Abstract

The invention provides an intelligent course arrangement method fusing learning characteristics of students and teaching nodes of teachers, and the method comprises the steps: obtaining original data used by daily equipment of the students, the use data comprises the use duration, the use frequency and the use time period, and obtaining a standardized equipment use data set through data preprocessing and abnormal value and noise data removal; the method comprises the following steps: acquiring use data of learnable electronic equipment, including login frequency, learning duration and interaction condition, cleaning and converting the data, and constructing a tag vector of the participation degree of a digital course; and establishing a tracking feedback mechanism of the electronic equipment use habit and the digital course participation degree of the student, regularly collecting equipment use data and course participation data of the student, updating and optimizing the electronic equipment use habit and digital course participation degree association model, and dynamically adjusting the course proportion.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to an intelligent class scheduling method integrating students' learning characteristics and teachers' teaching nodes. Background Art

[0002] In the digital teaching model, there is a complex relationship between students' electronic device usage habits and online course participation. On the one hand, the duration and mode of students' daily device use may affect their participation and learning outcomes in online courses and blended learning. Excessive use of electronic devices may also cause students to lose concentration and reduce their learning motivation in digital courses, thereby affecting learning outcomes. On the other hand, moderate use of electronic devices can help students better adapt to the digital learning environment and improve their learning efficiency. During the study, there were individual differences in the impact of increasing or decreasing screen time by 1 hour per day on digital course participation. Some students may be distracted by the extra screen time, resulting in a decrease in participation, while others may use the extra time to prepare and review the course, thereby improving participation. Such individual differences complicate the process of optimizing the ratio of online and offline courses. In addition, different types of electronic device use also have different effects on students' digital learning participation. Passive entertainment electronic device use (such as watching videos and playing games) may have a negative impact on learning participation, while active learning electronic device use (such as online discussions and completing interactive exercises) may promote increased participation. Therefore, when analyzing the relationship between electronic device usage habits and digital course participation, it is necessary to consider the type and quality of electronic device use, not just the duration of use. In short, the relationship between students' electronic device usage habits and digital course participation is affected by many factors, including duration of use, usage patterns, and individual differences. A deep understanding of the complex interactions between these factors is of great significance for optimizing digital teaching strategies and improving students' learning outcomes. Summary of the invention

[0003] The present invention provides an intelligent class scheduling method integrating student learning characteristics and teacher teaching nodes, which mainly includes:

[0004] Obtain the original data of students' daily device usage, including usage duration, usage frequency, and usage time period. Through data preprocessing, remove outliers and noise data to obtain a standardized device usage data set;

[0005] Perform feature engineering on the standardized device usage dataset to extract key features that reflect students' electronic device usage habits, including average daily usage time, high-frequency usage time periods, and device type distribution. Construct a feature vector of students' electronic device usage habits based on the key features.

[0006] Obtain learning electronic device usage data, including login frequency, learning time, and interaction status, clean and transform the data, and construct a label vector for digital course participation;

[0007] The characteristic vectors of students’ electronic device usage habits and the label vectors of digital course participation are associated, and a machine learning algorithm is used to train a prediction model of electronic device usage habits and participation to obtain the association rules and influence weights between the characteristic vectors of electronic device usage habits and the label vectors of digital course participation;

[0008] Conduct targeted analysis on the relationship between screen time and digital course participation, and count the distribution of participation under different screen time. If there is a significant difference, it is determined that there is a correlation between the electronic device usage habit feature vector and the digital course participation label vector, and determine the reasonable range of daily screen time.

[0009] According to the correlation model between electronic device usage habits and digital course participation, as well as the reasonable range of screen usage time, the ratio of online courses to offline courses is dynamically adjusted. When the predicted participation value is lower than the preset threshold, the proportion of offline courses is increased; when the predicted participation value is higher than the preset threshold, the proportion of online courses is increased. By balancing the proportion of online and offline courses, the hybrid teaching model is optimized.

[0010] Establish a tracking and feedback mechanism for students’ electronic device usage habits and digital course participation, regularly collect students’ device usage data and course participation data, and use it to update and optimize the correlation model between electronic device usage habits and digital course participation, and dynamically adjust the course ratio;

[0011] Based on the dynamic ratio of online and offline courses, combined with students' learning preferences and teachers' teaching time, an intelligent scheduling algorithm is used to automatically generate course schedules. Classes are scheduled regularly according to the latest updated course ratio, forming a dynamically updated intelligent scheduling mechanism to adapt to the changing trends in students' electronic device usage habits and digital course participation.

[0012] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0013] The present invention discloses an intelligent class scheduling method integrating student learning characteristics and teacher teaching nodes. The present invention collects students' daily equipment usage data and learning electronic equipment usage data, extracts key features reflecting students' electronic equipment usage habits, and constructs students' electronic equipment usage habit feature vectors and digital course participation label vectors. A machine learning algorithm is used to train the prediction model of electronic equipment usage habits and participation, and the association rules and influence weights between the two are obtained. The relationship between screen usage time and digital course participation is analyzed to determine the reasonable interval of screen usage time every day. According to the prediction model and the screen usage time interval, the proportion of online and offline courses is dynamically adjusted. When the participation prediction value is lower than the threshold, the proportion of offline courses is increased, and vice versa, the proportion of online courses is increased. A tracking feedback mechanism is established to regularly update the model and adjust the course ratio. Finally, based on the dynamic course ratio, an intelligent class scheduling algorithm is used to generate a course schedule. By analyzing the relationship between electronic equipment usage habits and course participation, the present invention realizes intelligent optimization of a hybrid teaching mode and improves teaching effects. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 The present invention is a flow chart of an intelligent class scheduling method that integrates student learning characteristics and teacher teaching nodes.

[0015] Figure 2 It is a schematic diagram of an intelligent class scheduling method of the present invention that integrates student learning characteristics and teacher teaching nodes.

[0016] Figure 3 This is another schematic diagram of an intelligent class scheduling method that integrates student learning characteristics and teacher teaching nodes of the present invention. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] like Figure 1-3 In this embodiment, an intelligent class scheduling method integrating student learning characteristics and teacher teaching nodes may specifically include:

[0019] Step S101, obtaining the original data of students' daily device usage, the usage data includes usage duration, usage frequency, and usage time period. Through data preprocessing, outliers and noise data are removed to obtain a standardized device usage data set.

[0020] The original record information sent by the device data acquisition interface is obtained, and the original record information includes the device startup time, end time, application identifier, operation interval duration, and number of startups per hour; outlier detection is performed using the box plot interquartile range according to the values ​​of each dimension in the original record information. If the value of any dimension in a single record is less than the lower limit of the interquartile range or greater than the upper limit of the interquartile range, it is determined to be an abnormal record; linear interpolation is used to obtain the filling value for the missing value of the duration in the abnormal record, and the nearest neighbor value is used to obtain the filling value for the missing value of the number of times; the filling postscript is interpolated through a sliding time window The records are segmented, the basic size of the sliding time window is a daily average unit, the number of device startups within the sliding time window is calculated to obtain usage frequency data, and the usage time within the sliding time window is accumulated to obtain daily average usage time data; the usage frequency data and usage time data are grouped using an affinity propagation clustering algorithm, and the affinity propagation clustering algorithm clusters the data based on a time similarity threshold; the number of startups and the usage time in each time period in the clustering results are calculated to obtain a time period activity score, and a normalized weight value is assigned to the time period whose time period activity score is greater than the activity threshold according to the score range.

[0021] Specifically, the original data records are collected from the student devices according to the fixed data sampling period, and the device startup time, end time, application identifier, operation interval, and number of startups per hour of each record are obtained from the device interface. The value type and storage format are set for each type of data, and a record set in a unified format is obtained through data standardization. For the standardized data records, the box plot interquartile range is used to detect outliers in multi-dimensional data. If the value of any dimension in a single record is less than Q1-1.5IQR or greater than Q3+1.5IQR, it is determined to be an outlier and removed. Linear interpolation is used to fill in different types of missing values. For the missing values ​​of the duration type, linear estimation is performed based on the previous and next data, and for the missing values ​​of the number type, the nearest neighbor value is used to fill. The device usage data is segmented by a sliding time window, the basic window size is set to 24 hours, the window sliding step is set to 1 hour, the number of device startups in the window is calculated to obtain the usage frequency data, and the usage time in the window is accumulated to obtain the daily average usage time data. The original window data is divided and merged based on the usage intensity to obtain multi-scale window data. The neighbor propagation clustering tool is used to group the time period data in the multi-scale window data, and the time similarity threshold is set to 0.85. The time period activity score is calculated based on the number of startups and usage time in each time period. The time period with an activity score greater than 0.75 is extracted, and the time period data is assigned a normalized weight value of 0.1-1.0 according to the activity score range of 0.75-1.0 to obtain the final data set. In the data collection of student devices, the interface sets the data sampling period to 300 seconds, continuously records the device startup status, and the interface collects data including timestamp, application ID, screen brightness, and application running status mark. The timestamp accuracy is milliseconds, the application ID uses a 32-bit string, the screen brightness value range is 0-255, and the application running status mark value is 0 or 1. The collected raw data is stored as a binary file in a unified format, and the data standardization process uses maximum and minimum value normalization, and the data value is mapped to the 0-1 range. The data outlier detection adopts the box plot method. By calculating the quartiles Q1, Q3 and the interquartile range IQR, the outlier judgment boundary is set. The detection results show that there are 87 abnormal records in the startup duration data and 92 abnormal records in the application switching frequency data. These abnormal records are eliminated and the missing data are filled with linear interpolation. The duration data is filled with the linear estimation value of the previous and next data points, and the frequency data is filled with the nearest neighbor value. The data segmentation processing adopts a sliding window. The basic window size is set to 24 hours and the sliding step is 1 hour. The device usage frequency data is accumulated and calculated in the window. The window data is divided into three scales based on the usage intensity: high frequency band, medium frequency band, and low frequency band. The corresponding window sizes are adjusted to 6 hours, 12 hours, and 24 hours respectively.The time period data are grouped using proximity propagation clustering, and the cluster similarity threshold is set to 0.85. The number of startups and usage duration in each time period are calculated to obtain the activity score. The time periods with scores greater than 0.75 are extracted, and weight values ​​are assigned to the time period data based on the activity scores. The weight values ​​are mapped linearly from activity scores of 0.75-1.0 to weight values ​​of 0.1-1.0.

[0022] Step S102, perform feature engineering on the standardized device usage data set to extract key features that reflect students' electronic device usage habits, including average daily usage time, high-frequency usage time periods, and device type distribution, and construct a student electronic device usage habit feature vector based on the key features.

[0023] A sliding time window is used to count the device usage data, and the window size of the sliding time window is a preset time. The number of times the device screen is turned on and off and the interactive behavior data are recorded by a counter; based on the device usage data, a Gaussian kernel density estimation tool is used to calculate the probability density distribution of the usage time, and the time interval whose usage frequency exceeds a set threshold is extracted from the probability density distribution to obtain the high-frequency usage time period data; the feature importance of the high-frequency usage time period data is sorted, and redundant features are eliminated by calculating the Pearson correlation coefficient, and the principal component tool is used to perform dimensionality reduction processing on the feature data to obtain reduced dimensionality feature data; statistical parameters are calculated based on the reduced dimensionality feature data, and the partitioning and clustering method is used to group the time period data to obtain time period features, and the device type with the highest proportion is extracted from the device type data; for the time period features and the device type, a feature vector is constructed in combination with the device interactive behavior data, and the feature vector includes duration statistical features, time period distribution features, device type features, interactive behavior features and scenario features.

[0024] Specifically, according to the timestamp mark of each record in the normalized data set, the device usage data is obtained by using a sliding time window. The window size is set to 24 hours, the sliding step is 1 hour, and the number of times the device screen is turned on and off is recorded by a counter. The device type identifier is recorded. The device type identifier is encoded with 16 digits. The continuous usage time within a single day is associated with the device interaction behavior data, and the original feature data set is constructed based on the scene information. The Gaussian kernel density estimation tool with a fixed bandwidth value of 0.5 is used to calculate the probability density distribution of the usage time data in the original feature data set. The interval with a usage frequency ratio of more than 0.6 is extracted from the time period data as a high-frequency period. The usage time ratio within 24 hours is calculated for different device types. The standardized feature data is obtained by normalization of the maximum and minimum values, and the click interval, sliding amplitude, and input speed features are extracted from the device interaction behavior data. The feature importance is sorted for the standardized feature data, the threshold of the Pearson correlation coefficient is set to 0.7, and redundant features with high correlation between features are eliminated. The principal component tool is used for feature dimensionality reduction, and the threshold of the feature contribution rate is set to 0.6. The feature dimensions with cumulative contribution rates exceeding the threshold are retained. The reduced-dimensional data are standardized in combination with the scene label information to obtain the reduced-dimensional feature data. The mean, standard deviation, and kurtosis value of the usage time are calculated based on the reduced-dimensional feature data. The number of clusters is set to 3. The usage time data is grouped by dividing the clusters to obtain the time period characteristics. The device type with the highest usage share is extracted from the device type data. A 20-dimensional feature vector is constructed by combining the device interaction characteristics and scene characteristics. The feature vector contains 4-dimensional duration statistical features, 6-dimensional time period distribution features, 4-dimensional device type features, 4-dimensional interactive behavior features, and 2-dimensional scene features. The device usage characteristics are obtained from the student device record data. The sliding time window size is set to 24 hours. The device status is recorded once every hour, including the screen switch count, usage time, and device identification. The device type identification adopts a 16-bit encoding format. The first 4 digits represent the device brand, the middle 8 digits represent the specific model, and the last 4 digits represent the operating system version. The number of screen switches recorded in a single day ranges from 0 to 288 times, and the continuous usage time ranges from 0 to 86400 seconds. The scene information includes three types of scene tags: indoor, outdoor, and transportation. The probability density distribution calculation uses the Gaussian kernel function, and the kernel function bandwidth is set to 0.5. The density of the 0-24 hour time period data is estimated. The calculation results show that the usage frequency in the three time periods of 8:00-12:00, 14:00-17:00, and 19:00-22:00 exceeds 0.6. These time periods are marked as high-frequency usage periods. The usage time ratio of the device type is calculated by the cumulative usage time of 24 hours. Among the device interaction behavior characteristics, the average click interval is 1.2 seconds, the average sliding amplitude is 320 pixels, and the average text input speed is 80 words per minute.The feature importance ranking shows that the Pearson correlation coefficients of duration statistical features, high-frequency time period features, device type features, interactive behavior features, and scene features are 0.85, 0.78, 0.72, 0.68, and 0.63, respectively. After removing redundant features with correlation coefficients greater than 0.7, the reduced-dimensional feature data are constructed into a 20-dimensional feature vector. The duration statistical features include four dimensions: mean, standard deviation, kurtosis, and coefficient of variation. The time period distribution features include six dimensions: frequency of use and duration in the morning, noon, and evening. The device type features include four dimensions: proportion of main device usage time, device switching frequency, device online time, and device offline time. The interactive behavior features include four dimensions: click frequency, sliding amplitude, input speed, and operation continuity. The scene features include two dimensions: scene change frequency and scene stay duration.

[0025] Step S103, obtaining the usage data of the learning electronic device, including the login frequency, learning time, and interaction status, cleaning and converting the data, and constructing a label vector of the digital course participation.

[0026] The login data sent by the terminal device is obtained, wherein the login data includes the login time, the logout time and the online time, and the number of test answers, the number of discussion replies, the resource browsing time, and the teacher-student interaction record are obtained from the teaching platform interface; according to the login data and the teaching platform data, the box plot quartile interval limit parameters are used for anomaly detection, and the abnormal data is replaced by forward data filling to obtain a standardized data matrix; for the standardized data matrix, a radial basis function regression tool is used to construct a participation quantification indicator, and the kernel function bandwidth of the radial basis function is set to a first preset threshold, and the regularization parameter is set to a second preset threshold to obtain a basic participation evaluation value; the basic participation evaluation value is subjected to bottom-up clustering processing, and the distance between participation features is calculated using the Euclidean distance to obtain five levels of participation grouping data;

[0027] For the participation grouping data, a linear weight value in a preset interval is assigned, and an eight-dimensional label vector is constructed by combining the online time proportion, interaction frequency proportion, resource access degree, and homework completion rate.

[0028] Specifically, according to the terminal records of the learning device, a 30-minute fixed time window is used to record login data, including login time, logout time, and online time. The number of test answers, number of discussion replies, resource browsing time, and teacher-student interaction records are obtained from the teaching platform data interface. The homework submission time and completion progress are recorded, and converted into a unified key-value pair format through the data interface. The data is stored as a binary file to obtain the original data set. Anomaly detection is performed on the numerical data in the original data set, and the box plot quartile interval boundary parameters Q1-1.5IQR to Q3+1.5IQR are set. Data outside the interval is eliminated, and forward data filling is used to replace missing values. The data is normalized with zero mean, converted into a row vector format, and a standardized data matrix is ​​constructed in combination with time series features. The radial basis function regression tool is used to construct a quantitative indicator of digital participation. The kernel function bandwidth is set to 0.8, the regularization parameter is set to 0.1, and the input features include online time ratio, interaction frequency score, resource access depth, and homework completion rate. The basic evaluation value of participation is calculated, and the indicator is corrected in combination with time series data to obtain basic label data. Clustering is performed on the basic label data, and bottom-up clustering tools are used to group data. The clustering distance uses the Euclidean distance, and the participation characteristics are divided into 5 levels. Linear weight values ​​of 0.2 to 1.0 are assigned to labels of different levels. An 8-dimensional label vector is constructed by combining the proportion of online time, the proportion of interaction frequency, the degree of resource access, and the homework compliance rate. The learning device data collection uses a 30-minute fixed window record. The device login and logout timestamps recorded in the window are accurate to seconds. Statistics show that the number of student logins in a single day is distributed in the range of 4-12 times, and the single online time is distributed in the range of 15-120 minutes. The data records generated by each login include 0-5 test answers, 0-8 discussion replies, 0-45 minutes of resource browsing time, and 0-6 teacher-student interactions. The homework submission record includes the submission timestamp and completion percentage. During the data cleaning phase, outliers outside the quartile range were identified. Data showed that records with online time exceeding 180 minutes, interaction frequency exceeding 10 times per hour, and resource access depth exceeding 60 minutes were marked as abnormal and removed. The previous record data was used to fill in missing values, and data standardization used the zero mean method to normalize the values ​​to the range of -1 to 1.The participation metric was quantified using radial basis function regression, with a kernel bandwidth of 0.8 and a regularization parameter of 0.1. The results showed that the score for online users who were online for more than 4 hours a day was 0.8-1.0, 2-4 hours was 0.6-0.8, 1-2 hours was 0.4-0.6, 0.5-1 hour was 0.2-0.4, and less than 0.5 hours was 0-0.2; the interaction frequency was 0.8-1.0 for more than 5 times per hour, 0.6-0.8 for 3-5 times, 0.4-0.6 for 2-3 times, and 0.1-2 times. 2-0.4, 0-0.2 for less than 1 time; 0.8-1.0 for resource access over 30 minutes, 0.6-0.8 for 20-30 minutes, 0.4-0.6 for 10-20 minutes, 0.2-0.4 for 5-10 minutes, 0-0.2 for less than 5 minutes; 0.8-1.0 for homework completion rate over 90%, 0.6-0.8 for 80-90%, 0.4-0.6 for 70-80%, 0.2-0.4 for 60-70%, 0-0.2 for less than 60%. The 8-dimensional label vector finally constructed contains 2 dimensions each for online time score, interaction frequency score, resource access score, and homework completion score, and the weight value of each dimension is linearly distributed in the range of 0.2-1.0.

[0029] Step S104, associating the student electronic device usage habit feature vector and the digital course participation label vector, using a machine learning algorithm to train a prediction model of electronic device usage habits and participation, and obtaining association rules and influence weights between the electronic device usage habit feature vector and the digital course participation label vector.

[0030] Obtain the usage habit feature vector and the participation label vector, and obtain normalized feature data through minimum and maximum value normalization processing; for the normalized feature data, use the Pearson correlation coefficient to calculate the significant correlation value, if the correlation coefficient is less than the preset threshold, then eliminate the feature dimension, and use the variance inflation factor test to eliminate the collinearity feature to obtain the preprocessed data set; stratify according to the participation label value interval in the preprocessed data set, use the stratified sampler to divide the training data set and the test data set, and train the training data set through the random forest regressor to obtain the prediction model; for the prediction model, use the Gini coefficient to calculate the feature purity change value, combine the information gain ratio to verify the feature importance, and calculate the feature combination importance through cross-validation to obtain the feature influence ranking result; for the feature influence ranking result, extract the association rules with weight values ​​higher than the preset threshold, use the decision tree based on the information gain ratio to perform path analysis on the association rules, and perform post-pruning through the error cost function to obtain the optimal rule subset.

[0031] Specifically, the minimum and maximum value standardization is used to normalize the usage habit feature vector and the engagement label vector. The two sets of vector data are aligned by timestamp, and the alignment window is set to 24 hours. The Pearson correlation coefficient is used to calculate the significant correlation value between the two sets of vectors. The feature dimensions with a correlation coefficient less than 0.6 are eliminated. The variance inflation factor test is used to eliminate collinear features, and the data consistency is checked to obtain the preprocessed data set. The preprocessed data set is stratified according to the engagement label value interval, and the training data set and the test data set are divided into a training data set and a test data set in an 8:2 ratio by a stratified sampler. The training data set is trained by a random forest regressor, and the number of trees is set to 200, the maximum depth is 12, and the minimum number of leaf node samples is 10. The loss function uses the mean square error. The root mean square error and the coefficient of determination are calculated on the test data set to evaluate the prediction accuracy. The feature importance score is calculated according to the prediction model, and the contribution weight of each dimension in the usage habit feature vector to the engagement label vector is calculated. The Gini coefficient is used to calculate the feature purity change value, and the feature importance is verified twice in combination with the information gain ratio. The feature combination importance is calculated by cross-validation to obtain the feature influence ranking result. The association rules with weight values ​​greater than 0.3 were extracted from the feature influence ranking results. The decision tree tool based on information gain ratio was used to perform path analysis on the high-weight feature rules. The maximum depth was set to 6, the minimum number of split samples was set to 20, and the error cost function was used for post-pruning. The optimal rule subset was selected through cross-validation to obtain the final association rule set. The usage habit feature vector and the engagement label vector were processed. The feature vector contained 20 dimensions and the label vector contained 8 dimensions. A 24-hour time window was used for data alignment. The data status was recorded once an hour. The Pearson coefficient calculation showed that the correlation coefficient between the usage time feature and the engagement was 0.82, the correlation coefficient between the usage frequency feature was 0.75, the correlation coefficient between the device switching feature was 0.68, the correlation coefficient between the interactive behavior feature was 0.64, and the correlation coefficient between the scene feature was 0.58. The correlation coefficients of other features were all lower than 0.6 and were eliminated. The variance inflation factor test results showed that there was collinearity between the usage time and the usage frequency, and the usage time feature with a higher correlation coefficient was retained. In the data stratified sampling, the engagement label values ​​are divided into 5 layers at an interval of 0.2, and the number of samples in each layer is 2000, 2500, 3000, 2800, and 2200 respectively. In the random forest training, the maximum depth of 200 decision trees is 12, the minimum number of leaf node samples is 10, the root mean square error on the test set is 0.15, and the determination coefficient is 0.83. The feature importance evaluation shows that the usage time feature importance score is 0.35, the interactive behavior feature score is 0.28, the device switching feature score is 0.22, and the scene feature score is 0.15. The feature purity change values ​​calculated by the Gini coefficient are 0.32, 0.25, 0.20, and 0.13 respectively, and the information gain ratio verification result is similar.In the decision tree rule extraction, the depth is set to 6, the minimum number of split samples is 20, 87 rule paths are retained after pruning by the error cost function, the cross-validation accuracy is 0.81, and finally 52 association rules with weight values ​​greater than 0.3 are extracted.

[0032] Step S105, conduct targeted analysis on the relationship between screen usage time and digital course participation, and count the distribution of participation under different screen usage times. If there is a significant difference, it is determined that there is a correlation between the electronic device usage habit feature vector and the digital course participation label vector, and determine the reasonable range of daily screen usage time.

[0033] A fixed time window is used to divide the screen usage data, and a duration data set is obtained through segmented statistics, wherein the duration data set includes a cumulative usage duration value, a usage count value, and a unit operation frequency value; the mean, standard deviation, kurtosis, and skewness of the engagement are calculated based on the duration data set, and a variance test is performed on the engagement data using a one-way analysis of variance tool. The significance of the interval is determined based on the F statistic to obtain a significant interval set; for the significant interval set, a reflection kernel density estimation tool is used to fit the engagement distribution, and the kernel function bandwidth of the reflection kernel density estimation tool is determined through cross-validation. Interval statistics are performed on the engagement distribution to obtain a data stability index; the data stability index is detected by a boundary through an adaptive sliding window, and the adaptive sliding window calculates the engagement change rate value and the stability index value during the detection process; a curvature-based inflection point detection tool is used to identify the boundary of the engagement change rate value, and the boundary position is corrected according to the stability index value to obtain a reasonable time interval set.

[0034] Specifically, the screen usage data is divided into fixed time windows, the window size is set to 24 hours, and the original data is segmented and counted at intervals of 15 minutes. The cumulative usage time, number of times used, and number of operations per unit time in each time period are calculated. The data noise is eliminated by moving average smoothing, and the data is tested for normality. The usage time distribution table is constructed to obtain the screen usage data set. According to the screen usage data set, the mean, standard deviation, kurtosis, and skewness of the participation in each time period are calculated. The one-way variance tool is used to test the homogeneity of variance of the participation data in different time intervals. The significance level is set to 0.01, and the significance of the difference between groups is calculated by the F statistic. The Tukey test is used for multiple comparisons, and the intervals with a significance greater than 0.01 are eliminated to obtain the significant interval set. For the significant interval set, the density estimation tool based on the reflection kernel is used to fit the participation distribution. The kernel function bandwidth is determined to be 0.4 through cross-validation. The interval statistics of the participation value distribution are performed, and the data stability index in each interval is calculated. The time interval with the highest participation mean and a volatility lower than 0.1 is extracted to obtain the optimal interval range. The optimal interval range is detected by adaptive sliding window. The initial window size is set to 30 minutes and the step length is 10 minutes. The participation change rate and stability index in the window are calculated. The curvature-based inflection point detection tool is used to identify the boundary points. The curvature threshold is set to 0.15. The boundary is corrected by the mean square error criterion to obtain a reasonable time interval. The student screen usage data is divided into time windows, with each 24 hours as a complete window. The window data is recorded in segments at 15-minute intervals. The statistical results show:

[0035] The average daily usage time is distributed in the range of 2-12 hours, and the number of operations per hour is distributed in the range of 20-180 times. The data fluctuations are eliminated by 5-point moving average. The data skewness coefficient is 0.85, the kurtosis coefficient is 2.4, and the P value of the normality test is 0.032. The participation data statistics show that the participation mean of 0-2 hours is 0.35 and the standard deviation is 0.12, the mean of 2-4 hours is 0.52 and the standard deviation is 0.15, the mean of 4-6 hours is 0.78 and the standard deviation is 0.11, the mean of 6-8 hours is 0.82 and the standard deviation is 0.09, the mean of 8-10 hours is 0.75 and the standard deviation is 0.14, and the mean of 10-12 hours is 0.61 and the standard deviation is 0.18. The F value of the variance homogeneity test is 23.6, and the P value is less than 0.01. The results of multiple comparisons show that the 4-8 hour interval is significantly different from other intervals. The participation density estimation uses the reflection kernel function, and the bandwidth is determined to be 0.4 through ten-fold cross validation. The participation distribution in the 4-8 hour interval is unimodal, with a peak at 6.5 hours. The data volatility in this interval is 0.08. The boundary detection uses a 30-minute window and a 10-minute step size for sliding. The calculated participation change rate curve has inflection points at 4.2 hours and 7.8 hours, with curvature values ​​of 0.17 and 0.16 respectively. The final boundary is determined to be 4-8 hours by minimizing the mean square error.

[0036] Step S106, dynamically adjust the ratio of online courses to offline courses based on the correlation model between electronic device usage habits and digital course participation, as well as the reasonable range of screen usage time. When the participation prediction value is lower than the preset threshold, increase the proportion of offline courses; when the participation prediction value is higher than the preset threshold, increase the proportion of online courses. By balancing the ratio of online and offline courses, the hybrid teaching mode is optimized.

[0037] A moving time window is used to collect electronic device usage data, and the electronic device usage data is processed by a stochastic gradient regressor to obtain a participation prediction value; a participation prediction sequence is generated according to the participation prediction value, and an exponentially weighted moving average is used to calculate the reference value for the participation prediction sequence; upper and lower thresholds are calculated for the reference value, and if the participation prediction value shows an upward trend for three consecutive times, the upper threshold is increased by a preset proportion, and the lower threshold is increased by a preset proportion to obtain an adjustment signal; a proportional-integral regulator is used to calculate a course proportion adjustment amount according to the adjustment signal, and if the participation prediction value is lower than the lower threshold, the online course proportion is reduced, and the reduction amount is proportional to the degree of deviation from the lower threshold; the course proportion adjustment amount is processed by a cubic spline smoother, and the course proportion adjustment amount is smoothed by an exponential moving average to obtain an actual course proportion.

[0038] Specifically, the participation prediction value is calculated in real time based on the electronic device usage data, and a moving time window is used for data collection. The window length is set to 3 days and the sliding step is 12 hours. The device usage characteristics are predicted in real time through a stochastic gradient regressor, and the learning rate is set to 0.01. The prediction model parameters are updated every hour, and the root mean square error is calculated to track the prediction accuracy. The data with a prediction error exceeding 0.15 is removed to obtain the participation prediction sequence. A dynamic threshold is set for the participation prediction sequence, and the reference value is calculated using an exponentially weighted moving average. The smoothing coefficient is set to 0.3. The upper and lower thresholds are calculated based on the reference value. The threshold interval is dynamically adjusted according to the participation change trend. If the predicted value shows an upward trend for three consecutive times, the upper threshold is increased by 0.05 and the lower threshold is increased by 0.03. If it shows a downward trend for three consecutive times, the upper and lower thresholds are correspondingly reduced. The predicted value is segmented to obtain an adjustment signal. The course proportion adjustment is calculated according to the adjustment signal, and the proportional integral regulator is used for incremental calculation. The integral time constant is set to 2 days, the proportional coefficient is set to 0.1, and the single adjustment is limited to 10%. If the participation rate is lower than the lower threshold, the online course proportion is adjusted downward, and the downward adjustment amount is proportional to the degree of deviation from the threshold. If it is higher than the upper threshold, it is adjusted upward, and the upward adjustment amount is proportional to the degree of exceeding the threshold, and the course proportion adjustment value is obtained. The course proportion adjustment value is processed by a cubic spline smoother, and the control point interval is set to 8 hours. The proportion of online and offline courses is calculated, and the upper limit of the online course proportion is set to 70%, and the lower limit is 30%. The adjusted proportion is constrained, and the exponential moving average with a buffer factor of 0.8 is used to achieve gradual adjustment of the course proportion, and the actual course proportion is updated every 4 hours. In the process of implementing participation prediction and course ratio adjustment, a 3-day sliding window is used to collect device usage data, and the data set is updated every 12 hours. The stochastic gradient regressor is trained in real time with a learning rate of 0.01. The prediction results show that the root mean square error of participation prediction fluctuates between 0.12-0.18, and the prediction accuracy reaches 83%. After removing abnormal prediction values ​​with an error of more than 0.15, the prediction sequence shows obvious periodic changes. In the calculation of participation threshold, the initial reference value of the exponential weighted moving average is set to 0.7. After trend adjustment, the upper threshold fluctuates in the range of 0.75-0.85, and the lower threshold fluctuates in the range of 0.55-0.65. Threshold adjustment is triggered when the prediction value changes by more than 0.05 for three consecutive times. The course ratio adjustment adopts proportional integral adjustment with an integral time constant of 2 days. When the participation is lower than the lower threshold of 0.2, a maximum of 10% reduction is triggered, and when it is lower than 0.1, a 5% reduction is triggered. When it is higher than the upper threshold of 0.1, a 5% increase is triggered, and when it is higher than 0.2, a 10% increase is triggered.In the proportion smoothing process, the cubic spline smoother calculates the adjustment curve at the control points at 8-hour intervals. The actual proportion of online courses fluctuates in the range of 30%-70%. The exponential moving average with a buffer factor of 0.8 is used to achieve gradual adjustment every 4 hours. The actual adjustment range of a single time is limited to 3%, and the entire adjustment process lasts about 48 hours.

[0039] Step S107, establish a tracking and feedback mechanism for students' electronic device usage habits and digital course participation, regularly collect students' device usage data and course participation data, which are used to update and optimize the electronic device usage habits and digital course participation association model, and dynamically adjust the course ratio.

[0040] The electronic device usage data and course participation data are obtained, wherein the electronic device usage data and course participation data include device usage time, number of interactions, resource access and homework submission rate, and a real-time data set is obtained through data consistency verification; based on the real-time data set, an exponentially weighted moving average is used to calculate the participation change rate, and an adaptive threshold calculator is used to determine the moment of significant participation change. If the participation prediction error exceeds a preset threshold, an associated model update signal is triggered; in response to the associated model update signal, a kernel function online learning regressor is used to update the model parameters, the learning rate is adjusted through exponential decay, and a model stability index is calculated based on real-time data; model performance is evaluated based on the model stability index, and if the model stability index is lower than a preset threshold value, a model rollback mechanism is triggered, and an updated model performance index is obtained through cross-validation; in response to the model performance index, an adaptive proportional-integral regulator is used to calculate the course proportion adjustment signal, and the course proportion is dynamically adjusted through Bezier curve spline smoothing to obtain a final course adjustment plan.

[0041] Specifically, electronic device usage data and course participation data are collected at minute-level time intervals, and the collection interval is set to 15 minutes. The device usage time, usage frequency, number of interactions, resource access, and homework submission rate are obtained through the data collection interface, and the participation score, completion progress, and interaction score are recorded. The collected data is timestamped and a version number is added. The data version is identified by a hash string, and abnormal records are removed through data consistency verification to obtain a real-time data set. The multi-dimensional participation change trend is calculated based on the real-time data set, and the participation change rate is calculated using an exponentially weighted moving average. The multi-level smoothing coefficient is set to 0.2 to 0.6, and the data points with a change rate exceeding 0.05 are recorded. Combined with the resource access change rate and the homework submission rate change trend, the adaptive threshold calculator is used to determine the moment of significant change in participation. When the participation prediction error exceeds 0.15, the associated model update signal is triggered. For the update signal of the associated model, the model parameters are updated by the online learning regressor based on the kernel function. The initial learning rate is set to 0.02, and the learning rate is adjusted to the range of 0.01 to 0.2 through exponential decay. The model stability index is calculated based on the latest 1000 data. If the stability index is lower than 0.8, the model rollback mechanism is triggered, and the performance of the updated model is evaluated through 5-fold cross validation. The course proportion is dynamically adjusted by the adaptive proportional integral regulator, and the integral time constant changes adaptively in the range of 12 to 36 hours. The adjustment signal is calculated based on the participation prediction value, change trend, and stability. The cubic spline smoothing based on the Bezier curve is used to achieve proportional adjustment. The minimum adjustment step is set to 2%, the maximum adjustment range is 15%, and the adjustment time span is controlled to be between 24 and 72 hours. In the process of tracking and feedback on the association between electronic device usage habits and engagement, the data collection interval is set to 15 minutes. The data items collected each time include: device usage time, screen unlock times, application switching frequency, resource access number, and homework submission number. The engagement-related data include: course completion, interaction score, and test score. The data version uses 32-bit hash coding, and the data consistency check eliminates 3.2% of abnormal records. The engagement trend calculation uses multi-level exponential smoothing, and the smoothing coefficient is set to 5 gears from 0.2 to 0.6. The calculation results show that the participation change rate fluctuates between 0.02 and 0.08, the resource access change rate fluctuates between 0.03 and 0.09, and the homework submission rate change trend fluctuates between -0.05 and 0.07. The adaptive threshold is dynamically adjusted from the initial 0.15 to 0.12. The kernel function of the online learning regressor adopts radial basis function. The initial learning rate of 0.02 is adjusted by exponential decay. When the prediction error exceeds 0.15, the learning rate increases to 0.2. When the prediction error is less than 0.05 for 100 consecutive times, the learning rate decreases to 0.01. The model stability index is maintained above 0.85.In the course proportion adjustment, the integral time constant varies between 12 and 36 hours according to the degree of participation fluctuation. The adjustment signal calculation adopts the predicted value with a weight of 0.4, the change trend with a weight of 0.3, and the stability index with a weight of 0.3. A smooth transition is achieved through the fourth-order Bezier curve. The actual adjustment range is controlled between 2% and 15%, and the complete adjustment cycle lasts an average of 48 hours.

[0042] Step S108, based on the dynamic ratio of online and offline courses, combined with students' learning preferences and teachers' teaching time, an intelligent scheduling algorithm is used to automatically generate a course schedule. Classes are scheduled regularly according to the latest updated course ratio, forming a dynamically updated intelligent scheduling mechanism to adapt to the changing trends of students' electronic device usage habits and digital course participation.

[0043] Obtain course constraints, prioritize the constraints according to their importance, set hard constraint weights and soft constraint weights for the constraints, and construct a constraint evaluation system through the hard constraint weights and the soft constraint weights to obtain basic course scheduling data; use an improved genetic algorithm to perform calculations based on the basic course scheduling data, set a real number encoding method and an initial population size for the improved genetic algorithm, calculate the population fitness value through a fitness function, and use a roulette wheel selection operator to select the population fitness value to obtain an initial course scheduling plan; use an adaptive local search tool for the initial course scheduling plan, and use the basic genetic algorithm to calculate the population fitness value, and use the roulette wheel selection operator to select the population fitness value. A search radius value is set for the total number of courses, and the time conflict and resource utilization in the initial course scheduling plan are detected. If the detection result exceeds a preset threshold, a local adjustment gain value is calculated; course participation data is obtained through a real-time monitor. If the change range of the course participation data exceeds a preset threshold value, a course scheduling update instruction is triggered, and a basic update cycle value is set according to the course scheduling update instruction; a time window incremental update method is adopted for the basic update cycle value, wherein a participation trend value is calculated through a sliding window, and an update cycle adjustment interval is determined according to the participation trend value, and the initial course scheduling plan is dynamically adjusted to obtain a final course scheduling plan.

[0044] Specifically, the scheduling constraints are constructed according to the dynamic ratio of online and offline courses, and the constraints are prioritized. Among them, teacher schedule occupancy, classroom resource conflict, and course time continuity are set as hard constraints, and student course selection preferences, teacher satisfaction, and resource utilization are set as soft constraints. The weights of the constraints are calculated through the hierarchical evaluation tool, and the hard constraint weight is set to 0.7 and the soft constraint weight is set to 0.3. The course time distribution, resource utilization, and student feedback data are statistically analyzed to construct a constraint evaluation system based on weighted scores to obtain the basic data for scheduling. The improved genetic algorithm is used to calculate the basic data for scheduling. The gene encoding uses real number encoding, the chromosome length is set to the total number of courses, the initial population size is 500, the crossover probability is dynamically adjusted from 0.6 to 0.9, the mutation probability range is 0.05 to 0.15, the number of iterations is 2000, and the fitness function includes three dimensions: time conflict, resource utilization, and satisfaction score, with weights of 0.5, 0.3, and 0.2 respectively. The initial scheduling plan is obtained by population evolution through the roulette selection operator. For the initial class scheduling plan, an adaptive local search tool is used for optimization. The search radius is dynamically set to 5% to 15% of the total time period based on the number of courses. Priority is given to areas where the time conflict exceeds 0.15 or the resource utilization rate is less than 0.6. The local adjustment gain is calculated based on the teacher schedule matching, course continuity, and resource concentration. The gain threshold is set to 0.1, and the adjustment plan that exceeds the threshold is selected for class scheduling optimization. The course participation data is detected through a real-time monitor. When the participation change exceeds 0.08 or is lower than the threshold of 0.75 twice in a row, the class schedule update is triggered. The basic update cycle is set to 7 days, and the participation trend is calculated through a sliding window. The update cycle is dynamically adjusted to the range of 5 to 10 days according to the trend change. The existing class scheduling plan is adjusted using an incremental update method based on a time window to maintain the continuity of the course before and after, and the number of courses adjusted at a single time is controlled to not exceed 20% of the total. In the process of implementing intelligent class scheduling, the priority division of constraints shows that: the teacher schedule occupancy weight is 0.3, the classroom resource conflict weight is 0.2, and the course continuity weight is 0.2 as hard constraints, and the student course selection preference weight is 0.1, the teacher satisfaction weight is 0.1, and the resource utilization weight is 0.1 as soft constraints. The hierarchical evaluation calculation results show that the scores of each constraint condition are distributed between 0.65 and 0.92. The genetic algorithm uses a 500-scale population for optimization, and the chromosome encoding uses 48-bit real numbers to represent the course schedule of 24 time periods. The initial value of the crossover probability is 0.8, which gradually decreases to 0.6 as the fitness increases. The initial value of the mutation probability is 0.05, which increases to 0.15 as the evolutionary generation increases. After 2000 iterations, the time conflict degree is reduced to 0.08, the resource utilization rate is increased to 0.82, and the satisfaction score reaches 0.88.In local search optimization, for the scheduling of 120 courses, the search radius was set to 12 time periods, focusing on optimizing 15 areas with time conflicts exceeding 0.15 and 8 areas with resource utilization rates below 0.6. The teacher schedule matching degree increased from 0.75 to 0.89, the course continuity index increased from 0.68 to 0.85, and the resource concentration increased from 0.72 to 0.88. Participation monitoring uses a 7-day basic cycle. When the participation rate drops from 0.85 to 0.76 or falls below 0.75 twice in a row, an update is triggered. Through trend analysis of the data in the last 30 days, the update cycle is dynamically adjusted to 5 to 10 days. A single update adjusts 18 courses, accounting for 15% of the total number of courses, and the course connectivity remains above 0.92.

[0045] The above only lists some preferred embodiments of the present invention, but the present invention is not limited thereto, and many improvements and changes can be made. As long as the improvements and changes are made on the basis of the basic principles of the present invention, they should be regarded as falling within the protection scope of the present invention.

Claims

1. An intelligent class scheduling method integrating students' learning characteristics and teachers' teaching nodes, characterized in that: The method comprises: Obtain the original data of students' daily device usage, including usage duration, usage frequency, and usage time period. Through data preprocessing, remove outliers and noise data to obtain a standardized device usage data set; Perform feature engineering on the standardized device usage dataset to extract key features that reflect students' electronic device usage habits, including average daily usage time, high-frequency usage time periods, and device type distribution. Construct a feature vector of students' electronic device usage habits based on the key features. Obtain learning electronic device usage data, including login frequency, learning time, and interaction status, clean and transform the data, and construct a label vector for digital course participation; The characteristic vectors of students’ electronic device usage habits and the label vectors of digital course participation are associated, and a machine learning algorithm is used to train a prediction model of electronic device usage habits and participation to obtain the association rules and influence weights between the characteristic vectors of electronic device usage habits and the label vectors of digital course participation; Conduct targeted analysis on the relationship between screen time and digital course participation, and count the distribution of participation under different screen time. If there is a significant difference, it is determined that there is a correlation between the electronic device usage habit feature vector and the digital course participation label vector, and determine the reasonable range of daily screen time. According to the correlation model between electronic device usage habits and digital course participation, as well as the reasonable range of screen usage time, the ratio of online courses to offline courses is dynamically adjusted. When the predicted participation value is lower than the preset threshold, the proportion of offline courses is increased; when the predicted participation value is higher than the preset threshold, the proportion of online courses is increased. By balancing the proportion of online and offline courses, the hybrid teaching model is optimized. Establish a tracking and feedback mechanism for students’ electronic device usage habits and digital course participation, regularly collect students’ device usage data and course participation data, and use it to update and optimize the correlation model between electronic device usage habits and digital course participation, and dynamically adjust the course ratio; Based on the dynamic ratio of online and offline courses, combined with students' learning preferences and teachers' teaching time, an intelligent scheduling algorithm is used to automatically generate course schedules. Classes are scheduled regularly according to the latest updated course ratio, forming a dynamically updated intelligent scheduling mechanism to adapt to the changing trends in students' electronic device usage habits and digital course participation.

2. The method according to claim 1, characterized in that The raw data of students' daily device usage is obtained, and the usage data includes usage duration, usage frequency, and usage time period. Through data preprocessing, outliers and noise data are removed to obtain a standardized device usage data set, including: Obtaining original record information sent by the device data acquisition interface, wherein the original record information includes the device startup time, end time, application identifier, operation interval duration, and number of startups per hour; An outlier detection is performed using the interquartile range of a box plot according to the values ​​of each dimension in the original record information. If the value of any dimension in a single record is less than the lower limit of the interquartile range or greater than the upper limit of the interquartile range, it is determined to be an abnormal record; For the missing values ​​of duration in the abnormal records, linear interpolation is used to obtain the filling value, and for the missing values ​​of frequency, the nearest neighbor value is used to obtain the filling value; The filled records are segmented by a sliding time window, the basic size of the sliding time window is a daily average unit, the number of device startups within the sliding time window is calculated to obtain the usage frequency data, and the usage time within the sliding time window is accumulated to obtain the daily average usage time data; The frequency of use data and the time of use data are grouped using an Affinity Propagation clustering algorithm, wherein the Affinity Propagation clustering algorithm clusters the data based on a time similarity threshold; The time period activity score is calculated based on the number of startups and the usage duration in each time period in the clustering result, and a normalized weight value is assigned to the time period whose time period activity score is greater than the activity threshold according to the score range.

3. The method according to claim 1, characterized in that The feature engineering is performed on the standardized device usage data set to extract key features that reflect the students' electronic device usage habits, including average daily usage time, high-frequency usage time period, and device type distribution. The feature vector of students' electronic device usage habits is constructed based on the key features, including: A sliding time window is used to collect statistics on device usage data. The window size of the sliding time window is a preset time. The number of times the device is turned on and off and the interactive behavior data are recorded through a counter. According to the device usage data, a Gaussian kernel density estimation tool is used to calculate the probability density distribution of the usage time, and the time interval in which the usage frequency exceeds a set threshold is extracted from the probability density distribution to obtain high-frequency usage period data; The feature importance of the high-frequency usage period data is sorted, redundant features are eliminated by calculating the Pearson correlation coefficient, and the principal component tool is used to perform dimensionality reduction processing on the feature data to obtain reduced-dimensional feature data; Calculate statistical parameters based on the dimension reduction feature data, group the time period data using a partitioning and clustering method to obtain time period features, and extract the device type with the highest proportion from the device type data; A feature vector is constructed for the time period characteristics and the device type in combination with the device interaction behavior data. The feature vector includes duration statistical characteristics, time period distribution characteristics, device type characteristics, interaction behavior characteristics, and scene characteristics.

4. The method according to claim 1, characterized in that The method of obtaining the usage data of the learning electronic device, including the login frequency, learning time, and interaction, cleaning and converting the data, and constructing a label vector of the digital course participation includes: Obtaining login data sent by the terminal device, the login data including login time and logout time and online time, and obtaining the number of test answers, number of discussion replies, resource browsing time and teacher-student interaction records from the teaching platform interface; According to the login data and teaching platform data, anomaly detection is performed using box plot quartile interval limit parameters, and abnormal data is replaced by forward data filling to obtain a standardized data matrix; For the standardized data matrix, a radial basis function regression tool is used to construct a participation quantification index, the kernel function bandwidth of the radial basis function is set to a first preset threshold, and the regularization parameter is set to a second preset threshold, to obtain a basic participation evaluation value; The basic evaluation values ​​of participation are clustered from bottom to top, and the distances between the participation features are calculated using the Euclidean distance to obtain five levels of participation grouping data; For the participation grouping data, a linear weight value in a preset interval is assigned, and an eight-dimensional label vector is constructed by combining the online time proportion, interaction frequency proportion, resource access degree and homework achievement rate.

5. The method according to claim 1, characterized in that: The method of associating the student electronic device usage habit feature vector with the digital course participation label vector, using a machine learning algorithm to train a prediction model of electronic device usage habits and participation, and obtaining association rules and influence weights between the electronic device usage habit feature vector and the digital course participation label vector includes: Obtain usage habit feature vectors and engagement label vectors, and obtain normalized feature data through minimum and maximum value normalization processing; For the normalized feature data, the Pearson correlation coefficient is used to calculate the significant correlation value. If the correlation coefficient is less than a preset threshold, the feature dimension is eliminated, and the collinearity feature is eliminated through the variance inflation factor test to obtain a preprocessed data set; Stratifying according to the engagement label value interval in the preprocessed data set, using a stratified sampler to divide the training data set and the test data set, and training the training data set by a random forest regressor to obtain a prediction model; For the prediction model, the Gini coefficient is used to calculate the feature purity change value, the feature importance is verified in combination with the information gain ratio, and the feature combination importance is calculated through cross-validation to obtain the feature influence ranking result; Based on the feature influence ranking results, association rules with weight values ​​higher than a preset threshold are extracted, and a decision tree based on information gain ratio is used to perform path analysis on the association rules. Post-pruning is performed through an error cost function to obtain an optimal rule subset.

6. The method according to claim 1, characterized in that The relationship between screen time and digital course participation is analyzed in a targeted manner, and the distribution of participation under different screen time is counted. If there is a significant difference, it is determined that there is a correlation between the electronic device usage habit feature vector and the digital course participation label vector, and a reasonable range of daily screen time is determined, including: The screen usage data is divided by using a fixed time window, and a duration data set is obtained by segmented statistics, wherein the duration data set includes a cumulative usage duration value, a usage count value, and a unit operation frequency value; Calculate the participation mean, standard deviation, kurtosis and skewness values ​​according to the duration data set, perform a variance test on the participation data using a one-way ANOVA tool, determine the significance of the intervals based on the F statistic, and obtain a set of significant intervals; For the significant interval set, a reflection kernel density estimation tool is used to fit the participation distribution, the kernel function bandwidth of the reflection kernel density estimation tool is determined by cross-validation, and interval statistics are performed on the participation distribution to obtain a data stability index; Performing boundary detection on the data stability index through an adaptive sliding window, wherein the adaptive sliding window calculates a participation change rate value and a stability index value during the detection process; A curvature-based inflection point detection tool is used to identify the boundary of the participation change rate value, and the boundary position is corrected according to the stability index value to obtain a reasonable time interval set.

7. The method according to claim 1, characterized in that The method dynamically adjusts the ratio of online courses to offline courses based on the correlation model between electronic device usage habits and digital course participation, as well as the reasonable interval of screen usage time. When the participation prediction value is lower than a preset threshold, the ratio of offline courses is increased. When the participation prediction value is higher than the preset threshold, the ratio of online courses is increased. By balancing the ratio of online and offline courses, the hybrid teaching mode is optimized, including: Using a moving time window to collect electronic device usage data, and processing the electronic device usage data through a stochastic gradient regressor to obtain a participation prediction value; Generate a participation prediction sequence according to the participation prediction value, and obtain a reference value by using an exponentially weighted moving average calculation for the participation prediction sequence; Calculating upper and lower thresholds for the reference value, if the participation prediction value shows an upward trend for three consecutive times, the upper threshold is increased by a preset ratio, and the lower threshold is increased by a preset ratio, and an adjustment signal is obtained; Calculating the course proportion adjustment amount by using a proportional-integral regulator according to the adjustment signal; if the participation prediction value is lower than the lower threshold, the online course proportion is adjusted downward, and the downward adjustment amount is proportional to the degree of deviation from the lower threshold; The course ratio adjustment amount is processed by a cubic spline smoother, and the exponential moving average is used to smooth the course ratio adjustment amount to obtain the actual course ratio.

8. The method according to claim 1, characterized in that The tracking and feedback mechanism for students’ electronic device usage habits and digital course participation is established to regularly collect students’ device usage data and course participation data for updating and optimizing the electronic device usage habits and digital course participation association model, and dynamically adjusting the course ratio, including: Obtaining electronic device usage data and course participation data, wherein the electronic device usage data and course participation data include device usage time, number of interactions, resource access, and homework submission rate, and obtaining a real-time data set through data consistency verification; Calculate the participation change rate using an exponentially weighted moving average based on the real-time data set, determine the moment of significant participation change using an adaptive threshold calculator, and trigger an associated model update signal if the participation prediction error exceeds a preset threshold; In response to the association model update signal, a kernel function online learning regressor is used to update the model parameters, the learning rate is adjusted by exponential decay, and the model stability index is calculated based on real-time data; Performing model performance evaluation according to the model stability index; if the model stability index is lower than a preset threshold value, triggering a model rollback mechanism, and obtaining an updated model performance index through cross-validation; According to the performance index of the model, an adaptive proportional-integral regulator is used to calculate the course ratio adjustment signal, and the course ratio is dynamically adjusted through Bezier curve spline smoothing to obtain the final course adjustment plan.

9. The method according to claim 1, characterized in that: Based on the dynamic ratio of online and offline courses, combined with students' learning preferences and teachers' teaching time, the intelligent scheduling algorithm is used to automatically generate course schedules. According to the latest updated course ratio, classes are scheduled regularly to form a dynamically updated intelligent scheduling mechanism to adapt to the changing trends of students' electronic device usage habits and digital course participation, including: Obtaining course constraints, prioritizing the constraints according to their importance, setting hard constraint weights and soft constraint weights for the constraints, and constructing a constraint evaluation system through the hard constraint weights and the soft constraint weights to obtain basic course scheduling data; An improved genetic algorithm is used to calculate based on the basic class scheduling data, wherein a real number encoding method and an initial population size are set for the improved genetic algorithm, a population fitness value is obtained by calculating the fitness function, and a roulette wheel selection operator is used to select the population fitness value to obtain an initial class scheduling plan; Adopting an adaptive local search tool for the initial course scheduling plan, wherein a search radius value is set based on the total number of courses, and the time conflict degree and resource utilization rate in the initial course scheduling plan are detected. If the detection result exceeds a preset threshold, a local adjustment gain value is calculated; Obtaining course participation data through a real-time monitor, and if the change range of the course participation data exceeds a preset threshold value, triggering a course scheduling update instruction, and setting a basic update cycle value according to the course scheduling update instruction; A time window incremental update method is adopted for the basic update cycle value, wherein the participation trend value is calculated through a sliding window, and the update cycle adjustment interval is determined according to the participation trend value, and the initial class scheduling plan is dynamically adjusted to obtain the final class scheduling plan.

Images