A method and system for assessing teachers' digital literacy based on process data

Abstract

This invention discloses a method and system for assessing teachers' digital literacy based on process data, belonging to the field of teacher evaluation. The method includes the following steps: S1, Data Acquisition and Structured Processing: Acquiring multi-source heterogeneous data such as classroom interaction data, online operation logs, resource usage trajectories, and audio / video data from teachers' multi-scenario teaching activities; S2, Cognitive Load and Behavioral Entropy Calculation: Based on the collected data, calculating the cognitive load index and behavioral entropy value of teachers during task execution; inputting semantic, behavioral, and contextual features into a dynamic evolutionary coding network, generating a temporal ability evolution vector through an adaptive attention mechanism and temporal difference reinforcement learning; S4, Contrastive Fusion and Latent Variable Generation; S5, Graph-Driven Assessment and Interpretable Output. This solution can comprehensively characterize teachers' actual operational abilities and strategy application in digital teaching.

Description

Technical Field

[0001] This invention relates to the field of teacher evaluation, and more specifically, to a method and system for assessing teachers' digital literacy based on process data. Background Technology

[0002] With the rapid development of information technology, digital education is becoming an important trend in teaching reform. Teachers' digital literacy levels directly impact their digital teaching abilities and are crucial to improving the quality of education. However, current teacher digital literacy assessments primarily rely on static methods such as traditional questionnaires or end-of-term assessments. These methods have significant limitations: traditional assessments mainly collect self-assessment data from teachers through paper questionnaires or online tests, lacking real-time observation and recording of the entire teaching process, and failing to comprehensively reflect teachers' actual digital operation abilities in the classroom and online teaching. Existing methods typically focus only on knowledge acquisition or skill scores, ignoring the dynamic process characteristics of teachers' behavior, operation frequency, cognitive load, and decision-making strategies in digital teaching, leading to biased and unobjective evaluation results. Traditional evaluation results are usually presented in the form of total scores or grades, making it difficult to identify teachers' strengths and weaknesses in specific ability dimensions, and failing to provide targeted improvement suggestions for subsequent training, thus limiting the practical value of the evaluation results. Teachers' digital abilities evolve over time, but existing evaluation systems lack continuity, failing to track changes in teacher abilities in real time, and cannot form a data-driven adaptive optimization mechanism. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for assessing teachers' digital literacy based on process data, so as to solve the problems mentioned in the background art.

[0004] Technical Solution: A method for assessing teachers' digital literacy based on process data, comprising the following steps:

[0005] S1. Data Acquisition and Structured Processing: In teachers' multi-scenario teaching activities, acquire multi-source heterogeneous data such as classroom interaction data, online operation logs, resource usage trajectories, and audio and video data, and construct time-series behavior clusters to achieve structured representation of process data;

[0006] S2. Cognitive Load and Behavioral Entropy Calculation: Based on the collected data, calculate the cognitive load index and behavioral entropy value of teachers during task execution to characterize the depth of information processing and the diversity of strategies.

[0007] S3. Dynamic Feature Encoding: Input semantic, behavioral and contextual features into a dynamic evolutionary coding network (DEE) and generate temporal capability evolution vectors through an adaptive attention mechanism and temporal difference reinforcement learning (TD-RL).

[0008] S4. Contrast Fusion and Latent Variable Generation: Feature alignment and fusion are performed through a hierarchical contrast fusion network (HCFN), and a digital literacy latent variable matrix is ​​generated by combining the similarity among teachers.

[0009] S5. Graph-driven assessment and interpretable output: Input the latent variable matrix into the graph-driven adaptive convolutional network (AGCN) to perform ability value reasoning and generate an interpretable indicator contribution graph, and output the teacher's total digital literacy score and dimensional distribution.

[0010] Preferably, the calculation of the cognitive load index and behavioral entropy value in S2 further includes: the operation complexity coefficient is calculated based on the number of operation steps, the frequency of tool switching, and the complexity of task type; the behavioral entropy value takes into account the probability distribution of action type and time weighting factors to enhance the sensitivity to short-term high-frequency operations; the cognitive load index and behavioral entropy value are standardized and input into the dynamic feature encoding module.

[0011] Preferably, the dynamic feature encoding in S3 further includes: independently encoding semantic, behavioral, and contextual features, and then performing cross-modal interactive encoding through a self-attention module; introducing a temporal difference reinforcement learning mechanism to dynamically adjust feature weights and capture the evolution trend of teachers' digital behavior; and outputting a time-series evolution vector of teachers' digital literacy to represent the dynamic changes in teachers' abilities over time.

[0012] Preferably, the comparison fusion in S4 further includes: constructing a teacher similarity matrix (cosine similarity or Mahalanobis distance); using a hierarchical comparison fusion network (HCFN) to perform semantic-behavioral alignment fusion of temporal features; and outputting a normalized latent variable matrix as an input to the assessment network as a vector of teacher digital literacy ability.

[0013] Preferably, the graph-driven assessment in S5 further includes: mapping the latent variable matrix to the nodes of the teacher's digital competence graph to construct a graph-structured adjacency matrix; using an adaptive graph convolutional network to propagate node information and perform competence value inference; and outputting the teacher's total digital literacy score and an interpretable indicator contribution chain.

[0014] Preferably, in step S6, the gradient is calculated based on the trend of teachers' digital literacy scores, and the model is retrained when the gradient is lower than the threshold; a contrastive learning mechanism is introduced to enhance the model's differentiation among similar teacher samples, thereby improving its generalization ability and stability; the optimized model parameters are fed back to the data collection and feature encoding process to achieve closed-loop adaptive optimization.

[0015] Preferably, S7 automatically matches differentiated ability improvement paths based on teachers' total digital literacy score and dimensional distribution; outputs personalized training suggestions, including digital teaching innovation ability, data-driven decision-making ability, and information security awareness improvement; and inputs the training results together with subsequent teaching behavior data to form a continuous ability tracking closed loop.

[0016] Preferably, in step S5, the generated latent variable matrix is ​​further subjected to the following operations: the marginal contribution of each feature in the latent variable matrix is ​​calculated using the Shapley value algorithm; a multi-dimensional visualization report is generated, including semantic, behavioral, and contextual dimension radar charts, and the contribution of indicators to the total score is marked; interactive queries and interpretable analysis are supported for teachers and administrators.

[0017] Preferably, the multi-source heterogeneous data in S1 is subjected to time synchronization, anomaly detection, missing data completion, and format unification; the text data is semantically vectorized and encoded; the behavior sequence is processed by sliding window to generate a behavior matrix; and the structured features are output as input to the dynamic feature encoding module.

[0018] A teacher digital literacy assessment system based on process data includes:

[0019] Data acquisition module: Collects data on teacher classroom interactions, online operations, resource access, and audio / video behavior.

[0020] Teaching behavior monitoring submodule: Real-time capture of screen sharing, courseware switching, interactive quizzes, and whiteboard writing behaviors;

[0021] Resource Tracking Submodule: Records teachers' resource access, document management, and online experiment operations on different platforms;

[0022] Physiological perception submodule: Collects physiological data such as teachers' eye movements, facial expressions, and heart rate to supplement cognitive state parameters;

[0023] Event anchoring mechanism: Automatically identify key teaching behavior nodes and construct time windows to form structured data clusters;

[0024] Data preprocessing module: performs synchronization, cleaning, missing data completion, and feature format unification on multi-source heterogeneous data;

[0025] Dynamic feature encoding module: Implements temporal encoding of semantic, behavioral, and contextual features, and outputs capability evolution vectors;

[0026] The comparison and fusion module performs hierarchical comparison and fusion based on the teacher similarity matrix to generate a latent variable matrix;

[0027] Assessment Modeling and Graph Module: Constructs a knowledge graph of teachers' digital competence, and performs competence reasoning and interpretable output based on an adaptive graph convolutional network;

[0028] Results Application and Optimization Module: Generates personalized training suggestions, closed-loop capability tracking, and dynamically optimizes model parameters.

[0029] Compared with the prior art, the advantages of this invention are:

[0030] (1) By collecting and processing classroom interaction data, online operation logs, resource usage trajectories, and audio-visual data of teachers in multi-scenario teaching activities through step S1, this invention can obtain process data of teachers' digital teaching. At the same time, semantic vectorization, sliding window processing, and format unification are performed on text, behavioral sequences, and contextual information to form high-dimensional structured features, providing a comprehensive and objective data foundation for subsequent dynamic analysis. Compared with traditional evaluation methods that rely on static questionnaires or single tests, this solution can comprehensively depict teachers' actual operational capabilities and strategy application in digital teaching.

[0031] (2) In step S3, this invention uses a dynamic evolutionary coding network (DEE) combined with an adaptive attention mechanism and temporal differential reinforcement learning (TD-RL) to encode semantic, behavioral, and contextual features across modalities and generate a time-series evolution vector of teachers' digital literacy. This design can capture the dynamic evolution trend of teachers' abilities over time, overcome the static and lagging problems of traditional evaluation methods, and make the evaluation results more accurate and reliable.

[0032] (3) Step S4 uses a hierarchical contrastive fusion network (HCFN) to achieve semantic-behavioral alignment fusion of teachers' temporal features and combines it with the similarity between teachers to generate a latent variable matrix. This mechanism can enhance the comparability and model stability among teachers, while eliminating noise caused by individual data differences and improving the interpretability and robustness of digital literacy ability vectors.

[0033] (4) In step S5, this invention maps the latent variable matrix to nodes in the teacher digital competency graph, uses an adaptive graph convolutional network (AGCN) for competency inference, and combines Shapley values ​​to calculate the marginal contribution of latent variable features to the total score and dimensional scores. Through the visualization of radar charts and indicator contribution chain diagrams, teachers and administrators can interactively analyze and intuitively identify competency strengths and weaknesses, enhancing the interpretability and practicality of the evaluation results. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the overall process of a teacher digital literacy assessment method based on process data according to the present invention. Detailed Implementation

[0035] For examples, please refer to Figure 1 A method for assessing teachers' digital literacy based on process data includes the following steps:

[0036] S1. Data Acquisition and Structured Processing: In teachers' multi-scenario teaching activities, acquire multi-source heterogeneous data such as classroom interaction data, online operation logs, resource usage trajectories, and audio / video data, and construct time-series behavior clusters to achieve structured representation of process data; perform time synchronization, anomaly detection, missing data completion, and format unification on the multi-source heterogeneous data in S1; perform semantic vectorization encoding on text data, and use sliding window processing on behavior sequences to generate behavior matrices; output structured features as input to the dynamic feature encoding module.

[0037] S1.1 Multi-source data acquisition

[0038] S1.1.1 Classroom Interaction Data Collection: Capture teacher screen sharing behavior, courseware switching, blackboard writing operations, interactive answer records, etc. through teaching platforms or classroom monitoring systems.

[0039] S1.1.2 Online Operation Log Collection: Records the teacher's operation trajectory on the digital platform, including login, clicks, file upload and download, tool switching, online experiment operation, etc.

[0040] S1.1.3 Resource Usage Tracking: Tracking the teaching resources (courseware, documents, multimedia content) used by teachers, their management operations, and their participation in teaching activities.

[0041] S1.1.4 Audio-visual and physiological perception data collection (optional): Collect teachers' classroom audio-visual data, facial expressions, eye movement trajectories, heart rate changes, etc., to assist in cognitive state analysis.

[0042] S1.2 Data Time Synchronization and Cleaning

[0043] S1.2.1 Time Synchronization: Timestamps are unified for data from different sources, and operation logs, interaction records, and sensor data are aligned.

[0044] S1.2.2 Anomaly Detection: Detect missing, duplicate, or abnormal operation events, and mark or remove abnormal data.

[0045] S1.2.3 Missing value imputation: For missing data, imputation is performed using mean imputation, interpolation, or model prediction methods.

[0046] S1.2.4 Format unification: Convert data from different sources into a unified format, such as converting text to string vectors, standardizing numerical data, and encoding event sequences.

[0047] S1.3 Semantic Vectorization of Text Data

[0048] Semantic vectorization is performed on the text data generated by teachers (such as teaching logs, class notes, and answer feedback).

[0049] Use word embedding (Word2Vec, GloVe), sentence embedding (Sentence-BERT), or other natural language processing methods to convert text into vector representations.

[0050] S1.4 Behavioral Sequence Processing

[0051] Organize the teacher's operation sequence or event sequence into a behavior sequence matrix according to the time sequence.

[0052] A sliding window process is applied to the behavioral sequence to generate a time window behavioral matrix, which is used to capture short-term behavioral patterns and long-term trends.

[0053] S1.5 Structured Feature Output

[0054] The vectors and matrices processed from various data are integrated to form a unified structured feature set.

[0055] The output structured features serve as input to the S3 dynamic feature encoding module, providing foundational data for subsequent cognitive load calculation, dynamic feature encoding, and evaluation analysis.

[0056] S2. Cognitive Load and Behavioral Entropy Calculation: Based on the collected data, the cognitive load index and behavioral entropy value of teachers during task execution are calculated to characterize the depth of information processing and the diversity of strategies. The calculation of the cognitive load index and behavioral entropy value in S2 further includes: the operation complexity coefficient is calculated based on the number of operation steps, the frequency of tool switching, and the complexity of task type; the behavioral entropy value considers the probability distribution of action type and time weighting factors to enhance the sensitivity to short-term high-frequency operations; the cognitive load index and behavioral entropy value are standardized and input into the dynamic feature encoding module.

[0057] S2.1 Calculation of Operation Complexity Coefficient

[0058] Count the number of steps teachers take during task execution. Number of tool switches and task type complexity (Task complexity can be assigned according to preset level 15).

[0059] Calculate the operation complexity coefficient :

[0060]

[0061] in, The number of operational steps actually performed by the teacher; : The maximum number of operation steps in a similar task (used for normalization); The number of times the teacher switches tools during a task; The maximum number of tool switches within the same type of task; : The complexity level of the current task type; Maximum complexity level; The weighting coefficients for operation steps, tool switching, and task complexity should meet the following requirements. ;

[0062] S2.2 Behavioral Entropy Calculation

[0063] Classify teacher behavior sequences and statistically analyze different action types. probability of occurrence (Action types may include clicking, dragging, writing on the whiteboard, switching courseware, etc.)

[0064] Introducing time weighting factors It enhances sensitivity to short-term, high-frequency movements.

[0065] Calculate weighted behavioral entropy :

[0066]

[0067] in, Total number of action types; Action type Probability of occurrence ; Time-weighted coefficient, which can be set according to the frequency of action and time interval; short-term, high-frequency actions have a larger weight.

[0068] The entropy value of the teacher's behavioral sequence is used to characterize the diversity of behavior and the degree of strategy dispersion.

[0069] S2.3 Cognitive Load Index Calculation

[0070] The operation complexity coefficient With behavioral entropy Comprehensive calculation of cognitive load index :

[0071]

[0072] Among them, CLI: Cognitive Load Index, reflects the information processing load of teachers during task execution; : Maximum behavioral entropy (used for normalization); : A coefficient that adjusts the weighting of the contribution of operational complexity and behavioral entropy. ;

[0073] S2.4 Standardization Processing

[0074] Cognitive Load Index With behavioral entropy Standardize it so that its value range is within :

[0075]

[0076]

[0077] Standardized and The output is sent to the S3 dynamic feature encoding module as input for capability evolution vector calculation.

[0078] S3. Dynamic Feature Encoding: Semantic, behavioral, and contextual features are input into a Dynamic Evolutionary Encoding Network (DEE). An adaptive attention mechanism and Temporal Differential Reinforcement Learning (TD-RL) are used to generate a temporal capability evolution vector. S3's dynamic feature encoding further includes: independently encoding semantic, behavioral, and contextual features, followed by cross-modal interactive encoding via a self-attention module; introducing a temporal differential reinforcement learning mechanism to dynamically adjust feature weights and capture the evolutionary trend of teachers' digital behavior; and outputting a temporal evolution vector of teachers' digital literacy to represent the dynamic changes in teachers' capabilities over time. S3.1 Input Feature Preparation

[0079] The structured features output by S1 and the cognitive load index after standardization by S2 are used to analyze the structured features output by S1 and the cognitive load index after standardization by S2. and behavioral entropy As input.

[0080] The input features are categorized into three modalities based on their type:

[0081] Semantic features: Text vectorization results, such as teaching logs and classroom answer feedback vectors;

[0082] Behavioral characteristics: teacher operation sequence matrix and behavioral time window encoding;

[0083] Contextual characteristics: task type complexity, classroom environment parameters, online teaching platform operation context, etc.;

[0084] S3.2 Independent Encoding

[0085] Each modal feature is processed using an independent encoder:

[0086] Semantic feature encoders: can use LSTM, GRU, or Transformer to map semantic vector sequences to latent space representations. ;

[0087] Behavioral feature encoder: Convolution or RNN encoding is performed on the behavior matrix of the sliding window to obtain the behavioral latent space representation. ;

[0088] Contextual Feature Encoder: Performs linear mapping or MLP encoding on task and environmental features to obtain a contextual latent space representation. ;

[0089] S3.3 Cross-modal interactive coding (self-attention mechanism)

[0090] The three types of latent space representations are input into the attention module to calculate the intermodal interaction relationships:

[0091]

[0092] in, : Query, key-value matrix, by generate; : Key vector dimension, used for normalization; output is the fused cross-modal feature vector. .

[0093] S3.4 Temporal Differential Reinforcement Learning (TDRL) Mechanism

[0094] To capture the evolutionary trends of teachers' digital behavior, temporal difference reinforcement learning is introduced to adjust feature weights:

[0095] Define state This indicates that the teacher is in time. The characteristic state;

[0096] Define Action Feature weight adjustment strategy;

[0097] Define rewards Optimize indicators for incremental or cognitive load in improving teachers' digital literacy;

[0098] Using TDRL update strategy function :

[0099]

[0100] in, : State action value function; Learning rate; Discount factor; The next time step state; the TDRL mechanism dynamically adjusts the weights of each feature mode in the final evolution vector, so that the feature encoding can reflect the changing trend of teacher ability over time.

[0101] S3.5 Output of the Temporal Evolution Vector of Teachers' Digital Literacy

[0102] The cross-modal feature sequences, weighted by self-attention and TDRL mechanisms, are integrated to generate a time-series evolution vector of teachers' digital literacy. ;

[0103] It represents the dynamic changes in teachers' abilities throughout the assessment cycle and can serve as input for S4 assessment graph reasoning and S5 interpretable output.

[0104] S4. Contrast Fusion and Latent Variable Generation: Feature alignment and fusion are performed through a Hierarchical Contrast Fusion Network (HCFN), and a digital literacy latent variable matrix is ​​generated by combining the similarity between teachers. The contrast fusion in S4 further includes: constructing a teacher similarity matrix (cosine similarity or Mahalanobis distance); performing semantic-behavioral alignment fusion of temporal features through the Hierarchical Contrast Fusion Network (HCFN); and outputting a normalized latent variable matrix as the input of the teacher digital literacy ability vector into the assessment network.

[0105] S4.1 Input Preparation

[0106] Input: Teacher digital literacy time-series evolution vector from S3 Each For teachers in time The capability vector.

[0107] The input also includes a sample set of teachers. This is used to calculate the similarity between teachers.

[0108] S4.2 Construction of Teacher Similarity Matrix

[0109] For the teacher sample set Calculate the similarity between the time-series vectors:

[0110] S4.2.1 Cosine Similarity Calculation

[0111]

[0112] in, :teacher With teachers Cosine similarity; :teacher and The temporal evolution vector; Vector L2 norm;

[0113] S4.2.2 Mahalanobis distance calculation (optional)

[0114]

[0115] in, :teacher and Mahalanobis distance; The covariance matrix of the ability vectors of all teachers; constructing the teacher similarity matrix using cosine similarity or Mahalanobis distance. .

[0116] S4.3 Hierarchical Contrastive Learning and Feature Alignment Fusion

[0117] Construct a hierarchical contrastive fusion network to perform semantic-behavioral-contextual alignment and fusion of teachers' temporal features:

[0118] S4.3.1 Hierarchical Division

[0119] First layer: Semantic-behavioral alignment

[0120] Second layer: Behavior-context alignment

[0121] Third layer: Overall feature fusion

[0122] S4.3.2 Contrastive Learning Loss Function

[0123]

[0124] in, : Set of positive sample pairs (highly similar teacher pairs); :sample The negative sample set (low-similarity teacher pairs); : Feature representation of HCFN output; Similarity functions (such as cosine similarity); Temperature coefficient is used to control the similarity distribution; HCFN achieves feature alignment and cross-modal fusion among teachers by minimizing contrast loss.

[0125] S4.4 Generation and Normalization of Latent Variable Matrix

[0126] The fused feature representations are integrated into a teacher digital literacy latent variable matrix. Each row corresponds to a teacher's ability vector, and each column corresponds to a latent variable dimension.

[0127] The matrix is ​​normalized so that the dimension of each latent variable ranges within a certain range. :

[0128]

[0129] in, : The first in the matrix Line 1 Column elements; , : No. Find the minimum and maximum values ​​of the column; output the normalized latent variable matrix. It serves as the vector input for teachers' digital literacy competence into the S5 assessment network.

[0130] S5. Graph-Driven Assessment and Interpretable Output: The latent variable matrix is ​​input into a graph-driven adaptive convolutional network (AGCN) to perform competency inference and generate an interpretable indicator contribution graph, outputting the teacher's total digital literacy score and dimensional distribution. Further, in step S5, the generated latent variable matrix is ​​processed as follows: the marginal contribution of each feature in the latent variable matrix is ​​calculated using the Shapley value algorithm; a multi-dimensional visualization report is generated, including semantic, behavioral, and contextual dimension radar charts, and the contribution of indicators to the total score is labeled; interactive queries and interpretable analysis are supported for teachers and administrators. The graph-driven assessment in S5 further includes: mapping the latent variable matrix to nodes in the teacher's digital competency graph, constructing a graph-structured adjacency matrix; using an adaptive graph convolutional network to propagate node information and perform competency inference; and outputting the teacher's total digital literacy score and an interpretable indicator contribution chain.

[0131] S5.1 Input Preparation

[0132] Input the normalized latent variable matrix from S4 ,in For the teacher sample size, For latent variable dimensions.

[0133] The input also includes information on the structure of the teacher digital competence graph (which can be built based on the school or platform) to map latent variables to graph nodes.

[0134] S5.2 Mapping the latent variable matrix to the graph nodes

[0135] Each row of the latent variable matrix Nodes mapped to the teacher digital competence map Each latent variable corresponds to a feature attribute of a graph node.

[0136] Constructing a graph-structured adjacency matrix The potential relationship or ability similarity between teachers can be represented by cosine similarity or Mahalanobis distance:

[0137] A ij =

[0138] in, :teacher and Similarity (from S4); The set similarity threshold;

[0139] S5.3 Adaptive Graph Convolutional Network (AGCN) Node Information Propagation

[0140] Using AGCN to perform convolutional propagation on graph node features:

[0141]

[0142] in, Initial node feature matrix; : Adding a self-loop adjacency matrix; The degree matrix; : No. The trainable weight matrix of the layer; Activation function : Final node representation, used for ability value reasoning.

[0143] S5.4 Reasoning about Teachers' Digital Literacy Ability

[0144] For each teacher node The final feature vector Perform linear mapping or MLP to predict teachers' total digital literacy score and scores in each dimension:

[0145]

[0146] in, : Linear mapping parameters obtained through training; :teacher The total score and scores for each dimension of digital literacy;

[0147] S5.5 Shapley value characteristic contribution calculation

[0148] For the latent variable matrix Calculate the marginal contribution for each column (feature dimension):

[0149]

[0150] in, : A set of latent variable dimensions; Using feature subsets Predicted digital literacy score; :feature Marginal contribution to the total score; Shapley values ​​can be used to explain the contribution of latent variables to the total score of teachers' digital literacy.

[0151] S5.6 Visualization and Interactive Output

[0152] Generate a multi-dimensional visualization report based on Shapley scores and scores across various dimensions:

[0153] Radar chart: Displays the ability scores for semantic, behavioral, and contextual dimensions, and their contribution to the total score.

[0154] Indicator Contribution Chain Diagram: Shows the correlation between latent variables and total score and dimensional scores.

[0155] Supports interactive queries by teachers and administrators, allowing them to select teachers, dimensions, or latent variables for analysis.

[0156] S6. Calculate the gradient based on the trend of teachers' digital literacy scores. When the gradient is lower than the threshold, the model is retrained. Introduce a contrastive learning mechanism to enhance the model's differentiation among similar teacher samples, and improve its generalization ability and stability. Feed the optimized model parameters back to the data collection and feature encoding process to achieve closed-loop adaptive optimization.

[0157] S7. Based on the total score and dimensional distribution of teachers' digital literacy, automatically match differentiated ability improvement paths; output personalized training suggestions, including digital teaching innovation ability, data-driven decision-making ability, and information security awareness improvement; and input the training results together with subsequent teaching behavior data to form a continuous ability tracking closed loop.

[0158] A teacher digital literacy assessment system based on process data includes:

[0159] Data acquisition module: Collects data on teacher classroom interactions, online operations, resource access, and audio / video behavior.

[0160] Teaching behavior monitoring submodule: Real-time capture of screen sharing, courseware switching, interactive quizzes, and whiteboard writing behaviors;

[0161] Resource Tracking Submodule: Records teachers' resource access, document management, and online experiment operations on different platforms;

[0162] Physiological perception submodule: Collects physiological data such as teachers' eye movements, facial expressions, and heart rate to supplement cognitive state parameters;

[0163] Event anchoring mechanism: Automatically identify key teaching behavior nodes and construct time windows to form structured data clusters;

[0164] Data preprocessing module: performs synchronization, cleaning, missing data completion, and feature format unification on multi-source heterogeneous data;

[0165] Dynamic feature encoding module: Implements temporal encoding of semantic, behavioral, and contextual features, and outputs capability evolution vectors;

[0166] The comparison and fusion module performs hierarchical comparison and fusion based on the teacher similarity matrix to generate a latent variable matrix;

[0167] Assessment Modeling and Graph Module: Constructs a knowledge graph of teachers' digital competence, and performs competence reasoning and interpretable output based on an adaptive graph convolutional network;

[0168] Results Application and Optimization Module: Generates personalized training suggestions, closed-loop capability tracking, and dynamically optimizes model parameters.

[0169] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples of the present invention and are not intended to limit the present invention. Various changes and modifications can be made to the present invention without departing from the spirit and scope thereof, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A teacher digital literacy assessment method based on process data, characterized in that, Comprise the following steps: S1, data acquisition and structured processing: in the multi-scene teaching activities of teachers, multi-source heterogeneous data such as classroom interaction data, online operation log, resource usage track and audio and video data are obtained, and time sequence behavior cluster is constructed to realize process data structured representation; S2, cognitive load and behavior entropy calculation: based on the collected data, the cognitive load index and behavior entropy value of the teacher in the task execution process are calculated, which are used to describe the information processing depth and strategy diversity; The calculation of cognitive load index and behavior entropy value in S2 further comprises: operation complexity coefficient is calculated based on operation step number, tool switching frequency and task type complexity; the behavior entropy value considers the probability distribution of action type and time weighting factor to enhance the sensitivity to short-term high-frequency operation; the cognitive load index is calculated by comprehensively calculating the operation complexity coefficient and the behavior entropy, which reflects the information processing load of the teacher in the task execution process; the cognitive load index and the behavior entropy value are standardized and input into the dynamic feature coding module; S3, dynamic feature coding: input semantic, behavior and context features into dynamic evolution coding network, generate teacher digital literacy time evolution vector through adaptive attention mechanism and time difference reinforcement learning; S4, comparison fusion and latent variable generation: feature alignment and fusion are carried out through hierarchical comparison fusion network, and digital literacy latent variable matrix is generated by combining similarity calculation between teachers; S5, atlas driven evaluation and interpretable output: input the latent variable matrix into the atlas driven adaptive convolution network, perform ability value reasoning and generate interpretable index contribution graph, output the total score and dimension distribution of teacher digital literacy.

2. The teacher digital literacy assessment method based on process data according to claim 1, characterized in that, The dynamic feature coding in S3 further comprises: the semantic, behavior and context features are independently coded, and then cross-modal interaction coding is carried out through self-attention module; the time difference reinforcement learning mechanism is introduced to dynamically adjust the feature weight and capture the evolution trend of teacher digital behavior; the teacher digital literacy time evolution vector is output, which is used to represent the dynamic change of teacher ability with time.

3. The teacher digital literacy assessment method based on process data according to claim 1, characterized in that, The comparison fusion in S4 further comprises: constructing teacher similarity matrix; hierarchical comparison learning is used for semantic-behavior alignment and fusion of time sequence features; output normalized latent variable matrix as teacher digital literacy ability vector input into evaluation network.

4. The teacher digital literacy assessment method based on process data according to claim 1, characterized in that, The atlas driven evaluation in S5 further comprises: mapping the latent variable matrix to the nodes of teacher digital ability atlas to construct graph structure adjacency matrix; node information is propagated by using adaptive graph convolution network to perform ability value reasoning; the total score and interpretable index contribution chain of teacher digital literacy are output.

5. The teacher digital literacy assessment method based on process data according to claim 1, characterized in that, S6, calculate the gradient based on the change trend of teacher digital literacy score, and trigger model retraining when the gradient is lower than the threshold; introduce contrast learning mechanism to strengthen the discrimination degree of the model among similar teachers, improve the generalization ability and stability; feedback the optimized model parameters to the data acquisition and feature coding process to realize closed-loop adaptive optimization.

6. The teacher digital literacy assessment method based on process data according to claim 1, characterized in that, S7, according to the total score and dimension distribution of teacher digital literacy, automatically match the differentiated ability improvement path; Output personalized training recommendations, including digital teaching innovation ability, data-driven decision-making ability, and information security awareness improvement; input training results and subsequent teaching behavior data to form a continuous ability tracking closed loop.

7. The teacher digital literacy assessment method based on process data according to claim 1, characterized in that, In step S5, further, the following operation is performed on the generated latent variable matrix: the marginal contribution of each feature in the latent variable matrix is calculated using the Shapley value algorithm; Generate multi-dimensional visualization reports, including semantic, behavioral, and situational dimension radar charts, and mark the contribution of indicators to the total score; support interactive query and interpretable analysis for teachers and managers.

8. The teacher digital literacy assessment method based on process data according to claim 1, characterized in that, Also includes: time synchronization, anomaly detection, missing data completion, and format unification of multi-source heterogeneous data in S1; semantic vectorization encoding of text data and sliding window processing of behavior sequences to generate behavior matrices; output structured features as input to the dynamic feature encoding module.

9. A teacher digital literacy assessment system based on process data, involving a teacher digital literacy assessment method based on process data according to any one of claims 1-8, characterized in that, Includes: Data acquisition module: collect teacher classroom interaction, online operation, resource calling, and audio and video behavior data; Teaching behavior monitoring sub-module: real-time capture of screen sharing, courseware switching, interactive answering, and blackboard writing behavior; Resource tracking sub-module: records the teacher's resource access, document management, and online experiment operation on different platforms; Physiological perception sub-module: collect teacher eye movement, facial expression, and heart rate physiological data to supplement cognitive state parameters; Event anchor labeling mechanism: automatically identify key teaching behavior nodes and build time windows to form structured data clusters; Data preprocessing module: synchronize, clean, complete missing data, and unify feature formats for multi-source heterogeneous data; Dynamic feature encoding module: realize the time series encoding of semantic, behavioral, and situational features, and output the ability evolution vector; Contrast fusion module: hierarchical contrast fusion based on teacher similarity matrix to generate latent variable matrix; Evaluation modeling and graph module: construct a teacher digital ability knowledge graph and perform ability reasoning and interpretable output based on an adaptive graph convolutional network; Result application and optimization module: generate personalized training recommendations, closed-loop ability tracking, and dynamically optimize model parameters.

Images