Self-adaptive intelligent teaching content recommendation system

Abstract

The invention discloses a self-adaptive intelligent teaching content recommendation system, and relates to the technical field of intelligent teaching, and the system comprises a data collection and analysis module which is used for collecting multiple types of data of students in the learning process, including learning behavior data, physiological signals, environment data and learning achievement cognition feedback data of the students, and after collection, sending the data to a database; performing preprocessing of noise reduction, standardization and analysis feature extraction on the multi-type data to generate multi-modal data; according to the method, the three-dimensional knowledge graph of knowledge points, error types and thinking paths is constructed, and the TCN analysis is combined, so that explicit knowledge defects can be identified, implicit knowledge vulnerabilities can be diagnosed, students can accurately know knowledge system vulnerabilities of themselves, and targeted defect checking, leak repairing and intensified training are carried out; learning interest information is extracted from multi-modal data, an explicit and implicit multi-dimensional interest model is established, and teaching content is screened in combination with a knowledge short board diagnosis result.

Description

Technical Field

[0001] The present invention relates to the field of intelligent teaching technology, and in particular to an adaptive intelligent teaching content recommendation system. Background Art

[0002] In today's education field, traditional teaching models are gradually shifting towards intelligence and personalization. Although many intelligent teaching systems have emerged, there are still technical bottlenecks in actual applications that make it difficult for students to achieve efficient and independent learning and precise improvement.

[0003] Existing technologies have the following shortcomings: For example, during the learning process, students often find it difficult to accurately identify gaps in their knowledge system. Traditional IRT / BKT models judge ability only by the accuracy of answering questions, which cannot capture complex knowledge gaps. Or, they trigger diagnosis only when students actively practice, failing to automatically identify potential weaknesses in unpracticed knowledge points. They lack in-depth analysis of the student learning process and cannot accurately locate students' knowledge gaps, making it difficult for students to conduct targeted assessments and fill gaps, and unable to provide sufficient intensive training for their weak links. Current intelligent teaching systems typically recommend content according to a preset difficulty gradient, often failing to accurately capture students' learning interests and ignoring the matching of content format with learners' interests. Recommendations are mainly based on course outlines and general learning needs, lacking the exploration of students' personalized interests. This can lead to students lacking enthusiasm and initiative in the recommended content, reducing their learning engagement. Furthermore, most existing intelligent teaching systems lack a comprehensive regular review mechanism and typically do not take into account individual memory decay characteristics. They are unable to rationally arrange review time and content based on students' learning status and forgetting curves. Students may forget important knowledge points shortly after learning new knowledge, affecting their long-term retention and application of knowledge.

[0004] Therefore, these existing technologies focus on the diagnosis of static knowledge and ignore the detection of students' knowledge gaps and potential weak areas. The lack of capturing students' interest and cognitive engagement, as well as the inadequate response to the forgetting curve, seriously restrict the effective recommendation of teaching content. An adaptive intelligent teaching content recommendation system is urgently needed to improve the shortcomings of existing technologies.

[0005] The above information disclosed in this Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not form the prior art that is already known to a person of ordinary skill in the art. Summary of the Invention

[0006] The purpose of the present invention is to provide an adaptive intelligent teaching content recommendation system. The present invention uses a three-dimensional knowledge graph combined with a temporal convolutional network (TCN) to analyze and identify explicit and implicit knowledge shortcomings, and solves the problems in the above-mentioned background technology by establishing an explicit and implicit multidimensional interest model and a model for predicting individual memory decay laws based on the Ebbinghaus forgetting curve.

[0007] To achieve the above objectives, the present invention provides the following technical solutions: an adaptive intelligent teaching content recommendation system, comprising a data acquisition and analysis module: collecting multiple types of data from students during the learning process, including students' learning behavior data, physiological signals, environmental data, and learning achievement cognitive feedback data; after collection, the multiple types of data are pre-processed by noise reduction, standardization, and feature extraction to generate multimodal data;

[0008] Multimodal knowledge diagnosis module: Based on multimodal data, the module uses the Neo4j graph database to construct a three-dimensional knowledge graph. The deep learning TCN network is used to analyze the three-dimensional knowledge graph to obtain knowledge gap diagnosis results, including identifying explicit knowledge deficiencies and diagnosing implicit knowledge gaps.

[0009] Interest-driven recommendation module: This module extracts learning interest information from multimodal data based on semantic analysis, establishes a multidimensional interest model that includes both explicit and implicit information, and uses a content recommendation algorithm to adaptively select and recommend teaching content that both meets students' interests and addresses their knowledge gaps, combined with the results of knowledge gap diagnosis.

[0010] Neural Memory Enhancement Module: Based on the memory prediction model of the Ebbinghaus forgetting curve, combined with students' learning time and environment data, it predicts individual memory decay patterns, dynamically adjusts review plans based on students' review feedback, and collaborates with the adaptive decision-making engine to trigger content regeneration;

[0011] Multi-source data fusion and adaptive decision engine module: Uses graph attention fusion algorithm to integrate characteristic data of knowledge gap diagnosis results, learning interest information, and individual memory decay rules, optimizes long-term learning benefits through reinforcement learning, and adaptively generates dynamic decisions on content recommendations and review plan combinations.

[0012] Optionally, the steps for collecting the multi-type data are as follows:

[0013] Through the logging function of the smart teaching app or by embedding HTML5 Video API in the smart teaching webpage, students’ learning behavior data is recorded and marked as Lbd;

[0014] Use smart bracelets, smart glasses, or medical-grade wearable devices for physiological monitoring to collect students' physiological signals in real time and calibrate them as Pss;

[0015] Light sensors, microphones, and UWB ultra-wideband positioning chips are deployed on the hardware mobile terminals of smart learning to collect real-time environmental data of students' learning places and calibrate it as Emd;

[0016] After studying the intelligent teaching content, the students' learning outcomes are counted and calibrated as Loc. At the same time, the WebAudioAPI is used to record the students' cognitive feedback voice information on intelligent learning and calibrate it as Cfb.

[0017] The expression of multi-type data is Mtd={Lbd, Pss, Emd, Loc, Cfb}, where Mtd represents multi-type data.

[0018] The steps for generating the multimodal data are as follows:

[0019] A hybrid Gaussian-wavelet threshold denoising model was developed to perform adaptive denoising on multi-type data, and the denoised data was normalized using Z-score normalization to obtain standardized data.

[0020] Principal component analysis (PCA) was used to extract behavioral, physiological, environmental, and cognitive features from the standardized data, and construct feature vectors.

[0021] After the feature vectors are time-synchronized and aligned according to the timestamps, the multimodal features are embedded and fused to output the multimodal fusion data.

[0022] The t-SNE projection method is used to reduce the dimensionality of multimodal fusion data, generate multimodal data and perform visualization. The multimodal data is calibrated as Mtd′, and Mtd′={Lbd′, Pss′, Emd′, Loc′, Cfb′}, where Lbd′, Pss′, Emd′, Loc′, and Cfb′ represent the preprocessed learning behavior data, physiological signals, environmental data, and learning outcome cognitive feedback data, respectively.

[0023] Optionally, the construction logic of the three-dimensional knowledge graph is as follows:

[0024] Extracting knowledge entities, attributes, and relationships based on multimodal data Mtd′. The extracted knowledge includes learning time and problem-solving efficiency for different knowledge points in learning behavior data Lbd, light intensity and ambient noise in learning environment data Emd, and learning performance and knowledge point mastery in learning outcomes Loc and cognitive feedback data Cfb.

[0025] The three dimensions of the three-dimensional knowledge graph are defined, including the knowledge point dimension, the error type dimension, and the thinking path dimension. The knowledge point dimension includes the subject knowledge point ontology and hierarchical relationships. The error type dimension includes the error types and associations of calculation errors, concept confusion, and step omissions. The thinking path dimension is the spatiotemporal evolution of students' cognitive paths in learning knowledge points.

[0026] The entities, attributes, and relationship information of knowledge extracted from multimodal data are mapped to the three dimensions of the three-dimensional knowledge graph: knowledge point dimension, error type dimension, and thinking path dimension;

[0027] Using the Neo4j graph database, we construct a three-dimensional knowledge graph with entities and attributes as nodes and relationships as edges. We also calculate the association strength of knowledge points, the association strength between knowledge points and error types, the association strength between knowledge points and thinking paths, and the influence of error propagation. The calculation formula for the association strength of knowledge points is: And L i 、L j ∈N + , L i 、L j =[1,5], where S ij It is expressed as the association strength between knowledge point i and knowledge point j, C ij It is expressed as the total number of occurrences of knowledge point i and knowledge point j in the error record, C i It is expressed as the number of times knowledge point i appears in the error record, C j It is represented by the number of times knowledge point j appears in the error record, λ is the cross-level attenuation factor, and λ∈(0,1), L i Expressed as the difficulty level of knowledge point i, L j Expressed as the difficulty level of knowledge point j, N + Expressed as a positive integer;

[0028] The calculation formula for the association strength between knowledge points and error types is: Where S ei Expressed as the association strength between error type e and knowledge point i, C ei It is expressed as the total number of occurrences of error type e and knowledge point i, C e It is expressed as the total number of times error type e appears alone, L e Expressed as the average difficulty level of error type e;

[0029] The calculation formula for the strength of the association between knowledge points and thinking paths is: Where S ip Expressed as the association strength between knowledge point i and thinking path p, C ip Expressed as the number of times knowledge point i appears in thinking path p, C′i The total number of occurrences of knowledge point i in all paths, C p Denotes the total number of steps in path p, γ denotes the path depth attenuation factor, and γ∈(0,1), D ip Expressed as the depth of knowledge point i in path p;

[0030] The formula for calculating the error propagation influence is: Where, I e It is expressed as the impact index of error type e on the knowledge system, i and I represent knowledge point i and the total number of knowledge points respectively, S ei Expressed as the association strength between error type e and knowledge point i, F i Expressed as the error frequency of knowledge point i, F total Expressed as the total number of system errors.

[0031] Optionally, the steps for obtaining the knowledge shortcoming diagnosis result are as follows:

[0032] For explicit knowledge defects, the DBSCAN algorithm is used to cluster error records and summarize error patterns. The core point condition of the DBSCAN algorithm clustering is N ∈ (Mtd′)≥MinPts, and MinPts=5, where N ∈ (Mtd') represents the number of neighborhood samples of the multimodal data Mtd' sample within the neighborhood radius ∈, ∈ represents the distance threshold that defines the neighborhood range of the sample, and MinPts represents the minimum number of neighborhood samples required to become a core point;

[0033] The impact of explicit knowledge defects is identified by calculating the defect severity score, where the calculation formula for the defect severity score is: And SC i ∈(0,1], where SC i is represented as the defect score of knowledge point i, and when SC i When it is >0.6, it is judged as a serious defect, and α is the error frequency F for knowledge point i. i The total number of system errors F total The weight coefficient, β is expressed as the impact index I of error type e on the knowledge system e The weight coefficient of

[0034] To identify implicit knowledge gaps, the 3D knowledge graph is fed into a deep learning temporal convolutional network (TCN). The dilated convolution of the TCN extracts features from nodes and edges in the knowledge graph by increasing the receptive field by the dilation rate, capturing the temporal dependencies and structural information between knowledge points and outputting a sequence of potential knowledge gaps.

[0035] The error type distribution is output through capsule network classification, and the three-dimensional knowledge graph is traced back to diagnose the association strength of knowledge points, calculate dynamic routing, and identify implicit knowledge loopholes. The calculation formula of dynamic routing is: and Where, d xy Expressed as the coupling coefficient from the lower-level capsule x to the higher-level capsule y, b xy is represented as the logarithmic prior probability in the dynamic routing iteration, z is represented as the total number of layers in the capsule network, Represented as the prediction vector of the high-level capsule y, ω xy Represented as outputting the lower-level capsule u x The learnable transformation matrix mapped to the high-level space, u x Represented as the output vector of the lower-level capsule x;

[0036] Then use the defect propagation algorithm to calculate the propagation probability of implicit knowledge vulnerabilities, where the calculation formula of the propagation probability is: and Where, P d is represented as the propagation probability of potential defect d, Γ(d) is represented as the set of knowledge points directly associated with defect d, S di Expressed as the correlation strength between defect d and knowledge point i, C di Expressed as the correlation strength between defect d and knowledge point i, C d Denotes the number of unique occurrences of latent defect d that were identified and diagnosed, C” i It is expressed as the number of times knowledge point i appears in the hidden defect, SC i It is represented as the defect score of knowledge point i.

[0037] Optionally, the steps for establishing the explicit and implicit multi-dimensional interest model are as follows:

[0038] Through semantic analysis, the learning content titles and descriptions are marked, and explicit interest information related to learning interests is directly extracted from the multimodal data Mtd′;

[0039] Based on the students’ historical learning content, hierarchical clustering is performed to build an explicit interest model, quantify the learning topic preference, and calculate the preference score of explicit interest. The calculation formula of the preference score is: Where s K is represented by the preference score of explicit interest in topic K, H is represented by the knowledge content, and H K It is represented as the set of knowledge content H belonging to topic K, dt H It is expressed as the residence time of knowledge content H, max(dt) is expressed as the maximum residence time of all knowledge contents, cc K It is expressed as the number of clicks on topic K;

[0040] The cognitive engagement index is calculated based on the gaze duration, the number of words in the notes, and the number of revisits to mine potential implicit interest information and build an implicit interest model. The calculation formula of the cognitive engagement index is: And ω1+ω2+ω3=1, where CEI represents the cognitive engagement index, ω1, ω2, and ω3 represent weight coefficients, a1, a2, and a3 represent the gaze duration, the number of words in the notes, and the number of revisits, respectively. A1, A2, and A3 represent the total duration of the learning content, the maximum number of words in all users' notes, and the maximum number of revisits, respectively.

[0041] Then, the explicit interest information and implicit interest information are interactively integrated and interest weights are assigned to reflect the importance of explicit and implicit interests. The calculation formula for explicit and implicit interest weights is ω m =ω E ·E m +ω I I m , and ω E +ω I =1, where ω m is the weight of the mth interest point, m is the number of interest points, ω E Expressed as the weight coefficient corresponding to the explicit interest score, E m Expressed as the explicit interest score of the mth interest point, ω I Expressed as the weight coefficient corresponding to the implicit interest score, I m It is expressed as the implicit interest score of the mth interest point.

[0042] Optionally, the steps for predicting the individual memory decay law are as follows:

[0043] The learning behavior data Lbd′ of students, including the duration of browsing course pages, time spent on answering questions and accuracy, the heart rate variability and cognitive load index of physiological signals Pss′, and the light intensity and environmental noise of environmental data Emd′, are obtained from the multimodal data Mtd′ to calculate the learning quality factor and environmental interference coefficient. The personalized memory strength parameters are optimized and fitted using maximum likelihood estimation based on historical learning data. The calculation formula of the learning quality factor is: And ω Q1 +ω Q2 =1, where Q0 is the learning quality factor, ω Q1 It is expressed as the weight coefficient corresponding to the correct rate, T% is expressed as the correct rate, t 反 Expressed as reaction time, t 基 Expressed as the historical average benchmark time for solving problems, ω Q2 It is expressed as the weight coefficient of the reaction time to the benchmark;

[0044] The calculation formula of environmental interference coefficient is: Lux max =1000 lux, W crit =65dB,ω Lux +ω W =1, where EI is the environmental interference coefficient, Lux t Expressed as the light intensity at the current moment t, Lux max Expressed as the upper limit of comfortable lighting, ω Lux Expressed as the weight coefficient of light intensity, W t Expressed as the noise decibel value at the current time t, W crit Denote the distraction threshold, ω W Expressed as the weight coefficient of environmental noise;

[0045] The calculation formula for optimizing the personalized memory strength parameter using maximum likelihood estimation is: In the formula, S0 represents the optimized personalized memory strength parameter, argmax represents the maximum likelihood function, and S represents the personalized memory strength parameter. Expressed as the actual measurement value of the n′th observation data point, Expressed as the probability density function of Gaussian distribution, R(t n′ )|S represents the personalized memory strength parameter S and time t n′ The generated model predictions;

[0046] Based on the Ebbinghaus forgetting curve, combined with the improvement of learning quality factor Q0, environmental interference coefficient EI and personalized memory strength parameter S0, a memory prediction model is constructed. The expression of the memory prediction model is: Where R(t) represents the memory retention rate after time t, e represents the exponential function, t represents the current moment, and S represents the personalized memory strength parameter;

[0047] Through LSTM network learning and training, the goal of minimizing the prediction error is defined, and the memory prediction model parameters are dynamically corrected. The calculation formula of the dynamic correction parameters is R'(t)=LSTM(R(t), ΔQ0, ΔEI, ΔS0, Δt), where R'(t) represents the memory retention rate after time t after dynamic correction of the LSTM network, LSTM represents the dynamic learning prediction of the LSTM network, and ΔQ0, ΔEI, ΔS0, and Δt all represent interference factors that quantify the impact on memory.

[0048] Optionally, the steps of integrating feature data of the graph attention fusion algorithm are as follows:

[0049] Using the graph attention fusion algorithm, the feature data of knowledge shortcoming diagnosis results, learning interest information, and individual memory decay rules are used as nodes to construct a heterogeneous graph;

[0050] The multi-head graph attention mechanism GAT is used to calculate the association weights between nodes;

[0051] According to the association weight, the features of adjacent nodes are aggregated using vector splicing operations, and the feature representation of each node is updated using the ELU activation function;

[0052] After fusing the feature data of all nodes, the integrated feature data representation is output.

[0053] Optionally, the adaptive generation logic flow of the dynamic decision is as follows:

[0054] The integrated feature data representation is used as the state input of the reinforcement learning network;

[0055] Define the action space, including content recommendation actions and review planning actions;

[0056] Design a reward function to evaluate the impact of content recommendation and review planning actions on long-term learning benefits, where long-term learning benefits include improved learning performance and knowledge mastery;

[0057] Use the proximal policy optimization algorithm PPO to learn the optimal decision-making strategy and select the action that can maximize the long-term reward based on the current state of the feature data;

[0058] Based on the learned optimal decision, dynamic decisions on the combination of content recommendation and review plan are adaptively generated according to the current state at each time step.

[0059] A computer device comprises: a memory and a processor; the memory stores a computer program, and the processor implements the steps of the above-mentioned adaptive intelligent teaching content recommendation system when executing the computer program.

[0060] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the above-mentioned adaptive intelligent teaching content recommendation system.

[0061] In the above technical solution, the technical effects and advantages provided by the present invention are:

[0062] The present invention constructs a three-dimensional knowledge graph of knowledge points, error types, and thinking paths combined with temporal convolutional network (TCN) analysis, which not only identifies explicit knowledge deficiencies but also diagnoses implicit knowledge gaps, allowing students to accurately know the gaps in their own knowledge system and conduct targeted gap-filling and intensive training. By extracting learning interest information from multimodal data, establishing an explicit and implicit multidimensional interest model, and screening teaching content based on the results of knowledge gap diagnosis, the present invention realizes the content recommendation function based on students' personalized interests, further improves students' enthusiasm and initiative for recommended content, and enhances learning participation. Based on the Ebbinghaus forgetting curve, by integrating learning behavior, environmental interference and physiological signals, the individual memory decay law is predicted, and the review plan is dynamically adjusted according to review feedback, which helps students to memorize and apply knowledge in the long term. In addition, through the graph attention mechanism, the data of knowledge gaps, learning interests, and memory decay are integrated, and reinforcement learning is used to optimize long-term learning benefits, adaptively generating dynamic decisions on content recommendation and review plan combinations, improving the intelligence and scientific nature of teaching content recommendation and review plan, and comprehensively improving teaching effectiveness and student learning efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction to the drawings required for use in the embodiments will be given below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0064] Figure 1 This is a module block diagram of the adaptive intelligent teaching content recommendation system of the present invention.

[0065] Figure 2 This is a flow chart of the adaptive intelligent teaching content recommendation method of the present invention. DETAILED DESCRIPTION

[0066] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these example embodiments are provided so that the description of this disclosure will be thorough and complete and will fully convey the concepts of the example embodiments to those skilled in the art.

[0067] The present invention provides Figure 1-2 The adaptive intelligent teaching content recommendation system shown includes a data acquisition and analysis module: it collects multiple types of data from students during the learning process, including students' learning behavior data, physiological signals, environmental data, and learning achievement cognitive feedback data. After collection, the multi-type data is pre-processed by noise reduction, standardization, and feature extraction to generate multimodal data;

[0068] Specifically, the steps for collecting multi-type data are as follows:

[0069] Through the logging function of the smart teaching APP or the HTML5 Video API embedded in the smart teaching webpage, students' learning behavior data is recorded and calibrated as Lbd, including login time, duration of browsing the course page, progress of watching the video (pause / fast forward / replay), time and accuracy of answering questions. Among them, the data collection of time and accuracy of answering questions can be used to calculate the students' answering efficiency. The calculation formula of answering efficiency is: Where p n Expressed as the efficiency of completing correct questions per unit time, c n It is the number of questions answered correctly by students in a period of time, t n It indicates the period of time for students to complete the total problem;

[0070] Using smart bracelets, smart glasses, or medical-grade physiological monitoring wearable devices, students' heart rate variability, brain waves, and eye movement trajectory physiological signals are collected in real time and calibrated as Pss. Among them, through the real-time collection of heart rate variability, brain waves, and eye movement trajectory data, students' learning concentration and cognitive load index can be calculated to associate students' learning status and learning effects. The calculation formula for concentration is: AC t ∈(0,1), where AC t It is expressed as the concentration of attention in the time window t, i and N represent the i-th eye movement and the total number of eye movements respectively. It is expressed as the gaze duration of the i-th eye movement hot zone, Expressed as the total learning time, is the blinking frequency, κ and θ are both Sigmoid parameters that adjust the blinking effect, with κ = 0.5, θ = 15, and e is an exponential function; the calculation formula of cognitive load index is and HRV max =150ms, where CLI is the cognitive load index, HRV is the heart rate variability, max Expressed as the reference maximum value of heart rate variability in healthy people, α EEG Expressed as the energy of alpha wave (8-12Hz) of EEG, beta EEG Expressed as the energy of beta waves (13-30Hz) of EEG;

[0071] Light sensors, microphones, and UWB ultra-wideband positioning chips are deployed on the hardware mobile terminal of smart learning to collect real-time environmental data on light intensity, ambient noise, and spatial positioning information of the student's learning place, calibrated as Emd. The environmental interference coefficient can be calculated by collecting light intensity and ambient noise, and then associated with the student's learning status. The calculation formula for the environmental interference coefficient is: Lux max =1000 lux, W crit =65dB, where EI is the environmental interference coefficient, Lux t Expressed as the light intensity at the current moment t, Lux max Expressed as the upper limit of comfortable lighting, ω Lux Expressed as the weight coefficient of light intensity, W t Expressed as the noise decibel value at the current time t, W crit Denote the distraction threshold, ω W Expressed as the weight coefficient of environmental noise;

[0072] After studying the intelligent teaching content, the students' learning outcomes are counted and labeled as Loc, including their learning scores and answer status in online tests, homework, exams, etc. At the same time, the WebAudio API is used to record students' cognitive feedback voice information on intelligent learning, labeled as Cfb, including students' independent learning experience, knowledge mastery, and evaluation of learning difficulties;

[0073] The expression of multi-type data is Mtd={Lbd, Pss, Emd, Loc, Cfb}, where Mtd represents multi-type data.

[0074] Specifically, the steps for generating multimodal data are as follows:

[0075] A hybrid Gaussian-wavelet threshold denoising model was developed to perform adaptive denoising on multi-type data, and the denoised data was normalized using Z-score normalization to obtain standardized data.

[0076] Principal component analysis (PCA) was used to extract behavioral, physiological, environmental, and cognitive features from the standardized data, and construct feature vectors.

[0077] After the feature vectors are time-synchronized and aligned according to the timestamps, the multimodal features are embedded and fused to output the multimodal fusion data.

[0078] The t-SNE projection method is used to reduce the dimensionality of multimodal fusion data, generate multimodal data and perform visualization to reduce redundant data. The multimodal data is calibrated as Mtd′, and Mtd′={Lbd′, Pss′, Emd′, Loc′, Cfb′}, where Lbd′, Pss′, Emd′, Loc′, and Cfb′ represent the preprocessed learning behavior data, physiological signals, environmental data, and learning outcome cognitive feedback data, respectively.

[0079] To further illustrate the working principle and effect of the above technical solution, by comprehensively collecting multi-type data of students in the learning process, it is possible to reflect the students' learning status and environmental factors from different angles, and by performing pre-processing operations such as noise reduction, standardization and feature extraction on multi-type data, different types of data can be converted into unified multimodal data, which can provide a more reliable basis for subsequent teaching content recommendations, student learning situation evaluation, etc.; in the pre-processing process, through noise reduction and standardization, the noise and dimensional differences in the data are removed, and the accuracy and comparability of the data are improved, and the feature extraction process can extract valuable features from a large amount of data, so that the students' learning characteristics and needs are explored, the pertinence and effectiveness of intelligent teaching are improved, and strong support is provided for personalized teaching content recommendation and learning plan formulation.

[0080] Multimodal knowledge diagnosis module: Based on multimodal data, the module uses the Neo4j graph database to construct a three-dimensional knowledge graph. The deep learning TCN network is used to analyze the three-dimensional knowledge graph to obtain knowledge gap diagnosis results, including identifying explicit knowledge deficiencies and diagnosing implicit knowledge gaps.

[0081] Specifically, the construction logic of the three-dimensional knowledge graph is as follows:

[0082] Extracting knowledge entities, attributes, and relationships based on multimodal data Mtd′. The extracted knowledge includes learning time and problem-solving efficiency for different knowledge points in learning behavior data Lbd, light intensity and ambient noise in learning environment data Emd, and learning performance and knowledge point mastery in learning outcomes Loc and cognitive feedback data Cfb.

[0083] The three dimensions of the three-dimensional knowledge graph are defined, including the knowledge point dimension, the error type dimension, and the thinking path dimension. The knowledge point dimension includes the subject knowledge point ontology and hierarchical relationships. The error type dimension includes the error types and associations of calculation errors, concept confusion, and step omissions. The thinking path dimension is the spatiotemporal evolution of students' cognitive paths in learning knowledge points, such as first understanding the definition, trying examples, and summarizing the rules.

[0084] The entities, attributes, and relationship information of knowledge extracted from multimodal data are mapped to the three dimensions of the three-dimensional knowledge graph: knowledge point dimension, error type dimension, and thinking path dimension;

[0085] Using the Neo4j graph database, we construct a three-dimensional knowledge graph with entities and attributes as nodes and relationships as edges. We also calculate the association strength of knowledge points, the association strength between knowledge points and error types, the association strength between knowledge points and thinking paths, and the influence of error propagation. The calculation formula for the association strength of knowledge points is: And L i 、L j ∈N + , L i 、L j =[1,5], where S ij It is expressed as the association strength between knowledge point i and knowledge point j, C ij It is expressed as the total number of occurrences of knowledge point i and knowledge point j in the error record, C i It is expressed as the number of times knowledge point i appears in the error record, C j It is represented by the number of times knowledge point j appears in the error record, λ is the cross-level attenuation factor, and λ∈(0,1), L i Expressed as the difficulty level of knowledge point i, L j Expressed as the difficulty level of knowledge point j, N + Expressed as a positive integer;

[0086] The calculation formula for the association strength between knowledge points and error types is: Where S ei Expressed as the association strength between error type e and knowledge point i, C ei It is expressed as the total number of occurrences of error type e and knowledge point i, C e It is expressed as the total number of times error type e appears alone, L e Expressed as the average difficulty level of error type e;

[0087] The calculation formula for the strength of the association between knowledge points and thinking paths is: Where S ip Expressed as the association strength between knowledge point i and thinking path p, C ip Expressed as the number of times knowledge point i appears in thinking path p, C′ i The total number of occurrences of knowledge point i in all paths, C p Denotes the total number of steps in path p, γ denotes the path depth attenuation factor, and γ∈(0,1), D ip Expressed as the depth of knowledge point i in path p;

[0088] The formula for calculating the error propagation influence is: Where, I e It is expressed as the impact index of error type e on the knowledge system, i and I represent knowledge point i and the total number of knowledge points respectively, S ei Expressed as the association strength between error type e and knowledge point i, F i Expressed as the error frequency of knowledge point i, F total Expressed as the total number of system errors.

[0089] Specifically, the steps for obtaining the knowledge gap diagnosis results are as follows:

[0090] For explicit knowledge defects, the DBSCAN algorithm is used to cluster error records and summarize error patterns. The core point condition of the DBSCAN algorithm clustering is N ∈ (Mtd′)≥MinPts, and MinPts=5, where N ∈ (Mtd') represents the number of neighborhood samples of the multimodal data Mtd' sample within the neighborhood radius ∈, ∈ represents the distance threshold that defines the neighborhood range of the sample, and MinPts represents the minimum number of neighborhood samples required to become a core point;

[0091] The impact of explicit knowledge defects is identified by calculating the defect severity score, where the calculation formula for the defect severity score is: And SC i ∈(0,1], where SC i is represented as the defect score of knowledge point i, and when SC i When it is >0.6, it is judged as a serious defect, and α is the error frequency F for knowledge point i. i The total number of system errors F total The weight coefficient, β is expressed as the impact index I of error type e on the knowledge system e The weight coefficient of

[0092] To identify implicit knowledge gaps, the 3D knowledge graph is fed into a deep learning temporal convolutional network (TCN). The dilated convolution of the TCN extracts features from nodes and edges in the knowledge graph by increasing the receptive field by the dilation rate, capturing the temporal dependencies and structural information between knowledge points and outputting a sequence of potential knowledge gaps.

[0093] The error type distribution is output through capsule network classification, and the three-dimensional knowledge graph is traced back to diagnose the association strength of knowledge points, calculate dynamic routing, and identify implicit knowledge loopholes. The calculation formula of dynamic routing is: and Where, d xy Expressed as the coupling coefficient from the lower-level capsule x to the higher-level capsule y, b xyis represented as the logarithmic prior probability in the dynamic routing iteration, z is represented as the total number of layers in the capsule network, Represented as the prediction vector of the high-level capsule y, ω xy Represented as outputting the lower-level capsule u x The learnable transformation matrix mapped to the high-level space, u x Represented as the output vector of the lower-level capsule x;

[0094] Then use the defect propagation algorithm to calculate the propagation probability of implicit knowledge vulnerabilities, where the calculation formula of the propagation probability is: and Where, P d is represented as the propagation probability of potential defect d, Γ(d) is represented as the set of knowledge points directly associated with defect d, S di Expressed as the correlation strength between defect d and knowledge point i, C di Expressed as the correlation strength between defect d and knowledge point i, C d Denotes the number of unique occurrences of latent defect d that were identified and diagnosed, C” i It is expressed as the number of times knowledge point i appears in the hidden defect, SC i It is represented as the defect score of knowledge point i.

[0095] To further illustrate the working principle and effect of the above technical solution, a three-dimensional knowledge graph is constructed based on multimodal data, and information such as students' learning behavior data, knowledge points and learning environment data is mapped into a three-dimensional graph structure of knowledge point dimension, error type dimension and thinking path dimension to intuitively represent the correlation between each other, which can comprehensively and accurately reflect the students' knowledge mastery. The knowledge graph is analyzed using a deep learning TCN network. The TCN network captures the time dependency and structural information in the knowledge graph through convolution operations, and learns the correlation patterns between different knowledge points, knowledge points and error types, and knowledge points and thinking paths. Again, through TCN network training and reasoning, the shortcomings of students in knowledge mastery are identified, and explicit knowledge deficiencies and implicit knowledge gaps are diagnosed. According to the diagnosis results, it helps intelligent teaching to formulate personalized teaching plans, provide targeted tutoring for students' shortcomings, and optimize and improve teaching effects.

[0096] Interest-driven recommendation module: This module extracts learning interest information from multimodal data based on semantic analysis, establishes a multidimensional interest model that includes both explicit and implicit information, and uses a content recommendation algorithm to adaptively select and recommend teaching content that both meets students' interests and addresses their knowledge gaps, combined with the results of knowledge gap diagnosis.

[0097] Specifically, the steps for establishing the explicit and implicit multidimensional interest model are as follows:

[0098] Through semantic analysis, we tag the titles and descriptions of learning content and directly extract explicit interest information related to learning interests from the multimodal data Mtd′. This includes pre-processed learning behavior data such as the length of time spent browsing course pages, the time and accuracy of answering questions, physiological signals of attention concentration, cognitive load index, learning achievement and students' autonomous learning experience.

[0099] Based on the students’ historical learning content, hierarchical clustering is performed to build an explicit interest model, quantify the learning topic preference, and calculate the preference score of explicit interest. The calculation formula of the preference score is: Where s K is represented by the preference score of explicit interest in topic K, H is represented by the knowledge content, and H K It is represented as the set of knowledge content H belonging to topic K, dt H It is expressed as the residence time of knowledge content H, max(dt) is expressed as the maximum residence time of all knowledge contents, cc K It is expressed as the number of clicks on topic K;

[0100] The cognitive engagement index is calculated based on the gaze duration, the number of words in the notes, and the number of revisits to mine potential implicit interest information and build an implicit interest model. The calculation formula of the cognitive engagement index is: And ω1+ω2+ω3=1, where CEI represents the cognitive engagement index, ω1, ω2, and ω3 represent weight coefficients, a1, a2, and a3 represent the gaze duration, the number of words in the notes, and the number of revisits, respectively. A1, A2, and A3 represent the total duration of the learning content, the maximum number of words in all users' notes, and the maximum number of revisits, respectively.

[0101] Then, the explicit interest information and implicit interest information are interactively integrated and interest weights are assigned to reflect the importance of explicit and implicit interests. The calculation formula for explicit and implicit interest weights is ω m =ω E ·E m +ω I I m , and ω E +ω I =1, where ω m is the weight of the mth interest point, m is the number of interest points, ω E Expressed as the weight coefficient corresponding to the explicit interest score, E m Expressed as the explicit interest score of the mth interest point, ω I Expressed as the weight coefficient corresponding to the implicit interest score, I m It is expressed as the implicit interest score of the mth interest point.

[0102] Specifically, the logical process of adaptive screening of the content recommendation algorithm is as follows:

[0103] Based on the knowledge gap diagnosis results, knowledge and interests are jointly screened to generate a candidate set of intelligently recommended teaching content;

[0104] Use the maximum marginal correlation method to control the diversity of the candidate set;

[0105] Establish a personalized sorting model and achieve adaptive sorting of intelligently recommended teaching content by defining target optimization.

[0106] To further illustrate the working principle and effect of the above technical solution, semantic analysis is used to identify emotional keywords, themes and emotional tendencies of multimodal data, extract students' learning interest information, and analyze students' mastery of various knowledge fields by establishing explicit and implicit multidimensional interest models. The multidimensional interest model is combined with the results of the knowledge gap diagnosis, and the content recommendation algorithm is used to screen the content in the teaching resource library, adaptively adjust the recommendation strategy, and continuously optimize the recommendation results. It can not only stimulate students' learning enthusiasm and initiative, participation and concentration, but also make up for students' knowledge gaps, help students consolidate weak knowledge points, improve the knowledge system, and further provide each student with personalized teaching content recommendations, meet the learning needs and interest preferences of different students, improve the utilization efficiency of teaching resources, and avoid students' time wasting in blindly searching in massive resources, so that teaching resources can be more reasonably allocated and utilized.

[0107] Neural Memory Enhancement Module: Based on the memory prediction model of the Ebbinghaus forgetting curve, combined with students' learning time and environment data, it predicts individual memory decay patterns, dynamically adjusts review plans based on students' review feedback, and collaborates with the adaptive decision-making engine to trigger content regeneration;

[0108] Specifically, the steps for predicting the individual memory decay law are as follows:

[0109] The learning behavior data Lbd′ of students, including the duration of browsing course pages, time spent on answering questions and accuracy, the heart rate variability and cognitive load index of physiological signals Pss′, and the light intensity and environmental noise of environmental data Emd′, are obtained from the multimodal data Mtd′ to calculate the learning quality factor and environmental interference coefficient. The personalized memory strength parameters are optimized and fitted using maximum likelihood estimation based on historical learning data. The calculation formula of the learning quality factor is: And ω Q1 +ω Q2 =1, where Q0 is the learning quality factor, ω Q1 It is expressed as the weight coefficient corresponding to the correct rate, T% is expressed as the correct rate, t反 Expressed as reaction time, t 基 Expressed as the historical average benchmark time for solving problems, ω Q2 It is expressed as the weight coefficient of the reaction time to the benchmark;

[0110] The calculation formula of environmental interference coefficient is: Lux max =1000 lux, W crit =65dB,ω Lux +ω W =1, where EI is the environmental interference coefficient, Lux t Expressed as the light intensity at the current moment t, Lux max Expressed as the upper limit of comfortable lighting, ω Lux Expressed as the weight coefficient of light intensity, W t Expressed as the noise decibel value at the current time t, W crit Denote the distraction threshold, ω W Expressed as the weight coefficient of environmental noise;

[0111] The calculation formula for optimizing the personalized memory strength parameter using maximum likelihood estimation is: In the formula, S0 represents the optimized personalized memory strength parameter, argmax represents the maximum likelihood function, and S represents the personalized memory strength parameter. Expressed as the actual measurement value of the n′th observation data point, Expressed as the probability density function of Gaussian distribution, R(t n′ )|S represents the personalized memory strength parameter S and time t n′ The generated model predictions;

[0112] Based on the Ebbinghaus forgetting curve, combined with the improvement of learning quality factor Q0, environmental interference coefficient EI and personalized memory strength parameter S0, a memory prediction model is constructed. The expression of the memory prediction model is: Where R(t) represents the memory retention rate after time t, e represents the exponential function, t represents the current moment, and S represents the personalized memory strength parameter;

[0113] Through LSTM network learning and training, the goal of minimizing the prediction error is defined, and the memory prediction model parameters are dynamically corrected. The calculation formula of the dynamic correction parameters is R'(t)=LSTM(R(t), ΔQ0, ΔEI, Δs0, Δt), where R'(t) represents the memory retention rate after time t after dynamic correction of the LSTM network, LSTM represents the dynamic learning prediction of the LSTM network, and ΔQ0, ΔEI, ΔS0, and Δt all represent interference factors that quantify the impact on memory.

[0114] To further illustrate the working principle and effect of the above technical solution, given that the Ebbinghaus forgetting curve describes the decay law of human memory over time, improvements are incorporated into learning time data and environmental data. Through learning quality factors, environmental interference coefficients and personalized memory strength parameters, a memory prediction model is established to predict the individual's memory decay law and determine the students' memory retention rate of the learned knowledge at different time points; the LSTM network is used to optimize the memory prediction model parameters, and then accurately predict and output the knowledge memory retention rate, so that according to the predicted memory decay law, an initial review plan is formulated for students, so that students can use their time more efficiently, improve their learning efficiency and maintain their mastery of knowledge for a long time.

[0115] Multi-source data fusion and adaptive decision engine module: Uses graph attention fusion algorithm to integrate characteristic data of knowledge gap diagnosis results, learning interest information, and individual memory decay rules, optimizes long-term learning benefits through reinforcement learning, and adaptively generates dynamic decisions on content recommendations and review plan combinations.

[0116] Specifically, the steps for integrating feature data in the graph attention fusion algorithm are as follows:

[0117] Using the graph attention fusion algorithm, the feature data of knowledge shortcoming diagnosis results, learning interest information, and individual memory decay rules are used as nodes to construct a heterogeneous graph;

[0118] The multi-head graph attention mechanism GAT is used to calculate the association weights between nodes;

[0119] According to the association weight, the features of adjacent nodes are aggregated using vector splicing operations, and the feature representation of each node is updated using the ELU activation function;

[0120] After fusing the feature data of all nodes, the integrated feature data representation is output.

[0121] Specifically, the adaptive generation logic flow of dynamic decision making is as follows:

[0122] The integrated feature data representation is used as the state input of the reinforcement learning network;

[0123] Define the action space, including content recommendation actions and review planning actions;

[0124] Design a reward function to evaluate the impact of content recommendation and review planning actions on long-term learning benefits, where long-term learning benefits include improved learning performance and knowledge mastery;

[0125] Use the proximal policy optimization algorithm PPO to learn the optimal decision-making strategy and select the action that can maximize the long-term reward based on the current state of the feature data;

[0126] Based on the learned optimal decision, dynamic decisions on the combination of content recommendation and review plan are adaptively generated according to the current state at each time step.

[0127] To further illustrate the working principle and effect of the above technical solution, the characteristic data of knowledge gap diagnosis results, learning interest information, and individual memory decay rules are represented as nodes in a heterogeneous graph, and the edges between nodes represent the association relationship between different data. The graph attention mechanism calculates the attention weight of each node to its neighboring nodes, and then calculates the attention coefficient through learnable parameters and activation functions, so that the model can automatically learn the association strength between different data, aggregate and update the characteristic data using vector splicing operations and output; then use reinforcement learning to input the updated characteristic data for learning, and evaluate the impact of each action on long-term learning benefits by defining actions and designing reward functions, and learn the optimal decision-making strategy through the PPO algorithm, so that dynamic decisions on content recommendation and review plan combinations are adaptively generated according to the current state at each time step; not only can highly personalized content recommendations and review plans be provided for each student, the continuous optimization process helps students better master knowledge and improve their academic performance, but also enhances students' learning ability and autonomous learning awareness, and adapts to changes in students' learning status.

[0128] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters in the formulas are set by technicians in this field according to actual conditions.

[0129] The above embodiments can be implemented in whole or in part by software, hardware, firmware or any other combination. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer program are loaded or executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center via wired or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that contains one or more available media sets. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state drive.

[0130] It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0131] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0132] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. An adaptive intelligent teaching content recommendation system, characterized in that: It includes a data collection and analysis module: it collects various types of data from students during the learning process, including students' learning behavior data, physiological signals, environmental data, and cognitive feedback data on learning outcomes. After collection, the multi-type data is pre-processed by noise reduction, standardization, and feature extraction to generate multimodal data; Multimodal knowledge diagnosis module: Based on multimodal data, the module uses the Neo4j graph database to construct a three-dimensional knowledge graph. The deep learning TCN network is used to analyze the three-dimensional knowledge graph to obtain knowledge gap diagnosis results, including identifying explicit knowledge deficiencies and diagnosing implicit knowledge gaps. Interest-driven recommendation module: This module extracts learning interest information from multimodal data based on semantic analysis, establishes a multidimensional interest model that includes both explicit and implicit information, and uses a content recommendation algorithm to adaptively select and recommend teaching content that both meets students' interests and addresses their knowledge gaps, combined with the results of knowledge gap diagnosis. Neural Memory Enhancement Module: Based on the memory prediction model of the Ebbinghaus forgetting curve, combined with students' learning time and environment data, it predicts individual memory decay patterns, dynamically adjusts review plans based on students' review feedback, and collaborates with the adaptive decision-making engine to trigger content regeneration; Multi-source data fusion and adaptive decision engine module: Uses graph attention fusion algorithm to integrate characteristic data of knowledge gap diagnosis results, learning interest information, and individual memory decay rules, optimizes long-term learning benefits through reinforcement learning, and adaptively generates dynamic decisions on content recommendations and review plan combinations.

2. The adaptive intelligent teaching content recommendation system according to claim 1, characterized in that: The steps for collecting the multi-type data are as follows: Through the logging function of the smart teaching app or by embedding HTML5 Video API in the smart teaching webpage, students’ learning behavior data is recorded and marked as Lbd; Use smart bracelets, smart glasses, or medical-grade wearable devices for physiological monitoring to collect students' physiological signals in real time and calibrate them as Pss; Light sensors, microphones, and UWB ultra-wideband positioning chips are deployed on the hardware mobile terminals of smart learning to collect real-time environmental data of students' learning places and calibrate it as Emd; After studying the intelligent teaching content, the students' learning outcomes are counted and calibrated as Loc. At the same time, the WebAudioAPI is used to record the students' cognitive feedback voice information on intelligent learning and calibrate it as Cfb. The expression of multi-type data is Mtd={Lbd, Pss, Emd, Loc, Cfb}, where Mtd represents multi-type data. The steps for generating the multimodal data are as follows: A hybrid Gaussian-wavelet threshold denoising model was developed to perform adaptive denoising on multi-type data, and the denoised data was normalized using Z-score normalization to obtain standardized data. Principal component analysis (PCA) was used to extract behavioral, physiological, environmental, and cognitive features from the standardized data, and construct feature vectors. After the feature vectors are time-synchronized and aligned according to the timestamps, the multimodal features are embedded and fused to output the multimodal fusion data. The t-SNE projection method is used to reduce the dimensionality of multimodal fusion data, generate multimodal data and perform visualization. The multimodal data is calibrated as Mtd′, and Mtd′={Lbd′, Pss′, Emd′, Loc′, Cfb′}, where Lbd′, Pss′, Emd′, Loc′, and Cfb′ represent the preprocessed learning behavior data, physiological signals, environmental data, and learning outcome cognitive feedback data, respectively.

3. The adaptive intelligent teaching content recommendation system according to claim 2, characterized in that: The construction logic of the three-dimensional knowledge graph is as follows: Extracting knowledge entities, attributes, and relationships based on multimodal data Mtd′. The extracted knowledge includes learning time and problem-solving efficiency for different knowledge points in learning behavior data Lbd, light intensity and ambient noise in learning environment data Emd, and learning performance and knowledge point mastery in learning outcomes Loc and cognitive feedback data Cfb. The three dimensions of the three-dimensional knowledge graph are defined, including the knowledge point dimension, the error type dimension, and the thinking path dimension. The knowledge point dimension includes the subject knowledge point ontology and hierarchical relationships. The error type dimension includes the error types and associations of calculation errors, concept confusion, and step omissions. The thinking path dimension is the spatiotemporal evolution of students' cognitive paths in learning knowledge points. The entities, attributes, and relationship information of knowledge extracted from multimodal data are mapped to the three dimensions of the three-dimensional knowledge graph: knowledge point dimension, error type dimension, and thinking path dimension; Using the Neo4j graph database, we construct a three-dimensional knowledge graph with entities and attributes as nodes and relationships as edges. We also calculate the association strength of knowledge points, the association strength between knowledge points and error types, the association strength between knowledge points and thinking paths, and the influence of error propagation. The calculation formula for the association strength of knowledge points is: And L i , L j ∈N + , L i , L j =[1,5], where S ij It is expressed as the association strength between knowledge point i and knowledge point j, C ij It is expressed as the total number of occurrences of knowledge point i and knowledge point j in the error record, C i It is expressed as the number of times knowledge point i appears in the error record, C j It is represented by the number of times knowledge point j appears in the error record, λ is the cross-level attenuation factor, and λ∈(0,1), L i Expressed as the difficulty level of knowledge point i, L j Expressed as the difficulty level of knowledge point j, N + Expressed as a positive integer; The calculation formula for the association strength between knowledge points and error types is: Where S ei Expressed as the association strength between error type e and knowledge point i, C ei It is expressed as the total number of occurrences of error type e and knowledge point i, C e It is expressed as the total number of times error type e appears alone, L e Expressed as the average difficulty level of error type e; The calculation formula for the strength of the association between knowledge points and thinking paths is: Where S ip Expressed as the association strength between knowledge point i and thinking path p, C ip Expressed as the number of times knowledge point i appears in thinking path p, C' i The total number of occurrences of knowledge point i in all paths, C p Denotes the total number of steps in path p, γ denotes the path depth attenuation factor, and γ∈(0,1), D ip Expressed as the depth of knowledge point i in path p; The formula for calculating the error propagation influence is: Where, I e It is expressed as the impact index of error type e on the knowledge system, i and I represent knowledge point i and the total number of knowledge points respectively, S ei Expressed as the association strength between error type e and knowledge point i, F i Expressed as the error frequency of knowledge point i, F total Expressed as the total number of system errors.

4. The adaptive intelligent teaching content recommendation system according to claim 3, characterized in that: The steps for obtaining the knowledge shortcoming diagnosis results are as follows: For explicit knowledge defects, the DBSCAN algorithm is used to cluster error records and summarize error patterns. The core point condition of the DBSCAN algorithm clustering is N ∈ (Mtd')≥MinPts, and MinPts=5, where N ∈ (Mtd') represents the number of neighborhood samples of the multimodal data Mtd' sample within the neighborhood radius ∈, ∈ represents the distance threshold that defines the neighborhood range of the sample, and MinPts represents the minimum number of neighborhood samples required to become a core point; The impact of explicit knowledge defects is identified by calculating the defect severity score, where the calculation formula for the defect severity score is: And SC i ∈(0,1], where SC i is represented as the defect score of knowledge point i, and when SC i When it is >0.6, it is judged as a serious defect, and α is the error frequency F for knowledge point i. i The total number of system errors F total The weight coefficient, β is expressed as the impact index I of error type e on the knowledge system e The weight coefficient of To identify implicit knowledge gaps, the 3D knowledge graph is fed into a deep learning temporal convolutional network (TCN). The dilated convolution of the TCN extracts features from nodes and edges in the knowledge graph by increasing the receptive field by the dilation rate, capturing the temporal dependencies and structural information between knowledge points and outputting a sequence of potential knowledge gaps. The error type distribution is output through capsule network classification, and the three-dimensional knowledge graph is traced back to diagnose the association strength of knowledge points, calculate dynamic routing, and identify implicit knowledge loopholes. The calculation formula of dynamic routing is: and Where, d xy Expressed as the coupling coefficient from the lower-level capsule x to the higher-level capsule y, b xy is represented as the logarithmic prior probability in the dynamic routing iteration, z is represented as the total number of layers in the capsule network, Represented as the prediction vector of the high-level capsule y, ω xy Represented as outputting the lower-level capsule u x The learnable transformation matrix mapped to the high-level space, u x Represented as the output vector of the lower-level capsule x; Then use the defect propagation algorithm to calculate the propagation probability of implicit knowledge vulnerabilities, where the calculation formula of the propagation probability is: and Where, P d is represented as the propagation probability of potential defect d, Γ(d) is represented as the set of knowledge points directly associated with defect d, S di Expressed as the correlation strength between defect d and knowledge point i, C di Expressed as the correlation strength between defect d and knowledge point i, C d Denotes the number of unique occurrences of latent defect d that were identified and diagnosed, C” i It is expressed as the number of times knowledge point i appears in the hidden defect, SC i It is represented as the defect score of knowledge point i.

5. The adaptive intelligent teaching content recommendation system according to claim 4, characterized in that: The steps for establishing the explicit and implicit multidimensional interest model are as follows: Through semantic analysis, the learning content titles and descriptions are marked, and explicit interest information related to learning interests is directly extracted from the multimodal data Mtd′; Based on the students’ historical learning content, hierarchical clustering is performed to build an explicit interest model, quantify the learning topic preference, and calculate the preference score of explicit interest. The calculation formula of the preference score is: Where s K is represented by the preference score of explicit interest in topic K, H is represented by the knowledge content, and H K It is represented as the set of knowledge content H belonging to topic K, dt H It is expressed as the residence time of knowledge content H, max(dt) is expressed as the maximum residence time of all knowledge contents, cc K It is expressed as the number of clicks on topic K; The cognitive engagement index is calculated based on the gaze duration, the number of words in the notes, and the number of revisits to mine potential implicit interest information and build an implicit interest model. The calculation formula of the cognitive engagement index is: And ω1+ω2+ω3=1, where CEI represents the cognitive engagement index, ω1, ω2, and ω3 represent weight coefficients, a1, a2, and a3 represent the gaze duration, the number of words in the notes, and the number of revisits, respectively. A1, A2, and A3 represent the total duration of the learning content, the maximum number of words in all users' notes, and the maximum number of revisits, respectively. Then, the explicit interest information and implicit interest information are interactively integrated and interest weights are assigned to reflect the importance of explicit and implicit interests. The calculation formula for explicit and implicit interest weights is ω m =ω E ·E m +ω I I m , and ω E +ω I =1, where ω m is the weight of the mth interest point, m is the number of interest points, ω E Expressed as the weight coefficient corresponding to the explicit interest score, E m Expressed as the explicit interest score of the mth interest point, ω I Expressed as the weight coefficient corresponding to the implicit interest score, I m It is expressed as the implicit interest score of the mth interest point.

6. The adaptive intelligent teaching content recommendation system according to claim 5, characterized in that: The steps for predicting the individual memory decay law are as follows: The learning behavior data Lbd′ of students, including the duration of browsing course pages, time spent on answering questions and accuracy, the heart rate variability and cognitive load index of physiological signals Pss′, and the light intensity and environmental noise of environmental data Emd′, are obtained from the multimodal data Mtd′ to calculate the learning quality factor and environmental interference coefficient. The personalized memory strength parameters are optimized and fitted using maximum likelihood estimation based on historical learning data. The calculation formula of the learning quality factor is: And ω Q1 +ω Q2 =1, where Q0 is the learning quality factor, ω Q1 It is expressed as the weight coefficient corresponding to the correct rate, T% is expressed as the correct rate, t 反 Expressed as reaction time, t 基 Expressed as the historical average benchmark time for solving problems, ω Q2 It is expressed as the weight coefficient of the reaction time to the benchmark; The calculation formula of environmental interference coefficient is: Lux max =1000 lux, W crit =65dB,ω Lux +ω W =1, where EI is the environmental interference coefficient, Lux t Expressed as the light intensity at the current moment t, Lux max Expressed as the upper limit of comfortable lighting, ω Lux Expressed as the weight coefficient of light intensity, W t Expressed as the noise decibel value at the current time t, W crit Denote the distraction threshold, ω W Expressed as the weight coefficient of environmental noise; The calculation formula for optimizing the personalized memory strength parameter using maximum likelihood estimation is: In the formula, S0 represents the optimized personalized memory strength parameter, argmax represents the maximum likelihood function, and S represents the personalized memory strength parameter. Expressed as the actual measurement value of the n'th observation data point, Expressed as the probability density function of Gaussian distribution, R(t n' )|S represents the personalized memory strength parameter S and time t n' The generated model predictions; Based on the Ebbinghaus forgetting curve, combined with the improvement of learning quality factor Q0, environmental interference coefficient EI and personalized memory strength parameter S0, a memory prediction model is constructed. The expression of the memory prediction model is: Where R(t) represents the memory retention rate after time t, e represents the exponential function, t represents the current moment, and S represents the personalized memory strength parameter; Through LSTM network learning and training, the goal of minimizing the prediction error is defined, and the memory prediction model parameters are dynamically corrected. The calculation formula of the dynamic correction parameters is R'(t)=LSTM(R(t), ΔQ0, ΔEI, ΔS0, Δt), where R'(t) represents the memory retention rate after time t after dynamic correction of the LSTM network, LSTM represents the dynamic learning prediction of the LSTM network, and ΔQ0, ΔEI, ΔS0, and Δt all represent interference factors that quantify the impact on memory.

7. The adaptive intelligent teaching content recommendation system according to claim 6, characterized in that: The steps of integrating feature data of the graph attention fusion algorithm are as follows: Using the graph attention fusion algorithm, the feature data of knowledge shortcoming diagnosis results, learning interest information, and individual memory decay rules are used as nodes to construct a heterogeneous graph; The multi-head graph attention mechanism GAT is used to calculate the association weights between nodes; According to the association weight, the features of adjacent nodes are aggregated using vector splicing operations, and the feature representation of each node is updated using the ELU activation function; After fusing the feature data of all nodes, the integrated feature data representation is output.

8. The adaptive intelligent teaching content recommendation system according to claim 7, characterized in that: The adaptive generation logic flow of the dynamic decision is as follows: The integrated feature data representation is used as the state input of the reinforcement learning network; Define the action space, including content recommendation actions and review planning actions; Design a reward function to evaluate the impact of content recommendation and review planning actions on long-term learning benefits, where long-term learning benefits include improved learning performance and knowledge mastery; Use the proximal policy optimization algorithm PPO to learn the optimal decision-making strategy and select the action that can maximize the long-term reward based on the current state of the feature data; Based on the learned optimal decision, dynamic decisions on the combination of content recommendation and review plan are adaptively generated according to the current state at each time step.

9. A computer device comprising: A memory and a processor; the memory stores a computer program, characterized in that when the processor executes the computer program, it implements the steps of an adaptive intelligent teaching content recommendation system according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of an adaptive intelligent teaching content recommendation system according to any one of claims 1 to 8 are implemented.

Images