Learning path intelligent recommendation system based on user behavior big data analysis

Abstract

This invention relates to the field of cloud platform technology, specifically to an intelligent learning path recommendation system based on big data analysis of user behavior. The system includes a behavior data construction module, used to collect data on mouse trajectory coordinates, pause / click frequency, and response time for answering questions generated by users on the online learning platform, constructing a user interaction timestamp sequence. In this invention, the calculated parameters are compared in real-time with preset attention loss rates and fatigue divergence thresholds. When attention decay is detected, a dynamic slicing mechanism is triggered, dividing the originally long teaching video into fragmented knowledge point nodes. This path reorganization strategy based on real-time psychological states ensures that the difficulty and duration of the teaching content dynamically adapt to the user's current cognitive capacity, effectively alleviating learning fatigue and maintaining a high level of focus during key knowledge acquisition periods, thereby improving the completion rate and knowledge absorption efficiency of online courses.

Description

Technical Field

[0001] This invention relates to the field of cloud platform technology, and in particular to an intelligent recommendation system for learning paths based on big data analysis of user behavior. Background Technology

[0002] Cloud platform technology mainly refers to Internet-based computing methods. Through core technologies such as virtualization, distributed computing, and parallel processing, it integrates a large number of physical hardware resources (such as servers, storage devices, and network devices) into a unified logical resource pool, enabling on-demand allocation and dynamic scheduling of computing power, storage space, and application services.

[0003] Existing cloud platform technologies focus on the virtualization and integration of underlying hardware resources and the on-demand allocation of computing and storage capabilities. In practice, they typically treat the interaction of online educational content as a standard data stream transmission task, prioritizing high concurrency and low latency of video streams while neglecting the cognitive interaction state of the receiving user. This operational model relies on a pre-arranged static syllabus, where the playback duration and order of teaching resources are fixed before the session begins. This makes it difficult for the resource distribution mechanism to perceive and respond to learners' real-time physiological and psychological fluctuations. When users experience cognitive overload or a significant decline in attention during prolonged learning, the underlying platform continues to push content according to a predetermined rigid plan. This mismatch between static resource delivery and dynamic cognitive abilities can easily lead to interruptions in information reception due to physiological fatigue, resulting in an ineffective learning state where the video continues to play while the user has ceased effective intake of information. Therefore, improvements are needed. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing an intelligent recommendation system for learning paths based on big data analysis of user behavior.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A learning path intelligent recommendation system based on user behavior big data analysis includes: The behavioral data construction module is used to collect data such as mouse trajectory coordinates, pause click frequency and answer response time generated by users on the online learning terminal, construct a user interaction timestamp sequence, set a fixed-length time sliding window to capture the user interaction timestamp sequence, and generate a set of windowed interaction intervals. The feature index calculation module is used to statistically analyze the dispersion of values ​​in the set of windowed interaction intervals, quantify the interval variance fluctuation value, statistically analyze the amount of interaction data per unit time, perform a linear fitting operation on the downward trend, calculate the slope of the fitted line, and associate and combine the slope with the interval variance fluctuation value to generate composite learning state feature parameters. The state determination and decision module is used to compare the degree of tilt in the composite learning state feature parameters with a preset attention loss rate threshold, and at the same time detect whether the interval variance fluctuation value exceeds the fatigue divergence threshold, generate an abnormal fatigue state signal, map the abnormal fatigue state signal to a preset rule scheduling table for strategy matching, and generate a content adaptation trigger instruction. The path reorganization and scheduling module is used to adapt trigger instructions according to the content, initiate slicing for the attention decay state, divide the originally planned long-term teaching video into fragmented knowledge point nodes, reorganize the original learning path according to the fragmented knowledge point nodes, and generate a dynamic learning recommendation queue.

[0006] Preferably, the step of obtaining the set of windowed interaction intervals is as follows: Collect mouse trajectory coordinates, pause click frequency, and response time of users on the online learning platform. Use the same timestamp format as the same benchmark, sort them in ascending order of time occurrence, remove duplicate timestamps while maintaining the first order, and associate the sorted timestamps with behavior identifiers to form a user interaction timestamp sequence. Based on the user interaction timestamp sequence, a fixed-length time sliding window boundary is set, the smallest timestamp is selected as the starting point, and the window is shifted backward by a fixed length and the left-closed and right-open rule is adopted to extract the timestamp subsequence within each time sliding window, forming a windowed user interaction timestamp sequence fragment set. Based on the set of windowed user interaction timestamp sequence segments, the time difference between adjacent timestamps is calculated for each segment. The minimum and maximum time differences within the segments are scaled proportionally to a uniform range. The scaled time differences are then merged in the order of the segments to form a set of windowed interaction intervals.

[0007] Preferably, the step of obtaining the interval variance fluctuation value is as follows: Based on the set of windowed interaction intervals, the data type and value range of each interaction interval value are verified, interval values ​​less than zero and missing values ​​are removed, and all valid interaction interval values ​​are retained to form a valid windowed interaction interval sequence. Based on the effective windowed interaction interval sequence, the number of interval values ​​in the sequence is counted, the average of all interval values ​​is calculated as the mean value, the square of the deviation between each interval value and the mean value is calculated and averaged to obtain the variance value, the square root of the variance value is taken to obtain the standard deviation value, and the average interaction interval value of the user's historical learning cycle is retrieved as the historical mean value, thus forming the mean value, variance value, standard deviation value and historical mean value. The interval variance fluctuation value is calculated based on the mean, variance, standard deviation and historical mean values.

[0008] Preferably, the steps for obtaining the composite learning state feature parameters are as follows: Based on the user interaction timestamp sequence, the data is divided into segments of equal width according to the unit time length. Each segment is counted to form a unit time interaction data sequence. The difference between adjacent counts is calculated, and a downward trend is determined by the proportion of consecutive negative differences exceeding a preset threshold value. A set of downward trend segments is extracted, and a straight line fitting operation is performed on the set of downward trend segments. The slope of the fitted straight line for each segment is calculated, and the slope of the largest negative fitted straight line is selected as the slope. The slope and the interval variance fluctuation value are concatenated into a binary vector in a fixed order to form a composite learning state feature parameter.

[0009] Preferably, the step of obtaining the abnormal fatigue state signal is as follows: Based on the composite learning state characteristic parameters, the tilt degree and interval variance fluctuation value are read item by item, and compared with the preset attention loss speed threshold and fatigue divergence threshold respectively. The name of the out-of-bounds item, the out-of-bounds direction, the out-of-bounds amplitude, and the trigger time and source identifier are added. The comparison results are merged into a unified abnormal structure to generate an abnormal fatigue state signal.

[0010] Preferably, the step of obtaining the content adaptation trigger instruction is as follows: Based on the abnormal fatigue state signal, the name of the out-of-bounds item, the direction of the out-of-bounds, the magnitude of the out-of-bounds, the trigger time and the source identifier are parsed. The matching row is retrieved in the preset rule scheduling table according to the field index. The path adjustment scheme identifier and the path adjustment scheme content are read to generate the path adjustment scheme. Based on the path adjustment scheme, organize the content target, playback order, insertion position, execution timing and effective conditions, encapsulate them into an action command header and parameter body, add the path adjustment scheme identifier and check code, set the trigger range and validity period, and generate content adaptation trigger commands.

[0011] Preferably, the steps for obtaining the fragmented knowledge point nodes are as follows: Based on the content adaptation trigger instruction, retrieve the teaching resource index table, extract the total duration value and basic slice duration value of the original long-term teaching video, read the interval variance fluctuation value in the composite learning state feature parameters, and generate a teaching video slice parameter set. Based on the set of teaching video slice parameters, calculate the dynamic duration value of each slice one by one; Based on the dynamic duration value, video segments are extracted one by one starting from the position where the original long-term teaching video has not been played. If the remaining video duration is longer than the current dynamic duration value, the segments are extracted according to the current duration. If it is shorter, the remaining video portion is divided into the last segment. The start and end times and knowledge points are extracted for each segment to form a set of segments with dynamically adjusted durations. When an attention decay status indicator is detected, the segments are loaded and played in the order of the dynamically adjusted duration set to obtain fragmented knowledge point nodes.

[0012] Preferably, the steps for obtaining the dynamic learning recommendation queue are as follows: Based on the fragmented knowledge point nodes or simple review resources, the unique identifier of the original teaching resource is read, and the corresponding original resource entry is located in the original learning path using the unique identifier. The original resource entry is removed from the learning path, and the fragmented knowledge point nodes or simple review resources are inserted into their original positions in order of playback priority. After the insertion is completed, the entire learning path is traversed, and a continuous playback order index and outline order index are regenerated from beginning to end to form a dynamic learning recommendation queue.

[0013] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, by collecting data on mouse trajectory coordinates, pause click frequency, and response time of users on the online learning platform, a continuous sequence of user interaction timestamps is constructed. A fixed-length time sliding window is used to truncate this sequence, generating a set of windowed interaction intervals. This captures the continuity of user behavior at a microscopic level. The variance fluctuation of the intervals is quantified by statistically analyzing the dispersion of values ​​within the window, and the degree of skewness is calculated by linear fitting of the amount of interaction data per unit time. The composite learning state feature parameters generated by combining these two factors more accurately map the user's cognitive load and attention stability compared to a single indicator. The calculated parameters are compared in real-time with preset attention loss rates and fatigue divergence thresholds. When attention decay is detected, a dynamic slicing mechanism is triggered, dividing the originally long teaching video into fragmented knowledge point nodes. This path reorganization strategy based on real-time psychological state ensures that the difficulty and duration of the teaching content dynamically adapt to the user's current cognitive capacity, effectively alleviating learning fatigue and maintaining a high level of focus during key knowledge acquisition periods, thereby improving the completion rate and knowledge absorption efficiency of online courses. Attached Figure Description

[0014] Figure 1 This is a system flowchart of the present invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0016] Please see Figure 1 This invention provides a technical solution: an intelligent recommendation system for learning paths based on big data analysis of user behavior, comprising: The behavioral data construction module is used to collect data such as mouse trajectory coordinates, pause click frequency and answer response time generated by users on the online learning terminal, construct a user interaction timestamp sequence, set a fixed-length time sliding window to capture the user interaction timestamp sequence, and generate a set of windowed interaction intervals. The feature index calculation module is used to statistically analyze the dispersion of values ​​in the windowed interaction interval set, quantify the interval variance fluctuation value, statistically analyze the amount of interaction data per unit time, perform linear fitting operation on the downward trend, calculate the slope of the fitted line, and associate and combine the slope with the interval variance fluctuation value to generate composite learning state feature parameters. The state determination and decision module is used to compare the degree of tilt in the composite learning state feature parameters with the preset attention loss rate threshold, and at the same time detect whether the interval variance fluctuation value exceeds the fatigue divergence threshold, generate an abnormal fatigue state signal, map the abnormal fatigue state signal to the preset rule scheduling table for strategy matching, and generate content adaptation trigger instructions. The path reorganization and scheduling module is used to initiate slicing based on content adaptation trigger instructions, start the attention decay state, divide the originally planned long-term teaching video into fragmented knowledge point nodes, reorganize the original learning path based on the fragmented knowledge point nodes, and generate a dynamic learning recommendation queue.

[0017] The steps to obtain the set of windowed interaction intervals are as follows: Collect mouse trajectory coordinates, pause click frequency, and response time of users on the online learning platform. Use the same timestamp format as the same benchmark, sort them in ascending order of time occurrence, remove duplicate timestamps while maintaining the first order, and associate the sorted timestamps with behavior identifiers to form a user interaction timestamp sequence. Based on the user interaction timestamp sequence, a fixed-length time sliding window boundary is set. The smallest timestamp is selected as the starting point, and the window is shifted backward by a fixed length and the left-closed and right-open rule is adopted to extract the timestamp subsequence within each time sliding window, forming a windowed user interaction timestamp sequence fragment set. Based on the set of windowed user interaction timestamp sequence segments, the time difference between adjacent timestamps is calculated for each segment. The minimum and maximum time differences within the segments are then scaled proportionally to a uniform range. The scaled time differences are then merged in the order of the segments to form a set of windowed interaction intervals.

[0018] Specifically, the system collects mouse trajectory coordinates, pause / click frequency, and response time for online learning. It deploys a full suite of event listener hooks on the front-end page, capturing screen pixel coordinates for the `mousemove` event at a sampling rate of 50 times per second, recording click times for the `pause` event in the video player, and calculating the time difference between the `load` and `submit` events in the question-answering interface. All raw time data is then converted to long integer Unix timestamp format, accurate to milliseconds (e.g., converting "2024-03-01 10:00:00.005" to "1709258400005"). All timestamp types are merged into a single array, which is then sorted in ascending order using either quicksort or mergesort. The sorted timestamp array is iterated through, and a deduplication threshold is set. The threshold, set in milliseconds, is based on the shortest physical reset time of the mouse microswitch when clicked by a human hand and the limit of neural reflex. The difference between two adjacent timestamps is calculated. If the difference is less than the threshold, it is judged as redundant data caused by mechanical jitter or accidental touch and is removed. Only the first valid trigger point within the time window is retained. A behavior type mapping dictionary is established, with the behavior identifier corresponding to mouse movement set as 1, the behavior identifier corresponding to pause clicking as 2, and the behavior identifier corresponding to answering questions as 3. According to the source attributes of the original data, each sorted and deduplicated timestamp is bound to the corresponding behavior identifier, and a binary structure containing time attributes and event attributes is constructed to form a user interaction timestamp sequence.

[0019] Based on the user interaction timestamp sequence, set a fixed-length sliding window boundary and set the window length parameter. The value is set based on the average period (approximately 5 minutes) of sustained high concentration in adults, as described in cognitive psychology. The sliding step size parameter is also set. Seconds are used to ensure an 80% data overlap between adjacent windows to capture continuous trends; the initial time point for initializing the window is [number of seconds]. Given the timestamp with the smallest value in the sequence, calculate the end time of the current window. for and The sum is obtained by iterating through the entire sequence of user interaction timestamps and comparing each timestamp in the sequence. Filter out all that meet the conditions Timestamp data, in which For the first in the sequence Each timestamp is selected and stored in a temporary subsequence container while maintaining its original chronological order. After extracting data from the current window, the starting time point is... Add a sliding step Once a new start time is obtained, the end time is recalculated, and the above filtering and extraction process is repeated until the start time exceeds the largest timestamp value in the sequence. All generated temporary subsequence containers are then stored in a list in the order of generation, forming a set of windowed user interaction timestamp sequence fragments.

[0020] Based on the set of windowed user interaction timestamp sequence segments, the time difference between adjacent timestamps is calculated for each segment. For each sequence segment, starting from the first timestamp, the previous timestamp is subtracted sequentially to obtain the interaction interval value, generating the original interval sequence for that segment. The maximum interval value is then found by traversing this original interval sequence. and minimum interval value The range transformation method is used to map each original interval value in the sequence to a closed interval between 0 and 1. The calculation formula is as follows: ,in, The normalized version A numerical value for the interaction interval. The first in the original interval sequence A number, This is the minimum interval value within the current segment. This represents the maximum interval value within the current segment. To prevent the denominator from being zero, a very small positive number (with a value of...) is set. This calculation eliminates the dimensional differences caused by different baseline operation frequencies in different learning periods. The normalized interval values ​​calculated for each segment are spliced ​​together end to end according to the original order of the segments in the set. Data in the overlapping areas of adjacent segments are not deduplicated or merged but are retained to reflect the characteristic changes of the sliding process, forming a windowed interactive interval set.

[0021] The steps to obtain the interval variance fluctuation value are as follows: Based on the set of windowed interaction intervals, the data type and value range of each interaction interval value are verified, interval values ​​less than zero and missing values ​​are removed, and all valid interaction interval values ​​are retained to form a valid windowed interaction interval sequence. Based on the effective windowed interaction interval sequence, the number of interval values ​​in the sequence is counted, the average of all interval values ​​is calculated as the mean value, the square of the deviation of each interval value from the mean value is calculated and averaged to obtain the variance value, the square root of the variance value is taken to obtain the standard deviation value, and the average interaction interval value of the user's historical learning cycle is retrieved as the historical mean value, forming the mean value, variance value, standard deviation value and historical mean value. The interval variance fluctuation is calculated based on the mean, variance, standard deviation, and historical mean values. The formula is as follows: ; Where F is the interval variance fluctuation value, A is the variance value, and M is the mean value. is the historical average value, and U is the number of interval values ​​in the effective windowed interactive interval sequence.

[0022] Specifically, based on the set of windowed interaction intervals, the original data segments corresponding to each time sliding window in the set are traversed. A data validity verification program is started, and each interaction interval value within the segment is scanned item by item. First, the data type of the value is determined to ensure that it is a double-precision floating-point format. If a non-numeric character or null pointer type is encountered, it is directly marked as invalid data. Then, a preset reasonableness judgment interval is retrieved. The interval is set based on the limits of human physiological reaction function and the normal operating rhythm in online learning scenarios, and a minimum valid interval threshold is set. The value is set based on the shortest conduction time of human nerve reflexes, approximately 50 milliseconds. Intervals shorter than this value are usually caused by device contact jitter or electrical signal noise. A maximum effective interval threshold is set. The value is based on an online learning attention maintenance model, which considers an inactivity interval of more than 10 minutes to indicate that the user has left the current learning environment or is in a prolonged idle state. The value for each interaction interval is... With the closed interval Compare and determine those that meet the requirements. or The values ​​are identified as noisy data and removed from the sequence. At the same time, missing values ​​(NaN) in the sequence are removed by omitting them in the same position without interpolation to avoid introducing human bias. After cleaning, the index order of the remaining values ​​is rearranged to ensure that the data is arranged in a continuous and compact manner in the time dimension, forming an effective windowed interactive interval sequence.

[0023] Based on the effective windowed interaction interval sequence, the total number of elements stored in the sequence is read and defined as the sample size. Each floating-point value in the sequence is iterated and summed. The sum is divided by the sample size to obtain the mean value reflecting the average operation frequency within the current window. Then, the sequence is iterated again, and the mean value calculated above is subtracted from each interval value to obtain the deviation value. Each deviation value is squared to eliminate the canceling effect of positive and negative values. All squared deviation values ​​are summed and divided by the sample size to obtain the variance value representing the degree of data dispersion. The square root function is then called to calculate the variance value to obtain the standard deviation value consistent with the original data units. Simultaneously... Establish a connection with the backend user profile database, retrieve the interaction logs of all learning units completed by the user in the past 30 natural days using the user's unique identifier, extract the average value of all valid interaction intervals in the historical logs, and if the historical data exceeds 1000 records, use a reservoir sampling algorithm to extract representative samples for calculation to ensure the reliability of the baseline. Define the average level of this long period as the historical mean value. Package and encapsulate the mean value, variance value, and standard deviation value calculated in real time in the current window with the historical mean value obtained by retrieval to provide basic statistical parameters for subsequent state fluctuation analysis, forming the mean value, variance value, standard deviation value and historical mean value.

[0024] In the formula for calculating the variance fluctuation value, a dynamic indicator that is highly sensitive to the stability of interaction is constructed by introducing the coefficient of variation and the logarithmic weighting of the sample size. The ratio of variance to mean is used to eliminate the influence of absolute time length. At the same time, historical benchmark values ​​are combined to correct the relative fatigue level of the current state. The formula structure design aims to amplify the contribution weight of high-frequency unstable operations to fatigue judgment, thereby proving that the formula can more accurately capture the subtle trend of users shifting from focus to distraction. The variance parameter represents the fluctuation range of the user's interaction interval within the current time window, reflecting the stability of the operation rhythm. It is obtained by first reading the effective windowed interaction interval sequence. Calculate the sequence mean Then, the square of the difference is calculated for each item. For example, in a certain test, the window contains 50 interactive interval data points, and the calculated mean is 1.5 seconds. The variance is obtained by averaging the squared fluctuations of each data point relative to 1.5 seconds. Here, it is set to be obtained through calculation. The value is 0.81; This is the mean value, representing the average interaction speed of the user within the current time window, reflecting the current learning activity level. It is obtained by summing all values ​​in the effective windowed interaction interval sequence and dividing by the total number of elements in the sequence. A smaller value indicates more frequent interaction; here, it is assumed to be obtained by averaging over 50 data points. The value is 1.5 (unit: seconds); This is a historical average value. This parameter represents the user's habitual interaction rhythm and serves as a personalized baseline. The steps to obtain it are: query the database for the interaction interval records of all learning sessions for the user over the past 30 days, remove long intervals during dormant periods, and calculate the average. Here, it is assumed to be calculated based on retrieved historical records. A value of 2.0 (in seconds) indicates that the user's usual operating pace is slightly slower than that of the current window; The number of interval values ​​in the effective windowed interactive interval sequence. This parameter represents the confidence capacity of the statistical sample and is used to weight the significance of fluctuation values ​​in the formula. The steps to obtain it are: directly count the number of sequence elements after cleaning, and here we set the number of valid data in the current window. It is 50; Calculations based on parameters: Substitute the above parameter values ​​into the formula to perform multi-level calculations: The first step is to calculate the ratio of the product of variance and mean: ; The second step is to calculate the weighted average of the coefficients of variation within the logarithmic terms: First, calculate the standard deviation: ; Calculate the coefficient of variation term: ; Calculate the logarithm of the sample size: ; Calculate the sum within the logarithm: ; The third step is to calculate the natural logarithm: ; Step 4: Calculate the final interval variance fluctuation value: ; The results indicate that current user interaction behavior exhibits a certain degree of dispersion. As a quantitative dimensionless indicator, the value of 0.1898 gradually increases over time, which means that the user's operation rhythm becomes increasingly unstable within a unit of time, including both rapid continuous clicks and unexpected pauses. This non-linear amplification of variance relative to the mean is usually strongly correlated with the distraction caused by cognitive fatigue.

[0025] The steps for obtaining the composite learning state feature parameters are as follows: Based on the user interaction timestamp sequence, the data is divided into segments with equal widths according to the unit time length. Each segment is counted to form a unit time interaction data sequence. The difference between adjacent counts is calculated, and a downward trend is determined by the proportion of consecutive negative differences exceeding a preset threshold value. A set of downward trend segments is extracted, and a straight line fitting operation is performed on the set of downward trend segments. The slope of the fitted straight line for each segment is calculated, and the slope of the largest negative fitted straight line is selected as the slope. The slope and the interval variance fluctuation value are concatenated into a binary vector in a fixed order to form a composite learning state feature parameter.

[0026] Specifically, based on the user interaction timestamp sequence, the unit time length parameter is set to 60 seconds. This value is based on the average period of short-term memory retention in cognitive psychology. The entire sequence is divided into multiple consecutive time buckets according to this length. The interaction data in each bucket is traversed and the number is counted to generate an interaction frequency sequence. The difference between two adjacent terms in the sequence is calculated, and the proportion threshold for judging a downward trend is defined as 0.75. This threshold is obtained by analyzing 500 historical normal learning curves, calculating the average proportion of curves with negative first derivatives, taking the mean and adding one standard deviation. The observation window length is set to 5 units of time. If the proportion of negative differences within the window is... If the data exceeds the preset threshold, the data within the window is marked as a candidate descent segment. All candidate segments are extracted, and the least squares method is used to perform linear regression fitting on the interaction frequency data within each segment. The objective function is to minimize the sum of squared residuals. The slope parameter of the fitted line is obtained by solving for the slope parameter, which represents the rate of attention loss. All calculated slopes are iterated, and the slope with the largest absolute value and a value less than zero is selected as the final tilt degree. The interval variance fluctuation value calculated in the previous step is read, and the tilt degree is used as the first dimension feature, and the interval variance fluctuation value is used as the second dimension feature. They are combined in order to form the composite learning state feature parameter.

[0027] The steps for obtaining abnormal fatigue state signals are as follows: Based on the composite learning state characteristic parameters, the tilt degree and interval variance fluctuation value are read item by item and compared with the preset attention loss speed threshold and fatigue divergence threshold respectively. The name of the out-of-bounds item, the out-of-bounds direction, the out-of-bounds amplitude, and the trigger time and source identifier are added. The comparison results are merged into a unified abnormal structure to generate an abnormal fatigue state signal.

[0028] Specifically, based on the composite learning state feature parameters, the feature vector is unpacked to obtain the tilt degree value and the interval variance fluctuation value. A preset threshold configuration file is called, where the attention loss rate threshold is set to -0.3, which is calculated based on the average interaction decline rate of a large number of users before giving up learning, and the fatigue divergence threshold is set to 0.25, which is determined by statistically analyzing the 95th percentile of the interval variance fluctuation value of users in a conscious state. The tilt degree value is compared with the attention loss rate threshold. If it is less than the threshold, it is judged as rapid attention decay. The interval variance fluctuation value is compared with the fatigue divergence threshold. If it is greater than the threshold, it is judged as cognitive fatigue dispersion. For items that meet the judgment conditions, their parameter names, the direction of exceeding the threshold, and the specific difference magnitude are recorded. At the same time, the current system timestamp is obtained as the trigger time, and the data source is marked as the real-time monitoring module. These discrete abnormal description information are aggregated in a standardized JSON format to form an abnormal fatigue state signal.

[0029] The steps to obtain the content adaptation trigger command are as follows: Based on the abnormal fatigue state signal, the name of the out-of-bounds item, the direction of the out-of-bounds, the magnitude of the out-of-bounds, the trigger time and the source identifier are parsed. The matching row is retrieved in the preset rule scheduling table according to the field index, the path adjustment scheme identifier and the path adjustment scheme content are read, and the path adjustment scheme is generated. Based on the path adjustment scheme, organize the content target, playback order, insertion position, execution timing and effective conditions, encapsulate them into an action command header and parameter body, add the path adjustment scheme identifier and check code, set the trigger range and validity period, and generate content adaptation trigger commands.

[0030] Specifically, based on the abnormal fatigue state signal, a JSON parser is used to extract core fields such as the name of the out-of-bounds item, the direction of the out-of-bounds out-of-bounds out-of-bounds out-of-bounds, and the magnitude of the out-of-bounds ...

[0031] Based on the path adjustment scheme, the scheme content is parsed to clarify the target ID of the content to be adjusted, the new playback order index, and the specific insertion time point. The execution timing is set to be executed immediately after the current video buffering ends, and the effective condition is set to the user's current network status being good. These control parameters are packaged, and an instruction header containing the instruction type code and length information is constructed. The parameter body contains the detailed parameters parsed above. To prevent transmission errors, the CRC32 checksum of the instruction packet is calculated and appended to the end. At the same time, the effective time window of the instruction is set to 300 seconds, that is, if the instruction is not responded to by the client within 300 seconds after it is issued, it will automatically become invalid. The triggering scope is set to the currently online learning terminals. Finally, all byte streams are combined to generate the content adaptation trigger instruction.

[0032] The steps for acquiring fragmented knowledge point nodes are as follows: Based on the content adaptation trigger command, retrieve the teaching resource index table, extract the total duration value and basic slice duration value of the original long-term teaching video, read the interval variance fluctuation value in the composite learning state feature parameters, and generate a teaching video slice parameter set. Based on the parameter set of the teaching video segments, the dynamic duration value of each segment is calculated one by one. The calculation formula is as follows: ; in, For the first The dynamic duration value of each slice. The base slice duration value, The state sensitivity coefficient, Here, represents the normalized state index value, and represents the normalized result of the interval variance fluctuation value. For personalized recovery coefficient values, The slice number; Based on the dynamic duration value, video segments are extracted one by one starting from the position where the original long-term teaching video has not been played. If the remaining video duration is longer than the current dynamic duration value, the segments are extracted according to the current duration. If it is shorter, the remaining video portion is divided into the last segment. The start and end times and knowledge points are extracted for each segment to form a set of segments with dynamically adjusted duration. When an attention decay status indicator is detected, the fragmented knowledge point nodes are loaded and played in the order of the dynamically adjusted duration set.

[0033] Specifically, based on the content adaptation trigger command, the binary data stream in the command package is parsed to obtain core control parameters. The unique identifier (ID) of the teaching resource is extracted, and this identifier is used to establish a database connection with the backend teaching resource index table. This index table is a pre-built NoSQL database that stores metadata for all course videos. The corresponding target video entry is located in the index table, its metadata fields are read, and the total duration of the originally planned long-duration teaching video is obtained. For example, the total duration is set to 2700 seconds (i.e., 45 minutes). Simultaneously, the preset basic slice duration is read. This value is based on the average time window for efficient information reception in adults according to cognitive load theory, typically configured as 300 seconds. Then, the composite learning state feature parameters calculated in the preceding steps are separated from the command parameter body, and the interval variance fluctuation value (denoted as...) is deconstructed and read. For example, the previously calculated value of 0.1898 is used to encapsulate the total duration value, the basic slice duration value, and the real-time interval variance fluctuation value to construct a temporary teaching video slice parameter set object. At the same time, the object is checked for integrity to ensure that all values ​​are not empty and are in the correct format, providing standardized data input for the subsequent dynamic calculation process and generating the teaching video slice parameter set.

[0034] In the dynamic duration calculation formula, the saturation characteristics of the hyperbolic tangent function are used to simulate the nonlinear compression effect of attention decay on learning duration. At the same time, a logarithmic growth recovery factor is introduced so that the generated slice duration can dynamically increase as the slice number increases, thereby achieving a "short first, long later" rhythm guidance to help users gradually recover from fatigue to a normal learning state. The base slice duration value represents the standard learning unit length that a user can maintain focus under ideal, fatigue-free conditions, measured in seconds. It is obtained by directly reading from the configuration items in the teaching resource index table. These configuration items are determined based on survival analysis of historical viewing behavior data from 100,000 users. The maximum duration segment with an average user viewing completion rate of over 80% is calculated, rounded, and used as the baseline. Here, it is set... Second; The state sensitivity coefficient is a dimensionless value that characterizes the intensity of the algorithm's response to the user's current fatigue level. A higher value means that the same fluctuation will result in a shorter data slice. It is obtained by calculating the average time required for the user to return to normal interaction after receiving intervention, based on the user's historical intervention response records. Compare it with the standard recovery time baseline. (Set for 60 seconds) Comparison is performed, and the calculation formula is as follows: ,in With the adjustment constant set to 0.5, and the average historical recovery time for users set to 180 seconds, then... ; This is a normalized state index value, a dimensionless parameter that represents the relative fatigue of mapping the original interval variance fluctuation value to a standard interval. It is obtained by reading the interval variance fluctuation value obtained in the preceding steps. And retrieve the maximum fluctuation value from the user's historical records. The calculation is performed using the linear normalization method, and the formula is as follows: If the calculated result is greater than 1, it will be truncated to 1. This is set here. Historical maximum value Then the calculation yields ; The personalized recovery coefficient is a dimensionless value that represents the ratio of a user's focus recovery speed after adapting to a fragmented learning pace to the standard speed. It is obtained by analyzing the slope of the user's interaction frequency recovery after completing fragmented tasks in the past week. Compare it with the average recovery slope of the group The comparison is performed, and the calculation formula is as follows: Here, we assume the user's recovery slope is slightly better than the average level, and the calculated ratio is... It is 1.2; The slice number is a dimensionless integer that represents the position of the currently generated slice in the recombined sequence. It starts from 1 and increments to introduce a time-dimensional recovery variable into the denominator of the formula, ensuring that the slice duration gradually approaches the baseline value as playback progresses.

[0035] Calculations based on parameters: Substitute the above-mentioned values ​​into the formula to calculate the dynamic duration of the first two slices: First, calculate the numerator (dimensionless) of the hyperbolic tangent term: ; ; For the first slice ( ): Calculate the denominator (dimensionless): ; Calculate the regulation factor (dimensionless): ; Calculate the final duration (seconds): Second; For the second slice ( ): Calculate the denominator (dimensionless): ; Calculate the regulation factor (dimensionless): ; Calculate the final duration (seconds): Second; This result indicates that in detecting fluctuations in user attention ( In the initial stage, the first slice is drastically compressed to about 102 seconds to retain users with very low time burden. Then, the second slice is quickly extended to about 192 seconds. This dynamic and incremental length arrangement conforms to the psychological cognitive law of users gradually recovering from fatigue.

[0036] Based on the dynamic duration value, the video stream capture engine is started, and a playback cursor is initialized to point to the current unplayed start position of the originally planned long-duration teaching video, for example, at 10 minutes and 02 seconds. The dynamic duration sequence generated above is then traversed. In each iteration, the required slice duration is first read, and the cursor position is increased by this duration to obtain the predetermined end position. The difference between the remaining total video duration and the required duration is calculated, and this difference is compared with the preset minimum fragment threshold (e.g., 30 seconds). If the remaining duration is longer than the current dynamic duration and the remaining portion exceeds the minimum fragment threshold, then physical or logical segmentation is performed strictly according to the current dynamic duration, and the video stream data within this time period is extracted. If the remaining duration is shorter than the current dynamic duration or the remaining portion is insufficient to form an independent slice, then the calculated dynamic duration is ignored, and all remaining video content is directly merged into a complete final slice. For each determined slice segment, keywords of the subtitle text within this time period are extracted using natural language processing technology. Combined with the start and end timestamps, a metadata structure containing a content summary and time index is constructed. These structures are stored in a list in the order of generation to form a dynamically adjusted slice set.

[0037] When an attention decay status indicator is detected, the reloading module of the front-end player is activated. It reads the dynamically adjusted duration collection of segments from memory, iterates through the metadata of each segment in the collection, and instantiates an independent playback node object for each segment. The streaming media playback address is configured in this object, which includes start and end time parameters accurate to milliseconds (e.g., #t=600,702.5). At the same time, a unique UUID is generated for each node as a node identifier. Based on the knowledge point content summary within the segment, the corresponding preview thumbnail and short title are obtained from the image and text generation interface. These visual element attributes are filled into the node object. The interaction attributes of the node are set to "automatic continuous playback" or "manual trigger required," which is configured according to the user's historical preferences. After encapsulating all attributes, these objects with complete playback capabilities and display information are concatenated according to the segmentation order to obtain fragmented knowledge point nodes.

[0038] The steps to obtain the dynamic learning recommendation queue are as follows: Based on fragmented knowledge point nodes or simple review resources, the unique identifier of the original teaching resource is read, and the corresponding original resource item is located in the original learning path using the unique identifier. The original resource item is removed from the learning path, and fragmented knowledge point nodes or simple review resources are inserted in their original positions in order of playback priority. After the insertion is completed, the entire learning path is traversed, and continuous playback order index and outline order index are regenerated from beginning to end to form a dynamic learning recommendation queue.

[0039] Specifically, based on fragmented knowledge point nodes or simple review resources, the current learning path execution thread is suspended, the playlist data structure is locked to prevent concurrent modifications, the unique identifier of the originally scheduled long-duration teaching video is read, the current user's learning path queue is traversed, the node position of the original long video is located by comparing the identifier, the deletion operation is performed to remove the original resource entry from the queue, the queue is kept locked, the index pointer of the removal position is obtained, and the previously generated fragmented knowledge point nodes are inserted into the index position in order of the predetermined playback priority. If there are simple review resources (such as test questions or summary cards), they are interspersed among the fragmented nodes to adjust the rhythm. After the insertion of all nodes is completed, the queue is unlocked, the full traversal program is started, and the consecutive playback order index (Sequence ID) and outline order index (Outline ID) are reallocated starting from the head of the queue to ensure that the newly inserted fragmented content is seamlessly connected with the original subsequent content in terms of logical sequence number. The path status in the database is updated to form a dynamic learning recommendation queue.

Claims

1. A learning path intelligent recommendation system based on user behavior big data analysis, characterized in that, The system includes: The behavioral data construction module is used to collect data such as mouse trajectory coordinates, pause click frequency and answer response time generated by users on the online learning terminal, construct a user interaction timestamp sequence, set a fixed-length time sliding window to capture the user interaction timestamp sequence, and generate a set of windowed interaction intervals. The feature index calculation module is used to statistically analyze the dispersion of values ​​in the set of windowed interaction intervals, quantify the interval variance fluctuation value, statistically analyze the amount of interaction data per unit time, perform a linear fitting operation on the downward trend, calculate the slope of the fitted line, and associate and combine the slope with the interval variance fluctuation value to generate composite learning state feature parameters. The state determination and decision module is used to compare the degree of tilt in the composite learning state feature parameters with a preset attention loss rate threshold, and at the same time detect whether the interval variance fluctuation value exceeds the fatigue divergence threshold, generate an abnormal fatigue state signal, map the abnormal fatigue state signal to a preset rule scheduling table for strategy matching, and generate a content adaptation trigger instruction. The path reorganization and scheduling module is used to adapt trigger instructions according to the content, initiate slicing for the attention decay state, divide the originally planned long-term teaching video into fragmented knowledge point nodes, reorganize the original learning path according to the fragmented knowledge point nodes, and generate a dynamic learning recommendation queue.

2. The intelligent recommendation system for learning paths based on big data analysis of user behavior according to claim 1, characterized in that, The steps for obtaining the set of windowed interaction intervals are as follows: Collect mouse trajectory coordinates, pause click frequency, and response time of users on the online learning platform. Use the same timestamp format as the same benchmark, sort them in ascending order of time occurrence, remove duplicate timestamps while maintaining the first order, and associate the sorted timestamps with behavior identifiers to form a user interaction timestamp sequence. Based on the user interaction timestamp sequence, a fixed-length time sliding window boundary is set, the smallest timestamp is selected as the starting point, and the window is shifted backward by a fixed length and the left-closed and right-open rule is adopted to extract the timestamp subsequence within each time sliding window, forming a windowed user interaction timestamp sequence fragment set. Based on the set of windowed user interaction timestamp sequence segments, the time difference between adjacent timestamps is calculated for each segment. The minimum and maximum time differences within the segments are scaled proportionally to a uniform range. The scaled time differences are then merged in the order of the segments to form a set of windowed interaction intervals.

3. The intelligent recommendation system for learning paths based on big data analysis of user behavior according to claim 1, characterized in that, The steps for obtaining the interval variance fluctuation value are as follows: Based on the set of windowed interaction intervals, the data type and value range of each interaction interval value are verified, interval values ​​less than zero and missing values ​​are removed, and all valid interaction interval values ​​are retained to form a valid windowed interaction interval sequence. Based on the effective windowed interaction interval sequence, the number of interval values ​​in the sequence is counted, the average of all interval values ​​is calculated as the mean value, the square of the deviation between each interval value and the mean value is calculated and averaged to obtain the variance value, the square root of the variance value is taken to obtain the standard deviation value, and the average interaction interval value of the user's historical learning cycle is retrieved as the historical mean value, thus forming the mean value, variance value, standard deviation value and historical mean value. The interval variance fluctuation value is calculated based on the mean, variance, standard deviation and historical mean values.

4. The intelligent recommendation system for learning paths based on big data analysis of user behavior according to claim 1, characterized in that, The steps for obtaining the composite learning state feature parameters are as follows: Based on the user interaction timestamp sequence, the data is divided into segments of equal width according to the unit time length. Each segment is counted to form a unit time interaction data sequence. The difference between adjacent counts is calculated, and a downward trend is determined by the proportion of consecutive negative differences exceeding a preset threshold value. A set of downward trend segments is extracted, and a straight line fitting operation is performed on the set of downward trend segments. The slope of the fitted straight line for each segment is calculated, and the slope of the largest negative fitted straight line is selected as the slope. The slope and the interval variance fluctuation value are concatenated into a binary vector in a fixed order to form a composite learning state feature parameter.

5. The intelligent recommendation system for learning paths based on big data analysis of user behavior according to claim 1, characterized in that, The steps for obtaining the abnormal fatigue state signal are as follows: Based on the composite learning state characteristic parameters, the tilt degree and interval variance fluctuation value are read item by item, and compared with the preset attention loss speed threshold and fatigue divergence threshold respectively. The name of the out-of-bounds item, the out-of-bounds direction, the out-of-bounds amplitude, and the trigger time and source identifier are added. The comparison results are merged into a unified abnormal structure to generate an abnormal fatigue state signal.

6. The intelligent recommendation system for learning paths based on big data analysis of user behavior according to claim 1, characterized in that, The steps for obtaining the content adaptation trigger command are as follows: Based on the abnormal fatigue state signal, the name of the out-of-bounds item, the direction of the out-of-bounds, the magnitude of the out-of-bounds, the trigger time and the source identifier are parsed. The matching row is retrieved in the preset rule scheduling table according to the field index. The path adjustment scheme identifier and the path adjustment scheme content are read to generate the path adjustment scheme. Based on the path adjustment scheme, organize the content target, playback order, insertion position, execution timing and effective conditions, encapsulate them into an action command header and parameter body, add the path adjustment scheme identifier and check code, set the trigger range and validity period, and generate content adaptation trigger commands.

7. The intelligent recommendation system for learning paths based on big data analysis of user behavior according to claim 1, characterized in that, The steps for obtaining the fragmented knowledge point nodes are as follows: Based on the content adaptation trigger instruction, retrieve the teaching resource index table, extract the total duration value and basic slice duration value of the original long-term teaching video, read the interval variance fluctuation value in the composite learning state feature parameters, and generate a teaching video slice parameter set. Based on the set of teaching video slice parameters, calculate the dynamic duration value of each slice one by one; Based on the dynamic duration value, video segments are extracted one by one starting from the position where the original long-term teaching video has not been played. If the remaining video duration is longer than the current dynamic duration value, the segments are extracted according to the current duration. If it is shorter, the remaining video portion is divided into the last segment. The start and end times and knowledge points are extracted for each segment to form a set of segments with dynamically adjusted durations. When an attention decay status indicator is detected, the segments are loaded and played in the order of the dynamically adjusted duration set to obtain fragmented knowledge point nodes.

8. The intelligent recommendation system for learning paths based on big data analysis of user behavior according to claim 1, characterized in that, The steps for obtaining the dynamic learning recommendation queue are as follows: Based on the fragmented knowledge point nodes or simple review resources, the unique identifier of the original teaching resource is read, and the corresponding original resource entry is located in the original learning path using the unique identifier. The original resource entry is removed from the learning path, and the fragmented knowledge point nodes or simple review resources are inserted into their original positions in order of playback priority. After the insertion is completed, the entire learning path is traversed, and a continuous playback order index and outline order index are regenerated from beginning to end to form a dynamic learning recommendation queue.

Images