A personalized learning path dynamic regulation method and system based on limited computing resources

Abstract

The application discloses a kind of based on the dynamic regulation and control method and system of personalized learning path of limited computing resources, including the acquisition student completes the learning interaction when the learning characteristic data, based on the learning characteristic data adopts linear weighting model calculation student's mastery to current knowledge point, constructs finite state automaton, according to the size relationship of the mastery and preset first threshold value, second threshold value, triggers state conversion, and according to current state executes corresponding learning path regulation action.The application guarantees the accuracy of learning condition evaluation under the constraint of limited computing resources, and improves stability.

Description

Technical Field

[0001] This invention relates to learning path control methods, and in particular to a method and system for dynamic control of personalized learning paths based on limited computing resources. Background Technology

[0002] Existing adaptive learning systems typically employ the following technical solutions: collecting basic data such as students' answer accuracy and learning time; recommending content based on collaborative filtering or deep learning models; and updating the learning path using unlimited backtracking or global recomputation.

[0003] When the above technical solution is actually deployed in a server environment with limited resources, the following technical contradictions exist:

[0004] (1) The contradiction between the single dimension of learning analysis and the accuracy of assessment: Most systems only count the accuracy rate or duration. The correlation coefficient between single feature assessment and teacher subjective evaluation is usually less than 0.65, which makes it impossible to accurately locate the individual student's knowledge point-level weaknesses.

[0005] (2) The contradiction between high-precision models and limited computing resources: Deep learning-based content recommendation models require high computing resources (GPU support, large memory). Actual tests show that when running deep learning models on resource-limited servers (4-core CPU, 8GB memory, 100Mbps bandwidth), a single inference takes about 200ms, the CPU utilization exceeds 95% when there are 30 concurrent users, and the response timeout rate is over 15%.

[0006] (3) The contradiction between unlimited backtracking and system stability: For systems that support knowledge point backtracking, when the knowledge graph depth is n, the worst case requires maintaining 2^n state variables. Taking n=10 as an example, the number of state variables reaches 1024, the memory usage surges from the baseline 200MB to 1.2GB, and the response time increases from 200ms to more than 5000ms, leading to system crash.

[0007] (4) The conversion efficiency of teacher-side learning data to teaching instructions is low: the existing system only provides raw data display, and teachers need to manually identify common weaknesses, which takes about 60 minutes per class. From the perspective of computer system, this is a problem of conversion efficiency of "data → information → executable instructions". Summary of the Invention

[0008] To address the shortcomings of the existing technologies, this invention provides a method for dynamically adjusting personalized learning paths based on limited computing resources, resolving the issue of balancing accuracy and stability in learning assessment under limited computing resource constraints. This invention also provides a system for dynamically adjusting personalized learning paths based on limited computing resources.

[0009] The technical solution of the present invention is as follows:

[0010] A method for dynamically adjusting personalized learning paths based on limited computing resources, comprising:

[0011] Step S1: Collect learning characteristic data of students when they complete the learning interaction. The learning characteristic data includes: correctness of answers, time spent answering questions, number of times they ask for help, and task completion rate.

[0012] Step S2: Calculate the student's mastery of the current knowledge point using a linear weighted model based on the learning characteristics data;

[0013] Step S3: Construct a finite state automaton with a state set S = {locked, learning, completed, intervention}. Based on the relationship between the mastery level and the preset first and second thresholds, trigger state transitions and execute corresponding learning path adjustment actions according to the current state.

[0014] Among them, the state "locked" indicates that the knowledge point is not unlocked and requires backtracking; the state "learning" indicates that the knowledge point is being learned; the state "completed" indicates that the knowledge point has been completed; and the state "intervention" indicates that human intervention is required.

[0015] In the learning path control, when the state transition requires backtracking to the previous knowledge point of the current knowledge point, a backtracking depth limit based on the maximum backtracking depth threshold is executed: the number of backtracking layers is recorded, and the backtracking operation is performed only when the number of backtracking layers does not exceed the preset maximum backtracking depth threshold, and the current knowledge point state is changed to locked; otherwise, the backtracking is stopped, and the state is changed to intervention.

[0016] Furthermore, the linear weighted model is as follows:

[0017] , ,

[0018] in, To control the degree, The ratio of correct answers to the total number of answers is denoted as _____. This is the ratio of the preset standard time to the average time spent answering questions. The value is set to 1 if the preset standard time taken is greater than the average time taken to answer questions. For task completion rate, , , , These are the weighting coefficients.

[0019] Furthermore, , , , The weighting combination was optimized using a simulated dataset of 500 students, resulting in a correlation coefficient of 0.87 between student assessment and teacher subjective evaluation.

[0020] Furthermore, before step S3, step S2-1 is performed, which divides the mastery level into three levels: "not mastered", "basically mastered" and "mastered" based on the relationship between the mastery level and the preset first threshold and second threshold. The corresponding knowledge point content is pushed according to the mastery level. The knowledge point content has a set difficulty value, and the deviation between the pushed knowledge point content and the mastery level is controlled within the deviation threshold range.

[0021] Furthermore, the knowledge points are pushed out in order of priority: practice > video > text and images.

[0022] Furthermore, the state transition includes the following rules:

[0023] When the current knowledge point is in the "learning" state, if the mastery level is "mastered", the state changes to "completed"; if the mastery level is "not mastered", the state changes to "locked"; if the mastery level is "basically mastered", the knowledge point content will be repeatedly pushed and the state will remain unchanged until the number of pushes reaches the set threshold, at which point the state changes to "intervention"; if the current knowledge point is in the "locked" state, and the mastery level of the preceding knowledge point is "mastered", the state changes to "learning".

[0024] Furthermore, step S4 involves periodically calculating the percentage of students whose mastery level for a single knowledge point is "not mastered." When this percentage exceeds a threshold, a reinforcement package containing practice questions with multiple difficulty levels corresponding to "not mastered" and "basically mastered" levels is generated and pushed to the teacher. By automatically identifying common weaknesses, the time spent by teachers in identifying class learning situations can be reduced, and the efficiency of converting learning data into teaching instructions can be improved.

[0025] Another technical solution of the present invention is: a personalized learning path dynamic control system based on limited computing resources, comprising:

[0026] Learning Situation Collection Module: Collects learning situation characteristic data of students when they complete learning interactions. The learning situation characteristic data includes: answer accuracy, answer time, number of times they ask for help, and task completion rate.

[0027] Mastery assessment module: Based on the learning characteristics data, a linear weighted model is used to calculate the student's mastery of the current knowledge point;

[0028] Path control module: Constructs a finite state automaton with a state set S = {locked, learning, completed, intervention}. Based on the relationship between the mastery level and preset first and second thresholds, it triggers state transitions and executes corresponding learning path control actions according to the current state.

[0029] Among them, the state "locked" indicates that the knowledge point is not unlocked and requires backtracking; the state "learning" indicates that the knowledge point is being learned; the state "completed" indicates that the knowledge point has been completed; and the state "intervention" indicates that human intervention is required.

[0030] In the learning path control, when the state transition requires backtracking to the previous knowledge point of the current knowledge point, a backtracking depth limit based on the maximum backtracking depth threshold is executed: the number of backtracking layers is recorded, and the backtracking operation is performed only when the number of backtracking layers does not exceed the preset maximum backtracking depth threshold, and the current knowledge point state is changed to locked; otherwise, the backtracking is stopped, and the state is changed to intervention.

[0031] Furthermore, it includes a personalized push module: based on the relationship between the mastery level and preset first and second thresholds, it divides the mastery level into three levels: "not mastered", "basically mastered" and "mastered", and pushes corresponding knowledge point content according to the mastery level. The knowledge point content has a set difficulty value, and the deviation between the pushed knowledge point content and the mastery level is controlled within the deviation threshold range.

[0032] Furthermore, the system includes an aggregation module that periodically calculates the percentage of students who have "not mastered" a single knowledge point. When the percentage exceeds a threshold, the system generates a reinforcement package containing practice questions with multiple difficulty levels corresponding to "not mastered" and "basically mastered" levels and pushes it to the teacher.

[0033] Compared with the prior art, the present invention has the following advantages:

[0034] This invention uses a lightweight weighted evaluation model (time complexity O(1)) to replace the deep learning model, and combines finite state automata and maximum backtracking depth limit to change the worst-case time complexity from exponential growth to linear growth. This ensures the accuracy of learning assessment while reducing computer resource consumption, significantly reducing system response time, and reducing system crashes. Attached Figure Description

[0035] Figure 1 This is a flowchart illustrating the dynamic adjustment method for personalized learning paths based on limited computing resources according to the present invention.

[0036] Figure 2 This is a schematic diagram of a personalized learning path dynamic adjustment method based on limited computing resources, as an example.

[0037] Figure 3 This is a schematic diagram of the state transitions of a finite state automaton.

[0038] Figure 4 This is a system architecture diagram for a personalized learning path dynamic adjustment system based on limited computing resources.

[0039] Figure 5 This is a data flow sequence diagram for a personalized learning path dynamic control system based on limited computing resources. Detailed Implementation

[0040] The present invention will be further described below with reference to embodiments, but these are not intended to limit the scope of the invention.

[0041] Please combine Figures 1 to 3 As shown in the figure, the personalized learning path dynamic control method based on limited computing resources according to an embodiment of the present invention includes the following steps:

[0042] Step S1: Collect student learning characteristic data when completing learning interactions. The resulting feature data tuple is: (StudentID, KnowledgePointID, IsCorrect, TimeCost_ms, HelpCount, TaskCompletionRate), where StudentID represents the student's identity, KnowledgePointID represents the specific knowledge point, and IsCorrect, TimeCost_ms, HelpCount, and TaskCompletionRate are learning characteristic data with the following specific meanings:

[0043] IsCorrect indicates the correctness of the answer (0 / 1);

[0044] TimeCost_ms represents the time taken to answer the question (in milliseconds);

[0045] HelpCount indicates the number of times help has been requested (click for tips / view details);

[0046] TaskCompletionRate represents the task completion rate (number of completed subtasks / total number of subtasks, with a value of [0,1]).

[0047] Step S2: Calculate students' mastery of the current knowledge point using a linear weighted model based on learning characteristics data; the specific linear weighted model is as follows:

[0048] , ,

[0049] in, To control the degree, The ratio of correct answers to the total number of answers is denoted as _____. This is the ratio of the preset standard time to the average time spent answering questions. The value is set to 1 if the preset standard time taken is greater than the average time taken to answer questions. For task completion rate, , , , These are the weighting coefficients.

[0050] The weighting coefficients take the following values: , , , The weight combination was optimized using a simulated dataset (500 students) to achieve a correlation coefficient of 0.87 between student assessment and teacher subjective evaluation.

[0051] Step S2-1: Based on the relationship between the level of mastery and preset first and second thresholds, the mastery level is divided into three categories: "not mastered," "basically mastered," and "mastered." Corresponding knowledge points are then pushed according to the mastery level. These knowledge points have a set difficulty value, and the deviation between the pushed knowledge points and the mastery level is controlled within a deviation threshold. Specifically, the first threshold is set to 0.75, the second threshold is set to 0.4, and the mastery level is determined by the following rules:

[0052] P ≥ 0.75 → Status = "Mastered";

[0053] 0.4 ≤ P < 0.75 → Status = “Basic Mastery”;

[0054] P < 0.4 → Status = "Not Mastered".

[0055] The knowledge points pushed are selected from the system's built-in content library. Each knowledge point is marked with (ContentID, KnowledgePointID, DifficultyLevel, ContentType). Among them, ContentID represents the knowledge point content number, KnowledgePointID represents the corresponding knowledge point number, DifficultyLevel is the difficulty value, which takes the value [0,1], and ContentType is the specific content.

[0056] The push notification method is based on the student's current mastery level P of the knowledge point, selecting content with a difficulty level D that satisfies |DP| ≤ 0.15, where 0.15 is the deviation threshold. When multiple knowledge point contents meet the above requirements, they are pushed in the order of priority: practice > video > text / image.

[0057] The correspondence between specific push notification content and the level of understanding is explained as follows:

[0058]

[0059] Step S3: Construct a finite state automaton with a state set S = {locked, learning, completed, intervention}. Trigger state transitions based on the relationship between the mastery level and the preset first and second thresholds, and execute corresponding learning path adjustment actions based on the current state.

[0060] Among them, the state "locked" indicates that the knowledge point is not unlocked and requires backtracking; the state "learning" indicates that the knowledge point is being learned; the state "completed" indicates that the knowledge point has been completed; and the state "intervention" indicates that human intervention is required.

[0061] In learning path control, when a state transition requires backtracking to the preceding knowledge point of the current knowledge point, a backtracking depth limit based on a maximum backtracking depth threshold is implemented: the number of backtracking levels is recorded, and the backtracking operation is performed only if the number of backtracking levels does not exceed the preset maximum backtracking depth threshold, and the current knowledge point's state is changed to locked; otherwise, backtracking stops, and the state is changed to intervention. The specific state transition rules are as follows:

[0062]

[0063] By limiting the number of backtracking layers through a maximum backtracking depth threshold (taking 3 layers as an example), the number of state variables in the worst-case scenario is constrained from 2^n to 1+2+4+8=15, and memory usage is reduced from 1.2GB to below 200MB. For the same knowledge point, content is pushed a maximum of 3 times. If the target is not met after 3 attempts, intervention is initiated to avoid infinite loops and keep the maximum computational cost of a single knowledge point at a constant level.

[0064] Step S4: Periodically count the percentage of students whose mastery level for a single knowledge point is "not mastered": R_k = (number of students who have not mastered / total number of students in the class) × 100%. When the percentage of students exceeds the threshold, such as 30%, generate a reinforcement package containing multiple difficulty values ​​corresponding to the mastery levels of "not mastered" and "basically mastered" (for example, 10 practice questions with a difficulty level of 0.4 to 0.6, forming a whole-class practice paper) and push it to the teacher.

[0065] Please combine Figure 4 and Figure 5 As shown, the hardware architecture of the personalized learning path dynamic adjustment system based on limited computing resources in this embodiment of the invention consists of a student terminal, a cloud server, and a teacher terminal. The cloud server performs dynamic adjustment of the personalized learning path and includes:

[0066] Learning Situation Collection Module: Collects learning situation characteristic data when students complete learning interactions, specifically executing step 1 of the aforementioned method.

[0067] Mastery assessment module: Based on the learning characteristics data, a linear weighted model is used to calculate the student's mastery of the current knowledge point, specifically by executing step 2 of the aforementioned method.

[0068] Path control module: Construct a finite state automaton with a state set S = {locked, learning, completed, intervention}. Based on the relationship between the mastery level and the preset first and second thresholds, trigger state transitions and execute corresponding learning path control actions according to the current state, specifically executing step 3 of the aforementioned method.

[0069] Personalized push module: Based on the mastery level and the relationship between the mastery level and the preset first threshold and second threshold, the mastery level is divided into "not mastered", "basically mastered" and "mastered". The corresponding knowledge point content is pushed according to the mastery level. The knowledge point content has a set difficulty value. The deviation between the pushed knowledge point content and the mastery level is controlled within the deviation threshold range. Specifically, step 2-1 of the aforementioned method is executed.

[0070] Aggregation module: Periodically count the percentage of students whose mastery level of a single knowledge point is "not mastered". When the percentage of students exceeds the percentage threshold, generate a reinforcement package containing multiple difficulty levels corresponding to the mastery levels of "not mastered" and "basically mastered" and push it to the teacher, specifically executing step 4 of the aforementioned method.

[0071] Based on the above embodiments, the system response verification in a resource-constrained scenario was carried out in the following environment: a standard cloud server (4-core CPU, 8GB memory), 30 students simultaneously learning the knowledge point of "loop structure" online, and 30 students simultaneously submitting initial assessment data.

[0072] The results are as follows:

[0073]

[0074] The typical value of the existing technology is a scheme that uses a deep learning model (deep factorization machine) for content recommendation and unlimited backtracking, which is theoretically calculated under the same deployment environment as this invention (4-core CPU, 8GB memory, 30 concurrent connections).

[0075] Another test result of this invention verifies the effectiveness of the 3-layer depth constraint mechanism:

[0076] Scenario: Student Xiaohua repeatedly makes mistakes on knowledge points with a depth of 5, and the P-value is consistently <0.4.

[0077] Unrestricted backtracking scheme: The system backtracks to depth 4→3→2→1→0, maintaining 2^5=32 states, and the response time increases to 5000ms before timeout.

[0078] The solution of this invention: the system stops after backtracking to depth 3 and enters the intervention state. The response time is always ≤500ms. The teacher can intervene manually after receiving the notification.

[0079] Furthermore, in a class of 30 students, after learning the "definition of variables," the system of this invention identified 12 students (40%) who had not mastered the concept, automatically marking them as common weaknesses. Teachers can generate 10 reinforcement exercises with a single click, reducing lesson preparation time from 60 minutes to 10 minutes, demonstrating that this invention can significantly improve the efficiency of converting student learning data into teaching instructions.

Claims

1. A method for dynamically adjusting personalized learning paths based on limited computing resources, characterized in that, include: Step S1: Collect learning characteristic data of students when they complete the learning interaction. The learning characteristic data includes: correctness of answers, time spent answering questions, number of times they ask for help, and task completion rate. Step S2: Calculate the student's mastery of the current knowledge point using a linear weighted model based on the learning characteristics data; Step S3: Construct a finite state automaton with a state set S = {locked, learning, completed, intervention}. Based on the relationship between the mastery level and the preset first and second thresholds, trigger state transitions and execute corresponding learning path adjustment actions according to the current state. Among them, the state "locked" indicates that the knowledge point is not unlocked and requires backtracking; the state "learning" indicates that the knowledge point is being learned; the state "completed" indicates that the knowledge point has been completed; and the state "intervention" indicates that human intervention is required. In the learning path control, when the state transition requires backtracking to the previous knowledge point of the current knowledge point, a backtracking depth limit based on the maximum backtracking depth threshold is executed: the number of backtracking layers is recorded, and the backtracking operation is performed only when the number of backtracking layers does not exceed the preset maximum backtracking depth threshold, and the current knowledge point state is changed to locked; otherwise, the backtracking is stopped, and the state is changed to intervention.

2. The method for dynamic adjustment of personalized learning paths based on limited computing resources according to claim 1, characterized in that, The linear weighted model is as follows: ， ， in, To control the degree, The ratio of correct answers to the total number of answers is denoted as _____. This is the ratio of the preset standard time to the average time spent answering questions. The value is set to 1 if the preset standard time taken is greater than the average time taken to answer questions. For task completion rate, , , , These are the weighting coefficients.

3. The method for dynamic adjustment of personalized learning paths based on limited computing resources according to claim 2, characterized in that, ， ， ， 。 4. The method for dynamic adjustment of personalized learning paths based on limited computing resources according to claim 1, characterized in that, The process includes step S2-1 before step S3, which divides the mastery level into three levels: "not mastered", "basically mastered" and "mastered" based on the relationship between the mastery level and the preset first threshold and second threshold. The process also pushes corresponding knowledge point content according to the mastery level. The knowledge point content has a set difficulty value, and the deviation between the pushed knowledge point content and the mastery level is controlled within the deviation threshold range.

5. The method for dynamic adjustment of personalized learning paths based on limited computing resources according to claim 4, characterized in that, The knowledge points are pushed out in order of priority: exercises > videos > text and images.

6. The method for dynamic adjustment of personalized learning paths based on limited computing resources according to claim 4, characterized in that, The state transitions include the following rules: When the current knowledge point is in the learning state, if the mastery level is "mastered", the state changes to completed; if the mastery level is "not mastered", the state changes to locked; if the mastery level is "basically mastered", the knowledge point content will be repeatedly pushed and the state will remain unchanged until the number of times the knowledge point content is pushed reaches the set threshold, at which point the state changes to intervention. When the current knowledge point is in the "locked" state, and the mastery level of the prerequisite knowledge point is "mastered", the state changes to "learning".

7. The method for dynamic adjustment of personalized learning paths based on limited computing resources according to claim 1, characterized in that, This includes step S4, which involves periodically calculating the percentage of students who have "not mastered" a single knowledge point. When the percentage exceeds a threshold, a reinforcement package containing practice questions with multiple difficulty levels corresponding to "not mastered" and "basically mastered" knowledge points is generated and pushed to the teacher.

8. A personalized learning path dynamic control system based on limited computing resources, characterized in that, include: Learning Situation Collection Module: Collects learning situation characteristic data of students when they complete learning interactions. The learning situation characteristic data includes: answer accuracy, answer time, number of times they ask for help, and task completion rate. Mastery assessment module: Based on the learning characteristics data, a linear weighted model is used to calculate the student's mastery of the current knowledge point; Path control module: Constructs a finite state automaton with a state set S = {locked, learning, completed, intervention}. Based on the relationship between the mastery level and preset first and second thresholds, it triggers state transitions and executes corresponding learning path control actions according to the current state. Among them, the state "locked" indicates that the knowledge point is not unlocked and requires backtracking; the state "learning" indicates that the knowledge point is being learned; the state "completed" indicates that the knowledge point has been completed; and the state "intervention" indicates that human intervention is required. In the learning path control, when the state transition requires backtracking to the previous knowledge point of the current knowledge point, a backtracking depth limit based on the maximum backtracking depth threshold is executed: the number of backtracking layers is recorded, and the backtracking operation is performed only when the number of backtracking layers does not exceed the preset maximum backtracking depth threshold, and the current knowledge point state is changed to locked; otherwise, the backtracking is stopped, and the state is changed to intervention.

9. The personalized learning path dynamic control system based on limited computing resources according to claim 8, characterized in that, The system includes a personalized push module: based on the relationship between the mastery level and preset first and second thresholds, the mastery level is divided into three levels: "not mastered", "basically mastered" and "mastered". The module pushes corresponding knowledge content according to the mastery level. The knowledge content has a set difficulty value, and the deviation between the pushed knowledge content and the mastery level is controlled within the deviation threshold range.

10. The personalized learning path dynamic control system based on limited computing resources according to claim 8, characterized in that, The system includes an aggregation module that periodically calculates the percentage of students who have "not mastered" a single knowledge point. When the percentage exceeds a threshold, it generates a reinforcement package containing practice questions with multiple difficulty levels corresponding to "not mastered" and "basically mastered" levels and pushes it to the teacher.

Images