Teaching large model knowledge graph updating method and system based on federal learning

Abstract

The invention discloses a teaching large model knowledge graph updating method and system based on federated learning. The method comprises the following steps: training a local problem solving strategy model reflecting a specific teaching strategy by each client, and generating structured learning effect data recording an application effect of the local problem solving strategy model; the central server performs functional clustering on the models by analyzing behavior characteristics of the local problem solving strategy models when the local problem solving strategy models process the preset probe problems so as to identify and divide client clusters representing different teaching genres; and finally, based on a clustering result and the learning effect data, performing weighted updating on a multi-dimensional logic knowledge graph, and storing objective effects of each teaching genre for different student portraits into the graph. According to the method, the diversity of teaching strategies can be effectively identified and reserved, and the quality of a knowledge graph on which a large teaching model depends and the accuracy of personalized teaching recommendation are remarkably improved through intelligent fusion based on objective effects.

Description

Technical Field

[0001] This invention relates to the field of large-scale teaching models based on artificial intelligence, and more specifically, to a method and system for updating the knowledge graph of a large-scale teaching model based on federated learning. Background Technology

[0002] With the deepening application of artificial intelligence technology in education, personalized teaching and learning have become important development directions. Knowledge graphs, as a core technology for structured representation of knowledge relationships, provide a foundation for personalized learning path planning and intelligent tutoring, and are a key component driving large-scale teaching models to perform logical reasoning and generate interpretable content. However, building high-quality teaching knowledge graphs requires integrating data from different educational institutions, which often involves sensitive information such as student learning behavior, raising data security concerns.

[0003] Federated learning, as a distributed machine learning paradigm, allows participating parties to collaboratively train a model without sharing the original data, providing a feasible technical path to address the aforementioned privacy issues. In federated learning, each participating party (client) trains the model locally using its private data, only uploading model update information (such as parameter gradients) to a central server for aggregation, thereby generating a globally superior model.

[0004] While federated learning offers significant advantages in terms of privacy protection, it faces the challenge of data heterogeneity in practical applications, particularly when dealing with data from diverse sources. Data heterogeneity, especially statistical heterogeneity, refers to the fact that the local data of each client is not independently identically distributed (Non-IID). This difference in data distribution can cause the local problem-solving strategy models trained on each client to optimize in different directions, leading to conflicts during aggregation and thus reducing the convergence speed and final performance of the global model.

[0005] To address the issue of data heterogeneity, existing technologies have proposed several solutions. For example, improving federated learning aggregation algorithms by introducing correction terms can mitigate the "drift" between local problem-solving strategy models; alternatively, personalized federated learning strategies can be employed, retaining a subset of personalized model parameters for each client; still others involve clustering clients before aggregation, grouping similar clients together, and then performing model aggregation within each cluster. For instance, some technical solutions propose clustering based on the similarity of client model update parameters.

[0006] However, existing technologies, when applied to the field of education, have failed to fully recognize the unique implications of heterogeneity in this context. In educational practice, different educational institutions, through long-term exploration, often develop distinctive teaching methods and problem-solving strategies for the same knowledge domain, which can be termed different teaching schools. These schools represent logically consistent but divergent cognitive paradigms and procedural knowledge. Taking the widely recognized teaching challenge of "constructing auxiliary lines" in plane geometry proofs as an example, how to add auxiliary lines not given in the problem is not a deterministic algorithmic process, but rather guided by a series of core heuristic strategies. These strategies, refined and summarized through long-term teaching practice, have formed several distinctive "teaching schools," such as: The similarity construction approach: its core heuristic strategy is to establish proportional relationships between the side lengths by creating similar triangles. Typical methods of constructing auxiliary lines include extending line segments to intersect, or drawing parallel lines to intercept line segments.

[0007] The circle and tangent approach: Its core idea is to use the geometric properties of circles to solve problems. Typical methods of constructing auxiliary lines are to connect the center of the circle with the point of tangency or intersection, or to construct auxiliary circles based on specific conditions.

[0008] Symmetry and Rotation School: Its core strategy is to transform the problem conditions by constructing symmetrical or rotated figures. Typical auxiliary line construction methods include drawing an axis of symmetry or rotating the figure around a certain point.

[0009] The equal area approach: Its core strategy is to construct shapes with equal areas and then substitute them. Typical methods for constructing auxiliary lines include connecting vertices to points on the sides, or drawing parallel lines to construct triangles of equal height.

[0010] Existing technologies tend to treat differences between clients as statistical noise that needs to be eliminated or averaged. If clustering is performed based on methods such as model parameter similarity, accidental parameter update similarities may incorrectly group models with vastly different teaching logics into the same category, or separate models with the same logic but different implementation details. Ultimately, through simple model parameter averaging or inappropriate clustering, these strategies with specific teaching logics (i.e., teaching styles) may be weakened or eliminated during the aggregation process. This results in the global model failing to reflect the advantages of specific strategies, reducing their effectiveness in specific scenarios, and failing to provide optimal guidance for specific types of students, thus violating the original intention of personalized education. Summary of the Invention

[0011] One aspect of this invention is to provide a method and system for updating the knowledge graph of a large teaching model based on federated learning, in order to solve the technical problem that existing technologies fail to effectively identify and utilize the logical heterogeneity represented by different teaching schools in educational application scenarios, resulting in poor model fusion effects.

[0012] To address the aforementioned technical problems, this invention provides a method for updating a knowledge graph of a large-scale teaching model based on federated learning. This method is executed in a system consisting of a central server and multiple client nodes, and includes the following steps: At each client node, a local problem-solving strategy model embodying a specific teaching strategy is trained using local teaching data. This teaching strategy generates personalized guidance information based on the student's problem-solving process. Based on the application effect of the local problem-solving strategy model, a structured learning effectiveness data is generated. This structured learning effectiveness data records the quantitative effectiveness indicators of the teaching strategy for a specific student profile and a specific knowledge transfer path. Then, the local problem-solving strategy model or its representative information, along with the structured learning effectiveness data, is sent to the central server.

[0013] The central server receives local problem-solving strategy models or their representative information from multiple client nodes, along with structured learning performance data. Next, it performs functional clustering on these local problem-solving strategy models to identify and divide client clusters representing different teaching styles (in this scheme, these client clusters actually represent the data structure abstraction of teaching styles during the knowledge graph update and construction process). This functional clustering is based on the behavioral characteristics exhibited by the problem-solving strategy models when handling preset probe problems. The probe problems refer to a preset set of baseline problems, each with multiple known problem-solving paths. Finally, based on the clustering results and the structured learning performance data, a multidimensional logical knowledge graph, serving as the core knowledge base of the large teaching model, is weighted and updated. The update process includes aggregating problem-solving strategy models within each client cluster, calculating the comprehensive performance score of each teaching style for different student profiles and knowledge transfer paths using the structured learning performance data, and storing this comprehensive performance score in the multidimensional logical knowledge graph.

[0014] One aspect of this invention is the provision of a functional clustering method to identify logical heterogeneity. Preferably, the step of functionally clustering local problem-solving strategy models specifically includes: loading each local problem-solving strategy model by the central server and having it process a preset set of standard probe questions containing typical cognitive difficulties; during processing, collecting and analyzing the behavioral characteristics of each local problem-solving strategy model to generate a behavioral signature vector, wherein the behavioral characteristics include at least one of feedback type distribution, intervention timing, content generation style, or error guidance pattern; finally, using a clustering algorithm to cluster the behavioral signature vectors corresponding to all local problem-solving strategy models, thereby grouping models with behavioral signature vectors within a preset distance threshold into the same client cluster. This invention's method does not directly rely on the model's internal parameters, but rather identifies essentially the same or similar teaching styles from a behavioral and functional perspective by analyzing the external behavioral characteristics exhibited by the model when processing specific tasks, thus avoiding the erroneous merging of valuable teaching strategies.

[0015] Another aspect of this invention is the proposal of a weighted aggregation mechanism based on objective effectiveness. Preferably, the structured learning effectiveness data is a structured data object that maps a student profile and a knowledge transfer process to a set of specific effectiveness metrics. The student profile categorizes student learning characteristics, the knowledge transfer path refers to the learning process from a prior knowledge point to a target knowledge point, and the effectiveness metrics include at least one of the following: mastery time reduction rate, first-attempt success rate increase, or error rate decrease percentage. By establishing an effectiveness data channel independent of model parameters, the updating of the knowledge graph is driven by quantifiable and realistic teaching effects, enabling the selection and integration of teaching strategies based on objective teaching effectiveness.

[0016] Furthermore, to enhance the protection of core intellectual property rights of clients and the credibility of the system, in some implementations, the step of functionally clustering the local problem-solving strategy model is replaced by a decentralized mechanism based on zero-knowledge proofs. This mechanism specifically includes: the central server broadcasting an encrypted challenge task to all client nodes; each client node executing the challenge task in its local secure environment and generating a zero-knowledge proof for the behavioral characteristics of its local problem-solving strategy model. This zero-knowledge proof proves that the behavioral characteristics satisfy specific conditions without revealing the model's specific behavior and parameters; the client nodes submit the zero-knowledge proof to a decentralized public ledger; on this public ledger, by executing a preset smart contract, the validity of all client-submitted zero-knowledge proofs is automatically verified, and a clustering operation is performed based on the behavioral characteristics represented by the verified proofs, thereby determining the client clustering results. This mechanism eliminates the need for the central server to access the client model or its specific behavioral data during the clustering process, achieving private knowledge protection and intellectual property protection, and improving the overall credibility of the system.

[0017] This invention also provides a method for dynamically applying the multidimensional logical knowledge graph to provide personalized teaching recommendations. The method includes: obtaining a student profile of the current user and determining the knowledge transfer path to be learned; querying the multidimensional logical knowledge graph for teaching information corresponding to the knowledge transfer path, providing the student profile as a query condition; the multidimensional logical knowledge graph, based on the student profile, retrieving the teaching method with the highest overall effectiveness score from multiple recorded teaching methods; and finally, calling the aggregated representative problem-solving strategy model associated with the teaching method with the highest overall effectiveness score to generate and deliver personalized learning content or guidance paths for the user.

[0018] Another aspect of the present invention is to provide a knowledge graph update system for a large teaching model based on federated learning, which includes multiple client nodes and a central server.

[0019] The client node is configured to: train a local problem-solving strategy model using local teaching data; generate structured learning performance data based on the model's application effect; and send the model or its representative information and performance data to the central server.

[0020] The central server is configured to: receive data from clients; perform functional clustering of problem-solving strategy models to identify teaching styles; and perform weighted updates of the multidimensional logical knowledge graph based on clustering results and effectiveness data.

[0021] Through the methods described above, this system can effectively preserve and utilize teaching diversity, achieve intelligent integration based on objective outcomes, and provide deeply personalized and explainable teaching recommendations, thereby promoting collaboration within the education ecosystem.

[0022] Compared with existing technologies, the method proposed in this invention, by introducing functional clustering based on model behavioral features, can penetrate the superficial appearance of model parameters, accurately identify and retain the logical heterogeneity with practical teaching value represented by each participant in federated learning, and overcome the technical defects of traditional aggregation methods that may lead to the erroneous averaging or elimination of high-quality teaching strategies. Simultaneously, by establishing a weighted update mechanism based on structured learning effectiveness data, the updating and optimization of the multidimensional logical knowledge graph has objective and quantifiable effectiveness evidence, ensuring that strategies with better teaching effects can be retained and applied. Ultimately, this invention can construct a high-quality knowledge graph that reflects teaching diversity and labels the applicable scenarios and objective effectiveness of various teaching strategies, thereby significantly improving the accuracy and effectiveness of personalized guidance when the large teaching model is used. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0024] Figure 1 This is a schematic diagram of the system structure according to an embodiment of the present invention; Figure 2 This is a flowchart of a method according to an embodiment of the present invention; Figure 3 This is a detailed flowchart of the functional clustering steps in an embodiment of the present invention; Figure 4 This is a schematic diagram comparing the performance of the embodiments of the present invention with existing technologies across multiple dimensions; Figure 5 This is a schematic diagram of the functional clustering process in a specific embodiment of the present invention. Detailed Implementation

[0025] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0026] Example 1: This embodiment of the invention provides a method for updating the knowledge graph of a large-scale teaching model based on federated learning. (Refer to...) Figure 1 This method is executed in a system consisting of a central server 100 and multiple client nodes 200. The central server 100 can be one or more physical servers, cloud servers, or server clusters. The client nodes 200 can be servers of educational institutions, personal computers, or other terminal devices with a certain computing power.

[0027] Reference Figure 2 The method provided in this embodiment may include the following steps: Step S1: On each client node 200, a local problem-solving strategy model embodying a specific teaching strategy is trained using local teaching data. The teaching strategy is used to generate personalized guidance information based on the student's problem-solving process. Based on the application effect of the local problem-solving strategy model, a structured learning effectiveness data is generated. This structured learning effectiveness data records the quantitative effectiveness indicators of the teaching strategy for a specific student profile and a specific knowledge transfer path. Then, the local problem-solving strategy model or its representative information, along with the structured learning effectiveness data, is sent to the central server 100.

[0028] Specifically, the local problem-solving strategy model can be an algorithmic model designed to guide students in solving problems in a specific domain. In educational scenarios, such as mathematical problem-solving, this model can be a sequence-to-sequence model based on the Transformer architecture. Its inputs are the problem description and the student's current solution steps, and its output is a hint or guiding question for the next step. Each client node 200, for example representing different educational institutions, uses its unique teaching materials, exercise solution records, teacher-student interaction logs, etc., which have been validated through long-term teaching practice, as local teaching data for model training. This data contains the institution's unique teaching methods and problem-solving approaches, i.e., teaching styles, and the trained local problem-solving strategy model thus becomes a digital representation of that teaching style.

[0029] When generating structured learning effectiveness data, the first step is to quantify the students and learning tasks. For example, a student profile can be defined as a multi-dimensional vector, whose dimensions may include: prior knowledge mastery (0-1 floating-point number), logical reasoning ability score (1-5 integer), learning initiative index (0-1 floating-point number), etc. A knowledge transfer path can be defined as a binary tuple, such as (knowledge point ID_A, knowledge point ID_B), representing the learning process from prior knowledge point A to target knowledge point B.

[0030] Based on this, client node 200 locally applies its problem-solving strategy model to a group of student samples with known profiles and records their performance on specific knowledge transfer paths. The generated structured learning effectiveness data is a key-value pair data object, where the key is a combination of "student profile" and "knowledge transfer path," and the value is a set of specific effectiveness metrics. Preferably, the effectiveness metrics include, but are not limited to: mastery time reduction rate, first-attempt success rate improvement value, or percentage reduction in key concept error rate. For example, an effectiveness data record could be: { (student profile vector [0.8, 4, 0.9], knowledge transfer path (KP034, KP056)): {mastery time reduction rate: 0.25, success rate improvement value: 0.15}}. This data shows that for students with solid prior knowledge, strong logical ability, and high initiative, when learning "the sum of the interior angles of a triangle" from "the sum of the interior angles of a polygon," using the teaching strategy of this client reduces learning time by 25% and increases the first-attempt success rate by 15%.

[0031] This step establishes an objective effect evaluation channel independent of the model's internal parameters, binding the abstract "teaching strategy" with its quantifiable "teaching effectiveness," providing an objective basis for subsequent valuable integration by the central server. This solves the problem that traditional federated learning relies solely on model parameters and cannot evaluate the merits of the underlying logic of the model.

[0032] As an alternative, local problem-solving strategy models can also employ reinforcement learning models, such as a policy gradient-based intelligent agent. Its state space represents the current problem and the student's problem-solving state, its action space represents the provision of different types of instructional interventions (such as heuristic, step-by-step, and conceptual), and its reward function is directly linked to the aforementioned performance metrics (such as problem-solving time and accuracy).

[0033] Step S2: At the central server 100, local problem-solving strategy models or their representative information, as well as structured learning performance data, are received from multiple client nodes 200.

[0034] The model representation information may be key parts of the model extracted to reduce communication overhead, such as parameters of the last few layers of the model, or model embedding vectors generated by a pre-trained autoencoder. The central server 100 receives this data through a secure network interface (e.g., a RESTful API based on TLS encryption) and establishes a temporary storage space for each client node 200 for subsequent processing.

[0035] Step S3: Perform functional clustering on the multiple local problem-solving strategy models to identify and classify client clusters representing different teaching styles.

[0036] This is one of the core steps of this invention: identifying logically similar teaching strategies at the behavioral and functional levels. (See also...) Figure 3 This step can be broken down into: Step S301: The central server 100 loads a pre-defined set of standard probe problems. This set of problems is designed by domain experts and contains several (e.g., 20-50) representative problems. These problems are characterized by the existence of multiple known, logically different, and valid solution paths, and contain cognitive difficulties that can easily confuse students. For example, in plane geometry, a probe problem may be solvable either by constructing similar triangles or by constructing an auxiliary circle.

[0037] Step S302: The central server 100 sequentially loads the local problem-solving strategy models received from each client node 200, and allows each model to independently "solve" all the problems in the probe problem set. During this virtual problem-solving process, the central server 100 collects and records the key behavioral characteristics of each model.

[0038] Preferably, the behavioral characteristics include: a) Feedback type distribution: The feedback information generated by the model is classified into "definition explanation", "example", "step breakdown", "negative example warning", etc. by a classifier (such as a BERT-based text classifier), and the frequency distribution of each category is statistically analyzed.

[0039] b) Intervention timing: Record the average number of interactions required for the model to provide a critical prompt when a student experiences a pause in their operation.

[0040] c) Content generation style: Extract TF-IDF feature vectors or sentence embedding vectors from the text content generated by the model to characterize its linguistic habits and complexity.

[0041] d) Error-guided pattern: When faced with specific error-prone points, the analysis model will identify and classify whether it leads students into a specific erroneous thinking path.

[0042] Step S303: For each local problem-solving strategy model, concatenate the various quantified behavioral features collected in step S302 to form a high-dimensional behavioral signature vector. For example, the behavioral signature vector of a model can be V = [v1, v2, ..., vn], where the first four dimensions represent the distribution frequency of the four feedback types, the fifth dimension represents the average intervention timing, and the subsequent dimensions are the vector representation of the content style.

[0043] Step S304: A clustering algorithm, such as Density-Based Spatial Clustering of Applications with Noise (DBSCAN), is used to cluster the behavioral signature vectors corresponding to all local problem-solving strategy models. The DBSCAN algorithm can divide them into different clusters based on the density of the vectors in the feature space and identify outliers as noise. For example, by setting an appropriate neighborhood radius eps and a minimum number of samples min_samples, the algorithm will automatically group models with similar behavioral signature vectors into the same client cluster. These models are classified as belonging to the same teaching school because they exhibit similar behavioral patterns when handling key problems.

[0044] The technical advantage of this step is that it does not rely on model parameters that may vary greatly due to training details. Instead, it judges the internal logic by the model's external behavior when solving problems, thus achieving accurate identification of the functional heterogeneity of teaching styles and avoiding the mutual cancellation of the effects of effective strategies with different logics due to aggregation.

[0045] Step S4: Based on the clustering results and the structured learning performance data, a multidimensional logical knowledge graph designed to provide decision-making support for a large-scale teaching model is updated with weights.

[0046] This step specifically includes: The aggregation here can employ conventional algorithms such as FedAvg, because the models within a cluster have similar logic, making parameter averaging reasonable. After aggregation, each cluster generates a "representative problem-solving strategy model".

[0047] Next, for each cluster (i.e., each teaching approach), the central server 100 uses the structured learning outcomes data sent by all clients to calculate the overall effectiveness score for that teaching approach. Specifically, for each (student profile, knowledge transfer path) combination, the central server searches for the effectiveness data submitted by all client nodes belonging to the current cluster, and performs a weighted average or takes the maximum value of the corresponding effectiveness metrics (such as the mastery time reduction rate) to obtain the overall effectiveness score of that teaching approach in that specific scenario. This score objectively reflects the effectiveness of this teaching approach for a specific type of student on a specific learning task.

[0048] Finally, the central server 100 updates a multi-dimensional logical knowledge graph. Nodes in this knowledge graph represent knowledge points, and edges represent the migration paths between knowledge points. In this invention, this knowledge graph serves as a core component of the large-scale teaching model. Each edge no longer stores a single relation, but rather a list. Each element in the list is a structure containing a teaching style ID, a comprehensive effectiveness score, and a pointer to a representative problem-solving strategy model of that style. During updates, the central server 100 updates or adds the corresponding teaching style information to the list of the corresponding edges in the knowledge graph based on the calculated comprehensive effectiveness score.

[0049] Reference Figure 4 Compared with the traditional federated averaging method, the method proposed in this invention effectively identifies and retains high-quality teaching strategies. As a result, the final global model (called through a knowledge graph) can achieve a higher level in key performance indicators such as personalized recommendation accuracy, and may have a faster convergence speed because it avoids model update conflicts.

[0050] Furthermore, to enhance the protection of core intellectual property rights on the client side and the credibility of the system, in some implementations, the functional clustering process in step S3 can be replaced by a decentralized mechanism based on zero-knowledge proofs. This mechanism specifically includes: A set of encrypted challenge questions (i.e., probe questions) is broadcast. Each client node 200 decrypts and executes the challenge question in its local secure execution environment (such as SGX, Software Guard Extensions), generating a zero-knowledge proof (such as zk-SNARK, Zero-Knowledge Succinct Non-Interactive Argument of Knowledge) based on the behavioral characteristics of its local problem-solving strategy model. This proof demonstrates that one or more components of its model's behavioral signature vector fall within a specific range without revealing the specific values ​​of the vectors or the model itself. For example, a proof might demonstrate that "the proportion of 'generalization' in my model's feedback type distribution is greater than 0.5," but the verifier cannot know whether the specific proportion is 0.6 or 0.7. The client node submits this zero-knowledge proof to a decentralized public ledger (such as a blockchain network based on Hyperledger Fabric). On this ledger, a pre-defined smart contract automatically verifies the validity of all proofs submitted by clients and, based on the common characteristics declared in the verified proofs, executes clustering logic to divide client addresses into different clusters, recording the clustering results on the ledger. Under this mechanism, the central server 100 does not need to receive models, but only acts as a task publisher and result reader, greatly improving the system's privacy protection level and mutual trust foundation.

[0051] This invention also provides a method for dynamically applying the multidimensional logical knowledge graph to provide personalized teaching recommendations. When a learner requests a teaching service, the system first obtains the user's student profile and determines their current knowledge transfer path. Then, using this student profile and knowledge transfer path as query criteria, the system queries the multidimensional logical knowledge graph. The knowledge graph returns all teaching approaches recorded along the transfer path and their corresponding comprehensive effectiveness scores. The system selects the teaching approach with the highest comprehensive effectiveness score for the student profile and calls the associated, aggregated, representative problem-solving strategy model to generate and deliver highly personalized and optimized learning content or guidance paths to the user.

[0052] Example 2: This embodiment of the invention also provides a knowledge graph update system for a large-scale teaching model based on federated learning, the structure of which is shown in the attached figure. Figure 1 As shown, this system is the execution entity of the aforementioned method embodiment. The system includes a central server 100 and multiple client nodes 200.

[0053] The client node 200 is configured as follows: A local problem-solving strategy model was trained using local teaching data, and this model reflects the client-specific teaching strategies. Based on the application effect of the local problem-solving strategy model, structured learning effectiveness data is generated, which quantitatively correlates the effectiveness of teaching strategies with student profiles and knowledge transfer paths. The local problem-solving strategy model or its representative information, along with the structured learning performance data, are sent to the central server 100 via a communication interface.

[0054] In specific implementation, the client node 200 may include: a model training module for performing the training task of the local problem-solving strategy model; an effectiveness evaluation module for applying the model and generating structured learning effectiveness data; and a communication module for data interaction with the central server 100.

[0055] The central server 100 is configured as follows: Receive local problem-solving strategy models or their representative information from multiple client nodes 200, as well as structured learning performance data; Functional clustering is performed on the multiple local problem-solving strategy models to identify and classify client clusters representing different teaching styles. The clustering is based on the behavioral characteristics of the models when handling probe problems, rather than the model parameters. Based on the clustering results and the structured learning performance data, a multidimensional logical knowledge graph is updated with weights. The update process includes intra-cluster model aggregation and calculation of the comprehensive utility of each school of thought based on the performance data.

[0056] In its specific implementation, the central server 100 may include: a data receiving module for receiving and managing data from various clients; and a functional clustering module for performing clustering operations. Figure 3 The clustering process is shown; and a knowledge graph update module is used to perform weighted update logic.

[0057] Through the collaborative work of the aforementioned client nodes and central server, this system can effectively identify, retain, and utilize the diversity of valuable teaching strategies in educational scenarios, establish an intelligent fusion mechanism based on objective results, and ultimately provide deeply personalized and interpretable teaching recommendations through an updated multidimensional logical knowledge graph, thereby constructing a technology platform that can promote collaborative innovation in the education ecosystem.

[0058] Example 3: To further clarify the application details and technical advantages of this invention in specific business scenarios, this example will take the classic teaching difficulty of "construction of auxiliary lines" in high school plane geometry as an example to describe in detail the complete execution process of the method of this invention. In this scenario, different problem-solving approaches (e.g., based on congruence and similarity construction, based on symmetry and rotation transformation, based on the relevant properties of circles, etc.) constitute logically significantly different "teaching styles". In this scenario, the large-scale teaching model requires a high-quality knowledge graph to provide accurate and diverse guidance for problem-solving approaches.

[0059] The federated learning system in this embodiment consists of a central server 100 deployed in the cloud and three client nodes 200 representing different types of top educational institutions. These are Client A (representing a traditional key middle school that emphasizes fundamentals and axiomatic deduction, referred to as the "Axiomatic School"), Client B (representing an online education platform that is good at inspiring students to perform graphical transformations and think flexibly, referred to as the "Transformation School" in this invention), and Client C (representing an institution that focuses on mathematics competition training and is proficient in advanced problem-solving techniques, referred to as the "Competition School").

[0060] Client A trained a local problem-solving strategy model, LM_A (LM stands for Local Model), using a large number of classic examples annotated by experienced teachers and student solution records. When students get stuck, this model tends to guide them to review fundamental theorems and attempt to construct congruent or similar triangles. Client B's training data includes numerous interactive geometric animations and teaching materials guiding students to "fold" and "rotate" shapes. Its trained model, LM_B, tends to provide hints such as "Try rotating △ABC 90 degrees around point A" or "Construct symmetry along a certain line segment." Client C's model, LM_C, is trained based on past competition problems and special problem-solving models, excelling at identifying hidden "concyclic" conditions or utilizing advanced techniques such as Menelaus's theorem.

[0061] Meanwhile, each client applies the model in its teaching practice and generates structured learning effectiveness data. For example, for a knowledge transfer path from "mastering the properties of isosceles triangles" to "solving complex geometric proofs with special angles," and a student profile P1 defined as "weak geometric intuition but strong algebraic calculation ability," client A, through a comparative experiment, recorded the application effectiveness of its model LM_A as: { (profile P1, transfer path (KP_Isosceles, KP_ComplexProof)): {first attempt success rate improvement: 0.35, average problem-solving time reduction rate: 0.20}}. Similarly, clients B and C also generated their own structured learning effectiveness data packages D_B and D_C for different student profiles and knowledge transfer paths, and encrypted and sent the parameters (or their embedding vectors) of the three models LM_A, LM_B, and LM_C along with the effectiveness data packages D_A, D_B, and D_C to the central server 100.

[0062] The server loads a pre-set set of probe questions, including a classic problem: "In quadrilateral ABCD, given AD=BC, M and N are the midpoints of AB and CD respectively, prove that the angles between MN and AD / BC are equal." This problem has multiple solutions, effectively identifying different problem-solving logics. The central server loads LM_A, LM_B, and LM_C sequentially and simulates the interaction between students and the model to collect behavioral features.

[0063] When dealing with this probe problem, the feedback sequence generated by the LM_A model is mainly: "Connect AC, take the midpoint P of AC, connect PM and PN", "Consider the relationship between PM and BC, and the relationship between PN and AD", "Use the midline theorem of triangles". Its behavioral characteristics are quantified, and in the dimension of "feedback type distribution", the "step decomposition" class has a very high weight.

[0064] Model LM_B generates completely different feedback: "Translate △ADN so that point D coincides with point C, resulting in △BCE", and "Observe the vector relationship MN and AE in the new figure". Its behavioral characteristics, in the dimension of "content generation style", show a strong orientation towards geometric transformation.

[0065] The model LM_C may provide less common hints, such as “consider complex or vector methods”, which is reflected in the “higher-order tool application” category in the “feedback type distribution”.

[0066] The central server generated behavioral signature vectors V_A, V_B, and V_C for each model. Since the distances between V_A and V_B in the feature space are significantly greater than their average vector distances to the historical "axiomatic" and "transformative" models, the DBSCAN clustering algorithm successfully identified and partitioned clients A and B into two distinct client clusters: Cluster 1 (axiomatic) and Cluster 2 (transformative). Client C, due to its unique solution, may be classified into a separate cluster or identified as an outlier, representing the still-developing "competitive" group.

[0067] Finally, the central server performs a weighted update of the multidimensional logical knowledge graph based on the clustering results and learning performance data. Since there is currently only client A in cluster 1, its representative problem-solving strategy model M_1 is LM_A. The representative model M_2 of cluster 2 is LM_B. The server then processes the performance data. For example, when processing the student profile P1 and the knowledge transfer path mentioned above, the server updates or adds the following information to the edge "(KP_Isosceles, KP_ComplexProof)" in the knowledge graph: {School ID: "Axiomatic School", Applicable Profile: "P1", Overall Performance Score: {Success Rate Increase: 0.35, Time Reduction: 0.20}, Representative Model Pointer: &M_1}. If client B's data shows that its "Transformation School" method is not effective for students with profile P1 (e.g., the success rate increase is only 0.05), then the corresponding record's overall performance score will be lower.

[0068] Through this complete process, the multidimensional logical knowledge graph ultimately constructed by this invention can serve as a high-quality knowledge source to empower a large-scale teaching model. It not only includes problem-solving strategy models from different teaching schools of thought, but more importantly, it accurately labels the true effectiveness of each school of thought (such as the "axiomatic school" and the "transformation school") for different types of students (such as those with weak geometric intuition but strong computational abilities) on different learning tasks (such as "from isosceles triangles to complex proofs"). When a new user similar to profile P1 encounters such a problem, the system can intelligently invoke the representative model M_1 of the "axiomatic school" based on the high-score records in the knowledge graph, providing the most suitable and efficient personalized learning guidance. This provides an interpretable personalized education method driven by a large model.

[0069] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A knowledge graph update method for a large-scale teaching model based on federated learning, executed in a system consisting of a central server and multiple client nodes, characterized in that... Includes the following steps: At each client node, a local problem-solving strategy model embodying a specific teaching strategy is trained using local teaching data. This teaching strategy generates personalized guidance information based on the student's problem-solving process. Based on the application effect of the local problem-solving strategy model, a structured learning effectiveness data is generated. This structured learning effectiveness data records the quantitative effectiveness indicators of the teaching strategy for a specific student profile and a specific knowledge transfer path. The local problem-solving strategy model or its representative information, along with the structured learning effectiveness data, is then sent to the central server. The representative information is at least one of the parameters of one or more network layers of the local problem-solving strategy model, the model embedding vector, and the parameter gradient. The central server receives local problem-solving strategy models or their representative information from multiple client nodes, as well as the structured learning performance data. It then performs functional clustering on the multiple local problem-solving strategy models to identify and divide them into different client clusters. This functional clustering is based on the behavioral characteristics exhibited by the local problem-solving strategy models when processing a preset set of benchmark problems, each with multiple known problem-solving paths. Finally, based on the clustering results and the structured learning performance data, a multidimensional logical knowledge graph is updated with weights.

2. The method according to claim 1, characterized in that, The steps for functional clustering of local problem-solving strategy models include: The central server loads each local problem-solving strategy model and makes it process a preset set of benchmark problems, each of which has multiple known problem-solving paths; During the processing, behavioral characteristics of each local problem-solving strategy model are collected and analyzed to generate a behavioral signature vector. The behavioral characteristics include at least one of feedback type distribution, intervention timing, content generation style, or error guidance mode. Clustering algorithms are used to cluster the behavior signature vectors corresponding to all local problem-solving strategy models, thereby grouping models with behavior signature vectors within a preset distance threshold into the same client cluster.

3. The method according to claim 1, characterized in that, The structured learning effectiveness data is a structured data object that maps a student profile and a knowledge transfer path to a set of specific effectiveness metrics. The student profile is a classification of student learning characteristics, the knowledge transfer path refers to the learning process from a prior knowledge point to a target knowledge point, and the effectiveness measurement indicators include at least one of the following: mastery time reduction rate, first attempt success rate increase value, or error rate decrease percentage.

4. The method according to claim 1, characterized in that, The specific steps for weighted updating of a multidimensional logical knowledge graph include: Within each client cluster, problem-solving strategy models are aggregated to generate representative problem-solving strategy models; Using the structured learning performance data, the comprehensive performance score of each client cluster for different student profiles and knowledge transfer paths is calculated; The comprehensive performance score and the representative problem-solving strategy model associated with each client cluster are stored in the multidimensional logical knowledge graph.

5. The method according to claim 1, characterized in that, The step of functionally clustering the local problem-solving strategy model is replaced by a decentralized mechanism based on zero-knowledge proofs, which includes: The central server broadcasts an encrypted challenge to all client nodes; Each client node generates a zero-knowledge proof locally for the behavioral characteristics of its local problem-solving strategy model. The zero-knowledge proof is used to prove that its behavioral characteristics satisfy specific conditions without revealing the specific behavior and parameters of the model. The client node submits the zero-knowledge proof to a decentralized public ledger; On this public ledger, a pre-defined smart contract is executed to perform clustering operations based on the behavioral characteristics represented by the verified proofs, in order to determine the partitioning results of client clusters.

6. The method according to claim 5, characterized in that, Each client node executes the challenge task in a local secure environment and generates the zero-knowledge proof.

7. The method according to claim 1 or 4, characterized in that, It also includes the step of applying the multidimensional logical knowledge graph to make personalized teaching recommendations, the steps of which include: Obtain the current user's student profile and determine the knowledge transfer path to be learned; The system queries the multidimensional logical knowledge graph and, based on the student profile, retrieves the client cluster with the highest overall performance score from among the recorded multiple client clusters. The system invokes a representative problem-solving strategy model associated with the client cluster that has the highest overall performance score to generate personalized learning content or guidance paths for the user.

8. The method according to claim 1, characterized in that, The local problem-solving strategy model is either a sequence-to-sequence model based on the Transformer architecture or a reinforcement learning model based on policy gradients.

9. The method according to claim 2, characterized in that, The clustering algorithm is the density-based DBSCAN algorithm for noisy application spaces.

10. A knowledge graph update system for a large-scale teaching model based on federated learning, comprising a central server and multiple client nodes, characterized in that: The client node is configured to: train a local problem-solving strategy model that reflects a specific teaching strategy using local teaching data; generate structured learning effectiveness data based on the application effect of the local problem-solving strategy model, wherein the structured learning effectiveness data records the quantitative effectiveness indicators of the teaching strategy for a specific student profile and a specific knowledge transfer path; and send the local problem-solving strategy model or its representative information along with the structured learning effectiveness data to the central server. The central server is configured to: receive data from multiple client nodes; perform functional clustering on the multiple local problem-solving strategy models to identify and divide different client clusters, wherein the functional clustering is based on the behavioral characteristics exhibited by the local problem-solving strategy models when processing a preset set of benchmark problems, each with multiple known problem-solving paths; and perform weighted updates on a multidimensional logical knowledge graph based on the clustering results and the structured learning performance data.

Images