A Method and System for Pushing Teaching Content Based on Multi-Source Data Analysis

Abstract

This invention discloses a method and system for pushing teaching content based on multi-source data analysis. The method includes the following steps: S1, collecting and preprocessing multi-source heterogeneous teaching data to obtain a standardized dataset; S2, fusing multi-source data using a graph neural network-based fusion model to generate a fusion feature vector, and constructing a multi-dimensional dynamic user profile based on this vector; S3, constructing a knowledge point intelligent graph, and analyzing the adaptability of teaching content from different dimensions in conjunction with the dynamic user profile; S4, generating a multi-objective push strategy and push scheme based on the teaching content adaptability analysis results, executing the push task and collecting feedback data, evaluating the push effect, and iteratively optimizing the push strategy. This invention, employing the above-mentioned method and system for pushing teaching content based on multi-source data analysis, can output systematic teaching content adapted to students' personalized needs and teaching progress, and continuously optimize the push strategy through a real-time feedback mechanism, thereby improving teaching efficiency and learning outcomes.

Description

Technical Field

[0001] This invention relates to the fields of educational technology and data analysis technology, and in particular to a method and system for pushing teaching content based on multi-source data analysis. Background Technology

[0002] With the rapid development of smart education, personalized teaching has become a core direction for breaking through the homogeneity of traditional classrooms and improving teaching quality. Currently, content delivery technology, as an important carrier of personalized teaching, is gradually shifting from experience-driven to data-driven approaches. However, existing content delivery solutions still have many shortcomings and are insufficient to meet the demands of the new era of education for precision, personalization, and end-to-end adaptation.

[0003] In existing technologies, some methods for pushing educational content rely solely on student exam scores and resource click-through rates for decision-making. This limitation in data sources leads to a one-sided user profile, failing to comprehensively reflect students' learning abilities, cognitive characteristics, learning preferences, and knowledge gaps. For example, pushing remedial content based solely on grades ignores students' learning styles, attention spans, and classroom interaction, easily resulting in a mismatch between the pushed content and students' actual needs, thus reducing learning efficiency. Another approach, while attempting to integrate multi-source data, suffers from insufficient data fusion depth and weak heterogeneous data processing capabilities.

[0004] Meanwhile, the push logic of existing systems is mostly driven by static rules, while the teaching process is dynamic and evolving. Students' knowledge status and learning needs change in real time with their learning progress. Existing systems struggle to quickly capture these dynamic changes and adjust their push strategies accordingly. Furthermore, they lack a comprehensive knowledge point association system, resulting in fragmented push content that fails to guide students in forming a systematic knowledge structure. In addition, some systems do not fully consider the diversity of teaching scenarios, such as blended online and offline teaching and the characteristics of different subjects, leading to poor universality and scalability of their push solutions.

[0005] Therefore, there is an urgent need for a teaching content delivery method and system that can deeply integrate multi-source heterogeneous teaching data, accurately construct dynamic user profiles, and achieve systematic delivery based on knowledge graphs with closed-loop optimization capabilities, in order to address the shortcomings of existing technologies and promote the large-scale implementation of personalized teaching. Summary of the Invention

[0006] The purpose of this invention is to provide a teaching content push method and system based on multi-source data analysis, to achieve deep integration of multi-source heterogeneous teaching data, to build a comprehensive and accurate dynamic user profile and knowledge point association system, to output systematic teaching content that adapts to students' personalized needs and teaching progress, and to continuously optimize the push strategy through a real-time feedback mechanism, thereby improving teaching efficiency and learning outcomes.

[0007] To achieve the above objectives, this invention provides a method for pushing teaching content based on multi-source data analysis, comprising the following steps: S1. Collect and preprocess multi-source heterogeneous teaching data to obtain a standardized dataset; S2. Multi-source data is fused using a graph neural network-based fusion model to generate a fused feature vector, and a multi-dimensional dynamic user profile is constructed based on this vector. S3. Construct an intelligent knowledge graph and analyze the suitability of teaching content from different dimensions by combining dynamic user profiles; S4. Generate multi-target push strategies and push schemes based on the results of teaching content adaptability analysis, execute push tasks and collect feedback data, evaluate the push effect and iteratively optimize the push strategy.

[0008] Preferably, in S1, the multi-source heterogeneous teaching data includes academic data, behavioral data, cognitive data, environmental data, and resource data; Academic data includes students' exam scores, homework completion status, knowledge mastery assessment results, error records, and correction data; Behavioral data includes students' online resource access patterns, classroom interaction behavior, eye-tracking data, and learning device operation behavior data; Cognitive data includes data on students' learning styles, personality traits, attention span, and thinking speed; Environmental data includes teaching progress, course type, class student distribution, learning time period, and equipment environment data; Resource data includes the type of teaching content, difficulty level, knowledge point tags, appropriate grade level, and resource rating data.

[0009] Preferably, in S1, the preprocessing operation specifically includes: The Isolation Forest algorithm is used to identify and remove outlier data, as shown in the following formula: ; In the formula, For a single data sample to be detected The abnormal value score ranges from [0,1]. This refers to a single data sample to be tested. The total sample size for training the isolated forest; For the sample Path length in an isolated tree; For the sample The expected value of the path length on all isolated trees; This represents the average expected value of the path length for normal samples. Set threshold ,when If the data sample is found to be abnormal, it will be removed.

[0010] Preferably, in S1, the preprocessing further includes converting the text data into vector form using the Word2Vec model. The text vector is represented as follows: ; In the formula, For text vectors, The number of words in the text. For the first Word vectors of words, For the first The weight of each word is expressed as: ; In the formula, For the first The word in the text The frequency in This represents the total number of texts. For the appearance of the first The number of words in the text; Time series data are aligned to a uniform time granularity, and categorical data is encoded using label encoding. By using the cosine similarity algorithm and Jaccard coefficient to remove redundant data, the cosine similarity formula is optimized as follows: ; In the formula, These are two independent multi-source data sample vectors. , These are the vector magnitudes, The smoothing factor for smoothing vector similarity results, when At that time, determine that one of the data is redundant and remove it; For missing data, the mean is used to fill numerical data, and the mode is used to fill categorical data; LSTM time series prediction is used to fill time series behavioral data.

[0011] Preferably, in S2, a multi-source data fusion model based on graph neural network (GNN) is constructed. The model structure includes an input layer, a feature alignment layer, a graph aggregation layer, and an output layer. The input layer consists of aligned feature vectors. , The total number of samples; The feature alignment layer outputs standardized feature vectors, and the graph aggregation layer uses the GraphSAGE algorithm to aggregate nodes, with the following formula: ; In the formula, For nodes No. The hidden state of the layer For the first The weight matrix of the layer, For the first Layer bias vector, For nodes The set of neighboring nodes, The number of neighboring nodes. This is a vector concatenation operation. For activation function, For nodes No. The hidden state of the layer; A dynamic weight adjustment mechanism is introduced, and the dynamic weight update formula is as follows: ; In the formula, For the first Dynamic weights of data sources The time decay coefficient, As a time-sensitive factor, As a reliability factor, To supplement the coefficients, For data integrity factors; Noise filtering is performed simultaneously during the fusion process, and L2 regularization is used to suppress redundant information interference. The regularization formula is: ; In the formula, To improve the L2 regularization loss, The regularization coefficient is . This represents the total number of weight matrices in the model. For the first Each weight matrix is ​​used to ultimately output the fused feature vector. , The total number of samples; Based on fused feature vectors, a multi-dimensional user profile model is constructed. The profile dimensions include knowledge, ability, preference, and state dimensions. The feature quantification formulas for each dimension are as follows: The knowledge dimension is quantified by knowledge point mastery, and the formula is: ; in, Knowledge point mastery is used to quantify students' overall grasp of knowledge points. The total number of knowledge points mastered; For the first The weight of each knowledge point is set according to the proportion of the course; For the first The mastery status of each knowledge point; For the first The difficulty level of each knowledge point; The ability dimension is quantified using learning efficiency, and the formula is: ; in, To measure learning efficiency, it is used to quantify the student's learning input-output ratio; Effective time for completing learning tasks; The total time spent completing a learning task includes both effective and ineffective time. This is a focus coefficient used to reward focused learning behavior; The preference dimension is quantified using resource type preference, and the formula is: ; in, Resource type preference is used to quantify students' preferences for the first resource type. The degree of preference for this type of teaching resource; For students to visit the The number of times a type of teaching resource is used; This refers to the resource adaptation coefficient. Student visit The number of times a type of teaching resource is used; For the first The fit coefficient of similar teaching resources This represents the total number of teaching resource categories. The state dimension is quantified using attention concentration, and the formula is: ; in, Attention concentration is used to quantify the level of focus students have during the learning process. Effective fixation time; Total study time; The smoothing factor, which is used to smooth attention concentration, is fixed at 1 to avoid calculation errors caused by a numerator of 0. Establish a dynamic profile update mechanism and use an incremental learning algorithm to update user profiles. The update formula is as follows: ; in, for User profiles at any time are used to characterize students. A constant state of learning; For historical portrait weighting coefficients; for Historical user profiles at any given moment; Incremental vectors for generating profiles from newly added data; This is the time decay factor.

[0012] Preferably, in S3, a graph structure model is used to construct a knowledge point intelligent graph. : ; in, It is a set of knowledge point node vectors. This represents the total number of knowledge points in the intelligent knowledge point graph. This is the set of edges connecting knowledge points. Let the set of edge weights represent the association strength, and the calculation formula is: ; In the formula, For knowledge points and The number of times they co-occur. A smoothing factor calculated for the strength of association between knowledge points. The correlation strength coefficient; Based on resource data and academic data, machine learning algorithms are used to mine implicit knowledge point associations, supplement the graph edge structure, and the mining formula adopts a support-confidence model for association rule mining. ; ; ; In the formula, , There are two knowledge points. To involve , The number of incorrect questions or resources, This refers to the total number of incorrect questions or resources. For those involved The number of incorrect questions or resources, For knowledge points Support To increase the degree, Confidence level; When support >0.2, confidence level >0.7 and lift When >1.2, determine and There is a significant implicit association.

[0013] Preferably, S4 includes the following steps: S41. Based on the content adaptation analysis results, construct a multi-objective push strategy model with the objective function as follows: ; in, To provide a comprehensive score for the push strategy, To enhance efficiency through knowledge. To enhance learning interest, The learning burden index, , , These are the weighting coefficients; Using heuristic graph search A The algorithm explores the optimal sequence of teaching content in the knowledge point intelligent graph. The evaluation function of the algorithm is: ; in, For nodes The evaluation value, From the starting node to the node The actual cost, For the node The estimated cost to reach the target node. For the estimated cost weighting coefficient; S42. The push system pushes teaching content to students through the learning terminal according to the generated push plan; S43. Collect student feedback data on the pushed content, and build a push effect evaluation model based on the feedback data. Quantitatively evaluate the content from three dimensions: content suitability, learning effectiveness improvement rate, and user experience. The content suitability evaluation formula is as follows: ; in, For knowledge adaptation, For capability adaptation, For preference fit, , , The weights for each item are: ; ; In the formula, A smoothing factor used to smooth out knowledge fit; The formula for evaluating the improvement rate of learning effectiveness is as follows: ; in, To assess the level of knowledge mastery after the push notification, To assess the level of knowledge mastery before pushing the content; The user experience rating formula is as follows: ; in, For the number of feedback recipients, For the first Student satisfaction ratings. As a weight for feedback credibility; Overall assessment score The evaluation indicators include content completion rate, knowledge point mastery improvement rate, user satisfaction score, and learning efficiency improvement rate. Based on the evaluation results, the push strategy was optimized using a reinforcement learning algorithm. The reinforcement learning model adopted the Q-learning algorithm, and the formula is as follows: ; in, For state Next action of value, For learning rate, As a reward value, As a discount factor, To perform the action The next state after that, For state The optimal action is as follows. To explore motivating factors.

[0014] This invention provides a teaching content delivery system based on multi-source data analysis, comprising: The data acquisition module is used to collect multi-source heterogeneous teaching data, including academic data acquisition unit, behavioral data acquisition unit, cognitive data acquisition unit, environmental data acquisition unit and resource data acquisition unit, and has access control and data encryption functions; The data processing and fusion module is used for data preprocessing, feature alignment and deep fusion, and outputs a fused feature vector. The user profile building module is used to build multi-dimensional dynamic user profiles, and supports real-time updates and historical review of profiles. The knowledge point graph construction module is used to organize the knowledge point system and generate and dynamically update the intelligent knowledge point graph. The push strategy generation module is used to generate personalized push solutions based on user profiles and knowledge point graphs, adapting to multiple teaching scenarios; The push execution module is used to execute push tasks, supports multi-channel push linkage, and monitors push status; The feedback optimization module is used to collect feedback data, evaluate the push effect, and iteratively optimize the strategy; The storage module adopts a distributed storage architecture to store various types of data and has encryption, backup, and efficient retrieval functions.

[0015] Preferably, the data processing and fusion module uses the TensorFlow framework to build a graph neural network fusion model, the push strategy generation module uses a heuristic graph search algorithm to generate push schemes, and the storage module uses a hybrid storage architecture of MySQL and MongoDB to support the classified storage of structured and unstructured data.

[0016] Therefore, the present invention employs the above-mentioned teaching content delivery method and system based on multi-source data analysis, and the beneficial effects are as follows: (1) The teaching content push method and system based on multi-source data analysis of the present invention can accurately locate students' knowledge gaps. The improvement rate of students' knowledge mastery in the experimental group is much higher than that of traditional single data push and random push, which significantly improves students' knowledge mastery level.

[0017] (2) This invention can effectively optimize the input-output ratio of learning. The average learning efficiency of the experimental group is better than that of the control group. It can reduce ineffective learning time, help students improve their academic ability efficiently, and highlight the advantages of multi-source data fusion.

[0018] (3) This invention can quickly adapt to the personalized teaching needs of junior high school mathematics. Relying on multi-source data fusion, multi-dimensional dynamic user profiles and knowledge intelligent graphs, the content pushed is highly targeted and has significant advantages over traditional push methods, providing reliable support for the implementation of personalized junior high school mathematics teaching.

[0019] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0020] Figure 1 This is an overall flowchart of an embodiment of the teaching content push method based on multi-source data analysis of the present invention; Figure 2 This is an overall system block diagram of an embodiment of the teaching content push system based on multi-source data analysis of the present invention; Figure 3 This is a comparison chart of experimental results for an embodiment of the teaching content delivery method and system based on multi-source data analysis of the present invention. Detailed Implementation

[0021] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0022] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0023] like Figure 1 As shown, the teaching content delivery method based on multi-source data analysis includes the following steps: S1. Multi-source teaching data collection and preprocessing to obtain a standardized dataset.

[0024] Collect multi-source heterogeneous teaching data from sources such as learning terminals, teaching terminals, campus management systems, and third-party educational resource platforms. The multi-source heterogeneous teaching data includes academic data, behavioral data, cognitive data, environmental data, and resource data.

[0025] The academic data includes student exam scores, homework completion status, knowledge mastery assessment results, error records, and correction data collected through the campus management system, online homework platform, and teacher grading terminals.

[0026] Behavioral data includes students' online resource access trajectories, classroom interaction behaviors, eye-tracking data, and learning device operation behavior data collected through learning terminal sensors, classroom interaction systems, and eye-tracking devices.

[0027] Cognitive data includes data on students' learning styles, personality traits, attention span, and thinking speed, obtained through standardized questionnaires, cognitive testing modules, and reverse inference from behavioral data.

[0028] Environmental data includes teaching progress, course type, class student distribution, learning time period, and equipment environment data collected through teaching terminals, campus management systems, and environmental sensors on learning terminals.

[0029] Resource data includes the type, difficulty level, knowledge point tags, appropriate grade level, and resource rating data of teaching content collected through third-party educational resource platforms and local resource libraries.

[0030] Preprocessing is performed on the collected multi-source heterogeneous data, specifically as follows: The Isolation Forest algorithm is used to identify and remove data with abnormal performance and behavioral anomalies. The formula is as follows: ; In the formula, For a single data sample to be detected The abnormal value score ranges from [0,1]. This refers to a single data sample to be tested. The total sample size for training the isolated forest; For the sample Path length in an isolated tree; For the sample The expected value of the path length on all isolated trees is used to reduce the randomness interference of individual trees; The average expected value of the path length for normal samples is calculated using the following formula: ; In the formula, For the first The harmonic number is approximately equal to the order of harmonics. ,in, 0.5772 is Euler's constant.

[0031] Set threshold =0.75, when If the data sample is found to be abnormal, it will be removed.

[0032] The preprocessing steps also include converting the text data into vector form using the Word2Vec model. The text vector is represented as follows: ; In the formula, For text vectors, The number of words in the text. For the first Word vectors of words, For the first The weight of each word is calculated using TF-IDF, and the expression is: ; In the formula, For the first The word in the text The frequency in This represents the total number of texts. For the appearance of the first The number of words in the text.

[0033] Time series data is aligned to a uniform time granularity by minute or class period, and categorized data is encoded using label encoding.

[0034] Furthermore, redundant data was removed using the cosine similarity algorithm and Jaccard coefficient to avoid data duplication interfering with the analysis results. The cosine similarity formula was optimized as follows: ; In the formula, These are two independent multi-source data sample vectors. , These are the vector magnitudes, The smoothing factor for smoothing vector similarity results, when At a value of 0.85, one of the data points is identified as redundant and removed.

[0035] For missing data, numerical data is filled with the mean, categorical data is filled with the mode, and time-series behavioral data is filled with LSTM time-series prediction. The LSTM time-series prediction formula is as follows: ; ; ; ; ; in, , , These are the input gate, forget gate, and output gate vectors, respectively. Let be the cell state vector. The hidden layer output vector. for Input vector at time step, , The weight matrix calculated by the input gate corresponds to... Time-input vector and Hidden layer output vector at each time step The linear mapping weights; , The weight matrix for the forget gate calculation corresponds to... Time-input vector and Hidden layer output vector at each time step The linear mapping weights; , The weight matrix representing cell state updates corresponds to... Time-input vector and Hidden layer output vector at each time step The linear mapping weights; , The weight matrix of the output gate corresponds to... Time-input vector and Hidden layer output vector at each time step The linear mapping weights; , , and These are the input gate bias vector, forget gate bias vector, cell state bias vector, and output gate bias vector. It is the Sigmoid activation function. For element-wise product operation, =0.1、 =0.1 is the gating adjustment coefficient, which is used to improve the stability of time series prediction and ensure the integrity of the dataset.

[0036] S2. A fusion model based on graph neural networks is used to fuse multi-source data, generate a fusion feature vector, and construct a multi-dimensional dynamic user profile based on this vector.

[0037] First, the structured data is normalized using the Min-Max method, as shown in the following formula: ; In the formula, For the normalized data, This is the original data. , These represent the maximum and minimum values ​​of the dataset, respectively. =1、 =0 represents the upper and lower limits of normalization, which are adjusted according to the model requirements. Here, it is mapped to the [0,1] interval.

[0038] For unstructured data, semantic features of text comments are extracted using natural language processing (NLP), spatial features of eye-tracking trajectories are extracted using computer vision technology, and then transformed into standardized feature vectors.

[0039] Then, an attention mechanism is used to assign initial weights to features from different sources, highlighting the weight proportions of academic data and core behavioral data. The attention weight allocation formula is as follows: ; In the formula, For the first Initial weights of the data source class For the first Importance rating of data source class For the first Importance rating of data source class The total number of categories in the data source, as described in this invention. =5, corresponding to academic, behavioral, cognitive, environmental, and resource data, introduced The weighted terms reinforce the differences in importance, with the initial weights assigned as follows: academic data =0.35, behavioral data =0.3, cognitive data =0.2, Environmental Data =0.1, resource data =0.05.

[0040] A multi-source data fusion model based on graph neural networks (GNNs) is constructed. The model structure includes an input layer, a feature alignment layer, a graph aggregation layer, and an output layer, where the input layer consists of aligned feature vectors. , The total number of samples; the feature alignment layer outputs a standardized feature vector with a dimension of 128; the graph aggregation layer uses the GraphSAGE algorithm to aggregate nodes, with the formula as follows: ; In the formula, For nodes No. The hidden state of the layer For the first The weight matrix of the layer, For the first Layer bias vector, For nodes The set of neighboring nodes, The number of neighboring nodes. This is a vector concatenation operation. For nodes No. The hidden state of the layer This is used as an activation function to enhance the model's nonlinear expressive power.

[0041] A dynamic weight adjustment mechanism is introduced to dynamically adjust the weights of each data source based on data reliability and timeliness. The dynamic weight update formula is as follows: ; In the formula, For the first Dynamic weights of data sources =0.6 is the time decay coefficient. For timeliness factors, recent data =1.2, historical data =0.8, For reliability factor (human verification data) =1.2, Predicted data =0.9); =0.2 is a supplementary coefficient. For data integrity factor (complete data) =1.0), complete the data ( =0.8); For example, after the weights of recent complete pre-study behavior data are updated: 0.3×(0.6×1.2+0.4×1.0+0.2×1.0)=0.36.

[0042] After the historical performance data weights were updated: 0.35×(0.6×0.8+0.4×1.2+0.2×1.0)=0.336.

[0043] Noise filtering is performed simultaneously during the fusion process, and L2 regularization is used to suppress redundant information interference. The regularization formula is: ; In the formula, This is an improved L2 regularization loss (to avoid excessive weight decay). =0.001 is the regularization coefficient. This represents the total number of weight matrices in the model. For the first Each weight matrix is ​​used to ultimately output a fused 256-dimensional feature vector. , This represents the number of samples.

[0044] Based on fused feature vectors, a multi-dimensional user profile model is constructed. The profile dimensions include knowledge, ability, preference, and state dimensions. The feature quantification formulas for each dimension are as follows: The knowledge dimension is quantified by knowledge point mastery, and the formula is: ; in, Knowledge point mastery is used to quantify students' overall grasp of knowledge points. The total number of knowledge points mastered; For the first The weight of each knowledge point is set according to the proportion of the course; For the first The mastery status of each knowledge point, including mastery, partial mastery, and no mastery. =1, partially mastered =0.5, not yet known =0; For the first The difficulty level of each knowledge point, basic level. =0.8, advanced level =1.0, Extended Level =1.2.

[0045] The ability dimension is quantified using learning efficiency, and the formula is: ; in, To measure learning efficiency, it is used to quantify the student's learning input-output ratio; Effective time for completing learning tasks; The total time spent completing a learning task includes both effective and ineffective time. This is the focus coefficient, with a fixed value of 1.1, used to reward focused learning behavior.

[0046] The preference dimension is quantified using resource type preference, as shown in the following formula: ; in, Resource type preference is used to quantify students' preferences for the first resource type. The degree of preference for this type of teaching resource; For students to visit the The number of times a type of teaching resource is used; The resource adaptation coefficient is used to match resources to students' learning styles. =1.2, ordinary resources =1.0; The total number of types of teaching resources; Student visit The number of times a type of teaching resource is used; For the first The fit coefficient of similar teaching resources This represents the total number of teaching resource categories.

[0047] The state dimension is quantified using attention concentration, and the formula is: ; in, Attention concentration is used to quantify the degree of focus of students during the learning process, with a value range of [0,1]. Effective fixation time, which is the time a student focuses on the learning content, is collected from eye-tracking data. Total learning time is the total duration of a student's participation in learning. This is a smoothing factor used to smooth attention concentration, with a fixed value of 1, to avoid calculation errors caused by a numerator of 0.

[0048] A dynamic user profile update mechanism is established, employing an incremental learning algorithm to update user profiles. The formula for dynamic user profile updates is as follows: ; in, for User profiles at any time are used to characterize students. A constant state of learning; The historical profile weighting coefficient is fixed at 0.7 and is used to retain historical learning status information. for Historical user profiles at any given moment; The incremental vector of the profile generated for the new data reflects the latest changes in the student's learning status; Time decay factor, real-time data =1.0, historical increment = This is used to adjust the weight of data from different time periods. The update cycle can be set to class period, daily, or weekly based on the teaching scenario to ensure that the profile remains consistent with the actual status of students.

[0049] S3. Construct an intelligent knowledge graph and analyze the suitability of teaching content from different dimensions by combining dynamic user profiles.

[0050] Intelligent knowledge graph constructed using graph structure model : ; in, It is a set of knowledge point node vectors. This represents the total number of knowledge points in the intelligent knowledge point graph. This is the set of edges connecting knowledge points. =1 indicates a knowledge point and There is a connection. =0 indicates no association). Let edge weights be the set of edge weights, where each edge weight represents the strength of the association. The calculation formula is as follows: ; In the formula, For knowledge points and The number of co-occurrences can be calculated based on incorrect question data and resource association data; the denominator is the maximum number of co-occurrences. =1 is the smoothing factor for calculating the association strength of knowledge points; For correlation strength coefficient, logical correlation =1.2, subordinate association =1.0, Extended Relationships =0.8, the weight range is mapped to [0,1], such as the association weight between the graph of a linear function and image translation. =0.82.

[0051] Based on resource data and academic data, machine learning algorithms are used to mine implicit knowledge point associations, including knowledge point combinations associated with frequently missed questions, to supplement the graph edge structure. The mining formula adopts a support-confidence model for association rule mining. ; ; ; In the formula, , There are two knowledge points. To involve , The number of incorrect questions or resources, This refers to the total number of incorrect questions or resources. For those involved The number of incorrect questions or resources, For knowledge points Support To improve the degree (measure the effectiveness of the association). The confidence level is the support level. >0.2, confidence level >0.7 and lift When >1.2, determine and Significant implicit relationships exist. The knowledge graph supports dynamic updates; when new teaching resources are added or curriculum standards are adjusted, the knowledge point nodes and their relationships are automatically updated to ensure the timeliness and completeness of the knowledge graph.

[0052] S4. Based on the results of the teaching content adaptability analysis, generate a multi-target push strategy and push plan, execute the push task and collect feedback data, evaluate the push effect and iteratively optimize the push strategy, including the following steps: S41. Based on the content adaptation analysis results, construct a multi-objective push strategy model with the objective function as follows: ; in, The overall score for the push strategy is in the range of 0-1. Enhance efficiency for knowledge (0-1). To assess the improvement in learning interest (0-1). The learning burden index is 0-1. =0.4、 =0.3、 =0.3 is the weighting coefficient, which transforms the burden term into a positive optimization objective and improves the rationality of the formula.

[0053] Using heuristic graph search A The algorithm explores the optimal sequence of teaching content in the knowledge point intelligent graph. The evaluation function of the algorithm is: ; in, For nodes The evaluation value, From the starting node to the node The actual cost, For the node The estimated cost to reach the target node. =0.8 is the estimated cost weighting coefficient, which balances the actual cost and the estimated cost. The algorithm generates the optimal teaching content sequence, i.e., the personalized push scheme.

[0054] Optimize push strategies for different teaching scenarios: For online self-study scenarios, a combined push mode of core content, reinforcement content, and extension content is adopted to give students the space to choose. For offline classroom scenarios, common content that is adapted to the class's learning situation and personalized supplementary content for individual differences are pushed. For blended teaching scenarios, online and offline content are pushed in a coordinated manner to ensure the continuity of the learning path.

[0055] S42. The push system pushes teaching content to students through learning terminals such as tablets, mobile phones, and smart blackboards according to the generated push plan; it supports multi-channel push linkage, including pop-up reminders, message notifications, and personalized homepage recommendations; it monitors the delivery status of content and the viewing status of students in real time during the push process to ensure that the pushed content is effectively delivered; it sets up a delayed reminder mechanism for content that has not been viewed in time to avoid missing key learning content. S43. Push notification performance feedback and strategy optimization: Collect student feedback data on the pushed content, including behavioral feedback such as content completion rate, dwell time, number of repeated views, and interactive operations; subjective feedback such as satisfaction ratings and suggestions; and effect feedback such as the improvement in the accuracy of knowledge point assessments.

[0056] A push notification effectiveness evaluation model is built based on feedback data, and quantitative evaluation is performed from three dimensions: content suitability, learning effect improvement rate, and user experience. The evaluation formulas for each dimension are as follows: The formula for evaluating content suitability is as follows: ; in, For knowledge adaptation, For capability adaptation, For preference fit, , , The weights for each item are: ; ; In the formula, This is a smoothing factor used to smooth out knowledge fit.

[0057] The formula for evaluating the improvement rate of learning effectiveness is as follows: ; in, For example, if the knowledge point mastery rate after the push is 0.72, The knowledge mastery level before the push was 0.42. =0.01 to avoid the denominator being 0, and to take 0 when the boost rate is negative.

[0058] The user experience rating formula is as follows: ; in, For the number of feedback recipients, For the first Student satisfaction ratings. To provide feedback on credibility weight, active users =1.2, for regular users =1.0.

[0059] Overall assessment score The evaluation indicators include content completion rate, knowledge point mastery improvement rate, user satisfaction score, and learning efficiency improvement rate.

[0060] Based on the evaluation results, the push strategy was optimized using a reinforcement learning algorithm. The reinforcement learning model adopted the Q-learning algorithm, and the formula is as follows: ; in, For state Next action of The value is used to measure the expected return of the strategy. =0.1 is the learning rate used to control the magnitude of policy updates. The reward value can be based on the comprehensive evaluation score. calculate, = / 10 is normalized to the 0-1 interval, such as =8.6 =0.86; =0.9 is the discount factor. To perform the action The next stage, such as the stage after students' mastery has improved; For state The optimal action corresponds to the adjusted push strategy. =1.1 is used to explore incentive factors, encourage strategies that adapt to new states, and avoid local optima.

[0061] for Strategies with a score of ≥8, indicating high adaptability and significant learning effectiveness, should be retained and promoted; for <6, meaning strategies for low adaptability and poor student feedback, include adjusting content difficulty, presentation format, or timing of push notifications; for cases where improvement in knowledge point mastery is not significant ( <0.1), re-analyze user profiles and knowledge gaps, optimize content adaptation logic; continuously iterate push model parameters to achieve dynamic optimization of push strategy.

[0062] like Figure 2 As shown, the teaching content push system based on multi-source data analysis of the present invention is used to implement the above method, including: The data acquisition module collects multi-source heterogeneous teaching data, including academic, behavioral, cognitive, environmental, and resource data collection units, and features access control and data encryption. The data processing and fusion module performs data preprocessing, feature alignment, and deep fusion, outputting a fused feature vector. The user profile building module constructs multi-dimensional dynamic user profiles, supporting real-time updates and historical data review.

[0063] The knowledge point graph construction module is used to organize the knowledge point system, generate and dynamically update the intelligent knowledge point graph. The push strategy generation module is used to generate personalized push plans based on user profiles and the knowledge point graph, adapting to multiple teaching scenarios. The push execution module is used to execute push tasks, supports multi-channel push linkage, and monitors push status.

[0064] The feedback optimization module collects feedback data, evaluates the push effect, and iteratively optimizes the strategy. The storage module employs a distributed storage architecture to store various types of data and features encryption, backup, and efficient retrieval capabilities.

[0065] The data processing and fusion module uses the TensorFlow framework to build a graph neural network fusion model, the push strategy generation module uses a heuristic graph search algorithm to generate push schemes, and the storage module uses a hybrid storage architecture of MySQL and MongoDB to support the classification and storage of structured and unstructured data.

[0066] Example 1: To simplify the verification of the effectiveness of this invention, 120 students from three parallel classes in the second year of junior high school were selected as experimental subjects for an 8-week comparative experiment, as detailed below: Three parallel classes with identical initial learning conditions (i.e., no significant difference in entrance scores and mathematical foundation) were selected and grouped as follows. All classes adopted the same teaching schedule and were taught by the same teachers: The experimental group consisted of 40 people, who used the push method and system described in this invention; the control group consisted of 40 people, who used single academic data (homework and exam scores only) to drive the push; and the control group consisted of 40 people, who used random push of teaching content.

[0067] One week prior to the experiment, initial data from three groups of students was collected, including academic data, behavioral data, cognitive data, environmental data, and resource data. This data was preprocessed according to step S1 of this invention to obtain a standardized dataset. The experiment lasted for eight weeks. Teaching content was pushed out daily after class in a corresponding manner, and a knowledge point assessment was conducted weekly to collect core feedback data.

[0068] After the experiment, two core indicators were selected: the improvement rate of knowledge point mastery and learning efficiency. The average values ​​of the data before and after the experiment were calculated, as shown in the table below. Figure 3 As shown: Table 1. Statistical Table of Experimental Data

[0069] As shown in Table 1, the experimental group achieved a 72.2% improvement rate in knowledge point mastery, significantly higher than the 37.6% in control group 1 and 18.1% in control group 2. This indicates that the present invention can accurately pinpoint knowledge gaps and deliver targeted content. Furthermore, the average learning efficiency of the experimental group was 0.81, superior to 0.64 in control group 1 and 0.50 in control group 2. This demonstrates that the method of the present invention can effectively achieve deep integration of multi-source heterogeneous teaching data, construct a comprehensive and accurate dynamic user profile and intelligent knowledge point graph, output teaching content adapted to students' personalized needs, and continuously improve the delivery effect through a closed-loop feedback optimization mechanism. Compared with traditional single-data-driven delivery methods and random delivery methods, the present invention has significant advantages in improving knowledge point mastery, optimizing learning efficiency, user satisfaction, and content adaptability. It can effectively overcome the limitations of homogeneity and one-sidedness in traditional teaching delivery, providing reliable methodological support and practical basis for the large-scale implementation of personalized teaching.

[0070] Therefore, the teaching content push method and system based on multi-source data analysis described above can effectively improve students' mastery of knowledge points and learning efficiency. It has significant advantages over traditional push methods and can quickly adapt to personalized teaching needs.

[0071] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A teaching content delivery method based on multi-source data analysis, characterized in that, Includes the following steps: S1. Collect and preprocess multi-source heterogeneous teaching data to obtain a standardized dataset; S2. Multi-source data is fused using a graph neural network-based fusion model to generate a fused feature vector, and a multi-dimensional dynamic user profile is constructed based on this vector. S3. Construct an intelligent knowledge graph and analyze the suitability of teaching content from different dimensions by combining dynamic user profiles; S4. Generate multi-target push strategies and push schemes based on the results of teaching content adaptability analysis, execute push tasks and collect feedback data, evaluate the push effect and iteratively optimize the push strategy.

2. The teaching content delivery method based on multi-source data analysis according to claim 1, characterized in that, In S1, multi-source heterogeneous teaching data includes academic data, behavioral data, cognitive data, environmental data, and resource data; Academic data includes students' exam scores, homework completion status, knowledge mastery assessment results, error records, and correction data; Behavioral data includes students' online resource access patterns, classroom interaction behavior, eye-tracking data, and learning device operation behavior data; Cognitive data includes data on students' learning styles, personality traits, attention span, and thinking speed; Environmental data includes teaching progress, course type, class student distribution, learning time period, and equipment environment data; Resource data includes the type of teaching content, difficulty level, knowledge point tags, appropriate grade level, and resource rating data.

3. The teaching content delivery method based on multi-source data analysis according to claim 2, characterized in that, In S1, the preprocessing operation is as follows: The Isolation Forest algorithm is used to identify and remove outlier data, as shown in the following formula: ； In the formula, For a single data sample to be detected The abnormal value score ranges from [0,1]. This refers to a single data sample to be tested. The total sample size for training the isolated forest; For the sample Path length in an isolated tree; For the sample The expected value of the path length on all isolated trees; This represents the average expected value of the path length for normal samples. Set threshold ,when If the data sample is found to be abnormal, it will be removed.

4. The teaching content delivery method based on multi-source data analysis according to claim 3, characterized in that, In S1, preprocessing also includes using the Word2Vec model to convert the text data into vector form, where the text vector is represented as: ； In the formula, For text vectors, The number of words in the text. For the first Word vectors of words, For the first The weight of each word is expressed as: ； In the formula, For the first The word in the text The frequency in This represents the total number of texts. For the appearance of the first The number of words in the text; Time series data are aligned to a uniform time granularity, and categorical data is encoded using label encoding. By using the cosine similarity algorithm and Jaccard coefficient to remove redundant data, the cosine similarity formula is optimized as follows: ； In the formula, These are two independent multi-source data sample vectors. , These are the vector magnitudes, The smoothing factor for smoothing vector similarity results, when At that time, determine that one of the data is redundant and remove it; For missing data, the mean is used to fill numerical data, and the mode is used to fill categorical data; LSTM time series prediction is used to fill time series behavioral data.

5. The teaching content delivery method based on multi-source data analysis according to claim 4, characterized in that, In S2, a multi-source data fusion model based on graph neural network (GNN) is constructed. The model structure includes an input layer, a feature alignment layer, a graph aggregation layer, and an output layer. The input layer consists of aligned feature vectors. , The total number of samples; The feature alignment layer outputs standardized feature vectors, and the graph aggregation layer uses the GraphSAGE algorithm to aggregate nodes, with the following formula: ； In the formula, For nodes No. The hidden state of the layer For the first The weight matrix of the layer, For the first Layer bias vector, For nodes The set of neighboring nodes, The number of neighboring nodes. This is a vector concatenation operation. For activation function, For nodes No. The hidden state of the layer; A dynamic weight adjustment mechanism is introduced, and the dynamic weight update formula is as follows: ； In the formula, For the first Dynamic weights of data sources The time decay coefficient, As a time-sensitive factor, As a reliability factor, To supplement the coefficients, For data integrity factors; Noise filtering is performed simultaneously during the fusion process, and L2 regularization is used to suppress redundant information interference. The regularization formula is: ； In the formula, To improve the L2 regularization loss, The regularization coefficient is . This represents the total number of weight matrices in the model. For the first Each weight matrix is ​​used to ultimately output the fused feature vector. , The total number of samples; Based on fused feature vectors, a multi-dimensional user profile model is constructed. The profile dimensions include knowledge, ability, preference, and state dimensions. The feature quantification formulas for each dimension are as follows: The knowledge dimension is quantified by knowledge point mastery, and the formula is: ； in, Knowledge point mastery is used to quantify students' overall grasp of knowledge points. The total number of knowledge points mastered; For the first The weight of each knowledge point is set according to the proportion of the course; For the first The mastery status of each knowledge point; For the first The difficulty level of each knowledge point; The ability dimension is quantified using learning efficiency, and the formula is: ； in, To measure learning efficiency, it is used to quantify the student's learning input-output ratio; Effective time for completing learning tasks; The total time spent completing a learning task includes both effective and ineffective time. This is a focus coefficient used to reward focused learning behavior; The preference dimension is quantified using resource type preference, and the formula is: ； in, Resource type preference is used to quantify students' preferences for the first resource type. The degree of preference for this type of teaching resource; For students to visit the The number of times a type of teaching resource is used; This refers to the resource adaptation coefficient. Student visit The number of times a type of teaching resource is used; For the first The fit coefficient of similar teaching resources This represents the total number of teaching resource categories. The state dimension is quantified using attention concentration, and the formula is: ； in, Attention concentration is used to quantify the level of focus students have during the learning process. Effective fixation time; Total study time; The smoothing factor, which is used to smooth attention concentration, is fixed at 1 to avoid calculation errors caused by a numerator of 0. Establish a dynamic profile update mechanism and use an incremental learning algorithm to update user profiles. The update formula is as follows: ； in, for User profiles at any time are used to characterize students. A constant state of learning; For historical portrait weighting coefficients; for Historical user profiles at any given moment; Incremental vectors for generating profiles from newly added data; This is the time decay factor.

6. The teaching content delivery method based on multi-source data analysis according to claim 5, characterized in that, In S3, a graph structure model is used to construct a knowledge point intelligent graph. : ； in, It is a set of knowledge point node vectors. This represents the total number of knowledge points in the intelligent knowledge point graph. This is the set of edges connecting knowledge points. Let the set of edge weights represent the association strength, and the calculation formula is: ； In the formula, For knowledge points and The number of times they co-occur. A smoothing factor calculated for the strength of association between knowledge points. The correlation strength coefficient; Based on resource data and academic data, machine learning algorithms are used to mine implicit knowledge point associations, supplement the graph edge structure, and the mining formula adopts a support-confidence model for association rule mining. ； ； ； In the formula, , There are two knowledge points. To involve , The number of incorrect questions or resources, This refers to the total number of incorrect questions or resources. For those involved The number of incorrect questions or resources, For knowledge points Support To increase the degree, Confidence level; When support >0.2, confidence level >0.7 and lift When >1.2, determine and There is a significant implicit association.

7. The teaching content delivery method based on multi-source data analysis according to claim 6, characterized in that, S4 includes the following steps: S41. Based on the content adaptation analysis results, construct a multi-objective push strategy model with the objective function as follows: ； in, To provide a comprehensive score for the push strategy, To enhance efficiency through knowledge. To enhance learning interest, The learning burden index, , , These are the weighting coefficients; Using heuristic graph search A The algorithm explores the optimal sequence of teaching content in the knowledge point intelligent graph. The evaluation function of the algorithm is: ； in, For nodes The evaluation value, From the starting node to the node The actual cost, For the node The estimated cost to reach the target node. For the estimated cost weighting coefficient; S42. The push system pushes teaching content to students through the learning terminal according to the generated push plan; S43. Collect student feedback data on the pushed content, and build a push effect evaluation model based on the feedback data. Quantitatively evaluate the content from three dimensions: content suitability, learning effectiveness improvement rate, and user experience. The content suitability evaluation formula is as follows: ； in, For knowledge adaptation, For capability adaptation, For preference fit, , , The weights for each item are: ； ； In the formula, A smoothing factor used to smooth out knowledge fit; The formula for evaluating the improvement rate of learning effectiveness is as follows: ； in, To assess the level of knowledge mastery after the push notification, To assess the level of knowledge mastery before pushing the content; The user experience rating formula is as follows: ； in, For the number of feedback recipients, For the first Student satisfaction ratings. As a weight for feedback credibility; Overall assessment score The evaluation indicators include content completion rate, knowledge point mastery improvement rate, user satisfaction score, and learning efficiency improvement rate. Based on the evaluation results, the push strategy was optimized using a reinforcement learning algorithm. The reinforcement learning model adopted the Q-learning algorithm, and the formula is as follows: ； in, For state Next action of value, For learning rate, As a reward value, As a discount factor, To perform the action The next state after that, For state The optimal action is as follows. To explore motivating factors.

8. A teaching content delivery system based on multi-source data analysis, used to implement the method described in any one of claims 1-7, characterized in that, include: The data acquisition module is used to collect multi-source heterogeneous teaching data, including academic data acquisition unit, behavioral data acquisition unit, cognitive data acquisition unit, environmental data acquisition unit and resource data acquisition unit, and has access control and data encryption functions; The data processing and fusion module is used for data preprocessing, feature alignment and deep fusion, and outputs a fused feature vector. The user profile building module is used to build multi-dimensional dynamic user profiles, and supports real-time updates and historical review of profiles. The knowledge point graph construction module is used to organize the knowledge point system and generate and dynamically update the intelligent knowledge point graph. The push strategy generation module is used to generate personalized push solutions based on user profiles and knowledge point graphs, adapting to multiple teaching scenarios; The push execution module is used to execute push tasks, supports multi-channel push linkage, and monitors push status; The feedback optimization module is used to collect feedback data, evaluate the push effect, and iteratively optimize the strategy; The storage module adopts a distributed storage architecture to store various types of data and has encryption, backup, and efficient retrieval functions.

9. The teaching content delivery system based on multi-source data analysis according to claim 8, characterized in that, The data processing and fusion module uses the TensorFlow framework to build a graph neural network fusion model, the push strategy generation module uses a heuristic graph search algorithm to generate push schemes, and the storage module uses a hybrid storage architecture of MySQL and MongoDB to support the classified storage of structured and unstructured data.

Images