An adaptive intelligent teaching content recommendation system

Abstract

The application discloses a kind of self-adapting intelligent teaching content recommendation system, it is related to intelligent teaching technical field, including data acquisition analysis module: collection student in the learning process multiple types of data, including student's learning behavior data, physiological signal, environmental data and learning achievement cognitive feedback data, after acquisition, multiple types of data are preprocessed with noise reduction, standardization and analysis feature extraction, generate multimodal data;The application is analyzed by constructing knowledge point, error type, three-dimensional knowledge graph of thinking path and time sequence convolution network TCN, not only can identify dominant knowledge defects, but also can diagnose implicit knowledge leaks, so that students can accurately know their own knowledge system leaks, targeted to check and fill the gap and strengthen training;By extracting learning interest information from multimodal data, an explicit and implicit multidimensional interest model is established, and the teaching content is filtered in combination with the knowledge gap diagnosis result.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent teaching, and particularly relates to a self-adaptive intelligent teaching content recommendation system. BACKGROUND

[0002] In the field of education today, the traditional teaching mode is gradually changing towards intelligence and individualization, although a large number of intelligent teaching systems have appeared, but there are still technical bottlenecks in actual application that lead to students' difficulty in achieving efficient self-learning and precise improvement.

[0003] The prior art has the following deficiencies: for example, in the learning process, students often have difficulty in accurately identifying the gaps in their knowledge system, the traditional IRT / BKT model only judges the ability through the correct rate of answering questions, cannot capture complex knowledge gaps, or only triggers diagnosis when students actively practice, cannot automatically identify potential weak areas of unpracticed knowledge points, lacks depth analysis of the learning process of students, cannot accurately locate the knowledge short board of students, leads to students' difficulty in targeted gap filling, and cannot sufficiently strengthen the training of their weak links; the current intelligent teaching system usually recommends content according to the preset difficulty gradient, often fails to accurately capture the learning interest of students and ignores the matching degree of content form and learner interest, the recommended content is mainly based on the curriculum outline and the general learning needs, lacks the mining of students' individualized interest, which makes students lack enthusiasm and initiative for the recommended content, reduces the participation in learning; and the existing intelligent teaching system mostly lacks a perfect regular review mechanism, usually does not combine the individual memory decay characteristics, cannot reasonably arrange the review time and content according to the learning situation and forgetting curve of students, students may forget important knowledge points soon after learning new knowledge, which affects the long-term memory and application of knowledge.

[0004] Therefore, these existing technologies diagnose static knowledge, ignore the detection of students' knowledge gaps and potential weak areas, lack the capture of students' interest and cognitive input, and have serious defects in the response to the forgetting curve, which seriously restricts the effective recommendation of teaching content, and an adaptive intelligent teaching content recommendation system is urgently needed to improve the deficiencies of the prior art.

[0005] The above information disclosed in the background section is only intended to strengthen the understanding of the background of the present disclosure, and therefore it can include information that does not constitute the prior art known to those of ordinary skill in the art. SUMMARY

[0006] The application aims to provide an adaptive intelligent teaching content recommendation system, which analyzes and identifies explicit and implicit knowledge short boards through a three-dimensional knowledge graph combined with a time sequence convolution network (TCN), and solves the problems in the background technology by establishing an explicit and implicit multi-dimensional interest model and constructing a prediction individual memory decay law model based on Ebbinghaus forgetting curve.

[0007] To achieve the above-mentioned purpose, the application provides the following technical scheme: an adaptive intelligent teaching content recommendation system, comprising a data collection and analysis module: collecting multiple types of data of students in the learning process, including learning behavior data, physiological signals, environmental data and learning achievement cognitive feedback data of students, after collection, pre-processing of multiple types of data for noise reduction, standardization and analysis feature extraction, generating multi-modal data;

[0008] A multi-modal knowledge diagnosis module: based on multi-modal data, a three-dimensional knowledge graph is constructed by using a Neo4j graph database, and a deep learning TCN network is used to analyze the three-dimensional knowledge graph to obtain knowledge short board diagnosis results, including identification of explicit knowledge defects and diagnosis of implicit knowledge gaps;

[0009] An interest-driven recommendation module: based on semantic analysis to extract learning interest information of multi-modal data, an explicit and implicit multi-dimensional interest model is established, combined with the knowledge short board diagnosis results, and a content recommendation algorithm is used to adaptively select teaching content that meets the students' interests and can make up for knowledge gaps for recommendation;

[0010] A neural memory enhancement module: based on the memory prediction model of Ebbinghaus forgetting curve, combined with the learning time and environmental data of students, the individual memory decay law is predicted, and the review plan is dynamically adjusted according to the review feedback of students, and the adaptive decision engine is coordinated to trigger content regeneration;

[0011] A multi-source data fusion and adaptive decision engine module: using a graph attention fusion algorithm to integrate the feature data of knowledge short board diagnosis results, learning interest information and individual memory decay law, and optimizing long-term learning benefits through reinforcement learning, and adaptively generating dynamic decisions of content recommendation and review plan combination.

[0012] Optionally, the collection step of the multiple types of data is as follows:

[0013] Through the log recording function of the intelligent teaching APP or the embedded HTML5 Video API in the intelligent teaching webpage, the learning behavior data of the students is recorded and labeled as Lbd;

[0014] Using smart bracelets, smart glasses or wearable devices for medical physiological monitoring, the physiological signals of the students are collected in real time and labeled as Pss;

[0015] Deploy light sensors, microphones and UWB ultra-wideband positioning chips on the hardware mobile terminal of intelligent learning, real-time collect the environmental data of the learning place where the student is located, and calibrate as Emd;

[0016] After learning the intelligent teaching content, the learning achievements of the students are counted and calibrated as Loc, and at the same time, the cognitive feedback voice information of the students on the intelligent learning is recorded using WebAudioAPI and calibrated as Cfb;

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

[0018] The generation steps of the multi-modal data are as follows:

[0019] A mixed Gaussian-wavelet threshold denoising model is developed to adaptively denoise the multi-type data, and the data after denoising is standardized by Z-score standardization to obtain standardized data;

[0020] The behavior characteristics, physiological characteristics, environmental characteristics and cognitive characteristics are extracted from the standardized data using principal component analysis (PCA), and a feature vector is constructed;

[0021] After time synchronization and alignment processing of the feature vector according to the time stamp, multi-modal feature embedding fusion is performed, and multi-modal fusion data is output;

[0022] The multi-modal fusion data is reduced in dimension using t-SNE projection method, and multi-modal data is generated and visualized, wherein the multi-modal data is calibrated as Mtd′, and Mtd′={Lbd′, Pss′, Emd′, Loc′, Cfb′}, wherein Lbd′, Pss′, Emd′, Loc′, Cfb′ represent the preprocessed learning behavior data, physiological signals, environmental data, learning achievement and cognitive feedback data, respectively.

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

[0024] According to the multi-modal data Mtd′, the entity, attribute and relationship information of the knowledge are extracted, wherein the extracted knowledge includes the learning time and the problem solving efficiency of different knowledge points in the learning behavior data Lbd, the illumination intensity and the environmental noise of the learning place environmental data Emd, the learning achievement and the knowledge point mastery in the learning achievement Loc and the cognitive feedback data Cfb;

[0025] Three dimensions of the three-dimensional knowledge graph are defined, including a knowledge point dimension, an error type dimension, and a thinking path dimension, wherein the knowledge point dimension includes subject knowledge point ontology and hierarchical relationship, the error type dimension includes error types and correlations of calculation errors, concept confusion, and step omission, and the thinking path dimension is a spatiotemporal evolution of a cognitive path of a student learning a knowledge point;

[0026] Entity, attribute, and relationship information extracted from multi-modal data are respectively mapped to three dimensions of the knowledge point dimension, the error type dimension, and the thinking path dimension of the three-dimensional knowledge graph;

[0027] A Neo4j graph database is used to construct the three-dimensional knowledge graph by taking entities and attributes as nodes and relationships as edges, and to calculate knowledge point correlation strength, knowledge point and error type correlation strength, knowledge point and thinking path correlation strength, and error propagation influence, wherein a calculation formula of the knowledge point correlation strength is and L i , L j ∈ N + , L i , L j = [1, 5], wherein S ij represents correlation strength of a knowledge point i and a knowledge point j, C ij represents a co-occurrence number of the knowledge point i and the knowledge point j in error records, C i represents a separate occurrence number of the knowledge point i in the error records, C j represents a separate occurrence number of the knowledge point j in the error records, and λ represents a cross-hierarchical attenuation factor, and λ ∈ (0, 1), L i represents a difficulty level of the knowledge point i, L j represents a difficulty level of the knowledge point j, and N + represents a positive integer;

[0028] A calculation formula of the knowledge point and error type correlation strength is wherein S ei represents correlation strength of an error type e and a knowledge point i, C ei represents a co-occurrence number of the error type e and the knowledge point i, C e represents a total number of separate occurrences of the error type e, L e represents an average difficulty level of the error type e;

[0029] A calculation formula of the knowledge point and thinking path correlation strength is wherein S ip represents correlation strength of a knowledge point i and a thinking path p, C ip represents a number of occurrences of the knowledge point i in the thinking path p, and C′i total number of occurrences of knowledge point i in all paths, C p total number of steps of path p, γ represents a path depth decay factor, and γ ∈ (0, 1), D ip represents the depth of knowledge point i in path p.

[0030] The calculation formula of error propagation influence is In the formula, I e represents the influence 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 represents the association strength of error type e and knowledge point i, F i represents the error frequency of knowledge point i, F total represents the total number of system errors.

[0031] Optionally, the knowledge gap diagnosis result acquisition step is as follows:

[0032] For explicit knowledge defects, DBSCAN algorithm is used to cluster error records and induce error patterns, wherein the core point condition formula of DBSCAN algorithm clustering is N ∈ (Mtd') ≥ MinPts, and MinPts = 5, wherein N ∈ (Mtd') represents the number of neighborhood samples of multi-modal data Mtd' samples within a neighborhood radius ∈, ∈ represents a 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] By calculating the defect severity score, the influence degree of explicit knowledge defects is identified, wherein the calculation formula of the defect severity score is and SC i ∈ (0, 1], wherein SC i represents the defect score of knowledge point i, and when SC i > 0.6, it is determined as a serious defect, α represents the weight coefficient of the error frequency F i of knowledge point i to the total number of system errors F total , and β represents the weight coefficient of the influence index I e of error type e on the knowledge system.

[0034] For implicit knowledge vulnerabilities, a three-dimensional knowledge graph is input into a deep learning time convolution network TCN, and the dilated convolution of the TCN network increases the receptive field through the operation of the expansion rate to extract features of the nodes and edges in the knowledge graph, capture the time dependence relationship and structure information between knowledge points, and output the potential defect sequence of knowledge.

[0035] The capsule network classifies the output error type distribution, traces the three-dimensional knowledge graph, diagnoses the knowledge point correlation strength, and calculates the dynamic routing to identify the implicit knowledge vulnerability, wherein the calculation formula of the dynamic routing is and In the formula, d xy represents the coupling coefficient of the low-layer capsule x to the high-layer capsule y, b xy represents the logarithmic prior probability in the dynamic routing iteration, z represents the total number of layers of the capsule network, represents the prediction vector of the high-layer capsule y, ω xy represents the learnable transformation matrix for mapping the low-layer capsule output u x to the high-layer space, u x represents the output vector of the low-layer capsule x.

[0036] The defect propagation algorithm is used to calculate the propagation probability of the implicit knowledge vulnerability, wherein the calculation formula of the propagation probability is and In the formula, P d represents the propagation probability of the potential defect d, Γ(d) represents the set of knowledge points directly associated with the defect d, S di represents the correlation strength of the defect d and the knowledge point i, C di represents the correlation strength of the defect d and the knowledge point i, C d represents the number of times the implicit defect d is identified and diagnosed alone, C” i represents the number of times the knowledge point i appears alone in the implicit defect, SC i represents the defect score of the knowledge point i.

[0037] Optionally, the establishment of the explicit and implicit multi-dimensional interest model is as follows:

[0038] The learning content title and description are marked through semantic analysis, and the explicit interest information associated with the learning interest is directly extracted from the multi-modal data Mtd';

[0039] The explicit interest model is constructed according to the hierarchical clustering of the student historical learning content, the learning theme preference is quantified, and the preference score of the explicit interest is calculated, wherein the calculation formula of the preference score is In the formula, s K represents the preference score of the explicit interest of the theme K, H represents the knowledge content, H K represents the set of knowledge content H belonging to the theme K, dt H represents the stay duration of the knowledge content H, max(dt) represents the maximum value of the stay duration among all knowledge contents, cc K represents the number of clicks of the theme K.

[0040] A cognitive engagement index is calculated according to the gaze duration, the number of note words, and the revisit times to mine potential implicit interest information and construct an implicit interest model, wherein the calculation formula of the cognitive engagement index is and ω1+ω2+ω3=1, wherein CEI represents the cognitive engagement index, ω1, ω2, and ω3 all represent weight coefficients, a1, a2, and a3 respectively represent the gaze duration, the number of note words, and the revisit times, A1, A2, and A3 respectively represent the total duration of learning content, the maximum value of the number of note words of all users, and the maximum value of the revisit times of all users;

[0041] The explicit interest information and the implicit interest information are interactively integrated, and interest weights are allocated to reflect the importance of the explicit and implicit interests, wherein the calculation formula of the explicit and implicit interest weights is ω m =ω E ·E m +ω I ·I m , and ω E +ω I =1, wherein ω m represents the weight of the mth interest point, m represents the number of interest points, ω E represents the weight coefficient corresponding to the explicit interest score, E m represents the explicit interest score of the mth interest point, ω I represents the weight coefficient corresponding to the implicit interest score, I m represents the implicit interest score of the mth interest point.

[0042] Optionally, the prediction step of the individual memory decay law is as follows:

[0043] The duration of browsing the course page, the time and accuracy of doing the questions, the heart rate variability and cognitive load index of the physiological signal Pss', and the light intensity and environmental noise of the environmental data Emd' of the learning behavior data Lbd' of the student are obtained from the multi-modal data Mtd' to calculate the learning quality factor and the environmental interference coefficient, and the personalized memory strength parameter is optimized and fitted by using the maximum likelihood estimation based on the historical learning data, wherein the calculation formula of the learning quality factor is and ω Q1 +ω Q2 =1, wherein Q0 represents the learning quality factor, ω Q1 represents the weight coefficient corresponding to the accuracy, T% represents the accuracy, t 反 represents the reaction time, t 基 represents the historical average reference time for solving the questions, and ω Q2 represents the weight coefficient of the reaction time for solving the questions relative to the reference;

[0044] The calculation formula of the environmental interference coefficient is and Lux max = 1000 lux, W crit = 65 dB, ω Lux + ω W = 1, wherein, EI represents the environmental interference coefficient, Lux t represents the illumination intensity at the current time t, Lux max represents the upper limit of comfortable light, ω Lux represents the weight coefficient of the illumination intensity, W t represents the noise decibel value at the current time t, W crit represents the attention distraction critical value, ω W represents the weight coefficient of the environmental noise.

[0045] The calculation formula for optimizing the personalized memory intensity parameter using maximum likelihood estimation is wherein, S0 represents the optimized personalized memory intensity parameter, argmax represents the maximum likelihood function, S represents the personalized memory intensity parameter, represents the actual measurement value of the nth observation data point, represents the probability density function of the Gaussian distribution, R(t n′ ) | S represents the generated model prediction value of the personalized memory intensity parameter S and the time t n′ .

[0046] Based on the Ebbinghaus forgetting curve, the learning quality factor Q0, the environmental interference coefficient EI and the improved personalized memory intensity parameter S0 are combined to construct a memory prediction model, wherein the expression of the memory prediction model is wherein, R(t) represents the memory retention rate after the time t, e represents the exponential function, t represents the current time, and S represents the personalized memory intensity parameter.

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

[0048] Optionally, the integration feature data step of the graph attention fusion algorithm is as follows:

[0049] The knowledge short board diagnosis result, learning interest information and feature data of individual memory attenuation law are fused as nodes to construct a heterogeneous graph using a graph attention fusion algorithm.

[0050] A multi-head graph attention mechanism GAT is used to calculate the correlation weight between nodes.

[0051] According to the correlation weight, the features of adjacent nodes are aggregated using a vector splicing operation, and the feature representation of each node is updated using an ELU activation function.

[0052] After the feature data of all nodes are fused, 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 input as the state of the reinforcement learning network.

[0055] An action space is defined, including a content recommendation action and a review plan action.

[0056] A reward function is designed to evaluate the impact of the content recommendation action and the review plan action on long-term learning benefits, wherein the long-term learning benefits include learning performance improvement and knowledge mastery.

[0057] An optimal decision strategy is learned using a proximal policy optimization algorithm PPO, and an action that can maximize the long-term reward is selected according to the current feature data state.

[0058] According to the learned optimal decision, a dynamic decision of content recommendation and review plan combination is adaptively generated at each time step according to the current state.

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

[0060] A computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the steps of the adaptive intelligent teaching content recommendation system.

[0061] In the above technical solution, the present application provides technical effects and advantages:

[0062] The application realizes not only identification of explicit knowledge defects, but also diagnosis of implicit knowledge defects, achieves that students can accurately know their own knowledge system defects, and carry out targeted defect filling and reinforcement training, extracts learning interest information from multi-modal data, establishes an explicit and implicit multi-dimensional interest model, and combines the knowledge short board diagnosis result to filter teaching content, realizes content recommendation function for personalized interest of students, further improves the enthusiasm and initiative of students to recommended content, and enhances learning participation; according to the Ebbinghaus forgetting curve, the individual memory decay law is predicted by fusing learning behavior, environmental interference and physiological signals, the review plan is dynamically adjusted according to the review feedback, which is helpful for long-term memory and application of students; and the data of knowledge short board, learning interest and memory decay are fused through the graph attention mechanism, and the long-term learning benefit is optimized by using reinforcement learning, and the dynamic decision of adaptive generation of content recommendation and review plan combination is generated, which improves the intelligence and scientificness of teaching content recommendation and review plan, and improves the teaching effect and student learning efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0063] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments described in the present application, and other drawings can also be obtained by those skilled in the art according to these drawings.

[0064] Fig. 1 The module block diagram of the adaptive intelligent teaching content recommendation system of the present application.

[0065] Fig. 2 The flow chart of the adaptive intelligent teaching content recommendation method of the present application. DETAILED DESCRIPTION

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

[0067] The present application provides an adaptive intelligent teaching content recommendation system as shown in Figs. 1-2 The present application provides an adaptive intelligent teaching content recommendation system as shown in

[0068] Specifically, the collection of multi-type data is as follows:

[0069] The learning behavior data of the student is recorded by the log recording function of the intelligent teaching APP or the embedded HTML5 Video API in the intelligent teaching webpage, and is labeled as Lbd, including login time, course page browsing time, video watching progress (pause / fast forward / playback), problem solving time and accuracy. The problem solving efficiency of the student can be calculated by collecting the problem solving time and accuracy data, and the calculation formula of the problem solving efficiency is wherein, p n represents the problem solving efficiency of the correct problems completed per unit time, c n represents the number of correct problems solved by the student in a period of time, t n represents the period of time for the student to complete all the problems; and

[0070] The physiological signals of the student, such as heart rate variability, brain waves and eye movement trajectories, are collected in real time by using smart bracelets, smart glasses or wearable devices for medical physiological monitoring, and are labeled as Pss. The attention concentration degree and cognitive load index of the student during learning can be calculated by collecting the heart rate variability, brain waves and eye movement trajectory data in real time, which are used to correlate the learning state and learning effect of the student. The calculation formula of the attention concentration degree is and AC t ∈(0, 1), wherein, AC t represents the attention concentration degree in the time window t, i and N represent the i th eye movement and the total number of eye movements, represents the gaze duration of the i th eye movement hot area, represents the total learning duration, represents the blink frequency, κ and θ represent the Sigmoid parameters for adjusting the influence of blinking, and κ = 0.5 and θ = 15, and e represents the exponential function; and the calculation formula of the cognitive load index is and HRV max = 150 ms, wherein, CLI represents the cognitive load index, HRV represents the heart rate variability, and HRV max represents the reference maximum value of the heart rate variability of healthy people, α EEG represents the α wave (8-12 Hz) energy of the brain waves EEG, β EEG represents the β wave (13-30 Hz) energy of the brain waves EEG;

[0071] The light sensor, microphone and UWB ultra-wideband positioning chip are deployed on the hardware mobile terminal of intelligent learning, and the environmental data of the light intensity, environmental noise and spatial positioning information of the learning place of the student are collected in real time, and are calibrated as Emd, wherein the environmental interference coefficient can be calculated through the collection of the light intensity and the environmental noise, and then the learning state of the student is associated, and the calculation formula of the environmental interference coefficient is and Lux max = 1000 lux, W crit = 65 dB, wherein, EI represents the environmental interference coefficient, Lux t represents the light intensity at the current t moment, Lux max represents the upper limit of comfortable light, ω Lux represents the weight coefficient of the light intensity, W t represents the noise decibel value at the current t moment, W crit represents the attention dispersion critical value, ω W represents the weight coefficient of the environmental noise.

[0072] After learning the intelligent teaching content, the learning achievements of the students are counted and calibrated as Loc, including the learning achievements and answering conditions in the forms of online test, homework and examination, at the same time, the cognitive feedback voice information of the students on the intelligent learning is recorded by using WebAudioAPI and is calibrated as Cfb, including the learning experience, knowledge mastery degree and learning difficulty evaluation of the students.

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

[0074] Specifically, the generation steps of the multi-modal data are as follows:

[0075] A mixed Gaussian-wavelet threshold denoising model is developed to adaptively denoise the multi-type data, and the data after denoising is standardized by Z-score standardization to obtain standardized data;

[0076] The behavior characteristics, physiological characteristics, environmental characteristics and cognitive characteristics are extracted from the standardized data by principal component analysis (PCA), and the feature vectors are constructed;

[0077] After the feature vectors are time-synchronized and aligned according to the time stamp, the multi-modal feature embedding fusion is performed, and the multi-modal fusion data is outputted;

[0078] The multi-modal fusion data is reduced in dimension by using a t-SNE projection method, multi-modal data is generated and visualized, and is used to reduce redundant data, wherein the multi-modal data is labeled as Mtd', and Mtd'={Lbd', Pss', Emd', Loc', Cfb'}; wherein Lbd', Pss', Emd', Loc', and Cfb' represent pre-processed learning behavior data, physiological signals, environmental data, learning achievement cognitive feedback data, respectively.

[0079] Further explanation of the working principle and effect of the above technical solution is that by comprehensively collecting multi-type data of students in the learning process, the learning state and environmental factors of students can be reflected from different angles, and by carrying out preprocessing operations such as noise reduction, standardization and feature extraction on multi-type data, different types of data are converted into unified multi-modal data, which can provide more reliable basis for subsequent teaching content recommendation, student learning condition evaluation, etc.; in the preprocessing process, by noise reduction and standardization processing, the noise and dimension difference in the data are removed, 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 learning characteristics and needs of students are mined, the pertinence and effectiveness of intelligent teaching are improved, and strong support is provided for personalized teaching content recommendation and learning scheme formulation.

[0080] The multi-modal knowledge diagnosis module: based on multi-modal data, a three-dimensional knowledge graph is constructed using a Neo4j graph database, and a deep learning TCN network is used to analyze the three-dimensional knowledge graph to obtain knowledge short board diagnosis results, including identifying explicit knowledge defects and diagnosing implicit knowledge gaps;

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

[0082] According to the multi-modal data Mtd', entity, attribute and relationship information of knowledge are extracted, wherein the extracted knowledge includes learning time and problem solving efficiency of different knowledge points in the learning behavior data Lbd, illumination intensity and environmental noise of the learning place environment data Emd, learning achievement Loc and knowledge point mastery in the cognitive feedback data Cfb;

[0083] Three dimensions of the three-dimensional knowledge graph are defined, including a knowledge point dimension, an error type dimension, and a thinking path dimension, wherein the knowledge point dimension includes subject knowledge point ontology and hierarchical relationship, the error type dimension includes calculation error, concept confusion, step omission error type and association, and the thinking path dimension is the spatio-temporal evolution of the cognitive path of students learning knowledge points, such as first understanding definition, trying examples and summarizing laws;

[0084] The entity, attribute and relationship information extracted from the multi-modal data are mapped to three dimensions of the three-dimensional knowledge graph, i.e., a knowledge point dimension, an error type dimension and a thinking path dimension;

[0085] Using a Neo4j graph database, the entity and attribute are taken as nodes, and the relationship is taken as an edge to construct the three-dimensional knowledge graph, and the knowledge point correlation strength, the knowledge point and error type correlation strength, the knowledge point and thinking path correlation strength and the error propagation influence are calculated, wherein the calculation formula of the knowledge point correlation strength is and L i , L j ∈N + , L i , L j =[1,5], wherein S ij represents the correlation strength of the knowledge point i and the knowledge point j, C ij represents the co-occurrence times of the knowledge point i and the knowledge point j in the error record, C i represents the individual occurrence times of the knowledge point i in the error record, C j represents the individual occurrence times of the knowledge point j in the error record, λ represents a cross-level attenuation factor, and λ∈(0, 1), L i represents the difficulty level of the knowledge point i, L j represents the difficulty level of the knowledge point j, and N + represents a positive integer;

[0086] The calculation formula of the knowledge point and error type correlation strength is wherein S ei represents the correlation strength of the error type e and the knowledge point i, C ei represents the co-occurrence times of the error type e and the knowledge point i, C e represents the total number of individual occurrences of the error type e, L e represents the average difficulty level of the error type e;

[0087] The calculation formula of the knowledge point and thinking path correlation strength is wherein S ip represents the correlation strength of the knowledge point i and the thinking path p, C ip represents the occurrence times of the knowledge point i in the thinking path p, C' i represents the total occurrence times of the knowledge point i in all paths, C p represents the total step number of the path p, and γ represents a path depth attenuation factor, and γ∈(0, 1), D ip represents the depth of the knowledge point i in the path p;

[0088] The calculation formula of the error propagation influence is wherein, I e is the influence index of error type e on the knowledge system, i and I represent the knowledge point i and the total number of knowledge points, respectively, S ei is the association strength of error type e and knowledge point i, F i is the error frequency of knowledge point i, F total is the total number of errors in the system.

[0089] Specifically, the acquisition step of the knowledge short board diagnosis result is as follows:

[0090] For the explicit knowledge defect, the DBSCAN algorithm is used to cluster the error records and induce the error mode, wherein the core point condition formula of the DBSCAN algorithm clustering is N ∈ (Mtd')≥MinPts, and MinPts=5, wherein, N ∈ (Mtd') represents the number of neighborhood samples of the multi-modal data Mtd' sample within the neighborhood radius ∈, ∈ represents the distance threshold value defining the neighborhood range of the sample, and MinPts represents the minimum number of neighborhood samples required to become a core point;

[0091] The influence degree of the explicit knowledge defect is identified by calculating the defect severity score, wherein the calculation formula of the defect severity score is and SC i ∈(0, 1], wherein, SC i represents the defect score of the knowledge point i, and when SC i >0.6, it is determined as a serious defect, α represents the weight coefficient of the error frequency F i of the knowledge point i in the total number of errors F total of the system, and β represents the weight coefficient of the influence index I e of the error type e on the knowledge system;

[0092] For the implicit knowledge vulnerability, the three-dimensional knowledge graph is input into the time sequence convolution network TCN of deep learning, the dilated convolution of the TCN network increases the receptive field through the operation of the expansion rate, and the features of the nodes and edges in the knowledge graph are extracted, the time dependence relationship and structure information between the knowledge points are captured, and the potential defect sequence of the knowledge is output;

[0093] The error type distribution is classified and output by the capsule network, the three-dimensional knowledge graph is traced back, the association strength of the knowledge points is diagnosed, and the dynamic routing is calculated to identify the implicit knowledge vulnerability, wherein the calculation formula of the dynamic routing is and wherein, d xy represents the coupling coefficient of the low-level capsule x to the high-level capsule y, b xydenoted as the logarithmic prior probability in the dynamic routing iteration, z is denoted as the total number of layers of the capsule network, denoted as the prediction vector of the high-level capsule y, ω xy denoted as the output vector of the low-level capsule x, u x denoted as the learnable transformation matrix mapping the low-level capsule output u to the high-level space, x denoted as the output vector of the low-level capsule x;

[0094] The propagation probability of the implicit knowledge gap is calculated again using the defect propagation algorithm, wherein the calculation formula of the propagation probability is and In the formula, P d denoted as the propagation probability of the latent defect d, Γ(d) is denoted as the set of knowledge points directly associated with the defect d, S di denoted as the association strength of the defect d and the knowledge point i, C di denoted as the association strength of the defect d and the knowledge point i, C d denoted as the number of times of identifying and diagnosing the implicit defect d alone, C" i denoted as the number of times of knowledge point i appearing alone in the implicit defect, SC i denoted as the defect score of the knowledge point i.

[0095] Further, the working principle and effects of the above technical solutions are as follows: based on the multi-modal data, a three-dimensional knowledge graph is constructed, and the learning behavior data, knowledge points and learning environment data of students and other information are mapped to the three-dimensional graph structure of the knowledge point dimension, error type dimension and thinking path dimension, so as to intuitively represent the association relationship between each other, which can comprehensively and accurately reflect the knowledge mastery of students. The knowledge graph is analyzed by using a deep learning TCN network, the TCN network captures the time dependence and structure information in the knowledge graph through convolution operation, learns the association mode between different knowledge points, knowledge points and error types, and knowledge points and thinking paths, and then the TCN network is trained and reasoned again to identify the short board of students in knowledge mastery, diagnose the explicit knowledge defects and implicit knowledge gaps, and according to the diagnosis result, the intelligent teaching can help to formulate individualized teaching plan, targetedly guide the students according to their short board, and optimize the teaching effect.

[0096] Interest-driven recommendation module: based on semantic analysis, the learning interest information of multi-modal data is extracted, an explicit and implicit multi-dimensional interest model is established, combined with the knowledge short board diagnosis result, the content recommendation algorithm is used to adaptively filter out the teaching content which meets the interest of students and can make up for the knowledge gap for recommendation;

[0097] Specifically, the establishment steps of the explicit and implicit multi-dimensional interest model are as follows:

[0098] The learning content title and description are marked by semantic analysis, explicit interest information associated with learning interest is directly extracted from the multi-modal data Mtd', including the time length of browsing course pages, the time and accuracy of doing exercises, the attention concentration degree of physiological signals, the cognitive load index, the learning achievement cognitive feedback data of the pre-processed learning behavior data, and the learning experience of students;

[0099] The explicit interest model is constructed by hierarchical clustering according to the historical learning content of students, the learning theme preference is quantified, and the preference score of explicit interest is calculated, wherein the calculation formula of the preference score is In the formula, s K represents the preference score of explicit interest for the theme K, H represents the knowledge content, and H K represents the set of knowledge content H belonging to the theme K, dt H represents the time length of staying for the knowledge content H, max(dt) represents the maximum value of the time length of staying in all knowledge contents, cc K represents the number of clicks for the theme K;

[0100] The cognitive investment index is calculated according to the gaze time, the number of note words and the revisit times, the potential implicit interest information is mined, and the implicit interest model is constructed, wherein the calculation formula of the cognitive investment index is and ω1+ω2+ω3=1, in the formula, CEI represents the cognitive investment index, ω1, ω2, ω3 all represent the weight coefficient, a1, a2, a3 respectively represent the gaze time, the number of note words and the revisit times, A1, A2, A3 respectively represent the total time length of learning content, the maximum value of all user note words and the maximum value of all user revisit times;

[0101] Then the explicit interest information and the implicit interest information are interactively integrated, and the interest weight is allocated to reflect the importance of explicit and implicit interest, wherein the calculation formula of the explicit and implicit interest weight is ω m =ω E ·E m +ω I ·I m , and ω E +ω I =1, in the formula, ω m represents the weight of the mth interest point, m represents the number of interest points, ω E represents the weight coefficient corresponding to the explicit interest score, E m represents the explicit interest score of the mth interest point, ω I represents the weight coefficient corresponding to the implicit interest score, I m represents the implicit interest score of the mth interest point.

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

[0103] According to the knowledge gap diagnosis result, the knowledge-interest is jointly screened to generate a candidate set of intelligent recommended teaching content;

[0104] The maximum boundary correlation method is used to control the diversity of the candidate set;

[0105] An individualized ranking model is established to realize the adaptive ranking of the intelligent recommended teaching content by defining the target optimization.

[0106] Further explanation of the working principle and effect of the above technical solution is that semantic analysis is used to identify keywords, topics and sentiment analysis about emotion on multi-modal data, and the learning interest information of students is extracted. By establishing an explicit and implicit multi-dimensional interest model, the mastery of students in each knowledge field is analyzed. The combination of the multi-dimensional interest model and the knowledge gap diagnosis result uses the content recommendation algorithm to screen the content in the teaching resource library, adaptively adjusts the recommendation strategy, and continuously optimizes the recommendation result. Not only can it stimulate students' learning enthusiasm and initiative, participation and concentration, but also can fill the knowledge gaps of students, help students consolidate weak knowledge points, perfect the knowledge system, further provide personalized teaching content recommendation for each student, meet the learning needs and interest preferences of different students, improve the utilization efficiency of teaching resources, and avoid the waste of time in the massive resources. Students blindly search, so that teaching resources can be more reasonably allocated and utilized.

[0107] Neural memory enhancement module: a memory prediction model based on Ebbinghaus forgetting curve, combining students' learning time and environmental data, predicting individual memory decay law, and dynamically adjusting the review plan according to students' review feedback, coordinating with the adaptive decision engine to trigger content regeneration;

[0108] Specifically, the prediction steps of the individual memory decay law are as follows:

[0109] From the multi-modal data Mtd' to obtain the learning behavior data Lbd' of the student, including the duration of browsing the course page, the time and accuracy of doing exercises, the heart rate variability and cognitive load index of the physiological signal Pss', and the light intensity and environmental noise of the environmental data Emd', to calculate the learning quality factor and the environmental interference coefficient, and through the historical learning data, the maximum likelihood estimation is used to optimize the fitting of the personalized memory intensity parameter, wherein the calculation formula of the learning quality factor is and ω Q1 + ω Q2 = 1, wherein Q0 represents the learning quality factor, ω Q1 represents the weight coefficient corresponding to the accuracy, T% represents the accuracy, and t反 denoted as reaction time, t 基 denoted as reference time of historical average solving, ω Q2 denoted as weight coefficient of reaction time of doing problems to reference;

[0110] The calculation formula of environmental interference coefficient is and Lux max = 1000 lux, W crit = 65 dB, ω Lux + ω W = 1, wherein, EI is denoted as environmental interference coefficient, Lux t denoted as current illumination intensity at time t, Lux max denoted as upper limit of comfortable illumination, ω Lux denoted as weight coefficient of illumination intensity, W t denoted as current noise decibel value at time t, W crit denoted as attention dispersion critical value, ω W denoted as weight coefficient of environmental noise;

[0111] The calculation formula of optimizing personalized memory intensity parameter using maximum likelihood estimation is wherein, S0 is denoted as optimized personalized memory intensity parameter, argmax is denoted as maximizing likelihood function, S is denoted as personalized memory intensity parameter, denoted as actual measurement value of the nth observation data point, denoted as probability density function of Gaussian distribution, R(t n′ )|S is denoted as personalized memory intensity parameter S and time t n′ generated model prediction value;

[0112] Based on Ebbinghaus forgetting curve, the improvement of learning quality factor Q0, environmental interference coefficient EI and personalized memory intensity parameter S0 is combined to construct a memory prediction model, wherein the expression of the memory prediction model is wherein, R(t) is denoted as memory retention rate after time t, e is denoted as exponential function, t is denoted as current time, and S is denoted as personalized memory intensity parameter;

[0113] Through LSTM network learning and training, the target of minimizing prediction error is defined, and the memory prediction model parameters are dynamically corrected, wherein the calculation formula of dynamically correcting parameters is R'(t) = LSTM(R(t), ΔQ0, ΔEI, Δs0, Δt), wherein R'(t) is denoted as memory retention rate after time t after LSTM network dynamic correction, LSTM is denoted as LSTM network dynamic learning prediction, and ΔQ0, ΔEI, ΔS0 and Δt all denote interference factors quantifying the influence on memory.

[0114] Further explanation of the working principle and effect of the above technical solution is that, in view of the fact that Ebbinghaus forgetting curve describes the decay law of human memory over time, the improvement of learning time data and environment data is fused, the memory prediction model is established by learning quality factor, environmental interference coefficient and personalized memory strength parameter to predict the memory decay law of individual, and the memory retention rate of students on learned knowledge at different time points is determined; the memory prediction model parameters are optimized by using LSTM network, and then the knowledge memory retention rate is accurately predicted and output, so as to make the initial review plan for students according to the predicted memory decay law, so that students can make more efficient use of time, improve learning efficiency and long-term retention of knowledge.

[0115] Multi-source data fusion and adaptive decision engine module: using graph attention fusion algorithm to integrate knowledge gap diagnosis results, learning interest information, and feature data of individual memory decay law, and optimizing long-term learning benefits through reinforcement learning to adaptively generate dynamic decisions of content recommendation and review plan combination.

[0116] Specifically, the integration feature data step of the graph attention fusion algorithm is as follows:

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

[0118] The multi-head graph attention mechanism GAT is used to calculate the correlation weight between nodes.

[0119] According to the correlation weight, the features of adjacent nodes are aggregated using vector concatenation operation, and the feature representation of each node is updated using 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 is as follows:

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

[0123] Define the action space, including content recommendation action and review plan action.

[0124] Design a reward function to evaluate the impact of content recommendation action and review plan action on long-term learning benefits, wherein the long-term learning benefits include learning achievement improvement and knowledge mastery degree.

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

[0126] The dynamic decision of the combination of the content recommendation and the review plan is adaptively generated according to the current state at each time step according to the learned optimal decision.

[0127] Further, the working principle and effects of the above technical solutions are as follows: the knowledge short board diagnosis result, the learning interest information and the characteristic data of the individual memory attenuation law are respectively represented as nodes in the heterogeneous graph, the association relationship between different data is represented according to the edges between the nodes, the graph attention mechanism calculates the attention weight of each node to its neighbor node, and then the attention coefficient is calculated through the learnable parameters and the activation function, so that the model can automatically learn the association strength between different data, the vector splicing operation is used to aggregate and update the characteristic data and output; the updated characteristic data is input into the reinforcement learning to learn, the influence of each action on the long-term learning benefit is evaluated by defining the action and designing the reward function, and the optimal decision strategy is learned through the PPO algorithm, so that the dynamic decision of the combination of the content recommendation and the review plan is adaptively generated according to the current state at each time step; not only can each student be provided with highly personalized content recommendation and review plan, but also the continuous optimization process helps students better master knowledge and improve learning performance, and at the same time, the learning ability and autonomous learning consciousness of students are enhanced, and the learning state of students is adapted.

[0128] The above formulas are all dimensionless values, and the formulas are obtained by software simulation of a large amount of data to obtain a formula of the latest real situation, and the preset parameters in the formula are set by a person skilled in the art according to the actual situation.

[0129] The above-described embodiments can be implemented in part or in whole through software, hardware, firmware or any combination thereof. When implemented in software, the above-described embodiments can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When loaded and executed by a computer, the computer instructions or computer programs cause the computer to perform the processes or functions described in the embodiments of the present application. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable apparatus. 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, from a website site, a computer, a server or a data center to another website site, a computer, a server or a data center through a wired or wireless (for example, infrared, wireless, microwave, etc.) manner. 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, a data center, etc. containing one or more available medium collections. The available medium can be a magnetic medium (for example, a floppy disk, a hard disk, a magnetic tape), an optical medium (for example, a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state disk.

[0130] It should be understood that the size of the sequence number of each process described above in various embodiments of the present application does not mean the order of execution, and the execution order of each process should be determined by its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

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

[0132] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. An adaptive intelligent teaching content recommendation system, characterized by, The data collection and analysis module includes collecting multiple types of data of students in the learning process, including learning behavior data, physiological signals, environmental data and learning achievement cognitive feedback data of students, after collection, multiple types of data are preprocessed by noise reduction, standardization and analysis feature extraction, and multi-modal data are generated; The multi-modal knowledge diagnosis module: based on multi-modal data, a three-dimensional knowledge graph is constructed by using Neo4j graph database, and a deep learning TCN network is used to analyze the three-dimensional knowledge graph to obtain knowledge short board diagnosis results, including identifying explicit knowledge defects and diagnosing implicit knowledge gaps; The construction logic of the three-dimensional knowledge graph is as follows: According to the multi-modal data Mtd', the entity, attribute and relationship information of knowledge is extracted, wherein the extraction knowledge includes the learning time and the problem solving efficiency of different knowledge points in the learning behavior data Lbd, the light intensity and the environmental noise of the learning place environmental data Emd, the learning achievement and the knowledge point mastery in the learning achievement Loc and the cognitive feedback data Cfb; Three dimensions of the three-dimensional knowledge graph are defined, including the knowledge point dimension, the error type dimension and the thinking path dimension, wherein the knowledge point dimension includes the subject knowledge point ontology and the hierarchical relationship, the error type dimension includes the error types and the association of calculation error, concept confusion and step omission, and the thinking path dimension is the spatio-temporal evolution of the cognitive path of students learning knowledge points; The entity, attribute and relationship information of knowledge extracted from the multi-modal data are respectively mapped to the three dimensions of the knowledge point dimension, the error type dimension and the thinking path dimension of the three-dimensional knowledge graph; Using Neo4j graph database, entities, attributes are taken as nodes, and relations are taken as edges to construct a three-dimensional knowledge graph, and to calculate knowledge point correlation strength, knowledge point and error type correlation strength, knowledge point and thinking path correlation strength and error propagation influence, wherein the calculation formula of knowledge point correlation strength is , and , wherein S ij represents the correlation strength of knowledge point i and knowledge point j, C ij represents the co-occurrence times of knowledge point i and knowledge point j in error records, C i represents the single occurrence times of knowledge point i in error records, C j represents the single occurrence times of knowledge point j in error records, λ represents a cross-level attenuation factor, and λ∈(0,1), L i represents the difficulty level of knowledge point i, L j represents the difficulty level of knowledge point j, and N + represents a positive integer; The calculation formula of the association strength between knowledge points and error types is , wherein S ei represents the association strength between the error type e and the knowledge point i, C ei represents the co-occurrence times of the error type e and the knowledge point i, C e represents the total times of the error type e appearing alone, and L e represents the average difficulty level of the error type e. The calculation formula of the association strength between the knowledge points and the thinking paths is , wherein S ip represents the association strength between the knowledge point i and the thinking path p, C ip represents the number of times that the knowledge point i appears in the thinking path p, C' i represents the total number of times that the knowledge point i appears in all paths, C p represents the total number of steps of the path p, and γ represents a path depth attenuation factor, γ ∈ (0, 1), D ip represents the depth of the knowledge point i in the path p. The calculation formula of the influence of error propagation is , wherein I e represents the influence index of the error type e on the knowledge system, i and I represent the knowledge point i and the total number of knowledge points, respectively ei represents the association strength of the error type e and the knowledge point i, F i represents the error frequency of the knowledge point i, F total represents the total number of errors in the system; The interest-driven recommendation module: based on semantic analysis, the learning interest information of multi-modal data is extracted, the explicit and implicit multi-dimensional interest model is established, the knowledge short board diagnosis result is combined, the content recommendation algorithm is used to adaptively select the teaching content which meets the students' interest and can make up the knowledge gap for recommendation; The neural memory enhancement module: based on the memory prediction model of Ebbinghaus forgetting curve, the individual memory decay law is predicted combined with the learning time and environmental data of students, and the review plan is dynamically adjusted according to the review feedback of students, which is coordinated with the adaptive decision engine to trigger content regeneration; The multi-source data fusion and adaptive decision engine module: the feature data of knowledge short board diagnosis result, learning interest information and individual memory decay law are integrated by using graph attention fusion algorithm, and the long-term learning benefit is optimized by reinforcement learning, and the dynamic decision of content recommendation and review plan combination is adaptively generated.

2. The self-adaptive intelligent teaching content recommendation system according to claim 1, wherein, The collection steps of the multiple types of data are as follows: The learning behavior data of students is recorded as Lbd through the log recording function of intelligent teaching APP or the embedded HTML5VideoAPI in intelligent teaching webpage; The physiological signals of students are collected in real time by using smart bracelet, smart glasses or wearable devices for medical physiological monitoring, and are marked as Pss; The environmental data of the learning place of students are collected in real time by deploying light sensor, microphone and UWB ultra-wideband positioning chip on the hardware mobile terminal of intelligent learning, and are marked as Emd; After learning the intelligent teaching content, the learning achievements of the students are counted and labeled as Loc, and the cognitive feedback voice information of the students on the intelligent learning is recorded using the WebAudioAPI and labeled as Cfb; The expression of the multi-type data is wherein Mtd represents the multi-type data; The generation of the multi-modal data is as follows: A hybrid Gaussian-wavelet threshold denoising model is developed to adaptively denoise the multi-type data, and the denoised data is standardized by Z-score standardization to obtain standardized data; Behavioral features, physiological features, environmental features and cognitive features are extracted from the standardized data using principal component analysis (PCA), and a feature vector is constructed; After time synchronization and alignment of the feature vector according to the time stamp, multi-modal feature embedding fusion is performed, and multi-modal fusion data is outputted; The multi-modal fusion data is reduced in dimension by using a t-SNE projection method, multi-modal data is generated, and visual display is performed, wherein the multi-modal data is calibrated as Mtd', and , wherein Lbd', Pss', Emd', Loc', and Cfb' represent pre-processed learning behavior data, physiological signals, environmental data, and learning achievement cognitive feedback data, respectively.

3. The self-adaptive intelligent teaching content recommendation system of claim 2, wherein, The acquisition of the knowledge short board diagnosis result is as follows: For explicit knowledge defects, the DBSCAN algorithm is used to cluster error records and induce error patterns, wherein the core point condition formula of the DBSCAN algorithm is , MinPts=5, wherein represents the number of neighborhood samples of the multi-modal data Mtd' sample within a neighborhood radius , represents a distance threshold value for defining the neighborhood range of the sample, and MinPts represents the minimum number of neighborhood samples required to become a core point. The influence degree of the dominant knowledge defect is identified by calculating a defect severity score, wherein a calculation formula of the defect severity score is , and SC i ∈(0, 1], in the formula, SC i represents a defect score of the knowledge point i, and when SC i > 0.6, it is determined as a serious defect, α represents a weight coefficient of a frequency F i of errors on the knowledge point i accounts for a total number F total of errors of the system, and β represents a weight coefficient of an influence index I e of the error type e on the knowledge system. For the implicit knowledge vulnerability, the three-dimensional knowledge graph is inputted into a deep learning time convolution network (TCN), and the dilated convolution of the TCN network increases the receptive field by expansion rate to extract features of the nodes and edges in the knowledge graph, capture the time dependence and structure information between knowledge points, and output a potential defect sequence of the knowledge; The capsule network classifies an output error type distribution and traces a three-dimensional knowledge graph to diagnose knowledge point correlation strength and calculate a dynamic route to identify implicit knowledge vulnerabilities, wherein a calculation formula of the dynamic route is , and , wherein d xy represents a coupling coefficient from a low-layer capsule x to a high-layer capsule y, b xy represents a logarithmic prior probability in dynamic route iteration, z represents a total number of layers of the capsule network, represents a prediction vector of the high-layer capsule y, and ω xy represents a learnable transformation matrix for mapping a low-layer capsule output u x to a high-layer space, and u x represents an output vector of the low-layer capsule x. The reuse defect propagation algorithm is used to calculate the propagation probability of the implicit knowledge vulnerability, wherein a calculation formula of the propagation probability is , and , wherein P d represents the propagation probability of the potential defect d, represents a knowledge point set directly associated with the defect d, S di represents the association strength of the defect d and the knowledge point i, C di represents the association strength of the defect d and the knowledge point i, C d represents the number of times of identifying and diagnosing the single appearance of the implicit defect d, C'' i represents the number of times of the single appearance of the knowledge point i in the implicit defect, SC i represents the defect score of the knowledge point i.

4. The self-adaptive intelligent teaching content recommendation system of claim 3, wherein, The establishment of the explicit and implicit multi-dimensional interest model is as follows: The learning content title and description are marked through semantic analysis, and the explicit interest information associated with the learning interest is directly extracted from the multi-modal data Mtd'; According to the hierarchical clustering of the historical learning content of the student, an explicit interest model is constructed, the learning theme preference is quantified, and the preference score of the explicit interest is calculated, wherein the calculation formula of the preference score is , wherein S K represents the preference score of the explicit interest of the theme K, H represents the knowledge content, H K represents the set of knowledge content H belonging to the theme K, dt H represents the stay time of the knowledge content H, max(dt) represents the maximum value of the stay time in all knowledge contents, and cc K represents the number of clicks of the theme K. According to the gaze duration, the note word number and the revisit number, a cognitive engagement index is calculated to mine potential implicit interest information and construct an implicit interest model, wherein the calculation formula of the cognitive engagement index is , and ω1+ω2+ω3=1, wherein CEI represents the cognitive engagement index, ω1, ω2 and ω3 all represent weight coefficients, a1, a2 and a3 respectively represent the gaze duration, the note word number and the revisit number, A1, A2 and A3 respectively represent the total duration of learning content, the maximum value of note word numbers of all users and the maximum value of revisit numbers of all users; The explicit interest information and the implicit interest information are interactively integrated, and interest weights are distributed to reflect the importance of the explicit and implicit interest, wherein a calculation formula of the explicit and implicit interest weights is , and ω E +ω1=1, wherein ω m represents a weight of the mth interest point, m represents a number of the interest points, ω E represents a weight coefficient corresponding to the explicit interest score, E m represents the explicit interest score of the mth interest point, ω1 represents a weight coefficient corresponding to the implicit interest score, and I m represents the implicit interest score of the mth interest point.

5. The self-adaptive intelligent teaching content recommendation system of claim 4, wherein, The prediction of the individual memory decay law is as follows: The length of browsing the course page, the time and accuracy of doing the questions, the heart rate variability and cognitive load index of the physiological signal Pss', and the light intensity and environmental noise of the environmental data Emd' of the student's learning behavior data Lbd' are obtained from the multi-modal data Mtd' to calculate the learning quality factor and the environmental interference coefficient, and the personalized memory strength parameter is optimized and fitted by using maximum likelihood estimation through the historical learning data, wherein the calculation formula of the learning quality factor is , and ω Q1 + ω Q2 = 1, wherein Q0 represents the learning quality factor, ω Q1 represents the weight coefficient corresponding to the accuracy, T% represents the accuracy, t 反 represents the reaction time, t 基 represents the historical average reference time of solving the questions, and ω Q2 represents the weight coefficient of the reaction time of the questions to the reference. The calculation formula of the environmental interference coefficient is , and , wherein, EI represents the environmental interference coefficient, Lux t represents the illumination intensity at the current time t, Lux max represents the upper limit of the comfortable illumination, ω Lux represents the weight coefficient of the illumination intensity, W t represents the noise decibel value at the current time t, W crit represents the attention distraction critical value, ω w represents the weight coefficient of the environmental noise; The calculation formula for optimizing the personalized memory strength parameter using maximum likelihood estimation is , wherein S0 represents the optimized personalized memory strength parameter, argmax represents maximizing the likelihood function, S represents the personalized memory strength parameter, represents the actual measurement value of the nth observation data point, represents the probability density function of the Gaussian distribution, represents the personalized memory strength parameter S and the time point t n' the generated model prediction value; Based on the Ebbinghaus forgetting curve, an improved memory prediction model is constructed by combining a learning quality factor Q0, an environmental interference coefficient EI and a personalized memory strength parameter S0, wherein an expression of the memory prediction model is , wherein R(t) represents a memory retention rate after time t, e represents an exponential function, t represents a current time, and S represents a personalized memory strength parameter. Through LSTM network learning training, a target of minimizing prediction error is defined, and a dynamic correction memory prediction model parameter is performed, wherein a calculation formula of the dynamic correction parameter is , wherein R'(t) represents a memory retention rate after time t after dynamic correction of the LSTM network, LSTM represents dynamic learning prediction of the LSTM network, all represent interference factors quantifying the influence on memory.

6. The self-adaptive intelligent teaching content recommendation system according to claim 5, wherein, The integration of the feature data by the graph attention fusion algorithm is as follows: The knowledge short board diagnosis result, the learning interest information and the feature data of the individual memory decay law are used as nodes to construct a heterogeneous graph by using the graph attention fusion algorithm; A multi-head graph attention mechanism (GAT) is used to calculate the correlation weight between nodes; According to the correlation weight, the features of adjacent nodes are aggregated by vector concatenation operation, and the feature representation of each node is updated by using an ELU activation function; After the feature data of all nodes are fused, the integrated feature data representation is outputted.

7. The self-adaptive intelligent teaching content recommendation system of claim 6, wherein, The adaptive generation logic flow of the dynamic decision is as follows: The integrated feature data representation is inputted as the state of the reinforcement learning network; An action space is defined, including a content recommendation action and a review plan action; A reward function is designed to evaluate the impact of the content recommendation action and the review plan action on the long-term learning benefit, wherein the long-term learning benefit includes learning achievement improvement and knowledge mastery degree; An on-policy optimization algorithm (PPO) is used to learn the optimal decision strategy, and an action that can maximize the long-term reward is selected according to the current feature data state; According to the learned optimal decision, a dynamic decision of the combination of content recommendation and review plan is adaptively generated according to the current state at each time step.

Images