Learning state evaluation method, system and device based on big data analysis

Abstract

This application provides a learning status assessment method, system, and device based on big data analytics, relating to the field of big data analytics. The method includes: collecting multi-source learning behavior data of students to construct individual knowledge graphs; collecting students' resting physiological signals to construct a physiological signal database; using the individual knowledge graphs as federated nodes and combining them with the physiological signal database to construct a group knowledge graph federation; constructing a teaching progress timeline based on knowledge points and calculating the progress offset between the student's current learning progress and the class's teaching progress; comprehensively evaluating the individual knowledge graphs, the group knowledge graph federation, and the progress offset to obtain a multi-dimensional comprehensive evaluation report; and using a multi-objective optimization algorithm to generate personalized learning plans and push them to students. This solves the technical problems of one-sided, subjective, and inaccurate learning status assessment results caused by a single evaluation dimension and neglect of the dynamic influence of class teaching progress.

Description

Technical Field

[0001] This application relates to the field of big data analytics, and in particular to methods, systems and devices for evaluating learning status based on big data analytics. Background Technology

[0002] In the education informatization industry, big data-based learning analytics technology has been widely used. Existing technologies typically collect students' surface-level learning behavior data, such as video viewing time, homework completion rate, and exercise accuracy, through online learning platforms or school internal teaching management systems. This data is stored in the form of structured logs and processed using statistical analysis methods such as knowledge tracking models and performance regression models, or simple machine learning models. Ultimately, the data is used to generate learning reports, performance alerts, or push unified review materials.

[0003] However, in practical applications, student A might have a low accuracy rate but demonstrate novel problem-solving approaches and strong transfer of thinking skills, while student B, despite achieving high scores through repeated practice, might not have truly grasped the essence of the knowledge. If the system judges solely based on accuracy, it might misclassify student A as a poor student and classify student B as an excellent student, which clearly contradicts the educational goal of individualized instruction. Furthermore, different teachers have different teaching plans, leading to varying learning paces among students. Using standardized, milestone-based learning check-ins to assess progress and completion could cause some students to engage in superficial or accelerated learning to meet the evaluation criteria, thus reducing overall learning absorption. Finally, due to differences in personality and psychological resilience, students often perform differently during exams. For example, student C might experience sleep deprivation due to the previous night's tension, resulting in physiological fatigue (abnormal ECG and EEG signals) during data collection or the exam, affecting their performance. However, the system might attribute this physiological fatigue to a poor learning attitude or a lack of mastery of the knowledge point. In summary, existing assessment methods and approaches, when used, fail to capture students' deep cognitive states due to their singular and superficial assessment dimensions, neglect of the dynamic impact of class teaching progress, and lack of integration of students' physiological state information, resulting in inaccurate assessment of learning status. Summary of the Invention

[0004] This application provides a learning status assessment method, system, and device based on big data analysis, which solves the technical problems of existing technologies, such as one-sidedness, subjectivity, and inaccuracy in learning status assessment results due to the single assessment dimension and neglect of the dynamic impact of class teaching progress.

[0005] To achieve the above objectives, this application adopts the following technical solution: Firstly, the learning status assessment method based on big data analysis includes: collecting multi-source learning behavior data of students to construct individual knowledge graphs, which include knowledge element layers, skill layers, and literacy layers; collecting students' resting physiological signals to construct a physiological signal database, which includes at least electrocardiogram (ECG) and electroencephalogram (EEG) signals; using individual knowledge graphs as federated nodes and combining them with the physiological signal database to construct a group knowledge graph federation; collecting actual teaching progress data for each class, constructing a teaching progress timeline based on knowledge points, and calculating the progress offset between the student's current learning progress and the class's teaching progress; comprehensively evaluating the personalized knowledge graphs, the group knowledge graph federation, and the progress offset to obtain a multi-dimensional comprehensive evaluation report; and using a multi-objective optimization algorithm to generate personalized learning plans based on the multi-dimensional comprehensive evaluation report and pushing them to students.

[0006] In conjunction with the first aspect mentioned above, in one possible implementation, the process of constructing an individual knowledge graph specifically includes: Multi-source learning behavior data of students is collected and feature extracted to generate a multi-dimensional feature space, including features of knowledge mastery, learning habits, knowledge application ability, thinking transfer, and learning style. A knowledge meta-layer is constructed using knowledge mastery features as nodes, consisting of nodes representing knowledge points and edges representing semantic relationships between knowledge points. The knowledge meta-layer is input into a first graph neural network, and the nodes are aggregated using knowledge application ability and thinking transfer features as constraints to obtain a skill layer, consisting of nodes representing core skills. The skill layer is then input into a second graph neural network, and the nodes are aggregated using learning habits and learning style features as constraints to obtain a competency layer, consisting of nodes representing comprehensive competencies. Hierarchical connections between the knowledge meta-layer, skill layer, and competency layer are established to generate a complete individual knowledge graph.

[0007] In conjunction with the first aspect mentioned above, one possible implementation involves constructing a group knowledge graph federation, specifically including: extracting students' knowledge element layer subgraphs, skill layer subgraphs, and resting physiological feature vectors from individual knowledge graphs and physiological signal databases, and inputting them into a cross-modal joint embedding model to obtain cross-modal joint embedding vectors. These vectors represent the intrinsic correlation patterns between students at the knowledge element layer, skill layer, and physiological state. Using each student as a federation node, the cross-modal joint embedding vectors are aggregated to obtain a group federation center vector. This center vector represents the common correlation patterns among students of the same grade in the region at the knowledge element layer, skill layer, and physiological state. Based on the group federation center vector and the cross-modal joint embedding vectors, horizontal evaluation indicators for students within the group are calculated. Historical cross-modal joint embedding vectors of students are collected to form a temporal embedding vector sequence. This temporal embedding vector sequence is compared and analyzed with the group federation center vector to calculate longitudinal development trajectory indicators. Finally, the horizontal evaluation indicators and longitudinal development trajectory indicators are integrated to generate the group knowledge graph federation.

[0008] In conjunction with the first aspect mentioned above, in one possible implementation, the cross-modal joint embedding model includes a graph encoding layer, a physiological encoding layer, a bilinear attention layer, and a joint embedding layer: The graph encoding layer comprises a first graph convolutional network and a second graph convolutional network. The first graph convolutional network takes the knowledge element layer subgraph as input to obtain the knowledge element layer graph embedding vector, and the second graph convolutional network takes the skill layer subgraph as input to obtain the skill layer graph embedding vector. The physiological encoding layer takes the resting physiological feature vector as input, performs a nonlinear transformation through a multilayer perceptron, and outputs a physiological feature embedding vector. The bilinear attention layer calculates the first attention weight between the knowledge element layer graph embedding vector and the physiological feature embedding vector, and the second attention weight between the skill layer graph embedding vector and the physiological feature embedding vector, respectively. The joint embedding layer performs weighted fusion of the knowledge element layer graph embedding vector and the physiological feature embedding vector based on the first attention weight, and performs weighted fusion of the skill layer graph embedding vector and the physiological feature embedding vector based on the second attention weight. The results of the two weighted fusions are concatenated and mapped through a fully connected layer to output the cross-modal joint embedding vector.

[0009] In conjunction with the first aspect mentioned above, one possible implementation involves calculating the progress offset between a student's current learning progress and the class's teaching progress. Specifically, this includes: obtaining the actual teaching progress sequence of the target class, arranged chronologically at the knowledge point level; extracting learning progress trajectories from individual knowledge graphs; inputting the learning progress trajectories and the actual teaching progress sequence into a progress decoupling comparison network, which includes a temporal coding layer and a dynamic alignment layer; the temporal coding layer extracting temporal features from both the actual teaching progress sequence and the learning progress trajectory to obtain a class teaching progress temporal vector and a student individual progress temporal vector; and the dynamic alignment layer employing a differential dynamic time warping algorithm to calculate the minimum cumulative distance between the class teaching progress temporal vector and the student individual progress temporal vector, using the warped path corresponding to the minimum cumulative distance as the progress offset representation path; and, based on the progress offset representation path, analyzing the target student's progress lead, progress lag, and knowledge point mastery drift relative to the class teaching progress at different time slices to generate the progress offset.

[0010] In conjunction with the first aspect mentioned above, one possible implementation involves comprehensively evaluating personalized knowledge graphs, federated group knowledge graphs, and progress offsets to obtain a multi-dimensional comprehensive evaluation report. Specifically, this includes: aligning the knowledge element layer nodes and skill layer nodes in the individual knowledge graph with the common association patterns in the federated group knowledge graph based on a temporal sequence, generating joint individual and group knowledge data. This joint knowledge data is then input into a cross-attention evaluation network, which includes a horizontal comparison module and a vertical tracking module. The horizontal comparison module calculates the horizontal deviation of each node in the individual knowledge graph, using the common association patterns as a reference. The vertical tracking module performs a temporal comparison between the cross-modal joint embedding vector and the group federation center vector, and calculates the vertical development trajectory deviation by combining the progress offset representation path. Finally, the horizontal deviation and the vertical development trajectory deviation are jointly decoded, and a multi-layer perceptron feature fusion and nonlinear mapping are used to generate the multi-dimensional comprehensive evaluation report.

[0011] In conjunction with the first aspect mentioned above, one possible implementation involves generating personalized learning plans for students based on a multi-dimensional comprehensive evaluation report using a multi-objective optimization algorithm and then pushing these plans to the students. Specifically, this includes: extracting subject ability indicators, learning effort indicators, learning progress indicators, and learning efficiency indicators from the multi-dimensional comprehensive evaluation report; constructing decision variables and constraints for the multi-objective optimization function; inputting these indicators into a pre-defined multi-objective optimization model; and using a non-dominated sorting genetic algorithm to solve for Pareto optimal solutions by maximizing subject balance, effort matching, progress synchronization, and efficiency. The Pareto optimal solution set contains multiple alternative daily subject learning time allocation schemes. Based on students' learning habit characteristics and historical behavioral data, the optimal learning plan is selected from the Pareto optimal solution set and pushed to the student's terminal.

[0012] Secondly, a learning status assessment system based on big data analysis is provided, including: an individual knowledge graph construction module, used to collect students' multi-source learning behavior data and construct an individual knowledge graph, which includes a knowledge element layer, a skill layer, and a literacy layer; a physiological signal database construction module, used to collect students' resting physiological signals and construct a physiological signal database, which includes at least electrocardiogram (ECG) and electroencephalogram (EEG) signals; a group knowledge graph federation construction module, used to construct a group knowledge graph federation using individual knowledge graphs as federation nodes and combining them with the physiological signal database; an offset calculation module, used to collect actual teaching progress data for each class, construct a teaching progress timeline based on knowledge points, and calculate the progress offset between the student's current learning progress and the class's teaching progress; a comprehensive evaluation module, used to comprehensively evaluate the personalized knowledge graph, the group knowledge graph federation, and the progress offset, and obtain a multi-dimensional comprehensive evaluation report; and a learning plan generation module, used to generate personalized learning plans based on the multi-dimensional comprehensive evaluation report using a multi-objective optimization algorithm and push them to students.

[0013] In conjunction with the second aspect mentioned above, in one possible implementation, the group knowledge graph federation construction module is specifically used to: extract students' knowledge element layer subgraphs, skill layer subgraphs, and resting physiological feature vectors from individual knowledge graphs and physiological signal databases, and input them into a cross-modal joint embedding model to obtain cross-modal joint embedding vectors. These vectors represent the intrinsic correlation patterns between students at the knowledge element layer, skill layer, and physiological state. Using each student as a federation node, the cross-modal joint embedding vectors are federated to obtain the group federation center vector. Based on the group federation center vector and the cross-modal joint embedding vectors, the horizontal evaluation index of students within the group is calculated. Historical cross-modal joint embedding vectors of students are collected to form a temporal embedding vector sequence. This temporal embedding vector sequence is compared and analyzed with the group federation center vector to calculate the vertical development trajectory index. The horizontal evaluation index and the vertical development trajectory index are then fused to generate the group knowledge graph federation.

[0014] Thirdly, a learning status assessment device based on big data analysis is provided, including a communication unit and a processing unit.

[0015] The communication unit is used to collect students' multi-source learning behavior data, resting physiological signals, and actual teaching progress data for each class.

[0016] The processing unit is used to construct individual knowledge graphs, which include knowledge element layers, skill layers, and literacy layers. A physiological signal database is constructed, including at least electrocardiogram (ECG) and electroencephalogram (EEG) signals at rest. Using the individual knowledge graphs as federated nodes, and combining them with the physiological signal database, a group knowledge graph federation is constructed. A teaching progress timeline is built based on knowledge points, calculating the progress offset between the student's current learning progress and the class's teaching progress. A comprehensive evaluation is performed on the personalized knowledge graphs, the group knowledge graph federation, and the progress offset, resulting in a multi-dimensional comprehensive evaluation report. Based on the multi-dimensional comprehensive evaluation report, a multi-objective optimization algorithm is used to generate personalized learning plans and push them to students.

[0017] This application provides a learning status assessment method, system, and device based on big data analysis. It can construct an individual knowledge graph containing knowledge element layers, skill layers, and literacy layers, overcoming the limitations of existing technologies that only collect surface-level learning behavior data. This accurately captures students' knowledge mastery, knowledge application ability, and deep cognitive state, solving the problem of single-dimensional and superficial assessment and providing comprehensive cognitive support. Furthermore, by fusing physiological signals and knowledge graphs through a cross-modal joint embedding model and constructing a federated group knowledge graph, it can capture students' physiological states such as fatigue and attention, providing an objective group reference system for individual assessment. This eliminates interference from physiological states, effectively distinguishes between ability problems and state problems, and avoids misjudging abnormal performance caused by poor physiological state as insufficient knowledge mastery, thus improving the accuracy and objectivity of the assessment. Finally, by calculating the progress offset through a progress decoupling comparison network, it achieves dynamic alignment between individual learning progress and class teaching progress, solving the problem of assessment distortion caused by ignoring differences in teaching progress. This allows for accurate judgment of the rationality of learning progress, avoiding a one-size-fits-all assessment, thereby changing the current one-sided and subjective assessment status and achieving a comprehensive, objective, and accurate assessment of students' learning status. Attached Figure Description

[0018] Figure 1 A flowchart illustrating the learning status evaluation method based on big data analysis provided in this application embodiment; Figure 2 A flowchart illustrating the individual knowledge graph construction steps in the learning state evaluation method based on big data analysis provided in this application embodiment; Figure 3 A flowchart illustrating the federated construction steps of a group knowledge graph in the learning state evaluation method based on big data analysis provided in this application embodiment; Figure 4 A flowchart illustrating the steps for calculating the progress offset between a student's current learning progress and the class's teaching progress in the learning status assessment method based on big data analysis provided in this application embodiment. Figure 5 The flowchart of the learning status evaluation method based on big data analysis provided in the embodiments of this application is as follows: a comprehensive evaluation of personalized knowledge graph, group knowledge graph federation and progress offset is performed to obtain a multi-dimensional comprehensive evaluation report. Figure 6 A system framework diagram of a learning status evaluation system based on big data analysis provided for embodiments of this application; Figure 7 A schematic diagram of the structure of the learning status evaluation device based on big data analysis provided in the embodiments of this application; Figure 8 This is a schematic diagram of the hardware structure of the learning status assessment device based on big data analysis provided in the embodiments of this application. Detailed Implementation

[0019] To address the issues of subjective and inaccurate assessment results in existing technologies, this application proposes a learning status assessment method based on big data analysis. This method constructs a multi-layered individual knowledge graph, a federated group knowledge graph integrating physiological signals, and a dynamic progress offset to conduct a multi-dimensional comprehensive assessment of students' learning status. Ultimately, it generates personalized learning plans, achieving objectivity, accuracy, and personalization of the assessment results.

[0020] To achieve the above objectives, such as Figure 1 As shown in the embodiments of this application, the learning status evaluation method based on big data analysis includes: Step 101: Collect students' multi-source learning behavior data and construct individual knowledge graphs, which include knowledge element layers, skill layers, and literacy layers.

[0021] Multi-source learning behavior data refers to various process data generated by students in digital learning environments. These data sources include, but are not limited to, log data from online learning platforms (such as video viewing time and page dwell time), answer records from homework and exam systems, and question and discussion data from teacher-student interaction systems. The knowledge element layer is the most basic level constituting an individual's knowledge graph. It consists of nodes representing isolated knowledge points and edges between nodes representing semantic relationships such as prerequisites and inclusion, used to accurately depict students' mastery of specific knowledge points. The skill layer is a higher-order cognitive ability layer built upon the knowledge element layer. Its nodes are formed by aggregating and abstracting multiple related knowledge elements, representing students' core abilities to apply knowledge to solve complex problems and achieve cognitive transfer. The literacy layer is the highest level of the graph. Its nodes are generated by a higher-order fusion and abstraction of multiple skill layer nodes, used to represent students' internalized, stable, and comprehensive learning qualities, such as logical reasoning literacy or self-directed learning literacy.

[0022] In some implementations, process data generated by students from multiple channels such as online learning platforms and homework systems is collected, and feature extraction algorithms are used to extract features from these data. These features are categorized into five dimensions: knowledge mastery, learning habits, knowledge application ability, thinking transfer, and learning style, which together form a multi-dimensional feature space.

[0023] Then, using knowledge mastery characteristics as node attributes and semantic relationships between knowledge points as edges, a knowledge element layer that forms the basis of an individual knowledge graph is constructed. The knowledge element layer is then input into the first graph neural network, which uses knowledge application capability features and thinking transfer features as loss functions for aggregation constraints. This guides the network to cluster and fuse closely related nodes in the knowledge element layer, thereby generating a skill layer composed of core skill nodes.

[0024] The generated skill layer can then be input into the second graph neural network. This network, constrained by learning habit features and learning style features, further abstracts and aggregates the skill nodes, ultimately forming a competency layer that represents comprehensive competency.

[0025] Finally, hierarchical connections are automatically established between the knowledge layer, skill layer, and literacy layer, connecting the three levels of nodes into a complete individual knowledge graph that can hierarchically reflect the depth of students' cognition.

[0026] For example, analyzing student A's math learning data revealed that their mastery of both "linear functions" and "quadratic equations" was 0.9, and their problem-solving records showed a strong tendency to combine numerical and graphical methods, indicating a high level of knowledge application ability. Therefore, when constructing an individual knowledge graph, these two knowledge point nodes (knowledge element layers) can be aggregated by the first graph neural network and combined with their high knowledge application ability to form the skill node "Application of Function Thinking" (skill layer). Subsequently, this skill node is combined with student A's demonstrated ability to "summarize question types" (learning habit characteristic) and "prefer graphical understanding" (learning style characteristic), and through aggregation by the second graph neural network, ultimately forms their "Mathematical Modeling Literacy" literacy node (literacy layer). This clearly demonstrates the complete path of student A from mastering specific function knowledge points to developing function application skills, and finally internalizing them into mathematical modeling literacy.

[0027] Step 102: Collect students' resting physiological signals and construct a physiological signal database. The resting physiological signals include at least electrocardiogram (ECG) signals and electroencephalogram (EEG) signals.

[0028] Among them, resting physiological signals refer to physiological data collected when students are in a quiet, relaxed state without specific learning tasks, which can reflect their basic physiological functions and nervous system state.

[0029] In some implementations, the raw ECG and EEG signals are received in real time and transmitted to the physiological signal library construction module in the processing unit. At this time, the module integrates a signal preprocessing submodule to perform noise reduction, filtering and baseline drift correction on the raw signals. For example, wavelet transform is used to remove electromyographic interference in the ECG signal and independent component analysis algorithm is used to remove electrooculography artifacts in the EEG signal.

[0030] This allows the preprocessed signal to enter the feature extraction submodule, which extracts heart rate variability indicators (including time-domain indicators SDNN and RMSSD and frequency-domain indicator LF / HF ratio) from the electrocardiogram signal, and extracts power spectral density of each frequency band and calculates features such as θ / β ratio from the electroencephalogram signal.

[0031] Finally, these extracted physiological feature vectors are structured and stored according to metadata such as student ID and collection timestamp, and an association index is established with the student ID output by the individual knowledge graph construction module to form a complete physiological signal library, providing the necessary physiological dimension support for subsequent cross-modal fusion.

[0032] It should be noted that the resting physiological feature vectors stored in this physiological signal database will be input into the cross-modal joint embedding model together with the knowledge element subgraph and skill subgraph in the individual knowledge graph, and the deep fusion of behavioral cognitive data and physiological state data will be achieved through the bilinear attention mechanism.

[0033] For example, during a data acquisition process, student Xiaoming's smart bracelet transmits 5 minutes of resting electrocardiogram (ECG) signals to the system via Bluetooth, while his EEG headband also simultaneously transmits resting EEG signals to the system. At this point, the physiological signal database construction module first performs wavelet denoising on the raw ECG signals to remove motion artifacts, then extracts the heart rate variability index, calculating SDNN as 45ms and RMSSD as 32ms. Simultaneously, after performing independent component analysis to remove blink artifacts from the EEG signal, the alpha wave power spectral density was extracted to be 12 μV² / Hz and the beta wave power spectral density to be 8 μV² / Hz, with the θ / β ratio calculated to be 1.2. These physiological feature vectors were structured and stored by the system according to the student ID "2024001" and the collection time "2024-05-20 14:30", and an association index was established with the student's individual knowledge graph.

[0034] These physiological feature vectors can then be retrieved during the subsequent construction of the group knowledge graph federation, and compared with Xiaoming's knowledge element subgraph and skill layer subgraph. Figure 1 The same input cross-modal joint embedding model is used to generate cross-modal joint embedding vectors that can comprehensively represent the intrinsic relationship between cognitive and physiological states.

[0035] Step 103: Using individual knowledge graphs as federated nodes and combining them with a physiological signal database, construct a group knowledge graph federation.

[0036] In some implementations, knowledge element layer subgraphs and skill layer subgraphs for each student are extracted from the individual knowledge graph, and corresponding resting physiological feature vectors are extracted from the physiological signal database. Then, these three types of heterogeneous data are input into a pre-trained cross-modal joint embedding model (including graph coding layer, physiological coding layer, bilinear attention layer and joint embedding layer). The graph coding layer maps the knowledge layer subgraphs to knowledge layer graph embedding vectors through a first graph convolutional network, and then maps the skill layer subgraphs to skill layer graph embedding vectors through a second graph convolutional network. Finally, the physiological coding layer maps the resting physiological feature vectors to physiological feature embedding vectors through a multilayer perceptron. The bilinear attention layer calculates the first attention weight between the knowledge layer graph embedding vectors and the physiological feature embedding vectors, and the second attention weight between the skill layer graph embedding vectors and the physiological feature embedding vectors. The joint embedding layer performs weighted fusion of the knowledge layer graph embedding vectors and the physiological feature embedding vectors based on the first attention weight, and weighted fusion of the skill layer graph embedding vectors and the physiological feature embedding vectors based on the second attention weight. The results of the two weighted fusions are concatenated and mapped through a fully connected layer to finally output the cross-modal joint embedding vector for each student.

[0037] At this point, each student can be treated as a federated node, and a federated learning architecture can be used to securely aggregate the cross-modal joint embedding vectors of each node. After each node calculates the embedding vector locally, it only uploads the encrypted model parameters to the central server. The central server calculates the group federated center vector representing the common pattern of the group through a secure aggregation algorithm.

[0038] Based on the group's federated center vector and each student's cross-modal joint embedding vector, cross-sectional evaluation metrics for each student within the group are obtained by calculating measures such as cosine similarity or Euclidean distance. Simultaneously, historical cross-modal joint embedding vectors for each student are collected from a historical database to form a temporal embedding vector sequence. This sequence is then compared and analyzed with the group's federated center vector to calculate each student's longitudinal development trajectory index relative to the group's common trajectory. Finally, the cross-sectional evaluation metrics and longitudinal development trajectory indexes are fused to generate a complete group knowledge graph federation, providing a group reference system for subsequent multi-dimensional comprehensive evaluation.

[0039] Step 104: Collect actual teaching progress data for each class, construct a teaching progress timeline based on knowledge points, and calculate the progress offset between the student's current learning progress and the class's teaching progress.

[0040] In some implementations, the teaching plan data of the school's teaching management system is periodically synchronized through the communication unit, or the progress information uploaded by the teacher's terminal through the mobile communication network is received to collect the actual teaching progress data of each class. Then, a teaching progress timeline is constructed with knowledge points as the unit and the teaching time sequence is followed, and the actual teaching progress sequence arranged in time sequence is extracted from the timeline.

[0041] Simultaneously, from the individual knowledge graph, the knowledge points that students have completed learning and reached a set mastery threshold (e.g., 0.8) and their corresponding timestamp information are extracted to form the student's learning progress trajectory. The actual teaching progress sequence of the class and the student's learning progress trajectory can then be input into the progress decoupling comparison network. At this point, the network first extracts the temporal features of the two input sequences through its internal temporal coding layer. The temporal coding layer adopts a long short-term memory network or a Transformer architecture, which can capture the sequential dependencies and teaching rhythm features of each knowledge point in the sequence, and output the class teaching progress temporal vector and the student individual progress temporal vector.

[0042] Then, the two time-series vectors are input into the dynamic alignment layer of the progress decoupling comparison network. A differential dynamic time warping algorithm is used to calculate the minimum cumulative distance between the two time-series vectors through dynamic programming. That is, the optimal alignment path is found under the condition that the time axis is allowed to stretch elastically, and the warping path corresponding to the minimum cumulative distance is used as the progress offset representation path.

[0043] Finally, the progress offset representation path is analyzed. For each time slice in the path, the time difference between the class progress and the student progress is calculated to obtain the progress advance or progress lag. At the same time, the knowledge points that are skipped or repeatedly aligned during the alignment process are analyzed and their mastery drift is calculated. Finally, these three sub-dimensions are integrated to generate a comprehensive progress offset.

[0044] Step 105: Conduct a comprehensive evaluation of the personalized knowledge graph, the federated group knowledge graph, and the progress offset to obtain a multi-dimensional comprehensive evaluation report.

[0045] In some implementations, by receiving personalized knowledge graphs, federated group knowledge graphs, and progress offsets, the knowledge element layer nodes and skill layer nodes in the personalized knowledge graphs can be granularly aligned with the common association patterns of the group in the federated group knowledge graphs based on the time sequence. This ensures that individual data and group data are comparable at the same timestamp and knowledge point granularity, thereby generating joint knowledge data for individuals and groups.

[0046] The combined knowledge data of this individual group is then input into a pre-trained cross-attention evaluation network (containing two parallel processing channels: a horizontal comparison module and a vertical tracking module). In the horizontal comparison module, using the common association patterns of the group in the group knowledge graph federation as a reference benchmark, the network calculates the Euclidean distance or cosine similarity between each knowledge element layer node and skill layer node in the individual knowledge graph and its corresponding group node. This yields the horizontal deviation of each node, reflecting the student's advantages or disadvantages relative to their peers in terms of mastering specific knowledge points and forming core skills.

[0047] In the longitudinal tracking module, the student's historical cross-modal joint embedding vector sequence is retrieved and compared with the temporal evolution trajectory of the group's federated center vector. At the same time, the progress offset representation path in the progress offset is introduced as a constraint condition for temporal alignment. By calculating the dynamic time warping distance between the two temporal sequences, the longitudinal development trajectory deviation is obtained to reflect the degree of fit between the student's personal growth trajectory and the common development law of the group.

[0048] Finally, the horizontal deviation output by the horizontal comparison module and the vertical development trajectory deviation output by the vertical tracking module are input into the joint decoding layer. Multilayer perceptron is used for feature fusion and nonlinear mapping to transform the multi-dimensional deviation indicators into interpretable evaluation dimensions, generating a multi-dimensional comprehensive evaluation report that includes subject ability indicators, learning effort indicators, learning progress indicators, and learning efficiency indicators.

[0049] Step 106: Based on the multi-dimensional comprehensive evaluation report, a multi-objective optimization algorithm is used to generate a personalized learning plan and push it to the student.

[0050] The process of generating personalized learning plans and pushing them to students based on a multi-dimensional comprehensive evaluation report using a multi-objective optimization algorithm includes: extracting subject ability indicators, learning effort indicators, learning progress indicators, and learning efficiency indicators from the multi-dimensional comprehensive evaluation report to construct decision variables and constraints for the multi-objective optimization function; inputting the subject ability indicators, learning effort indicators, learning progress indicators, and learning efficiency indicators into a pre-set multi-objective optimization model (a mathematical programming model constructed based on the multi-dimensional comprehensive evaluation report, using the subject ability indicators, learning effort indicators, learning progress indicators, and learning efficiency indicators as inputs for decision variables and constraints); using a non-dominated sorting genetic algorithm (a multi-objective optimization algorithm based on population evolution, which searches for a Pareto optimal solution set in the solution space through non-dominated sorting and crowding distance calculation); and solving for the Pareto optimal solution set with the four objective functions of maximizing subject balance, effort matching, progress synchronization, and efficiency maximization as optimization objectives. The Pareto optimal solution set contains multiple alternative daily subject learning time allocation schemes; and selecting the optimal learning plan from the Pareto optimal solution set based on students' learning habit characteristics and historical behavior data, and pushing the optimal learning plan to the student's terminal.

[0051] In some implementation methods, key indicators extracted from multi-dimensional comprehensive evaluation reports include subject ability indicators reflecting the degree of mastery of each subject, learning energy indicators reflecting students' physiological state and attention level, learning progress indicators reflecting synchronization with the class teaching progress, and learning efficiency indicators reflecting the rate of knowledge acquisition per unit time.

[0052] Based on these indicators, decision variables and constraints for a multi-objective optimization function can be constructed. The decision variables are the daily learning time allocation schemes for each subject, and the constraints include that the total learning time does not exceed the preset upper limit and the continuous learning time for a single subject does not exceed the anti-fatigue threshold.

[0053] The subject ability index, learning effort index, learning progress index, and learning efficiency index are then input into a pre-defined multi-objective optimization model. A non-dominated sorting genetic algorithm is then used for iterative solution. The model first randomly generates an initial population, with each individual representing a learning time allocation scheme. Then, the performance of each scheme on four objective functions is calculated. Objective function one is to maximize subject balance, which means reducing the gap in mastery between subjects by prioritizing the allocation of time to weaker subjects. Objective function two is to maximize effort adaptation, which means arranging high cognitive load learning tasks during the individual's peak effort period. Objective function three is to maximize progress synchronization, which means prioritizing learning content that helps catch up with the class's progress. Objective function four is to maximize efficiency, which means prioritizing the learning of knowledge points with high mastery rates.

[0054] Individuals in the population can be divided into Pareto fronts of different levels through non-dominated sorting. The diversity of solutions is maintained by crowding distance calculation. After multiple iterative evolutions through selection, crossover, and mutation operations, a set of Pareto optimal solutions is finally converged. Each solution corresponds to a specific daily learning time allocation scheme for each subject.

[0055] Then, the student's learning habits and historical behavior data are retrieved, and the optimal learning plan that best matches the student's daily routine and learning preferences is selected from the Pareto optimal solution set. The plan includes the daily allocation of study time for each subject, specific study time arrangements, suggestions on the types of study content, and suggestions on rest reminders.

[0056] Finally, the generated personalized learning plan is pushed to the student's terminal device in a structured data format through the mobile communication network or the Internet of the communication unit, and the status changes of the student after implementing the plan are continuously tracked in the subsequent learning process, forming a closed-loop optimization mechanism of evaluation, recommendation, execution and feedback.

[0057] For example, after generating a multi-dimensional comprehensive assessment report for Xiaoming, a second-year junior high school student, the student proceeds to the learning plan generation stage. Indicators extracted from the report show: a percentile of 87% in mathematics (a strong subject), a percentile of 32% in English (a weak subject), a learning energy indicator showing peak energy levels between 9-11 AM, a learning progress indicator showing synchronization with the class's teaching progress, and a learning efficiency indicator showing an English vocabulary memorization efficiency of 85 points and a math problem-solving efficiency of 72 points.

[0058] These indicators are then input into a multi-objective optimization model, and after 200 generations of iteration using a non-dominated sorting genetic algorithm, a Pareto optimal solution set is obtained. Combined with Xiaoming's learning habits (preferring to learn new knowledge in the morning and review and consolidate in the evening), the optimal solution is selected: a total study time of 2.5 hours each evening, including 1.5 hours of English (18:30-20:00, prioritizing weaker subjects and maximizing energy allocation) and 1 hour of mathematics (20:15-21:15, consolidating stronger subjects). For English learning, interactive vocabulary quizzes are recommended (maximizing efficiency), and for mathematics, extended application problems are recommended (synchronizing with class progress). Finally, this plan is pushed to Xiaoming via mobile network.

[0059] Based on the above technical solution, by constructing an individual knowledge graph comprising a knowledge element layer, a skill layer, and a competency layer, the limitations of existing technologies that only collect surface-level learning behavior data are overcome. The knowledge element layer accurately captures students' mastery of basic knowledge points, the skill layer reflects students' ability to apply knowledge, and the competency layer uncovers students' thinking habits, transfer abilities, and other deeper cognitive states. This effectively solves the problem of single, superficial assessment dimensions and the inability to capture deep cognitive states, providing comprehensive cognitive support for accurate assessment. Furthermore, by combining this with a physiological signal database, the shortcomings of existing technologies in integrating physiological state information are addressed. This allows for the accurate capture of students' fatigue levels, attention levels, and other physiological states, avoiding misjudging abnormal learning performance due to poor physiological states as insufficient knowledge mastery, further improving assessment accuracy. Simultaneously, by using the individual knowledge graph as federated nodes and combining it with the physiological signal database to construct a group knowledge graph federation, comparative analysis between individuals and groups can be achieved while ensuring data privacy. This provides a reasonable reference for individual assessment and assists in optimizing assessment results. Finally, actual teaching progress data for each class is collected, and a teaching progress timeline is constructed based on knowledge points. The progress offset is calculated to solve the problem that existing technologies ignore the dynamic impact of class teaching progress. Based on the actual teaching progress of the student's class, the rationality of the student's learning progress can be accurately judged, avoiding a one-size-fits-all assessment that is detached from the teaching context. A comprehensive assessment of individual knowledge graphs, group knowledge graph federation, and progress offset can be obtained to produce a multi-dimensional comprehensive assessment report. This report integrates information from multiple aspects such as cognitive state, physiological state, and teaching context, completely changing the one-sided and subjective status quo of existing technology assessments and achieving a comprehensive, objective, and accurate assessment of students' learning status.

[0060] In another possible implementation of the embodiments of this application, combined with Figure 1-2 As shown, the construction process of an individual knowledge graph can be achieved through the following steps 201 to 205, which are explained in detail below: Step 201: Collect students' multi-source learning behavior data and extract features to generate a multi-dimensional feature space.

[0061] The multidimensional feature space includes features related to knowledge mastery, learning habits, knowledge application, cognitive transfer, and learning style. Learning habits describe patterns in students' study time regularity, review frequency, and pre-study behavior. Knowledge application measures students' ability to apply learned knowledge to solve complex problems and complete higher-order tasks. Cognitive transfer reflects students' ability to apply knowledge flexibly in different contexts. Learning style characteristics characterize students' preferences in information reception and processing, such as a preference for visual, auditory, or kinesthetic learning.

[0062] In some implementations, communication units connect to various learning platform databases to collect raw, multi-source learning behavior data generated by students periodically or in real-time. This allows for the application of various algorithms to process the heterogeneous data. For the answer data, the probability of mastering each knowledge point is calculated using the Item Response Theory model to form a knowledge mastery characteristic.

[0063] For time-series behavior logs (such as learning periods and clickstreams), statistical analysis methods are used to extract learning patterns and rhythms, forming characteristics of learning habits.

[0064] It can analyze the problem-solving steps and error types in students' homework and exams, use pre-trained deep learning models to evaluate the strategy and complexity of their problem-solving, and generate knowledge application capability features.

[0065] Then, by analyzing students' performance transfer across different knowledge points and question types, we can use association rule mining or graph neural networks to calculate the patterns and breadth of their knowledge transfer, thereby obtaining the characteristics of thinking transfer.

[0066] Finally, based on students' interaction preferences and completion status with different media types (video, text, and audio), a Bayesian personalized ranking algorithm is used to infer students' learning style characteristics. These five types of feature vectors can then be concatenated and normalized to form a multi-dimensional feature space, which serves as the basic input for constructing an individual knowledge graph.

[0067] It should be noted that the multidimensional feature space will periodically re-trigger the feature extraction process as students continuously generate new learning behavior data, updating the feature values ​​of each dimension in an incremental learning manner to ensure that it can reflect the students' latest cognitive state and behavioral patterns in real time and accurately.

[0068] Step 202: Construct a knowledge meta-layer using knowledge mastery features as nodes. The knowledge meta-layer consists of nodes representing knowledge points and edges representing semantic relationships between knowledge points.

[0069] In some implementation methods, based on national curriculum standards, textbook catalogs, and expert experience, a knowledge point system for a subject is predefined, the granularity of all knowledge points and their unique identifiers are clarified, and the semantic association types between knowledge points are sorted out to construct a standard knowledge point relationship template library.

[0070] This allows us to extract all knowledge mastery feature values ​​corresponding to each knowledge point from the previously constructed multidimensional feature space. Then, using each knowledge point as a node, its knowledge mastery feature value as the node's core attribute, and the semantic associations defined in the knowledge point relationship template library as edges, we begin constructing a graph structure.

[0071] During the construction process, a unique ID is assigned to each knowledge point node, and the mastery degree value is stored in the node as an initial attribute. For edges, different weights are assigned according to the preset relationship type (such as predecessor, containment, similarity). For example, edges with "predecessor relationship" have higher weights to reflect their strong dependency characteristics.

[0072] Finally, all nodes and edges are integrated to form a knowledge element layer subgraph with knowledge points as the granularity, mastery as the attribute, and semantic association as the skeleton. At this point, the subgraph completely maps the student's mastery status of all isolated knowledge points and their internal logical structure at the current moment.

[0073] Step 203: Input the knowledge element layer into the first graph neural network, and aggregate the nodes of the knowledge element layer to obtain the skill layer, using the knowledge application capability feature and the thinking transfer feature as constraints.

[0074] The first graph neural network is a graph convolutional network used to aggregate nodes in the lower-level knowledge layer into higher-order skill layers. Its core mechanism is to update the representation of the central node by iteratively aggregating information from neighboring nodes. The skill layer consists of nodes representing core skills.

[0075] In some implementations, the knowledge meta-layer subgraph is fed into the first graph neural network. Since the first graph neural network consists of multiple stacked graph convolutional layers, each layer performs a neighbor node feature aggregation operation: for each knowledge point node in the knowledge meta-layer, the network collects the features (i.e., knowledge mastery) of all its neighboring nodes (connected by semantically related edges), performs a linear transformation using a learnable weight matrix, and then performs a non-linear mapping using an activation function (such as ReLU), thereby updating the node's representation. After multiple convolutional layers, the originally isolated knowledge point representation has incorporated the contextual information of surrounding knowledge points.

[0076] When training the first graph neural network, the loss function consists of two parts: one part is the traditional graph reconstruction loss, which ensures that the aggregated node representations can retain the structural information of the original graph; the other part is a constraint-based regularization term, which calculates the difference (such as mean square error) between the candidate skill node representations generated after aggregation and the knowledge application ability feature vectors and thinking transfer feature vectors extracted from student behavior data. Through the backpropagation algorithm, the aggregation method learned by the network is forced to produce node representations consistent with these two higher-order ability features.

[0077] In practical applications after training, the knowledge layer of the current student is input into the trained first graph neural network. The network automatically performs aggregation operations, and the final output node is the core skill node of the skill layer. Each node represents a high-order ability, and its vector representation contains the combination of knowledge points supporting the ability and their weights.

[0078] Taking the "Ohm's Law" unit in junior high school physics as an example, a student's knowledge layer contains multiple knowledge point nodes such as "voltage," "current," "resistance," "Ohm's Law formula," "series circuit," and "parallel circuit," each with a mastery attribute. After inputting this subgraph into the first trained neural network, the network, through multi-layer aggregation, discovers that the nodes "voltage," "current," "resistance," and "Ohm's Law formula" are frequently activated simultaneously and are closely related. Therefore, under the constraints of knowledge application ability characteristics (the student performs well in experimental questions) and thinking transfer characteristics (the student can transfer circuit knowledge to practical problems), the network aggregates these nodes into a skill node representing "basic circuit calculation ability."

[0079] Meanwhile, the network further aggregates nodes such as "series circuits" and "parallel circuits" with the aforementioned basic knowledge points to form another skill node representing "complex circuit analysis ability." Ultimately, the originally scattered knowledge point nodes are aggregated into two core skill nodes, clearly reflecting the two key abilities the student has developed in physics.

[0080] Step 204: Input the skill layer into the second neural network. Using learning habit features and learning style features as constraints, aggregate the nodes of the skill layer to obtain the literacy layer.

[0081] The second graph neural network is a graph convolutional network used to further aggregate skill layer nodes into a higher-order competency layer. Its core mechanism is similar to that of the first graph neural network, but its target layer is higher. The competency layer consists of nodes representing comprehensive competencies.

[0082] In some implementations, the skill layer subgraph is fed into a pre-trained second graph neural network. This network also consists of multiple stacked graph convolutional layers, each performing a neighbor node feature aggregation operation. For each skill node in the skill layer (representing a core competency, such as "application of functional thinking" or "complex circuit analysis ability"), the network collects features from other skill nodes connected to it via semantically related edges, performs a linear transformation using a learnable weight matrix, and then performs a non-linear mapping using an activation function to update the representation of the current skill node. After multiple convolutional layers, the representation of a single skill node has incorporated information from other related skill nodes.

[0083] Meanwhile, when training the second graph neural network, the loss function consists of two parts: one part is the graph reconstruction loss, which ensures that the aggregated node representation can retain the structural information of the skill layer subgraph; the other part is a constraint-based regularization term, which calculates the difference between the candidate literacy node representation generated after aggregation and the learning habit feature vector and learning style feature vector extracted from student behavior data. Through the backpropagation algorithm, the aggregation method learned by the network is forced to produce node representations consistent with these two stable personality features.

[0084] When the training is completed and applied in practice, the current student's skill layer subgraph is input into the trained second neural network. The network automatically performs the aggregation operation, and the final output node is the comprehensive literacy node of the literacy layer. Each node represents an internalized learning quality, and its vector representation contains multiple core skills that support the literacy and their combination.

[0085] Taking a high school student as an example, their skill layer includes multiple skill nodes such as "application of function thinking," "geometric intuition," "data analysis ability," and "mathematical modeling ability." After inputting this skill layer subgraph into a trained second-graph neural network, the network, through multi-layer aggregation, discovered that the student not only excelled in "application of function thinking" and "geometric intuition," but also that these skill nodes exhibited frequent co-activation patterns with the "data analysis ability" node (e.g., simultaneous activation when solving complex problems). Furthermore, combining the student's learning habits (daily review at fixed times, regular summarization of incorrect answers) and learning style characteristics (preference for logical reasoning, field-independent learning), the network aggregated these related skill nodes into a competency node representing "mathematical logical reasoning literacy."

[0086] Meanwhile, the network further aggregated skill nodes such as "mathematical modeling ability" and "data analysis ability" with the student's ability to transform practical problems into mathematical problems, forming another competency node representing "mathematical modeling literacy." Ultimately, multiple skill nodes were aggregated into two core competency nodes, clearly reflecting the student's comprehensive learning qualities in mathematics.

[0087] Step 205: Establish hierarchical connections between the knowledge layer, skill layer, and literacy layer to generate a complete individual knowledge graph.

[0088] Among them, the hierarchical association edge is a special edge that connects nodes at different graph levels. It is used to represent the subordinate or support relationship between knowledge point nodes in the knowledge element layer and skill nodes in the skill layer, as well as between skill nodes in the skill layer and literacy nodes in the literacy layer. The existence of these edges connects the originally isolated three-layer structure into an organic and traceable hierarchical knowledge network.

[0089] In some implementations, the knowledge element layer subgraph, skill layer subgraph, and literacy layer subgraph are obtained. For each skill node in the skill layer, the set of knowledge element layer nodes corresponding to it during the aggregation process of the first graph neural network is traced back (these nodes are identified as core knowledge points supporting the skill through the attention weights of the graph convolutional network). Then, based on the contribution weights assigned to each knowledge element layer node in the last layer of the network during aggregation, directed hierarchical edges are established from these knowledge element layer nodes to the skill node. The weight of the edge is the contribution weight, representing the importance of the knowledge point in forming the skill.

[0090] Similarly, for each competency node in the competency layer, backtrack the set of skill layer nodes corresponding to it during the aggregation process of the second graph neural network, and establish directed hierarchical association edges from these skill layer nodes to the competency node based on the contribution weights during aggregation.

[0091] After establishing all inter-layer edges, the three subgraphs—knowledge element layer, skill layer, and competency layer—along with all intra-layer semantic association edges and inter-layer hierarchical association edges, are integrated to form a multi-level, fully connected heterogeneous graph structure. This allows for the assignment of a unique identifier to each node and edge, and the storage of all attributes (such as mastery level and contribution weight) in the graph database, ultimately generating a complete, queryable, and computable individual knowledge graph.

[0092] Based on the above technical solutions, this approach overcomes the limitations of surface-level learning behavior data through multi-dimensional feature space extraction, comprehensively characterizing student learning behavior from multiple dimensions, including knowledge mastery, learning behavior patterns, knowledge application ability, thinking flexibility, and information processing preferences. The knowledge meta-layer accurately depicts students' mastery of specific knowledge points and the logical connections between them, solving the problem that existing assessments cannot clearly present the composition of students' knowledge systems. The first graph neural network transforms fragmented knowledge point mastery into representations of students' core skills, capturing students' higher-order cognitive abilities to apply knowledge to solve problems and draw inferences, thus addressing the shortcomings of existing assessments that only focus on knowledge point mastery while neglecting knowledge application and thinking abilities. The second graph neural network further abstracts from the skill level to students' comprehensive learning qualities, uncovering students' internalized, stable, deep cognitive and behavioral characteristics, solving the problem that existing technologies cannot reach the level of student learning literacy assessment. This allows for the establishment of hierarchical connections between the three levels, linking the knowledge element layer, skill layer, and literacy layer into an organic whole. This forms a hierarchical cognitive map from specific knowledge points to core skills and then to comprehensive literacy, enabling a full-dimensional representation of students' learning status from surface behavior to deep cognition, and from single knowledge points to overall literacy. This allows assessments to penetrate surface behavioral data and accurately capture students' deep cognitive states such as knowledge structure, application ability, thinking patterns, and comprehensive literacy, changing the current assessment's single dimension and one-sided interpretation of students' learning status.

[0093] In another possible implementation of the embodiments of this application, combined with Figure 1-3 As shown, the construction process of the group knowledge graph federation can be achieved through the following steps 301 to 305, which are explained in detail below: Step 301: Extract the student's knowledge element subgraph, skill subgraph, and resting physiological feature vector from the individual knowledge graph and physiological signal database, and input them into the cross-modal joint embedding model to obtain the cross-modal joint embedding vector.

[0094] The cross-modal joint embedding model includes a graph encoding layer, a physiological encoding layer, a bilinear attention layer, and a joint embedding layer. The graph encoding layer consists of a first graph convolutional network and a second graph convolutional network. The first graph convolutional network takes the knowledge element layer subgraph as input to obtain the knowledge element layer graph embedding vector, and the second graph convolutional network takes the skill layer subgraph as input to obtain the skill layer graph embedding vector. The physiological encoding layer takes the resting physiological feature vector as input, performs nonlinear transformation through a multilayer perceptron, and outputs the physiological feature embedding vector. The bilinear attention layer calculates the first attention weight between the knowledge element layer graph embedding vector and the physiological feature embedding vector, and the second attention weight between the skill layer graph embedding vector and the physiological feature embedding vector. The joint embedding layer performs weighted fusion of the knowledge element layer graph embedding vector and the physiological feature embedding vector based on the first attention weight, and performs weighted fusion of the skill layer graph embedding vector and the physiological feature embedding vector based on the second attention weight. The results of the two weighted fusions are concatenated and mapped through a fully connected layer to output the cross-modal joint embedding vector.

[0095] In some implementations, data is processed in parallel through graph coding layers. The first graph convolutional network performs multi-layer convolution and pooling operations on the input knowledge element layer subgraph, aggregates neighbor node information, and finally outputs a fixed-dimensional knowledge element layer graph embedding vector, which encodes the distribution and structure of knowledge point mastery. Simultaneously, the second graph convolutional network performs the same operation on the skill layer subgraph, outputting a skill layer graph embedding vector that encodes the representation of the core skill. Meanwhile, the physiological coding layer uses a multilayer perceptron (MLP) to perform a nonlinear transformation on the input resting physiological feature vector, learning its high-order abstract representation and outputting a physiological feature embedding vector.

[0096] Subsequently, these three embedding vectors are fed into a bilinear attention layer, where bilinear pooling operations are used to calculate the attention weights (first attention weights) between the knowledge element layer graph embedding vector and the physiological feature embedding vector, and the attention weights (second attention weights) between the skill layer graph embedding vector and the physiological feature embedding vector, thereby quantifying the correlation between different modalities.

[0097] Finally, the joint embedding layer performs a weighted summation of the knowledge element layer graph embedding vector and the physiological feature embedding vector based on the first attention weight to achieve fusion. Similarly, the skill layer graph embedding vector and the physiological feature embedding vector are weighted and fused based on the second attention weight. The results of the two fusions are then concatenated into vectors and passed through a fully connected layer for final dimensional transformation and nonlinear mapping, outputting a single, dimension-unified cross-modal joint embedding vector.

[0098] For example, the system extracts a knowledge element layer subgraph about the "quadratic equation" unit from Xiaoming's individual knowledge graph. The knowledge points "formula method" and "completing the square method" have high mastery levels and are related, forming a knowledge element layer graph embedding vector. Simultaneously, it extracts a skill layer subgraph of "algebraic computation ability" aggregated from these knowledge points, forming a skill layer graph embedding vector. Next, the system retrieves Xiaoming's most recent resting state physiological feature vector from his physiological signal database, which includes indicators such as a heart rate variability (HRV) of 45ms and an EEG theta / β ratio of 1.5.

[0099] These vectors can then be input into a cross-modal joint embedding model. By using a bilinear attention layer, it was found that Xiaoming's low HRV (possibly due to mild fatigue) was highly correlated with his "algebraic computation ability" skill node. Therefore, a higher weight was assigned to the second attention layer to explore the impact of fatigue on core skills.

[0100] At this point, the joint embedding layer fuses these weights to generate a cross-modal joint embedding vector that integrates Xiaoming's current performance in "algebraic computation ability" and his fatigue state. This vector will serve as the unique identifier for the "federation node," used in subsequent construction of the group knowledge graph federation and for horizontal comparison with other students.

[0101] Step 302: Using each student as a federation node, perform federation aggregation on the cross-modal joint embedding vectors to obtain the group federation center vector.

[0102] In some implementations, an independent federated node is established for each student, and each node locally stores the student's cross-modal joint embedding vector generated by the cross-modal joint embedding model. At the start of each round of federated aggregation, the central server issues the current global model parameters or aggregation instructions to all participating federated nodes.

[0103] After receiving the instruction, each node encrypts its local cross-modal joint embedding vector (such as homomorphic encryption or differential privacy perturbation), and only uploads the encrypted gradient information or vector parameters to the central server, while the original vector remains locally and does not leave the node.

[0104] Once the central server has collected the encrypted information from all participating nodes, it performs a weighted summation and average of the uploaded encrypted parameters based on the number of students represented by each node or a preset weight coefficient, resulting in a new aggregate vector. This aggregate vector is the group federation center vector for the current round, integrating information from all participating nodes and reflecting the common association patterns of the entire student group.

[0105] At this point, the central server redistributes the newly generated group federation center vector to all federation nodes, and each node uses the center vector to update its local model or for subsequent horizontal comparison calculations.

[0106] Suppose that in a math learning assessment for eighth-grade students in a certain region, the system sets up a federated node for each of 500 students, including Xiaoming and Xiaohong. Each node locally stores a cross-modal joint embedding vector representing its own learning status. In a certain round of aggregation, the central server issues an aggregation command to these 500 nodes. Each node uploads its local cross-modal joint embedding vector to the server after homomorphic encryption. At this time, the server calculates the group federated center vector for this round by weighted averaging based on the representativeness of each student (node) (e.g., a weighted average based on their recent learning activity). For example, this vector might be [0.65, 0.32, 0.78, ...], representing the common pattern of these 500 students' mastery of the "quadratic function" knowledge point and their average physiological state (e.g., average HRV value) at the current stage. Simultaneously, the server broadcasts this vector to all nodes.

[0107] After receiving the central vector, Xiaoming's node compares it with its own local vector to gain a preliminary understanding of its relative position within the group. Through multiple rounds of such iterations and aggregations, a stable and representative group federation central vector is finally obtained, serving as an objective reference benchmark for the learning status of all eighth-grade students in the region in this subject.

[0108] Step 303: Calculate the cross-evaluation index of students in the group based on the group federated central vector and cross-modal joint embedding vector.

[0109] In some implementations, the group's federated center vector is obtained and stored on a central server or broadcast to each federated node as a common reference benchmark for students of the same grade in the current region. This allows each federated node to retrieve the student's unique cross-modal joint embedding vector locally.

[0110] Therefore, the cosine similarity between the two vectors is calculated first, and a value between -1 and 1 is obtained, which is used to measure the consistency between individuals and groups in direction.

[0111] Simultaneously, the Euclidean distance is calculated to obtain a non-negative value, which is used to measure the absolute difference between individuals and groups in the numerical space.

[0112] Calculate the dimension-wise deviation of the individual vector from the group center vector in each dimension to obtain the deviation value of mastery of each knowledge point and the deviation value of each physiological indicator.

[0113] This allows us to use pre-built population distribution statistics (such as mean, standard deviation, and percentile thresholds for each dimension) to map each dimension of an individual vector to its relative position in the population distribution. Specifically, this is achieved by calculating the Z-score (individual value minus the population mean, then divided by the population standard deviation) or percentile rank (the percentage of individuals whose value is lower than that in the population).

[0114] Finally, the results of the above multiple measurements are combined to generate a multi-dimensional cross-sectional evaluation index vector. This vector includes sub-indicators such as the individual's group percentile in knowledge mastery, deviation level in skill formation, and relative position in physiological state, which comprehensively reflects the student's cross-sectional comparison in the group.

[0115] Step 304: Collect students' historical cross-modal joint embedding vectors to form a temporal embedding vector sequence. Compare and analyze the temporal embedding vector sequence with the group's federated center vector to calculate the longitudinal development trajectory index.

[0116] Among them, the historical cross-modal joint embedding vector refers to the historical record of cross-modal joint embedding vectors generated by the system for the student at multiple time points in the past (such as weekly or monthly). Each vector represents a comprehensive representation of the student's knowledge, skills, and physiological state at the corresponding time point. The temporal embedding vector sequence is a set of historical cross-modal joint embedding vectors of the same student arranged in chronological order, forming a sequence of data that can reflect the evolution trajectory of the student's cognitive and physiological state over time.

[0117] In some implementations, all historical cross-modal joint embedding vectors generated by the student over a period of time (such as a semester) are retrieved from the local storage of each federated node and arranged in chronological order of the timestamps collected to construct a temporal embedding vector sequence. At this time, each time point in the sequence corresponds to a vector, which fully records the evolution of the individual's state.

[0118] The group federation center vectors generated by each round of federation aggregation within the same time period are obtained from the central server and arranged in chronological order to form a time sequence of group federation center vectors, which reflects the dynamic evolution trajectory of the group's common patterns.

[0119] The dynamic time warping algorithm is used to align the two sequences temporally, eliminating potential misalignment between individuals and the group on the time axis. This involves calculating the difference vector between the individual vector and the group center vector at each aligned time point.

[0120] The cumulative difference trend across the entire sequence can be obtained by calculating the consistency of the offset angle, offset magnitude, and offset direction of individual trajectories relative to the group trajectories.

[0121] Meanwhile, in order to quantify the longitudinal development trajectory indicators, the trajectory similarity measurement method is adopted to calculate the Pearson correlation coefficient between the individual time series and the group time series, and obtain a value between -1 and 1, which is used to measure the correlation between the individual development trend and the group development trend. Calculate the Frescher distance between two sequences to measure the geometric similarity between the two trajectory curves.

[0122] Finally, by combining the progress offset representation path in the progress offset, the longitudinal development trajectory is corrected to generate the final longitudinal development trajectory index. This index includes sub-indicators such as the synchronicity score of individual development trajectory and group common trajectory, the early warning level of development trend deviation, and the comparison results of growth rate.

[0123] Suppose that student Xiaoming has 6 cross-modal joint embedding vector records in the first semester of his second year of junior high school, from t1 to t6 in chronological order, forming his temporal embedding vector sequence. During the same period, the regional group federal center vector also has 6 records, from T1 to T6 in chronological order.

[0124] The system first uses a dynamic time warping algorithm to align Xiaoming's trajectory with the group's trajectory on the time axis, finding that Xiaoming's trajectory is basically synchronized and requires no significant adjustment. Then, it calculates the difference at each time point: at t1, Xiaoming is 0.15 below the group average in the "quadratic function" knowledge point; at t2, the gap narrows to 0.10; at t3, the gap is 0.08; at t4, he surpasses the group by 0.02; at t5, he leads by 0.05; and at t6, he leads by 0.08. The Pearson correlation coefficient for the entire sequence is calculated to be 0.92, indicating a high positive correlation between Xiaoming's development trend and the group's; the Fraser distance is calculated to be 0.18, indicating that the two trajectory curves are similar in shape. Simultaneously, combined with the progress offset, it is found that Xiaoming's class teaching progress lagged behind the regional average from t1 to t3, therefore his lag at that time is attributed to differences in teaching progress.

[0125] The final longitudinal development trajectory index is: "The individual development trajectory is highly synchronized with the group's common trajectory (correlation coefficient 0.92), showing a trend of catching up and surpassing, with a growth rate better than the group average, and no abnormal development warnings." This comprehensively reflects Xiaoming's dynamic growth process throughout the semester and his relationship with the group.

[0126] Step 305: Integrate horizontal evaluation indicators and vertical development trajectory indicators to generate a federation of group knowledge graphs.

[0127] In some implementations, by obtaining the horizontal evaluation index vector and the vertical development trajectory index vector, the horizontal evaluation index can be stored in the form of node attributes, including the student's group percentile in each knowledge point, the deviation level of each skill dimension, and the relative position of each physiological indicator; the vertical development trajectory index is also stored in the form of node attributes, including the synchronicity score of individual development trajectory and group common trajectory, the warning level of development trend deviation, and the comparison results of growth rate.

[0128] This allows for the mapping of these federation nodes with horizontal and vertical indicator attributes to the group federation center vector, constructing semantic association edges between nodes. Specifically, it establishes a "horizontal reference edge" for each federation node pointing to the group federation center vector, with the edge weight being the comprehensive score of the horizontal evaluation indicators; and establishes a "temporal evolution edge" between each federation node and its historical version nodes, forming a graph representation of the individual development trajectory. Establish "group association edges" based on similarity between different federated nodes. When the horizontal evaluation index or vertical trajectory similarity between two nodes exceeds a preset threshold, an association edge is automatically established.

[0129] Finally, all federated nodes, group federated center vectors, various associated edges, and all additional horizontal and vertical indicator attributes are uniformly integrated to form a complete, queryable, and computable group knowledge graph federated structure. This structure supports both group distribution queries by knowledge point dimension and horizontal comparison and vertical tracking analysis by individual dimension.

[0130] Based on the above technical solution, a joint embedding vector is obtained by fusing individual knowledge graph subgraphs and resting physiological feature vectors through a cross-modal joint embedding model. This vector is then federated to obtain a group federation center vector. Combined with horizontal and vertical indicators, a group knowledge graph federation is generated, solving the technical problems of existing assessments lacking group reference, failing to integrate physiological states, and being unable to track individual development trajectories. Specifically, cross-modal fusion realizes the intrinsic relationship between knowledge cognition and physiological states, compensating for the one-sidedness of single behavioral data assessment; federated aggregation constructs a common group reference system while protecting data privacy, providing an objective group comparison standard for individual assessment; horizontal indicators clarify the individual's relative position within the group, and vertical indicators capture the deviation between the individual's development trajectory and the group's; the federated graph generated by the fusion of these two indicators enables horizontal group comparison of individual learning states and vertical development tracking, significantly improving the objectivity and comprehensiveness of the assessment and effectively distinguishing between ability problems and performance problems caused by physiological states.

[0131] In another possible implementation of the embodiments of this application, combined with Figure 1-4 As shown, the process of calculating the progress offset between the student's current learning progress and the class's teaching progress can be achieved through the following steps 401 to 405, which are explained in detail below: Step 401: Obtain the actual teaching progress sequence of the target class.

[0132] The actual teaching progress sequence is arranged in chronological order with knowledge points as the granularity. The granularity of knowledge points means that the teaching progress is recorded and tracked at the smallest knowledge unit, rather than at the traditional chapter or lesson level.

[0133] In some implementations, teaching progress records submitted by teachers in the school's teaching management system are collected in real time, or the currently taught knowledge points marked in the classroom interaction system are automatically captured. An initial sequence of actual teaching progress is generated according to the completion date of each knowledge point. This sequence is then preprocessed, including imputing missing knowledge points, removing abnormal time points, and standardizing the sequence length to ensure that each knowledge point corresponds to a unique timestamp. This allows the construction of a semantic association matrix between knowledge points, transforming isolated knowledge point sequences into a graph structure sequence with dependencies. Finally, the processed sequence is stored in a time-series database, with class and time indexes established. The data is also periodically compared and verified with the teaching syllabus to ensure the accuracy and real-time nature of the progress data.

[0134] Step 402: Extract the learning progress trajectory from the individual knowledge graph, and input the learning progress trajectory and the actual teaching progress sequence into the progress decoupling comparison network.

[0135] The learning progress trajectory refers to the time-series representation of a student's actual learning path extracted from an individual's knowledge graph, with nodes representing the time points when knowledge points are completed and reach a mastery threshold. The progress decoupling comparison network is a neural network model used to compare and analyze the dynamic relationship between individual learning trajectories and class teaching progress. It includes a temporal encoding layer and a dynamic alignment layer. The temporal encoding layer is responsible for extracting temporal features from the input actual teaching progress sequence and learning progress trajectory, capturing the temporal dependencies and knowledge point sequence features in the sequences through recurrent neural networks or Transformer structures. The dynamic alignment layer, based on the feature vector output by the temporal encoding layer, uses a differentiable dynamic time warping algorithm to calculate the optimal alignment path and minimum cumulative distance between the two sequences.

[0136] In some implementations, knowledge points that students have completed learning and whose mastery level has reached a preset threshold (e.g., 0.8) and their corresponding timestamps are extracted from individual knowledge graphs, and a student's learning progress trajectory is constructed in chronological order. This trajectory is then input into a decoupled progress comparison network in parallel with the actual teaching progress sequence of the class.

[0137] Step 403: The temporal coding layer extracts temporal features from the actual teaching progress sequence and the learning progress trajectory to obtain the class teaching progress temporal vector and the student individual progress temporal vector.

[0138] In some implementations, each knowledge point in the actual teaching progress sequence is mapped to a corresponding embedding vector (obtained through a pre-trained knowledge point embedding table). Each knowledge point corresponds to a fixed-dimensional distributed representation, and a positional encoding is added at each time step to preserve temporal order information. These embedding vector sequences are then input into a bidirectional long short-term memory network (consisting of forward and backward LSTM layers). The forward LSTM captures the historical dependencies of the sequence in chronological order, while the backward LSTM captures the future context information of the sequence in reverse order. At each time step, the hidden states from both directions are concatenated to obtain temporal features that fuse bidirectional information.

[0139] The same processing method is used for the learning progress trajectory. The same knowledge point embedding table is used to ensure semantic space consistency. Feature extraction is performed through a bidirectional LSTM network with shared weights, and the hidden state sequence of each time step is output.

[0140] This allows us to perform average pooling on the hidden states of the entire sequence or take the hidden state of the last time step to obtain a fixed-dimensional time-series vector of class teaching progress and a time-series vector of individual student progress.

[0141] The bidirectional LSTM network is trained using a self-supervised learning approach, with sequence prediction as the pre-training target. In practical applications, it is fine-tuned and optimized using the loss function of the progress alignment task.

[0142] Step 404: The dynamic alignment layer uses a differential dynamic time warping algorithm to calculate the minimum cumulative distance between the class teaching progress time sequence vector and the student individual progress time sequence vector, and uses the warped path corresponding to the minimum cumulative distance as the progress offset representation path.

[0143] Among them, the differential dynamic time warping algorithm is a differentiable version improved from the traditional dynamic time warping algorithm. By introducing a smooth approximation technique, it enables gradient backpropagation in the originally non-differentiable alignment path calculation, thus supporting end-to-end neural network training. The minimum cumulative distance refers to the minimum total cost accumulated when matching two time series according to the optimal path during the alignment process, reflecting the overall degree of difference between the two sequences. The warped path is the set of correspondences connecting the time points of the two time series, describing which time points of the class teaching progress time series vector match the individual student progress time series vector under the optimal alignment scheme. The progress offset representation path refers to using the warped path corresponding to the minimum cumulative distance as the basic mapping relationship for subsequent analysis of progress lead, lag, and mastery drift.

[0144] In some implementations, the class teaching progress time-series vector sequence output by the time-series coding layer is obtained. Student individual progress time sequence vector sequence , where m and n are the lengths of the two sequences, respectively.

[0145] Then construct an m x n cumulative distance matrix D, where each element in the matrix... Initialize to and The Euclidean distance between them is used as the local cost. Then, a dynamic programming algorithm is used to calculate the cumulative distance matrix, with the recursive formula as follows: The entire matrix is ​​filled sequentially. To achieve differentiability, a softmin function is used instead of the original hard minimum selection when calculating the minimum. That is, the cumulative distances of the three candidate paths are weighted by a softmax function with a temperature parameter, so that the gradient can be backpropagated through the minimum selection operation. After obtaining the cumulative distance matrix, the path that minimizes the cumulative distance is found by backtracking from D(m,n). This path consists of a series of coordinate pairs (p1,q1), (p2,q2),...,(pk,qk), where p represents the time point index of the class teaching progress time series vector and q represents the time point index of the student individual progress time series vector. It satisfies the monotonicity and continuity constraints, which is the optimal regular path.

[0146] This path is then used as the progress offset representation path output, and D(m,n) is the minimum cumulative distance.

[0147] Step 405: Based on the progress offset representation path, analyze the progress advance, progress lag, and knowledge point mastery drift of the target student relative to the class teaching progress at different time slices, and generate the progress offset.

[0148] Among them, the knowledge point mastery drift refers to the degree of attenuation or transfer deviation in the quality of mastery of a knowledge point when there is a misalignment between the knowledge points mastered by students and the knowledge points taught in the class during the progress alignment process. It reflects the changes in the quality of knowledge mastery caused by abnormal learning order or excessive time interval.

[0149] In some implementations, the progress offset representation path (consisting of a series of coordinate pairs (p1, q1), (p2, q2), ..., (pk, qk), where p represents the time point index of the class teaching progress time series vector and q represents the time point index of the student's individual progress time series vector) can be obtained by traversing each aligned coordinate pair and calculating the difference in time point indices as the basic offset. When q is greater than p, it means that the student's progress is ahead of the class's progress, and the amount of the advance is the difference between q and p multiplied by the unit time step. When q is less than p, it means that the student's progress is lagging behind the class's progress. The lag is the difference between p and q multiplied by the unit time step.

[0150] For jump points or many-to-one alignments in the path, identify the skipped knowledge points or the repeatedly aligned knowledge points, and master the calculation of drift degree for these knowledge points.

[0151] Mastery drift is obtained by comparing the difference between the mastery of a knowledge point in an individual student's progress trajectory and the expected mastery of the knowledge point in the class's teaching progress. At the same time, the time interval decay factor is considered. The longer the time interval, the greater the mastery drift.

[0152] Finally, the lead-up, lag, and mastery drift of each abnormal alignment knowledge point on each time slice are vectorized and concatenated, and organized in chronological order to generate a structured progress offset vector. This vector contains three sub-vectors corresponding to the lead-up sequence, lag sequence, and drift sequence, as well as the corresponding knowledge point identifier and timestamp information.

[0153] Based on the above technical solution, this method calculates the progress offset by combining a progress decoupling comparison network with a differential dynamic time warping algorithm, addressing the technical problems of existing technologies neglecting the dynamic impact of class teaching progress and the disconnect between assessment and actual teaching. It first obtains the class teaching progress sequence at the knowledge point level and extracts the student learning progress trajectory. A temporal encoding layer extracts temporal features to obtain corresponding vectors. Then, a dynamic alignment layer uses a differential dynamic time warping algorithm to achieve dynamic elastic alignment of the sequence, calculates the minimum cumulative distance to obtain the progress offset representation path, and finally analyzes the advance amount, lag amount, and mastery drift to generate the progress offset. This method abandons the simple progress bar comparison approach, accurately quantifies the dynamic offset relationship between individual learning pace and class teaching pace, making the assessment more aligned with the real teaching context. It can distinguish whether a student's knowledge is insufficient or the teaching progress is not covered, avoiding a one-size-fits-all assessment that is detached from actual teaching, and improving the accuracy and objectivity of learning progress assessment.

[0154] In another possible implementation of the embodiments of this application, combined with Figure 1-5 As shown, the process of comprehensively evaluating personalized knowledge graphs, federated group knowledge graphs, and progress offsets to obtain a multi-dimensional comprehensive evaluation report can be achieved through the following steps 501 to 505, which are explained in detail below: Step 501: Based on the temporal sequence, granularly align the knowledge element layer nodes and skill layer nodes in the individual knowledge graph with the common association patterns in the group knowledge graph federation to generate joint knowledge data of individuals and groups, and input it into the cross-attention evaluation network.

[0155] Granular alignment refers to the precise matching of fine-grained knowledge element layer nodes and skill layer nodes in an individual knowledge graph with statistical features (such as group mastery distribution and group learning trajectory) representing common association patterns in the group knowledge graph federation, at both the knowledge point granularity and time dimensions. The common association patterns in the group are fine-grained statistical features parsed from the group federation center vector of the group knowledge graph federation, specifically manifested as the group mastery distribution parameters (such as mean, standard deviation, and percentiles) corresponding to each knowledge element layer node and the group skill intensity distribution corresponding to each skill layer node. The cross-attention evaluation network includes a horizontal comparison module and a vertical tracking module.

[0156] In some implementations, snapshots of the target student's individual knowledge graph across multiple historical time slices are retrieved from a time-series database (containing attribute values ​​of knowledge element and skill layer nodes at each moment). Simultaneously, the group federation center vector corresponding to the time slice is extracted from the group knowledge graph federation, and a decoder parses it into common group association patterns that can be compared at the node level. For example, it parses the mean and standard deviation of mastery of each knowledge point within the group, as well as typical learning paths between knowledge points.

[0157] Perform granular alignment: Pair each knowledge element node and skill layer node in the individual knowledge graph with its corresponding knowledge point and time slice in the group's common association pattern, according to its knowledge point identifier and timestamp. To handle possible minor misalignments between individuals and the group on the timeline, this process uses linear interpolation to smooth the group's common pattern over time, ensuring that each individual node's evaluation time has a corresponding, precisely matched group reference benchmark.

[0158] After the above pairing is completed, the originally independent individual nodes and group reference features can be encapsulated to generate a joint data unit containing individual node attributes (such as individual mastery and skill strength) and related group statistical features (such as group mean and percentile). All these units are organized according to the original knowledge graph structure to finally form structured individual and group joint knowledge data.

[0159] Step 502: The horizontal comparison module uses the common association pattern of the group as a reference to calculate the horizontal deviation of each node in the individual knowledge graph.

[0160] In some implementations, the individual node attributes, after granular alignment, and their corresponding common association patterns within the group are first obtained from the joint knowledge data of individuals and groups. For each knowledge element node, the individual's knowledge point mastery value is read, along with the mean μ and standard deviation σ of that knowledge point's mastery within the group. Subsequently, the standardized deviation of the node is calculated using the Z-score formula (Z = (individual mastery - μ) / σ), where the sign of the Z-score indicates how high or low the individual is relative to the group mean, and the absolute value indicates the magnitude of the deviation.

[0161] For skill-level nodes, the benchmark is set as the mean and standard deviation of the group's skill strength. Using similar calculation logic, after calculating the basic Z-score for each node, it can be mapped to an intuitive percentile ranking. By querying the standard normal cumulative distribution function table, the Z-score is converted into a ranking value that the individual surpasses by how many percent of students in the group on that knowledge point.

[0162] Simultaneously, nodes are labeled with deviation levels based on preset thresholds (such as an absolute Z-score greater than 1.5 or a percentile rank lower than 15% or higher than 85%), such as "significantly lower than the group", "within the normal range", or "significantly higher than the group". The lateral deviation calculation results of all nodes are organized into a deviation graph that is completely consistent with the individual knowledge graph structure. Each node is attached with three attributes: Z-score, percentile rank, and deviation level, as the complete output of lateral deviation.

[0163] Step 503: The longitudinal tracking module compares the cross-modal joint embedding vector with the group federation center vector in time series, and calculates the longitudinal development trajectory deviation by combining the progress offset characterization path.

[0164] In some implementations, all historical cross-modal joint embedding vectors of the target student over a period of time (such as a semester) are retrieved from the time-series database and arranged in chronological order to form an individual time-series embedding vector sequence; at the same time, all historical group federation center vectors at the corresponding time points are obtained from the group knowledge graph federation to form a group time-series center vector sequence.

[0165] By taking these two sequences as input and introducing a progress offset representation path as an auxiliary constraint for temporal alignment, a dynamic time warping algorithm can be used to flexibly align the two sequences. Furthermore, by constructing a cumulative distance matrix, the minimum cumulative distance between the two sequences under all possible alignment methods is calculated, thereby finding the optimal temporal matching relationship.

[0166] In the calculation process, the progress offset representation path is incorporated as a soft constraint into the distance calculation: for time slices marked as ahead or behind in the progress offset representation path, a penalty coefficient is introduced when calculating the vector distance at the corresponding time point, so that the alignment result can reasonably explain the misalignment of individual and group states caused by differences in teaching progress.

[0167] After alignment, the cosine similarity of vectors on each alignment point pair is calculated to obtain the state consistency score between the individual and the group at each time slice. Subsequently, statistical analysis is performed on the similarity sequence along the entire alignment path, calculating its mean, standard deviation, and trend slope to obtain an overall synchronicity index between the individual's development trajectory and the group's development trajectory.

[0168] Simultaneously, the Fraser distance between two sequences can be calculated to measure the geometric similarity between the two trajectory curves. Finally, the synchronicity index, Fraser distance, and the cumulative values ​​of progress lead / lag extracted from the progress deviation representation path are weighted and fused to generate a comprehensive longitudinal development trajectory deviation index vector, which includes three sub-dimensions: synchronicity score, trend deviation warning level, and growth rate comparison results.

[0169] Step 504: Jointly decode the lateral deviation and the longitudinal development trajectory deviation, and generate a multi-dimensional comprehensive evaluation report by using feature fusion and nonlinear mapping of the multilayer perceptron.

[0170] In some implementations, the joint decoding layer can be activated by receiving the lateral deviation vector H and the longitudinal development trajectory deviation vector V, and H and V can be connected in the feature dimension to form a joint feature vector U=[H;V] that integrates lateral and longitudinal information. At this time, the dimension of the vector is dim(H)+dim(V).

[0171] Subsequently, this joint feature vector is input into a pre-trained multilayer perceptron network. This multilayer perceptron consists of an input layer, two hidden layers, and an output layer. The number of neurons in the input layer is equal to the dimension of the joint feature vector; the first hidden layer contains 128 neurons and uses the ReLU activation function for non-linear transformation to learn higher-order interaction patterns between horizontal and vertical features. The second hidden layer contains 64 neurons and also uses the ReLU activation function to further abstract the feature representation.

[0172] The output layer contains the same number of neurons as the preset evaluation dimensions (such as subject ability, learning effort, learning progress, and learning efficiency). The Sigmoid activation function is used to compress the output values ​​to the 0-1 range, forming standardized scores for each dimension.

[0173] The training process of this multilayer perceptron employs supervised learning: it collects historical students' horizontal and vertical feature data as input, uses multi-dimensional expert-annotated scores as labels, employs mean squared error as the loss function, and optimizes the network weight parameters through backpropagation. After obtaining standardized scores for each dimension, a joint decoding layer further calls a pre-defined report generation template to map these values ​​into highly readable text descriptions.

[0174] For example, a subject ability score of 0.85 is mapped to "excellent mastery of mathematics, exceeding 85% of peers", and a learning progress score of 0.32 is mapped to "current learning progress is lagging behind class teaching, and it is recommended to catch up first".

[0175] Simultaneously, the module integrates a visualization generator that automatically draws a six-dimensional capability radar chart based on the scoring data and generates SWOT analysis text (strengths, weaknesses, opportunities, and threats) based on the relative levels of each dimension. Finally, it calls a trend prediction model to predict possible changes in mastery levels across subjects over the next two weeks based on growth rate information from longitudinal development trajectory deviations, generating trend curves and prediction text. All this structured and unstructured data is packaged into a complete, multi-dimensional comprehensive evaluation report in JSON format.

[0176] Based on the above technical solution, this method aligns individual knowledge graph nodes with common association patterns within a group through temporal sequence granular alignment. Combined with the horizontal and vertical module analysis of the cross-attention evaluation network and the fusion of the joint decoding layer, it addresses the problems of biased and inaccurate assessments caused by the lack of group reference, failure to incorporate dynamic teaching progress, and insufficient fusion of multi-source data in existing learning status assessments. Specifically, granular alignment achieves precise matching of individual and group data at the knowledge point and temporal granularity, laying a comparable foundation for assessment. The internal horizontal comparison module calculates node deviation using common group characteristics as a reference, clarifying the individual's relative position within the group and compensating for the isolation of individual assessments. The vertical tracking module performs temporal comparisons based on progress offset representation paths, integrating the dynamic context of teaching progress to avoid judgments detached from actual teaching. Finally, the joint decoding layer fuses horizontal and vertical indicators through a multilayer perceptron, achieving deep integration and nonlinear mapping of multi-source heterogeneous data, ultimately generating a multi-dimensional comprehensive assessment report. This allows the assessment to combine horizontal group reference with individual vertical development perspectives, while also aligning with actual teaching progress, significantly improving the comprehensiveness, objectivity, and accuracy of the assessment.

[0177] like Figure 6As shown, this embodiment provides a learning status assessment system based on big data analysis, which is applied to the learning status assessment methods based on big data analysis described in the above embodiments. The system includes a communication unit and a processing unit, and is logically divided into an individual knowledge graph construction module, a physiological signal database construction module, a group knowledge graph federated construction module, an offset calculation module, a comprehensive evaluation module, and a learning plan generation module, thereby achieving fully automated processing from data acquisition, graph construction, federated aggregation to evaluation decision-making.

[0178] Among them, the individual knowledge graph construction module serves as the system's data entry point and basic modeling unit. It connects to various learning platform databases through the communication unit to collect students' multi-source learning behavior data and construct individual knowledge graphs, which include knowledge element layers, skill layers, and literacy layers.

[0179] In actual operation, this module is not only responsible for data ETL (Extract, Transform, Load) processing, but also for the dynamic updating of the knowledge graph. For example, after a student completes an assignment on the online platform, the module captures behavioral data in real time, updates the mastery attribute of the corresponding node in the knowledge element layer through feature extraction algorithms, and triggers the aggregation calculation of the graph neural network, synchronously updating the status of the skill layer and the literacy layer. This ensures that the individual knowledge graph can reflect the student's latest cognitive status in real time, providing a dynamic and up-to-date data foundation for subsequent federated aggregation and comprehensive assessment.

[0180] The physiological signal database construction module establishes a data connection with smart wearable devices (such as smart bracelets and EEG headbands) through a communication unit to collect students' resting physiological signals (including at least ECG and EEG signals) and build a physiological signal database. This module also considers the complexity of the physiological signal acquisition environment, and therefore typically integrates a signal preprocessing submodule to perform noise reduction, filtering, and feature extraction on the collected raw signals.

[0181] For example, heart rate variability (HRV) indicators are extracted from raw electrocardiogram (ECG) signals, and the power spectral density ratio of alpha waves to beta waves is extracted from electroencephalogram (EEG) signals. The constructed physiological signal library is stored in a structured form and indexed with student IDs output by the individual knowledge graph construction module. This ensures that physiological state data of specific students can be quickly accessed in subsequent processing, providing the necessary physiological dimension support for cross-modal fusion.

[0182] When constructing the group knowledge graph federation, the module first maps heterogeneous graph data and physiological data into a unified feature space using its internally integrated pre-trained cross-modal joint embedding model (such as the model in the aforementioned embodiment that includes a graph coding layer, a physiological coding layer, and a bilinear attention layer). This allows for the simultaneous processing of embedding requests from multiple students through a parallel computing architecture, generating their respective cross-modal joint embedding vectors. Furthermore, this module successfully constructs a multi-dimensional reference coordinate system that includes both horizontal comparison and vertical tracking, ensuring that the system's output evaluation result is no longer an isolated absolute value but an objective assessment with relative group significance.

[0183] The offset calculation module collects the actual teaching progress data of each class by periodically synchronizing the progress data of the school's teaching management system through the communication unit or receiving progress information from the teacher's terminal. It then constructs a teaching progress timeline based on knowledge points and calculates the progress offset between the student's current learning progress and the class's teaching progress.

[0184] The comprehensive assessment module acts as the brain of the system, responsible for coordinating and processing multi-source heterogeneous data from other modules. Based on the cross-attention assessment network integrated within this module, it matches individual data with group data through a granular alignment mechanism and performs in-depth analysis through two sub-channels: horizontal comparison and vertical tracking. This generates a multi-dimensional comprehensive assessment report, which is stored in the form of structured data. The report covers multiple dimensions such as subject ability, learning effort, and progress synchronization, providing a comprehensive diagnostic basis for subsequent decision-making.

[0185] Therefore, through the collaborative work of the above six modules, this embodiment constructs a fully functional and logically rigorous learning status evaluation system. The modules interact through standardized data interfaces, ensuring both the independence of the processing flow and efficient data transfer. In particular, the refined design of the group knowledge graph federation construction module provides a solid group reference foundation for the system, effectively solving the technical problems of isolated evaluation results and lack of horizontal comparison in existing technologies.

[0186] like Figure 7-8 As shown, this embodiment provides a learning status assessment device based on big data analysis. This device is applied to the learning status assessment methods based on big data analysis in the above embodiments. The device specifically includes a communication unit and a processing unit, and may also include a storage unit.

[0187] For example, the communication unit is used to collect multi-source learning behavior data of students, resting physiological signals, and actual teaching progress data of each class. The communication unit is the hardware interface component that allows the device to interact with external data sources; it is responsible not only for data input but also for data output.

[0188] The processing unit is used to construct individual knowledge graphs, which include knowledge element layers, skill layers, and literacy layers; construct a physiological signal database, including at least electrocardiogram (ECG) and electroencephalogram (EEG) signals at rest; construct a group knowledge graph federation using the individual knowledge graphs as federated nodes and the physiological signal database; construct a teaching progress timeline based on knowledge points, and calculate the progress offset between the student's current learning progress and the class's teaching progress; comprehensively evaluate the personalized knowledge graphs, the group knowledge graph federation, and the progress offset to obtain a multi-dimensional comprehensive evaluation report; and generate personalized learning plans based on the multi-dimensional comprehensive evaluation report using a multi-objective optimization algorithm.

[0189] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely exemplary illustrations of this application as defined by the appended claims, and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from the spirit and scope of this application. Thus, if such modifications and variations of this application fall within the scope of the claims of this application and their equivalents, this application is also intended to include such modifications and variations.

Claims

1. A learning status assessment method based on big data analysis, characterized in that, include: Collect students’ multi-source learning behavior data to construct individual knowledge graphs, which include knowledge element layers, skill layers, and literacy layers. Collect students' resting physiological signals and construct a physiological signal database. The resting physiological signals include at least electrocardiogram (ECG) signals and electroencephalogram (EEG) signals. Using the individual knowledge graph as federation nodes and combining it with the physiological signal database, a group knowledge graph federation is constructed. Collect actual teaching progress data for each class, construct a teaching progress timeline based on knowledge points, and calculate the progress offset between the student's current learning progress and the class's teaching progress. A comprehensive evaluation of the personalized knowledge graph, the federated group knowledge graph, and the progress offset is conducted to obtain a multi-dimensional comprehensive evaluation report. Based on the aforementioned multi-dimensional comprehensive evaluation report, a multi-objective optimization algorithm is used to generate a personalized learning plan and push it to the student.

2. The learning status evaluation method based on big data analysis according to claim 1, characterized in that, The process of constructing an individual knowledge graph specifically includes: Collect students' multi-source learning behavior data and extract features to generate a multi-dimensional feature space, which includes knowledge mastery features, learning habit features, knowledge application ability features, thinking transfer features, and learning style features. A knowledge meta-layer is constructed using the knowledge mastery feature as nodes. The knowledge meta-layer consists of nodes representing knowledge points and edges representing semantic relationships between knowledge points. The knowledge element layer is input into the first graph neural network. The nodes of the knowledge element layer are aggregated to obtain the skill layer, which is composed of nodes representing core skills, and the knowledge application capability feature and the thinking transfer feature are used as constraints. The skill layer is input into the second graph neural network. The nodes of the skill layer are aggregated under the constraints of the learning habit features and the learning style features to obtain the literacy layer, which is composed of nodes representing comprehensive literacy. Establish hierarchical connections between the knowledge element layer, skill layer, and literacy layer to generate a complete individual knowledge graph.

3. The learning status evaluation method based on big data analysis according to claim 1, characterized in that, The process of constructing the federated knowledge graph specifically includes: The knowledge element layer subgraph, skill layer subgraph, and resting physiological feature vector of the student are extracted from the individual knowledge graph and physiological signal database, and input into the cross-modal joint embedding model to obtain the cross-modal joint embedding vector. The cross-modal joint embedding vector is used to characterize the intrinsic correlation pattern between the student and the knowledge element layer, skill layer and physiological state. Using each student as a federation node, the cross-modal joint embedding vector is federated to obtain a group federation center vector. The group federation center vector is used to characterize the common association patterns among students of the same grade in the region between the knowledge meta-layer, the skill layer and the physiological state. Based on the group federated central vector and cross-modal joint embedding vector, the cross-evaluation index of students in the group is calculated; Collect students' historical cross-modal joint embedding vectors to form a temporal embedding vector sequence. Compare and analyze the temporal embedding vector sequence with the group's federated center vector to calculate the longitudinal development trajectory index. By integrating the horizontal evaluation indicators and the vertical development trajectory indicators, a group knowledge graph federation is generated.

4. The learning status evaluation method based on big data analysis according to claim 3, characterized in that, The cross-modal joint embedding model includes a graph coding layer, a physiological coding layer, a bilinear attention layer, and a joint embedding layer: The graph coding layer includes a first graph convolutional network and a second graph convolutional network. The first graph convolutional network takes the knowledge element layer subgraph as input to obtain the knowledge element layer graph embedding vector, and the second graph convolutional network takes the skill layer subgraph as input to obtain the skill layer graph embedding vector. The physiological coding layer takes the resting physiological feature vector as input, performs nonlinear transformation through a multilayer perceptron, and outputs a physiological feature embedding vector. The bilinear attention layer calculates the first attention weight between the knowledge element layer graph embedding vector and the physiological feature embedding vector, and the second attention weight between the skill layer graph embedding vector and the physiological feature embedding vector, respectively. The joint embedding layer performs weighted fusion of the knowledge element layer graph embedding vector and the physiological feature embedding vector based on the first attention weight, and performs weighted fusion of the skill layer graph embedding vector and the physiological feature embedding vector based on the second attention weight. The results of the two weighted fusions are then concatenated and mapped through a fully connected layer to output the cross-modal joint embedding vector.

5. The learning status evaluation method based on big data analysis according to claim 1, characterized in that, The process of calculating the deviation between the student's current learning progress and the class's teaching progress specifically includes: Obtain the actual teaching progress sequence of the target class, wherein the actual teaching progress sequence is arranged in chronological order with knowledge points as the granularity. The learning progress trajectory is extracted from the individual knowledge graph, and the learning progress trajectory and the actual teaching progress sequence are input into the progress decoupling comparison network, which includes a temporal coding layer and a dynamic alignment layer. The temporal coding layer extracts temporal features from the actual teaching progress sequence and the learning progress trajectory to obtain the class teaching progress temporal vector and the student individual progress temporal vector. The dynamic alignment layer uses a differential dynamic time warping algorithm to calculate the minimum cumulative distance between the class teaching progress time sequence vector and the student individual progress time sequence vector, and uses the warped path corresponding to the minimum cumulative distance as the progress offset representation path. Based on the progress offset representation path, the progress advance, progress lag, and knowledge point mastery drift of the target student relative to the class teaching progress in different time slices are analyzed to generate the progress offset.

6. The learning status evaluation method based on big data analysis according to claim 3, characterized in that, The process of comprehensively evaluating the personalized knowledge graph, the federated group knowledge graph, and the progress offset to obtain a multi-dimensional comprehensive evaluation report specifically includes: Based on the time sequence, the knowledge element layer nodes and skill layer nodes in the individual knowledge graph are granularly aligned with the common association patterns of the group knowledge graph federation to generate joint knowledge data of individuals and groups. The joint knowledge data of the individual groups is input into the cross-attention evaluation network, which includes a horizontal comparison module and a vertical tracking module. The horizontal comparison module uses the common association pattern of the group as a reference to calculate the horizontal deviation of each node in the individual knowledge graph; The longitudinal tracking module performs a time-series comparison between the cross-modal joint embedding vector and the group federation center vector, and calculates the longitudinal development trajectory deviation by combining the progress offset characterization path. The lateral deviation and the longitudinal development trajectory deviation are input into the joint decoding layer, and a multi-dimensional comprehensive evaluation report is generated by using feature fusion and nonlinear mapping of the multilayer perceptron.

7. The learning status assessment method based on big data analysis according to claim 1, characterized in that, The process of generating a personalized learning plan based on the multi-dimensional comprehensive evaluation report and pushing it to the student using a multi-objective optimization algorithm specifically includes: Subject ability indicators, learning effort indicators, learning progress indicators, and learning efficiency indicators are extracted from the multi-dimensional comprehensive evaluation report to construct decision variables and constraints for a multi-objective optimization function. The subject ability indicators, learning effort indicators, learning progress indicators and learning efficiency indicators are input into a preset multi-objective optimization model. The Pareto optimal solution set is obtained by using a non-dominated sorting genetic algorithm with the four objective functions of maximizing subject balance, effort matching, progress synchronization and efficiency maximization as the optimization objectives. The Pareto optimal solution set contains multiple alternative daily subject learning time allocation schemes. Based on students' learning habits and historical behavior data, the optimal learning plan is selected from the Pareto optimal solution set and then pushed to the student's terminal.

8. A learning status assessment system based on big data analysis, characterized in that, The learning state assessment system, applied to the learning state assessment method based on big data analysis as described in any one of claims 1-7, specifically includes: a communication unit and a processing unit. The individual knowledge graph construction module is used to collect students' multi-source learning behavior data and construct an individual knowledge graph, which includes a knowledge element layer, a skill layer, and a literacy layer. A physiological signal library construction module is used to collect students' resting physiological signals to construct a physiological signal library, wherein the resting physiological signals include at least electrocardiogram (ECG) signals and electroencephalogram (EEG) signals. The group knowledge graph federation construction module is used to construct a group knowledge graph federation by using the individual knowledge graph as federation nodes and combining it with the physiological signal database. The offset calculation module is used to collect the actual teaching progress data of each class, construct a teaching progress timeline with knowledge points as the unit, and calculate the progress offset between the student's current learning progress and the class's teaching progress. The comprehensive evaluation module is used to comprehensively evaluate the personalized knowledge graph, the federated group knowledge graph, and the progress offset to obtain a multi-dimensional comprehensive evaluation report. The learning plan generation module is used to generate personalized learning plans based on the multi-dimensional comprehensive evaluation report, using a multi-objective optimization algorithm, and then push them to students.

9. The learning status evaluation system based on big data analysis according to claim 8, characterized in that, The federated construction module for the group knowledge graph is specifically used for: The knowledge element layer subgraph, skill layer subgraph, and resting physiological feature vector of the student are extracted from the individual knowledge graph and physiological signal database, and input into the cross-modal joint embedding model to obtain the cross-modal joint embedding vector. The cross-modal joint embedding vector is used to characterize the intrinsic correlation pattern between the student and the knowledge element layer, skill layer and physiological state. Using each student as a federation node, the cross-modal joint embedding vector is federated to obtain the group federation center vector; Based on the group federated central vector and cross-modal joint embedding vector, the cross-evaluation index of students in the group is calculated; Collect students' historical cross-modal joint embedding vectors to form a temporal embedding vector sequence. Compare and analyze the temporal embedding vector sequence with the group's federated center vector to calculate the longitudinal development trajectory index. By integrating the horizontal evaluation indicators and the vertical development trajectory indicators, a group knowledge graph federation is generated.

10. A learning status assessment device based on big data analysis, characterized in that, The learning state assessment device, which is applied to the learning state assessment method based on big data analysis according to any one of claims 1-7, specifically includes: a communication unit and a processing unit; The communication unit is used to collect students' multi-source learning behavior data, resting physiological signals, and actual teaching progress data of each class; The processing unit is used to construct an individual knowledge graph, which includes a knowledge element layer, a skill layer, and a literacy layer. Construct a physiological signal library, wherein the resting physiological signals include at least electrocardiogram (ECG) signals and electroencephalogram (EEG) signals; Using the individual knowledge graph as federation nodes and combining it with the physiological signal database, a group knowledge graph federation is constructed. A teaching progress timeline is constructed based on knowledge points, and the progress offset between the student's current learning progress and the class's teaching progress is calculated. A comprehensive evaluation of the personalized knowledge graph, the federated group knowledge graph, and the progress offset is conducted to obtain a multi-dimensional comprehensive evaluation report. Based on the aforementioned multi-dimensional comprehensive evaluation report, a multi-objective optimization algorithm is used to generate a personalized learning plan, which is then pushed to the student.

Images