Educational resource image intelligent recommendation and multi-scale matching system and method based on machine learning

Abstract

The invention discloses an educational resource image intelligent recommendation and multi-scale matching system and method based on machine learning, and relates to the technical field of intelligent education, a permeability evaluation model is constructed by collecting interactive behavior data of students and educational resource images to quantify cognitive stability and knowledge mastery parameters; synchronously performing multi-scale analysis on the image to generate a visual feature, a knowledge association feature and a cognitive guide feature; and determining an optimal recommendation level based on dynamic granularity selection processing, generating an enhanced image with a permeability feedback mark, and interactively generating learning evaluation data through the enhanced image. According to the method, cognitive dynamic evaluation and multi-scale feature matching are fused, so that the technical bottleneck of a traditional recommendation system on image granularity selection and cognitive adaptation is solved, and the educational resource recommendation accuracy and learning efficiency are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent education technology, and in particular to a machine learning-based intelligent recommendation and multi-scale matching system for educational resource images and a method thereof. Background Art

[0002] Traditional educational resource image applications generally have technical limitations, making it difficult to achieve accurate personalized recommendations. Existing methods have a single dimension for evaluating the user's knowledge mastery status, relying only on simple behavioral statistics or static test results. They are unable to dynamically quantify the depth of knowledge penetration and the difference in cognitive stability, making it difficult for the recommendation system to capture the dynamic fluctuation characteristics of the cognitive state. At the same time, image feature processing technology is mostly limited to visual content analysis and lacks modeling of knowledge logical associations and cognitive guidance capabilities. It is unable to construct semantic associations between multi-level features, nor can it combine cognitive state assessment results with feature analysis depth, resulting in a mismatch between recommendation results and users' actual learning needs. More prominently, the existing system adopts a fixed hierarchical display strategy for information granularity control. It is unable to dynamically adjust knowledge density according to the user's cognitive state, and it is difficult to balance the needs of global cognitive guidance and local knowledge focus. As a result, high-penetration users repeatedly receive redundant information and low-penetration users face the dilemma of knowledge overload.

[0003] In terms of dynamic granularity selection mechanisms, traditional methods have significant technical flaws. Existing recommendation systems usually use preset rules to divide image information levels. They neither establish a dynamic mapping relationship between penetration evaluation values and the proportion of feature effective areas, nor do they explore the teaching value of the synergistic effects of multi-scale features. For example, the feature mask generation rules under the fixed granularity division strategy are rigid and cannot adaptively shrink or expand the area of interest based on cognitive stability parameters, resulting in a misalignment between recommended content and cognitive needs. In addition, directional consistency metrics mostly use static similarity calculations, and do not introduce a penetration weighting mechanism to strengthen the feature matching priority of cognitive weak links, causing the recommendation level selection to deviate from the user's actual knowledge absorption rhythm. This technical defect makes it impossible for the recommendation system to achieve dynamic adaptation of "cognitive state-feature granularity-knowledge density", exacerbating the low efficiency of teaching resource utilization.

[0004] To address the above issues, this field urgently needs to build a dynamic coupling mechanism between permeability assessment and multi-scale features, and achieve accurate matching of recommendation granularity levels through regional adaptive adjustment of feature masks and cognitive weighting strategies of directional consistency measurement. Summary of the Invention

[0005] In view of the problem that the existing technology is difficult to adapt to the fluctuation of cognitive state during the learning process, the present invention provides an educational resource image intelligent recommendation and multi-scale matching system and method based on machine learning.

[0006] In order to achieve the above object, the technical solution of the present invention is as follows:

[0007] In a first aspect, the present invention discloses a method for intelligent recommendation and multi-scale matching of educational resource images based on machine learning, comprising the following steps:

[0008] Obtaining interactive behavior data between users and educational resource images;

[0009] Calculating a permeability evaluation value of a knowledge content-based permeability evaluation model according to the interactive behavior data;

[0010] Simultaneously perform multi-scale analysis on educational resource images to generate a multi-level image feature set including visual features, knowledge association features, and cognitive guidance features;

[0011] The permeability evaluation value and the multi-level image feature set are subjected to a dynamic granularity selection process to obtain the optimal recommended granularity level. The specific steps of the dynamic granularity selection process are as follows:

[0012] generating a feature mask for each scale level of the multi-level image feature set, wherein the feature mask dynamically adjusts the effective area ratio according to the permeability evaluation value;

[0013] Calculating the directional consistency measure of the feature mask at each level and the current knowledge state vector, and applying a permeability weighting coefficient;

[0014] According to the product of the directional consistency metric value and the permeability weighted coefficient,

[0015] selecting the optimal recommendation granularity level;

[0016] An enhanced recommendation image with a permeability feedback mark is generated according to the optimal recommendation granularity level, and learning evaluation data of the user is obtained after the user interacts with the enhanced recommendation image.

[0017] In a second aspect, the present invention discloses a system for intelligent recommendation and multi-scale matching of educational resource images based on machine learning, which uses the above-mentioned method for intelligent recommendation and multi-scale matching of educational resource images based on machine learning, including:

[0018] The data collection module is used to obtain user interaction behavior data and subsequent evaluation data in real time, and encapsulate them into structured data packets through preset protocols;

[0019] a permeability evaluation module, configured to calculate a permeability evaluation value of the knowledge content according to the structured data packet;

[0020] The multi-scale analysis module is used to simultaneously perform multi-scale analysis on educational resource images and generate a multi-level image feature set including visual features, knowledge association features, and cognitive guidance features;

[0021] The dynamic recommendation decision module generates the optimal recommended granularity level through dynamic granularity selection processing based on the permeability evaluation value and the multi-level image feature set;

[0022] An image enhancement module, configured to generate an enhanced recommended image with a permeability feedback mark according to the optimal recommendation granularity level;

[0023] The evaluation module is used to generate enhanced recommendation images with permeability feedback tags based on the optimal recommendation granularity level, and obtain user learning evaluation data after the user interacts with the enhanced recommendation image;

[0024] A feedback reinforcement module is used to collect learning evaluation data of the user after obtaining the enhanced recommended image recommendation, and to update the learning parameters of the permeability assessment model through a feedback reinforcement mechanism to form a closed-loop optimization system;

[0025] Among them, the data interfaces among the data acquisition module, the permeability assessment module, the multi-scale analysis module, the dynamic recommendation decision module, the image enhancement module and the evaluation module adopt a predefined standardized protocol for feature alignment.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] 1. The present invention achieves a breakthrough in the cognitive adaptability of the educational resource recommendation system through a dynamic granularity selection mechanism. Different from the traditional fixed-level recommendation strategy, a multi-level feature mask is dynamically generated based on the permeability evaluation value, so that the display ratio of the effective area of the image is adjusted in real time with the user's cognitive stability: when the permeability evaluation value is low, the global composition area is expanded to strengthen the knowledge framework cognition; when it is high, the mechanism annotation details are focused to deepen knowledge penetration, effectively solving the problem of knowledge redundancy or cognitive overload caused by the rigidity of information density in traditional methods. Combining the collaborative calculation of directional consistency measurement and permeability weighting coefficient, the matching priority of each level of features and the user's knowledge status is accurately quantified, breaking through the mechanical matching limitations of traditional similarity calculations, and realizing the dynamic fit between the recommendation granularity level and the cognitive absorption rhythm.

[0028] 2. By integrating graph neural networks associated with knowledge graphs and attention mechanisms, a three-dimensional feature space of visual features, logical relationships between knowledge points, and cognitive guidance signals is constructed. By integrating the knowledge association matrix, visual semantic features, and cognitive guidance vectors, a multi-level feature system with teaching logic coherence is constructed. This enables the recommendation system to not only analyze the visual content of the image, but also capture the topological relationships and cognitive evolution paths between knowledge points, transforming the abstract knowledge penetration state into an interactive and explicit guidance interface, prompting the active presentation of knowledge points in low-penetration areas, and improving the efficiency of user cognitive correction, which is significantly higher than traditional static labeling systems.

[0029] 3. This invention forms a closed-loop optimization system based on dynamic granularity selection and feedback reinforcement mechanisms. Feature mask generation rules and a permeability-weighted similarity calculation strategy enable intelligent adaptation of image display granularity to cognitive state. This technology upgrades educational resource recommendations from a static matching system to a dynamic evolutionary one, significantly reducing cognitive load while improving knowledge transfer efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] The disclosure of the present invention is described with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. In the drawings, the same reference numerals are used to refer to the same components. Among them:

[0031] Figure 1 is a flow chart of the steps of the present invention;

[0032] Figure 2 A flow chart of feature mask generation rules for the present invention;

[0033] Figure 3 Enhanced feedback logic diagram for the method of the present invention;

[0034] Figure 4 is a system flow chart of the present invention;

[0035] Figure 5 It is a system flow demonstration diagram of the present invention. DETAILED DESCRIPTION

[0036] It is easy to understand that according to the technical solution of the present invention, without changing the essential spirit of the present invention, a person skilled in the art can propose a variety of interchangeable structural modes and implementation modes. Therefore, the following specific embodiments and drawings are only exemplary descriptions of the technical solution of the present invention and should not be regarded as the entire invention or as a limitation or restriction of the technical solution of the present invention.

[0037] Application Overview:

[0038] Existing technologies often rely on static profile matching or single-dimensional behavioral analysis to personalize educational resources, making it difficult to dynamically capture the evolving patterns of cognitive states. Traditional methods employ a fixed granularity level for feature analysis, failing to adjust the density of recommended information based on students' real-time knowledge penetration. This leads to a mismatch between cognitive load and knowledge needs. Existing systems lack multidimensional quantitative assessments of interactive behaviors. When unusual behaviors such as frequent zooming and hotspot jumping occur, recommendation results are prone to adaptation bias, making it difficult to meet personalized learning needs.

[0039] The inventors discovered a complementary relationship between cognitive state representations in interactive behaviors, including zoom stability, hotspot dwell time, and annotation correction frequency. By constructing a penetration assessment model, they achieved collaborative analysis of multi-dimensional behavioral characteristics. The study found that image zoom variance is negatively correlated with cognitive stability, that hotspot dwell time increases exponentially with knowledge mastery, and that annotation correction frequency reflects the intensity of cognitive conflict above a specific threshold. This led to the proposal of a three-level feature fusion mechanism based on dynamic granularity selection, which overcomes the limitations of single-dimensional matching in traditional recommendation systems.

[0040] Specifically, the interactive behavior data stream including image zoom trajectory, hotspot residence hotspots and annotation correction logs is first captured in real time. The cognitive stability parameter (zoom level variance) and knowledge mastery parameter (logarithmic transformation value of residence time + correction frequency segment weight) are calculated through the penetration evaluation model to generate a multi-dimensional state vector. The educational resource image is simultaneously analyzed at multiple scales to extract visual semantic features, knowledge graph association features and cognitive guidance features. When it is detected that the penetration evaluation value is in different intervals of the dynamic threshold, the granularity selection process is triggered: the directional consistency measure of the features of each level and the current knowledge state is calculated, and weighted fusion is performed after applying the penetration attenuation factor to automatically select the recommendation level with the optimal information density. The enhanced recommendation image finally generated is superimposed with thermal feedback marks and knowledge key points floating layers, and the evaluation data is learned through feedback with the enhanced recommendation image.

[0041] Compared with existing technologies, traditional recommendation systems use fixed feature weights and static user profiles, making it difficult to adapt to cognitive state fluctuations during the learning process. This solution establishes a dynamic feature fusion mechanism driven by penetration, achieves intelligent matching of multi-scale features through directional consistency measurement, and combines dual-channel feedback (behavioral data + evaluation results) to build a closed-loop optimization system. Compared with single-dimensional analysis methods, this solution improves recommendation accuracy, enhances cognitive load adaptability, and enhances recommendation stability in abnormal behavior scenarios, providing a reliable technical paradigm for the precise delivery of personalized educational resources.

[0042] After introducing the basic concept of the present invention, embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0043] Example 1:

[0044] like Figure 1 As shown in FIG, the method for intelligent recommendation and multi-scale matching of educational resource images based on machine learning includes the following steps:

[0045] Obtaining user interaction data with educational resource images, including image zoom level records, duration of stay in hot spots, and number of annotation modifications;

[0046] Calculate the penetration evaluation value of the knowledge content-based penetration evaluation model based on the interactive behavior data;

[0047] Simultaneously perform multi-scale analysis on educational resource images to generate a multi-level image feature set including visual features, knowledge association features, and cognitive guidance features;

[0048] The permeability evaluation value and the multi-level image feature set are processed through dynamic granularity selection to obtain the optimal recommended granularity level. The specific steps of dynamic granularity selection are as follows:

[0049] Generate a feature mask for each scale level of the multi-level image feature set, and dynamically adjust the effective area ratio of the feature mask according to the permeability evaluation value;

[0050] Calculate the directional consistency measure between the feature mask at each level and the current knowledge state vector, and apply the permeability weighting coefficient;

[0051] The optimal recommended granularity level is selected based on the product of the directional consistency metric value and the permeability weighting coefficient;

[0052] An enhanced recommendation image with permeability feedback mark is generated according to the optimal recommendation granularity level, and the user's learning evaluation data is obtained after the user interacts with the enhanced recommendation image.

[0053] Among them, the cognitive stability parameter is calculated based on the variance of the image zoom level, and the knowledge mastery parameter is determined jointly by the duration of hotspot stay and the frequency of annotation modification.

[0054] Among them, interactive behavior data refers to the digital operation trajectory recorded when students interact with educational resource images, specifically including image zoom level records, hot spot area residence time distribution and annotation modification frequency statistics, which can be obtained through browser event monitoring API or mobile touch log collection technology, and is used to quantify students' attention intensity and cognitive depth to knowledge units. Image zoom level records refer to the numerical information recorded by the system when users zoom in / out while browsing images, reflecting the user's exploration depth of image details (high-frequency zooming indicates cognitive confusion). Hot spot area residence time refers to the user's gaze or mouse hover time on a specific area of the image (such as a formula derivation step diagram), which is used to quantify the user's attention intensity to key knowledge points. Annotation modification frequency refers to the number of times users edit preset annotations (such as arrows and text descriptions) on the image, reflecting the intensity of users' understanding conflicts of knowledge points.

[0055] Among them, the penetration assessment model refers to a composite assessment framework constructed based on cognitive stability parameters and knowledge mastery parameters. The specific implementation method is: cognitive stability is represented by calculating the variance value of the image zoom level, and knowledge mastery is represented by the fusion value of the logarithmic transformation value of the hotspot residence time and the segmented weighting of the annotation modification frequency. The weight coefficient is trained using a multivariate linear regression model or XGBoost algorithm to dynamically evaluate the degree of students' absorption of knowledge content.

[0056] Among them, multi-scale analysis refers to the technical process of cross-dimensional feature extraction of educational resource images. The specific implementation steps include: extracting global visual semantic features through the pre-trained ResNet-50 network; generating a knowledge association matrix based on knowledge graph node embedding; using the attention mechanism to fuse the permeability assessment value and image features to generate cognitive guidance features, forming a three-level feature system including composition hierarchy, functional zoning, and mechanism annotation.

[0057] Dynamic granularity selection refers to a decision-making strategy that adaptively adjusts the density of recommended information based on cognitive state. The specific workflow is as follows: generating an adjustable binary mask matrix for each feature level; calculating the cosine similarity between the feature vectors of each level and the current knowledge state vector; applying a permeability attenuation coefficient for priority sorting, and using a random forest classifier or gradient boosting decision tree to implement level switching decisions, which is used to balance knowledge focus requirements and cognitive load thresholds. The formula is as follows:

[0058]

[0059] Among them, f l is the lth layer feature vector, s k is the knowledge state vector, P k is the permeability evaluation value, and δ is the preset normalization factor.

[0060] Among them, enhanced recommended images refer to the optimized output results of superimposed dynamic visual marks. The specific generation methods include: superimposing a semi-transparent heat map on the basic image to display the historical hotspot distribution; using SVG vector graphics to render the floating layer of key knowledge points; triggering the annotation flashing animation when a permeability mutation area is detected, and realizing real-time interactive display through the WebGL rendering engine to strengthen the visual guidance of cognitive weaknesses.

[0061] Among them, learning evaluation data refers to the evaluation data feedback mechanism, and the specific implementation logic is: the difference in the test score change rate before and after the recommendation is input into the hyperbolic tangent function for normalization; the parameter adjustment gradient is calculated based on the actual usage time of the recommended image, and the incremental learning algorithm or online learning framework is used to implement model iteration to maintain the dynamic adaptability of the recommendation strategy, and obtain learning evaluation data of the user's learning effect after enhancing the recommended image.

[0062] This application constructs a multi-scale feature dynamic matching mechanism based on penetration assessment. By integrating quantitative analysis of cognitive states, multi-dimensional semantic analysis of images, and closed-loop feedback optimization, it breaks through the technical limitations of static portrait matching and fixed information density of traditional education recommendation systems, and achieves precise adaptation of educational resources and cognitive dynamics.

[0063] The working process and principles of this solution are as follows: first, real-time data on students' interactions with educational resource images is collected, and cognitive stability and knowledge mastery parameters are calculated using a permeability assessment model. Simultaneously, multi-scale analysis of educational resource images is performed to generate visual features, knowledge association features, and cognitive guidance features. Dynamic granularity selection is used to calculate directional consistency metrics between features and cognitive states at each level, and the optimal recommendation granularity level is selected using a permeability weighting coefficient. Enhanced recommendation images with thermal feedback tags are generated. Finally, learning assessment data is obtained from the enhanced recommendation images, forming a learning assessment process of "behavior collection - state assessment - feature matching - feedback data." This technical system significantly improves the personalized adaptation efficiency of educational resources and the effectiveness of knowledge transfer by real-time sensing of cognitive state fluctuations, intelligently adjusting information density, and implementing learning recommendation strategies.

[0064] This application further proposes that the penetration evaluation value includes a cognitive stability parameter and a knowledge mastery parameter. The calculation formula of the penetration evaluation value is as follows:

[0065]

[0066] Among them, D i is the standard stay time of the i-th knowledge unit, τ is the actual stay time, represents the quality of focus item, ΔS is the rate of change of the associated assessment score, β·log(1+ΔS) represents the learning gain item, Var(Z) is the scaled level variance, γ·Var(Z) represents the cognitive fluctuation item, and α, β, and γ are preset weight coefficients.

[0067] The calculation of the penetration assessment value is a composite indicator that quantifies cognitive stability and knowledge mastery, with a weighting coefficient of α + β + γ = 1. The penetration assessment value is a composite indicator that quantifies a student's understanding of the content. Its calculation incorporates multiple characteristics, including the duration of behavioral engagement, the complexity of interactive operations, and the rate of change in assessment scores. The penetration assessment model constructs a dynamic quantitative system by integrating cognitive stability parameters with knowledge mastery parameters. Its calculation process supports the collaborative analysis of multidimensional behavioral characteristics and an adaptive weight adjustment mechanism.

[0068] The cognitive stability parameter refers to the degree of cognitive fluctuation, as measured by the variance of the zoom levels in image interaction behaviors. This is implemented using a sliding window variance calculation algorithm. For example, the discreteness of the zoom level sequence is statistically analyzed within a 30-second time window, combined with double exponential smoothing to eliminate transient operational noise. This allows for real-time monitoring of students' cognitive stability within a knowledge unit. The knowledge mastery parameter is a quantitative indicator of learning effectiveness that combines the dwell time of a hotspot with annotation correction behavior. This is achieved by applying a logarithmic transformation to the raw dwell time to suppress the influence of extreme values, while also applying a piecewise weighting strategy to the frequency of annotation corrections (low-frequency corrections are assigned a high weight, while high-frequency corrections trigger an attenuation factor). Finally, a 0-1 standardized score is generated through linear fusion. The dynamic adjustment mechanism for weight coefficients is a closed-loop system that automatically optimizes the parameter contribution ratio based on the characteristics of the learning stage. This is implemented using a reinforcement learning framework: a feature vector is constructed that includes the rate of change of assessment scores, the difficulty level of the knowledge unit, and historical behavior patterns. The DQN algorithm is then used to dynamically output the optimal weight combination of α, β, and γ, ensuring that the evaluation model remains consistently adapted to the current learning scenario.

[0069] Specifically, the system captures raw data streams such as zoom levels, residence hotspots, and annotation logs in real time through an embedded behavior acquisition module. The cognitive stability calculation unit calculates the variance of zoom operations with millisecond-level timing accuracy. When it detects that the variance value exceeds the adaptive threshold, it automatically activates the noise filtering algorithm to eliminate interference from false touch operations. The knowledge mastery analysis engine runs a dual-channel processing flow simultaneously: the residence time analysis channel uses time series decomposition technology to extract effective learning periods, and the annotation correction analysis channel uses an operation intention recognition model to distinguish effective corrections from redundant operations. After receiving the above parameters and environmental features, the weight decision module generates a dynamic weight coefficient based on the pre-trained deep Q network, and finally outputs a penetration assessment value. This value is pushed to the recommendation decision system in real time through the message queue, triggering feature matching and hierarchical selection operations, and is written to the time series database at the same time.

[0070] Compared with existing technologies, traditional cognitive assessment models mostly use static weight allocation strategies, which are difficult to adapt to the changing needs of different knowledge types and learning stages. This solution introduces a dynamic weight adjustment mechanism and realizes intelligent optimization of parameter contribution ratios through a reinforcement learning framework. Compared with the fixed weight system, this solution improves assessment accuracy and improves assessment stability several times in interdisciplinary knowledge scenarios. Existing systems usually process various behavioral characteristics independently and lack time series correlation analysis and noise filtering mechanisms, resulting in a decrease in the evaluation signal-to-noise ratio in high-interaction frequency scenarios. This solution improves the anti-interference ability of the evaluation results through the collaborative design of sliding window variance calculation and dual-channel filtering technology.

[0071] Through the above technical solution, this application has constructed a dynamic penetration evaluation system with environmental perception capabilities. Cognitive stability parameters and knowledge mastery parameters are independently calculated through orthogonalized feature extraction channels to avoid evaluation bias caused by dimensional coupling; the weight coefficient dynamic adjustment module relies on the deep reinforcement learning framework to enable the evaluation model to autonomously adapt to complex scenarios such as knowledge difficulty transitions and learning mode switching; the combination of multi-level noise filtering mechanism and real-time data stream processing architecture ensures that the evaluation process has both high sensitivity and strong anti-interference capabilities. This solution enables the real-time performance of cognitive state assessment to reach millisecond-level response accuracy, providing a highly reliable quantitative decision-making basis for personalized learning path planning.

[0072] This application further proposes that multi-scale analysis includes:

[0073] Obtain visual features and basic image features of educational resource images through a pre-trained visual feature extraction network;

[0074] Graph neural network based on knowledge graph association extracts knowledge association features of educational resource images;

[0075] The attention mechanism is used to fuse the basic image features of the permeability evaluation value with the knowledge association features, and generate cognitive guidance features;

[0076] Among them, the edge weights associated with the knowledge graph are dynamically adjusted according to the dependency relationships of knowledge points preset in the teaching syllabus.

[0077] The visual feature extraction network is a pre-trained convolutional neural network optimized based on course knowledge distillation. It employs a ResNet-152 backbone network combined with course difficulty grading labels for transfer learning. Through a feature pyramid structure, it outputs multi-granular visual semantic features covering global composition, local functional areas, and pixel-level mechanism annotations, capturing the spatial semantic distribution patterns of educational resource images. The knowledge association feature extraction module is a dynamic graph neural network built on the syllabus knowledge graph. It is implemented by encoding knowledge point dependencies into a weighted adjacency matrix, aggregating cross-chapter knowledge point embedding vectors using a graph attention mechanism, and fusing temporal teaching progress features via a gated recurrent unit to generate a spatiotemporal fusion feature matrix reflecting knowledge topological associations. The cognitively guided feature generation mechanism is an attention fusion model driven by permeability assessment. This is implemented through a multi-headed cross-attention layer: Using permeability assessment values as query vectors, visual features as key vectors, and knowledge association features as value vectors, a cross-modal attention weight matrix is constructed to adaptively focus on instructional semantic regions that match the current cognitive state.

[0078] Specifically, a three-step parsing process is executed in parallel through a distributed computing framework: the visual parsing channel uses the quantized, accelerated TensorRT engine to extract 1024-dimensional visual feature vectors in real time; the knowledge association channel dynamically loads knowledge point topological relationships based on the Neo4j graph database and utilizes the DGL graph neural network framework to generate 768-dimensional knowledge embedding vectors; and the cognitive guidance channel deploys a lightweight Transformer model to encode the permeability assessment value sequence into a 128-dimensional state vector, dynamically weighting the visual and knowledge features through a cross-modal attention mechanism. After the output features of each channel are aligned in dimension through a normalization layer, they are input into a hierarchical selection decision system to generate a multi-scale feature tensor. Finally, a feature compression algorithm is used to generate a cognitive guidance heat map adapted to the display device resolution.

[0079] Compared with existing technologies, traditional educational resource parsing solutions often rely on single visual features or static knowledge labels, making it difficult to dynamically adapt instructional semantics to cognitive states. This solution, however, leverages a feature fusion mechanism and cognitive guidance model to improve the alignment of recommended content with learners' cognitive states. Existing graph neural network processing methods use fixed knowledge point edge weights and are unable to adapt to dynamic adjustments to the syllabus. This solution utilizes a real-time adjacency matrix update algorithm to reduce the latency in updating knowledge association features.

[0080] Through the above technical solutions, this application constructs a multi-scale parsing system for teaching scene perception. The visual feature extraction network breaks through the traditional model's reliance on general image features and obtains visual semantic understanding capabilities unique to the education field through course knowledge distillation. The dynamic graph neural network expands the association of static knowledge points into spatiotemporal fusion features that include teaching progress, knowledge difficulty, and historical mastery. The cognitive guidance mechanism constructs a dynamic mapping relationship between cognitive state and teaching semantics through penetration evaluation values, achieving precise feature focus of "what you see is what you need".

[0081] This application further proposes that the dynamic adjustment of edge weights associated with the knowledge graph includes:

[0082] Generate initial edge weights based on the pre-dependencies of knowledge points in the syllabus;

[0083] Based on the knowledge path selection data of the student group, the edge weights are dynamically updated through the reinforcement learning algorithm;

[0084] Among them, the knowledge path selection data is calculated by counting the frequency of students jumping across knowledge points and the ratio of their stay time.

[0085] The initial edge weight generation mechanism is a structured quantitative modeling method based on the dependencies between knowledge points in the syllabus. This is achieved by encoding the pre-dependency graph of knowledge points into a weighted adjacency matrix. A knowledge distillation algorithm is used to map teaching parameters such as course difficulty, teaching phase division, and cognitive load level into initial weight vectors. Multi-factor normalization is then used to precisely adapt the weight values to the actual teaching scenario. The reinforcement learning dynamic update algorithm is an online optimization model based on the deep deterministic policy gradient (DDPG) framework. Its state space is constructed from group behavioral characteristics such as knowledge point jump frequency, dwell time distribution, and the rate of change in associated assessment scores. The action space is designed as a matrix of incremental adjustment coefficients for edge weights. The reward function integrates three perspectives: knowledge path selection efficiency, learning effect improvement rate, and abnormal jump suppression. The knowledge path selection data mining module is a real-time analysis engine for large-scale learning behavior logs. It uses a time series pattern mining algorithm to extract cross-knowledge point jump trajectory features. An improved PageRank algorithm is used to quantify the correlation weight between knowledge point dwell time and learning effect. This constructs a standardized student behavior feature tensor for the reinforcement learning model to use.

[0086] Specifically, a distributed stream processing platform ingests student interaction behavior data in real time. A window aggregation engine compiles a statistical matrix of knowledge point jump frequencies, which is then combined with dwell time records in a time series database to generate group behavior signatures. Simultaneously, the system receives structured syllabus data and uses the Analytic Hierarchy Process (AHP) to calculate the initial weight distribution of dependencies between knowledge points. By regularly acquiring the latest behavior signatures and the current edge weight status, the policy network outputs a weight adjustment gradient within a certain step size, which is then updated by the executor in the knowledge graph adjacency matrix.

[0087] Compared with existing technologies, traditional knowledge graph construction methods rely on manually set fixed edge weights, which limits the mining of interdisciplinary knowledge associations and makes them unable to adapt to the evolution of group learning patterns. However, this solution improves the matching degree between knowledge point recommendation paths and actual learning behaviors through a dynamic adjustment mechanism driven by reinforcement learning. This solution uses an online reinforcement learning framework to achieve continuous tuning at the minute level, thereby improving the rationality score of knowledge point path jumps. Existing behavioral data analysis methods are limited to single jump frequency statistics and ignore the temporal correlation characteristics in the time dimension. This solution improves the accuracy of path rationality prediction through time series pattern mining and behavioral feature tensor modeling.

[0088] Through the above technical solutions, this application constructs a knowledge graph self-evolution system for intelligent education scenarios. The initial weight generation mechanism transforms the experience of teaching experts into a standardized quantitative model, the reinforcement learning engine deeply analyzes the temporal evolution of group behavior, and the weight reconstruction module achieves dynamic alignment of knowledge topology and learning behavior characteristics. This application can be deployed on an online education platform with 10,000 students, with a high accuracy rate for matching students' first learning paths and improving the efficiency of discovering interdisciplinary knowledge associations, effectively solving the problems of lagging personalized recommendations and rigid paths caused by the static nature of traditional knowledge systems.

[0089] This application further proposes that the generation of cognitive guidance features includes:

[0090] Gridding the knowledge association features to generate a knowledge association feature matrix;

[0091] Perform matrix product operation on the permeability evaluation value vector of the permeability evaluation value and the knowledge association feature matrix, and normalize the product result to obtain the attention weight matrix;

[0092] The attention weight matrix is probabilized through the Softmax function;

[0093] The dimensions of the knowledge association feature matrix and the dimensions of the penetration evaluation value vector are matched through preset dimension alignment rules.

[0094] The generation formula of cognitive guidance features includes:

[0095] Perform matrix multiplication of the penetration evaluation value vector and the knowledge association feature to obtain the attention weight matrix:

[0096]

[0097] Among them, P k is the penetration evaluation value vector, K is the knowledge association feature matrix, and d is the preset dimension scaling factor.

[0098] Among them, the gridding processing of knowledge-related features refers to mapping the output features of unstructured graph neural networks into a regularized semantic matrix. Specifically, the spatial pyramid pooling technology is used to convert the variable-length knowledge embedding vector into a three-dimensional feature grid of fixed size (such as 8×8×256). The topological relationship information of knowledge points is injected through a learnable position encoder to form a unified spatiotemporal semantic representation space. Among them, the dynamic attention weight calculation refers to a collaborative similarity measurement method based on cognitive state and knowledge semantics. The specific implementation method is: mapping the permeability evaluation value vector to a high-dimensional semantic space through a multi-layer perceptron, performing multi-head cross-attention calculation with the knowledge-related feature matrix, introducing an adaptive scaling factor based on the difficulty of the knowledge unit, and suppressing the interference of the noise dimension on the attention distribution. Among them, the dimensional alignment optimization mechanism refers to dynamically adjusting the dimensional matching relationship of the feature space. It is specifically implemented through a differentiable resampling algorithm: constructing a bidirectional LSTM network to predict the dimensional difference coefficient between the permeability vector and the knowledge matrix, and using a dynamic routing algorithm to generate a dimensional transformation matrix to eliminate the dimensional mismatch problem during cross-modal feature fusion.

[0099] Specifically, the knowledge graph parsing pipeline generates a dynamic knowledge association matrix in real time. This matrix is fed into a gridded processing unit after extracting multi-scale semantic features through a three-dimensional convolutional network. The permeability assessment model outputs a time-series state vector, which is encoded into a query vector via a gated attention network. The dynamic dimension alignment engine continuously monitors dimensional differences in the feature space and generates a real-time scaling factor to adjust the attention calculation process. Finally, the knowledge association feature matrix and the permeability vector perform multi-head attention calculations within the optimized semantic space. Probabilistic weight allocation focuses on key knowledge areas, generating a cognitively guided heat map that directly drives the recommendation system's visualization rendering engine.

[0100] Compared with existing technologies, traditional cognitive feature fusion methods generally use static weight allocation or simple dot product similarity calculation, resulting in large errors in the identification of knowledge hotspot areas. However, this solution improves feature matching accuracy through the collaborative design of dynamic dimension alignment and adaptive scaling factors. The existing attention mechanism requires manual setting of fixed scaling factors when processing variable-dimensional features. This solution uses differentiable resampling technology to achieve autonomous optimization of dimensional space, thereby improving the feature information retention rate. Existing grid processing methods rely on fixed pooling strategies and cannot adapt to the dynamic evolution of knowledge topology. This solution reduces the spatiotemporal consistency error of knowledge association features by two orders of magnitude through the joint optimization of spatial pyramid pooling and position encoding.

[0101] Through the above technical solution, this application constructs a dynamic attention fusion system driven by cognitive state. The grid processing of knowledge-related features breaks through the dimensional limitations of traditional feature expression, the dynamic alignment mechanism realizes autonomous adaptation of cross-modal semantic space, and the probabilistic weight distribution accurately focuses on cognitively weak areas. In the deployment of this technology on an online education platform with 10,000 users, the accuracy of knowledge point recognition is improved, the visual salience score of cognitive guidance features is improved, and the matching efficiency of recommended content with cognitive needs is improved, verifying the engineering practicality of the technical solution in complex educational scenarios.

[0102] like Figure 2 As shown, the present application further proposes that the generation rules of the feature mask include:

[0103] When the permeability evaluation value is lower than a first preset threshold, all features of the global composition area of the image are activated;

[0104] When the permeability evaluation value is between the first preset threshold and the second preset threshold, retaining the edge contour features of the functional partition;

[0105] When the permeability evaluation value is higher than a second preset threshold, only the mechanism annotation detail feature area is activated;

[0106] Among them, the first preset threshold and the second preset threshold are critical values preset according to the teaching scenario type.

[0107] The generation rules of feature masks construct an adaptive matching system between cognitive state and knowledge presentation density through a dynamic hierarchical activation mechanism driven by penetration, solving the common problem of information overload or insufficiency in the presentation of traditional educational resources.

[0108] The specific generation logic for the first preset threshold (T1) and the second preset threshold (T2) is as follows: Based on the teaching objectives of the knowledge unit (such as concept understanding, skill training, and comprehensive application), the pre-trained neural network model predicts the student's cognitive load limit, and combines hardware parameters such as device display resolution and ambient light intensity to generate a dynamic threshold surface. For example, in theoretical teaching scenarios, (T1 = 0.3) and (T2 = 0.7) focus on global concept guidance; while in experimental operation scenarios, (T1 = 0.5) and (T2 = 0.9) strengthen the focus on detailed steps.

[0109] Among them, global composition region feature activation refers to the strategy of retaining the overall semantic structure of educational resource images. The specific implementation is: when (P KWhen (T1), the Spatial Pyramid Pooling technique is used to extract multi-scale global features of the image. The feature vectors are mapped to the full image space through the region growing algorithm to generate a binary mask matrix covering the entire image area (with all mask values being 1), ensuring that students establish a complete understanding of the knowledge framework.

[0110] Among them, the retention of the edge contour features of the functional partition refers to the refined information screening strategy for the medium cognitive state. When (T1 ≤ P K ≤ T2), an improved Canny edge detection algorithm is deployed, combined with the Non-local Attention mechanism to enhance the edges of the image functional partitions (such as the operation area of experimental instruments and the logical chain of formula derivation). The noise contours are removed through morphological opening operations to generate a sparse mask matrix containing only the key structural boundaries (non-edge areas are set to 0, and edge areas are set to 1).

[0111] Among them, the activation of the detailed feature regions of the mechanism annotation refers to the sub-pixel level information focusing strategy for the high cognitive state. When (P K > T2), the U-Net semantic segmentation network is run to perform pixel-level recognition on the mechanism annotation regions in the image (such as chemical molecular structural formulas and physical formula derivation steps). The boundary accuracy is optimized through CRF (Conditional Random Field) post-processing to generate a refined mask matrix containing only the detailed annotation regions (annotation regions are set to 1, and the remaining regions are set to 0), guiding students to deeply understand the knowledge core. <​​​​​​​​​​

[0116] Among them, the group average mastery is calculated based on the mean of the penetration assessment values of all students, and the individual difference index is calculated based on the deviation between the current student's historical interaction behavior data and the group data.

[0117] The generation of enhanced recommended images is achieved through the integration of dynamic visualization driven by penetration and group intelligent feedback technology, building an intelligent rendering system with cognitive guidance and social reference functions.

[0118] The superposition of dynamic visualization elements adopts the "penetration-visual parameter" mapping mechanism. Specifically, the penetration evaluation value (reflecting the depth of students' mastery of the current knowledge content, with a value range of 0-1, 0 for complete lack of mastery, and 1 for complete mastery) is used as a driving variable to establish a positive correlation function relationship with the display intensity of the visualization element (such as color saturation, animation frequency, and layer transparency). For example, when the penetration evaluation value is 0.3, the visualization element may be presented as a low-saturation light blue flash (frequency 0.5Hz); when the evaluation value is increased to 0.8, the element turns into a high-saturation warm red that is always on (transparency 20%). This mapping is achieved through a pre-trained visual parameter calibration model to ensure that the visual intensity changes corresponding to different penetration intervals are consistent with the sensitive threshold of human visual perception (such as the brightness change step is controlled within 5 nits to avoid visual fatigue).

[0119] In a specific embodiment, the interactive permeability prompt control adopts a "dual-indicator-multi-dimensional" display architecture. The control is integrated into the edge of the image (such as a floating window in the upper right corner) and contains two core data:

[0120] Group average mastery: By calling the education platform database in real time, the average penetration assessment value of all students in the current course is calculated (formula: group mean = Σ(individual penetration value) / total number of students). The data is updated every 5 minutes (taking into account both real-time performance and computational efficiency), and the error is controlled within ±0.02.

[0121] Individual Difference Index: This index is based on the Mahalanobis distance between the current student's recent 30 interactive behavior data (including 20+ dimensional features such as image browsing duration, dwell time in key areas, and number of annotation modifications) and the group benchmark data (the group's typical behavior feature vector extracted through principal component analysis). The formula is: Difference Index = √[(X-μ)^TΣ-1(X-μ)], where X is the individual feature vector, μ is the group mean vector, and Σ is the group covariance matrix. This index quantifies the degree of deviation between the current student's learning behavior and the group's typical pattern. A larger value indicates a more personalized learning path.

[0122] The generation of enhanced recommendation images is not a one-way output; rather, it is deeply embedded in a closed-loop process of "data collection-evaluation-visualization-interaction-re-evaluation." The intensity parameters of dynamic visualization elements are directly derived from the real-time output of the penetration assessment model (which is trained based on historical student interaction data and updates its parameters with each learning task completed). The group average mastery and individual difference index are synchronized with the education platform database in real time via an API. When a new student completes a learning task and generates a penetration value, the database triggers a recalculation of the average. When a current student performs a new interaction (such as zooming in on a specific area of the image), the system automatically extracts new features and updates the individual difference index. When a student clicks an interactive control, the system triggers a "drill-down query" function. For example, clicking on the group average mastery value reveals the distribution of subgroup means across different classes and learning stages. Clicking on the individual difference index reveals specific deviations in the behavioral dimension (e.g., "Time spent in a key area is 30% longer than the group average"). These interaction data are recorded and fed back into the penetration assessment model to optimize subsequent assessment accuracy (for example, if a student frequently clicks on the difference index control, the model can increase the weight of that student's historical behavior data).

[0123] Compared with traditional static recommendation images (which only display knowledge content and lack learning status feedback), this application uses the collaborative design of dynamic visualization and interactive controls. The dynamic visualization elements convert abstract penetration values into "visible" learning progress. Students do not need to consult complex reports and can quickly judge their own mastery level by the color and brightness changes of elements in the image. The real-time display of the group average mastery and individual difference index constructs a two-dimensional reference system of "self-group". Students can not only locate their position in the group (such as "My mastery is 15% higher than the class average"), but also identify personalized learning characteristics (such as "I prefer to repeatedly review image details rather than the overall structure"), so as to adjust their learning strategies in a targeted manner. The interaction data between users and controls is reversely input into the penetration evaluation model, forming a positive cycle of "evaluation-feedback-optimization" - the model analyzes the different dimensions that students care about (such as whether they care more about dwell time or number of annotations), and dynamically adjusts the weights of evaluation indicators, so that subsequent penetration calculations are more in line with students' actual learning needs.

[0124] Through the enhanced recommendation image generation technology of the above technical solution, through the deep integration of dynamic visual translation, real-time group comparison and interactive feedback mechanism, not only the functional upgrade of recommendation images is achieved, but also an efficient interactive bridge between students, systems and knowledge content is built, providing a new technical paradigm for intelligent recommendation in personalized education scenarios.

[0125] The present application further proposes that the dynamic visualization elements include:

[0126] Display key knowledge animations in areas where the permeability assessment value is lower than a preset threshold. The playback frequency of key knowledge animations is negatively correlated with the permeability assessment value. Display key knowledge animations in areas where the permeability assessment value is lower than a preset threshold.

[0127] Generate color depth gradient markers based on the hotspot dwell time distribution, where the color depth is proportional to the hotspot dwell time;

[0128] The correlation formula derivation path pops up in the permeability evaluation value mutation area, and the mutation area is determined by detecting the second-order derivative change of the permeability evaluation value through the sliding window algorithm.

[0129] The dynamic triggering mechanism for key knowledge animations is an adaptive frequency control algorithm based on a Gaussian mixture model (GMM). This is achieved by inputting the penetration assessment value into a pre-trained neural network prediction model, which outputs the initial playback frequency of the key knowledge animation. By monitoring the animation's click-through rate and dwell time in real time, the dynamic time warping (DTW) algorithm is used to adjust the negative correlation slope coefficient between playback frequency and penetration value, ensuring that the animation trigger frequency in low-penetration areas (P < 0.3) reaches 2-5Hz, creating sustained cognitive stimulation. Generating color depth gradient labels based on the dwell time distribution of hotspots is a color mapping engine that incorporates a spatiotemporal attention mechanism. This engine utilizes HSV color space conversion technology to nonlinearly map the dwell time (t) of hotspots to two channels: hue (H) and value (V). The H channel linearly allocates the color spectrum (blue → red) according to the t value, while the V channel exponentially enhances the brightness of high-dwelling areas using a function (V = 1-exp(-kt)). Sub-pixel gradient heatmaps are generated in real time using WebGL shaders. The mutation perception pop-up formula derivation refers to an intelligent prompt framework based on differential dynamic logic. Specifically, the second-order derivative absolute value sequence of the permeability evaluation value is calculated through a sliding window (window size N = 10). When a threshold value (|d 2 P / dt 2 |>0.1), a pop-up window showing the derivation path of the associated formula is rendered. The Bezier curve path planning algorithm is used to dynamically generate a formula derivation flowchart, and touch gesture recognition technology is combined to achieve interactive expansion and folding of the derivation steps.

[0130] Specifically, the system collaboratively processes visualization tasks through a distributed rendering cluster: After receiving the permeability assessment stream, the animation engine uses a particle system to simulate the emergence of key knowledge points. The particle motion trajectories are dynamically adjusted by predicting the learner's visual focus shift trend through an LSTM network. The hotspot residency matrix in the behavior log is simultaneously parsed, and the color gradient distribution for each frame is generated through CUDA-accelerated parallel computing. The permeability assessment value runs a real-time differential calculation pipeline, using an FPGA hardware-accelerated second-order derivative filter (cutoff frequency 5Hz) to quickly identify cognitive transition points. Finally, light field fusion technology is superimposed on the base image, and alpha blending and depth buffer optimization are used to eliminate the interference of overlapping visual elements, completing 4K resolution frame synthesis within 8ms.

[0131] Compared with existing technologies, traditional visualization solutions use static thresholds to trigger animation playback, resulting in a mismatch between cognitive stimulation intensity and actual needs. This solution improves the effective cognitive awakening rate through a dynamic frequency tuning mechanism. Existing thermal rendering methods rely on a single color channel, which lacks visual recognition in complex educational images. This solution uses HSV dual-channel encoding and brightness enhancement technology to significantly improve the visual salience score of hotspot areas. Existing formula prompt systems use fixed pop-up window positions and content, which often cause key knowledge points to be obscured. This solution reduces the visual interference rate to a minimum through dynamic adaptation of eye tracking and path planning.

[0132] Through the above technical solutions, this application has built a multi-dimensional cognitive visualization enhancement system. The dynamic animation engine realizes penetration-driven attention guidance, the thermal gradient mapping accurately depicts the spatiotemporal distribution of learning behavior, and the formula deduction missile window intelligently responds to cognitive transition needs. It can be deployed on an online experimental platform with a scale of 10,000 people, thereby improving the first-time learning and mastery rate of knowledge points, extending the average dwell time in key and difficult areas, and increasing the interactive click-through rate of formula derivation paths, verifying the significant advantages of the technical solution in improving learning efficiency and deep cognitive construction.

[0133] like Figure 3 As shown, this application further proposes that the method for intelligent recommendation and multi-scale matching of educational resource images based on machine learning also includes a feedback reinforcement mechanism:

[0134] The user's interaction behavior data before recommendation and the learning evaluation data after recommendation are used to generate the difference in the evaluation score change rate, and then input into the nonlinear constraint function to generate the score change constraint value;

[0135] Calculate the ratio of the usage time of the enhanced recommended image to the total learning time and generate a time ratio value;

[0136] Combine the performance change constraint value, duration ratio value and preset learning rate to generate the model parameter adjustment value;

[0137] The sliding window algorithm is used to calculate the moving average of historical parameter adjustments and update the weight coefficients in the penetration assessment model.

[0138] Among them, the weight coefficients include the weight of the cognitive stability parameter, the weight of the knowledge mastery parameter and the scaled variance penalty coefficient.

[0139] The feedback reinforcement mechanism package establishes a positive feedback correlation model between the rate of change of assessment scores and the effect of recommendations:

[0140]

[0141] Where η is the learning rate, T use T is the recommended image usage time. total is the total learning time; and the weight coefficients α, β, and γ in the penetration evaluation formula are updated through the sliding window mechanism.

[0142] The evaluation score change rate difference modeling refers to a nonlinear constraint function constructed based on the cognitive state transfer theory. The specific implementation method is: the evaluation score change rate difference before and after the recommendation (ΔS) is input into the hyperbolic tangent function (tanh) for normalization, and the nonlinear mapping relationship (S) is adjusted by the adaptive slope coefficient. c =tanh(k·ΔS)), where the slope coefficient k is dynamically adjusted by the difficulty level of the knowledge unit and the duration of the learning phase, to eliminate the interference of subject differences and time accumulation effects on the evaluation results. The recommendation effect refers to the time ratio analysis of the integrated temporal behavior characteristics, specifically using the window sliding integral algorithm to calculate the effective usage time ratio of the enhanced recommended image. Recommended usage time / total learning time) is combined with the forgetting curve model to perform exponential decay weighting on historical duration data to generate a time-weighted recommendation effectiveness indicator. The parameter dynamic optimization engine refers to an incremental learning framework based on a sliding window. Specifically, it extracts the spatiotemporal characteristics of historical parameter adjustments through a time series convolutional network (TCN). A moving average and fluctuation suppression algorithm is used to generate a robust weight update gradient, achieving coordinated tuning of the cognitive stability weight (α), knowledge mastery weight (β), and the scaled variance penalty coefficient (γ), satisfying the constraint of α+β+γ=1.

[0143] Specifically, the system collects assessment score streams and behavior log streams in real time. The score analysis pipeline performs a change rate difference calculation every 5 minutes, normalizes the original score data with the Z-score, and then inputs it into the dynamically adjusted hyperbolic tangent function to generate a constraint value between -1 and 1. The behavior analysis engine simultaneously collects statistics on the actual usage data of the recommended images, converts discrete operation events into continuous duration distribution curves through the time alignment algorithm, and calculates the timeliness weighted R t The parameter optimization module receives Sc With R t After the data flow, the parameter adjustment amount Δθ=η·(S c *R t ), where η is the learning rate (initial value 0.01, decay coefficient 0.95), which is dynamically adjusted based on the model's convergence state. The sliding window mechanism maintains historical data of Δθ for the last N cycles (default N = 30). After extracting cross-cycle dependency features through the TCN network, it outputs smoothed weight update instructions to the penetration assessment model, completing the online iteration of α, β, and γ.

[0144] Compared with existing technologies, traditional feedback mechanisms use fixed formulas to calculate parameter adjustments, resulting in high model divergence rates in interdisciplinary scenarios. This solution improves the stable convergence rate of the model through the collaborative design of dynamic hyperbolic tangent functions and time-effectiveness weighting mechanisms. Existing parameter update methods ignore the spatiotemporal correlation of historical adjustments, and the weight oscillation amplitude exceeds ±0.15. This solution suppresses weight fluctuations within the range of ±0.03 through TCN-driven sliding window optimization. The existing system learning rate setting relies on manual experience, and the response lag is more than 6 hours. This solution uses a convergence state perception algorithm to achieve minute-level dynamic adjustment of the learning rate, thereby improving model training efficiency. In addition, frequent manual intervention is required when dealing with multiple constraints (α+β+γ=1). This solution automatically meets the constraints through a projected gradient descent algorithm.

[0145] Through the above technical solutions, this application constructs an intelligent closed-loop optimization system driven by educational big data. Nonlinear constraint functions overcome the limitations of traditional linear normalization, accurately depicting the complex relationship between assessment changes and recommendation effectiveness. A temporal convolution sliding window captures the long-term dependencies of weight adjustments and suppresses interference from short-term data fluctuations. A dynamic learning rate mechanism achieves an optimal balance between model convergence speed and stability.

[0146] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0147] Specific implementation scenario: Recommending educational resources for the chapter "Electromagnetic Induction" in high school physics

[0148] A high school physics teacher uses an intelligent education platform to push teaching images for the "Electromagnetic Induction" chapter (including diagrams of Faraday's experimental setup, Lenz's law, and a diagram for deducing the direction of induced current). Based on students' interactions with the images, the teacher dynamically adjusts the information density and visual feedback of the recommended images to improve learning efficiency. The following is the specific implementation process:

[0149] Interaction behavior data collection:

[0150] When student A browses the "Electromagnetic Induction Experimental Device Diagram" on the platform, the system collects the following interactive data in real time through the browser-side event monitoring SDK:

[0151] Zoom level recording: Student A zooms in on the "coil winding direction" area three times (zoom level from 1 → 3 → 2), and records the timestamps of the zoom trajectory (10:05:10, 10:05:15, 10:05:20) and the level sequence [1, 3, 2].

[0152] Hotspot dwell time: 68 seconds in the "Lenz's law annotation area" (image coordinates (200,300)-(400,500)) (standard dwell time is 50 seconds), and 22 seconds in the "experimental circuit connection area" (standard dwell time is 30 seconds).

[0153] Mark the number of modifications: Modify the "Induction Current Direction Arrow" twice (first to the left, second to the right), and record the modification time (10:06:00, 10:06:10).

[0154] Permeability assessment:

[0155] The penetration evaluation module calculates the penetration evaluation value of student A based on the interaction data. The formula is:

[0156]

[0157] Among them, D i is the standard length of stay in the i-th knowledge unit, τ is the actual length of stay, which represents the quality of focus item, ΔS is the rate of change of the associated assessment score, which represents the learning gain item, Var(Z) is the scaled level variance, γ·Var(Z) represents the cognitive fluctuation item, and α, β, and γ are the preset weight coefficients.

[0158] Focus on quality items Take the standard stay time D of the "Lenz's law marked area" i = 50 seconds, the actual stay time τ = 68 seconds, α = 0.4, then focus on the quality item

[0159] Learning gain term (β·log(1+ΔS)): Student A’s previous assessment score was 70 points, and this assessment score was 85 points. The rate of change ΔS = (85-70) / 70 = 0.214, β = 0.4, then the learning gain term = 0.0780.4×log(1+0.214)≈0.4×0.194=0.078.

[0160] Cognitive fluctuation term (γ·Var(Z)): For the variance Var(Z) of the scaling level sequence [1, 3, 2] = 0.667 and γ = 0.2, then γ·Var(Z) = 0.133, 0.2×0.667 = 0.133.

[0161] Final P K = 0.294 + 0.078 - 0.133 = 0.239 (lower than the first preset threshold T1 = 0.3).

[0162] Multi-scale analysis:

[0163] The multi-scale analysis module synchronously processes the "Electromagnetic Induction Experiment Device Diagram":

[0164] Visual features: Extract global composition features (overall structure of the experimental device), local functional area features (coil, magnet, ammeter), and mechanism annotation features (arrow of current direction, text description of Lenz's law) through the pre-trained ResNet-50 network.

[0165] Knowledge association features: Based on the teaching syllabus knowledge graph ("Electromagnetic Induction" → "Faraday's Law" → "Lenz's Law"), generate an association matrix through a graph neural network, and the edge weights are dynamically adjusted according to the cross-knowledge point jump frequency of the student group (for example, the edge weight between "Electromagnetic Induction" and "Lenz's Law" is 0.8 because 80% of the students jump to "Lenz's Law" after learning "Electromagnetic Induction").

[0166] Cognitive guidance features: Grid the knowledge association features (8×8×256 matrix), perform matrix multiplication with the penetration vector (P K = 0.239), and after Softmax probability processing, generate an attention weight matrix, focusing on the "Lenz's Law Annotation Area" (weight 0.7).

[0167] Dynamic granularity selection:

[0168] Because P K = 0.239 < T1 = 0.3, the feature mask rule activates all features in the global composition area (the mask matrix is all 1). The dynamic granularity selection process calculates the direction consistency metric value (cosine similarity) between each level feature and the knowledge state vector (the current knowledge state of student A is the "basic concept of electromagnetic induction"), and after applying the penetration weighting coefficient ( δ = 0.5), selects the "global composition layer level" as the optimal recommended granularity (the highest score).

[0169] Enhanced recommended image generation and learning evaluation:

[0170] The image enhancement module generates enhanced recommended images based on the global composition layer level, with dynamic visualization elements: Because P K=0.239 (low permeability), a light blue flashing animation (frequency 0.5 Hz) is superimposed on the “Lenz’s law marking area”, and a dark purple gradient mark is displayed in the “Lenz’s law area” where the hotspot stays for a long time (the color depth is positively correlated with the stay time).

[0171] The interactive control displays a floating window in the upper right corner, showing the group's average mastery (for a class of 30 students, the average penetration is 0.35) and the individual difference index (Student A's interaction eigenvector has a Mahalanobis distance of 0.2 from the group mean, indicating that "the time spent in the Lenz's law zone is 36% longer than the class mean"). When a student clicks the control, the subgroup data is expanded.

[0172] After using the enhanced image, student A's assessment score increased to 85 points (change rate ΔS = 0.214). The recommended image usage time is 12 minutes (total learning time is 15 minutes, ratio R t =0.8).

[0173] The feedback reinforcement module calculates the parameter adjustment amount:

[0174]

[0175] Δθ=0.01×tanh(0.214)×0.8≈0.01×0.211×0.8=0.0017 The weight coefficient of the penetration model is updated through sliding window (N=30) moving average calculation (such as α is adjusted from 0.4 to 0.42 to strengthen the influence of the "focus on quality item").

[0176] Through the above process, Student A's learning efficiency in the "Electromagnetic Induction" chapter was significantly improved. The global composition display of the recommended images helped him establish a complete knowledge framework (traditional fixed-granularity recommendations are prone to fragmentation); dynamic animation and gradient marking guided him to focus on the weak points of "Lenz's Law" (experimental data showed that the residence time in this area was extended to 82 seconds, 30 seconds longer than traditional recommendations); group comparison of interactive controls prompted him to adjust his learning strategy (reduce invalid zooming in subsequent learning and increase stay in key areas); after feedback optimization, the system's recommendation accuracy for similar students increased from 75% to 92% (according to platform log statistics).

[0177] This implementation scenario completely covers the entire process of interactive data collection, permeability assessment, multi-scale analysis, dynamic granularity selection, enhanced image generation and feedback optimization, verifying the feasibility and effectiveness of the technical solution in actual teaching.

[0178] Example 2:

[0179] like Figure 4 and Figure 5As shown, this embodiment proposes a system for intelligent recommendation and multi-scale matching of educational resource images based on machine learning, which uses the above-mentioned method for intelligent recommendation and multi-scale matching of educational resource images based on machine learning, including:

[0180] The data collection module is used to obtain user interaction behavior data and subsequent evaluation data in real time, and encapsulate them into structured data packets through preset protocols;

[0181] A penetration evaluation module, used for calculating a penetration evaluation value of knowledge content according to the structured data package;

[0182] The multi-scale analysis module is used to simultaneously perform multi-scale analysis on educational resource images and generate a multi-level image feature set including visual features, knowledge association features, and cognitive guidance features;

[0183] The dynamic recommendation decision module generates the optimal recommended granularity level through dynamic granularity selection processing based on the permeability evaluation value and the multi-level image feature set;

[0184] An image enhancement module is used to generate an enhanced recommended image with permeability feedback tags according to the optimal recommendation granularity level;

[0185] an evaluation module, configured to generate an enhanced recommendation image with a permeability feedback mark according to the optimal recommendation granularity level, and obtain user learning evaluation data after the user interacts with the enhanced recommendation image;

[0186] The feedback reinforcement module is used to collect learning evaluation data after users obtain enhanced recommended images, and update the learning parameters of the penetration assessment model through the feedback reinforcement mechanism to form a closed-loop optimization system;

[0187] Among them, the data interfaces between the data acquisition module, permeability assessment module, multi-scale analysis module, dynamic recommendation decision module, image enhancement module and evaluation module use predefined standardized protocols for feature alignment.

[0188] The data acquisition module is a hybrid data processing architecture for complex educational scenarios. Its implementation involves: using a browser-side event monitoring SDK to capture image zoom trajectories (recording coordinate sequences, velocity distribution, and hotspots), and transmitting interaction logs in real time via the WebSocket protocol. An embedded eye tracking module acquires visual focus coordinates at a 120Hz sampling rate and combines data from touch device pressure sensors to construct a multidimensional behavioral feature vector. The assessment data access layer automatically parses standardized test system interfaces to extract knowledge point mastery distribution matrices. All data flows are processed by a timestamp alignment engine and then encapsulated into Apache Arrow-formatted data packets containing time series metadata, ensuring zero-copy transmission across modules.

[0189] The penetration evaluation model refers to a composite parameter model that integrates cognitive stability and knowledge mastery. It is implemented through an improved LSTM-GRU hybrid network: a bidirectional LSTM layer is constructed to capture the temporal fluctuation characteristics of scaling behavior and calculate the cognitive stability parameter (σ = sliding window variance); a gated recurrent unit (GRU) analyzes the coupling relationship between the hotspot residence time and the assessment results to generate the knowledge mastery parameter (μ = tanh (Σt i ·s i )); Finally, the permeability evaluation value is output through the dynamic weight fully connected layer (α+β+γ=1), and the weight coefficient is dynamically optimized by the reinforcement learning framework according to the teaching scenario.

[0190] Among them, the multi-scale feature parsing architecture includes three parallel processing channels: 1) The visual parsing channel uses the DenseNet-201 network to extract multi-level visual features (global composition → functional partitioning → mechanism labeling); 2) The knowledge association channel aggregates knowledge graph node embeddings through the graph attention network (GAT) to generate an interdisciplinary association matrix; 3) The cognitive guidance channel deploys a lightweight Transformer model to encode the permeability evaluation sequence into a 128-dimensional state vector, and realizes the dynamic mapping of cognitive state and knowledge features through a multi-head attention mechanism.

[0191] Among them, the dynamic recommendation decision module refers to an intelligent hierarchical selection system based on directional consistency measurement. The specific workflow is: calculate the cosine similarity between the feature vector of each level and the cognitive state vector, apply the permeability attenuation factor to suppress the low-correlation level; use the deep deterministic policy gradient (DDPG) algorithm to dynamically allocate hierarchical weights and generate a three-dimensional decision tensor containing the recommendation granularity priority.

[0192] Among them, the enhanced recommended image refers to a real-time rendering system for visualizing cognitive status, specifically integrating: 1) a thermal layer rendering engine, which uses WebGL shader technology to achieve penetration-driven transparency gradient overlay (α = 1-exp(-k·P)); 2) an SVG vector floating layer generator, which dynamically draws knowledge point association paths and formula derivation flowcharts; and 3) a Lottie animation interface, which inserts particle-guided animations in cognitively weak areas. Its playback frequency is negatively correlated with the penetration assessment value.

[0193] Among them, the feedback reinforcement module refers to the parameter dynamic update architecture based on federated learning. Specifically, the edge computing nodes collect desensitized behavior data, and the central server aggregates the global feature distribution, and then uses the momentum optimization algorithm to update the model parameters: design the hyperbolic tangent constraint function (Δθ=η·tanh(ΔS)·R t ) adjusts the learning rate and combines the sliding window mechanism to maintain historical gradient data to ensure that the weight update process meets the rigid constraint condition of α+β+γ=1.

[0194] Specific implementation process:

[0195] The browser SDK batches interaction events at a 50ms cycle and transmits raw data via a dedicated USB channel. The assessment system asynchronously pushes score data via a REST API, and all data is time-aligned to the millisecond level via Kafka message queues. The visual parsing pipeline uses a TensorRT-accelerated DenseNet network to achieve 60fps feature extraction. The knowledge graph service incrementally updates the association matrix every 5 minutes, and the cognitive state encoder generates real-time state vectors using a CUDA-accelerated Transformer model. The directional consistency metric calculates layer priorities every 200ms. The DDPG decision network combines historical decision performance data to generate noise-resistant recommendation strategies. The output decision tensor is serialized in Protobuf and pushed to the rendering engine. The WebGL engine dynamically blends base images and visualization layers based on layer weights. SVG overlays achieve sub-pixel rendering with CSS hardware acceleration. The particle animation system utilizes WebWorker multi-threading technology to ensure 60fps smoothness. Edge nodes upload desensitized feature distributions every 15 minutes. The central server aggregates the global model through horizontal federated learning, employing the Elastic Weight Consolidation (EWC) algorithm to prevent knowledge loss. Update packages are digitally signed and pushed to end devices.

[0196] Compared with traditional static recommendation systems (which rely on fixed user portraits and single feature matching), this system achieves accurate quantification of cognitive status through the collaboration of multiple modules. Traditional evaluation relies on single evaluation data. This system integrates three-dimensional features: scaled variance (cognitive stability), hotspot retention (knowledge mastery), and evaluation change (learning gain), and the accuracy of permeability evaluation is improved; this system realizes dynamic matching of multi-scale features. Traditional recommendations use a fixed granularity level. This system uses permeability-driven feature masks and directional consistency measurements to improve the matching degree between the information density of recommended images and cognitive needs; this system realizes multi-dimensional fusion of interactive feedback. Traditional feedback only displays evaluation results. This system improves students' cognitive accuracy of their own learning status through dynamic visualization (color / animation) and interactive controls (group / individual comparison); this system realizes the continuous evolution of recommendation strategies: the parameters of traditional systems are fixed, and this system uses a feedback reinforcement mechanism to achieve minute-level parameter tuning, which reduces the model prediction error rate.

[0197] Through the above technical solutions, this system has built an "intelligent ecosystem" for educational resource recommendations - from behavioral perception to cognitive assessment, from feature analysis to recommendation decisions, from interactive feedback to model evolution, each module is linked together, enabling the recommendation of educational resource images to be upgraded from "experience-driven" to a precise "data-driven + cognitive adaptation" model, significantly improving students' learning efficiency and depth of knowledge absorption.

[0198] The technical scope of the present invention is not limited to the contents of the above description. Those skilled in the art can make various deformations and modifications to the above embodiments without departing from the technical idea of the present invention, and these deformations and modifications should all fall within the protection scope of the present invention.

Claims

1. An intelligent recommendation and multi-scale matching method for educational resource images based on machine learning, characterized by: The following steps are involved: Obtaining interactive behavior data between users and educational resource images; Calculating a permeability evaluation value of a knowledge content-based permeability evaluation model according to the interactive behavior data; Simultaneously perform multi-scale analysis on educational resource images to generate a multi-level image feature set including visual features, knowledge association features, and cognitive guidance features; The permeability evaluation value and the multi-level image feature set are subjected to a dynamic granularity selection process to obtain the optimal recommended granularity level. The specific steps of the dynamic granularity selection process are as follows: generating a feature mask for each scale level of the multi-level image feature set, wherein the feature mask dynamically adjusts the effective area ratio according to the permeability evaluation value; Calculating the directional consistency measure of the feature mask at each level and the current knowledge state vector, and applying a permeability weighting coefficient; Selecting the optimal recommended granularity level according to a product result of the directional consistency metric value and the permeability weighted coefficient; An enhanced recommendation image with a permeability feedback mark is generated according to the optimal recommendation granularity level, and learning evaluation data of the user is obtained after the user interacts with the enhanced recommendation image.

2. The method for intelligent recommendation and multi-scale matching of educational resource images based on machine learning according to claim 1, characterized in that: The calculation formula of the permeability evaluation value is as follows: Among them, D i is the standard stay time of the i-th knowledge unit, τ is the actual stay time, represents the quality of focus item, ΔS is the rate of change of the associated assessment score, β·log(1+ΔS) represents the learning gain item, Var(Z) is the scaled level variance, γ·Var(Z) represents the cognitive fluctuation item, and α, β, and γ are preset weight coefficients.

3. The method for intelligent recommendation and multi-scale matching of educational resource images based on machine learning according to claim 1, characterized in that: The multi-scale analysis includes: Obtaining the visual features and basic image features of the educational resource image through a pre-trained visual feature extraction network; Extracting the knowledge association features of the educational resource image based on a graph neural network associated with a knowledge graph; Using an attention mechanism to fuse the permeability evaluation value, the basic image features, and the knowledge association features, and generate the cognitive guidance features; Among them, the edge weights associated with the knowledge graph are dynamically adjusted according to the preset knowledge point dependencies of the teaching syllabus.

4. According to the machine learning-based intelligent recommendation and multi-scale matching method for educational resource images of claim 3, the dynamic adjustment of the edge weights associated with the knowledge graph includes: Generate initial edge weights based on the pre-dependencies of the knowledge points in the syllabus; Based on the knowledge path selection data of the student group, the edge weights are dynamically updated through the reinforcement learning algorithm; The knowledge path selection data is calculated by counting the frequency of students jumping across knowledge points and the ratio of their stay time.

5. The method for intelligent recommendation and multi-scale matching of educational resource images based on machine learning according to claim 3, characterized in that: The generation of the cognitive guidance feature includes: Gridding the knowledge association features to generate a knowledge association feature matrix; Performing a matrix product operation on the permeability evaluation value vector of the permeability evaluation value and the knowledge association feature matrix, and normalizing the product result to obtain an attention weight matrix; Probabilistically processing the attention weight matrix through the Softmax function; The dimension of the knowledge association feature matrix and the dimension of the permeability evaluation value vector are matched by a preset dimension alignment rule.

6. The method for intelligent recommendation and multi-scale matching of educational resource images based on machine learning according to claim 1, characterized in that: The generation rules of the feature mask include: When the permeability evaluation value is lower than a first preset threshold, activating all features of the global composition area of the image; When the permeability evaluation value is between the first preset threshold and the second preset threshold, retaining the edge contour features of the functional partition; When the permeability evaluation value is higher than the second preset threshold, only the mechanism annotation detail feature area is activated; Among them, the first preset threshold and the second preset threshold are critical values preset according to the teaching scenario type.

7. The method for intelligent recommendation and multi-scale matching of educational resource images based on machine learning according to claim 1, characterized in that: The generation of the enhanced recommendation image includes: superimposing a dynamic visualization element on the basic image feature, wherein the display intensity of the dynamic visualization element is positively correlated with the permeability evaluation value; Embed interactive penetration prompt controls to display the group's average mastery and individual difference index in real time; The group average mastery is calculated based on the mean of the penetration evaluation values of all students, and the individual difference index is calculated based on the deviation between the current student's historical interaction behavior data and the group data.

8. The method for intelligent recommendation and multi-scale matching of educational resource images based on machine learning according to claim 7, characterized in that: The dynamic visualization elements include: Displaying a key knowledge point animation in an area where the permeability evaluation value is lower than a preset threshold, wherein the playing frequency of the key knowledge point animation is negatively correlated with the permeability evaluation value; Generate a color depth gradient mark based on the hotspot residence time distribution, where the color depth is proportional to the hotspot residence time; A correlation formula derivation path pops up in the permeability evaluation value mutation area, and the mutation area is determined by detecting the second-order derivative change of the permeability evaluation value through a sliding window algorithm.

9. The method for intelligent recommendation and multi-scale matching of educational resource images based on machine learning according to claim 1, characterized in that: The method for intelligent recommendation and multi-scale matching of educational resource images based on machine learning also includes a feedback reinforcement mechanism: The interaction behavior data before the recommendation and the learning evaluation data after the recommendation are used to generate the evaluation score change rate difference, and the difference is input into a nonlinear constraint function to generate a score change constraint value; Counting the proportion of the usage time of the enhanced recommended image to the total learning time, and generating a time ratio value; The performance change constraint value, the duration ratio value and the preset learning rate are combined to generate a model parameter adjustment amount; Performing a moving average calculation on the historical parameter adjustment amount by a sliding window algorithm to update the weight coefficient in the permeability assessment model; The weight coefficient includes the weight of the cognitive stability parameter, the weight of the knowledge mastery parameter, and a scaled variance penalty coefficient.

10. An intelligent recommendation and multi-scale matching system for educational resource images based on machine learning, characterized by: The method for intelligent recommendation and multi-scale matching of educational resource images based on machine learning according to any one of claims 1 to 9 is used, comprising: The data collection module is used to obtain user interaction behavior data and subsequent evaluation data in real time, and encapsulate them into structured data packets through preset protocols; a permeability evaluation module, configured to calculate a permeability evaluation value of the knowledge content according to the structured data packet; Multi-scale analysis module, used to simultaneously perform multi-scale analysis on educational resource images and generate a multi-level image feature set including visual features, knowledge association features, and cognitive guidance features; The dynamic recommendation decision module generates the optimal recommended granularity level through dynamic granularity selection processing based on the permeability evaluation value and the multi-level image feature set; An image enhancement module, configured to generate an enhanced recommended image with a permeability feedback mark according to the optimal recommendation granularity level; an evaluation module, configured to generate an enhanced recommendation image with a permeability feedback mark according to the optimal recommendation granularity level, and obtain learning evaluation data of the user after the user interacts with the enhanced recommendation image; A feedback reinforcement module is used to collect learning evaluation data of the user after obtaining the enhanced recommended image recommendation, and to update the learning parameters of the permeability assessment model through a feedback reinforcement mechanism to form a closed-loop optimization system; Among them, the data interfaces among the data acquisition module, the permeability assessment module, the multi-scale analysis module, the dynamic recommendation decision module, the image enhancement module and the evaluation module adopt a predefined standardized protocol for feature alignment.

Images