Multi-modal dynamic optimization educational resource recommendation system and method

Abstract

The invention relates to a multi-modal dynamic optimization educational resource recommendation system and method, and the system comprises the following modules: a multi-modal data collection module integrates video behaviors, answer tracks, physiological signals and other data through edge calculation, and constructs a learning feature map; the student portrait module adopts an LSTM-Attention network in combination with a graph neural network to dynamically model knowledge mastery and learning styles; the resource matching engine realizes multi-objective optimization of knowledge gain, cognitive load and interest matching based on reinforcement learning and knowledge graph analysis; the tag adaptive module dynamically adjusts resource weights through causal inference and comparative learning, the personalized recommendation module generates a dynamic learning path and pushes adaptive resources based on student portraits and real-time behavior data, and the learning progress tracking module monitors a learning state in real time and feeds back the learning state to the resource matching engine to optimize a recommendation strategy in a closed loop mode. The technical defects that resource recommendation of a traditional education platform is rigid and personalized adaptation is lacked are overcome.

Description

Technical Field

[0001] The present invention belongs to the technical field of artificial intelligence education, and specifically relates to an educational resource recommendation system and method with multi-modal dynamic optimization. Background Art

[0002] The educational platform resource recommendation system is an intelligent system designed to accurately match the most suitable learning resources for students by analyzing their learning behaviors and characteristics. Its core lies in using advanced artificial intelligence technologies and algorithms to achieve a deep understanding of students' learning needs and personalized satisfaction. This system not only focuses on students' learning progress and abilities but also fully considers their interest preferences and learning styles, thereby providing students with a learning path and resource recommendation that better suits their individual characteristics. By continuously tracking students' learning status and adjusting the recommendation strategy, the educational platform resource recommendation system can ensure that students always maintain an efficient and coherent learning state, thereby improving learning effects and learning satisfaction.

[0003] Traditional recommendation systems have the following defects when in use:

[0004] (1) Static label dependence: The label system is rigid and cannot dynamically adapt to changes in students' knowledge states;

[0005] (2) Single recommendation strategy: Only relying on the collaborative filtering algorithm (see the paper "Learning Resource Recommendation Based on Collaborative Filtering" in the Journal of Educational Technology in 2020), ignoring the balance between cognitive load and interest;

[0006] (3) Lack of effectiveness evaluation: Lack of quantitative evaluation and feedback loop for the actual teaching effects of resources;

[0007] (4) Limited data dimension: Relying on single behavioral data (such as answering records) and not integrating multi-modal learning features (such as video viewing behavior, physiological signals). Summary of the Invention

[0008] The purpose of the present invention is to provide an educational resource recommendation system and method with multi-modal dynamic optimization to solve the problems raised in the above background art.

[0009] To achieve the above purpose, the present invention provides the following technical solutions:

[0010] A multi-modal dynamic optimization educational resource recommendation system, which is characterized by including the following core modules:

[0011] Multi-modal data acquisition module: Integrate video viewing behavior, question answering trajectories, interaction preferences, and physiological signals through edge computing and spatio-temporal alignment technologies to construct a learning feature map;

[0012] Multi-modal Data Acquisition Module: Integrate video viewing behavior, question answering trajectory, interaction preferences, and physiological signals through edge computing and spatio-temporal alignment technology to construct a learning feature map;

[0013] Student Portrait Construction Module: Adopt an LSTM-Attention network combined with a temporal convolutional network to dynamically model the knowledge forgetting curve and spatio-temporal behavior features, and calculate the mastery degree of knowledge points through the graph neural network propagation algorithm;

[0014] Resource Dynamic Matching Engine: Based on reinforcement learning and knowledge graph analysis, achieve multi-objective optimization of knowledge gain, cognitive load, and interest matching, and generate personalized learning paths;

[0015] Label Adaptive Update Module: Dynamically adjust the resource label weights through causal inference and contrast learning techniques, and update the label system in combination with the decay factor and the knowledge graph propagation mechanism;

[0016] Personalized Recommendation Module: Based on the student portrait and real-time behavior data, generate a dynamic learning path and push adapted resources, specifically including: generating a phased learning plan by combining knowledge gaps, interest preferences, and cognitive load; monitoring the resource completion rate and the mid-way jump-out rate, and dynamically adjusting the recommendation order; adapting to the PC / mobile interface layout and interaction logic;

[0017] Learning Progress Tracking Module: Real-time monitor the learning status and feedback it to the resource matching engine to optimize the recommendation strategy in a closed loop.

[0018] The multi-modal data acquisition module is signal-connected to the student portrait construction module, the student portrait construction module is signal-connected to the resource dynamic matching engine, the resource dynamic matching engine is signal-connected to the label adaptive update module, the label adaptive update module is signal-connected to the personalized recommendation module, and the personalized recommendation module is signal-connected to the learning progress update module.

[0019] The multi-modal data acquisition module further includes:

[0020] Video Viewing Behavior Analysis Unit: Calculate the concentration index, where the concentration index = effective viewing duration / total duration × interference times correction coefficient;

[0021] Question Answering Trajectory Analysis Unit: Mark the difficult points and record the answering path;

[0022] Interaction Preference Processing Unit: Construct user interaction preference labels to provide a personalized decision basis for resource dynamic matching;

[0023] Physiological Signal Processing Unit: Evaluate the cognitive load through heart rate variability analysis;

[0024] Data preprocessing unit: Perform standardized processing of timestamp alignment, data format unification, and outlier filtering, classify and store behavior logs, knowledge point data, and resource metadata, perform multi-modal feature fusion on spatio-temporal alignment and semantic level mapping, and the quality verification must be complete, consistent, and compliant.

[0025] The student portrait construction module further includes:

[0026] Dynamic update mechanism: Retain the portraits of the last 10 historical versions, support version comparison and rollback;

[0027] Knowledge point mastery evaluation model: A dual-channel GRU network processes the sequence of correct rates and the correlation of wrong questions, combines the Bayesian knowledge tracing algorithm to calculate the probability of knowledge point mastery, and outputs the probability value P(mastery t ) = P(mastery t-1 ) + (1 - P(mastery t-1 )) · γ · correct t ;

[0028] Where γ is the learning rate parameter, correct t is the correct question flag; t is the current time;

[0029] Learning style classification model: Through principal component analysis (PCA) for dimensionality reduction and K-means clustering, it is divided into visual, auditory, and practical types, and the corresponding resource strategies are matched;

[0030] Portrait credibility verification mechanism: The teacher gives an artificial score, or uses the cross-validation method or A / B test to compare the portrait effects. If the teacher's artificial score ≥ 85 points, it passes. The cross-validation method compares the difference between the predicted score and the actual exam score of the portrait. The A / B test is to compare the effects of different portrait versions.

[0031] The resource dynamic matching engine further includes:

[0032] Multi-dimensional scoring model: A weighted comprehensive score based on a knowledge gain weight coefficient of 0.45, a cognitive load weight coefficient of 0.30, and an interest matching weight coefficient of 0.25;

[0033] Knowledge graph enhanced matching strategy: Construct a knowledge graph of knowledge points, resources, learning objectives, and relevance. The initial value of the edge weight is calculated by the TF-IDF algorithm and dynamically updated through collaborative filtering;

[0034] Three - layer associated retrieval mechanism: Screen resources within the three - layer association range of the user's knowledge gap nodes; the three layers are specifically: the direct association layer, the indirect association layer, and the interdisciplinary expansion layer; the core threshold range of the direct association layer is 0.5 ≤ weight ≤ 1.0, and wrong questions are used for dynamic adjustment. When the number of repetitions of wrong questions ≥ 3, the weight drops to 0.4 to trigger the associated retrieval of knowledge points; the core threshold range of the indirect association layer is 0.3 ≤ weight < 0.5. When students prefer video learning, the video weight is multiplied by 1.2 to increase the priority, and graphical explanation videos are dynamically appended; the core threshold of the interdisciplinary expansion layer is 0.25 ≤ weight < 0.4, and an interest - driven threshold adjustment mechanism is adopted. When the student's interest score > 70, the lower limit of the threshold is adjusted to 0.2;

[0035] Exclude nodes with a core threshold of each association layer < 0.6; Adopt the priority and each association layer in the order of: direct association layer > indirect association layer > interdisciplinary expansion layer. For example, "polynomial operation" with a weight of 0.5 is preferentially processed according to the direct association layer, calculate the optimal learning path, and preferentially push resources associated with weak knowledge points;

[0036] Reinforcement learning dynamic adjustment strategy: The state space is the portrait feature vector and the environmental state, the action space is the adjustment of the resource sorting weight, and the reward function R = αΔK + β(1 - F)+γI; where, ΔK = knowledge gain, F = fatigue degree, I = interest matching; α represents the weight of the knowledge gain ΔK, and the typical value is in 0.4 ≤ α ≤ 0.6, which is dynamically adjusted through AB testing; β represents the weight of the fatigue degree F, and the typical value is in 0.2 ≤ β ≤ 0.3, γ represents the weight of the interest matching degree I, and the typical value is in 0.1 ≤ γ ≤ 0.2;

[0037] Cognitive load monitoring indicators: Attention dispersion, learning efficiency ratio, and physiological signals;

[0038] Attention dispersion = standard deviation of the number of operations per unit time / average number of operations per unit time; The standard deviation of the number of operations per unit time and the average number of operations per unit time are obtained from the click - stream event log. Set the warning threshold. When the attention dispersion > 1.0 times the mean value, it is mild dispersion; > 1.5 times the mean value, it is moderate dispersion; > 2.5 times the mean value, it is severe dispersion and intervention is required;

[0039] Learning efficiency ratio = effective learning time / total learning duration × 100%, and the determination standard of effective learning time: continuous concentration duration ≥ 5 minutes and no ineffective operations; Set the warning threshold. When the learning efficiency ratio < 40%, the reason needs to be diagnosed;

[0040] Physiological signals are detected by abnormal heart rate variability. Set the warning threshold. When SDNN < 50ms, the reason needs to be diagnosed;

[0041] Pareto Frontier Screening Mechanism: Retain the non-dominated solution set, control resource diversity, with an interval of ≥ 3 items of the same type, and a new resource exposure rate of ≥ 15%;

[0042] Time Period Adaptation Strategy: Match resource combinations of different difficulties and types according to morning / afternoon / evening time periods.

[0043] The label adaptive update module further includes:

[0044] Causal Effect Calculation Unit: Calculate the net impact of resources on performance improvement through a counterfactual reasoning framework. An average treatment effect ATE ≥ 0.15 and a significance test p-value < 0.05 are considered valid;

[0045] Semantic Space Reconstruction Unit: Optimize vector representations through contrastive learning, and fully refresh label weights every 24 hours;

[0046] Label Propagation Mechanism: Based on knowledge graph propagation, locate the core knowledge point nodes of resources, and propagate label weights w along the graph edges chlid = w parent × δ depth ;

[0047] where δ is the attenuation factor, depth is the propagation level, and w parent is the initial weight of the label;

[0048] Dual Verification System of Algorithm Self-Check and Manual Sampling Inspection: The system algorithm conducts real-time self-check, with a confidence level ≥ 80%, and manual sampling inspection is carried out weekly, with a sampling accuracy rate ≥ 90%;

[0049] Abnormal Label Isolation Rule: Automatically isolate when the evaluation contribution degree < 0.05 for 3 consecutive times, the association path with the core knowledge point ≥ 3 levels, or the number of user complaints > 5 times;

[0050] Version Rollback Strategy: Load historical version difference analysis and repair when the label causes the click-through rate to drop > 15% or the association with key knowledge points is lost.

[0051] The personalized recommendation module The personalized recommendation module further includes:

[0052] Path generation algorithm: Adopt the hierarchical progressive method, dynamic pruning method, interest weighting method, and reinforcement learning to explore and match resources in the zone of proximal development, prioritize pushing content related to weak knowledge points, use topological sorting of directed acyclic graph (DAG) to ensure that prerequisite knowledge points are learned before subsequent knowledge points, and generate personalized recommendation sequences; in the hierarchical progressive method, when the weight > 0.5, first supplement the content in the direct association layer, and then expand to the indirect association layer; use a logistic regression model based on student behavior data to predict the effectiveness of resources, and implement the dynamic pruning method to filter exercises with a historical correct rate > 90% in real time; introduce the interest weighting method to increase the weight of resources in the interdisciplinary expansion layer according to subject interests, where the interest score = historical click-through rate × 0.6 + average dwell time × 0.4; the average dwell time is calculated in minutes; use reinforcement learning to explore inserting 10% interdisciplinary challenge questions for ZPD boundary knowledge points with 0.6 ≤ student mastery probability < 0.7 to test transfer ability, specifically optimize the path at the ZPD boundary through the Q-Learning model and the reward function, and the reward function = mastery improvement × 0.7 + interest matching × 0.3;

[0053] Emergency switching mechanism: When it detects that a resource skip occurs continuously 3 times, it automatically triggers the manual intervention process and generates alternative solutions.

[0054] The learning progress tracking module continuously tracks the student's learning progress and adjusts the recommendation strategy in a timely manner to ensure that the student always progresses along the most suitable learning path.

[0055] A method for a multi-modal dynamic optimization educational resource recommendation system, characterized by including the following steps:

[0056] Step A: Multi-modal data collection and preprocessing, integrating video viewing behavior, question answering trajectories, interaction preferences, and physiological signals through edge computing and spatio-temporal alignment technologies to construct a learning feature map;

[0057] Step B: Adopt an LSTM-Attention network combined with a temporal convolutional network to dynamically model the knowledge forgetting curve and behavioral spatio-temporal characteristics of the student portrait, and calculate the knowledge mastery degree through the graph neural network propagation algorithm;

[0058] Step C: Dynamic resource matching, realizing multi-objective optimization of knowledge gain, cognitive load, and interest matching through reinforcement learning and knowledge graph analysis, and generating a personalized learning path;

[0059] Step D: Label adaptive update, dynamically adjust the resource label weights through causal inference and contrast learning technologies, and update the label system in combination with the decay factor and the knowledge graph propagation mechanism;

[0060] Step E: Personalized Recommendation. Based on the student profile and real-time behavior data, generate a dynamic learning path and push appropriate resources. Specifically, it includes: generating a phased learning plan by combining knowledge gaps, interest preferences, and cognitive load; monitoring the resource completion rate and mid-course dropout rate, and dynamically adjusting the recommendation order; adapting the PC / mobile interface layout and interaction logic.

[0061] Step F: Learning Progress Tracking. Real-time monitor the learning status and feedback it to the resource matching engine to optimize the recommendation strategy in a closed loop.

[0062] The multi-modal data preprocessing in Step A includes: performing standardized processing such as timestamp alignment, data format unification, and outlier filtering, with a timestamp alignment error < 50ms; classifying and storing behavior logs, knowledge point data, and resource metadata, and performing multi-modal feature fusion on spatio-temporal alignment and semantic level mapping; and the quality verification must be complete, consistent, and compliant.

[0063] Resource Efficiency Evaluation. Calculate and evaluate according to the direct contribution degree, indirect correlation degree, and group adaptability. The total resource efficiency score = direct contribution degree × 0.5 + indirect correlation degree × 0.3 + group adaptability × 0.2. If the total resource efficiency score ≥ 0.8, it is an excellent resource; if the total resource efficiency score is between 0.6 - 0.79, it is a good resource; if the total resource efficiency score < 0.6, it is a resource that needs to be optimized.

[0064] Optimize the resource library according to the total resource efficiency score, eliminate resources with a total resource efficiency score < 0.4, and recommend resources with a total resource efficiency score ≥ 0.6; perform personalized placement according to the total resource efficiency score: resources with a total resource efficiency score > 0.8 are resources with high group adaptability, automatically match the corresponding student labels, and resources with a total resource efficiency score < 0.4 are resources with low indirect correlation, only triggered when the direct correlation layer is weak.

[0065] Label Dynamic Adjustment. The adjustment strategy strengthens effective labels and weakens ineffective labels. The formula for strengthening a label is: The formula for weakening a label is: w new is the label value after strengthening or weakening, w old is the initial label value; S base is the benchmark effect value, taking the average label value of similar resources in the historical period, e is the natural constant, ΔS is the effect improvement amplitude, λ is the attenuation coefficient, T un is the ineffective duration, and the label types are knowledge dimension, ability dimension, and form dimension.

[0066] Map the resource usage records to knowledge points and quantify the effect indicators. The specific steps are as follows:

[0067] The binding of resources to knowledge points means that each resource of videos, exercises, or experiments is labeled with a knowledge point ID, and the mapping between resource usage records and knowledge points is completed through manual annotation by the staff of the teaching and research team.

[0068] The quantification of effectiveness indicators is divided into short-term effectiveness quantification within 0 - 7 days, medium-term effectiveness quantification within 1 - 3 months, and long-term effectiveness quantification at the semester level.

[0069] The short-term effectiveness quantification indicators within 0 - 7 days include resource completion rate, interaction depth, and immediate test scores.

[0070] Resource completion rate = (number of completions / number of recommendations) × 100%. If the number of mid-course exits is ≥ 3, then the total number of completions is multiplied by a coefficient of 0.7. The number of completions and number of recommendations are obtained from the resource open / close event logs. A high completion rate indicates users' recognition of the content, and a low completion rate requires investigation of resource difficulty or recommendation strategies.

[0071] Interaction depth = Σ(weight of single behavior × number of occurrences). Example of weights: play / pause = 0.3, redo questions = 0.7, note annotation = 1.0. The weight of single behavior and number of occurrences are obtained from clickstream data or button event tracking. A high proportion of high-weight behaviors indicates deep learning rather than passive browsing.

[0072] Immediate test score = ((current score - class average) / class standard deviation) × 10 + 50. The current score, class average, and class standard deviation are obtained from exercise submission records and class statistical databases. A score > 70 indicates exceeding 70% of classmates, and < 30 requires priority intervention.

[0073] The medium-term effectiveness quantification indicators within 1 - 3 months include the improvement rate of unit tests, the repetition rate of wrong questions, and the knowledge consolidation index.

[0074] Improvement rate of unit tests = ((current test score - previous test score) / previous test total score) × 100%, only associated with tests within the same knowledge cluster. An improvement rate < 20% indicates that the teaching plan needs to be optimized, and > 50% can be used as an excellent case for analysis.

[0075] Repetition rate of wrong questions = (number of repeated wrong knowledge points / total number of wrong knowledge points) × 100%. Wrong questions need to satisfy the condition that the interval between wrong questions of the same knowledge point ≤ 14 days. A repetition rate > 40% indicates that a personalized error correction plan needs to be launched, and < 10% indicates that weak points have been effectively broken through.

[0076] Knowledge consolidation index = probability of mastery × e^(-0.1 × number of interval days), where e is the natural constant, number of interval days = number of days from the last learning to the test day. The probability of mastery and number of interval days are obtained from the output of the Bayesian Knowledge Tracing (BKT) model and learning behavior timestamps. A knowledge consolidation index < 0.3 indicates the need for re-learning, and > 0.7 indicates that the knowledge point has been stably mastered.

[0077] The semester-level long-term effect quantification indicators include semester grade trends, knowledge transfer ability, and learning resilience index;

[0078] The semester grade trend is represented by the slope of linear regression. A slope ≥ 0.2 indicates a significant improvement, and ≤ -0.1 indicates a decline. Dimension analysis is carried out by aggregating according to knowledge point clusters. A negative trend requires warning of subject weaknesses, and a positive trend is used for teaching quality evaluation;

[0079] The knowledge transfer ability is represented by the cross-disciplinary score ratio. The cross-disciplinary score ratio = (cross-disciplinary application score / score of this subject) × 100%. The cross-disciplinary associations need to be marked. A ratio > 60% indicates outstanding high-order thinking ability, and < 30% requires strengthening interdisciplinary integration teaching;

[0080] The learning resilience index is represented by the frustration resistance. The frustration resistance = Σ(number of wrong questions corrected within 48 hours) ÷ total number of wrong questions × 100%. Record the time difference from the first error to the final correct answer. An index > 50% reflects strong frustration resistance, and < 20% requires attention to learning motivation and psychological intervention.

[0081] Compared with the prior art, the beneficial effects of the present invention are:

[0082] The system of the present invention continuously adjusts the label weights of knowledge associations, difficulty levels, and form preferences according to students' learning feedback and resource usage effects through the label adaptive update module to continuously optimize the recommendation effect. The present invention solves the technical defects of rigid resource recommendation and lack of personalized adaptation in traditional education platforms. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] Figure 1 is a schematic diagram of the system of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0084] Next, in combination with the embodiments of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0085] Please refer to Figure 1 , the present invention provides a multimodal dynamic optimization educational resource recommendation system. A multimodal dynamic optimization educational resource recommendation system is characterized by including the following core modules:

[0086] Multimodal data acquisition module: Integrate video viewing behavior, question answering trajectories, interaction preferences, and physiological signals through edge computing and spatio-temporal alignment technologies to construct a learning feature map;

[0087] Student Portrait Construction Module: It uses an LSTM-Attention network combined with a temporal convolutional network to dynamically model the knowledge forgetting curve and behavioral spatio-temporal features, and calculates the mastery degree of knowledge points through the graph neural network propagation algorithm;

[0088] Resource Dynamic Matching Engine: Based on reinforcement learning and knowledge graph analysis, it realizes multi-objective optimization of knowledge gain, cognitive load, and interest matching, and generates personalized learning paths;

[0089] Label Adaptive Update Module: It dynamically adjusts the resource label weights through causal inference and contrast learning techniques, and updates the label system by combining the decay factor and the knowledge graph propagation mechanism;

[0090] Personalized Recommendation Module: Based on the student portrait and real-time behavior data, it generates dynamic learning paths and pushes adapted resources. Specifically, it includes: generating a phased learning plan by combining knowledge gaps, interest preferences, and cognitive load; monitoring the resource completion rate and the mid-way dropout rate, and dynamically adjusting the recommendation order; adapting the PC / mobile interface layout and interaction logic;

[0091] Learning Progress Tracking Module: It monitors the learning status in real time and feeds it back to the resource matching engine to optimize the recommendation strategy in a closed loop.

[0092] The multi-modal data acquisition module is signal-connected to the student portrait construction module, the student portrait construction module is signal-connected to the resource dynamic matching engine, the resource dynamic matching engine is signal-connected to the label adaptive update module, the label adaptive update module is signal-connected to the personalized recommendation module, and the personalized recommendation module is signal-connected to the learning progress update module.

[0093] The multi-modal data acquisition module further includes:

[0094] Video Watching Behavior Analysis Unit: It calculates the concentration index, and the concentration index = effective watching duration / total duration × interference times correction coefficient;

[0095] Question Answering Trajectory Analysis Unit: It marks the difficult points and records the answering path;

[0096] Interaction Preference Processing Unit: It constructs user interaction preference labels to provide a personalized decision basis for resource dynamic matching;

[0097] Physiological Signal Processing Unit: It evaluates the cognitive load through heart rate variability analysis;

[0098] Data Preprocessing Unit: It performs standardized processing such as timestamp alignment, data format unification, and outlier filtering, classifies and stores behavioral logs, knowledge point data, and resource metadata, performs multi-modal feature fusion on spatio-temporal alignment and semantic level mapping, and the quality verification must be complete, consistent, and compliant.

[0099] The standard deviation of heart rate variability (SDNN) is the core indicator for evaluating cognitive load. The lower the SDNN value (unit: ms), the greater the stress on the autonomic nervous system and the higher the cognitive load. The time window is dynamically adjusted according to the learning stage (60 seconds for video learning / 30 seconds for practice) to achieve data alignment. Two types of indicators are converted into comparable numerical values in the range of [0,1] through Z-score standardization and interval mapping. Dynamic weights are assigned according to the learning scenario (video weight = 0.7 / practice weight = 0.5) for weighted fusion. A three-level intervention mechanism is triggered based on the fusion score. When the dynamic weight is above 0.8, the difficulty is increased, and when the dynamic weight is below 0.5, stepped intervention is carried out. Through the three-dimensional matching mechanism of time-space-scene, this scheme realizes the collaborative monitoring of attention and cognitive load, and improves the state recognition accuracy by 21.4%.

[0100] The student portrait construction module further includes:

[0101] Dynamic update mechanism: Retain the portraits of the most recent 10 historical versions, supporting version comparison and rollback;

[0102] Knowledge point mastery evaluation model: The two-channel GRU network processes the time series of correct rates and the correlation of wrong questions, and combines the Bayesian knowledge tracing algorithm to calculate the probability of mastering knowledge points, and outputs the probability value P(mastery t ) = P(mastery t-1 ) + (1 - P(mastery t-1 )) · γ · correct t ;

[0103] where γ is the learning rate parameter, and correct t is the correct question flag; t is the current time;

[0104] Learning style classification model: Through principal component analysis (PCA) for dimensionality reduction and K-means clustering, it is divided into visual, auditory, and practical types, and corresponding resource strategies are matched;

[0105] Portrait credibility verification mechanism: The teacher's manual scoring, cross-validation method, or A / B test is used to compare the portrait effects. The teacher's manual scoring ≥ 85 points is considered passed. The cross-validation method compares the difference between the predicted score and the actual exam score of the portrait. The A / B test is to compare the effects of different portrait versions.

[0106] The resource dynamic matching engine further includes:

[0107] Multi-dimensional scoring model: Weighted comprehensive scoring based on a knowledge gain weight coefficient of 0.45, a cognitive load weight coefficient of 0.30, and an interest matching weight coefficient of 0.25;

[0108] Knowledge Graph Enhanced Matching Strategy: Construct a knowledge graph of knowledge points, resources, learning objectives, and relevance. The initial value of the edge weight is calculated by the TF-IDF algorithm and dynamically updated through collaborative filtering.

[0109] Three-Layer Association Retrieval Mechanism: Screen resources within the three-layer association range of the user's knowledge gap nodes. The three layers are: the direct association layer, the indirect association layer, and the interdisciplinary extension layer. The core threshold range of the direct association layer is 0.5 ≤ weight ≤ 1.0, and dynamic adjustment is made using wrong questions. When the number of repetitions of wrong questions ≥ 3, the weight drops to 0.4 to trigger the retrieval of associated knowledge points. The core threshold range of the indirect association layer is 0.3 ≤ weight < 0.5. When students prefer video learning, the video weight is multiplied by 1.2 to increase the priority, and diagrammatic videos are dynamically appended. The core threshold of the interdisciplinary extension layer is 0.25 ≤ weight < 0.4, and an interest-driven threshold adjustment mechanism is adopted. When the student's interest score > 70, the lower limit of the threshold is adjusted to 0.2.

[0110] Exclude nodes with a core threshold < 0.6 in each association layer; adopt the priority and each association layer in the order of: direct association layer > indirect association layer > interdisciplinary extension layer. For example, "polynomial operation" with a weight of 0.5 is preferably processed according to the direct association layer, calculate the optimal learning path, and preferentially push resources associated with weak knowledge points.

[0111] Reinforcement Learning Dynamic Adjustment Strategy: The state space is the portrait feature vector and the environmental state, the action space is the adjustment of the resource sorting weight, and the reward function R = αΔK + β(1 - F) + γI; where, ΔK = knowledge gain, F = fatigue degree, I = interest matching; α represents the weight of the knowledge gain ΔK, and the typical value is in 0.4 ≤ α ≤ 0.6, dynamically adjusted through AB testing; β represents the weight of the fatigue degree F, and the typical value is in 0.2 ≤ β ≤ 0.3, γ represents the weight of the interest matching degree I, and the typical value is in 0.1 ≤ γ ≤ 0.2.

[0112] Cognitive Load Monitoring Metrics: Attention dispersion, learning efficiency ratio, and physiological signals.

[0113] Attention dispersion = standard deviation of the number of operations per unit time / average number of operations per unit time; the standard deviation of the number of operations per unit time and the average number of operations per unit time are obtained from the click stream event log. Example: A student clicks 12 times in 1 minute, and the standard deviation is 3 times, then the dispersion = 3 / 12 = 0.25; set the warning threshold, when the attention dispersion > 1.0 times the mean value, it is mild dispersion; > 1.5 times the mean value, it is moderate dispersion; > 2.5 times the mean value, it is severe dispersion and intervention is required.

[0114] Learning efficiency ratio = effective learning time / total learning duration × 100%. Criteria for determining effective learning time: continuous concentration duration ≥ 5 minutes and no ineffective operations; if the learning efficiency ratio < 40%, the reasons need to be diagnosed.

[0115] Physiological signals are detected by abnormal heart rate variability. If SDNN < 50ms, the reasons need to be diagnosed.

[0116] Pareto front screening mechanism: retain the non-dominated solution set, control resource diversity, the interval of the same type ≥ 3 items, and the exposure rate of new resources ≥ 15%.

[0117] Time period adaptation strategy: match different difficulty and type of resource combinations according to morning / afternoon / evening time periods.

[0118] The label adaptive update module further includes:

[0119] Causal effect calculation unit: calculate the net impact of resources on performance improvement through a counterfactual reasoning framework. An average treatment effect ATE ≥ 0.15 and a significance test p-value < 0.05 are considered effective.

[0120] Semantic space reconstruction unit: optimize the vector representation through contrastive learning and fully refresh the label weights every 24 hours.

[0121] Label propagation mechanism: based on the knowledge graph for propagation, locate the core knowledge point nodes of resources, and propagate the label weight w along the edges of the graph chlid = w parent × δ depth ;

[0122] where δ is the attenuation factor, depth is the propagation level, and w parent is the initial weight of the label;

[0123] Dual verification system of algorithm self-check and manual sampling inspection: the system algorithm performs real-time self-check, with a confidence level ≥ 80%, and manual sampling inspection is carried out weekly, with a sampling accuracy rate ≥ 90%.

[0124] Abnormal label isolation rule: automatically isolate when the contribution degree is < 0.05 for 3 consecutive evaluations, the association path with the core knowledge point ≥ 3 layers, or the number of user complaints > 5 times.

[0125] Version rollback strategy: load historical version difference analysis and repair when the label causes the click-through rate to drop > 15% or the association with key knowledge points is lost.

[0126] The personalized recommendation module The personalized recommendation module further includes:

[0127] Path generation algorithm: The hierarchical progressive method, dynamic pruning method, interest weighting method, and reinforcement learning are used to explore and match the resources in the zone of proximal development, and the content related to weak knowledge points is preferentially pushed. The topological sorting of the directed acyclic graph (DAG) is used to ensure that the prerequisite knowledge points are learned before the subsequent knowledge points, and a personalized recommendation sequence is generated. In the hierarchical progressive method, when the weight > 0.5, the content of the direct association layer is preferentially supplemented, and then extended to the indirect association layer. The logical regression model based on the student behavior data predicts the resource effectiveness, and the dynamic pruning method is implemented to filter the exercises with a historical correct rate > 90% in real time. The interest weighting method is introduced to increase the weight of the resources in the interdisciplinary extension layer according to the subject interest score. The interest score = historical click-through rate × 0.6 + average stay duration × 0.4. The average stay duration is calculated in minutes. Reinforcement learning is used to explore inserting 10% interdisciplinary challenge questions for the ZPD boundary knowledge points with 0.6 ≤ student mastery probability < 0.7 to test the transfer ability. Specifically, the Q-Learning model and the reward function are used to optimize the path at the ZPD boundary. The reward function = improvement in mastery × 0.7 + interest matching × 0.3;

[0128] Emergency switching mechanism: When it is detected that the resource jumps continuously for 3 times, the manual intervention process is automatically triggered and alternative solutions are generated.

[0129] The learning progress tracking module continuously tracks the learning progress of students, adjusts the recommendation strategy in a timely manner, and ensures that students always move forward along the most suitable learning path for themselves.

[0130] A method for a multi-modal dynamic optimization education resource recommendation system, characterized by including the following steps:

[0131] Step A: Multi-modal data collection and preprocessing, integrating video viewing behavior, question answering trajectories, interaction preferences, and physiological signals through edge computing and spatio-temporal alignment technologies to construct a learning feature map;

[0132] Step B: Use the LSTM-Attention network combined with the temporal convolutional network to dynamically model the knowledge forgetting curve and behavioral spatio-temporal characteristics of the student portrait, and calculate the knowledge mastery degree through the graph neural network propagation algorithm;

[0133] Step C: Resource dynamic matching, through reinforcement learning and knowledge graph analysis, realizing multi-objective optimization of knowledge gain, cognitive load, and interest matching, and generating a personalized learning path;

[0134] Step D: Label adaptive update, dynamically adjusting the resource label weights through causal inference and contrast learning technologies, and updating the label system in combination with the decay factor and the knowledge graph propagation mechanism;

[0135] Step E: Personalized Recommendation. Based on the student profile and real-time behavior data, generate a dynamic learning path and push appropriate resources. Specifically, it includes: combining knowledge gaps, interest preferences, and cognitive load to generate a phased learning plan; monitoring the resource completion rate and mid-course dropout rate, and dynamically adjusting the recommendation order; adapting the PC / mobile interface layout and interaction logic;

[0136] Step F: Learning Progress Tracking. Real-time monitor the learning status and feedback it to the resource matching engine to optimize the recommendation strategy in a closed loop.

[0137] The multi-modal data preprocessing in Step A includes: performing standardized processing such as timestamp alignment, data format unification, and outlier filtering, with a timestamp alignment error < 50ms; classifying and storing behavior logs, knowledge point data, and resource metadata, and performing multi-modal feature fusion on spatio-temporal alignment and semantic level mapping; and the quality verification must be complete, consistent, and compliant.

[0138] The specific content of the label adaptive update in Step D includes:

[0139] Learning Effect Tracking. Input data related to short-term, medium-term, and long-term effects. Data related to short-term effects are the resource completion rate, interaction depth, and immediate test scores; data related to medium-term effects are the improvement rate of unit tests and the repetition rate of wrong questions; data related to long-term effects are the semester grade trend and knowledge transfer ability; Data Association: Map the resource usage records to knowledge points and quantify the effect indicators;

[0140] Resource Efficiency Evaluation. Calculate and evaluate according to the direct contribution degree, indirect correlation degree, and group adaptability. The total resource efficiency score = direct contribution degree × 0.5 + indirect correlation degree × 0.3 + group adaptability × 0.2. When the total resource efficiency score ≥ 0.8, it is an excellent resource; when the total resource efficiency score is between 0.6 - 0.79, it is a good resource; when the total resource efficiency score < 0.6, it is a resource that needs to be optimized;

[0141] Optimize the resource library according to the total resource efficiency score, eliminate resources with a total resource efficiency score < 0.4, and recommend resources with a total resource efficiency score ≥ 0.6; Perform personalized placement according to the total resource efficiency score: Resources with a total resource efficiency score > 0.8 are resources with high group adaptability, automatically match the corresponding student labels, and resources with a total resource efficiency score < 0.4 are resources with low indirect correlation, and are only triggered when the direct correlation layer is weak;

[0142] Label Dynamic Adjustment. The adjustment strategy is to strengthen effective labels and weaken ineffective labels. The formula for strengthening labels is: The formula for weakening labels is: w new is the label value after strengthening or weakening, and w old is the initial label value; S baseTaking the average tag value of similar resources within the historical period as the benchmark effect value, the historical period is generally 30 days, e is the natural constant, ΔS is the effect improvement amplitude, λ is the decay coefficient, and T un is the invalid duration, and the tag types are knowledge dimension, ability dimension, and form dimension.

[0143] Map the resource usage records to knowledge points and quantify the effect indicators. The specific steps are as follows:

[0144] Binding resources to knowledge points means that each resource of video / exercise / experiment is labeled with a knowledge point ID. Through the manual labeling of the staff in the teaching and research team, the resource usage records are mapped one by one to the knowledge points, such as VIDEO_203 → knowledge point K12-MATH-3.2;

[0145] The quantification of effect indicators is divided into short-term effect quantification from 0 to 7 days, medium-term effect quantification from 1 to 3 months, and long-term effect quantification at the semester level;

[0146] The short-term effect quantification indicators from 0 to 7 days include resource completion rate, interaction depth, and immediate test score.

[0147] Resource completion rate = number of completions ÷ number of recommendations × 100%. If the number of mid-course exits is ≥ 3, then the total number of completions is multiplied by a coefficient of 0.7; the number of completions and the number of recommendations are obtained from the resource open / close event logs; a high completion rate indicates user recognition of the content, and a low completion rate requires investigation of resource difficulty or recommendation strategy;

[0148] Interaction depth = Σ (weight of single behavior × number of occurrences). Example of weights: play / pause = 0.3, redo questions = 0.7, note marking = 1.0; the weight of single behavior and the number of occurrences are obtained from clickstream data or button event tracking; a high proportion of high-weight behaviors indicates deep learning rather than passive browsing;

[0149] Immediate test score = (current test score - class average score) / class standard deviation × 10 + 50. The current test score, class average score, and class standard deviation are obtained from the exercise submission records and the class statistics database; a score > 70 indicates exceeding 70% of classmates, and < 30 requires priority intervention;

[0150] The medium-term effect quantification indicators from 1 to 3 months include the improvement amplitude of unit tests, the repetition rate of wrong questions, and the knowledge consolidation index;

[0151] Improvement amplitude of unit test = (current test score - previous test score) / previous test total score × 100%, only related to tests within the same knowledge cluster; an improvement amplitude < 20% indicates that the teaching plan needs to be optimized, and > 50% can be used as an excellent case for analysis;

[0152] Error question repetition rate = number of repeated error knowledge points / total number of error knowledge points × 100%. Error questions need to meet the condition that the interval between error questions of the same knowledge point ≤ 14 days; a repetition rate > 40% indicates that a personalized error correction plan needs to be initiated, and < 10% indicates that the weak points have been effectively broken through;

[0153] Knowledge consolidation index = probability of mastery × e^(-0.1 × interval days), where e is the natural constant, and the interval days = the number of days from the last learning to the test day. The probability of mastery and the interval days are obtained from the output of the Bayesian Knowledge Tracing (BKT) model and the learning behavior timestamp. A knowledge consolidation index < 0.3 indicates the need for re-learning, and > 0.7 indicates that the knowledge point has been stably mastered;

[0154] Quantitative indicators of long-term semester-level effects include semester grade trend, knowledge transfer ability, and learning resilience index;

[0155] The semester grade trend is represented by the slope of linear regression. A slope ≥ 0.2 indicates a significant improvement, and ≤ -0.1 indicates a decline. Dimension analysis is carried out by aggregating according to knowledge point clusters. A negative trend requires warning of subject shortboards, and a positive trend is used for teaching quality evaluation;

[0156] Knowledge transfer ability is represented by the cross-disciplinary score ratio. Cross-disciplinary score ratio = (cross-disciplinary application score / score of this subject) × 100%. The cross-disciplinary association needs to be marked. A ratio > 60% indicates outstanding high-order thinking ability, and < 30% requires strengthening interdisciplinary integration teaching;

[0157] The learning resilience index is represented by the resilience. Resilience = Σ(number of error questions corrected within 48 hours) ÷ total number of error questions × 100%. Record the time difference from the first error to the final correct answer. An index > 50% reflects strong resilience, and < 20% requires attention to learning motivation and psychological intervention.

[0158] Although the present invention has been described in detail with reference to the foregoing embodiments, for those skilled in the art, they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A multi-modal dynamic optimization educational resource recommendation system, characterized in that It includes the following core modules: Multimodal data acquisition module: Integrate video viewing behavior, question answering trajectory, interaction preferences, and physiological signals through edge computing and spatio-temporal alignment technology to construct a learning feature map; Student portrait construction module: Use an LSTM-Attention network combined with a temporal convolutional network to dynamically model the knowledge forgetting curve and spatio-temporal behavior characteristics, and calculate the knowledge mastery degree through the graph neural network propagation algorithm; Resource dynamic matching engine: Based on reinforcement learning and knowledge graph analysis, achieve multi-objective optimization of knowledge gain, cognitive load, and interest matching, and generate a personalized learning path; Label adaptive update module: Dynamically adjust the resource label weights through causal inference and contrast learning technology, and update the label system by combining the decay factor and the knowledge graph propagation mechanism; Personalized recommendation module: Generate a dynamic learning path and push adapted resources based on the student portrait and real-time behavior data. Specifically, it includes: generating a phased learning plan by combining knowledge gaps, interest preferences, and cognitive load; monitoring the resource completion rate and mid-course jump rate, and dynamically adjusting the recommendation order; adapting to the PC / mobile interface layout and interaction logic; Learning progress tracking module: Real-time monitor the learning status and feedback it to the resource matching engine to optimize the recommendation strategy in a closed loop.

2. The multimodal dynamic optimization-based educational resource recommendation system according to claim 1, wherein The multimodal data acquisition module further includes: Video viewing behavior analysis unit: Calculate the concentration index, where the concentration index = effective viewing duration / total duration × interference frequency correction coefficient; Question answering trajectory analysis unit: Mark the difficult points and record the answering path; Physiological signal processing unit: Evaluate the cognitive load through heart rate variability analysis; Interaction preference processing unit: Construct user interaction preference labels to provide a personalized decision basis for resource dynamic matching; Data preprocessing unit: Perform standardized processing of timestamp alignment, data format unification, and outlier filtering, classify and store behavior logs, knowledge point data, and resource metadata, perform multimodal feature fusion on spatio-temporal alignment and semantic level mapping, and the quality verification must be complete, consistent, and compliant.

3. A multimodal dynamic optimization-based educational resource recommendation system according to claim 1, characterized in that, The student portrait construction module further includes: Dynamic update mechanism: Retain the portraits of the last 10 historical versions, support version comparison and rollback; Knowledge Point Mastery Evaluation Model: The dual-channel GRU network processes the time series of correct rates and the correlation of wrong questions, combines the Bayesian knowledge tracing algorithm to calculate the probability of knowledge point mastery, and outputs the probability value P(mastery t ) = P(mastery t-1 ) + (1 - P(mastery t-1 )) · γ · correct t ; where γ is the learning rate parameter, and correct t is the correct identification of the question; t is the current time; Learning style classification model: Divide into visual, auditory, and practical types through principal component analysis PCA dimensionality reduction and K-means clustering, and match the corresponding resource strategies; Portrait credibility verification mechanism: Manually score by teachers, use the cross-validation method or A / B test to compare the portrait effects. Pass if the teacher's manual score ≥ 85 points. Use the cross-validation method to compare the difference between the predicted score and the actual exam score of the portrait. The A / B test is to compare the effects of different portrait versions.

4. A multimodal dynamic optimization-based educational resource recommendation system according to claim 1, characterized in that, The resource dynamic matching engine further includes: Multi-dimensional scoring model: Weighted comprehensive scoring based on a knowledge gain weight coefficient of 0.45, a cognitive load weight coefficient of 0.30, and an interest matching weight coefficient of 0.25; Knowledge graph enhanced matching strategy: Construct a knowledge graph of knowledge points, resources, learning objectives, and relevance. The initial value of the edge weight is calculated by the TF-IDF algorithm and dynamically updated through collaborative filtering; Three - layer associated retrieval mechanism: Screen resources within the three - layer association range of the user's knowledge gap nodes; the three layers are specifically: direct association layer, indirect association layer, and interdisciplinary expansion layer; the core threshold range of the direct association layer is 0.5 ≤ weight ≤ 1.0, and wrong questions are used for dynamic adjustment. When the number of repetitions of wrong questions ≥ 3, the weight drops to 0.4 to trigger the associated retrieval of knowledge points; the core threshold range of the indirect association layer is 0.3 ≤ weight < 0.

5. When students prefer video learning, the video weight is multiplied by 1.2 to increase the priority, and illustrated videos are dynamically appended; the core threshold of the interdisciplinary expansion layer is 0.25 ≤ weight < 0.4, and an interest - driven threshold adjustment mechanism is adopted. When the student's interest score > 70, the lower limit of the threshold is adjusted to 0.2; Exclude nodes with a core threshold < 0.6 in each associated layer; calculate the optimal learning path according to the priority and each associated layer in the order of: direct association layer > indirect association layer > interdisciplinary expansion layer, and preferentially push resources related to weak knowledge points; Reinforcement learning dynamic adjustment strategy: The state space is the portrait feature vector and the environmental state, the action space is the adjustment of the resource sorting weight, and the reward function R = αΔK + β(1 - F)+γI; where, ΔK = knowledge gain, F = fatigue, I = interest matching; α represents the weight of the knowledge gain ΔK, and the typical value ranges from 0.4 ≤ α ≤ 0.6, which is dynamically adjusted through AB testing; β represents the weight of the fatigue F, and the typical value ranges from 0.2 ≤ β ≤ 0.3, γ represents the weight of the interest matching degree I, and the typical value ranges from 0.1 ≤ γ ≤ 0.2; Cognitive load monitoring indicators: Attention dispersion, learning efficiency ratio, and physiological signals; Attention dispersion = standard deviation of the number of operations per unit time / average number of operations per unit time; the standard deviation of the number of operations per unit time and the average number of operations per unit time are obtained from the click - stream event log. Set the warning threshold. When the attention dispersion > 1.0 times the mean value, it is mild dispersion; > 1.5 times the mean value, it is moderate dispersion; > 2.5 times the mean value, it is severe dispersion and intervention is required; Learning efficiency ratio = effective learning time / total learning duration × 100%; the determination standard of effective learning time: continuous concentration duration ≥ 5 minutes and no ineffective operations; set the warning threshold. When the learning efficiency ratio < 40%, the reason needs to be diagnosed; Physiological signals are detected by abnormal heart rate variability. Set the warning threshold. When SDNN < 50ms, the reason needs to be diagnosed; Pareto - front screening mechanism: Retain the non - dominated solution set, control resource diversity, the interval of the same type ≥ 3 items, and the exposure rate of new resources ≥ 15%; Time - period adaptation strategy: Match resource combinations of different difficulties and types according to morning / afternoon / evening time periods.

5. A multimodal dynamic optimization-based educational resource recommendation system according to claim 1, characterized in that: The label adaptive update module further includes: Causal effect calculation unit: Calculate the net impact of resources on performance improvement through the counterfactual reasoning framework. When the average treatment effect ATE ≥ 0.15 and the significance test p - value < 0.05, it is regarded as effective; Semantic space reconstruction unit: Optimize the vector representation through contrastive learning, and fully refresh the label weight every 24 hours; Label propagation mechanism: Based on the propagation of the knowledge graph, locate the core knowledge point nodes of the resources, and propagate the label weight w along the edges of the graph chlid = w parent × δ depth ; where δ is the attenuation factor, depth is the propagation level, and w parent is the initial weight of the label; Dual verification system of algorithm self-check and manual sampling inspection: The system algorithm conducts real-time self-checking with a confidence level ≥ 80%, and manual sampling inspection is carried out weekly with a sampling accuracy rate ≥ 90%. Abnormal label isolation rule: Automatically isolate when the contribution degree is continuously evaluated < 0.05 for 3 times, the association path with the core knowledge points ≥ 3 layers, or the number of user complaints > 5 times. Version rollback strategy: When the click-through rate drops > 15% caused by the label or the association of key knowledge points is lost, load the historical version difference analysis and repair.

6. A multimodal dynamic optimization-based educational resource recommendation system according to claim 4, characterized in that: The personalized recommendation module further includes:[[]]END]] Path generation algorithm: Adopt the hierarchical progressive method, dynamic pruning method, interest weighting method and reinforcement learning to explore and match the resources in the zone of proximal development, give priority to pushing the content related to weak knowledge points, and use the topological sorting of the directed acyclic graph (DAG) to ensure that the precursor knowledge points are learned before the subsequent knowledge points, and generate a personalized recommendation sequence; in the hierarchical progressive method, when the weight > 0.5, first supplement the content of the direct association layer, and then expand to the indirect association layer; predict the resource effectiveness based on the logistic regression model of student behavior data, and implement the dynamic pruning method to filter the exercises with a historical correct rate > 90% in real time; introduce the interest weighting method to increase the weight of the resources in the interdisciplinary expansion layer according to the subject interest score, and the interest score = historical click-through rate × 0.6 + average stay time × 0.4; the average stay time is calculated in minutes; adopt reinforcement learning to explore and insert 10% interdisciplinary challenge questions for the ZPD boundary knowledge points with 0.6 ≤ student mastery probability < 0.7 to test the transfer ability, and specifically optimize the path at the ZPD boundary through the Q-Learning model and the reward function, and the reward function = mastery improvement × 0.7 + interest matching × 0.

3. Emergency switching mechanism: When it is detected that the resource jumps continuously for 3 times, automatically trigger the manual intervention process and generate an alternative plan.

7. A method for a multimodal dynamic optimization educational resource recommendation system according to any one of claims 1-6, characterized in that Including the following steps: Step A: Multi-modal data collection and preprocessing, integrate video viewing behavior, question answering trajectory, interaction preferences and physiological signals through edge computing and spatio-temporal alignment technology to construct a learning feature map. Step B: Adopt the LSTM-Attention network combined with the temporal convolutional network to dynamically model the knowledge forgetting curve and behavioral spatio-temporal characteristics of the student portrait, and calculate the knowledge mastery degree through the graph neural network propagation algorithm. Step C: Resource dynamic matching, through reinforcement learning and knowledge graph analysis, realize the multi-objective optimization of knowledge gain, cognitive load and interest matching, and generate a personalized learning path. Step D: Label adaptive update, dynamically adjust the resource label weight through causal inference and contrast learning technology, and update the label system in combination with the attenuation factor and the knowledge graph propagation mechanism. Step E: Personalized recommendation, generate a dynamic learning path and push the adapted resources based on the student portrait and real-time behavior data, specifically including: generate a phased learning plan by combining knowledge gaps, interest preferences and cognitive load; monitor the resource completion rate and the mid-way jump-out rate, and dynamically adjust the recommendation order; adapt to the PC / mobile interface layout and interaction logic. Step F: Learning progress tracking, real-time monitor the learning status, and feedback to the resource matching engine to optimize the recommendation strategy in a closed loop.

8. A method for a multimodal dynamic optimization educational resource recommendation system according to claim 7, characterized in that, The multimodal data preprocessing in step A includes: standardization of timestamp alignment, data format unification and outlier filtering, with timestamp alignment error less than 50ms; classified storage of behavior logs, knowledge point data and resource metadata, and multimodal feature fusion of spatiotemporal alignment and semantic level mapping; and quality verification must be complete, consistent and compliant.

9. A method for a multimodal dynamic optimization educational resource recommendation system according to claim 7, characterized in that, The label adaptive update in step D specifically includes: Learning effect tracking: input data related to short-term, medium-term and long-term effects. Data related to short-term effects include resource completion rate, interaction depth and instant test scores; data related to medium-term effects include unit test improvement and wrong question repetition rate; data related to long-term effects include semester grade trends and knowledge transfer ability; data association: map resource usage records with knowledge points and quantify effect indicators; Resource efficiency evaluation is calculated and evaluated based on direct contribution, indirect correlation, and group adaptability. The total resource efficiency score = direct contribution × 0.5 + indirect correlation × 0.3 + group adaptability × 0.

2. A total resource efficiency score of ≥ 0.8 indicates an excellent resource; a total resource efficiency score of 0.6-0.79 indicates a good resource; a total resource efficiency score of < 0.6 indicates a resource that needs to be optimized. Optimize the resource library based on the total resource efficiency score, eliminate resources with a total resource efficiency score of <0.4, and recommend resources with a total resource efficiency score ≥0.6; perform personalized delivery based on the total resource efficiency score: resources with a total resource efficiency score >0.8 are highly group-adaptive resources, which are automatically matched to corresponding student labels; resources with a total resource efficiency score <0.4 are low-indirect-association resources, which are triggered only when the direct association layer is weak; The label is dynamically adjusted. The adjustment strategy strengthens effective labels and weakens ineffective labels. The formula for strengthening a label is: The formula for weakening a label is: w new is the label value after strengthening or weakening, and w old is the initial label value; S base is the benchmark effect value, which is the average label value of similar resources in the historical period. e is the natural constant, ΔS is the effect improvement amplitude, λ is the decay coefficient, and T un is the ineffective duration. The label types are knowledge dimension, ability dimension, and form dimension.

10. The method of a multimodal dynamic optimization education resource recommendation system according to claim 9, characterized in that, Map resource usage records to knowledge points and quantify effect indicators. The specific steps are as follows: Binding resources and knowledge points means labeling each video / exercise / experiment resource with a knowledge point ID, and manually labeling by the teaching and research team staff to complete the one-to-one mapping of resource usage records and knowledge points; The effect index quantification is divided into 0-7 days short-term effect quantification, 1-3 months medium-term effect quantification and semester-level long-term effect quantification; The quantitative indicators of short-term effects in 0-7 days include resource completion rate, interaction depth and instant test results. Resource completion rate = completion times ÷ recommendation times × 100%. If the user quits ≥ 3 times, the total completion times × 0.7 coefficient. The completion times and recommendation times are obtained from the resource opening / closing event log. A high completion rate indicates the user's recognition of the content. A low completion rate requires checking the resource difficulty or recommendation strategy. Depth of interaction = Σ(weight of a single action × number of occurrences), where the weight of a single action and the number of occurrences are obtained from click flow data or button event tracking. A high proportion of high-weighted actions indicates deep learning rather than passive browsing. Instant test score = (this score - class average score) / class standard deviation × 10 + 50. This score, class average score, and class standard deviation are obtained from the exercise submission record and class statistics database. A score > 70 points indicates that you are ahead of 70% of your classmates, and a score < 30 points requires priority intervention. The quantitative indicators of the mid-term effect in 1-3 months include the improvement of unit tests, the repetition rate of wrong questions, and the knowledge consolidation index; Unit test improvement rate = (Current test score - Previous test score) / Total score of previous test × 100%, only related to tests within the same knowledge cluster; An improvement rate < 20% indicates that the teaching plan needs to be optimized, and > 50% can be used as an excellent case for analysis; Error repetition rate = Number of repeated error knowledge points / Total number of error knowledge points × 100%, and the wrong questions need to meet the condition that the interval between wrong questions of the same knowledge point ≤ 14 days; A repetition rate > 40% indicates that a personalized error correction plan needs to be initiated, and < 10% indicates that the weak points have been effectively broken through; Knowledge consolidation index = Mastery probability × e^(-0.1 × Interval days), where e is the natural constant, Interval days = Number of days from the last learning to the test day, and the mastery probability and interval days are obtained from the output of the Bayesian Knowledge Tracing (BKT) model and the learning behavior timestamp. A knowledge consolidation index < 0.3 indicates that re - learning is required, and > 0.7 indicates that the knowledge point has been stably mastered; Quantitative indicators of long - term semester - level effects include semester grade trend, knowledge transfer ability, and learning resilience index; The semester grade trend is represented by the slope of linear regression. A slope ≥ 0.2 indicates significant improvement, and ≤ - 0.1 indicates a decline. Dimension analysis is carried out by aggregating according to knowledge point clusters. A negative trend requires warning of subject weaknesses, and a positive trend is used for teaching quality assessment; Knowledge transfer ability is represented by the cross - disciplinary score ratio. Cross - disciplinary score ratio = (Cross - disciplinary application score / Score of this subject) × 100%, and the cross - disciplinary association needs to be marked. A ratio > 60% indicates outstanding high - order thinking ability, and < 30% requires strengthening cross - disciplinary integration teaching; The learning resilience index is represented by the frustration resistance. Frustration resistance = Σ (Number of wrong questions corrected within 48 hours) ÷ Total number of wrong questions × 100%, record the time difference from the first error to the final correct answer. An index > 50% reflects strong frustration resistance, and < 20% requires attention to learning motivation and psychological intervention.

Images