Adaptive test difficulty control system based on reinforcement learning and knowledge graph

Abstract

This invention belongs to the field of computer application technology and discloses an adaptive test difficulty control system based on reinforcement learning and knowledge graphs. The system constructs a five-dimensional quantitative model of cognitive complexity, knowledge relevance, computational complexity, context novelty, and data heterogeneity adaptability. It accurately calculates features such as knowledge point path length and node centrality through federated knowledge graphs, and generates a comprehensive difficulty vector by combining attention mechanisms. It also achieves dynamic difficulty correction by relying on federated Bayesian updates. For new test questions, it uses transfer learning and prediction models to complete pre-calibration. The system constructs a three-dimensional cognitive model based on an improved federated graph attention network, which includes knowledge mastery status, cognitive ability level, and learning behavior characteristics. The cognitive ability level dimension includes five core abilities such as logical reasoning, calculation, and memory. It also deeply mines multimodal process data such as answer time, mouse trajectory, and option modification.

Description

Technical Field

[0001] This invention belongs to the field of computer application technology, specifically an adaptive test difficulty control system based on reinforcement learning and knowledge graphs. Background Technology

[0002] Computer-Adaptive Testing (CAT) is applied in various scenarios such as online assessment and personalized learning. Traditional adaptive testing systems are mainly based on methods such as Item Reaction Theory (IRT), Knowledge Space Theory (KST), or Bayesian networks, which can dynamically adjust the difficulty of test questions based on the test taker's performance, achieving a certain degree of personalized testing. However, existing technologies still have the following technical problems: The difficulty level of existing systems is mostly defined manually in advance by domain experts or calculated based on limited historical statistical data. They lack dynamic adjustment capabilities and cannot be adapted in real time according to the cognitive characteristics and answering habits of different test takers. They also cannot adapt to the differences in test difficulty among test takers in different regions and with different teaching progress, resulting in a low degree of match between the difficulty of the test questions and the actual ability of the test takers.

[0003] Most systems adopt a flat knowledge point tagging system, which simply marks the knowledge points corresponding to the test questions. This fails to fully reflect the hierarchical structure, dependencies, and transfer effects between knowledge points, making it impossible to build a complete knowledge network, accurately capture the weak links and cognitive gaps in the test taker's knowledge system, and provide precise support for personalized question selection.

[0004] Traditional CAT systems often rely on single-objective optimization for question selection. When the test length is short, the accuracy of estimating candidates' abilities drops significantly, and they are prone to getting trapped in local optima. They tend to focus excessively on certain knowledge points or difficulty ranges, failing to comprehensively cover candidates' knowledge mastery and resulting in insufficient comprehensiveness and reliability of test results.

[0005] The existing system makes decisions based solely on whether the candidate answers correctly or incorrectly, ignoring important information that reflects the candidate's true cognitive state, such as answering time, answering pattern, and error type. It cannot accurately determine whether the candidate has truly mastered the knowledge points, nor can it capture the fluctuations in the candidate's state during the test, thus affecting the scientific nature of the decision-making. Summary of the Invention

[0006] The purpose of this invention is to provide an adaptive test difficulty control system based on reinforcement learning and knowledge graphs to solve one or more problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: an adaptive test difficulty control system based on reinforcement learning and knowledge graphs, comprising the following modules: Furthermore, the federated graph module constructs a federated knowledge graph containing knowledge point entities, test question entities, cognitive level entities, and multiple types of relationships. It adopts a horizontal federated architecture, distributing and storing knowledge point and test question data from different educational scenarios, and achieving cross-node graph collaborative construction only through parameter sharing. The knowledge point entities include data source identifiers and cross-domain association weight attributes; the test question entities include data heterogeneity labels and adaptive scenario type attributes; and the relationships include cross-domain association strength and dynamic confidence. An improved graph convolutional neural network is used to learn the embedding representation of the federated knowledge graph. An attention mechanism and federated distillation technique are introduced to map each knowledge point and question into a low-dimensional dense vector, while calculating the topological features of each knowledge point. The federated distillation technology is specifically optimized for knowledge graph embedding scenarios. It uses a globally collaboratively trained graph model as the teacher model and the local models of each distributed node as student models. The distillation process only transmits the feature distribution and association weight information of the graph nodes, without transmitting the original knowledge points and test data. The student model completes local optimization by fitting the feature output of the teacher model.

[0008] It supports incremental updates based on user feedback and answer data, and designs a federated incremental update algorithm. When a node discovers a new knowledge point association or test question relationship, it only updates the graph fragment and related parameters of that node. The global graph is dynamically optimized through a federated synchronization mechanism.

[0009] Furthermore, the difficulty calibration module, combined with heterogeneous data processing technology, constructs a five-dimensional comprehensive calibration system for test question difficulty, including cognitive complexity, knowledge relevance, computational complexity, contextual novelty, and data heterogeneity adaptability. The cognitive complexity is based on Bloom's revised taxonomy of cognition and is dynamically adjusted in conjunction with test taker behavior data. The knowledge relevance is quantified by calculating the path length, node centrality, and cross-domain association strength of the knowledge points tested in the test question in the federated knowledge graph. The computational complexity is determined based on the number of reasoning steps required by the test question, the amount of computation, and the difficulty of data processing. The contextual novelty is calculated based on the similarity between the test question context and the test taker's common training context, and a similarity matching algorithm is used to optimize the calculation accuracy. The data heterogeneity adaptability is used to adapt test question data from different sources and in different formats. An attention mechanism is used to integrate five dimensions of difficulty features to generate a comprehensive difficulty vector for the test questions. A dynamic difficulty adjustment mechanism based on federated Bayes updates is established. According to the answer data of different test takers, the difficulty values ​​of each dimension of the test questions are adjusted in real time through parameter collaboration between federated nodes. Five-dimensional test question comprehensive difficulty vector:

[0010]

[0011] This represents the overall difficulty vector of the test questions, which is the final difficulty score after standardization. to The adaptive attention weights, representing the five-dimensional difficulty features, are automatically assigned based on the subject attributes, testing scenarios, and characteristics of the test-taker group. Represents cognitive complexity, based on a standardized score determined according to Bloom's revised taxonomy of cognitive objectives; The degree of knowledge relevance is quantified by the path length of knowledge points, node centrality, and cross-domain relevance strength in the federated knowledge graph. The standardized score represents the computational complexity, determined based on the computational complexity and data processing difficulty of the reasoning steps in the test question. It represents the novelty of the context, and is a quantitative value calculated based on the degree of difference between the test context and the regular training context; This represents the degree of data heterogeneity adaptation, a quantitative score that measures the compatibility between the source and presentation of test questions.

[0012] For newly added questions in the question bank, the difficulty is pre-labeled by combining similar question information in the federated knowledge graph with transfer learning algorithms, and a question difficulty prediction model trained based on machine learning algorithms is introduced.

[0013] Furthermore, the cognitive modeling module estimates the examinee's mastery of each knowledge point based on a federated knowledge graph, and protects user privacy through collaborative learning of federated nodes. Based on the examinee's mastery of prior and subsequent knowledge points, a three-dimensional cognitive state model is constructed, which includes knowledge mastery status, cognitive ability level, and learning behavior characteristics. The cognitive ability level includes a data processing ability dimension, which includes five types of abilities: logical reasoning, spatial imagination, calculation, memory, and information processing. The model is evaluated by analyzing the examinee's performance on different types of test questions. The learning behavior characteristics include answer time distribution, answer patterns, error type preferences, attention changes, and data interaction behavior. By mining and extracting data from the candidates' answering process, an improved long short-term memory network is used to model the temporal changes in the candidates' cognitive state. A temporal anomaly detection algorithm is introduced to capture the candidates' ability improvement, state fluctuations, and abnormal changes in real time during the test.

[0014] Furthermore, the topic selection decision module models the adaptive testing process as a distributed Markov decision process, using the candidate's three-dimensional cognitive state vector, current testing progress, and data heterogeneity adaptation as the state space, and all available questions in the question bank as the action space. The design incorporates a multi-objective reward function that comprehensively considers ability estimation accuracy, knowledge point coverage, testing efficiency, test taker experience, and data processing efficiency. An adaptive weight allocation algorithm is adopted to adjust the weights of each objective in real time according to the test scenario and test taker status. A distributed dual-deep Q-network (DDQN) architecture is combined with the PPO algorithm for fusion optimization. Priority experience replay mechanism, objective network separation technology, and federated distillation technology are also introduced. The priority experience replay mechanism calculates priority based on the contribution of samples to the model's decision-making. Samples whose answers can significantly change cognitive state, cover weak knowledge points, and correct difficulty biases are given higher priority, while regular, unfluctuating answer samples have lower priority. During replay, high-value samples are selected first to participate in model training, while avoiding training redundancy caused by repeatedly sampling low-value samples. In the horizontal federated architecture, each node independently maintains its local experience replay pool, synchronizing only the priority calculation rules rather than the original sample data.

[0015] We designed a course learning strategy based on federated knowledge graphs. We first pre-trained on simulated test-taker data and then fine-tuned it in a distributed manner on real test data. We introduced an exploration-balancing mechanism that combines Thompson sampling and ε-greedy strategy to explore test questions of different types and difficulty ranges. At the same time, we improved the efficiency of question selection decision-making through distributed computing.

[0016] Furthermore, the data processing module collects multimodal answering process data of candidates in real time, including answering time, mouse trajectory, keyboard input, option modification records, abandonment of answers, and handwritten answer fragments for each question. It uses computer vision, natural language processing, and data mining technologies to perform federated analysis on the collected multimodal data. Each node independently processes local data and only shares the analysis result parameters. Key information including answer patterns, cognitive load, emotional state, and data interaction characteristics is extracted. A federated anomaly detection algorithm based on a combination of rules and machine learning is designed to identify anomalies including guessing, plagiarism, fatigue, lack of concentration, and data tampering, and to perform collaborative verification by federated nodes. The federated node collaborative verification adopts a weighted voting judgment rule. Nodes with larger data scale and higher historical confidence have higher voting weight. When more than half of the nodes determine that it is abnormal, the final abnormal state is confirmed. An abnormality marked by a single node needs to be reviewed by adjacent high-confidence nodes to avoid misjudgment caused by local data deviation.

[0017] For abnormal answers, the difficulty and type of subsequent test questions are automatically adjusted, and appropriate reminders are sent to the test taker. At the same time, based on the test taker's real-time performance, the cognitive state model of the test taker is dynamically updated through a federated synchronization mechanism, and the updated results are fed back to the decision interpretation module.

[0018] Furthermore, the decision explanation module is designed with a federated hierarchical interpretable mechanism. Based on the reasoning ability of the federated knowledge graph, it tracks the complete basis of the distributed reinforcement learning topic selection decision, clarifies the question selection logic, the knowledge points tested by the questions, the relationship with the candidate's current cognitive state and the data processing process, and generates a decision explanation in natural language form. This approach combines template generation with a fine-tuned large language model to generate explanatory texts that conform to the language habits of educational scenarios. The texts include two types of explanations: algorithmic decision-making logic and data processing procedures. For teachers, it provides a decision analysis report, including an analysis of students' knowledge mastery, the difficulty adjustment trajectory during the test, explanations of the selection criteria, records of algorithm parameter adjustments, and data processing logs. For students, it provides feedback, pointing out knowledge gaps, offering learning suggestions, and addressing data interaction issues during the answering process. It also supports interactive queries based on interactive interface technology, making it convenient for users to ask in-depth questions about specific test questions, decision points, or data processing processes.

[0019] Furthermore, the transfer optimization module uses a meta-federated learning method to train the basic reinforcement learning model, enabling the model to adapt to different subjects, grades, and testing scenarios. In the global knowledge graph federated sharing mechanism, an attention-based cross-domain knowledge graph alignment algorithm is designed to enable cross-domain association and alignment of knowledge graphs from different disciplines and scenarios. Anonymized data from all test scenarios are collected for federated learning training. The federated knowledge graph after cross-domain alignment adopts a sharing rule that combines scheduled synchronization and triggered synchronization. Scheduled synchronization synchronizes the aligned cross-domain relationships to all nodes at fixed intervals. Triggered synchronization is executed immediately when new knowledge points are added, new test questions are added, or the confidence of the relationship is updated. During the synchronization process, only incremental relationship data is transmitted, and the entire graph is not transmitted. Each node completes the update based on the local graph and incremental data.

[0020] Design an automated testing framework based on scheduling technology that can automatically evaluate the effects of different algorithm parameters, model structures, and data processing strategies.

[0021] The beneficial effects of this invention are as follows: 1. This invention constructs a five-dimensional quantitative model encompassing cognitive complexity, knowledge relevance, computational complexity, contextual novelty, and data heterogeneity adaptability. It accurately calculates features such as knowledge point path length and node centrality using a federated knowledge graph, and integrates these features with an attention mechanism to generate a comprehensive difficulty vector. Dynamic difficulty correction is achieved through federated Bayesian updates. For new test questions, transfer learning and prediction models are used for pre-calibration. The horizontal federated architecture enables distributed collaborative construction of educational data across multiple scenarios, fully preserving the hierarchy and dependencies of knowledge points. This allows for adaptation to different student groups, teaching progress, and regional differences, improving the match between test difficulty and students' actual abilities.

[0022] 2. The system of this invention constructs a three-dimensional cognitive model based on an improved federated graph attention network, encompassing knowledge mastery status, cognitive ability level, and learning behavior characteristics. The cognitive ability level dimension includes five core abilities: logical reasoning, calculation, and memory. It deeply mines multimodal process data such as answer time, mouse trajectory, and option modification. It uses an improved long short-term memory network to model cognitive temporal changes and combines temporal anomaly detection to identify abnormal states such as fatigue, guessing, and plagiarism. The federated learning architecture avoids centralized transmission of raw data. It can accurately locate knowledge weaknesses and cognitive gaps, and realistically reproduce the fluctuations in the examinee's abilities.

[0023] 3. The system of this invention models the test as a distributed Markov decision process, designs a multi-objective reward function and a distributed dual-deep Q-network fusion PPO algorithm, which takes into account the accuracy of ability estimation, knowledge point coverage and testing efficiency, and avoids the selection of topics getting stuck in local optima; the federated hierarchical interpretable mechanism tracks the basis of topic selection based on knowledge graph, and provides teachers and students with targeted explanations in natural language form; meta-federated learning and cross-domain graph alignment realize rapid adaptation to different subjects, grades and scenarios, and the automated testing framework continuously iterates and optimizes the system performance. Attached Figure Description

[0024] Figure 1 This is the main flowchart of the system of the present invention; Figure 2 This is a flowchart of the topic selection decision-making process for this invention; Figure 3 This is a flowchart of the cognitive modeling and anomaly handling sub-process of the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] like Figures 1 to 3As shown, this embodiment of the invention provides an adaptive test difficulty control system based on reinforcement learning and knowledge graphs, including the following modules: In this embodiment of the invention, the federated graph module constructs a federated knowledge graph containing knowledge point entities, test question entities, cognitive level entities, and multiple types of relationships. It adopts a horizontal federated architecture, distributing and storing knowledge point and test question data from different schools, disciplines, and other educational scenarios. Cross-node graph collaborative construction is achieved solely through parameter sharing. The knowledge point entities include data source identifiers and cross-domain association weight attributes; the test question entities include data heterogeneity tags and adaptive scenario type attributes; and the relationships include cross-domain association strength and dynamic confidence, thereby improving the structured association dimensions between knowledge points and test questions. The horizontal federated architecture adopts a decentralized distributed node networking mode. Each educational scenario corresponds to an independent federated node. The node stores the knowledge points and test questions of its own region locally and does not transmit any privacy information to the outside. When building across nodes, each node first independently completes the initialization and feature extraction of its local knowledge graph, and then uploads the parameter gradient of the graph embedded in the model to the global coordination node in encryption. The global node securely aggregates the parameters of each node and then issues updated parameters. Each node synchronously optimizes its local graph structure according to the global parameters. The nodes adopt an asynchronous communication mechanism, which can complete local updates without waiting for all nodes to be online.

[0027] The dynamic confidence of the association is updated in real time based on the test accuracy of the candidates, the frequency of knowledge point association verification, and the cross-domain adaptation effect. Associations that are repeatedly verified by the test data have gradually increased confidence, while associations that have not been verified for a long time or conflict with the test results have automatically decreased confidence. The confidence value provides a basis for the calculation of graph topological features and the quantification of test difficulty.

[0028] An improved graph convolutional neural network is used to learn the embedding representation of the federated knowledge graph. An attention mechanism and federated distillation technique are introduced to map each knowledge point and question into a low-dimensional dense vector. At the same time, the topological features such as betweenness centrality, proximity centrality and eigenvector centrality of each knowledge point are calculated. The improved graph convolutional neural network adopts a four-level cascaded structure consisting of a two-layer graph convolutional layer, a multi-head attention fusion layer, and a federated feature distillation layer. Fully connected residual connections between layers are used to avoid the gradient vanishing problem. The first graph convolutional layer has an input dimension of 256 (the original feature dimension of the knowledge graph nodes) and an output dimension of 128. The second graph convolutional layer has an input dimension of 128 and an output dimension of 64. The multi-head attention layer uses eight parallel attention heads to capture three types of topological features: knowledge point hierarchy dependency, question association strength, and cross-domain node similarity. The federated distillation layer uses a global center model as the teacher network and distributed node models as student networks. Feature collaboration is achieved between layers through parameter soft links. Model training follows a four-step process: The first step is to initialize the features using the node attribute matrix, adjacency matrix, and cross-domain association weight matrix of the federated knowledge graph as joint inputs. The second step uses the AdamW optimizer for gradient updates, with an initial learning rate set to 1×10⁻⁶. -3 Weight decay coefficient 1×10 -5 The batch size is 32, and the node embedding convergence is achieved through 500 rounds of iterative training. The third step introduces topological feature constraint loss and simultaneously calculates three indices: betweenness centrality of knowledge points, proximity centrality, and eigenvector centrality. The fourth step achieves a smooth fusion of global and local features by using a federated distillation temperature coefficient τ=5, ultimately outputting a knowledge point embedding vector with a dimension of 64 and a question embedding vector with a dimension of 128.

[0029] It supports incremental updates based on user feedback and answer data, and designs a federated incremental update algorithm. When a node discovers a new knowledge point association or test question relationship, it only updates the graph fragment and related parameters of that node. The federated synchronization mechanism realizes dynamic optimization of the global graph, reduces resource consumption during the update process, and improves update efficiency.

[0030] The federated incremental update algorithm adopts the execution logic of local fragment update and global asynchronous synchronization. First, the local node monitors the input of new knowledge points, new test questions or new relationships, automatically locates the local fragments in the federated knowledge graph that need to be updated, and only modifies the node features, association weights and topological attributes of the fragment without affecting other areas of the graph. After the update is completed, a local incremental update package is generated, which includes the changed node identifier, new feature values, and association adjustment values. This package is then transmitted to the global synchronization node through an encrypted channel. Upon receiving the update package, the global node performs conflict detection. If multiple nodes modify the same area simultaneously, they are prioritized according to timestamp and data confidence. After conflict resolution, the update is effectively synchronized to all distributed nodes, ensuring that the global graph structure is consistent and the features are accurate.

[0031] In this embodiment of the invention, the difficulty calibration module constructs a five-dimensional comprehensive calibration system for test question difficulty by combining heterogeneous data processing technology. This system includes cognitive complexity, knowledge relevance, computational complexity, contextual novelty, and data heterogeneity adaptability. The cognitive complexity is based on Bloom's revised taxonomy of cognition and is dynamically adjusted in conjunction with test taker behavior data. The knowledge relevance is quantified by calculating the path length, node centrality, and cross-domain association strength of the knowledge points tested in the test question in the federated knowledge graph. The computational complexity is determined based on the number of reasoning steps, computational load, and data processing difficulty required by the test question. The contextual novelty is calculated based on the similarity between the test question context and the test taker's common training context, and a similarity matching algorithm is used to optimize the calculation accuracy. The data heterogeneity adaptability is used to adapt test question data from different sources and in different formats, such as structured test questions and unstructured subjective question fragments. The five dimensions of difficulty features were all calculated using a standardized quantitative process. Cognitive complexity was based on Bloom's six-level cognitive taxonomy, combined with the memory, comprehension, application, analysis, evaluation, and creativity levels tested in the test questions, and the corresponding scores were matched and dynamically adjusted based on the test takers' correct answer rate. Computational complexity is scored by comprehensively considering the number of reasoning steps, operation types, and data transformations required for each question; the more steps and the more complex the operations, the higher the score. Contextual novelty is achieved by extracting the theme, scene, and questioning style features of the question context and comparing them with the features of conventional questions in the local question bank; the lower the similarity, the higher the novelty score. Data heterogeneity fit is scored based on the format, source, and presentation of the questions; structured questions have lower fit scores, while unstructured subjective questions and multimodal questions have higher fit scores. All five dimensions are normalized to the same range.

[0032] An attention mechanism is used to integrate five dimensions of difficulty features to generate a comprehensive difficulty vector for the test questions. A dynamic difficulty adjustment mechanism based on federated Bayes updates is established. According to the answer data of different test takers, the difficulty values ​​of each dimension of the test questions are adjusted in real time through parameter collaboration between federated nodes, thereby adapting to different test takers and testing scenarios. The federated Bayesian update model uses the Beta distribution as the prior distribution of the five-dimensional difficulty features. The prior parameters are uniformly set to α=2 and β=2. The correct answer rate, average answer time, and error distribution rate of different test taker groups are used as joint likelihood functions to construct a high-precision posterior difficulty estimation model. The update mechanism adopts a distributed node independent calculation and global central weighted aggregation mode. A local node triggers a local posterior update every 100 valid response data collected, and calculates the posterior expectation and variance of the single-dimensional difficulty. In the global aggregation stage, weights are allocated according to the proportion of data volume of each node, and abnormal node data that deviates from twice the global mean are automatically removed to ensure the reliability of the difficulty correction results.

[0033] The model sets three types of dynamic triggering conditions: the overall correct answer rate of the test takers fluctuates by more than 15%, the difference in cross-regional node data exceeds 20%, and the cumulative number of answers after the new test questions are pre-calibrated reaches 50. After the update, a standardized five-dimensional difficulty vector in the interval [0, 1] is output, and the final comprehensive difficulty score is generated by combining the attention weight.

[0034] Federated Bayesian dynamic difficulty posterior estimation formula:

[0035] This represents the posterior estimate of the i-th dimension difficulty feature after being updated by federated Bayes, used for dynamic and precise correction of the difficulty feature; The initial prior parameters of the distribution are represented by fixed values ​​set by the system to construct the prior distribution of the difficulty features; The initial prior parameters of the distribution, together with α, constitute the basic prior distribution of the difficulty feature; This indicates the total number of correct answers given by the target test-taker group for this question; This represents the dimension number of the difficulty feature, i, which indicates the total number of responses generated by the question throughout the entire test process, corresponding to different calculation dimensions of the five-dimensional difficulty.

[0036] The attention mechanism implements adaptive weighted fusion of five-dimensional difficulty features. The system will automatically allocate the weight of each dimension feature according to the current test scenario, the type of candidate group, and the subject attribute. For science questions, the weight of computational complexity and knowledge relevance will be increased. For humanities questions, the weight of cognitive complexity and situational novelty will be increased. For multimodal questions, the weight of data heterogeneity adaptation will be increased. The process automatically filters out low-contribution features and highlights core difficulty factors, resulting in a comprehensive difficulty vector that accurately reflects the true difficulty level of the questions in the corresponding scenarios.

[0037] For newly added questions in the question bank, the difficulty is pre-calibrated by combining similar question information in the federated knowledge graph with transfer learning algorithms. A question difficulty prediction model trained based on machine learning algorithms is introduced to improve the accuracy of pre-calibration.

[0038] The pre-calibration of the difficulty of the new test questions is based on the matching of similar test questions using a federated knowledge graph. Mature test questions with the same knowledge points, cognitive dimensions, and context types are selected first, and their stable difficulty characteristics are transferred to the new test questions as a benchmark. Then, the initial calibration is completed by combining the difficulty prediction model with the five-dimensional basic features of the new test questions. The calibration results are rapidly iterated and corrected with the first batch of answer data for the new test questions.

[0039] In this embodiment of the invention, the cognitive modeling module estimates the candidate's mastery of each knowledge point based on a federated knowledge graph, and avoids centralized transmission of the candidate's original answer data through collaborative learning of federated nodes, thereby protecting user privacy. The assessment of knowledge point mastery is based on the knowledge point dependencies of the federated knowledge graph, combined with the candidate’s performance on the corresponding knowledge point questions and the prerequisite knowledge point questions. Answering highly relevant questions correctly will improve the score of the corresponding knowledge point mastery, while answering them incorrectly will lower the score. The estimation process is completed independently on the local federation node. Only the mastery level score is encrypted and uploaded to participate in global collaboration, without disclosing the specific content of the candidate's answer. At the same time, the score is adjusted based on the depth of the knowledge point level. The higher the level and the more complex the relationship of the knowledge point, the more the estimation result focuses on reasoning ability and comprehensive application performance.

[0040] Based on the examinee's mastery of prior and subsequent knowledge points, a three-dimensional cognitive state model is constructed, which includes knowledge mastery status, cognitive ability level, and learning behavior characteristics. The cognitive ability level includes a data processing ability dimension, which includes five types of abilities: logical reasoning, spatial imagination, calculation, memory, and information processing. The model is evaluated by analyzing the examinee's performance on different types of test questions. The three-dimensional cognitive state model generates a unified state vector by using feature splicing and adaptive weighted fusion. First, the independent features of knowledge mastery, cognitive ability level, and learning behavior characteristics are standardized and spliced. Then, fusion weights are assigned according to the current testing stage and the candidate's performance. In the early stage of the test, the weight of learning behavior characteristics is increased to quickly perceive the candidate's state. In the middle and later stages of the test, the weight of knowledge mastery and cognitive ability is increased to accurately assess the level. Finally, the fusion results in a coherent cognitive state vector.

[0041] Three-dimensional candidate cognitive state fusion vector:

[0042]

[0043] This represents the candidate's three-dimensional cognitive state vector, which serves as the core state input for adaptive question selection decision-making. The adaptive fusion weights, representing the characteristics of knowledge mastery status, are dynamically adjusted as the testing progresses. Adaptive fusion weights representing cognitive ability level characteristics are automatically assigned based on the examinee's real-time status; The adaptive fusion weights represent the learning behavior features and adapt to the state changes throughout the testing process. This represents a knowledge mastery state vector, comprehensively reflecting the candidate's mastery of each knowledge point in the federated knowledge graph; A vector representing cognitive ability levels, comprising scores for five core abilities: logical reasoning, spatial imagination, computation, memory, and information processing. It represents the feature vector of learning behavior, and fully quantifies the candidate's answering time, answering pattern, error type, attention and data interaction behavior.

[0044] The learning behavior characteristics include answer time distribution, answer patterns, error type preferences, attention changes, and data interaction behaviors, such as mouse trajectory and keyboard input rhythm. By mining and extracting data from the candidates' answering process, an improved long short-term memory network is used to model the temporal changes in the candidates' cognitive state. A temporal anomaly detection algorithm is introduced to capture in real time the candidates' ability improvement, state fluctuations, and abnormal changes during the test, such as sudden changes in cognitive state caused by fatigue or lack of concentration.

[0045] The improved Long Short-Term Memory (LSTM) network adds cognitive state gating and temporal attention gating to the traditional structure, forming a five-gate collaborative architecture consisting of input gate, forget gate, output gate, cognitive state gate, and temporal attention gate. The network adopts a two-layer stacked structure with a single hidden layer dimension of 256, and dropout regularization is added between layers to improve generalization ability. The temporal input window is fixed at 20 time steps, and each time step includes five features: response time, answer accuracy, error type, operation trajectory, and cognitive load. Model training consists of two steps: local modeling and federated collaboration. The first step uses the sequential characteristics of the examinee's continuous responses as input, and employs the MSE loss function to fit the trajectory of cognitive state changes. The Adam optimizer has an initial learning rate of 5×10⁻⁶. -4 Iterates back along the time step for 300 rounds; The second step combines a Gaussian mixture model to detect temporal anomalies. The anomaly threshold is set at three times the standard deviation of the feature deviation from the mean. Three types of situations are identified: sudden changes in cognitive state, fatigued responses, and abnormal fluctuations. Finally, a 15-dimensional three-dimensional cognitive state temporal vector is output, and the anomaly type and confidence level are labeled simultaneously.

[0046] Once an abnormal temporal state is identified, the cognitive state model is immediately corrected. The regular temporal fitting calculation logic is paused, and the current cognitive state value is recalculated after the anomaly compensation algorithm is used to remove interference data. The corrected state data is then synchronized to the federated node.

[0047] In this embodiment of the invention, the topic selection decision module models the adaptive testing process as a distributed Markov decision process, using the candidate's three-dimensional cognitive state vector, the current testing progress, and the data heterogeneity adaptation degree as the state space, and all available questions in the question bank as the action space. The state transition of the distributed Markov decision process is executed based on the candidate's real-time answer results and changes in cognitive state. Each time a question is answered, a state transition is triggered. The cognitive state vector is updated based on the correctness of the answer, the answer duration, and the abnormal state marker. The test progress is advanced synchronously, and the data heterogeneity adaptation is adjusted. The state transition process does not rely on global centralized computation. Each federated node independently completes the local state update, and then global state consistency is achieved through parameter encryption synchronization.

[0048] The design incorporates a multi-objective reward function that comprehensively considers ability estimation accuracy, knowledge point coverage, testing efficiency, test taker experience, and data processing efficiency. An adaptive weight allocation algorithm is adopted to adjust the weights of each objective in real time according to the test scenario and test taker status. A distributed dual-depth Q-network architecture is combined with the PPO algorithm for fusion optimization. Priority experience replay mechanism, objective network separation technology, and federated distillation technology are introduced to improve the stability, convergence speed, and sample efficiency of the algorithm. The fusion algorithm employs a dual-parallel structure of a distributed DDQN backbone network and a PPO policy optimization branch. DDQN includes two independent branches: an evaluation network and a target network, each with three fully connected layers of dimensions 512, 256, and 128 respectively. The activation function is GELU, and the output layer corresponds to the probability distribution of test actions. The PPO branch has a pruning coefficient ε=0.2 and a policy learning rate of 3×10⁻⁶. -4 Value learning rate 1×10 -3 Experience replay pool capacity 1×10 5 Priority sampling weight α=0.6, importance sampling coefficient β=0.4; The training process follows federated distributed logic: it takes an 18-dimensional state space as input and a question bank as the action space; it constructs a multi-objective reward function and adaptively allocates weights for ability accuracy, knowledge point coverage, testing efficiency, and examinee experience; it synchronizes dual network parameters every 100 steps and shares them through federated distillation every 50 rounds; it adopts Thompson sampling and ε-greedy balance exploration and utilization to finally output the optimal question number and question selection confidence.

[0049] The adaptive weight allocation algorithm dynamically adjusts the reward weights according to the test progress stage. In the early stage of the test, it focuses on the coverage of knowledge points and the breadth of exploration, expanding the scope of the test questions to quickly locate the candidate's ability range. In the middle stage of the test, it focuses on the accuracy of ability estimation, accurately matching the candidate's cognitive level to improve the accuracy of the assessment. In the later stage of the test, it focuses on the test efficiency and candidate experience, shortening the answering time and avoiding sudden increases or decreases in difficulty. At the same time, it temporarily adjusts the weights for abnormal answering states to reduce the interference of abnormal data on reward calculation.

[0050] We designed a course learning strategy based on federated knowledge graphs. We first pre-trained on simulated test-taker data and then fine-tuned it in a distributed manner on real test data. We introduced an exploration-balancing mechanism that combines Thompson sampling and ε-greedy strategy to explore different types of test questions with different difficulty ranges, avoiding over-concentration on certain knowledge points or difficulty ranges. At the same time, we improved the efficiency of question selection decision-making through distributed computing.

[0051] The exploration-utilization balancing mechanism employs a dynamic switching rule. In the early stages of testing, an ε-greedy strategy is primarily used to maintain a high exploration probability, selecting a wide range of questions with different knowledge points and difficulties to quickly build a profile of the candidate's abilities. In the later stages of testing, the strategy switches to Thompson sampling as the primary method, selecting questions precisely based on the candidate's cognitive state to reduce the exploration probability and improve utilization efficiency. Simultaneously, the switching nodes are fine-tuned in real time based on the stability of the candidate's answers, appropriately increasing the exploration ratio when the state fluctuates greatly and enhancing utilization accuracy when the state is stable.

[0052] The course learning strategy strictly follows the knowledge point hierarchy and dependencies of the federated knowledge graph, and promotes model training in a cognitive progression order from easy to difficult and from basic to comprehensive. In the pre-training stage, gradient simulation data is generated based on the graph to enable the model to first master the basic knowledge point selection logic. In the fine-tuning stage, the model is adapted to the cognitive distribution of real test takers.

[0053] In this embodiment of the invention, the data processing module collects multimodal answering process data of the examinee in real time, including the answering time, mouse trajectory, keyboard input, option modification records, abandonment of answering, and handwritten answer fragments for each question. The collected multimodal data is analyzed in a federated manner using computer vision, natural language processing, and data mining techniques. Each node independently processes local data and only shares the analysis result parameters. The federated multimodal data analysis is performed in modules and steps. The computer vision module processes handwritten answer fragments and operation trajectories to extract features of writing coherence and trajectory stability; the natural language processing module processes text answer content to extract features of answer completeness and logical correlation; and the data mining module processes structured data such as answer duration, option modification, and input rhythm to extract behavioral pattern features. All analysis steps are completed on the local node. Only the extracted high-dimensional feature parameters are encrypted and transmitted, and the original images, text, and operation records are not retained.

[0054] Key information including answer patterns, cognitive load, emotional state, and data interaction characteristics is extracted. A federated anomaly detection algorithm based on a combination of rules and machine learning is designed to identify anomalies such as guessing, plagiarism, fatigue, lack of concentration, and data tampering. The algorithm also performs collaborative verification by federated nodes to improve the accuracy of anomaly detection. The abnormal answer detection algorithm adopts a two-layer collaborative architecture of rule engine and lightweight classification model. The upper rule engine is responsible for quickly identifying obvious anomalies and generating preliminary anomaly markers based on preset answer duration thresholds, option modification frequency, operation trajectory continuity, and answer jump logic. The lower machine learning model uses a lightweight random forest classifier, taking multimodal answer features as input, including 12-dimensional features such as answer duration distribution, mouse movement speed, keyboard input interval, number of option switching, proportion of abandoned answers, and handwritten answer coherence. The model completes training and inference on local nodes and only uploads the anomaly probability value to global nodes. The federated collaborative verification adopts a multi-node voting mechanism. When the confidence level of a single node marking an anomaly is greater than 0.8, it is directly determined to be an anomaly. When the confidence level is between 0.5 and 0.8, it triggers cross-validation of data from adjacent nodes. The final anomaly status is determined by combining the results from multiple parties.

[0055] For abnormal answers, the difficulty and type of subsequent test questions are automatically adjusted, and appropriate reminders are sent to the test taker. At the same time, based on the test taker's real-time performance, the cognitive state model of the test taker is dynamically updated through a federated synchronization mechanism, and the updated results are fed back to the decision interpretation module.

[0056] In this embodiment of the invention, the decision explanation module is designed with a federated hierarchical interpretable mechanism. Based on the reasoning ability of the federated knowledge graph, it tracks the complete basis of the distributed reinforcement learning topic selection decision, clarifies the question selection logic, the knowledge points tested by the questions, the correlation with the candidate's current cognitive state and the data processing process, and generates a decision explanation in natural language form. The federated hierarchical interpretable mechanism is divided into three levels: basic interpretation, in-depth interpretation, and source interpretation. The basic interpretation is for test takers and only explains the reasons for the matching of the knowledge points and difficulty of the test questions. The in-depth interpretation is for teachers and supplements the algorithm logic of question selection, the basis for matching cognitive state, and the difficulty adjustment process. The source interpretation is for system operation and maintenance personnel and presents the data processing flow, model parameter changes, and federated collaboration records. The three layers of explanation share the same set of decision-making criteria, and the information details are unfolded layer by layer by linking test questions, knowledge points, and cognitive state data through a federated knowledge graph.

[0057] The method combines template generation with a fine-tuned large language model to generate explanatory texts that conform to the language habits of educational scenarios. It includes two types of explanatory content: algorithm decision logic and data processing process. For teacher users, it provides decision analysis reports, including analysis of candidates' knowledge mastery, difficulty adjustment trajectory during the test, explanation of the basis for question selection, algorithm parameter adjustment records and data processing logs. The explanation generation mechanism adopts a combination of hierarchical template matching and large-scale model semantic enhancement. The bottom layer builds a fixed explanation template library, covering four major scenarios: topic selection basis, difficulty adjustment, knowledge point examination, and cognitive state matching. Each type of template has a preset standard sentence framework and variable filling position. The upper layer uses a large language model that has been finely tuned in the education field to perform semantic polishing and logical optimization on the text after template filling, so that the explanation content is more in line with the reading habits of teachers and students. The model takes as input parameters the selection of questions, the characteristics of the knowledge points in the questions, and the cognitive state data of the test takers. The output is a hierarchical natural language explanation. The explanation for students focuses on the knowledge gaps and answer suggestions, and is concise, easy to understand and free of technical jargon. The explanation for teachers supplements the details of the algorithm decision-making process, the parameter adjustment process and the data basis, and also supports interactive follow-up questions and responses.

[0058] For student users, provide concise and clear feedback, pointing out knowledge gaps, learning suggestions, and data interaction issues during the answering process, such as the correlation between answering speed, error types, and data processing ability; It also supports interactive queries based on interactive interface technology, making it convenient for users to ask in-depth questions about specific test questions, decision points, or data processing processes.

[0059] In this embodiment of the invention, the transfer optimization module uses a meta-federated learning method to train the basic reinforcement learning model, which combines the rapid adaptability of meta-learning with the privacy protection advantages of federated learning, enabling the model to adapt to different subjects, different grades, and different testing scenarios. The meta-federated learning training process is divided into three stages: global meta-training, local node fine-tuning, and federated parameter aggregation. In the first stage, the basic reinforcement learning model is meta-trained using multi-disciplinary general simulation data at the global central node. This process teaches the model general question selection logic and difficulty adaptation rules, enabling it to master core decision-making capabilities across scenarios. In the second stage, the trained basic model is distributed to each local node. Each node uses local subject-specific data for small-sample fine-tuning, requiring only a small amount of real test data to quickly adapt to the knowledge point system and question characteristics of the current scenario. In the third stage, each node uploads the fine-tuned model parameters, and the global node uses a weighted average method to complete federated aggregation and update the global basic model.

[0060] In the global knowledge graph federated sharing mechanism, an attention-based cross-domain knowledge graph alignment algorithm is designed to enable cross-domain association and alignment of knowledge graphs from different disciplines and scenarios, realize the fusion and transfer of cross-disciplinary and cross-scenario knowledge, collect anonymized data from all test scenarios for federated learning training, and optimize the overall performance of the system. The cross-domain knowledge graph alignment algorithm is based on a single-head adaptive attention alignment layer and uses node embedding cosine similarity as the matching criterion. The similarity threshold is set to 0.75. The algorithm executes a four-step standardized process: First, it extracts three types of features from the source and target graphs: node embeddings, topological relationships, and hierarchical attributes. Second, it assigns differentiated weights to three types of entities—knowledge points, test questions, and cognitive levels—through an attention mechanism to enhance the matching accuracy of core knowledge points. Third, it traverses all node pairs and selects node pairs with a similarity ≥ 0.75 to establish temporary cross-domain associations. Fourth, it synchronizes the effective associations to the global graph through federated incremental updates, updates the cross-domain association strength and dynamic confidence, and automatically removes weak associations with a confidence level below 0.3. The algorithm combines a meta-federated learning framework, supporting rapid adaptation to cross-disciplinary, cross-grade, and cross-regional scenarios. In new scenarios, stable inference results can be achieved with only 50 rounds of fine-tuning.

[0061] Design an automated A / B testing framework based on scheduling technology, which can automatically evaluate the effects of different algorithm parameters, model structures and data processing strategies, and achieve continuous self-optimization of the system.

[0062] The automated A / B testing framework adopts a combined operation mode of timed scheduling and trigger-based scheduling. Timed scheduling compares the evaluation accuracy and topic selection efficiency of different algorithm parameters at fixed intervals, while trigger-based scheduling starts testing when the system introduces new disciplines or new scenarios. The framework automatically assigns test and control groups, collects multiple evaluation indicators such as capability estimation accuracy, knowledge point coverage, test duration, and candidate satisfaction, determines the optimal solution through weighted scoring of indicators, and automatically synchronizes the optimal parameters and structure to all federated nodes.

[0063] All artificial intelligence models within the system have established standardized input and output data interfaces and dimensional specifications, clarifying the inherent relationship between data flow between modules. All data has undergone de-identification preprocessing and normalization mapping processing, strictly adhering to the educational data privacy protection and federated learning transmission specifications. The global input data of the model includes five categories: anonymous test taker identifiers, knowledge point mastery sequence features, multimodal answer process data, original five-dimensional difficulty features of test questions, and topological structure features of federated knowledge graphs; the global output data of the model includes six categories: comprehensive difficulty vector of test questions, three-dimensional cognitive state vector of test takers, optimal question selection number, cognitive abnormal state marker, ability assessment report, and natural language explanation of question selection decision. The core data dimensions are uniformly standardized as follows: 1×15 three-dimensional cognitive state vector, 1×5 five-dimensional difficulty vector of test questions, 1×64 knowledge point embedding vector, and 1×128 test question embedding vector. The modules are transmitted via direct vector connection.

[0064] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0065] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An adaptive test difficulty control system based on reinforcement learning and knowledge graphs, characterized in that, It includes a federated graph module, a difficulty labeling module, a cognitive modeling module, a topic selection decision module, a data processing module, a decision interpretation module, and a transfer optimization module; The federated graph module constructs a federated knowledge graph containing entities and multiple types of relationships. It adopts a horizontal federated architecture to achieve cross-node collaborative construction and distributed storage, completes graph embedding learning and topological feature calculation, and realizes dynamic optimization of the global graph based on answer data and user feedback. The difficulty calibration module constructs a multi-dimensional comprehensive difficulty calibration system for test questions, generates a comprehensive difficulty vector, and establishes a dynamic difficulty adjustment mechanism to pre-calibrate the difficulty of new test questions. The cognitive modeling module constructs a federated three-dimensional cognitive state model of the examinee, estimates the examinee’s mastery of knowledge points based on the federated knowledge graph, and captures the temporal changes in cognitive state. The topic selection decision module is designed with a distributed reinforcement learning architecture that integrates knowledge graphs. It models adaptive testing as a distributed Markov decision process and uses a fusion optimization algorithm combining a distributed dual-deep Q network and PPO to optimize topic selection decisions. The data processing module uses a federated analysis method to process candidates' multimodal response data, enabling abnormal response detection and dynamic updating of the cognitive state model; The decision interpretation module is designed with a federated hierarchical interpretable mechanism, which tracks the decision-making basis for topic selection based on a federated knowledge graph and generates natural language interpretation. The migration optimization module designs a meta-federated learning architecture and a cross-domain knowledge graph alignment algorithm to achieve cross-scenario knowledge fusion and migration and continuous system optimization.

2. The adaptive test difficulty control system based on reinforcement learning and knowledge graph as described in claim 1, characterized in that, The federated graph module constructs a federated knowledge graph containing knowledge point entities, test question entities, cognitive level entities, and multiple types of relationships. It adopts a horizontal federated architecture to distribute and store knowledge point and test question data from different educational scenarios. It achieves cross-node graph collaborative construction by sharing only model parameters and not transmitting raw data. The knowledge point entity includes a data source identifier and a cross-domain association weight attribute; the test question entity includes a data heterogeneity label and an adaptation scenario type attribute; and the association relationship includes cross-domain association strength and dynamic confidence.

3. The adaptive test difficulty control system based on reinforcement learning and knowledge graph as described in claim 2, characterized in that, The federated graph module uses an improved graph convolutional neural network optimized for knowledge graph embedding to learn the embedding representation of the federated knowledge graph. It introduces an attention mechanism and federated distillation technology to map each knowledge point and question into a low-dimensional dense vector, while calculating the topological features of each knowledge point. The designed federated incremental update algorithm updates only the graph fragments and related parameters of the corresponding node when a node discovers new knowledge point associations or test question examination relationships. The global graph is dynamically optimized through the federated synchronization mechanism.

4. The adaptive test difficulty control system based on reinforcement learning and knowledge graph as described in claim 3, characterized in that, The difficulty calibration module combines heterogeneous data processing technology to construct a five-dimensional comprehensive calibration system for test question difficulty, including cognitive complexity, knowledge relevance, computational complexity, contextual novelty, and data heterogeneity adaptability. The cognitive complexity is based on Bloom's revised taxonomy of cognitive objectives and is dynamically adjusted in conjunction with test taker behavior data; the knowledge relevance is quantified by calculating the path length, node centrality, and cross-domain association strength of the knowledge points tested in the test questions in the federated knowledge graph.

5. The adaptive test difficulty control system based on reinforcement learning and knowledge graphs according to claim 4, characterized in that, The difficulty calibration module uses an attention mechanism to fuse five dimensions of difficulty features to generate a comprehensive difficulty vector for the test questions. It establishes a dynamic difficulty adjustment mechanism based on federated Bayesian updates. According to the answer data of different test takers, the difficulty values ​​of each dimension of the test questions are corrected in real time through federated Bayesian updates and inter-node parameter coordination. For newly added test questions, the difficulty pre-calibration is completed by using similar test question information in the federated knowledge graph and combining it with transfer learning algorithms.

6. The adaptive test difficulty control system based on reinforcement learning and knowledge graph as described in claim 5, characterized in that, The cognitive modeling module estimates the candidate's mastery of each knowledge point based on a federated knowledge graph, and protects user privacy through collaborative learning of federated nodes. A three-dimensional cognitive state model is constructed, which includes knowledge mastery status, cognitive ability level, and learning behavior characteristics. The cognitive ability level includes five types of abilities: logical reasoning, spatial imagination, calculation, memory, and information processing. The learning behavior characteristics include dimensions such as answering time, answering pattern, and error type. Based on a federated knowledge graph, the assessment is completed by analyzing the candidate's performance on different types of test questions.

7. The adaptive test difficulty control system based on reinforcement learning and knowledge graph as described in claim 6, characterized in that, The cognitive modeling module obtains learning behavior characteristics, including answer time distribution, answer patterns, error type preferences, attention changes, and data interaction behavior, by mining and extracting data from the multimodal answering process of test takers. An improved long short-term memory network is used to model the temporal changes in the cognitive state of test takers. A temporal anomaly detection algorithm oriented towards cognitive state is introduced, and the improved long short-term memory network is used to capture the temporal changes and abnormal fluctuations in the cognitive state of test takers in real time.

8. The adaptive test difficulty control system based on reinforcement learning and knowledge graph as described in claim 7, characterized in that, The topic selection decision module models the adaptive testing process as a distributed Markov decision process, with the candidate's three-dimensional cognitive state vector, the current testing progress and the data heterogeneity adaptation degree forming the system state space, and all available questions in the question bank as the action space. The design incorporates a multi-objective reward function that comprehensively considers ability estimation accuracy, knowledge point coverage, testing efficiency, test taker experience, and data processing efficiency. An adaptive weight allocation algorithm is used to adjust the weights of each objective in real time based on the testing scenario and the test taker's status.

9. The adaptive test difficulty control system based on reinforcement learning and knowledge graph as described in claim 8, characterized in that, The topic selection decision module adopts a distributed dual-depth Q network architecture combined with the PPO algorithm for fusion optimization, and introduces a priority experience replay mechanism, target network separation technology and federated distillation technology. We designed a course learning strategy based on federated knowledge graphs, which was first pre-trained on simulated test-taker data and then fine-tuned in a distributed manner on real test data. We introduced an exploration-utilization balance mechanism that combines Thompson sampling and ε-greedy algorithm.

10. The adaptive test difficulty control system based on reinforcement learning and knowledge graph as described in claim 9, characterized in that, The data processing module collects multimodal answer process data of candidates in real time, and uses computer vision, natural language processing and data mining technology to perform federated analysis on the collected multimodal data. Each node processes local data independently and only shares the analysis result parameters. Design a federated anomaly detection algorithm based on a combination of rules and machine learning, which can identify anomaly responses and perform collaborative verification by federated nodes; The decision explanation module uses a combination of template generation and fine-tuned large language model to generate explanation text, providing differentiated feedback for different users and supporting interactive queries; The transfer optimization module uses a meta-federated learning method to train the basic reinforcement learning model, designs a cross-domain knowledge graph alignment algorithm based on the attention mechanism, establishes a global knowledge graph federated sharing mechanism, and designs an automated testing framework to realize the automated evaluation and iteration of algorithm parameters and model structure.

Images