An AI-enabled intelligent teaching personalized service method and system

Abstract

The present invention provides an AI-enabled smart teaching personalized service method and system, which relates to the field of smart teaching technology. The method includes periodically collecting multimodal data of current students; performing association analysis on learning data and behavioral data in the multimodal data to generate analysis results; constructing a weight matrix based on group labels and ability prediction models in the analysis results, and dividing all knowledge nodes into hierarchical courses based on the weight matrix; determining the core knowledge proportion and the extended knowledge proportion based on the ability prediction model and the style of study vector, and fusing them to generate a fusion ratio; screening out intensive training sequences and extended content sequences from the core module and the extended module respectively, and cross-arranging the intensive training sequences and the extended content sequences according to the fusion ratio to generate a dynamic learning path. The present invention achieves precision and personalization of teaching services through multimodal data fusion, deep association analysis, hierarchical course construction and dynamic path planning.

Description

Technical Field

[0001] The present invention relates to the field of smart teaching technology, and in particular to an AI-enabled smart teaching personalized service method and system. Background Art

[0002] The traditional education model has long been centered around "standardized teaching," using fixed teaching content, teaching schedules, and assessment standards. This makes it difficult to adapt to the significant differences among students in terms of learning ability, knowledge base, and interests. This is specifically manifested in the following ways:

[0003] Teachers need to take into account the overall level of the class, which results in some students "not getting enough" (those with stronger abilities lack challenges) or "not keeping up" (those with weak foundations find it difficult to digest).

[0004] The course design lacks the ability to dynamically adjust and cannot optimize the learning sequence based on students' real-time learning status;

[0005] Teaching decisions often rely on experience-based judgments and lack quantitative analysis of students' learning behaviors and knowledge mastery.

[0006] The traditional model makes it difficult to achieve "teaching students in accordance with their aptitude", resulting in uneven teaching results and limited individual development of students.

[0007] Therefore, it is necessary to provide an AI-enabled intelligent teaching personalized service method and its system to solve the above technical problems. Summary of the Invention

[0008] To solve the above technical problems, the present invention provides an AI-enabled intelligent teaching personalized service method and system, which realizes the precision and personalization of teaching services through multimodal data fusion, deep correlation analysis and dynamic path planning.

[0009] The present invention provides an AI-enabled intelligent teaching personalized service method, the method comprising the following steps:

[0010] Periodically collecting multimodal data of the current student, wherein the multimodal data includes learning data, behavioral data, and a knowledge graph, wherein the knowledge graph is composed of multiple knowledge nodes containing difficulty coefficients and dependency relationships;

[0011] Performing association analysis on the learning data and the behavior data to generate analysis results, wherein the association analysis includes cluster classification, time series prediction, and vector mapping, and the analysis results include group labels, ability prediction models, and learning style vectors corresponding to the association analysis;

[0012] Constructing a weight matrix according to the group labels and the ability prediction model, and dividing all knowledge nodes into hierarchical courses according to the weight matrix, wherein the hierarchical courses include core modules and expansion modules;

[0013] Based on the ability prediction model and the study style vector, the core knowledge proportion and the extended knowledge proportion are determined respectively and integrated to generate an integrated ratio;

[0014] Reinforced training sequences and expanded content sequences are respectively selected from the core module and the expanded module, and the reinforced training sequences and the expanded content sequences are cross-arranged according to the fusion ratio to generate a dynamic learning path.

[0015] Preferably, the periodic collection of multimodal data of the current student includes:

[0016] Periodically collect learning data of current students, wherein the learning data includes performance indicators, progress records, class grades, and homework quality grading;

[0017] Periodically collect the current student's behavioral data, where the behavioral data includes a semantic feature vector composed of interactive text and platform logs, and an interest intensity value;

[0018] A knowledge graph consisting of multiple knowledge nodes is established based on the subject knowledge system.

[0019] Preferably, performing correlation analysis on the learning data and the behavior data to generate analysis results includes:

[0020] Cluster classification: Based on the performance indicators and classroom scores in the learning data, combined with the semantic feature vectors generated by the interactive text and platform logs in the behavioral data, the group labels are generated through a spectral clustering algorithm;

[0021] Time series prediction: using the progress records in the learning data, combined with the difficulty coefficients and dependencies of the knowledge nodes, to construct a time series prediction model and output the ability prediction model;

[0022] Vector mapping: Analyze the interest intensity value in the behavioral data and the homework quality grading score in the learning data, and generate the learning style vector through a cross-modal association algorithm.

[0023] Preferably, the weight matrix construction process specifically includes:

[0024] Calculate the group mastery weight of each knowledge node according to the group distribution density of the group label;

[0025] Based on the probability distribution of weak links in the capability prediction model, generating individual urgency weights of each knowledge node;

[0026] The group mastery weight and the individual urgency weight are combined into a weight matrix.

[0027] Preferably, the process of dividing the courses into different levels specifically includes:

[0028] Sort the knowledge nodes according to the weight matrix, and define the knowledge nodes that exceed the preset core knowledge threshold as core modules;

[0029] After evaluating the remaining knowledge nodes based on their dependencies and difficulty coefficients with the knowledge nodes of the core module, the knowledge nodes that meet the evaluation conditions are divided into expansion modules.

[0030] Preferably, the process of generating the fusion ratio specifically includes:

[0031] Extracting the urgency coefficient of the weak links of each knowledge node in the capability prediction model and generating the core knowledge proportion through normalization processing;

[0032] Extracting the interest relevance weight of the study style vector and adjusting it by a preset interest gain coefficient to generate an expanded knowledge proportion;

[0033] The core knowledge proportion and the extended knowledge proportion are integrated to generate an integrated ratio, where the integration formula is:

[0034]

[0035]

[0036] in, The proportion of core knowledge, To expand the knowledge share, is the integration ratio of core knowledge, To expand the integration ratio of knowledge.

[0037] Preferably, the process of generating the dynamic learning path includes:

[0038] From the core module, N nodes are selected in descending order of the urgency coefficient of the weak links of the knowledge nodes to generate a reinforcement training sequence;

[0039] From the expansion module, M nodes are selected in descending order of interest relevance weight of the knowledge nodes to generate an expansion content sequence;

[0040] According to the fusion ratio, the reinforcement training sequence and the expansion content sequence are cross-arranged according to the arrangement rules to generate a dynamic learning path;

[0041] The arrangement rule is to allocate Core nodes are inserted expansion nodes, wherein the core node is a knowledge node in the core module, and the expansion node is a knowledge node in the expansion module, is the ceiling function.

[0042] Preferably, the process of determining N and M includes:

[0043] The total number of nodes in a single learning path is preset to T, where the total number of nodes T is dynamically set according to the average daily effective learning time of the current student;

[0044] N and M are calculated according to the fusion ratio and the total number of nodes using the formula:

[0045]

[0046]

[0047] in, , is the ceiling function.

[0048] The present invention also provides an AI-enabled intelligent teaching personalized service system for executing the AI-enabled intelligent teaching personalized service method. The system includes:

[0049] A multimodal data collection module, configured to periodically collect multimodal data of the current student, wherein the multimodal data includes learning data, behavioral data, and a knowledge graph, wherein the knowledge graph is composed of multiple knowledge nodes containing difficulty coefficients and dependency relationships;

[0050] an association analysis module, configured to perform association analysis on the learning data and the behavioral data to generate analysis results, wherein the association analysis includes cluster classification, time series prediction, and vector mapping, and the analysis results include group labels, ability prediction models, and learning style vectors corresponding to the association analysis;

[0051] A knowledge node stratification module is used to construct a weight matrix based on the group labels and the ability prediction model, and to divide all knowledge nodes into stratified courses based on the weight matrix, wherein the stratified courses include core modules and expansion modules;

[0052] A knowledge proportion fusion module is used to determine the core knowledge proportion and the extended knowledge proportion based on the ability prediction model and the study style vector, and fuse them to generate a fusion ratio;

[0053] The dynamic path generation module is used to filter out the enhanced training sequence and the extended content sequence from the core module and the extended module respectively, and cross-arrange the enhanced training sequence and the extended content sequence according to the fusion ratio to generate a dynamic learning path.

[0054] Compared with related technologies, the AI-enabled intelligent teaching personalized service method and system provided by the present invention have the following beneficial effects:

[0055] This invention systematically integrates multimodal data (covering learning data, behavioral data, and knowledge graphs) and builds a unified data representation system. Combined with AI deep correlation analysis technology, it can accurately mine the implicit correlations between individual students' learning abilities, knowledge mastery, and interest preferences, thereby generating highly targeted tiered courses and dynamic learning paths.

[0056] Specifically, the program quantifies the group mastery and individual urgency of knowledge nodes through a weight matrix, realizes the differentiated design of core knowledge modules and extended knowledge modules, and dynamically adjusts the path content based on students' real-time learning status to ensure the organic combination of intensive training of weak links and interest expansion, and ultimately forms a closed-loop optimization mechanism of "evaluation-feedback-adjustment".

[0057] Compared with the traditional education model, this invention significantly improves the precision and intelligence of teaching services, effectively solves the practical problem of "teaching students in accordance with their aptitude", provides quantifiable technical support for the digital transformation of education, and ultimately promotes the paradigm shift of smart education from "experience-driven" to "data-driven". BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 A flowchart of an AI-enabled intelligent teaching personalized service method provided by the present invention;

[0059] Figure 2 This is a module structure diagram of the AI-enabled intelligent teaching personalized service system provided by the present invention. DETAILED DESCRIPTION

[0060] The present invention will be further described in detail below with reference to the accompanying drawings and examples. It will be understood that the specific embodiments described herein are intended only to illustrate the present invention and are not intended to limit the present invention. It should also be noted that, for ease of description, the accompanying drawings only illustrate portions relevant to the present invention, not all of the structures. Furthermore, the embodiments of the present invention and the features of the embodiments may be combined with one another unless there is a conflict.

[0061] It should also be noted that, for ease of description, only portions relevant to the present invention are shown in the accompanying drawings, rather than all of the contents. Before discussing the exemplary embodiments in more detail, it should be noted that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the various operations (or steps) as sequential processes, many of the operations can be performed in parallel, concurrently, or simultaneously. In addition, the order of the various operations can be rearranged. The process can be terminated when its operations are completed, but may also have additional steps not included in the accompanying drawings. The process can correspond to a method, function, procedure, subroutine, subprogram, etc.

[0062] Example 1

[0063] This invention provides an AI-enabled intelligent teaching personalized service method, referring to Figure 1 As shown, the method includes the following steps:

[0064] S1: Periodically collect multimodal data of the current student, wherein the multimodal data includes learning data, behavioral data and a knowledge graph, wherein the knowledge graph is composed of multiple knowledge nodes containing difficulty coefficients and dependency relationships.

[0065] Specifically, step S1 includes the following contents:

[0066] Periodically collect learning data of current students, wherein the learning data includes performance indicators, progress records, class scores and homework quality grading.

[0067] In this embodiment, more specifically, the features included in the learning data are defined as follows:

[0068] Performance Indicators: Student test score data is extracted from the teaching management system, including the score percentage for each knowledge point, the distribution of error types, and the time taken to complete the test. The score percentage is calculated as: number of correctly answered questions / total number of questions × 100%. Error types include conceptual errors, calculation errors, and question-reading errors, and are identified through teacher annotation.

[0069] Progress Recording: A logging system records students' learning trajectory for each knowledge node, including the first learning time, number of review sessions, duration of each study session, and the number of days between sessions. Learning progress data is stored in a time series format, with the type of learning action (preview, study, review) and duration recorded at each time point.

[0070] Classroom Scoring: Instructors grade students based on their performance in class. Scoring criteria include: number of active questions asked (0.5 points each, capped at 3 points), accuracy of answers (1-3 points each), and contribution to group discussions (1-5 points). Scoring data is stored in association with specific knowledge nodes.

[0071] Assignment Quality Grading: The assignment grading system automatically grades assignments based on pre-set criteria. Objective questions are scored based on accuracy, while subjective questions are graded based on completeness (40%), accuracy (40%), and standardization (20%). Assignments are graded into four grades: A (85-100 points), B (70-84 points), C (60-69 points), and D (0-59 points).

[0072] The current student's behavior data is collected periodically, wherein the behavior data includes a semantic feature vector composed of interactive text and platform logs, and an interest intensity value.

[0073] In this embodiment, more specifically, the features included in the learning data are defined as follows:

[0074] Interactive text processing: We extract text from online learning platform discussion forums, Q&A areas, and private chat logs. We first clean the data to remove non-text content such as advertisements and emoticons. We then use a pre-trained Chinese BERT model to convert the text into a 768-dimensional semantic vector. Finally, we perform PCA dimensionality reduction to obtain a 50-dimensional semantic feature vector.

[0075] Platform log analysis: We collect student activity logs on the learning platform, including resource access records (video viewing duration, document downloads), interactive behaviors (page dwell time, click heatmaps), and system operations (saving, sharing, and taking notes). We then compile and encode these behavioral data to generate behavioral feature vectors.

[0076] Calculation of interest intensity: Combining explicit interest data (students' interest questionnaire scores, 1-5 points) and implicit interest data (resource access frequency, interactive participation), the interest intensity value is calculated using a weighted formula: Interest intensity value = questionnaire score 0.6 + Standardized Visit Frequency 0.3+Interactive Engagement 0.1. Standardized visit frequency = actual number of visits / average number of visits in the class.

[0077] A knowledge graph consisting of multiple knowledge nodes is established based on the subject knowledge system.

[0078] In this embodiment, the definition of knowledge nodes is as follows: Based on the subject curriculum standards, knowledge points are decomposed into the smallest teaching units as knowledge nodes. Each node contains the following attributes:

[0079] Node ID: A unique identifier, such as "math_k001".

[0080] Node name: the name of the knowledge point, such as "quadratic function graph".

[0081] Cognitive difficulty: Based on Bloom's taxonomy, it is divided into six levels: memory (1), understanding (2), application (3), analysis (4), synthesis (5), and evaluation (6).

[0082] Dynamic difficulty coefficient: dynamically calculated based on students' actual performance, the formula is: 1 (Class average score 0.7 Average learning efficiency 0.3).

[0083] Dependency establishment: Analyze the logical relationship between knowledge points and construct a directed acyclic graph.

[0084] Prerequisite relationship: prerequisite knowledge points that must be mastered.

[0085] Parallel relationship: knowledge points that can be learned in parallel.

[0086] Advanced relationship: in-depth knowledge points on the same topic.

[0087] Dependencies are annotated by subject experts and verified through teaching practice data.

[0088] Knowledge Graph Storage: A graph database is used to store the knowledge graph. Each node stores the aforementioned attributes, and edges store the relationship type and strength. The dynamic difficulty coefficient is updated weekly. A change of more than 10% triggers a change in the teaching strategy.

[0089] S2: Perform association analysis on the learning data and behavior data to generate analysis results, wherein the association analysis includes cluster classification, time series prediction and vector mapping, and the analysis results include group labels, ability prediction models and academic style vectors corresponding to the association analysis.

[0090] Specifically, step S2 includes the following steps:

[0091] Cluster classification: Based on the performance indicators and classroom scores in the learning data, and combined with the semantic feature vectors generated by the interactive text and platform logs in the behavioral data, the group labels are generated through a spectral clustering algorithm.

[0092] More specifically, during the clustering and classification process, the performance indicators in the learning data were first standardized, converting each subject's grades into a unified Z-score. Classroom rating data was then aggregated by knowledge node. Semantic feature vectors were then converted to three-dimensional data using dimensionality reduction techniques for easier analysis.

[0093] A spectral clustering algorithm was used to construct a student similarity matrix. A Gaussian kernel function was used to calculate the similarity between student eigenvectors, and a similarity threshold was set to filter out low-correlation data. By calculating the Laplace matrix and performing eigendecomposition, key eigenvectors were selected for K-means clustering. The number of clusters was dynamically determined using the silhouette coefficient method. Finally, meaningful group labels were generated for each cluster, such as "theoretical" and "practical," and the typical characteristic distribution and recommended teaching strategies for each category were recorded.

[0094] Time series prediction: Utilizing the progress records in the learning data, combined with the difficulty coefficients and dependencies of the knowledge nodes, a time series prediction model is constructed to output the capability prediction model.

[0095] In this embodiment, the time series prediction stage extracts key time series features from the learning progress records, including dynamic information such as sliding window statistics and knowledge node conversion patterns.

[0096] A predictive model based on the Transformer architecture is constructed by combining static features such as node difficulty coefficients and the mastery of preceding nodes in the knowledge graph. This model uses an encoder structure with four Transformer blocks and eight attention heads. It takes a 64-dimensional feature vector as input and outputs a probability distribution of mastery over the next three learning cycles. The training process uses a weighted cross-entropy loss function to assign greater weight to weak link predictions. The AdamW optimizer is used with a learning rate decay strategy and an early stopping mechanism to prevent overfitting. The model ultimately outputs the probability of a weak link and the urgency coefficient for each knowledge node, providing a basis for subsequent teaching decisions.

[0097] Vector mapping: Analyze the interest intensity value in the behavioral data and the homework quality grading score in the learning data, and generate the learning style vector through a cross-modal association algorithm.

[0098] In this embodiment, the vector mapping process realizes cross-modal data alignment by constructing a dual-tower model, in which the interest tower processes the interest intensity value and the effectiveness tower processes the work quality data, and the two are jointly represented in a 32-dimensional shared embedding space.

[0099] The model is trained using a contrastive learning approach, using pairs of interest-assignment data from the same student as positive samples and randomly sampled data as negative samples. A modified loss function is used to optimize model parameters. The resulting learning style vector contains three key dimensions: the correlation coefficient between interest and achievement, reflecting the degree of alignment between learning outcomes and interests; the distribution of interest preferences across knowledge areas, indicating students' interests; and learning style classification, identifying students as visual, auditory, or hands-on learners. These vector features provide important insights for generating personalized learning paths.

[0100] S3: Construct a weight matrix according to the group label and the ability prediction model, and divide all knowledge nodes into hierarchical courses according to the weight matrix, wherein the hierarchical courses include core modules and expansion modules.

[0101] Specifically, the construction process of the weight matrix in step S3 specifically includes:

[0102] First, the group mastery weight of each knowledge node is calculated according to the group distribution density of the group labels.

[0103] In this embodiment, the group mastery weight reflects the overall mastery of a knowledge node by a specific student group.

[0104] First, determine the group category to which the student belongs based on the group label, and then count the historical performance data of all students in the group at each knowledge node.

[0105] The calculation process involves extracting the average score rate for each knowledge node among students in the group, combining it with a weighted average of class participation (with a weight of 0.7 for score rate and 0.3 for participation), and normalizing it to a weight between 0 and 1 using a sigmoid function. For new knowledge nodes or insufficient data, the weights of similar knowledge nodes are interpolated for estimation. The final output, a group mastery weight matrix, contains the weight values ​​for each knowledge node. Higher values ​​indicate a better mastery of the knowledge point by the group.

[0106] Next, based on the probability distribution of weak links in the capability prediction model, the individual urgency weight of each knowledge node is generated.

[0107] In this embodiment, the individual urgency weight represents the learning urgency of a specific student at each knowledge node. Based on the probability distribution of weak links output by the ability prediction model, the probability values ​​are first standardized to eliminate the dimensional differences between different knowledge fields. Then, the dynamic difficulty coefficient of the knowledge node is combined to perform weighted adjustment. The adjustment coefficient is (That is, the higher the difficulty, the greater the urgency weight.) Furthermore, considering the temporal characteristics of learning progress, knowledge nodes with recently declining learning outcomes are assigned an additional urgency increment of 0.1 to 0.3. In the resulting individual urgency weight matrix, the weight of each knowledge node ranges from 0 to 1.5, with larger values ​​indicating greater urgency and importance for the current student.

[0108] Finally, the group mastery weight and the individual urgency weight are combined into a weight matrix.

[0109] In this embodiment, the group mastery weight and the individual urgency weight are combined into a unified weight matrix.

[0110] The combination process adopts a dynamic weighting strategy: for basic knowledge nodes (number of dependencies ≥ 3), the group weight accounts for 60% and the individual weight accounts for 40%.

[0111] For extended knowledge nodes (dependency number ≤ 1), the group weight accounts for 30% and the individual weight accounts for 70%; other knowledge nodes adopt a balanced ratio of 50% to 50%.

[0112] The combined weight matrix is ​​normalized to ensure that the final weight value of each knowledge node is between 0 and 1. This matrix not only reflects the importance ranking of knowledge points, but also retains the original weight source information for subsequent strategic adjustments when dividing tiered courses.

[0113] Specifically, the process of dividing the courses into different levels in step S3 includes:

[0114] First, the knowledge nodes are sorted according to the weight matrix, and the knowledge nodes that exceed the preset core knowledge threshold are defined as core modules.

[0115] More specifically, the detailed implementation process of the core module division is as follows:

[0116] First, calculate the quartiles of all knowledge node weights in the weight matrix, determine the first quartile (Q1) and the third quartile (Q3), and calculate the interquartile range (IQR = Q3 - Q1). The core knowledge threshold is set to Q3 + 1.5 IQR, the algorithm can adapt to data distribution and avoid interference from extreme values.

[0117] In the initial node screening phase, all knowledge nodes are sorted in descending order of weight, and candidate nodes with weights exceeding the threshold are selected. For example, if Q3 = 0.65 and IQR = 0.2, the threshold is 0.65 + 0.3 = 0.95, and nodes with weights greater than or equal to 0.95 are included in the candidate list.

[0118] Then carry out the three-level verification mechanism. The specific verification process includes:

[0119] Predependency Verification: Traverse all the candidate node's predependency nodes and calculate their average weight. If the average weight of any predependency node is less than 0.4, the candidate node is marked as "missing dependency" and requires additional learning of the predependency content before inclusion in the core module.

[0120] Verification of Difficulty Adaptability: An integrated model (LSTM + Graph Attention Network + Gradient Boosting Tree) is used to predict the probability of a student mastering a node. The model inputs include: time series data on the student's historical learning behavior, knowledge graph relationship features, and the dynamic difficulty coefficient of the node.

[0121] Verification criteria: The predicted mastery probability must be within ±0.3 standard deviations of the student's current ability level. For example, if the student's mean ability is 0.7 and the standard deviation is 0.1, the node difficulty coefficient must be between 0.4 and 1.0.

[0122] Course Foundation Verification: This verifies the foundational level of each node by connecting to the course standards library. Only nodes marked as "Basic (Level 1)" or "Advanced (Level 2)" are allowed to pass, while "Expanded (Level 3)" nodes are excluded.

[0123] Nodes that pass all verifications are marked as "core modules" and are highlighted in red in the knowledge graph.

[0124] Next, the remaining knowledge nodes are evaluated based on the dependency relationship and difficulty coefficient between the knowledge nodes and the core module, and the knowledge nodes that meet the evaluation conditions are divided into expansion modules.

[0125] In this embodiment, the depth assessment method of the expanded module division includes:

[0126] First, perform dependency analysis: construct a knowledge node association graph and use the Dijkstra algorithm to calculate the shortest path length from the remaining nodes to the core module.

[0127] Direct dependencies (those with a clear prerequisite relationship with the core node) are assigned a basic weight of 0.7; the weight of indirect dependencies decays according to the path length, using the formula: For example, the indirect dependency weight for a path length of 2 is ,in is a natural constant (approximately equal to 2.71828), Represents a natural constant Exponential operation.

[0128] Then, difficulty adaptability evaluation is performed: the integrated model predicts the probability of students mastering the remaining nodes. The model training process includes:

[0129] Data preprocessing: SMOTE-ENN hybrid sampling is used for unbalanced data, and the time series features are aligned through dynamic time warping.

[0130] Training strategy: First pre-train the graph attention network (learning rate 0.001, epochs = 50), and then jointly fine-tune the LSTM and gradient boosting tree (learning rate 0.0005, early stopping patience = 10).

[0131] Mark high-risk nodes: Nodes with a predicted probability less than 0.4 trigger a "high difficulty warning" and require priority allocation of auxiliary teaching resources.

[0132] Next, we perform interest correlation detection: Based on collaborative filtering technology, we calculate the cosine similarity between the student interest vector and the node content. For example, the student interest vector is , the node content vector is , then the similarity is:

[0133] .

[0134] Combined with the class hot spot data, an additional weighting coefficient of 0.2 is added to popular nodes (visit frequency is more than twice the class average).

[0135] Finally, comprehensive scoring and teaching review will be conducted, including:

[0136] Extended score calculation: 0.5 Dependency Weight 0.3 (1-prediction difficulty coefficient) 0.2 Interest similarity. For example, if a node has a dependency weight of 0.6, a difficulty coefficient of 0.3, and an interest similarity of 0.8, then the score is 0.5 0.6+0.3 0.7+0.2 0.8=0.67.

[0137] Teaching value review: Nodes with a score greater than or equal to 0.6 are pushed to the teacher's end, displaying their associated core nodes, predicted mastery probability, and class hot spot rankings. After the teacher confirms the content quality, it will be included in the expansion module.

[0138] S4: Based on the ability prediction model and the study style vector, the core knowledge proportion and the extended knowledge proportion are determined respectively and integrated to generate an integration ratio.

[0139] Specifically, the process of generating the fusion ratio in step S4 includes:

[0140] S41: Extracting the weak link urgency coefficient of each knowledge node in the capability prediction model, and generating the core knowledge proportion through normalization processing.

[0141] In this embodiment, the process of generating the core knowledge weight includes data cleaning, segment normalization, and teaching priority calibration.

[0142] First, the urgency coefficient of the weak links of each knowledge node is extracted from the capability prediction model (ranging from 0 to 1.5). Outliers exceeding ±3 standard deviations from the mean (e.g., extreme data with a coefficient greater than 2.0) are eliminated using the Z-score method. Valid data is processed in segments: the high-urgency interval (coefficient greater than or equal to 1.0) is converted to a weight value of 0.8 to 1.0 through linear mapping, and the regular interval (coefficient less than 1.0) is directly normalized to a range of 0 to 1. The weight is then adjusted based on the curriculum standard label of the knowledge node: the weight of basic knowledge points (level 1) is increased by 20%, the weight of expanded knowledge points (level 3) is reduced by 20%, and the weight of improved knowledge points (level 2) remains the same. For example, the original weight of a basic node was 0.8, which was increased to 0.96 after calibration to ensure that key knowledge points are strengthened first.

[0143] S42: extracting the interest relevance weight of the study style vector, and generating the extended knowledge proportion after adjusting it through a preset interest gain coefficient.

[0144] In this embodiment, the generation of the extended knowledge proportion is based on the interest relevance weight of the learning style vector, combined with dynamic adjustment of learning fatigue and class collaborative filtering.

[0145] First, the original weight of interest relevance (0 to 1) is extracted and dynamically adjusted based on student fatigue (average weekly learning interruption rate): the weight is increased by 20% for low fatigue (less than 30%) and decreased by 20% for high fatigue (greater than or equal to 60%). An exponential gain formula (weight × e^(1.5 × weight)) is then used to amplify high interest values. For example, a weight of 0.6 is amplified to 1.475 and then truncated to 1.2. Finally, an additional weight of 0.15 is added to popular content in the class (top 20% in visits and with more than 10 interactions per week) to ensure that interest guidance aligns with group trends. A "Data Analysis" node originally had a weight of 0.9, but this was adjusted to 1.05 to meet the popularity criteria.

[0146] S43: The core knowledge proportion and the extended knowledge proportion are integrated to generate an integrated ratio, wherein the integration formula is:

[0147]

[0148]

[0149] in, The proportion of core knowledge, To expand the knowledge share, is the integration ratio of core knowledge, To expand the integration ratio of knowledge.

[0150] In this embodiment, the fusion ratio is calculated as follows: , expansion ratio . Input data needs to be checked for integrity. If , then reset to 0.5; yes If the value exceeds the limit (>1.2), it will be truncated. A boundary protection mechanism is added to the calculation results: the core ratio is forced to be ≥10%, and the expansion ratio is ≥5% to prevent extreme deviations. For example, when =0.8, D=0.5, =61.5%, = 38.5%, the generated path is cross-arranged with the core and extension content in a ratio of 1.6:1; if =0.3, D=1.2, triggering the lower limit constraint, adjusted to =10%, =90%.

[0151] S5: Filtering out enhanced training sequences and extended content sequences from the core module and the extended module respectively, and cross-arranging the enhanced training sequences and the extended content sequences according to the fusion ratio to generate a dynamic learning path.

[0152] Specifically, before step S5 begins, it is necessary to determine the screening numbers N and M. The determination process includes:

[0153] First, the total number of nodes of a single learning path is preset to T, where the total number of nodes T is dynamically set according to the average daily effective learning time of the current student.

[0154] In this embodiment, the total number of nodes T is determined based on the average daily effective learning time of students using a standardized calculation method.

[0155] First, extract the student's learning data from the last seven days from the learning platform log and calculate their average daily effective learning time (excluding any non-learning interruptions exceeding 5 minutes). For example, a student's average daily effective learning time is 120 minutes. Then, use the preset standard learning time of 40 minutes per node as the benchmark value. This time has been verified by educational experiments and can balance learning effect and cognitive load. The total number of nodes T is calculated by dividing the average daily effective learning time by the standard time and rounding down. For example, 120 minutes corresponds to 3 nodes (120 / 40=3).

[0156] Secondly, N and M are calculated according to the fusion ratio and the total number of nodes using the formula:

[0157]

[0158]

[0159] in, , is the ceiling function.

[0160] In this embodiment, the number of core nodes Round up to expand the number of nodes After rounding up, for example, when , , hour, , , and satisfies .

[0161] Specifically, the process of generating the dynamic learning path in step S5 includes:

[0162] S51: From the core module, N nodes are selected in descending order according to the urgency coefficient of the weak links of the knowledge nodes to generate a reinforcement training sequence.

[0163] In this embodiment, the process of selecting N nodes from the core module to generate an enhanced training sequence includes three key steps.

[0164] First, all nodes of the core module are arranged in descending order according to the weak link urgency coefficient, which comprehensively considers the weak probability output by the capability prediction model (50% weight), the dynamic difficulty coefficient of the knowledge node (30% weight), and the learning time factor (20% weight).

[0165] Secondly, the system automatically excludes two types of nodes: duplicate nodes that have been selected for three consecutive times and nodes whose predecessor dependencies do not meet the standards (average mastery of predecessor nodes <60%).

[0166] Finally, each selected node is intelligently matched with three types of learning resources: targeted instructional videos (automatically selected based on a library of error types), a step-by-step practice set (covering three levels of difficulty: basic, advanced, and challenging), and a knowledge point association map (visually displaying prerequisite relationships). If the number of selectable nodes is less than N, the next most urgent node (with a coefficient greater than or equal to 0.6) is selected and triggers an alert on the teacher's end.

[0167] S52: From the expansion module, M nodes are selected in descending order of interest relevance weights of the knowledge nodes to generate an expansion content sequence.

[0168] In this embodiment, the generation of the extended content sequence is performed through intelligent screening based on the interest relevance weight.

[0169] First, the comprehensive weight value of each expansion node is calculated. This value consists of three parts: the interest similarity of the learning style vector (basic value), the class popularity coefficient (reflecting the number of visits / interactions), and the personalized gain factor (1 + historical completion rate). 0.5).

[0170] All nodes are then sorted in descending order of weight, applying two core filtering criteria: no more than two nodes from the same knowledge domain (to ensure diversity) and a difference in difficulty between adjacent nodes of less than 0.3 (to ensure a smooth transition). Selected nodes are automatically assigned three types of supplementary resources: engaging learning resources (such as interactive simulations and gamification), interdisciplinary application cases, and high-quality learning community content (the top three most-liked discussion threads). The system intelligently filters out nodes marked "not interesting" by students and prioritizes content highly relevant to the core module when popular nodes have similar weights (a difference of less than 0.05).

[0171] S53: Cross-arranging the enhanced training sequence and the expanded content sequence according to the fusion ratio and the arrangement rules to generate a dynamic learning path;

[0172] The arrangement rule is to allocate Core nodes are inserted expansion nodes, wherein the core node is a knowledge node in the core module, and the expansion node is a knowledge node in the expansion module, is the ceiling function.

[0173] The dynamic learning path is a personalized learning sequence formed by cross-arranging core knowledge nodes and expanded knowledge nodes in a scientific proportion. Its core features include:

[0174] (1) Adaptive structure: according to students' real-time ability ( ) and interest preferences ( ) Dynamically adjust the node ratio.

[0175] (2) Progressive reinforcement: core nodes are arranged in descending order of urgency to ensure that knowledge gaps are remedied first.

[0176] (3) Interest integration: expansion nodes are arranged according to interest weights to maintain learning motivation.

[0177] In this embodiment, for example, if (Core accounts for 60%), (Expansion accounts for 40%), the core allocation interval is 2, and the expansion insertion interval is 3, that is, after allocating 2 core nodes, 3 expansion nodes are inserted, forming a cyclic unit: core → core → expansion → expansion → expansion → expansion.

[0178] When there are insufficient core or expansion nodes, the current loop unit length is automatically shortened. For example, if there is one core node and two expansion nodes remaining, a short sequence of core → expansion → expansion is generated.

[0179] Dynamic learning paths build an efficient and sustainable learning process for current students through an intelligent node arrangement mechanism. Based on the students' real-time ability level (through the core knowledge integration ratio), ) and interest preferences (by expanding the proportion of knowledge integration ), dynamically adjust the ratio of core and extension content. For example, when students have obvious knowledge deficiencies ( ), the paths are arranged according to the rhythm of "2 core + 3 expansion", giving priority to strengthening weak knowledge points (such as loopholes in solving quadratic function problems), and inserting interest-oriented expansion content (such as mathematical gamification application cases) after each high-load core training to effectively balance cognitive pressure and learning motivation.

[0180] If the student's recent learning data shows a significant increase in interest ( If the number of nodes increases from 0.4 to 0.5, the interval between expansion nodes is automatically shortened, increasing the exposure frequency of interesting content. If there are insufficient remaining nodes, the path generation module intelligently compresses the loop units (for example, "1 core + 2 expansions") to avoid learning interruptions caused by mechanically adhering to a fixed ratio. This dynamic mechanism ensures that students consistently learn in a virtuous cycle of "targeted reinforcement" and "interest-driven" learning.

[0181] Example 2

[0182] The present invention also provides an AI-enabled intelligent teaching personalized service system for executing the AI-enabled intelligent teaching personalized service method. Figure 2 As shown, the system includes:

[0183] The multimodal data collection module 100 is used to periodically collect the multimodal data of the current student, wherein the multimodal data includes learning data, behavioral data and a knowledge graph, wherein the knowledge graph is composed of multiple knowledge nodes containing difficulty coefficients and dependency relationships.

[0184] The association analysis module 200 is used to perform association analysis on the learning data and behavior data to generate analysis results, wherein the association analysis includes cluster classification, time series prediction and vector mapping, and the analysis results include group labels, ability prediction models and academic style vectors corresponding to the association analysis.

[0185] The knowledge node stratification module 300 is used to construct a weight matrix according to the group label and the ability prediction model, and divide all knowledge nodes into stratified courses according to the weight matrix, wherein the stratified courses include core modules and expansion modules.

[0186] The knowledge proportion fusion module 400 is used to determine the core knowledge proportion and the extended knowledge proportion based on the ability prediction model and the study style vector, and fuse them to generate a fusion ratio.

[0187] The dynamic path generation module 500 is used to select the enhanced training sequence and the extended content sequence from the core module and the extended module respectively, and cross-arrange the enhanced training sequence and the extended content sequence according to the fusion ratio to generate a dynamic learning path.

[0188] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems) and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0189] Those skilled in the art will appreciate that all or part of the steps in the various methods of the above embodiments can be completed by instructing related hardware through a program. The program can be stored in a computer-readable storage medium, including a read-only memory (ROM), a random access memory (RAM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), a one-time programmable read-only memory (OTPROM), an electronically erasable programmable read-only memory (EEPROM), a compact disc read-only memory (CD-ROM), or other optical disc storage, magnetic disk storage, or magnetic tape storage, or any other computer-readable medium capable of carrying or storing data.

[0190] It should also be noted that the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, commodity, or apparatus that includes a series of elements includes not only those elements but also other elements not explicitly listed, or includes elements inherent to such process, method, commodity, or apparatus. In the absence of further limitations, an element defined by the phrase "comprises a..." does not exclude the presence of other identical elements in the process, method, commodity, or apparatus that includes the element.

Claims

1. An AI-enabled intelligent teaching personalized service method, characterized by: The method comprises the following steps: Periodically collecting multimodal data of the current student, wherein the multimodal data includes learning data, behavioral data, and a knowledge graph, wherein the knowledge graph is composed of multiple knowledge nodes containing difficulty coefficients and dependency relationships; Performing association analysis on the learning data, the knowledge graph, and the behavioral data to generate analysis results, wherein the association analysis includes cluster classification, time series prediction, and vector mapping, and the analysis results include group labels, ability prediction models, and learning style vectors corresponding to the association analysis; The association analysis specifically includes: Cluster classification: Based on the performance indicators and classroom scores in the learning data, combined with the semantic feature vectors generated by the interactive text and platform logs in the behavioral data, the group labels are generated through a spectral clustering algorithm; Time series prediction: using the progress records in the learning data, combined with the difficulty coefficients and dependencies of the knowledge nodes, to construct a time series prediction model and output the ability prediction model; Vector mapping: analyzing the interest intensity value in the behavioral data and the homework quality rating in the learning data, and generating the learning style vector through a cross-modal association algorithm; Constructing a weight matrix according to the group labels and the ability prediction model, and dividing all knowledge nodes into hierarchical courses according to the weight matrix, wherein the hierarchical courses include core modules and expansion modules; The weight matrix construction process specifically includes: Calculate the group mastery weight of each knowledge node according to the group distribution density of the group label; Based on the probability distribution of weak links in the capability prediction model, generating individual urgency weights of each knowledge node; Combining the group mastery weight and the individual urgency weight into a weight matrix; The process of tiered course division specifically includes: Sort the knowledge nodes according to the weight matrix, and define the knowledge nodes that exceed the preset core knowledge threshold as core modules; After evaluating the remaining knowledge nodes based on their dependencies and difficulty coefficients with the knowledge nodes of the core module, the knowledge nodes that meet the evaluation criteria are divided into expansion modules; Based on the ability prediction model and the study style vector, the core knowledge proportion and the extended knowledge proportion are determined respectively and integrated to generate an integrated ratio; Reinforced training sequences and expanded content sequences are respectively selected from the core module and the expanded module, and the reinforced training sequences and the expanded content sequences are cross-arranged according to the fusion ratio to generate a dynamic learning path.

2. The AI-enabled intelligent teaching personalized service method according to claim 1, characterized in that: The periodic collection of multimodal data of the current student includes: Periodically collect learning data of current students, wherein the learning data includes performance indicators, progress records, class grades, and homework quality grading; Periodically collect the current student's behavioral data, where the behavioral data includes a semantic feature vector composed of interactive text and platform logs, and an interest intensity value; A knowledge graph consisting of multiple knowledge nodes is established based on the subject knowledge system.

3. The AI-enabled intelligent teaching personalized service method according to claim 2, characterized in that: The generation process of the fusion ratio specifically includes: Extracting the urgency coefficient of the weak links of each knowledge node in the capability prediction model and generating the core knowledge proportion through normalization processing; Extracting the interest relevance weight of the study style vector and adjusting it by a preset interest gain coefficient to generate an expanded knowledge proportion; The core knowledge proportion and the extended knowledge proportion are integrated to generate an integrated ratio, where the integration formula is: in, The proportion of core knowledge, To expand the knowledge share, is the integration ratio of core knowledge, To expand the integration ratio of knowledge.

4. The AI-enabled intelligent teaching personalized service method according to claim 3 is characterized in that: The generation process of the dynamic learning path includes: From the core module, N nodes are selected in descending order of the urgency coefficient of the weak links of the knowledge nodes to generate a reinforcement training sequence; From the expansion module, M nodes are selected in descending order of interest relevance weight of the knowledge nodes to generate an expansion content sequence; According to the fusion ratio, the reinforcement training sequence and the expansion content sequence are cross-arranged according to the arrangement rules to generate a dynamic learning path; The arrangement rule is to allocate Core nodes are inserted expansion nodes, wherein the core node is a knowledge node in the core module, and the expansion node is a knowledge node in the expansion module, is the ceiling function.

5. The AI-enabled intelligent teaching personalized service method according to claim 4 is characterized in that: The process of determining N and M includes: The total number of nodes in a single learning path is preset to T, where the total number of nodes T is dynamically set according to the average daily effective learning time of the current student; N and M are calculated according to the fusion ratio and the total number of nodes using the formula: in, , is the ceiling function.

6. An AI-enabled intelligent teaching personalized service system, used to execute an AI-enabled intelligent teaching personalized service method according to any one of claims 1 to 5, characterized in that: The system comprises: A multimodal data collection module, configured to periodically collect multimodal data of the current student, wherein the multimodal data includes learning data, behavioral data, and a knowledge graph, wherein the knowledge graph is composed of multiple knowledge nodes containing difficulty coefficients and dependency relationships; an association analysis module, configured to perform association analysis on the learning data and the behavioral data to generate analysis results, wherein the association analysis includes cluster classification, time series prediction, and vector mapping, and the analysis results include group labels, ability prediction models, and learning style vectors corresponding to the association analysis; A knowledge node stratification module is used to construct a weight matrix based on the group labels and the ability prediction model, and to divide all knowledge nodes into stratified courses based on the weight matrix, wherein the stratified courses include core modules and expansion modules; A knowledge proportion fusion module is used to determine the core knowledge proportion and the extended knowledge proportion based on the ability prediction model and the study style vector, and fuse them to generate a fusion ratio; The dynamic path generation module is used to filter out the enhanced training sequence and the extended content sequence from the core module and the extended module respectively, and cross-arrange the enhanced training sequence and the extended content sequence according to the fusion ratio to generate a dynamic learning path.

Images