Big data-based intelligent education guidance evaluation system and method

Abstract

The application discloses a big data-based intelligent education guidance evaluation system and method, belongs to the technical field of intelligent education and big data analysis, and comprises a multi-source heterogeneous education data acquisition module, a data preprocessing and feature fusion module, a multi-dimensional capacity evaluation module, a learning behavior sequence analysis module, an individualized guidance recommendation module and a multi-role visual display module. The method integrates multi-source data through a time sequence self-adaptive alignment algorithm, realizes five-dimensional comprehensive capacity evaluation by adopting a subjective and objective fusion weighting algorithm and a double-flow attention knowledge tracking algorithm, performs learning behavior time sequence modeling by using a key behavior node identification algorithm, and generates an individualized learning path by a cognitive load perception path optimization algorithm. The application solves the problems of single evaluation dimension, fragmented data acquisition and lack of data support for individualized guidance in the existing education evaluation system.

Description

Technical Field

[0001] This invention relates to the fields of smart education and big data analysis technology, and in particular to a smart education guidance and evaluation system and method based on big data. Background Technology

[0002] With the deepening of educational informatization, various types of educational data are experiencing explosive growth. IoT sensors in smart classrooms continuously collect student attendance records, classroom interaction behaviors, and group collaboration performance; online learning platforms accumulate massive amounts of video viewing logs, chapter test results, and forum discussion content; and student management systems store students' basic learning profiles and historical academic data. This diverse and heterogeneous educational data contains rich information on learning patterns and development potential. How to effectively mine and utilize this data to support precise educational evaluation and personalized learning guidance has become a key issue that urgently needs to be addressed in the field of smart education.

[0003] In the prior art, Chinese invention patent CN113095969A discloses an immersive flipped classroom teaching system based on multiple virtual avatars and its working method. This system includes an immersive teaching environment generation module, an immersive flipped classroom teaching module based on multiple virtual avatars, and a teaching service module. In the teaching service module, the solution uses a testing submodule to detect learners' mastery of knowledge, introduces an AI testing assistant to provide step-by-step guidance, and pushes relevant resources to learners based on a hybrid collaborative intelligent recommendation algorithm. However, this existing technology has the following shortcomings: First, the solution is mainly geared towards VR immersive teaching scenarios, and its data collection and integration capabilities for conventional smart classrooms and online learning platforms are limited, making it difficult to construct a complete learner profile covering the entire process before, during, and after class. Second, the learning analysis of this solution mainly relies on simple statistics of test results and process data, lacking the ability to deeply model the sequence of learning behaviors and failing to identify key behavioral nodes that affect learning outcomes. Third, the resource recommendation of this solution adopts a hybrid collaborative recommendation based on user and project recommendation algorithms, without considering the learner's cognitive load, which may lead to a mismatch between recommended content and the learner's current learning capacity. Fourth, the evaluation dimensions of this solution are relatively singular, mainly focusing on knowledge mastery and operational proficiency, lacking a multi-dimensional assessment of comprehensive qualities such as thinking ability development, collaborative communication skills, and self-learning ability.

[0004] Furthermore, traditional educational evaluation methods typically rely on periodic exam scores as the primary evaluation criterion. This outcome-oriented approach suffers from significant lag, failing to promptly reflect problems and progress in the learning process. While some intelligent education systems incorporate learning analytics, they often employ simple statistical indicators and rule-matching methods, struggling to capture the complex temporal characteristics and individual differences in learning behavior. Regarding personalized learning guidance, existing recommendation systems primarily focus on learners' knowledge gaps, with less consideration for overall learning path optimization and dynamic adjustment of cognitive load, sometimes resulting in recommended content exceeding the learner's current cognitive capacity. In terms of presenting evaluation results, existing systems mostly display statistical indicators directly in the form of data reports, lacking differentiated presentation and natural language interpretation for different user roles, increasing the difficulty for non-specialist users to understand the evaluation results. Summary of the Invention

[0005] To address the aforementioned shortcomings of existing technologies, this invention provides a smart education guidance and evaluation system and method based on big data, aiming to solve technical problems in the existing education evaluation system such as single evaluation dimensions, fragmented data collection, evaluation results lagging behind the teaching process, and lack of data support for personalized guidance.

[0006] The technical solution of the present invention includes: a smart education guidance and evaluation system based on big data, comprising a multi-source heterogeneous education data acquisition module, a data preprocessing and feature fusion module, a multi-dimensional ability assessment module, a learning behavior sequence analysis module, a personalized guidance and recommendation module, and a multi-role visualization display module.

[0007] The multi-source heterogeneous education data acquisition module is used to acquire classroom attendance data, interactive response data, and group discussion participation data through IoT terminals deployed in smart classrooms. It also connects to online learning platforms to obtain video viewing progress, chapter test scores, homework submission time distribution, and forum post content. Furthermore, it integrates basic learning profiles from the student registration management system to generate a raw education dataset covering the entire process of pre-class preparation, in-class participation, and post-class consolidation.

[0008] The data preprocessing and feature fusion module receives the original educational dataset, uses a time-adaptive alignment algorithm to unify asynchronous data streams from different sources to a standard time axis, uses the isolated forest algorithm to identify and label outlier data points, fills missing values ​​through multiple imputation methods, and introduces a pre-trained language model to extract semantic feature vectors from text data, outputting standardized feature vectors.

[0009] The multi-dimensional ability assessment module receives standardized feature vectors and constructs an ability assessment framework based on five primary indicators: knowledge mastery level, learning engagement, thinking ability development, collaborative communication ability, and self-learning ability. It uses a subjective and objective weighting algorithm to determine the weight of each indicator and updates the learner's mastery probability estimate of each knowledge point in real time through a dual-stream attention knowledge tracking algorithm, outputting an ability assessment result vector and a knowledge mastery probability distribution.

[0010] The learning behavior sequence analysis module receives semantic feature vectors, uses sequence pattern mining algorithms to discover high-frequency behavior patterns from the learning trajectory, uses association rules to extract the potential correlation between effective learning strategies and academic performance, and uses key behavior node identification algorithms to perform time-series modeling of the learning behavior sequence, automatically identifies key behavior nodes that affect learning effectiveness, and outputs key behavior node sequences and behavior pattern features.

[0011] The personalized guidance recommendation module receives the ability assessment result vector and key behavior node sequence. It comprehensively considers the learner's current knowledge status, historical learning preferences, and goal achievement gap. It uses a hybrid strategy of collaborative filtering and content recommendation to match learning materials from the resource library. The recommendation strategy is continuously adjusted through a cognitive load-aware path optimization algorithm. The learning path optimization signal is fed back to the multi-dimensional ability assessment module to update the assessment parameters and output a list of recommended learning resources.

[0012] The multi-role visualization module receives knowledge mastery probability distribution, behavioral pattern characteristics, and recommended learning resource lists. It generates differentiated data dashboards for four roles: teachers, students, parents, and administrators, and uses natural language generation technology to transform statistical indicators into natural language evaluation reports.

[0013] This invention also provides a smart education guidance and evaluation method based on big data, including a multi-source heterogeneous education data collection step, a data preprocessing and feature fusion step, a multi-dimensional ability assessment step, a learning behavior sequence analysis step, a personalized guidance and recommendation step, and a multi-role visualization display step.

[0014] The beneficial effects of this invention include: by constructing a multi-source heterogeneous educational data acquisition module, automated collection and integration of learning data throughout the entire learning process—before, during, and after class—is achieved, overcoming the limitations of fragmented data collection in existing technologies; by designing a subjective-objective fusion weighting algorithm and a dual-stream attention knowledge tracking algorithm, comprehensive evaluation of five-dimensional ability indicators and real-time tracking of knowledge mastery status are realized, overcoming the problems of single-dimensionality and lag in traditional evaluation methods; by introducing a key behavior node identification algorithm, deep temporal modeling of learning behavior sequences is achieved, enabling the identification of key behavior nodes affecting learning outcomes and providing interpretable analytical basis; by designing a cognitive load perception path optimization algorithm, dynamic matching of learning resource recommendations and learners' cognitive capacity is achieved, improving the accuracy and effectiveness of personalized guidance; and by constructing a multi-role visualization module, differentiated data presentation and natural language interpretation for different user roles are achieved, lowering the understanding threshold of evaluation results. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the overall architecture of the system of the present invention.

[0016] Figure 2 This is a flowchart illustrating the method of the present invention.

[0017] Figure 3 This is a schematic diagram of the data preprocessing and feature fusion module.

[0018] Figure 4 This is a structural diagram of the multi-dimensional capability assessment module.

[0019] Figure 5 This is a structural diagram of the learning behavior sequence analysis module.

[0020] Figure 6 This is a structural diagram of the personalized guidance and recommendation module. Detailed Implementation

[0021] Please refer to the attached document. Figures 1-6 The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that the following embodiments are only used to illustrate the present invention and not to limit the scope of protection of the present invention.

[0022] like Figure 1 As shown, the smart education guidance and evaluation system based on big data provided by this invention includes a multi-source heterogeneous education data acquisition module 1, a data preprocessing and feature fusion module 2, a multi-dimensional ability assessment module 3, a learning behavior sequence analysis module 4, a personalized guidance and recommendation module 5, and a multi-role visualization display module 6. Each module is described in detail below.

[0023] The multi-source heterogeneous educational data acquisition module 1 corresponds to the multi-source heterogeneous educational data acquisition steps in the embodiments. This module includes a smart classroom data acquisition unit, an online learning platform docking unit, and a student registration management system docking unit. The smart classroom data acquisition unit establishes a connection with the terminal devices deployed in the smart classroom through the Internet of Things communication protocol, and receives classroom attendance data, interactive response data, and group discussion participation data in real time. The online learning platform docking unit exchanges data with mainstream online learning platforms through RESTful API, and periodically pulls video viewing progress, chapter test scores, homework submission time distribution, and forum post content. The student registration management system docking unit accesses the data tables of the student registration management system through the database connection protocol, and synchronizes the learners' basic learning files and historical grade records. The data collected by the three units are aggregated to form the raw educational dataset, and then sent to the data preprocessing and feature fusion module 2.

[0024] The data preprocessing and feature fusion module 2 corresponds to the data preprocessing and feature fusion steps in the implementation examples. This module includes a timestamp alignment unit, an anomaly detection unit, a missing value imputation unit, and a semantic feature extraction unit. The timestamp alignment unit implements the function of a time-adaptive alignment algorithm, unifying asynchronous data streams from different sources onto a standard timeline. The anomaly detection unit implements the function of the isolated forest algorithm, identifying and marking anomalous data points. The missing value imputation unit implements the function of a multiple imputation method, filling in missing values ​​caused by equipment failure or network interruption. The semantic feature extraction unit deploys a pre-trained language model to extract semantic feature vectors from text-based data. The standardized feature vectors output by this module are sent to the multi-dimensional capability assessment module 3 and the learning behavior sequence analysis module 4, respectively.

[0025] The functions and methods of the multi-dimensional ability assessment module 3 correspond to the multi-dimensional ability assessment steps in the embodiments. This module includes an ability assessment framework unit, a subjective and objective weighting unit, and a dual-stream attention knowledge tracking unit. The ability assessment framework unit stores the configuration information of the five-dimensional ability assessment index system. The subjective and objective weighting unit implements the weight calculation function combining the analytic hierarchy process (AHP) and the entropy weighting method. The dual-stream attention knowledge tracking unit deploys a graph attention network and a bidirectional Transformer to achieve real-time estimation of the probability of knowledge mastery. The ability assessment result vector output by this module is sent to the personalized guidance and recommendation module 5, and the knowledge mastery probability distribution is sent to the multi-role visualization module 6.

[0026] The functions of the learning behavior sequence analysis module 4 correspond to the learning behavior sequence analysis steps in the method embodiment. This module includes a behavior pattern mining unit, an association rule extraction unit, and a key behavior node identification unit. The behavior pattern mining unit implements the PrefixSpan sequence pattern mining algorithm. The association rule extraction unit implements the FP-Growth association rule mining algorithm. The key behavior node identification unit deploys an attention-weighted bidirectional LSTM network to achieve automatic identification of key behavior nodes. The key behavior node sequence output by this module is sent to the personalized guidance recommendation module 5, and the behavior pattern features are sent to the multi-role visualization display module 6.

[0027] The personalized guidance recommendation module 5 corresponds to the personalized guidance recommendation steps in the method embodiment. This module includes a hybrid recommendation strategy unit and a cognitive load-aware path optimization unit. The hybrid recommendation strategy unit implements a fusion recommendation function combining collaborative filtering and content recommendation. The cognitive load-aware path optimization unit deploys a deep Q-network to achieve dynamic optimization of the learning path. This module also includes a feedback channel for sending the learning path optimization signal to the multi-dimensional ability assessment module 3 to update the assessment parameters. The recommended learning resource list output by this module is sent to the multi-role visualization module 6.

[0028] The multi-role visualization module 6 corresponds to the multi-role visualization steps in the implementation examples. This module includes a differentiated data dashboard generation unit and a natural language report generation unit. The differentiated data dashboard generation unit dynamically loads different visualization templates based on the user's role, generating a class learning heatmap for teachers, an individual ability radar chart for students, a phased growth report for parents, and a regional education quality benchmarking analysis for administrators. The natural language report generation unit uses a combination of template filling and neural network generation to transform statistical indicators into easily understandable text descriptions.

[0029] This invention employs a microservice architecture, with loosely coupled functional modules that support elastic scaling to handle concurrent access pressure during peak teaching periods. The system ensures learner privacy and security through data anonymization and tiered access control, forming a closed-loop intelligent education evaluation system driven by data, providing intelligent analysis, precise guidance, and continuous optimization.

[0030] like Figure 2 As shown, the smart education guidance and evaluation method based on big data provided by this invention includes steps such as multi-source heterogeneous education data collection, data preprocessing and feature fusion, multi-dimensional ability assessment, learning behavior sequence analysis, personalized guidance and recommendation, and multi-role visualization. Each step is described in detail below.

[0031] I. Multi-source heterogeneous educational data acquisition steps: The core objective of multi-source heterogeneous educational data acquisition steps is to construct a data acquisition system covering the entire learning process, providing a comprehensive data foundation for subsequent analysis and evaluation. In one embodiment of this invention, this step achieves synchronous acquisition of multi-source data through three parallel data acquisition channels.

[0032] In the smart classroom data acquisition channel, the system collects classroom behavior data through IoT terminal devices deployed in the classroom. Specifically, facial recognition cameras collect student attendance information, recording each student's arrival and departure times, with a sampling frequency set to once per second to ensure the accuracy of attendance times. The interactive response system collects the frequency of students raising their hands, response time, and accuracy rate during class question-and-answer sessions, with response time recording accuracy down to the millisecond level. The group collaboration monitoring module uses audio sensors and visual analysis algorithms to collect data on students' speaking time percentage, interaction rounds, and body language activity during group discussions, with the statistical error for speaking time controlled within 500 milliseconds. The collected classroom behavior data is stored as a time-series sequence in JSON format, with each record containing four fields: student identifier, timestamp, behavior type, and behavior parameters.

[0033] In the online learning platform data collection channel, the system connects to mainstream online learning platforms through standardized application programming interfaces (APIs) to obtain learners' digital learning trajectories. The video learning module collects video viewing progress, pause counts, replay segments, and playback speed adjustment records. Viewing progress is recorded as a percentage with an accuracy of 1%. The chapter test module collects the answer time, submission time, score, and number of times incorrect questions were retaken for each question, with answer time accurate to the second. The homework submission module collects the submission time, number of modifications, and final score for each assignment. The submission time distribution is used to analyze learners' time management habits. The forum interaction module collects post content, reply relationships, and the number of likes. Post content is preserved in its original text for subsequent semantic analysis. All of the above learning trajectory data is organized in the form of event logs. Each log entry contains four core fields: learner identifier, event type, event timestamp, and event details.

[0034] Within the student registration management system interface, the system integrates learners' basic academic records. The basic information module retrieves the learner's grade, class, subject selection, and learning goals. The historical grades module retrieves the learner's midterm and final exam scores for each subject and semester, stored in a standardized percentage format. The comprehensive quality evaluation module retrieves the learner's evaluation levels in moral education, physical education, aesthetic education, and labor education. All of the above student registration data is stored in structured tables, organized by student dimension, supporting cross-semester vertical data queries.

[0035] After completing data collection from the three channels, the system aggregates the collected multi-source data to form a raw educational dataset. This dataset covers online learning behavior during the pre-class preparation stage, classroom interaction during the in-class participation stage, and homework completion and review during the post-class consolidation stage, forming a raw data pool of complete digital profiles of learners. In this embodiment, for a class with 40 students, the system collects approximately 2GB of raw data in one semester, of which classroom behavior data accounts for approximately 15%, online learning trajectory data accounts for approximately 70%, and student record data accounts for approximately 15%.

[0036] II. Data Preprocessing and Feature Fusion Steps: The core objective of these steps is to transform multi-source, heterogeneous raw data into a unified feature representation format, providing high-quality input data for subsequent intelligent analysis. For example... Figure 3 As shown, this step includes four sub-processes: timestamp alignment, anomaly detection, missing value imputation, and semantic feature extraction.

[0037] During timestamp alignment, the system employs a time-adaptive alignment algorithm to unify asynchronous data streams from different data sources onto a standard timeline. Because the data acquisition frequencies and clock precisions of the smart classroom terminal, online learning platform, and student registration management system differ, direct data fusion would lead to temporal discrepancies. The time-adaptive alignment algorithm designed in this invention uses a dynamic time window mechanism, adaptively adjusting the width of the alignment window based on the sampling characteristics of each data source.

[0038] The calculation process of the time-adaptive alignment algorithm can be described by the following formula. Let the... The original timestamp sequence of each data source is The sampling point sequence of the standard time axis is The sampling interval is For sampling points on the standard time axis The algorithm first calculates the width of the dynamic time window:

[0039] ,

[0040] in, For the first The time window width for each sampling point For the first The average sampling frequency of each data source For the first The standard deviation of the timestamp interval of each data source The frequency adaptive coefficient has a value range of 0.8 to 1.2. The fluctuation tolerance coefficient ranges from 0.5 to 1.5. In this embodiment, Set to 1.0, Set to 1.0, Set to 1 second.

[0041] After determining the time window width, the algorithm uses a weighted average to calculate the aligned feature values:

[0042] ,

[0043] in, For the standard timeline Alignment feature values ​​of each sampling point For the first The first data source Feature values ​​of each original sampling point The distance weight is calculated using the following formula: This weighting function ensures that the raw data points whose timestamps are closer to the standard sampling point contribute more to the alignment result.

[0044] In embodiments of the present invention, the width of the time window is set to range from 100 milliseconds to 500 milliseconds. When the data source sampling frequency is high, such as the sampling frequency of a classroom interactive response system reaching 10Hz, the time window width is automatically adjusted to a smaller value to preserve fine-grained temporal characteristics; when the data source sampling frequency is low, such as the video viewing progress of an online learning platform updating every 5 seconds, the time window width is automatically adjusted to a larger value to ensure smooth data transition.

[0045] In the anomaly detection subprocess, the system employs the Isolation Forest algorithm to identify and label anomalous data points in the original data. Isolation Forest is an unsupervised anomaly detection method based on random partitioning; its core idea is that anomalous data points are more easily isolated through random partitioning. In this embodiment, the number of trees in the Isolation Forest is set to 100, and the number of subsamples is set to 256. The algorithm calculates anomaly scores for each data point, with scores ranging from 0 to 1; a higher score indicates a greater likelihood of the data point being an anomaly. The system sets an anomaly score threshold between 0.6 and 0.8; data points with anomaly scores greater than this threshold are marked as anomalous. Data points marked as anomalous are not directly deleted but are either downweighted or replaced with imputed values ​​in subsequent analysis.

[0046] In the missing value imputation process, the system employs a multiple imputation method to fill in missing values ​​caused by equipment failure or network interruption. The multiple imputation method first establishes an imputation model based on the missing data pattern and the correlation between variables, then generates multiple sets of possible imputed values, and finally combines the multiple imputed results to provide the final imputed value and uncertainty estimate. In this embodiment, the number of imputation iterations is set to 5, with each iteration generating one set of imputed values. For numerical variables, a chain equation-based multiple imputation method is used; for categorical variables, a logistic regression imputation method is used. When the missing rate exceeds 30% for a certain time period, the data for that time period will be marked as an unreliable interval and given lower weight in subsequent evaluations.

[0047] In the semantic feature extraction subprocess, the system introduces a pre-trained language model to encode textual data. Learners' posts in forum discussions, textual answers in assignments, and transcribed text from group discussions are all unstructured textual data that needs to be converted into a computable numerical representation. In this embodiment, the system uses a Chinese pre-trained language model with 110M parameters for text encoding, transforming each text segment into a 768-dimensional semantic feature vector. During the encoding process, the system first performs word segmentation and sequence truncation on the original text, setting the maximum sequence length to 512 tokens; then, it inputs the token sequence into the pre-trained model, taking the pooling result of the last hidden state as the semantic feature vector of the text.

[0048] After the above four sub-processes, the original educational dataset is transformed into a standardized feature vector. The standardized feature vector uses... Stored in matrix form, where For the number of learners, The feature dimensions are defined as follows. In this embodiment, the numerical feature dimensions are 48, including 6 dimensions for attendance features, 12 dimensions for classroom interaction features, 8 dimensions for video learning features, 10 dimensions for test performance features, 6 dimensions for homework submission features, and 6 dimensions for forum interaction features; the semantic feature dimensions are 768. After concatenating the numerical and semantic feature vectors, the total feature dimensions reach 816.

[0049] Third, the multi-dimensional ability assessment steps. The core objective of the multi-dimensional ability assessment steps is to construct a comprehensive learner ability profile, realizing the transformation from single-knowledge mastery evaluation to comprehensive quality assessment. For example... Figure 4 As shown, this step includes three sub-processes: constructing a capability assessment framework, empowering through a combination of subjective and objective assessments, and knowledge tracking.

[0050] In the process of constructing the competency assessment framework, the system builds a multi-dimensional competency assessment framework based on five primary indicators. The knowledge mastery level indicator assesses learners' understanding and memorization of course knowledge points, with observation data derived from chapter test scores, homework scores, and final exam results. The learning engagement level indicator assesses learners' time and effort invested in learning activities, with observation data derived from video viewing time, platform login frequency, and timely homework submission. The thinking ability development indicator assesses learners' analytical reasoning and innovative thinking abilities, with observation data derived from performance on open-ended questions, in-depth analysis of forum discussions, and the degree of innovation in project assignments. The collaboration and communication ability indicator assesses learners' ability to collaborate with others to complete learning tasks, with observation data derived from group discussion participation, forum interaction activity, and the quality of peer evaluation. The self-directed learning ability indicator assesses learners' ability to independently plan and manage their learning process, with observation data derived from pre-study completion rate, learning plan execution rate, and proactive help-seeking behavior. Each primary indicator has 3 to 5 secondary observation points, forming a hierarchical assessment indicator system.

[0051] In the subjective-objective fusion weighting process, the system uses the subjective-objective fusion weighting algorithm designed in this invention to determine the comprehensive weight of each indicator. This algorithm includes two stages: subjective weight calculation and objective weight calculation, achieving the organic fusion of subjective and objective weights through weighted averaging.

[0052] In the subjective weight calculation stage, the analytic hierarchy process (AHP) combined with expert experience is used to determine the initial weight vector. The system invites senior experts in the education field to construct pairwise comparison judgment matrices between each ability indicator, with matrix elements using a 1-9 scale to represent relative importance. After obtaining the judgment matrices, the subjective weight vector is calculated using eigenvalue decomposition. The consistency check is performed to ensure the logical rationality of the judgment. A consistency ratio threshold is set to 0.1; when the consistency ratio exceeds this threshold, the judgment matrix needs to be readjusted. In this embodiment, the subjective weight vector given by the expert is... These correspond to five indicators: knowledge mastery level, learning engagement level, thinking ability development, collaborative communication ability, and self-learning ability.

[0053] In the objective weight calculation stage, the entropy weight method is used to determine the information entropy weight vector based on the dispersion of each indicator's data. The core idea of ​​the entropy weight method is that indicators with lower information entropy contain more effective information and should be assigned higher weights. Let the first... Each indicator in The sequence of observations on each learner is as follows: First, normalization is performed to obtain Then calculate the information entropy of the indicator:

[0054] ,

[0055] in, For the first Information entropy of each indicator For the number of learners, For the first The learner in the first Normalized observations on each indicator. When When, define After obtaining the information entropy of each indicator, the formula for calculating the objective weight vector is:

[0056] ,

[0057] in, For the first The objective weight of each indicator, For the first Information entropy of each indicator.

[0058] After obtaining the subjective weight vector and the objective weight vector, the system calculates the final weight vector by weighted average:

[0059] ,

[0060] in, For the first The final weight of each indicator, The weighting coefficient for subjective weights ranges from 0.4 to 0.6. In this embodiment, Setting it to 0.5 achieves a balanced integration of subjective and objective weights. By dynamically collecting observation data for each indicator and recalculating the objective weights, the system can adaptively adjust the weight configuration based on the actual distribution characteristics of the data.

[0061] In the knowledge tracking subprocess, the system employs the dual-stream attention knowledge tracking algorithm designed in this invention to update the learner's mastery probability estimate for each knowledge point in real time. This algorithm integrates two parallel information processing channels: concept relationship graph attention stream and temporal evolution attention stream. It can capture the structured relationships between knowledge points and model the dynamic evolution of the learner's knowledge state.

[0062] The core of concept relationship graph attention flow is to use graph attention networks to model the structured relationships between knowledge points. Let the set of knowledge points be... The prerequisite and similarity relationships between knowledge points constitute a knowledge graph. ,in Let it be a set of edges. (Regarding the knowledge point...) Its representation update formula in the graph attention network is:

[0063] ,

[0064] in, For knowledge points In the Hidden representation of layers, For knowledge points The set of neighboring nodes, For the first The learnable weight matrix of the layer, LeakyReLU is used as the activation function. The formula for calculating the attention coefficient is:

[0065] ,

[0066] in, For learnable attention vectors, This represents a vector concatenation operation. In this embodiment, the graph attention network has 2 layers, 128 hidden dimensions, and 4 attention heads.

[0067] The core of temporal evolutionary attention flow is the use of a bidirectional Transformer to model the dynamic evolution of the learner's knowledge state. Let the learner at time step... The learning interaction record is ,in Number the questions. The value of the response is either 0 or 1. Learning interaction sequences. Encoded as a sequence of input vectors ,in , Embed the matrix for the problem. Embed the answer results into a matrix. For position encoding.

[0068] The formula for calculating the self-attention of a bidirectional Transformer is:

[0069] ,

[0070] in, , , These are the query matrix, key matrix, and value matrix, respectively. is the dimension of the key vector. In this embodiment, the number of Transformer layers is set to 4, the hidden dimension is set to 256, and the number of attention heads is set to 8.

[0071] After fusing the outputs of the two attention streams, the system calculates the learner's mastery probability for each knowledge point. Let the knowledge point be... The graph attention representation is as follows Learners in time steps Temporal attention is represented as The formula for calculating probability is:

[0072] ,

[0073] in, For the Sigmoid function, and This is a learnable parameter. When the probability of mastering a knowledge point is greater than the mastery probability threshold, the knowledge point is considered to have been mastered. The mastery probability threshold ranges from 0.7 to 0.9, and is set to 0.8 in this embodiment.

[0074] IV. Steps for Learning Behavior Sequence Analysis

[0075] The core objective of learning behavior sequence analysis is to uncover valuable patterns and rules from learners' behavioral trajectories, providing interpretable evidence to support personalized instruction. For example... Figure 5 As shown, this step includes three sub-processes: behavior pattern mining, association rule extraction, and key behavior node identification.

[0076] In the behavior pattern mining sub-process, the system employs a sequence pattern mining algorithm to discover high-frequency behavior patterns from massive learning trajectories. Learning behavior sequences consist of a series of timestamped behavioral events, including video playback, video pause, video replay, question answering, assignment submission, forum posting, forum replying, and resource downloading. The system uses the PrefixSpan algorithm to mine frequent sequence patterns that meet a minimum support threshold of 0.05. In this embodiment, the system mined 326 frequent sequence patterns from 10,000 learning behavior sequences, including 87 positive patterns associated with high academic performance and 64 negative patterns associated with low academic performance.

[0077] In the association rule extraction sub-process, the system utilizes association rule mining algorithms to extract potential associations between effective learning strategies and academic performance. The form of the association rules is as follows: ,in The antecedent represents the learning behavior pattern. The consequent represents the academic performance level. The system uses the FP-Growth algorithm to generate candidate rules and filters high-quality rules using two metrics: confidence and lift. The confidence threshold is set to 0.6, and the lift threshold is set to 1.5. In this embodiment, the system filters out 48 high-quality association rules.

[0078] In the key behavior node identification subprocess, the system uses the key behavior node identification algorithm designed in this invention to perform temporal modeling of the learning behavior sequence. This algorithm uses an attention-weighted bidirectional long short-term memory network to encode the learning behavior sequence, and identifies behavior nodes that significantly affect the learning effect by analyzing the attention weights at each time step.

[0079] Let the input representation of the learning behavior sequence be... ,in For the first The feature vector of each behavioral event. The hidden state calculation process of bidirectional LSTM is as follows:

[0080] ,

[0081] ,

[0082] ,

[0083] in, For the forward LSTM at time step The hidden state, For backward LSTM at time step The hidden state, This involves concatenating the bidirectional hidden states. In this embodiment, the hidden dimension of the LSTM is set to 128 dimensions.

[0084] The formula for calculating attention weights is:

[0085] ,

[0086] ,

[0087] in, For the first Attention score at each time step The attention weights are normalized. As a vector of academic performance, , , and These are learnable parameters.

[0088] The system marks action nodes with attention weights greater than 1.5 times the average attention weight as key action nodes. The formula for identifying key action nodes is as follows:

[0089] ,

[0090] in, This represents the average attention weight. Behavioral events marked as key behavioral nodes will be included in the key behavioral node sequence for subsequent visualization and personalized guidance.

[0091] V. Personalized Guidance and Recommendation Steps: The core objective of these steps is to generate the optimal personalized learning path based on the learner's knowledge status and behavioral characteristics. For example... Figure 6 As shown, this step includes two sub-processes: a hybrid recommendation strategy and cognitive load-aware path optimization.

[0092] In the hybrid recommendation strategy sub-process, the system comprehensively employs both collaborative filtering and content recommendation methods to match learning materials from the resource library. Collaborative filtering recommends based on the similarity between learners. The system calculates the cosine similarity between the current learner and historical learners on their ability assessment result vectors, selects the K most similar neighbor learners, and recommends learning resources that these neighbor learners have used and highly rated. Content recommendation recommends based on the matching degree between learning resources and learners' knowledge gaps. The system compares the knowledge point tags of learning resources with the learner's knowledge mastery probability distribution, prioritizing the recommendation of learning resources covering knowledge points with low mastery probabilities. The results of the two recommendation methods are weighted and fused to obtain a preliminary recommendation list, with collaborative filtering weighted at 0.4 and content recommendation weighted at 0.6.

[0093] In the cognitive load-aware path optimization subprocess, the system uses the cognitive load-aware path optimization algorithm designed in this invention to dynamically adjust the recommendation strategy. This algorithm models the learning path planning problem as a Markov decision process, employs a deep Q-network as the policy optimizer, and continuously adjusts the recommendation strategy during the learning process to approximate the individual's optimal learning path configuration.

[0094] The state space of a deep Q-network is designed as a triplet. ,in This represents the learner's current knowledge state vector. This is an estimate of cognitive load. The target knowledge set is defined as follows. Cognitive load estimation is obtained by analyzing learners' interactive behavior data, including page dwell time, video playback frequency, question answering time, and frequency of seeking help. The formula for calculating cognitive load is:

[0095] ,

[0096] in, The time spent on the page. For the expected learning time, For the number of times the video can be replayed, Time allotted for answering the questions. The average response time, For the number of times a request for help is made, , , , For the weighting coefficients to satisfy In this embodiment, the weighting coefficients are set to 0.25, 0.25, 0.30, and 0.20, respectively.

[0097] The action space of a deep Q-network is a set of candidate learning resources. Each action corresponds to a recommended learning resource. The reward function is designed to comprehensively consider knowledge gain, changes in cognitive load, and learning duration.

[0098] ,

[0099] in, Knowledge gain is the increase in the probability of mastering knowledge after learning. This is an estimate of cognitive load. The preset cognitive load threshold ranges from 0.6 to 0.8. For study time, , , The reward weighting coefficients are set to 1.0, 0.5, and 0.1, respectively, and the preset cognitive load threshold is set to 0.7.

[0100] The Deep Q-Network employs an experience replay and target network mechanism for updates. The experience replay buffer has a capacity of 10000, and each training iteration randomly samples 64 samples at a time. The target network synchronizes its parameters with the online network every 1000 steps. The exploration rate is... - Greedy strategy, initial exploration rate set to 1.0, final exploration rate set to 0.1, decay steps set to 10000 steps.

[0101] When the estimated cognitive load exceeds a preset cognitive load threshold, the deep Q-network prioritizes recommending learning resources with lower difficulty to avoid learners experiencing frustration due to cognitive overload. The learning path optimization signal is transmitted to the multi-dimensional ability assessment module through a feedback channel to update assessment parameters and adjust subsequent ability assessment strategies, forming a closed-loop adaptive learning system.

[0102] VI. Multi-role Visualization Presentation Steps: The core objective of these steps is to present evaluation and analysis results to users with different roles in a differentiated manner, thereby lowering the barrier to data interpretation. This step includes two sub-processes: generating differentiated data dashboards and generating natural language reports.

[0103] In the process of generating differentiated data dashboards, the system designs different visualization interfaces for four roles: teachers, students, parents, and administrators. The teacher's data dashboard presents a class learning heatmap and an individual warning list. The heatmap displays the distribution of knowledge mastery across the class, with knowledge points on the horizontal axis and students on the vertical axis. The warning list lists students with a knowledge mastery probability below 0.5 or a significant decline in learning engagement. The student's data dashboard displays an individual ability radar chart and learning suggestion cards. The radar chart visually presents the relative level of five abilities, and the suggestion cards list recommended learning resources and areas for improvement. The parent's data dashboard pushes a periodic progress report, showing the child's progress trend across each ability dimension in a timeline format. The administrator's data dashboard summarizes regional education quality benchmarking analysis, supporting horizontal comparisons and vertical trend analysis between different schools and classes.

[0104] In the natural language report generation process, the system uses natural language generation technology to transform statistical indicators into easily understandable textual descriptions. Report generation combines template filling with neural network generation. First, a predefined template is used to generate the basic structure and key data statements of the report. Then, a fine-tuned language model is used to supplement explanatory notes and personalized suggestions. The generated evaluation report is kept under 500 words to ensure conciseness and readability.

Claims

1. A smart education guidance and evaluation system based on big data, characterized in that: include: The multi-source heterogeneous education data acquisition module is used to acquire classroom attendance data, interactive response data, and group discussion participation data through IoT terminals deployed in smart classrooms. It also connects to online learning platforms to obtain video viewing progress, chapter test scores, homework submission time distribution, and forum post content. Furthermore, it integrates basic learning profiles from the student management system to generate a raw education dataset covering the entire process of pre-class preparation, in-class participation, and post-class consolidation. The data preprocessing and feature fusion module is used to receive the original educational dataset, use a time-adaptive alignment algorithm to unify asynchronous data streams from different sources to a standard time axis, use the isolated forest algorithm to identify and mark outlier data points, fill missing values ​​through multiple imputation methods, and introduce a pre-trained language model to extract semantic feature vectors from text data and output standardized feature vectors. The multi-dimensional ability assessment module is used to receive the standardized feature vector, construct an ability assessment framework based on five primary indicators: knowledge mastery level, learning engagement level, thinking ability development, collaborative communication ability, and self-learning ability. It uses a subjective and objective weighting algorithm to determine the weight of each indicator, and updates the learner's mastery probability estimate of each knowledge point in real time through a dual-stream attention knowledge tracking algorithm, and outputs the ability assessment result vector and knowledge mastery probability distribution. The learning behavior sequence analysis module is used to receive the semantic feature vector, use a sequence pattern mining algorithm to discover high-frequency behavior patterns from the learning trajectory, use association rules to extract the potential correlation between effective learning strategies and academic performance, and use a key behavior node identification algorithm to perform time-series modeling of the learning behavior sequence, automatically identify key behavior nodes that affect learning effect, and output key behavior node sequence and behavior pattern features. Specifically, the key behavior node identification algorithm uses an attention-weighted bidirectional long short-term memory network to encode the learning behavior sequence, and identifies behavior nodes that have a significant impact on learning effect by calculating the attention weight at each time step. The personalized guidance recommendation module receives the ability assessment result vector and the key behavior node sequence, comprehensively considers the learner's current knowledge state, historical learning preferences, and goal achievement gap, and uses a hybrid strategy of collaborative filtering and content recommendation to match learning materials from the resource library. It continuously adjusts the recommendation strategy through a cognitive load-aware path optimization algorithm, feeds back the learning path optimization signal to the multi-dimensional ability assessment module to update the assessment parameters, and outputs a recommended learning resource list. Specifically, the cognitive load-aware path optimization algorithm uses a deep Q-network as the policy optimizer. The state space of the deep Q-network includes the learner's current knowledge state vector, the estimated cognitive load, and the target knowledge point set. The action space of the deep Q-network is the candidate learning resource set. The reward function of the deep Q-network comprehensively considers knowledge gain, cognitive load changes, and learning time. The exploration rate of the deep Q-network ranges from 0.1 to 0.

3. The multi-role visualization module receives the knowledge mastery probability distribution, the behavioral pattern characteristics, and the recommended learning resource list. It generates differentiated data dashboards for four roles: teachers, students, parents, and administrators, and uses natural language generation technology to convert statistical indicators into natural language evaluation reports.

2. The system according to claim 1, characterized in that: In the data preprocessing and feature fusion module, the temporal adaptive alignment algorithm adopts a dynamic time window mechanism. The width of the dynamic time window is adaptively adjusted according to the data sampling frequency, and the width range of the dynamic time window is 100 milliseconds to 500 milliseconds. The outlier score threshold of the isolated forest algorithm is set to 0.6 to 0.

8. When the outlier score of a data point is greater than the outlier score threshold, it is marked as an outlier data point.

3. The system according to claim 1, characterized in that: The data preprocessing and feature fusion module also includes a semantic feature extraction unit. The semantic feature extraction unit uses a pre-trained language model to encode the learners' discussion speeches and assignment texts to generate a 768-dimensional semantic feature vector. The semantic feature vector is concatenated with the numerical feature vector and then input into the learning behavior sequence analysis module.

4. The system according to claim 1, characterized in that: In the multi-dimensional capability assessment module, the subjective and objective weighting algorithm includes two stages: subjective weight calculation and objective weight calculation. The subjective weight calculation uses the analytic hierarchy process (AHP) combined with expert experience to determine the initial weight vector. The objective weight calculation uses the entropy weight method to calculate the information entropy weight vector based on the dispersion of each indicator data. The initial weight vector and the information entropy weight vector are weighted and averaged to obtain the final weight vector, wherein the weighting coefficient of the subjective weight ranges from 0.4 to 0.

6.

5. The system according to claim 1, characterized in that: In the multi-dimensional ability assessment module, the dual-stream attention knowledge tracking algorithm includes a concept graph attention stream and a temporal evolution attention stream. The concept graph attention stream uses a graph attention network to capture the structured relationships between knowledge points, while the temporal evolution attention stream uses a bidirectional Transformer to model the dynamic evolution of the learner's knowledge state. When the mastery probability of a certain knowledge point in the knowledge mastery probability distribution is greater than the mastery probability threshold, it is determined that the knowledge point has been mastered. The mastery probability threshold ranges from 0.7 to 0.

9.

6. The system according to claim 1, characterized in that: In the learning behavior sequence analysis module, a behavior node is marked as a key behavior node when its attention weight is greater than 1.5 times the average attention weight.

7. The system according to claim 1, characterized in that: The cognitive load estimate is obtained by analyzing learners' interactive behavior data, which includes page dwell time, video playback times, question answering time, and help frequency. When the cognitive load estimate exceeds a preset cognitive load threshold, the deep Q network prioritizes recommending learning resources with lower difficulty. The preset cognitive load threshold ranges from 0.6 to 0.

8.

8. The system according to claim 1, characterized in that: In the multi-role visualization module, the teacher's data dashboard presents a class learning heatmap and an individual warning list, the student's data dashboard displays a personal ability radar chart and learning suggestion cards, the parent's data dashboard pushes a phased growth report, and the administrator's data dashboard summarizes the regional education quality benchmarking analysis. The natural language assessment report uses a combination of template filling and neural network generation to convert the ability assessment results into a text description of no more than 500 words.

9. A smart education guidance and evaluation method based on big data, using the system described in any one of claims 1-8, characterized in that, include: Multi-source heterogeneous education data collection steps: acquire classroom attendance data, interactive response data, and group discussion participation data through IoT terminals deployed in smart classrooms; simultaneously connect with online learning platforms to obtain video viewing progress, chapter test scores, homework submission time distribution, and forum post content; and integrate basic learning profiles from the student management system to generate a raw education dataset covering the entire process of pre-class preparation, in-class participation, and post-class consolidation. Data preprocessing and feature fusion steps: Receive the original educational dataset, use a time-adaptive alignment algorithm to unify asynchronous data streams from different sources to a standard time axis, use the isolated forest algorithm to identify and label outlier data points, fill missing values ​​using a multiple imputation method, and introduce a pre-trained language model to extract semantic feature vectors from text data, outputting standardized feature vectors. Multi-dimensional ability assessment steps: Receive the standardized feature vector, construct an ability assessment framework based on five primary indicators: knowledge mastery level, learning engagement, thinking ability development, collaborative communication ability, and self-learning ability, determine the weight of each indicator using a subjective and objective weighting algorithm, and update the learner's mastery probability estimate of each knowledge point in real time through a dual-stream attention knowledge tracking algorithm, and output the ability assessment result vector and knowledge mastery probability distribution. Learning behavior sequence analysis steps: Receive the semantic feature vector, use the sequence pattern mining algorithm to discover high-frequency behavior patterns from the learning trajectory, use association rules to extract the potential correlation between effective learning strategies and academic performance, and use the key behavior node identification algorithm to perform time-series modeling of the learning behavior sequence, automatically identify key behavior nodes that affect learning effect, and output key behavior node sequence and behavior pattern features. Personalized guidance recommendation steps: Receive the ability assessment result vector and the key behavior node sequence, comprehensively consider the learner's current knowledge status, historical learning preferences and goal achievement gap, use a hybrid strategy of collaborative filtering and content recommendation to match learning materials from the resource library, continuously adjust the recommendation strategy through the cognitive load perception path optimization algorithm, feed back the learning path optimization signal to the multi-dimensional ability assessment step to update the assessment parameters, and output a recommended learning resource list; Multi-role visualization steps: Receive the knowledge mastery probability distribution, behavioral pattern characteristics, and recommended learning resource list; generate differentiated data dashboards for four roles: teachers, students, parents, and administrators; and use natural language generation technology to convert statistical indicators into natural language evaluation reports.

Images