AI teaching evaluation and post self-adaption method based on multi-layer dynamic graph mapping
The AI-based teaching evaluation method using multi-layer dynamic graph mapping solves the problem of the disconnect between teaching evaluation and job requirements. It enables real-time mapping of students' abilities to job requirements and personalized learning path recommendations, thereby improving teaching efficiency and personalized training.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- JILIN JUDICIAL POLICE OFFICER VOCATIONAL COLLEGE
- Filing Date
- 2026-04-15
- Publication Date
- 2026-07-14
AI Technical Summary
The existing teaching evaluation system cannot map students' various abilities to the specific job requirements in the industry chain in real time, resulting in a separation between learning and application, data silos and a single modality, and a lack of process-oriented intelligent intervention.
An AI-based teaching evaluation method based on multi-layer dynamic graph mapping is adopted. Through natural language processing and knowledge graph technology, a multi-layer dynamic graph from the industry chain to knowledge points is constructed. Multimodal behavioral data is collected to generate student learning profiles, which are then vectorized and matched with job requirement models to output learning path suggestions.
It has achieved precise evaluation of industry-education integration, generated a comprehensive ability radar chart, reduced the management burden on teachers, improved teaching efficiency and personalized training level, and realized the leap from learning to ability.
Smart Images
Figure CN122390923A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of teaching evaluation technology, and in particular to an AI-based teaching evaluation and job adaptation method based on multi-layer dynamic graph mapping. Background Technology
[0002] Current management technologies mainly fall into two categories: traditional academic affairs management systems, which primarily record student attendance, homework grades, and exam scores, offering only a single data dimension and representing outcome-based evaluation; and single-scenario online learning platforms, which, while recording online learning time and click-through rates, fail to capture the true learning status of offline classrooms and lack relevance to actual job requirements.
[0003] Problems and shortcomings of existing technologies: (1) Disconnect between evaluation and job requirements: Existing teaching evaluation systems often only focus on the mastery of textbook knowledge and fail to map students’ various abilities with the specific job requirements in the industry chain in real time, resulting in the separation of learning and application; (2) Data silos and single modality: Classroom behavior data, online learning data and data from third-party training systems are usually stored in different systems, which cannot be integrated and analyzed, making it difficult to form a complete student profile; (3) Lack of process-oriented intelligent intervention: Existing technologies are mostly post-event analysis and lack an immediate warning and dynamic intervention mechanism based on real-time behavior.
[0004] Therefore, proposing an AI teaching evaluation and job adaptation method based on multi-layer dynamic graph mapping to solve the difficulties of existing technologies is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0005] The purpose of this invention is to provide an AI-based teaching evaluation and job adaptation method based on multi-layer dynamic graph mapping, which enables highly intelligent adaptive intervention and can automatically generate learning prescriptions, classroom pace suggestions, and academic risk warnings, greatly reducing the management burden on teachers and improving teaching efficiency and personalized training levels.
[0006] To achieve the above objectives, the present invention provides the following solution: An AI-based teaching evaluation and job adaptation method based on multi-layer dynamic graph mapping includes the following steps: S1. Based on natural language processing and knowledge graph technology, automatically construct a multi-layered dynamic graph from the industry chain to knowledge points; S2. Collect multimodal behavioral data of students in different scenarios through classroom video analysis, online learning behavior capture and integration with the training system; S3. Based on the collected multimodal behavioral data and multi-layer dynamic graph, the pre-trained large model is fine-tuned in the domain to generate student learning profiles and academic risk warning categories; S4. Integrate multi-source data to generate a radar chart of students' comprehensive abilities, and perform vectorized matching with job requirement models to output job matching degree and ability gap analysis. Finally, generate learning path suggestions that include strengthening knowledge points and skills.
[0007] Preferably, in S1, the automated construction of a multi-layered dynamic graph from the industry chain to knowledge points specifically includes: Entity recognition: First, BERT pre-trained models are used to encode word vectors into massive amounts of teaching and industry data to obtain contextual semantic features; second, BiLSTM bidirectional long short-term memory networks are used to capture long-distance dependencies in the text; finally, CRF conditional random field layers are used for sequence labeling to extract job positions, skills, knowledge entities and the relationships between them. Keyword auto-extraction: The TF-IDF and TextRank algorithms are used to extract high-frequency keywords from industry documents, and Word2Vec is used to calculate word vector similarity to automatically remove common stop words and retain core skill vocabulary.
[0008] Preferably, in S2, multimodal behavioral data of students in different scenarios is collected through classroom video analysis, online learning behavior capture, and integration with the practical training system. Specifically, this includes: Face detection is performed using the RetinaFace model, combined with the PnP algorithm to detect the Euler angles of students' heads, and the head-up rate in class is calculated based on the Euler angles, which include pitch, yaw, and roll angles. A facial expression recognition model based on ResNet-50 is used to identify students' emotion categories and confidence levels. OpenPose or YOLOv8-Pose skeletal keypoint detection algorithms are used to capture 18 key points of the human body in real time, and a spatiotemporal graph convolutional network is constructed to identify action categories.
[0009] Preferably, S2 also includes: Students' classroom attention scores are calculated using a pre-set formula, as follows:
[0010] in, Head-up rate per unit time For the confidence score of positive sentiment extracted based on CNN, Frequency of fatigued movements The weighting coefficient of head-up rate per unit time. The weighting coefficients for positive sentiment confidence. This is the weighting coefficient for the frequency of fatigued actions.
[0011] Preferably, in S3, based on the collected multimodal behavioral data and multi-layer dynamic graph, the pre-trained large model is fine-tuned in a specific domain, including: The pre-trained large model is a base model using the Transformer architecture. It uses LoRA technology for low-rank fine-tuning. By freezing the weights of the pre-trained large model and training only the low-rank matrix, it injects education-specific data, multimodal behavioral data, and multi-layer dynamic graphs into the pre-trained large model, enabling it to have the ability to reason in the education domain.
[0012] Preferably, in S4, a comprehensive ability radar chart of students is generated by integrating multi-source data, and this chart is then vectorized and matched with the job requirement model to output the job matching degree and ability gap analysis, specifically including: A rule-based weighted scoring method is adopted, and a set of ability dimensions is defined. The assessment values from different data sources and across different ability dimensions are weighted and fused to generate a radar chart of the student's comprehensive ability, using the following formula:
[0013] in, For the first Each data source in dimension Normalized values on This represents the confidence weight of the data source; The student's competency profile and job requirement model are mapped to vectors in the same high-dimensional feature space, and the cosine similarity algorithm is used to calculate the job matching degree between the two. When the job matching degree is lower than a preset threshold, the difference between the student's competency vector and the job requirement vector is analyzed to identify competency gaps, and learning path suggestions containing knowledge points and skills to be strengthened are generated based on the multi-layer dynamic graph.
[0014] Preferably, the cosine similarity algorithm is used to calculate the job matching degree between the two, as shown in the following formula:
[0015] in, For student ability vectors, As a vector of job requirements, Let Match_Rate be the angle between two vectors in space. When Match_Rate is lower than a set threshold, calculate the top k dimensions with the largest positive values in the vector difference. These are the student's weak points that need to be addressed, and a personalized recommendation improvement path is generated accordingly.
[0016] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements an AI teaching evaluation and job adaptation method based on multi-layer dynamic graph mapping as described above.
[0017] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: (1) The method of this invention achieves precise evaluation of industry-education integration. Compared with traditional methods, this invention directly links students' learning evaluation to the job competency model through a multi-layer dynamic graph, realizing a leap from "what they have learned" to "what jobs they can do." This invention achieves full-link data closure and multi-modal fusion, innovatively integrating multi-source data such as facial expressions, body movements, online trajectories, and practical training operations. In particular, by connecting with training systems such as drones and police systems, it solves the problem of separation between theoretical and practical data, and the generated comprehensive competency radar chart is more realistic and comprehensive.
[0018] (2) This invention achieves highly intelligent adaptive intervention. Through the AI private big model, it can automatically generate learning prescriptions, classroom rhythm suggestions and academic risk warnings, which greatly reduces the management burden of teachers and improves teaching efficiency and personalized training level. Attached Figure Description
[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0020] Figure 1 This is a flowchart illustrating the AI teaching evaluation and job adaptation method based on multi-layer dynamic graph mapping provided by the present invention. Detailed Implementation
[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0023] like Figure 1 As shown, the present invention provides an AI teaching evaluation and job adaptation method based on multi-layer dynamic graph mapping, comprising the following steps: S1. Based on natural language processing and knowledge graph technology, automatically construct a multi-layered dynamic graph from the industry chain to knowledge points; S2. Collect multimodal behavioral data of students in different scenarios through classroom video analysis, online learning behavior capture and integration with the training system; S3. Based on the collected multimodal behavioral data and multi-layer dynamic graph, the pre-trained large model is fine-tuned in the domain to generate student learning profiles and academic risk warning categories; S4. Integrate multi-source data to generate a radar chart of students' comprehensive abilities, and perform vectorized matching with job requirement models to output job matching degree and ability gap analysis. Finally, generate learning path suggestions that include strengthening knowledge points and skills.
[0024] Furthermore, in S1, the automated construction of a multi-layered dynamic graph from the industry chain to knowledge points specifically includes: Entity recognition: First, BERT pre-trained models are used to encode word vectors into massive amounts of teaching and industry data to obtain contextual semantic features; second, BiLSTM bidirectional long short-term memory networks are used to capture long-distance dependencies in the text; finally, CRF conditional random field layers are used for sequence labeling to extract job positions, skills, knowledge entities and the relationships between them. Keyword auto-extraction: The TF-IDF and TextRank algorithms are used to extract high-frequency keywords from industry documents, and Word2Vec is used to calculate word vector similarity to automatically remove common stop words and retain core skill vocabulary.
[0025] Specifically, the selection and storage architecture of the graph database are as follows: Database Selection: The system uses Neo4j, a native graph database, as its core storage engine, storing massive amounts of teaching data based on the LabeledPropertyGraph (LPG) model. Compared to traditional relational databases, Neo4j offers millisecond-level response times in handling multi-hop queries and path association analysis.
[0026] Query and computation: The graph traversal is performed using the Cypher query language, and the GraphDataScience (GDS) library is integrated to run graph algorithms such as PageRank (importance node mining) and Louvain (community discovery).
[0027] Graph Schema Definition and Core Node / Edge Attribute Design: To achieve accurate mapping from industry to education, the system defines a strict graph schema, with the core fields designed as follows: Core Node Attribute Design: Job node: includes Job_ID (unique identifier), Job_Name (job name, such as: penetration testing engineer), Company_Ref (company ID), Salary_Avg (average salary), Post_Hot (job popularity index), and Update_Time (fetching time).
[0028] Competency node: contains Comp_ID, Comp_Name (capability name, such as SQL injection defense), Type (type: hard skill / soft skill), and Complexity (complexity weight: 0-1).
[0029] Knowledge node: includes KP_ID, KP_Name (knowledge point name), Source_Chapter (corresponding textbook chapter), Difficulty (difficulty level), and Teaching_Target (teaching objective: memorization / understanding / application).
[0030] Student node: includes Stu_ID, Major, Learning_Style, and Realtime_Risk.
[0031] Core Relationships attribute design: Requirement Relationships: Connect job positions with skills. Attributes include Weight (indicating the importance of the skill to the job, 0.1-1.0) and Necessity (Must / Optional).
[0032] Precursor relationships (PREREQUISITE): Connect knowledge points to each other. Attributes include Strength (dependency strength) and Gap_Days (suggested learning interval in days).
[0033] Mastered Relationship: Connects students with skills / knowledge. Attributes include Proficiency_Score (Mastery score: 0-100), Last_Assess_Time (Last assessment time), and Data_Source (Classroom / Assignment / Practical Training).
[0034] Multi-layer graph logical hierarchy: Industry chain / industry node map: Define macro-level industry demand.
[0035] Job system map: Clearly defines the specific skill requirements for each job.
[0036] Ability / Knowledge / Problem / Ideological and Political Education System Map: Breaking down job requirements into specific teaching knowledge points and ability points.
[0037] AI-powered intelligent multi-layer graph construction engine: This engine has built-in entity recognition (NER) and relation extraction (RE) algorithms.
[0038] Automatic skill keyword extraction: The TF-IDF and TextRank algorithms are used to extract high-frequency keywords from industry documents, and Word2Vec is used to calculate word vector similarity. Common stop words are automatically removed, and core skill vocabulary is retained.
[0039] Automatic breakdown of course knowledge points: Using OCR technology to identify unstructured textbook content, paragraph clustering is performed through LDA topic modeling to automatically divide knowledge point levels.
[0040] Furthermore, in S2, through classroom video analysis, online learning behavior capture, and integration with the training system, multimodal behavioral data of students in different scenarios is collected, specifically including: High-precision face detection is performed using RetinaFace, and head pose is solved using the PnP (Perspective-n-Point) algorithm. OpenPose or YOLOv8-Pose skeletal keypoint detection algorithms are used to capture 18 key points of the human body in real time (including: top of head, center of forehead, left / right sides of head, left / right eyes, neck, upper thoracic vertebrae, middle spine, center of abdomen, left / right shoulders, left / right elbows, left / right wrists, and left / right fingertips). A spatiotemporal graph convolutional network (ST-GCN) is then constructed to identify action categories (such as raising hand, leaning on a table, and resting chin on hand). Facial Expression Recognition: The FER (Facial Expression Recognition) model, which is an improvement on ResNet-50, is used to classify facial expressions into 7 basic emotions (happy, neutral, sad, etc.) and output the emotion confidence vector.
[0041] Head-up ratio definition and calculation logic: The system calculates the Euler angles of the head in three-dimensional space using the PnP algorithm: pitch, yaw and roll.
[0042] Judgment criteria: When the following conditions are met and When the head is raised / focused, it is judged as a "head down / wandering" state; otherwise, it is judged as a "head down / wandering" state.
[0043] Head-up rate formula: = ,in The cumulative frame duration for determining the head-up state. This represents the total course duration.
[0044] Furthermore, S2 also includes: Students' classroom attention scores are calculated using a pre-set formula, as follows:
[0045] in, Head-up rate per unit time For the confidence score of positive sentiment extracted based on CNN, Frequency of fatigued movements The weighting coefficient of head-up rate per unit time. The weighting coefficients for positive sentiment confidence. This is the weighting coefficient for the frequency of fatigued actions.
[0046] Furthermore, in S3, based on the collected multimodal behavioral data and multi-layer dynamic graph, the pre-trained large model is fine-tuned in a specific domain, including: Generative pre-trained models based on the Transformer architecture are deeply adapted for the education vertical.
[0047] Base model selection: DeepSeekV3 (671B parameters) was chosen as the base model. These models possess powerful Chinese semantic understanding and reasoning capabilities, making them suitable for private deployment.
[0048] Privatization fine-tuning strategy: Training data composition: Construct a hybrid dataset containing three types of data (education-specific data, multimodal behavioral data, and multi-layer dynamic graphs). Unique data in the education field: covering courses, textbook PDFs, lesson plans Word documents, classroom recordings, experiments and teacher / student data, and professional papers (approximately 50GB), enhancing domain knowledge.
[0049] Instructions for fine-tuning data: Construct "question-reasoning-answer" pairs (e.g., "What are the possible reasons why a student failed three assignments in a row?", "According to the graph analysis, the student has a gap in the knowledge point of 'loop structure'..."), totaling 100,000 entries.
[0050] Time-series behavioral data: Transforming students' structured behavioral logs into natural language descriptions as contextual input.
[0051] Training Process and Techniques: LoRA low-rank adaptive fine-tuning technique is employed. The principal parameter weights W0 of the pre-trained model are frozen, and only the side-path low-rank matrices A and B are trained (W = W0 + ΔW = W0 + BA), significantly reducing memory usage loss function. Cross-entropy loss function modeled using causal language:
[0052] in, For the currently predicted token, Historical context, These are the model parameters. The system minimizes this loss function, enabling the model to accurately predict academic risks and recommend learning paths.
[0053] Furthermore, in S4, multi-source data is integrated to generate a radar chart of students' comprehensive abilities, which is then vectorized and matched with a job requirement model to output job matching degree and ability gap analysis, specifically including: The Vector Space Model (VSM) is used to perform quantitative calculations of person-job matching; a rule-based weighted scoring method is adopted, and a set of ability dimensions is defined. The assessment values from different data sources and across different ability dimensions are weighted and fused to generate a radar chart of the student's comprehensive ability, using the following formula:
[0054] in, For the first Data sources (such as final exams, practical training grades, and classroom performance) are categorized into dimensions. Normalized values on Assign confidence weights to the data source (preset by the AHP analytic hierarchy process, for example, training data has a higher weight than regular homework). The student's competency profile and job requirement model are mapped to vectors in the same high-dimensional feature space, and the cosine similarity algorithm is used to calculate the job matching degree between the two. When the job matching degree is lower than a preset threshold, the difference between the student's competency vector and the job requirement vector is analyzed to identify competency gaps, and learning path suggestions containing knowledge points and skills to be strengthened are generated based on the multi-layer dynamic graph.
[0055] Furthermore, the "student ability model" and "job requirement model" are mapped to vectors in the same multidimensional feature space, and the cosine similarity algorithm is used to calculate the job matching degree between the two. The closer the value is to 1, the higher the matching degree. The formula is as follows:
[0056] in, For student ability vectors, As a vector of job requirements, The angle between two vectors in space is used to calculate the top k dimensions with the largest positive values in the vector difference when Match_Rate is lower than a set threshold (e.g., 0.8). These dimensions represent the student's weak points that need improvement, and personalized recommendation improvement paths are generated accordingly.
[0057] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements an AI teaching evaluation and job adaptation method based on multi-layer dynamic graph mapping as described above.
[0058] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.
[0059] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.
Claims
1. A method for AI-based teaching evaluation and job adaptation based on multi-layer dynamic graph mapping, characterized in that, Includes the following steps: S1. Based on natural language processing and knowledge graph technology, automatically construct a multi-layered dynamic graph from the industry chain to knowledge points; S2. Collect multimodal behavioral data of students in different scenarios through classroom video analysis, online learning behavior capture and integration with the training system; S3. Based on the collected multimodal behavioral data and multi-layer dynamic graph, the pre-trained large model is fine-tuned in the domain to generate student learning profiles and academic risk warning categories; S4. Integrate multi-source data to generate a radar chart of students' comprehensive abilities, and perform vectorized matching with job requirement models to output job matching degree and ability gap analysis. Finally, generate learning path suggestions that include strengthening knowledge points and skills.
2. The AI teaching evaluation and job adaptation method based on multi-layer dynamic graph mapping according to claim 1, characterized in that, In S1, the automated construction of a multi-layered dynamic graph from the industry chain to knowledge points specifically includes: Entity recognition: First, BERT pre-trained models are used to encode word vectors into massive amounts of teaching and industry data to obtain contextual semantic features; second, BiLSTM bidirectional long short-term memory networks are used to capture long-distance dependencies in the text; finally, CRF conditional random field layers are used for sequence labeling to extract job positions, skills, knowledge entities and the relationships between them. Keyword auto-extraction: The TF-IDF and TextRank algorithms are used to extract high-frequency keywords from industry documents, and Word2Vec is used to calculate word vector similarity to automatically remove common stop words and retain core skill vocabulary.
3. The AI teaching evaluation and job adaptation method based on multi-layer dynamic graph mapping according to claim 1, characterized in that, In step S2, multimodal behavioral data of students in different scenarios is collected through classroom video analysis, online learning behavior capture, and integration with the practical training system. Specifically, this includes: Face detection is performed using the RetinaFace model, combined with the PnP algorithm to detect the Euler angles of students' heads, and the head-up rate in class is calculated based on the Euler angles, which include pitch, yaw, and roll angles. A facial expression recognition model based on ResNet-50 is used to identify students' emotion categories and confidence levels. OpenPose or YOLOv8-Pose skeletal keypoint detection algorithms are used to capture 18 key points of the human body in real time, and a spatiotemporal graph convolutional network is constructed to identify action categories.
4. The AI teaching evaluation and job adaptation method based on multi-layer dynamic graph mapping according to claim 3, characterized in that, S2 also includes: Students' classroom attention scores are calculated using a pre-set formula, as follows: in, Head-up rate per unit time For the confidence score of positive sentiment extracted based on CNN, Frequency of fatigued movements The weighting coefficient for head-up rate per unit time. The weighting coefficients for positive sentiment confidence. This is the weighting coefficient for the frequency of fatigued actions.
5. The AI teaching evaluation and job adaptation method based on multi-layer dynamic graph mapping according to claim 1, characterized in that, In step S3, based on the collected multimodal behavioral data and multi-layer dynamic graph, the pre-trained large model is fine-tuned in the domain, specifically including: The pre-trained large model is a base model using the Transformer architecture. It uses LoRA technology for low-rank fine-tuning. By freezing the weights of the pre-trained large model and training only the low-rank matrix, it injects education-specific data, multimodal behavioral data, and multi-layer dynamic graphs into the pre-trained large model, enabling it to have the ability to reason in the education domain.
6. The AI teaching evaluation and job adaptation method based on multi-layer dynamic graph mapping according to claim 1, characterized in that, In step S4, multi-source data is integrated to generate a radar chart of students' comprehensive abilities, and this chart is then vectorized and matched with a job requirement model to output job matching degree and ability gap analysis, specifically including: A rule-based weighted scoring method is adopted, and a set of ability dimensions is defined. The assessment values from different data sources and across different ability dimensions are weighted and fused to generate a radar chart of the student's comprehensive ability, using the following formula: in, For the first Each data source in dimension Normalized values on This represents the confidence weight of the data source; The student's competency profile and job requirement model are mapped to vectors in the same high-dimensional feature space, and the cosine similarity algorithm is used to calculate the job matching degree between the two. When the job matching degree is lower than a preset threshold, the difference between the student's competency vector and the job requirement vector is analyzed to identify competency gaps, and learning path suggestions containing knowledge points and skills to be strengthened are generated based on the multi-layer dynamic graph.
7. The AI teaching evaluation and job adaptation method based on multi-layer dynamic graph mapping according to claim 6, characterized in that, The cosine similarity algorithm is used to calculate the job matching degree between the two, and the formula is as follows: in, For student ability vectors, As a vector of job requirements, Let Match_Rate be the angle between two vectors in space. When Match_Rate is lower than a set threshold, calculate the top k dimensions with the largest positive values in the vector difference. These are the student's weak points that need to be addressed, and a personalized recommendation improvement path is generated accordingly.
8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements an AI teaching evaluation and job adaptation method based on multi-layer dynamic graph mapping as described in any one of claims 1-7.