Network school student personalized learning path recommendation system

Abstract

The invention discloses a network school student personalized learning path recommendation system, and relates to the technical field of education, and the recommendation system comprises a data collection module which obtains the micro-expression and limb details of a student during learning based on a camera, collects the learning behavior data of the student, and judges the knowledge ability data of the student through image analysis and limb analysis; the student portrait construction module obtains the knowledge weakness direction of the current student based on the knowledge ability data of the student, and constructs a multi-dimensional student portrait; and the knowledge graph construction module is used for constructing an association graph of the subject knowledge system, and the association graph comprises a preposed dependency relationship and a difficulty level between knowledge points. According to the invention, learning is reinforced and learning rhythm is adjusted in real time through a recommendation algorithm layer, proper challenges are introduced into an anti-fragility mechanism to enhance knowledge mastering toughness, path output is coupled with real-time states of students, course design and teaching can be fed back, learning efficiency and achievement sense of students are improved, and technical barriers of network schools are enhanced.

Description

Technical Field

[0001] This invention relates to the field of educational technology, specifically to a personalized learning path recommendation system for online school students. Background Technology

[0002] With the deep integration of internet technology and the education industry, online schools, with their advantages of breaking through time and space limitations and having extensive resource coverage, have become an important part of the modern education system. However, current online school learning systems still have significant shortcomings in meeting students' personalized learning needs, making it difficult to adapt to different students' knowledge bases, learning abilities, and cognitive habits. This leads to problems such as low learning efficiency and poor learning experience, specifically manifested in the following aspects:

[0003] Students have varying knowledge bases and learning abilities. A uniform learning path can lead to high-achieving students finding the content too easy and losing interest, while low-achieving students may fall behind and give up. Students also have different learning goals and needs; some want remedial learning while others want advanced learning. Uniform learning content cannot meet these diverse needs. Furthermore, the lack of effective tracking and feedback on students' learning processes makes it impossible to adjust learning paths in a timely manner to improve learning efficiency. To address these issues, we propose a personalized learning path recommendation system for online school students. Summary of the Invention

[0004] To address the aforementioned technical issues, a personalized learning path recommendation system for online school students is provided. This technical solution resolves the problems described above.

[0005] To achieve the above objectives, the technical solution adopted by this invention is: a personalized learning path recommendation system for online school students, the recommendation system comprising:

[0006] The data acquisition module uses a camera to capture students' micro-expressions and body language details while they are learning, collects students' learning behavior data, and uses image analysis and body language analysis to determine students' knowledge and ability data.

[0007] The student profile building module, based on students' knowledge and ability data, identifies the areas where students currently have weak knowledge.

[0008] Constructing multi-dimensional student profiles;

[0009] The knowledge graph construction module builds a relational graph of the subject knowledge system, including the prerequisite dependencies and difficulty levels between knowledge points;

[0010] The recommendation algorithm module is connected to the student profile construction module and the knowledge graph construction module. Based on the multi-model fusion algorithm, the learning path is optimized in real time through inner-loop reinforcement learning, and the outer loop introduces cognitive meta-strategies to generate heuristic learning branches. An antifragile recommendation mechanism is embedded to generate adaptive personalized learning paths.

[0011] The path output and adjustment module is used to display recommended paths and make dynamic adjustments.

[0012] Preferably, the camera is deployed in the classroom to capture the panoramic view of the classroom in real time. The acquired data is preprocessed, and the problems of uneven lighting and motion blur are optimized by image enhancement algorithm. The target detection model is used to locate the student's facial area and limb skeleton, and the consecutive frames are aligned.

[0013] Micro-expression analysis uses a facial key point detection algorithm to locate facial feature points and calculate dynamic parameters such as eye opening and closing, eyebrow height, and mouth corner offset.

[0014] By capturing the rate of change in micro-expressions using the temporal difference method, we can identify the transition patterns of emotions such as surprise, confusion, and relief.

[0015] Establish an expression feature vector library, compare real-time data with preset benchmark templates for focus, confusion and fatigue, and output the probability value of emotional state.

[0016] Preferably, limb analysis uses skeletal key point recognition technology to extract skeletal node coordinates and calculate parameters such as trunk tilt angle, arm range of motion, and relative distance between the hand and face;

[0017] By matching action sequence patterns, we can identify learning behaviors such as resting one's chin on one's hand in thought, frequently twirling a pen, and leaning forward.

[0018] Construct a physical activity index to quantify the frequency and amplitude of movements and distinguish between active interaction and passive slackness;

[0019] By employing an attention mechanism to integrate facial expressions and body features, the system strengthens the judgment weight for actively solving the problem when it detects the combination of frowning, leaning forward, and pointing to the screen; and increases the confidence level of cognitive fatigue for the feature combination of wandering eyes, leaning back, and low body activity.

[0020] By using a time-series sliding window, a comprehensive behavioral feature vector for that period is generated. A machine learning model is trained based on a labeled dataset to establish a mapping relationship between behavioral features and cognitive states.

[0021] When a person consistently displays highly confused facial expressions and frequently adjusts their body language, it is determined to be a comprehension disorder of the knowledge point.

[0022] When low concentration is accompanied by low physical activity exceeding a threshold, it is marked as a decline in learning interest.

[0023] By combining specific learning scenarios, cognitive states are transformed into quantifiable knowledge and ability indicators, generating knowledge and ability data that includes the degree of mastery of each knowledge point, the type of cognitive gaps, and the best absorption methods.

[0024] Preferably, the steps for constructing the student profile building module are as follows:

[0025] Tag knowledge and skills data, establish a three-level indicator system of subject-module-knowledge point, and form a structured matrix;

[0026] By using horizontal comparison, vertical tracking, and correlation mining, we screened out knowledge points and weak clusters that were below average and had been stagnant for a long time, and determined the top 5 weak areas.

[0027] Expanding the three dimensions of cognitive style, learning strategy, and psychological state, and combining micro-expression and behavioral data to extract learning preferences, strategy efficiency, and psychological thresholds;

[0028] The weights of each dimension are dynamically adjusted according to the learning stage. Beginners focus on knowledge and cognitive style, intermediate learners strengthen strategies, and the final sprint focuses on weak connections.

[0029] The five-dimensional scoring, weak connection paths, and concrete tags are presented visually using radar charts and knowledge graphs.

[0030] By iterating through a dual-mode profile system of real-time micro-updates and periodic full updates, cross-period difference reports are generated, resulting in multi-dimensional student profiles.

[0031] Preferably, the three-level indicator system includes subject ability, knowledge module and specific knowledge point. Each indicator is assigned a quantitative value to form a structured knowledge ability matrix. Weaknesses are identified by calculating the standard deviation of mastery of each knowledge point within the same knowledge module and selecting knowledge points that are 1.5 standard deviations below the module average.

[0032] Preferably, the construction steps for the knowledge graph construction module are as follows:

[0033] The knowledge system is broken down, the core areas of the discipline are sorted out, and a three-level structure of chapters, knowledge modules and specific knowledge points is broken down, clarifying the definition, scope and application scenarios of each knowledge point;

[0034] Establish prerequisite dependencies, draw a knowledge point association network through subject logic analysis and teaching experience annotation, rely on polynomial operations through the solution of quadratic equations, and use graph algorithms to verify the completeness and rationality of the dependency chain. Divide the difficulty levels, combine curriculum standards and learning data, divide the knowledge points into three levels of cognitive complexity: basic, intermediate and advanced, and assign a quantitative difficulty coefficient.

[0035] A dynamic calibration mechanism is introduced to collect student learning data in real time. If the error rate of a certain knowledge point continues to be higher than the threshold, the difficulty coefficient is increased and the prerequisite dependencies are optimized.

[0036] Generate a visual graph, with node size representing difficulty and line thickness reflecting dependency intensity, and support filtering and viewing by subject and module.

[0037] Preferably, when verifying the pre-dependency relationship of knowledge points, the actual frequency of occurrence of the preset dependency path between each knowledge point is first counted, and the degree of consistency between the frequency and the theoretical derivation frequency is calculated. The higher the degree of consistency, the more reasonable the dependency relationship is. At the same time, it is checked whether there are isolated knowledge points or closed-loop dependencies to ensure that all knowledge points can be traced back to the basic starting point through a limited number of steps.

[0038] When classifying difficulty levels, the baseline for the basic level is determined based on the mastery requirements of knowledge points in the curriculum standards. Then, the average mastery time and error rate of the knowledge point in the learning data are analyzed. The higher the combined ratio of the two, the greater the upward adjustment of the difficulty level. In combination with cognitive complexity, the amount and closeness of the prerequisite knowledge required to understand the knowledge point are considered. The more prerequisite knowledge and the more complex the connection, the higher the difficulty coefficient.

[0039] Preferably, the real-time optimization learning path steps of inner-loop reinforcement learning in the recommendation algorithm module are as follows:

[0040] The state space is defined by using the student's current mastery of knowledge points, learning time, and real-time feedback as state parameters to construct a multi-dimensional state vector.

[0041] The action space design allows for optional actions such as continuing to practice the current knowledge point, jumping to previous basic knowledge points, and switching question types and difficulties. Each action corresponds to a clear learning content adjustment plan.

[0042] The reward mechanism is designed with the rate of improvement in knowledge mastery as the core reward, and dynamically weighted by learning time efficiency and emotional positivity to form an instant reward value.

[0043] The strategy is iteratively optimized. After each learning unit is completed, the action selection probability is updated based on the reward value. Path adjustment strategies that improve the reward value are retained first. An exploration factor is introduced to avoid path solidification.

[0044] Preferably, the steps in the recommendation algorithm module that introduce cognitive meta-strategies in the outer loop to generate heuristic learning branches are as follows:

[0045] Learning bottleneck identification involves continuously monitoring the duration of stagnation in mastery of knowledge points in the inner loop output and the repetition rate of similar errors. When any of these indicators exceeds a threshold, it is determined that a learning bottleneck has occurred.

[0046] Meta-strategy matching calls a preset cognitive meta-strategy library to match and adapt strategies based on the bottleneck type.

[0047] Cross-domain knowledge association: Based on the cross-disciplinary association network of knowledge graphs, it retrieves content from other disciplines that share common thinking with the current knowledge point and generates heuristic side content;

[0048] Branch integration and exit: Heuristic branches are inserted into the main learning path, a completion threshold is set for the branch, and the branch automatically returns to the main path after the threshold is reached. At the same time, the contribution of the branch to the bottleneck breakthrough is recorded.

[0049] The embedded antifragile recommendation mechanism first assesses the student's resilience baseline, defines the challenge boundary, and builds a three-level challenge content pool of light, medium and heavy, based on 10%-35% beyond the current ability. Regular and challenging content are pushed alternately, and the level is dynamically adjusted based on feedback. Breakthrough points are marked to strengthen transfer, and the challenge is re-imposed after addressing weaknesses. The baseline is updated weekly, and learning paths are recommended based on the student's own knowledge mastery.

[0050] Preferably, the path output and adjustment module uses a dual-view display of knowledge graph and progress dashboard. When a student makes three consecutive mistakes on a certain knowledge point, a preliminary basic review sub-line is automatically inserted. If the student completes the goal ahead of schedule and the accuracy rate exceeds 90%, more difficult extension content is pushed.

[0051] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0052] This invention uses a data acquisition layer to capture students' micro-expressions and body language details with a camera. Combined with image analysis, it transforms unstructured behavioral data into quantifiable indicators of focus and confusion, extending knowledge and ability assessment from "outcome judgment" to "process perception," laying the foundation for accurate student profiling. The student profiling module, based on process data, incorporates dynamic features such as learning status fluctuations and cognitive load thresholds to more precisely pinpoint areas of knowledge weakness. The recommendation algorithm layer adopts an "inner loop + outer loop + antifragile" model. The inner loop reinforces learning and adjusts the learning pace in real time, while the outer loop generates heuristic branches that match different cognitive styles through cognitive meta-strategies. The antifragile mechanism introduces appropriate challenges to enhance the resilience of knowledge acquisition. The path output is coupled with the student's real-time status and can also feed back into curriculum design and teaching, improving students' learning efficiency and sense of accomplishment while strengthening the technological barriers of online schools. Attached Figure Description

[0053] Figure 1 The recommended system framework diagram for this invention is shown below. Detailed Implementation

[0054] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0055] Reference Figure 1 As shown, a personalized learning path recommendation system for online school students includes:

[0056] The data acquisition module uses a camera to capture students' micro-expressions and body language details while they are learning, collects students' learning behavior data, and uses image analysis and body language analysis to determine students' knowledge and ability data.

[0057] The student profile building module, based on students' knowledge and ability data, identifies the areas where students currently have weak knowledge.

[0058] Constructing multi-dimensional student profiles;

[0059] The knowledge graph construction module builds a relational graph of the subject knowledge system, including the prerequisite dependencies and difficulty levels between knowledge points;

[0060] The recommendation algorithm module is connected to the student profile construction module and the knowledge graph construction module. Based on the multi-model fusion algorithm, the learning path is optimized in real time through inner-loop reinforcement learning, and the outer loop introduces cognitive meta-strategies to generate heuristic learning branches. An antifragile recommendation mechanism is embedded to generate adaptive personalized learning paths.

[0061] The path output and adjustment module is used to display recommended paths and make dynamic adjustments.

[0062] Cameras are deployed in the classroom to capture a panoramic view of the classroom in real time. The acquired data is preprocessed, and image enhancement algorithms are used to optimize the problems of uneven lighting and motion blur. A target detection model is used to locate the students' facial areas and limb skeletons, and consecutive frames are aligned.

[0063] Micro-expression analysis uses a facial key point detection algorithm to locate facial feature points and calculate dynamic parameters such as eye opening and closing, eyebrow height, and mouth corner offset.

[0064] By capturing the rate of change in micro-expressions using the temporal difference method, we can identify the transition patterns of emotions such as surprise, confusion, and relief.

[0065] Establish an expression feature vector library, compare real-time data with preset benchmark templates for focus, confusion and fatigue, and output the probability value of emotional state.

[0066] The Retinex algorithm is used to optimize the uneven illumination problem by decomposing the illumination and reflection components of the image to suppress illumination interference. For motion blur, a blind deconvolution algorithm is used in conjunction with a motion velocity statistical model in a student learning scenario to estimate the blur kernel and perform image restoration. A noise filtering step is added to the preprocessing process, using Gaussian filtering to remove image sensor noise, and grayscale normalization is used to map the image pixel values ​​to the [0, 255] range to ensure consistency in subsequent analysis.

[0067] The target detection uses the YOLOv8-nano model, which is lightweight and optimized for classroom scenes (model parameters ≤ 10M) to ensure real-time performance (detection frame rate ≥ 25fps). The accuracy of facial region localization is ≥ 98%. The limb skeleton localization uses the MediaPipe Pose model, extracting 33 skeletal key points with a coordinate localization error ≤ 2 pixels.

[0068] Facial key point detection uses the Dlib68 point model to enhance the annotation of feature points in key areas such as the palpebral fissure, brow peak, and corner of the mouth. The opening and closing degree of the palpebral fissure is calculated by the Euclidean distance between the key points of the eyelid and the lower eyelid in the vertical direction. The height of the brow peak is based on the vertical distance difference with the brow bone base point as the reference. The offset of the corner of the mouth is based on the horizontal distance difference with the midline of the face as the reference. The parameter calculation frequency is consistent with the video frame rate (25 times / second).

[0069] Limb analysis uses skeletal key point recognition technology to extract skeletal node coordinates and calculate parameters such as trunk tilt angle, arm range of motion, and relative distance between the hand and face.

[0070] By matching action sequence patterns, we can identify learning behaviors such as resting one's chin on one's hand in thought, frequently twirling a pen, and leaning forward.

[0071] Construct a physical activity index to quantify the frequency and amplitude of movements and distinguish between active interaction and passive slackness;

[0072] By employing an attention mechanism to integrate facial expressions and body features, the system strengthens the judgment weight for actively solving the problem when it detects the combination of frowning, leaning forward, and pointing to the screen; and increases the confidence level of cognitive fatigue for the feature combination of wandering eyes, leaning back, and low body activity.

[0073] By using a time-series sliding window, a comprehensive behavioral feature vector for that period is generated. A machine learning model is trained based on a labeled dataset to establish a mapping relationship between behavioral features and cognitive states.

[0074] When a person consistently displays highly confused facial expressions and frequently adjusts their body language, it is determined to be a comprehension disorder of the knowledge point.

[0075] When low concentration is accompanied by low physical activity exceeding a threshold, it is marked as a decline in learning interest.

[0076] By combining specific learning scenarios, cognitive states are transformed into quantifiable knowledge and ability indicators, generating knowledge and ability data that includes the degree of mastery of each knowledge point, the type of cognitive gaps, and the best absorption methods.

[0077] A random forest classification model was adopted with 120 decision trees and a maximum tree depth of 15. A mapping relationship was established. The model input was a 128-dimensional comprehensive behavioral feature vector, including 64-dimensional facial features and 64-dimensional body features. The output was a label of five cognitive states: knowledge point comprehension difficulties, learning interest decline, and normal learning.

[0078] The training dataset contains classroom behavior data of 1,000 students from different grade levels (one-third each in primary, junior high, and senior high schools). The sample size for each cognitive state is ≥5,000. The sample collection time covers the entire teaching cycle (40 minutes / class, 200 classes in total). The data annotation was cross-validated by three educational psychology experts, and the annotation consistency Kappa coefficient is ≥0.85.

[0079] The steps for building the student profile module are as follows:

[0080] Tag knowledge and skills data, establish a three-level indicator system of subject-module-knowledge point, and form a structured matrix;

[0081] By using horizontal comparison, vertical tracking, and correlation mining, we screened out knowledge points and weak clusters that were below average and had been stagnant for a long time, and determined the top 5 weak areas.

[0082] Expanding the three dimensions of cognitive style, learning strategy, and psychological state, and combining micro-expression and behavioral data to extract learning preferences, strategy efficiency, and psychological thresholds;

[0083] The weights of each dimension are dynamically adjusted according to the learning stage. Beginners focus on knowledge and cognitive style, intermediate learners strengthen strategies, and the final sprint focuses on weak connections.

[0084] The five-dimensional scoring, weak connection paths, and concrete tags are presented visually using radar charts and knowledge graphs.

[0085] By iterating through a dual-mode profile system of real-time micro-updates and periodic full updates, cross-period difference reports are generated, resulting in multi-dimensional student profiles.

[0086] The subject-specific ability dimension (weight 40%) includes knowledge mastery (weight 60%), application ability (weight 30%), and transfer ability (weight 10%). The knowledge module dimension (weight 35%) is divided by subject chapters, such as the algebra and geometry modules in mathematics. The knowledge point dimension (weight 25%) under each module is quantified from 0 to 100 points, where below 60 points is not mastered, 60-80 points is basically mastered, and above 80 points is proficient. The quantified value is calculated by weighting the knowledge point test score (60%) and the behavioral feature matching degree (40%).

[0087] When selecting weak areas, in addition to calculating the standard deviation of knowledge mastery within the module, a time decay coefficient is added (coefficient of knowledge points learned in the last week = 1.0, coefficient of knowledge points learned 1-2 weeks ago = 0.8, coefficient of knowledge points learned more than 2 weeks ago = 0.6) to correct the bias in judging knowledge points that have been stagnant for a long time and ensure that the top 5 weak areas are more in line with the current learning status.

[0088] The three-level indicator system includes subject ability, knowledge modules, and specific knowledge points. Each indicator is assigned a quantitative value to form a structured knowledge ability matrix. Weaknesses are identified by calculating the standard deviation of mastery of each knowledge point within the same knowledge module and selecting knowledge points that are 1.5 standard deviations below the module average.

[0089] Cognitive style is divided into three sub-dimensions: field-dependent / field-independent, impulsive / reflective, and visual / auditory / kinesthetic. It is determined by analyzing micro-expressions (e.g., visual learners' eye movement frequency ≥ 3 times / minute, auditory learners' head turning towards the sound source frequency ≥ 2 times / minute) and body behavior (kinesthetic learners' hand operation frequency ≥ 5 times / minute).

[0090] The learning strategy dimension includes retelling strategy, elaboration strategy, organization strategy, and resource management strategy. It is quantitatively scored (0-100 points) by monitoring students' note-taking frequency (≥3 times / 10 minutes indicates a high tendency towards retelling strategy), the number of questions asked (≥2 times / 20 minutes indicates a high tendency towards elaboration strategy), and knowledge point association annotation behavior (≥1 time / 15 minutes indicates a high tendency towards organization strategy).

[0091] The steps for building a knowledge graph construction module are as follows:

[0092] The knowledge system is broken down, the core areas of the discipline are sorted out, and a three-level structure of chapters, knowledge modules and specific knowledge points is broken down, clarifying the definition, scope and application scenarios of each knowledge point;

[0093] Establish prerequisite dependencies, draw a knowledge point association network through subject logic analysis and teaching experience annotation, rely on polynomial operations through the solution of quadratic equations, and use graph algorithms to verify the completeness and rationality of the dependency chain. Divide the difficulty levels, combine curriculum standards and learning data, divide the knowledge points into three levels of cognitive complexity: basic, intermediate and advanced, and assign a quantitative difficulty coefficient.

[0094] A dynamic calibration mechanism is introduced to collect student learning data in real time. If the error rate of a certain knowledge point continues to be higher than the threshold, the difficulty coefficient is increased and the prerequisite dependencies are optimized.

[0095] Generate a visual graph, with node size representing difficulty and line thickness reflecting dependency intensity, and support filtering and viewing by subject and module.

[0096] Real-time micro-updates are implemented, collecting the latest behavioral data every 30 minutes to dynamically adjust knowledge and ability indicators and cognitive status labels, with the adjustment range not exceeding 10% of the current score to avoid excessive fluctuations in the profile; a full-dimensional update is performed every Sunday at 24:00, integrating all learning data of the week (including classroom behavior, test scores, and homework completion), recalculating the weights and scores of each dimension, and generating a cross-cycle difference report (including the rate of change of weak areas and the core indicator of cognitive style matching degree changes between this week and last week).

[0097] When verifying the prerequisite dependencies of knowledge points, first count the actual frequency of the preset dependency paths between each knowledge point, and calculate the degree of agreement between this frequency and the theoretical derivation frequency. The higher the degree of agreement, the more reasonable the dependency relationship. At the same time, check whether there are isolated knowledge points or closed-loop dependencies to ensure that all knowledge points can be traced back to the basic starting point through a limited number of steps.

[0098] When classifying difficulty levels, the baseline for the basic level is determined based on the mastery requirements of knowledge points in the curriculum standards. Then, the average mastery time and error rate of the knowledge point in the learning data are analyzed. The higher the combined ratio of the two, the greater the upward adjustment of the difficulty level. In combination with cognitive complexity, the amount and closeness of the prerequisite knowledge required to understand the knowledge point are considered. The more prerequisite knowledge and the more complex the connection, the higher the difficulty coefficient.

[0099] The Depth-First Search (DFS) algorithm is used to traverse the knowledge point association network, check isolated knowledge points (nodes with no predecessor dependencies and no successor dependencies), and verify the existence of closed-loop dependencies through the topology sorting algorithm (if the topology sorting result contains all nodes, then there is no closed loop), to ensure the integrity of the dependency chain.

[0100] The dependency fit calculation is based on the theoretical frequency of the preset dependency path, which is based on the subject teaching syllabus (the theoretical frequency of "polynomial operation → solution of quadratic equation in one variable" is 100%). The actual frequency is obtained by analyzing 1,000 teacher teaching cases and 5,000 student learning path data. Fit = (actual frequency / theoretical frequency) × 100%. A fit ≥ 85% is considered a reasonable dependency, and a fit below 60% requires a re-examination of the dependency logic.

[0101] The formula for calculating the difficulty level (0-10 points, with 10 points being the highest difficulty) is as follows:

[0102] Difficulty level = Basic score (3 points for basic level, 6 points for intermediate level, 9 points for advanced level) + (Average mastery time deviation rate × 2) + (Error rate × 3);

[0103] Among them, the average mastery time deviation rate = (actual average mastery time of this knowledge point - standard mastery time of knowledge points at the same level) / standard mastery time of knowledge points at the same level;

[0104] Error rate = Number of incorrect answers to this knowledge point / Total number of answers;

[0105] The dynamic calibration threshold is set up so that if the error rate of a certain knowledge point is ≥40% for 7 consecutive days, or the average mastery time exceeds the standard time by 50%, the difficulty coefficient is increased (increase = 0.5 × error rate deviation rate, error rate deviation rate = (actual error rate - standard error rate) / standard error rate). At the same time, its prerequisite dependencies are re-verified. If a dependency is found to be missing, a new related node is added.

[0106] The steps for real-time optimization of the learning path in the inner-loop reinforcement learning of the recommendation algorithm module are as follows:

[0107] The state space is defined by using the student's current mastery of knowledge points, learning time, and real-time feedback as state parameters to construct a multi-dimensional state vector.

[0108] The multidimensional state vector includes: knowledge point mastery, cumulative learning time, real-time emotional positivity, accuracy rate of the last 5 quizzes, progress of the current learning unit, average mastery of prior knowledge points, interference of the learning environment, cognitive fatigue index, knowledge point relevance, quiz time deviation rate, note completion rate, question frequency, number of resource calls, strategy adaptability, and feedback response speed.

[0109] Reward function formula:

[0110] Instant reward value = (Mastery improvement rate × 0.5) + (1 / Learning time efficiency × 0.3) + (Emotional positivity × 0.2)

[0111] Among them, the rate of improvement of mastery = (current mastery - mastery of the previous unit) / learning time (minutes), and the efficiency of learning time = actual learning time / standard learning time (the standard time is set based on the difficulty coefficient of the knowledge point, such as a difficulty of 5 points corresponding to a standard time of 30 minutes).

[0112] The action space design allows for optional actions such as continuing to practice the current knowledge point, jumping to previous basic knowledge points, and switching question types and difficulties. Each action corresponds to a clear learning content adjustment plan.

[0113] The reward mechanism is designed with the rate of improvement in knowledge mastery as the core reward, and dynamically weighted by learning time efficiency and emotional positivity to form an instant reward value.

[0114] The strategy is iteratively optimized. After each learning unit is completed, the action selection probability is updated based on the reward value. Path adjustment strategies that improve the reward value are retained first. An exploration factor is introduced to avoid path solidification.

[0115] The initial value of the exploration factor is set to 0.3, and an exponential decay strategy is adopted for adjustment: the exploration factor decays by 0.02 for every 10 learning units completed, with a minimum decay to 0.05; when the learning path is detected to be solidified, the exploration factor is temporarily increased to 0.4, and the decay rhythm is restored after 2 learning units to ensure a balance between path diversity and optimization efficiency.

[0116] The steps in the recommendation algorithm module for introducing cognitive meta-strategies to generate heuristic learning branches in the outer loop are as follows:

[0117] Learning bottleneck identification involves continuously monitoring the duration of stagnation in mastery of knowledge points in the inner loop output and the repetition rate of similar errors. When any of these indicators exceeds a threshold, it is determined that a learning bottleneck has occurred.

[0118] Meta-strategy matching calls a preset cognitive meta-strategy library to match and adapt strategies based on the bottleneck type.

[0119] Cross-domain knowledge association: Based on the cross-disciplinary association network of knowledge graphs, it retrieves content from other disciplines that share common thinking with the current knowledge point and generates heuristic side content;

[0120] Branch integration and exit: Heuristic branches are inserted into the main learning path, a completion threshold is set for the branch, and the branch automatically returns to the main path after the threshold is reached. At the same time, the contribution of the branch to the bottleneck breakthrough is recorded.

[0121] The embedded antifragile recommendation mechanism first assesses the student's resilience baseline, defines the challenge boundary, and builds a three-level challenge content pool of light, medium and heavy, based on 10%-35% beyond the current ability. Regular and challenging content are pushed alternately, and the level is dynamically adjusted based on feedback. Breakthrough points are marked to strengthen transfer, and the challenge is re-imposed after addressing weaknesses. The baseline is updated weekly, and learning paths are recommended based on the student's own knowledge mastery.

[0122] The knowledge graph visualization uses the Neo4j graph database to store knowledge graph data. The front end uses the ECharts Graph component to display the visualization, supporting zoom (zoom range 10%-200%), panning, and node dragging interaction. The node size is positively correlated with the difficulty coefficient (a coefficient of 10 points corresponds to a node diameter of 20px, and a coefficient of 3 points corresponds to 10px). The thickness of the connection is positively correlated with the dependency strength (dependency strength = fit × 10, connection width = dependency strength / 10px).

[0123] The filtering function supports multi-dimensional filtering by subject (mathematics, Chinese), module (algebra, geometry), and difficulty level (basic / intermediate / advanced). The filtering results are updated in real time with a graph display, and a pop-up window with details of knowledge points is provided, including definitions, application scenarios, and links to typical examples.

[0124] The path output and adjustment module uses a dual-view display of knowledge graph and progress dashboard. When a student makes three consecutive mistakes on a certain knowledge point, a preliminary basic review branch is automatically inserted. If the student completes the goal ahead of schedule and the accuracy rate exceeds 90%, more difficult extension content is pushed.

[0125] The knowledge graph view is a visualization of the knowledge graph building module, overlaid with learning path markers, and allows users to click on nodes to view the learning plan for that knowledge point.

[0126] The daily / weekly learning progress is displayed in the form of a Gantt chart. The horizontal axis represents time, the vertical axis represents the name of the knowledge point, and the color of the progress bar corresponds to the mastery level. The progress bar can be dragged to adjust the learning plan, and the system automatically checks the compatibility of the adjusted plan with the knowledge graph dependencies.

[0127] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A personalized learning path recommendation system for online school students, characterized in that, Recommendation systems include: The data acquisition module uses a camera to capture students' micro-expressions and body language details while they are learning, collects students' learning behavior data, and uses image analysis and body language analysis to determine students' knowledge and ability data. The student profile building module uses students' knowledge and ability data to identify the current areas of weakness in students' knowledge and constructs a multi-dimensional student profile. The knowledge graph construction module builds a relational graph of the subject knowledge system, including the prerequisite dependencies and difficulty levels between knowledge points; The recommendation algorithm module is connected to the student profile construction module and the knowledge graph construction module. Based on the multi-model fusion algorithm, the learning path is optimized in real time through inner-loop reinforcement learning, and the outer loop introduces cognitive meta-strategies to generate heuristic learning branches. An antifragile recommendation mechanism is embedded to generate adaptive personalized learning paths. The path output and adjustment module is used to display recommended paths and make dynamic adjustments.

2. The personalized learning path recommendation system for online school students according to claim 1, characterized in that: Cameras are deployed in the classroom to capture a panoramic view of the classroom in real time. The acquired data is preprocessed, and image enhancement algorithms are used to optimize the problems of uneven lighting and motion blur. A target detection model was used to locate the student's facial region and limb skeleton, and the consecutive frames were aligned. Micro-expression analysis uses a facial key point detection algorithm to locate facial feature points and calculate dynamic parameters such as eye opening and closing, eyebrow height, and mouth corner offset. By capturing the rate of change in micro-expressions using the temporal difference method, we can identify the transition patterns of emotions such as surprise, confusion, and relief. Establish an expression feature vector library, compare real-time data with preset benchmark templates for focus, confusion and fatigue, and output the probability value of emotional state.

3. The personalized learning path recommendation system for online school students according to claim 1, characterized in that: Limb analysis uses skeletal key point recognition technology to extract skeletal node coordinates and calculate parameters such as trunk tilt angle, arm range of motion, and relative distance between the hand and face. By matching action sequence patterns, we can identify learning behaviors such as resting one's chin on one's hand in thought, frequently twirling a pen, and leaning forward. Construct a physical activity index to quantify the frequency and amplitude of movements and distinguish between active interaction and passive slackness; By employing an attention mechanism to integrate facial expressions and body features, the system strengthens the judgment weight for actively solving the problem when it detects the combination of frowning, leaning forward, and pointing to the screen; and increases the confidence level of cognitive fatigue for the feature combination of wandering eyes, leaning back, and low body activity. By using a time-series sliding window, a comprehensive behavioral feature vector for that period is generated. A machine learning model is trained based on a labeled dataset to establish a mapping relationship between behavioral features and cognitive states. When a person consistently displays highly confused facial expressions and frequently adjusts their body language, it is determined to be a comprehension disorder of the knowledge point. When low concentration is accompanied by low physical activity exceeding a threshold, it is marked as a decline in learning interest. By combining specific learning scenarios, cognitive states are transformed into quantifiable knowledge and ability indicators, generating knowledge and ability data that includes the degree of mastery of each knowledge point, the type of cognitive gaps, and the best absorption methods.

4. The personalized learning path recommendation system for online school students according to claim 1, characterized in that, The steps for building the student profile module are as follows: Tag knowledge and skills data, establish a three-level indicator system of subject-module-knowledge point, and form a structured matrix; By using horizontal comparison, vertical tracking, and correlation mining, we screened out knowledge points and weak clusters that were below average and had been stagnant for a long time, and determined the top 5 weak areas. Expanding the three dimensions of cognitive style, learning strategy, and psychological state, and combining micro-expression and behavioral data to extract learning preferences, strategy efficiency, and psychological thresholds; The weights of each dimension are dynamically adjusted according to the learning stage. Beginners focus on knowledge and cognitive style, intermediate learners strengthen strategies, and the final sprint focuses on weak connections. The five dimensions of scoring, weak connection paths, and concrete tags are presented visually using radar charts and knowledge graphs. By iterating through a dual-mode profile system of real-time micro-updates and periodic full updates, cross-period difference reports are generated, resulting in multi-dimensional student profiles.

5. The personalized learning path recommendation system for online school students according to claim 4, characterized in that: The three-level indicator system includes subject ability, knowledge modules, and specific knowledge points. Each indicator is assigned a quantitative value to form a structured knowledge ability matrix. Weaknesses are identified by calculating the standard deviation of mastery of each knowledge point within the same knowledge module and selecting knowledge points that are 1.5 standard deviations below the module average.

6. The personalized learning path recommendation system for online school students according to claim 1, characterized in that, The steps for building a knowledge graph construction module are as follows: The knowledge system is broken down, the core areas of the discipline are sorted out, and a three-level structure of chapters, knowledge modules and specific knowledge points is broken down, clarifying the definition, scope and application scenarios of each knowledge point; Establish prerequisite dependencies, draw a knowledge point association network through subject logic analysis and teaching experience annotation, rely on polynomial operations through the solution of quadratic equations, and use graph algorithms to verify the completeness and rationality of the dependency chain. Divide the difficulty levels, combine curriculum standards and learning data, divide the knowledge points into three levels of cognitive complexity: basic, intermediate and advanced, and assign a quantitative difficulty coefficient. A dynamic calibration mechanism is introduced to collect student learning data in real time. If the error rate of a certain knowledge point continues to be higher than the threshold, the difficulty coefficient is increased and the prerequisite dependencies are optimized. Generate a visual graph, with node size representing difficulty and line thickness reflecting dependency intensity, and support filtering and viewing by subject and module.

7. The personalized learning path recommendation system for online school students according to claim 6, characterized in that: When verifying the prerequisite dependencies of knowledge points, first count the actual frequency of the preset dependency paths between each knowledge point, and calculate the degree of agreement between this frequency and the theoretical derivation frequency. The higher the degree of agreement, the more reasonable the dependency relationship. At the same time, check whether there are isolated knowledge points or closed-loop dependencies to ensure that all knowledge points can be traced back to the basic starting point through a limited number of steps. When dividing the difficulty levels, the baseline for the basic level is determined based on the mastery requirements of the knowledge points in the curriculum standards. Further analysis of the learning data reveals that the average mastery time and error rate for this knowledge point are the highest combined proportions, and the greater the upward adjustment of the difficulty level. Taking into account cognitive complexity, the amount and relevance of the prerequisite knowledge required to understand the knowledge point are considered. The more prerequisite knowledge there is and the more complex the relationship, the higher the difficulty level.

8. The personalized learning path recommendation system for online school students according to claim 1, characterized in that, The steps for real-time optimization of the learning path in the inner-loop reinforcement learning of the recommendation algorithm module are as follows: The state space is defined by using the student's current mastery of knowledge points, learning time, and real-time feedback as state parameters to construct a multi-dimensional state vector. The action space design allows for optional actions such as continuing to practice the current knowledge point, jumping to previous basic knowledge points, and switching question types and difficulties. Each action corresponds to a clear learning content adjustment plan. The reward mechanism is designed with the rate of improvement in knowledge mastery as the core reward, and dynamically weighted by learning time efficiency and emotional positivity to form an instant reward value. The strategy is iteratively optimized. After each learning unit is completed, the action selection probability is updated based on the reward value. Path adjustment strategies that improve the reward value are retained first. An exploration factor is introduced to avoid path solidification.

9. The personalized learning path recommendation system for online school students according to claim 1, characterized in that, The steps in the recommendation algorithm module for introducing cognitive meta-policies to generate heuristic learning branches in the outer loop are as follows: Learning bottleneck identification involves continuously monitoring the duration of stagnation in mastery of knowledge points in the inner loop output and the repetition rate of similar errors. When any of these indicators exceeds a threshold, it is determined that a learning bottleneck has occurred. Meta-strategy matching calls a preset cognitive meta-strategy library to match and adapt strategies based on the bottleneck type. Cross-domain knowledge association: Based on the cross-disciplinary association network of knowledge graphs, it retrieves content from other disciplines that share common thinking with the current knowledge point and generates heuristic side content; Branch integration and exit: Heuristic branches are inserted into the main learning path, a completion threshold is set for the branch, and the branch automatically returns to the main path after the threshold is reached. At the same time, the contribution of the branch to the bottleneck breakthrough is recorded. The embedded antifragile recommendation mechanism first assesses the student's resilience baseline, defines the challenge boundary, and builds a three-level challenge content pool of light, medium and heavy, based on 10%-35% beyond the current ability. Regular and challenging content are pushed alternately, and the level is dynamically adjusted based on feedback. Breakthrough points are marked to strengthen transfer, and the challenge is re-imposed after addressing weaknesses. The baseline is updated weekly, and learning paths are recommended based on the student's own knowledge mastery.

10. The personalized learning path recommendation system for online school students according to claim 1, characterized in that: The path output and adjustment module uses a dual-view display of knowledge graph and progress dashboard. When a student makes three consecutive mistakes on a certain knowledge point, a preliminary basic review branch is automatically inserted. If the student completes the goal ahead of schedule and the accuracy rate exceeds 90%, more difficult extension content is pushed.

Images