A knowledge graph topology structure dynamic evolution method, system and device

Abstract

This invention discloses a method, system, and device for the dynamic evolution of knowledge graph topology. The method includes: performing anomaly detection for group learning to identify defects in the graph structure; upon detecting anomalies, triggering a reflective multi-agent team workflow to perform attribution analysis on the anomalous knowledge point nodes and generate missing hypothesis; without disrupting the original global graph topology, dynamically executing a graph database query language using a native graph database engine to reconstruct the local graph topology and generate isolated shadow branches; for the newly generated local graph topology, introducing a gray-scale release mechanism for gray-scale verification; if verification passes, the new topology is globally implemented; if verification fails, the topology is rolled back and the shadow branches are deleted. This invention aims to endow educational knowledge graphs with the ability of "self-healing and self-growth," improving the global scientific rigor and success rate of adaptive learning paths.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of artificial intelligence, digital education, knowledge graphs, and adaptive learning. Specifically, it relates to a method, system, and device for dynamically evolving the knowledge graph topology by utilizing group learning behavior data and multi-agent deep reasoning mechanisms to automatically discover defects in the topology of educational knowledge graphs, generate remedial hypotheses, and perform gray-scale verification and dynamic reconstruction. Background Technology

[0002] In adaptive learning systems and intelligent tutoring systems (ITS), knowledge graphs are the core foundation for cognitive diagnosis and path planning. However, existing knowledge graph construction techniques (whether manually constructed or generated in a single step using a large model) suffer from the following serious inherent static defects: 1. The "Curse of Knowledge" and the Lack of Implicit Prerequisite Knowledge: Knowledge graphs are typically generated by domain experts or large models based on standard teaching syllabi. Experts often suffer from the "curse of knowledge" (difficulty in empathizing with the cognitive blind spots of beginners), causing the graph to easily omit certain extremely basic but crucial "implicit prerequisite knowledge points." For example, when learning "L'Hôpital's rule," the root cause of many students' failures may be their lack of mastery of the prerequisite knowledge point "simplification of complex fractions," but this prerequisite node is not included in the graph.

[0003] 2. Lack of data feedback and self-correction capabilities: Although traditional systems collect massive amounts of user test-taking data, this data is only used to "evaluate users" (i.e., score students) and never used to "evaluate and repair the learning path itself." When a certain learning path has an abnormally high failure rate for a long period of time, traditional systems will only make students "blindly do practice problems" at that node repeatedly, and will not be able to discover and correct "unreasonable pre-dependencies" at the structural level.

[0004] 3. Fixed graph structure and high update risk: Even if defects in the graph are discovered manually, global modifications to the graph topology (adding or deleting nodes and edges) can easily cause system logic chaos, and there is a lack of a smooth evolution mechanism for safe verification (A / B testing) in a real teaching environment. Summary of the Invention

[0005] The present invention provides a method, system and device for dynamic evolution of knowledge graph topology based on group learning feedback that can overcome the defect that existing static knowledge graphs cannot self-correct, and can at least solve one of the above-mentioned technical problems.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: A method for dynamic evolution of knowledge graph topology includes the following steps: S1. Initiate anomaly detection for group learning, aiming to discover abnormal defects in the graph topology of group learning; S2. After detecting an anomaly in the graph topology, trigger the reflective multi-agent team workflow to perform attribution analysis on the abnormal nodes in the graph topology and generate missing hypothesis results. S3, the native graph database engine, without destroying the original global graph topology, reconstructs the local graph topology based on the missing hypothesis results and generates isolated shadow branches. S4. For the newly generated local graph topology, a gray-scale release mechanism is introduced for gray-scale verification. If the verification passes, the new topology will be globally effective. If the verification fails, the topology will be rolled back and the shadow branch will be deleted.

[0007] Furthermore, in S1, for group learning, the micro-course knowledge graph stored in the native graph database is continuously monitored, the micro-course knowledge graph is structurally transformed to obtain the micro-course knowledge graph topology, the dynamic user learning progress attributes of each knowledge point node in the graph topology are aggregated and extracted, and the group cognitive diagnosis failure rate of any knowledge point node is calculated. If, for a given knowledge point node, the group cognitive diagnosis failure rate of 10% or more users exceeds a preset threshold, and the group cognitive diagnosis failure rate of the preceding knowledge point nodes is lower than the preset threshold, then the knowledge point node is determined to have a structural cognitive gap and is identified as an abnormal knowledge point node. At the same time, a graph evolution mechanism for local graph topology reconstruction is triggered for the abnormal knowledge point node.

[0008] Furthermore, in S1, for knowledge point nodes... Group cognitive diagnosis failure rate The calculation expression is:

[0009] in, Indicates at the knowledge point node The total number of users who generate learning behavior. Indicates user index, Indicates the first Individual users at knowledge point nodes Cognitive diagnostic scores on This indicates the preset passing or mastery threshold. This is an indicator function, representing a value of 1 when the condition is met, i.e., the score is below the threshold, and a value of 0 otherwise.

[0010] Furthermore, in S2, the reflective agent team includes an attribution agent and a hypothesis-generating agent, and the reflective agent team workflow includes: S2.1 Extract the text of incorrect questions, the text of solution steps, and the dialogue interaction logs of users who failed at abnormal knowledge point nodes, and use them as data input; S2.2 Attribution agent uses the deep semantic analysis capabilities of large language models to cluster and extract common features from massive amounts of incorrect question texts, find the underlying error logic that leads to the failure of a group of users, and obtain the attribution results; S2.3 Assume that the generated agent infers the implicit prior knowledge points missing from the abnormal knowledge point nodes based on the attribution results.

[0011] Furthermore, in S3, the local graph topology is dynamically reconstructed, and isolated shadow branches are generated, specifically including: S3.1 For the abnormal knowledge point node A, insert the implicit prerequisite knowledge point node A0 generated by the hypothesis generating agent into the micro-course knowledge graph. The implicit prerequisite knowledge point node A0 contains the knowledge point name and the corresponding exercise. S3.2 Modify the edge pointing relationship in the micro-course knowledge graph. If the original edge pointing relationship is B→A, then generate a parallel test path B→A0→A in the original graph database and regenerate the local graph topology structure. Here, the edge represents the logical dependency relationship between knowledge point nodes.

[0012] Furthermore, the closed-loop verification process in S4 further includes: S4.1 When a new user reaches knowledge point node B, the traffic is distributed according to the set traffic ratio. A portion of the new users are directed to a new exploration path, i.e., B→A0→A, while the remaining new users continue to follow the original path, i.e., B→A. S4.2 Track and compare the first pass rate and cognitive diagnosis score of the two groups of new users on the final target knowledge point node A in real time, and perform a significance test on the knowledge point node A. S4.3. Let the group that passes through the implicit prior knowledge point node A0 be the gray group. If the mastery of the gray group on knowledge point node A exceeds the preset threshold, then the newly generated local graph topology structure is officially applied globally, and the old edge B→A is deleted. If it does not exceed the preset threshold, then the hypothesis of the agent generated by the hypothesis is determined to be invalid, the topology is rolled back, and the shadow branch, i.e., the implicit prior knowledge point node A0, is deleted.

[0013] Furthermore, in S4, the first-pass rate is the number of first-time passers divided by the total number of attempts. The cognitive diagnostic score is obtained by the large language model, and the significance test calculation formula is:

[0014] in, The Z-statistic, or standard score, represents the significance test. and These are the average scores of the test group (which passed through the new knowledge point node A0) and the control group (which followed the original path) at the final target knowledge point node A, respectively. and Let V be the variance of the two sets of scores. and These are the sample sizes for the two groups; pass The degree of mastery corresponding to the value calculation value:

[0015] in, The cumulative distribution function of the standard normal distribution. For integration variables; When mastering Less than the preset threshold When the newly generated local graph topology is deemed valid, the new topology is officially applied globally.

[0016] A dynamic evolution system for knowledge graph topology, applicable to the dynamic evolution method for knowledge graph topology, adds an evolution control center to the two-layer adaptive learning knowledge graph system; The two-layer adaptive learning knowledge graph system includes an AI engine layer, a hybrid data persistence layer, a presentation layer, and an application service layer. The evolutionary control center includes: An anomaly monitoring engine periodically performs batch queries on the native graph database, calculates the failure rate of group cognitive diagnosis for each node based on aggregation functions, and identifies high-frequency abnormal nodes. The reflective agent team, which is connected to the anomaly monitoring engine and is independent of the multi-agent team in the AI ​​engine layer, is used for error analysis, knowledge point reasoning, and generation of evolution graph database query language code. The gray-scale routing and evaluation module is connected to the reflective agent team. It maintains the path distribution strategy at the application service layer, records the comparison data of the gray-scale release mechanism, and performs significance tests.

[0017] Furthermore, the AI ​​engine layer encapsulates the client based on the AutoGen framework, directly interacts with the deep inference large language model, and the AI ​​engine layer adopts a polling group dialogue mode for orchestration and collaboration. The hybrid data persistence layer is connected to the AI ​​engine layer and includes a relational database and a native graph database. The relational database is used to store the graph attribute metadata of the macro career path graph and the micro course knowledge graph. The native graph database is used to store the structured graph data of the macro career path graph and the micro course knowledge graph, and the two levels of graphs are isolated by the node labels in the structured graph data. The presentation layer, as the front-end interface, is built on React and connected to the hybrid data persistence layer. It is used to retrieve data from the back-end and provide users with an intuitive knowledge graph visualization and interactive interface. The application service layer, which serves as the backend service framework, interacts with the three layers mentioned above to manage, process, invoke, and distribute data.

[0018] A computer device includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the above-described method for dynamic evolution of knowledge graph topology.

[0019] The beneficial effects of this invention are reflected in: 1. Break away from expert dependence and endow the map with self-growth capability: no longer relying on manual periodic review and updates, but able to "learn from mistakes" in teaching practice, and independently discover and fill in implicit knowledge points not explicitly stated in textbooks or syllabi.

[0020] 2. Transforming “Individual Failure” into “Group Dividend”: In traditional methods, a user’s failure is just individual data. However, in this invention, the failure data of the first batch of users will be transformed into nutrients for a complete graph, enabling subsequent users to automatically avoid the cognitive trap and achieve true collective intelligence co-construction.

[0021] 3. Safe and reliable gray-scale trial and error mechanism: By constructing parallel shadow paths in the native graph database for gray-scale verification, it is ensured that the "erroneous assumptions" that the Large Language Model (LLM) may generate will not disrupt the global teaching order, thus guaranteeing extremely high engineering safety and fault tolerance in the evolution process.

[0022] 4. Deep integration of dynamic attributes and graph topology: It fully leverages the architectural advantages of storing user data directly as graph node attributes, so that "group data statistics" can be completed without complex cross-table join queries, but only with low-latency Cypher attribute aggregation operations, supporting real-time topology evolution. Attached Figure Description

[0023] The accompanying drawings, which are provided to further understand this application and constitute a part of this application, illustrate exemplary embodiments of this application and are used to explain this application, but do not constitute an undue limitation of this application.

[0024] Figure 1 This is a schematic diagram of the overall process of the dynamic evolution method of knowledge graph topology structure according to an embodiment of the present invention.

[0025] Figure 2 This is a detailed flowchart illustrating the dynamic evolution method of knowledge graph topology in an embodiment of the present invention.

[0026] Figure 3 This is a schematic diagram of the reflective multi-agent team architecture and attribution analysis data flow according to an embodiment of the present invention.

[0027] Figure 4 This is a schematic diagram of native graph database topology reconstruction and grayscale routing in an embodiment of the present invention. Detailed Implementation

[0028] 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 a part of the embodiments of the present invention, and not all of them. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. 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.

[0029] It should be noted that the meaning of "and / or" throughout the text includes three parallel solutions. Taking "A and / or B" as an example, it includes solution A, solution B, or a solution that simultaneously satisfies A and B. Furthermore, "multiple" refers to two or more. Additionally, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0030] This invention provides a dynamic evolution system for knowledge graph topology, which adds an evolution control center to a two-layer adaptive learning knowledge graph system.

[0031] The dual-layer adaptive learning knowledge graph system comprises an AI engine layer, a hybrid data persistence layer, a presentation layer, and an application service layer, specifically: The AI ​​engine layer encapsulates the client based on the AutoGen framework, directly interacts with deep inference large language models (such as Deepseek-reasoner), and is configured with different multi-agent teams (including agents such as search agents, summarizing agents, planning agents, and graph building agents) with different System Messages and Function Tools. The AI ​​engine layer also uses a round-robin group chat mode to orchestrate collaboration. The hybrid data persistence layer is connected to the AI ​​engine layer and includes a relational database (SQLite) and a native graph database (Neo4j). The relational database is used to store graph attribute metadata (such as graph ID, name, parent-child relationship, etc.) of the macro career path graph and the micro course knowledge graph. The native graph database is used to store the structured graph data of the macro career path graph and the micro course knowledge graph, and the two levels of graphs are isolated by the node labels in the structured graph data. In the structured graph data generated by the macro-career path graph using the graph database query language (Cypher script), node labels are defined as courses, and edge labels are defined as prerequisite dependencies between courses. In the structured graph data generated by the micro-course knowledge graph using the graph database query language, node labels are defined as knowledge points, and edge labels are defined as logical dependencies between knowledge points. "Prerequisite dependencies between courses" is a core concept in the education curriculum system. It defines the knowledge base that must be mastered before learning a specific course. It usually exists in the form of "course A is a prerequisite for course B". This relationship not only directly affects the formulation of teaching plans and the planning of students' learning paths, but is also a classic application scenario of graph theory algorithms in computer science. It is often used to detect circular dependencies in course arrangement and generate feasible learning sequences. The logical dependencies between knowledge points are similar to the prerequisite dependencies between courses mentioned above, and are also a core element in the construction of a knowledge system. They are mainly divided into the following four categories: 1) Progressive relationship One knowledge point is the foundation of another; you must master the former before you can understand the latter. For example, in mathematics, the derivation of theorems requires understanding the definitions and axioms before you can apply the theorems. 2) Relationship Although the knowledge points are not directly dependent on each other, they complement each other. Understanding one helps to deepen the understanding of another, such as the relationship between historical events and cultural background. 3) Organizational relationships (part-whole) In a knowledge base, knowledge points are organized in a hierarchical structure, such as the relationship between chapters and sub-chapter in a subject framework; 4) Logical dependencies (support relationships) There is a logical order between the knowledge points, and they need to be learned in a specific order, such as the dependency order of function calls in programming; The presentation layer, as the front-end interface, is built on React and connected to the hybrid data persistence layer. It is used to retrieve data from the back-end and provide users with an intuitive knowledge graph visualization and interactive interface. The application service layer, as a backend service framework, interacts with the three layers mentioned above to manage, process, call, and distribute data (such as managing user logins, reading data from the hybrid data persistence layer based on user operations sent from the front end, and submitting tasks such as building graphs to the AI ​​engine layer).

[0032] In the dynamic evolution system of the knowledge graph topology, the evolution control center includes an anomaly monitoring engine, a reflective agent team, and a gray-scale routing and evaluation module, specifically: The anomaly monitoring engine periodically performs batch queries on the native graph database, calculates the attribute distribution of each node based on the aggregation function (calculation of the failure rate of group cognitive diagnosis), and identifies high-frequency abnormal nodes. The reflective agent team is connected to the anomaly monitoring engine and is independent of the multi-agent team in the AI ​​engine layer. The reflective agent team includes an attribution agent and a hypothesis generation agent, which are used for error analysis, knowledge point reasoning, and generation of evolution graph database query language code. The gray-scale routing and evaluation module is connected to the reflective agent team, maintains the path distribution strategy at the application service layer, records the comparison data of the gray-scale release mechanism (A / B testing), and performs significance testing (P-value analysis).

[0033] It should be noted that the "intelligent agent" involved in this invention refers to an intelligent entity in the field of artificial intelligence that can perceive the environment through sensors (i.e., tool invocation), make autonomous decisions with its own intelligence, and act on the environment through actuators. Its core features are autonomy, responsiveness, proactiveness, and sociality. It can proactively plan and complete goals without continuous human intervention. It can be a software form of AI assistant or intelligent customer service, or a hardware entity such as autonomous driving or robot. Currently, intelligent agents based on large models have the ability to remember, invoke tools, and process complex tasks.

[0034] See Figures 1-2 This invention provides a method for the dynamic evolution of knowledge graph topology, demonstrating a closed loop from anomaly detection → hypothesis generation → dynamic reconstruction → grayscale verification. The method is implemented using the aforementioned dynamic evolution system for knowledge graph topology and includes the following steps: S1. Initiate anomaly detection for group learning, aiming to discover abnormal defects in the graph topology of group learning; S2. After detecting an anomaly in the graph topology, the Reflective Agents workflow is triggered to perform attribution analysis on the abnormal nodes in the graph topology and generate missing hypothesis results. S3, the native graph database engine, without destroying the original global graph topology, reconstructs the local graph topology based on the missing hypothesis results and generates isolated shadow branches. S4. For the newly generated local graph topology, a gray-scale release mechanism (A / B testing) is introduced for gray-scale verification. If the verification passes, the new topology will be globally effective. If the verification fails, the topology will be rolled back and the shadow branches will be deleted.

[0035] In this embodiment, in step S1, for group learning, the micro-course knowledge graph stored in the native graph database (such as Neo4j) is continuously monitored, the micro-course knowledge graph is structurally transformed to obtain the micro-course knowledge graph topology, the dynamic user learning progress attributes (such as p_user1, p_user2) on each knowledge point node in the graph topology are aggregated and extracted, and the group cognitive diagnosis failure rate on any knowledge point node is calculated. If, for a certain knowledge point node, the group cognitive diagnosis failure rate of 10% or more users exceeds the preset threshold, and the group cognitive diagnosis failure rate of the preceding knowledge point nodes is lower than the preset threshold, then the knowledge point node is determined to have a structural cognitive gap and is judged as an abnormal knowledge point node. At the same time, a graph evolution mechanism for local graph topology reconstruction is triggered for the abnormal knowledge point node.

[0036] In this embodiment, in step S1, the knowledge point nodes are... Group cognitive diagnosis failure rate The calculation expression is:

[0037] in, Indicates at the knowledge point node The total number of users who generate learning behavior. Indicates user index, Indicates the first Individual users at knowledge point nodes Cognitive diagnostic scores on This indicates the preset passing or mastery threshold. This is an indicator function, representing a value of 1 when the condition is met, i.e., the score is below the threshold, and a value of 0 otherwise.

[0038] See Figure 3In this embodiment, the reflective agent team workflow in step S2 includes: S2.1 Extract the text of incorrect questions, the text of solution steps, and the dialogue interaction logs of users who failed at abnormal knowledge point nodes, and use them as data input; S2.2 Attribution Agent: Based on the deep semantic analysis capabilities of a large language model, it clusters and extracts common features from massive amounts of incorrect question texts to find the underlying error logic that leads to the failure of a group of users and obtains the attribution results. S2.3 Hypothesis Agent: Based on the attribution results, the Hypothesis Agent infers the implicit prior knowledge node missing from the abnormal knowledge node, denoted as A0. For example, it is assumed that "before learning knowledge node A, knowledge node A0 must be learned first".

[0039] See Figure 4 In this embodiment, step S3, which involves dynamic reconstruction of the local graph topology and generation of isolated shadow branches, specifically includes: S3.1 For the abnormal knowledge point node A, insert the implicit prerequisite knowledge point node A0 generated by the hypothesis generating agent into the micro-course knowledge graph. The implicit prerequisite knowledge point node A0 contains the knowledge point name and the corresponding exercise. S3.2 Modify the edge pointing relationship in the micro-course knowledge graph. If the original edge pointing relationship is B→A, then generate a parallel test path B→A0→A in the original graph database and regenerate the local graph topology structure. Here, the edge represents the logical dependency relationship between knowledge point nodes.

[0040] See Figure 4 In this embodiment, the closed-loop verification process in S4 further includes: S4.1 When a new user reaches knowledge point node B, the traffic is distributed according to the set traffic ratio. A portion of the new users are directed to a new exploration path, i.e., B→A0→A, while the remaining new users continue to follow the original path, i.e., B→A. S4.2. Track and compare the first pass rate and cognitive diagnosis score of the two groups of new users on the final target knowledge point node A in real time, and perform a significance test on knowledge point node A. S4.3. Let the group that passes through the implicit prior knowledge point node A0 be the gray group. If the mastery of the gray group on knowledge point node A exceeds the preset threshold, then the newly generated local graph topology structure is officially applied globally, and the old edge B→A is deleted. If it does not exceed the preset threshold, then the hypothesis of the agent generated by the hypothesis is determined to be invalid, the topology is rolled back, and the shadow branch, i.e., the implicit prior knowledge point node A0, is deleted.

[0041] In this embodiment, in step S4, the first-pass rate is the number of first-time passers divided by the total number of attempts. The cognitive diagnostic score is obtained by the large language model, and the significance test calculation formula is:

[0042] in, The Z-statistic, or standard score, represents the significance test. and These are the average scores of the test group (which passed through the new knowledge point node A0) and the control group (which followed the original path) at the final target knowledge point node A, respectively. and Let V be the variance of the two sets of scores. and These are the sample sizes for the two groups; pass The degree of mastery corresponding to the value calculation value:

[0043] in, The cumulative distribution function of the standard normal distribution. For integration variables; When mastering Less than the preset threshold When the newly generated local graph topology is deemed valid, the new topology is officially applied globally.

[0044] To further verify the feasibility of this method, the present invention will be illustrated with the following practical application examples.

[0045] Example 1: Specific Implementation of the Group Anomaly Detection Mechanism For micro-level course knowledge graphs, a dynamic evolution system for the knowledge graph topology is set to trigger a monitoring script every morning.

[0046] 1. Data Aggregation: For any node Node_A in the graph, the system performs a Cypher query using Neo4j to match all dynamic attributes of the form p_user*. The group cognitive diagnosis failure rate F(A) for this node is calculated using the following formula:

[0047] in, Indicates at the knowledge point node The total number of users who generate learning behavior. Indicates user index, Indicates the first Individual users at knowledge point nodes Cognitive diagnostic scores on This represents the preset passing or mastery threshold, which is set to 60 points in this embodiment. This is an indicator function, representing a value of 1 when the condition is met, i.e., the score is below the threshold, and a value of 0 otherwise.

[0048] 2. Failure rate calculation: If there are more than 100 user attribute records on node_A, and the proportion of records with scores below 60 is more than 45% (i.e., failure rate > 45%), the system will further query the scores of these failed users on the predecessor node Node_B of node_A.

[0049] 3. Trigger an alarm: If the average score on the predecessor node Node_B is >85, it indicates that "the prerequisite knowledge has been mastered, but a large-scale failure has occurred on the current node", which is a "structural anomaly". The system will mark node Node_A as Status: Needs_Evolution and suspend the evolution task.

[0050] Pseudocode for calculation: / / Match the preceding node B with the target node A, and aggregate user scores to calculate the failure rate and the average score of the preceding nodes. MATCH (b:KnowledgeNode)-[:PRE_REQUISITE_FOR]->(a:KnowledgeNode {id: 'Node_A'}) MATCH (u:User)-[r_a:LEARNED]->(a) MATCH (u)-[r_b:LEARNED]->(b) WITH a, b, COUNT(u) AS total_users, SUM(CASE WHEN r_a.score<60 THEN 1 ELSE 0 END) * 1.0 / COUNT(u) ASfail_rate, AVG(r_b.score) AS pre_avg_score WHERE total_users>100 AND fail_rate>0.45 AND pre_avg_score>85 SET a.Status = 'Needs_Evolution' RETURN a.id, fail_rate, pre_avg_score Example 2: Collaborative Reasoning Process of Reflective Agents Taking the anomaly at a node in the "Law of Rigid Body Rotation about a Fixed Axis" in university physics as an example: 1. Error Extraction: The system extracts the interactive text of the solutions from 50 students who answered incorrectly in the "Law of Rigid Body Rotation about a Fixed Axis".

[0051] 2. Attribution Agent Intervention: Inputting incorrect examples into a large language model, the model analysis reveals that over 70% of students do not actually lack understanding of physics formulas. It is not calculating the moment of inertia. At that time, errors occurred in the integral calculation for complex shapes.

[0052] 3. Hypothesis Agent's deduction: Upon receiving the attribution report, the large language model determines that the current micro-knowledge graph lacks the mathematical prerequisite of "complex geometric volume integration techniques" before "fixed-axis rotation law". Hypothesis Agent automatically generates a new knowledge node name "rotational inertia integral project" and calls the graph generation tool to generate JSON data for constructing the node and its accompanying basic exercises.

[0053] pseudocode: / / Pseudocode for a Reflection on the Evolutionary Workflow of Multi-Agent Teams Input: error_logs_A / / 50 error logs for the "Rigid Body Rotation Laws About a Fixed Axis" node Input: Node_A_prerequisites / / The set of currently known predecessor nodes Function Reflective_Evolution_Process(error_logs_A, Node_A_prerequisites): / / Step 1: Attribution Agent intervenes in deep semantic analysis prompt_attribution = "Analyze the following physics error log, extract non-formula memorization errors, and find underlying mathematical / logical gaps..." root_cause = LLM_Attribution.analyze(prompt_attribution, error_logs_A) / / Analysis results: root_cause = "70% of students cannot correctly calculate the integral of rotational inertia for complex shapes" / / Step 2: Hypothesis Agent generates hidden predecessor nodes If root_cause is clearly_identified: prompt_hypothesis = "Based on the root cause of the error, infer the missing transitional knowledge points between the known predecessor node and the target node." missing_knowledge_A0 = LLM_Hypothesis.infer(root_cause, Node_A_prerequisites) / / Deduction result: Missing knowledge point A0 = "Integral of rotational inertia" / / Step 3: The content generation agent outputs executable new node data new_node_json = LLM_ContentGen.create_curriculum(missing_knowledge_A0) return new_node_json / / Returns a dataset containing names, concept explanations, and accompanying exercises. End If return Null Example 3: Implementation of Topology Reconstruction and Cypher Gray-Scale Verification 1. Dynamic Insertion: After the reflective agent team generates a new node A0 (rotational inertia integral project), the native graph database engine dynamically executes the Cypher script to perform local reconstruction, creating a new node A0 in Neo4j and creating the relation TEST_PRE_REQUISITE with a specific test prefix. Note that the original standard dependency relation (b)-[:PRE_REQUISITE_FOR]->(a) is still retained, ensuring that the global topology is not destroyed.

[0054] The pseudocode for the core Cypher execution logic is as follows: / / Match the original start point B and end point A MATCH (b:KnowledgeNode {name: 'Fundamentals of Rigid Body Kinematics'}) MATCH (a:KnowledgeNode {name: 'Law of Rigid Body Rotation about a Fixed Axis'}) / / Create a new supplementary knowledge blind spot node A0 CREATE (a0:KnowledgeNode {name: 'Integral Project of Moment of Inertia', status: 'TESTING'}) / / Construct parallel shadow test paths CREATE (b)-[:TEST_PRE_REQUISITE]->(a0)-[:TEST_PRE_REQUISITE]->(a) 2. Gray-scale traffic generation: When it is necessary to recommend a path to new users, random sampling is performed. If 20% of the test group is hit, the backend path planning algorithm will prioritize identifying the TEST_PRE_REQUISITE relationship when traversing the graph and guide the user to the A0 node for learning; the 80% control group ignores the TEST relationship and continues to traverse the original edge (B)-[:PRE_REQUISITE_FOR]->(A).

[0055] pseudocode: Function Path_Routing(user_u, current_node_B, target_node_A): / / Generate a random number between 0.0 and 1.0 rand_val = generate_random() If rand_val <= 0.20: / / Hit 20% of the test group, prioritize identifying test edges path = GraphDB.find_path(current_node_B, target_node_A, edge_type="TEST_PRE_REQUISITE") user_u.assign_group("TEST") Else: / / 80% control group, following standard procedure path = GraphDB.find_path(current_node_B, target_node_A, edge_type="PRE_REQUISITE_FOR") user_u.assign_group("CONTROL") return path 3. Result Convergence and Structure Consolidation: After the system has accumulated a preset number (e.g., 200) of valid grayscale samples, the evaluation module automatically triggers a result convergence determination. To avoid misjudgments caused by the randomness of small samples, the system not only compares the absolute mean scores but also performs a statistical significance test (Z-test) on the diagnostic scores of the two groups of users at the final target node A. The calculation formula is as follows:

[0056] in, The Z-statistic, or standard score, represents the significance test. and These are the average scores of the test group (which passed through the new knowledge point node A0) and the control group (which followed the original path) at the final target knowledge point node A, respectively. and Let V be the variance of the two sets of scores. and These are the sample sizes for the two groups; The system's decision-making logic is as follows: assuming the Z-value is calculated using the above formula, and then converted into the corresponding level of mastery. value:

[0057] in, The cumulative distribution function of the standard normal distribution. For integration variables; If you have a good grasp of the degree , The threshold is set, meaning there is a statistically significant difference, and The system determines that the graph evolution is successful and the newly generated local graph topology is valid. The new topology is officially applied globally. The system deletes the old PRE_REQUISITE_FOR relationship and renames the TEST_PRE_REQUISITE relationship to the official PRE_REQUISITE_FOR, thus establishing B→A0→A as the global standard teaching path. Conversely, if If the assumption of the agent is invalid, the system rolls back, deletes node A0, and restores the original topology.

[0058] This invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the above-described knowledge graph topology dynamic evolution method.

[0059] This invention also provides a computer device, including a memory and a processor. The memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the knowledge graph topology dynamic evolution method described above.

[0060] This invention also provides a computer program product containing instructions that, when run on a computer, causes the computer to execute the steps of the above-described knowledge graph topology dynamic evolution method.

[0061] It is understood that the systems, devices, and storage media provided in the embodiments of the present invention correspond to the methods provided in the embodiments of the present invention, and the explanations, examples, and beneficial effects of the relevant content can be referred to the corresponding parts of the above-mentioned dynamic evolution method of knowledge graph topology.

[0062] It should be noted that those skilled in the art will understand that all or part of the steps implemented in the embodiments of the present invention can be implemented entirely or partially by software, hardware, firmware, or any combination thereof. When implemented in hardware, it can be implemented entirely or partially by purchasing standard parts or modifications. When implemented in software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid state disks (SSDs)).

[0063] In summary, this invention overcomes the limitation of existing static knowledge graphs in their inability to self-correct by providing a method and system for the dynamic evolution of knowledge graph topology based on group learning feedback. This method and system aim to endow educational knowledge graphs with the ability of "self-healing and self-growth." By monitoring the failure rate of group cognitive diagnosis, utilizing reflective multi-agent teams to uncover implicit knowledge blind spots, and through dynamic reconstruction and gray-scale verification of the graph database, a complete closed loop of "user data feeding back into the graph structure" is formed, fundamentally improving the global scientific rigor and success rate of adaptive learning paths.

[0064] It should be understood that the examples and embodiments described herein are for illustrative purposes only and are not intended to limit the invention. Those skilled in the art can make various modifications or changes based on them. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the protection scope of the invention.

Claims

1. A method for dynamic evolution of knowledge graph topology, characterized in that, Includes the following steps: S1. Initiate anomaly detection for group learning, aiming to discover abnormal defects in the graph topology of group learning; S2. After detecting an anomaly in the graph topology, trigger the reflective multi-agent team workflow to perform attribution analysis on the abnormal nodes in the graph topology and generate missing hypothesis results. S3, the native graph database engine, without destroying the original global graph topology, reconstructs the local graph topology based on the missing hypothesis results and generates isolated shadow branches. S4. For the newly generated local graph topology, a gray-scale release mechanism is introduced for gray-scale verification. If the verification passes, the new topology will be globally effective. If the verification fails, the topology will be rolled back and the shadow branch will be deleted.

2. The method for dynamic evolution of knowledge graph topology as described in claim 1, characterized in that, In S1, for group learning, the micro-course knowledge graph stored in the native graph database is continuously monitored, the micro-course knowledge graph is structurally transformed to obtain the micro-course knowledge graph topology, the dynamic user learning progress attributes of each knowledge point node in the graph topology are aggregated and extracted, and the group cognitive diagnosis failure rate of any knowledge point node is calculated. If, for a given knowledge point node, the group cognitive diagnosis failure rate of 10% or more users exceeds a preset threshold, and the group cognitive diagnosis failure rate of the preceding knowledge point nodes is lower than the preset threshold, then the knowledge point node is determined to have a structural cognitive gap and is identified as an abnormal knowledge point node. At the same time, a graph evolution mechanism for local graph topology reconstruction is triggered for the abnormal knowledge point node.

3. The method for dynamic evolution of knowledge graph topology as described in claim 2, characterized in that, In S1, the knowledge point nodes are... Group cognitive diagnosis failure rate The calculation expression is: in, Indicates at the knowledge point node The total number of users who generate learning behavior. Indicates user index, Indicates the first Individual users at knowledge point nodes Cognitive diagnostic scores on This indicates the preset passing or mastery threshold. This is an indicator function, representing a value of 1 when the condition is met, i.e., the score is below the threshold, and a value of 0 otherwise.

4. The method for dynamic evolution of knowledge graph topology as described in claim 1, characterized in that, In S2, the reflective agent team includes an attribution agent and a hypothesis-generating agent, and the reflective agent team workflow includes: S2.1 Extract the text of incorrect questions, the text of solution steps, and the dialogue interaction logs of users who failed at abnormal knowledge point nodes, and use them as data input; S2.2 Attribution agent uses the deep semantic analysis capabilities of large language models to cluster and extract common features from massive amounts of incorrect question texts, find the underlying error logic that leads to the failure of a group of users, and obtain the attribution results; S2.3 Assume that the generated agent infers the implicit prior knowledge points missing from the abnormal knowledge point nodes based on the attribution results.

5. The method for dynamic evolution of knowledge graph topology as described in claim 4, characterized in that, In step S3, the local graph topology is dynamically reconstructed, and isolated shadow branches are generated, specifically including: S3.1 For the abnormal knowledge point node A, insert the implicit prerequisite knowledge point node A0 generated by the hypothesis generating agent into the micro-course knowledge graph. The implicit prerequisite knowledge point node A0 contains the knowledge point name and the corresponding exercise. S3.2 Modify the edge pointing relationship in the micro-course knowledge graph. If the original edge pointing relationship is B→A, then generate a parallel test path B→A0→A in the original graph database and regenerate the local graph topology structure. Here, the edge represents the logical dependency relationship between knowledge point nodes.

6. The method for dynamic evolution of knowledge graph topology as described in claim 5, characterized in that, The closed-loop verification process in S4 further includes: S4.1 When a new user reaches knowledge point node B, the traffic is distributed according to the set traffic ratio. A portion of the new users are directed to a new exploration path, i.e., B→A0→A, while the remaining new users continue to follow the original path, i.e., B→A. S4.2 Track and compare the first pass rate and cognitive diagnosis score of the two groups of new users on the final target knowledge point node A in real time, and perform a significance test on the knowledge point node A. S4.

3. Let the group that passes through the implicit prior knowledge point node A0 be the gray group. If the mastery of the gray group on knowledge point node A exceeds the preset threshold, then the newly generated local graph topology structure is officially applied globally, and the old edge B→A is deleted. If it does not exceed the preset threshold, then the hypothesis of the agent generated by the hypothesis is determined to be invalid, the topology is rolled back, and the shadow branch, i.e., the implicit prior knowledge point node A0, is deleted.

7. The method for dynamic evolution of knowledge graph topology as described in claim 6, characterized in that, In S4, the first-pass rate is the number of first-time passers divided by the total number of attempts. The cognitive diagnostic score is obtained by the large language model. The significance test calculation formula is as follows: in, The Z-statistic, or standard score, represents the significance test. and These are the average scores of the test group (which passed through the new knowledge point node A0) and the control group (which followed the original path) at the final target knowledge point node A, respectively. and Let V be the variance of the two sets of scores. and These are the sample sizes for the two groups; pass The degree of mastery corresponding to the value calculation value: in, The cumulative distribution function of the standard normal distribution. For integration variables; When mastering Less than the preset threshold When the newly generated local graph topology is deemed valid, the new topology is officially applied globally.

8. A dynamic evolution system for knowledge graph topology, applicable to the dynamic evolution method for knowledge graph topology as described in any one of claims 1-7, characterized in that, An evolution control center is added to the two-layer adaptive learning knowledge graph system. The two-layer adaptive learning knowledge graph system includes an AI engine layer, a hybrid data persistence layer, a presentation layer, and an application service layer. The evolutionary control center includes: An anomaly monitoring engine periodically performs batch queries on the native graph database, calculates the failure rate of group cognitive diagnosis for each node based on aggregation functions, and identifies high-frequency abnormal nodes. The reflective agent team, which is connected to the anomaly monitoring engine and is independent of the multi-agent team in the AI ​​engine layer, is used for error analysis, knowledge point reasoning, and generation of evolution graph database query language code. The gray-scale routing and evaluation module is connected to the reflective agent team. It maintains the path distribution strategy at the application service layer, records the comparison data of the gray-scale release mechanism, and performs significance tests.

9. The knowledge graph topology dynamic evolution system as described in claim 8, characterized in that, The AI ​​engine layer encapsulates the client based on the AutoGen framework, directly interacts with the deep inference large language model, and the AI ​​engine layer adopts a polling group dialogue mode for orchestration and collaboration. The hybrid data persistence layer is connected to the AI ​​engine layer and includes a relational database and a native graph database. The relational database is used to store the graph attribute metadata of the macro career path graph and the micro course knowledge graph. The native graph database is used to store the structured graph data of the macro career path graph and the micro course knowledge graph, and the two levels of graphs are isolated by the node labels in the structured graph data. The presentation layer, as the front-end interface, is built on React and connected to the hybrid data persistence layer. It is used to retrieve data from the back-end and provide users with an intuitive knowledge graph visualization and interactive interface. The application service layer, which serves as the backend service framework, interacts with the three layers mentioned above to manage, process, invoke, and distribute data.

10. A computer device, characterized in that, It includes a memory and a processor, the memory storing a computer program, which, when executed by the processor, causes the processor to perform the steps of the knowledge graph topology dynamic evolution method as described in any one of claims 1-7.

Images