Classroom personalized learning path dynamic generation system based on artificial intelligence

Abstract

The invention discloses a classroom personalized learning path dynamic generation system based on artificial intelligence, and relates to the technical field of artificial intelligence and educational informationization, the system comprises a multi-modal sensing module, a neural symbol causal diagram construction module, an anti-factual reasoning and robustness evaluation module and a dynamic path generation module, and adopts a federal learning architecture. The multi-mode sensing module collects related data of physiological states, environments and learning behaviors of students; the neural symbol causal graph construction module fuses symbol rules and a neural network, and constructs a causal graph model containing multiple types of nodes; the anti-factual reasoning and robustness evaluation module peels off the influence of confusion factors and outputs a robustness mastering index and a knowledge gap root; a dynamic path generation module generates and adjusts personalized learning paths according to scenes; according to the method, the scientificity and the safety of personalized teaching are improved, and the method is suitable for a classroom personalized teaching scene.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and educational informatization, specifically to an AI-based system for dynamically generating personalized learning paths in the classroom. Background Technology

[0002] With the deepening application of artificial intelligence technology in education, personalized learning has become a key approach to improving teaching effectiveness. Its core lies in dynamically generating learning paths tailored to individual student characteristics by analyzing student learning data, thereby breaking the limitations of traditional standardized teaching and achieving individualized instruction. Currently, many technical solutions are dedicated to analyzing students' knowledge mastery using knowledge graphs, deep learning algorithms, and other methods, and recommending learning content accordingly. However, existing systems still face significant challenges in achieving true personalization and robustness.

[0003] In existing technologies, numerous related patents offer different solutions. For example, patent CN120338082A provides an "Intelligent Personalized Learning Path Recommendation System." This system synchronizes cross-platform learning behavior data in real time through a multi-source data acquisition module, predicts learning needs using a cognitive feature analysis engine, and finally dynamically adjusts the path through a reinforcement learning framework. Its shortcoming lies in the fact that this system mainly relies on explicit behavioral data for analysis and fails to fully consider the interference of various confounding factors in the classroom environment (such as students' real-time physiological state, emotional fluctuations, environmental noise, etc.) on the evaluation of learning performance, which may lead to misjudgments of students' true knowledge status. Another patent, CN120256721A, involves an "Artificial Intelligence-Based Personalized Learning Path Recommendation System." This solution uses knowledge graph and graph neural network technology to mine the associations of knowledge points and uses a sequence recommendation model to predict learning paths. However, this patent also fails to solve the problem of how to distinguish between a student's temporary decline in performance due to external environmental interference and a persistent weakness due to a lack of mastery of the essence of knowledge; the robustness and interpretability of its recommendation logic need to be strengthened.

[0004] Therefore, there is an urgent need for a new technological solution that can effectively address the core issue of how environmental and physiological confounding factors interfere with the assessment of knowledge status. An ideal system should be able to isolate the influence of these external factors, accurately assess students' true level of knowledge mastery, and trace the root causes of knowledge gaps, thereby providing a solid foundation for generating highly reliable and personalized learning paths.

[0005] In summary, existing technologies largely focus on analyzing historical behavioral data and the correctness of answers, lacking a deep fusion of multimodal information and causal perception of the learning process, particularly in distinguishing the influence of confusion factors from the true cognitive state. This limits the accuracy, robustness, and interpretability of personalized learning system recommendations. This invention aims to overcome these shortcomings. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an AI-based dynamic generation system for personalized classroom learning paths. This system collects data on students' physiological states, environment, and learning behaviors through a multimodal perception module. It then integrates symbolic rules of prior educational knowledge with neural network reasoning through a neural symbolic causal graph construction module to build a causal graph model containing multiple types of nodes. Next, a counterfactual reasoning and robustness assessment module intervenes, calculates counterfactual results, removes the influence of confounding factors, outputs a robustness mastery index, and identifies the root causes of knowledge gaps. Simultaneously, a dynamic path generation module generates and adjusts personalized learning paths according to different scenarios, while a federated learning architecture ensures data privacy. This allows for accurate assessment of students' true knowledge mastery levels and the generation of suitable personalized learning paths.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: On the one hand, an artificial intelligence-based dynamic generation system for personalized classroom learning paths, the system comprising: a multimodal perception module, a neural symbol causal graph construction module, a counterfactual reasoning and robustness assessment module, and a dynamic path generation module;

[0008] The multimodal perception module is used to collect students' physiological state data, environmental data, and learning behavior data.

[0009] The neural symbol causal graph construction module is used to construct and maintain a causal graph model containing knowledge point nodes, physiological state nodes, environmental nodes, and learning performance nodes, wherein the causal relationships between nodes are constrained by predefined symbol rules and causal discovery algorithms.

[0010] The counterfactual reasoning and robustness assessment module is used to perform intervention operations based on the causal graph model, calculate counterfactual results, eliminate the influence of confounding factors, and output the student's robustness mastery index.

[0011] The dynamic path generation module is used to dynamically generate and adjust a personalized learning path sequence based on the robustness mastery index and the causal root causes of knowledge gaps located by counterfactual reasoning.

[0012] Furthermore, the neural symbol causal graph construction module specifically includes:

[0013] The neural network feature extraction unit is used to extract features from the raw data collected by the multimodal perception module to obtain corresponding concept vectors. The concept vectors include knowledge point mastery vectors, physiological state vectors, environmental state vectors, and learning performance vectors.

[0014] The symbol rule injection unit stores symbol rules for prior educational knowledge. These symbol rules are used to define the logical relationships between knowledge points and the constraints of physiological and environmental factors on learning performance. The symbol rule injection unit converts the symbol rules into graph structure constraints.

[0015] The causal graph synthesis unit receives the concept vector and the graph structure constraints, and constructs the neural symbolic causal graph model by combining it with the causal discovery algorithm. Nodes in the causal graph model correspond to the concept vectors, and edges represent the causal relationships between nodes. The causal graph synthesis unit achieves the fusion of neural reasoning and symbolic logic through the following computational process:

[0016] In the message passing process of a graph neural network, nodes to neighboring nodes The message being delivered Not only by the neural network parameter weights and node features , The decision is made, and is also constrained by the symbolic rule function. The output of , whose message calculation formula is:

[0017]

[0018] in, It is a non-linear activation function. This represents a vector concatenation operation. The sigmoid function is used to map the output of a rule function to weight coefficients. It is a function based on predefined symbol rules, if the node and If at least one of the causal relationship and logical relationship defined by the rules exists, output a positive value; otherwise, output a negative value.

[0019] Furthermore, the counterfactual reasoning and robustness assessment module specifically includes:

[0020] An intervention execution unit is used to perform intervention operations on a specified confusion factor node in the neural symbol causal graph model, and set the value of the confusion factor node to a specific intervention value, thereby simulating the state in an ideal, interference-free target environment.

[0021] The counterfactual calculation unit is used to perform forward reasoning based on the modified graphical model after the intervention is implemented, and to calculate the predicted value of learning performance under counterfactual conditions.

[0022] The effect quantification unit is used to compare the actual observed learning performance value with the predicted learning performance value under the counterfactual condition, and to calculate the average causal effect of the confusion factor.

[0023] The robustness mastery index calculation unit is used to assess the stability of students' knowledge mastery under various simulated intervention environments. The robustness mastery index calculation unit calculates knowledge points using the following formula. Robustness mastery index :

[0024]

[0025] in, This represents the total number of simulated environmental intervention scenarios. Indicates the environment variable Implementation of the Intervention, setting its value to , This indicates the knowledge points under the intervention. The predicted performance value, To represent the acceptable threshold, This is an indicator function that returns 1 if the condition within the parentheses is true, and 0 otherwise.

[0026] Furthermore, the counterfactual reasoning and robustness assessment module also includes a causal root cause tracing unit. This unit is used to locate the root cause knowledge points leading to the knowledge gap. This unit achieves tracing by iteratively generating counterfactual questions and performing reasoning. The process is as follows:

[0027] For the target knowledge points that are not yet mastered The system automatically generates a hypothetical question, the core of which is: assuming the student has fully mastered the prerequisite knowledge points. To determine whether they can master the target knowledge points. ;

[0028] Subsequently, the knowledge point nodes in the neural symbolic causal graph model were analyzed. Implement intervention, setting its level of mastery to the maximum value;

[0029] Based on the intervention, the target knowledge points are re-reasoned. The degree of mastery;

[0030] If after intervention If the mastery reasoning result significantly improves and exceeds a predetermined threshold, then it is determined that... for The causal roots of the knowledge gap;

[0031] This process proceeds iteratively in reverse along the cause-effect graph until the most fundamental weakness in prior knowledge is found.

[0032] Furthermore, the decision-making logic of the dynamic path generation module is based on the robustness mastery index and the output of the causal root cause tracing unit;

[0033] The dynamic path generation module first determines whether the low robustness mastery index is mainly caused by the confusion factor effect;

[0034] If the average causal effect of the confusion factor is greater than the first set threshold, it is determined that the current knowledge state assessment is significantly disturbed. In this case, the generated path will prioritize the recommendation of auxiliary learning content to reduce the impact of the confusion factor.

[0035] If the average causal effect of the confusion factor is less than the first set threshold and the robustness mastery index is lower than the second set threshold, then it is determined that there is a real lack of knowledge mastery. At this time, the dynamic path generation module receives the root cause knowledge point located by the causal root cause tracing unit and generates a backtracking learning path sequence that starts from the root cause knowledge point and gradually leads to the target knowledge point.

[0036] Furthermore, the effect quantification unit employs a dual machine learning model to estimate the average causal effect of the confounding factor in order to achieve unbiased estimation;

[0037] The estimation process of the dual machine learning model consists of the following steps:

[0038] The first step is to use a machine learning model. Fitting Confusion Factor For covariates Based on the dependencies, the predicted value is obtained. And calculate the residuals ;

[0039] The second step is to use another machine learning model. Fitting outcome variables For covariates Based on the dependencies, the predicted value is obtained. And calculate the residuals ;

[0040] The third step is to use the residuals. For residuals Perform linear regression and obtain the regression coefficients This is the confusion factor. Regarding the results The estimated average causal effect is calculated using the following formula:

[0041]

[0042] in, Represents the confounding factor to be evaluated. Represents learning performance. Representative except In addition to all other covariates, including other confounding factors and knowledge state, Describing covariance, Indicates variance.

[0043] Furthermore, the multimodal sensing module includes:

[0044] A non-invasive vision sensor is used to acquire video streams of students' faces and calculates the head posture offset angle and eye gaze direction in real time through a lightweight facial feature point detection model and an eye state classification model.

[0045] An audio sensor is used to collect classroom environment sounds and student voices, and to distinguish between teacher voices and student voices and identify the emotional polarity in student voices through a voice activity detection and emotion recognition model.

[0046] Environmental sensors are used to collect data on classroom temperature, humidity, and ambient noise levels in decibels.

[0047] The physiological signal sensor uses a low-power EEG signal acquisition unit integrated into a head-mounted device to collect raw EEG signals from students and calculates attention index and cognitive load index in real time through embedded algorithms.

[0048] The learning behavior log collection unit is used to obtain data on students' answers, answer time, and number of interactions on the learning platform in real time through the application programming interface.

[0049] Furthermore, the system employs a federated learning architecture for model training and updating;

[0050] The federated learning architecture comprises a central server and multiple edge clients distributed across different schools;

[0051] The edge client locally stores the desensitized multimodal data of its students and uses the local data to train the neural network components in the neural symbol causal graph model to generate gradient information of the model parameters.

[0052] The central server periodically aggregates encrypted gradient information from each edge client, updates the global model parameters using a secure averaging algorithm, and then distributes the updated global model parameters to each edge client.

[0053] This process is repeated cyclically, enabling the system to optimize performance using data from multiple sources while ensuring that the original student data remains local and does not compromise privacy.

[0054] Compared with existing technologies, this AI-based system for dynamically generating personalized learning paths in the classroom has the following advantages:

[0055] I. This invention collects students' physiological state data, environmental data, and learning behavior data through a multimodal perception module. It then integrates symbolic rules of prior educational knowledge with data-driven neural network reasoning using a neural symbolic causal graph construction module to construct a causal graph model containing knowledge points, physiological states, environment, and learning performance nodes. Finally, a counterfactual reasoning and robustness assessment module intervenes in the confounding factor nodes within the causal graph model, calculating counterfactual learning performance predictions and comparing them with actual performance values ​​to eliminate the influence of confounding factors. This results in a robustness mastery index and pinpoints the causal root causes of knowledge gaps, thereby accurately assessing students' true knowledge mastery level. This effectively solves the technical problems of existing personalized learning systems, such as biased knowledge state assessments due to insufficient consideration of environmental and physiological confounding factors, and the inability to accurately trace the root causes of knowledge gaps. It provides accurate data and model support for generating reliable personalized learning paths, enhancing the scientific rigor and effectiveness of personalized teaching.

[0056] Second, this invention generates personalized learning paths based on the robustness mastery index and the causal root causes of knowledge gaps through a dynamic path generation module, categorized by scenario. When confusion factors have a significant impact, auxiliary learning content that reduces interference is prioritized. When real knowledge gaps exist, a backtracking path is generated, gradually leading from root knowledge points to target knowledge points, and can be dynamically adjusted according to students' real-time learning status. Simultaneously, a federated learning architecture is adopted, where edge clients train models locally and upload encrypted gradients, while the central server aggregates and updates the global model before distributing it, ensuring that original student data remains locally. This effectively solves the technical problems of insufficient adaptability of learning paths in existing systems, inability to dynamically adjust, and privacy risks associated with multi-source data sharing. It improves the matching degree between learning paths and individual student needs, ensures educational data security, complies with data compliance requirements in the education field, and enhances the applicability and reliability of the system in actual classroom scenarios.

[0057] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0058] 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. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0059] Figure 1 This is a schematic diagram of the operation process of the present invention;

[0060] Figure 2 This is a schematic diagram of the system composition and data flow of the present invention;

[0061] Figure 3 This is a schematic diagram of the causal analysis and path decision-making logic of the present invention. Detailed Implementation

[0062] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0063] Example 1

[0064] like Figure 2 and Figure 3 As shown, this embodiment aims to elaborate on the specific implementation of an AI-based dynamic generation system for personalized classroom learning paths. It addresses the technical problems in existing personalized learning systems, such as biased knowledge state assessment due to insufficient consideration of confounding factors like environmental and physiological factors, and inaccurate identification of the causal roots of knowledge gaps. This embodiment collects comprehensive learning process data through a multimodal perception module, integrates prior educational knowledge with data-driven reasoning to construct a causal model using a neural symbolic causal graph construction module, utilizes counterfactual reasoning and robustness assessment modules to remove the influence of confounding factors and locate the root causes of knowledge gaps, and combines this with a dynamic path generation module to generate learning paths tailored to each individual. Simultaneously, a federated learning architecture ensures student data privacy. The entire implementation process revolves around the functional implementation of each module, unit collaboration, algorithm principles, and formula application, ensuring the system can be stably and accurately applied to personalized classroom teaching scenarios, providing technical support for improving teaching effectiveness and achieving individualized instruction.

[0065] System Overall Architecture Description: The AI-based classroom personalized learning path dynamic generation system described in this embodiment consists of four core functional modules: a multimodal perception module, a neural symbol causal graph construction module, a counterfactual reasoning and robustness assessment module, and a dynamic path generation module. It also employs a federated learning architecture to complete the training and updating of the model. A closed-loop flow of data and instructions is formed between the modules: the multimodal perception module, as the data input, collects data related to students' physiological state, learning environment, and learning behavior, and transmits the raw data to the neural symbol causal graph construction module; the neural symbol causal graph construction module extracts features from the raw data and constructs causal models, generating a neural symbol causal graph containing knowledge points, physiological state, environment, and learning performance nodes, providing a model foundation for subsequent reasoning; the counterfactual reasoning and robustness assessment module performs interventions and calculates counterfactual results based on the above causal graph, outputs students' robustness mastery index and the causal root causes of knowledge gaps, and feeds the results back to the dynamic path generation module; the dynamic path generation module generates and adjusts personalized learning paths based on the robustness mastery index and the root causes of knowledge gaps; the federated learning architecture runs through the entire model lifecycle, enabling multi-client collaborative optimization of model performance while ensuring data privacy, and ensuring the system's adaptability and accuracy in different classroom scenarios.

[0066] Detailed Implementation of the Multimodal Perception Module: The core function of the multimodal perception module is to collect multi-dimensional raw data related to student learning in a classroom setting, providing data support for feature extraction and model building in subsequent modules. Its specific implementation includes the collaborative work of the following five units:

[0067] Implementation of Non-Invasive Visual Sensors: Non-invasive visual sensors utilize high-definition cameras deployed at the front and sides of the classroom to ensure coverage of all students' learning areas. The sensors operate without physical contact with students, avoiding interference with normal learning. The sensors capture real-time video streams of students' faces. This video stream data is first transmitted to a local data processing unit for feature extraction using a lightweight facial feature point detection model. This model, based on the MobileNet architecture, features low computational cost and high detection accuracy, and can locate 68 key facial feature points (such as the corners of the eyes, mouth, and nose tip) in real time. Subsequently, an eye state classification model calculates the student's head posture offset angle and eye gaze direction based on the positional changes of these feature points. The head posture offset angle is calculated through the coordinate changes of the feature points in three-dimensional space, reflecting whether the student's head is facing the learning area (e.g., whether they are looking down or turning their head). The eye gaze direction is determined by the relative positional relationship between the pupil center and eye feature points (such as the iris edge), judging whether the student is focused on the learning content (e.g., whether they are looking at the blackboard, textbook, or deviating from the learning area). The head posture offset angle and eye gaze direction data obtained from the above calculations are used as one of the raw data of students' physiological state (related to attention) and transmitted to the neural symbol causal graph construction module.

[0068] Audio sensor implementation: The audio sensor employs an omnidirectional microphone array, deployed on the classroom ceiling and desks to achieve comprehensive acquisition of classroom ambient sound and student speech. The acquired audio data is first preprocessed (including noise reduction and echo cancellation), then input into a speech activity detection model. This model analyzes the energy, spectral characteristics, and zero-crossing rate of the audio signal to distinguish between Silence and Speech states. The Speech state is further subdivided into teacher speech and student speech (based on differences in acoustic characteristics such as fundamental frequency and speech rate; teacher speech typically has a lower fundamental frequency and a relatively stable speech rate than student speech). Simultaneously, an emotion recognition model assesses the emotional polarity of student speech. This model, based on an LSTM architecture, inputs Mel-frequency cepstral coefficients (MFCC) features of the speech and outputs three emotional polarities: "positive," "neutral," and "negative" (e.g., a student's high-pitched voice corresponds to "positive," while a low-pitched voice corresponds to "negative"). Environmental sound data (such as noise intensity-related features) and student speech emotional polarity data are transmitted to subsequent modules as components of environmental data and physiological state data, respectively.

[0069] Implementation of Environmental Sensors: The environmental sensors employ integrated temperature, humidity, and noise detection modules, deployed in different areas of the classroom (e.g., near the podium, in the middle of the classroom, and in corners) to ensure the spatial representativeness of the collected environmental data. This module collects real-time data on classroom temperature, humidity, and ambient noise levels in decibels: temperature ranges from 15℃ to 35℃ (consistent with typical classroom temperatures), humidity ranges from 30% to 70% (avoiding the impact of extreme humidity on learning), and noise levels range from 30dB to 100dB (covering the range from normal classroom interaction to noise interference). The collected data is converted to digital signals via analog-to-digital conversion, and after mean filtering (eliminating the impact of instantaneous fluctuations), it is transmitted as environmental status data to the neural symbolic causal graph construction module.

[0070] Implementation of the physiological signal sensor: The physiological signal sensor is integrated into a head-mounted device with a lightweight design to ensure student comfort and without hindering head movement. The core of the sensor is a low-power EEG signal acquisition unit, which collects raw EEG signals from students via contact electrodes. The acquisition frequency meets the requirements of physiological signal analysis (ensuring the signal contains...). Wave, Wave, (Frequency bands related to cognitive states, such as wave patterns). The raw EEG signal is first amplified and filtered (to remove power frequency interference and electromyographic interference), and then feature analysis is performed using an embedded algorithm—this algorithm is based on power spectral density analysis to calculate... Wave (8Hz-13Hz, related to a relaxed state) The energy percentage of wavelengths (14Hz-30Hz, related to focus) is used to derive a focus index; simultaneously, through analysis... The energy changes of the brainwave (4Hz-7Hz, associated with excessive cognitive load) are combined with the characteristics of the student's EEG signals during answering questions to calculate the cognitive load index. The focus index and cognitive load index, as core physiological state data, are transmitted to subsequent modules.

[0071] Implementation of the Learning Behavior Log Collection Unit: The learning behavior log collection unit establishes a data connection with the learning platform used in the classroom (such as an online quiz system or interactive teaching platform) through an application programming interface (API) to acquire students' operational data on the learning platform in real time. Specific data collected includes the correctness of answers (reflecting students' immediate grasp of knowledge points), the time spent answering questions (reflecting students' speed of understanding and perception of difficulty of knowledge points), and the number of interactions (such as the number of times hands are raised to answer questions and the number of times students participate in group discussions, reflecting students' learning initiative). After data cleaning (removing invalid data, such as duplicate answer records caused by erroneous operations), the collected data is transmitted as learning behavior data to the neural symbol causal graph construction module, providing a basis for feature extraction of learning performance nodes.

[0072] Detailed Implementation of the Neural Symbolic Causal Graph Construction Module: The core function of the neural symbolic causal graph construction module is to transform the raw data collected by the multimodal perception module into a structured causal graph model, integrating prior educational knowledge (symbolic rules) with data-driven neural network reasoning to ensure that the causal graph is both interpretable and adaptable to the learning data characteristics of different students. Its specific implementation includes the collaborative work of the following three units:

[0073] Implementation of the neural network feature extraction unit: The neural network feature extraction unit receives all the raw data transmitted by the multimodal perception module, and performs feature extraction using an appropriate neural network architecture for different types of data, ultimately generating four types of concept vectors (knowledge point mastery vector, physiological state vector, environmental state vector, and learning performance vector):

[0074] Extraction of the knowledge point mastery vector: Taking the answer correctness and answer time data in the learning behavior log as input, a multilayer perceptron (MLP) architecture is adopted. The input layer is the feature of the answer data (such as the answer accuracy rate and average time of a certain knowledge point), the hidden layer is transformed nonlinearly through the ReLU activation function, and the output layer is a vector with the same dimension as the number of knowledge points in the course. The value of each element in the vector is in the range of 0-1, representing the student's mastery of the corresponding knowledge point (the closer the value is to 1, the higher the mastery).

[0075] Extraction of physiological state vectors: The physiological state vectors are generated by fusing head posture offset angle and eye gaze direction data from non-invasive visual sensors, student voice emotion polarity data from audio sensors, and attention index and cognitive load index from physiological signal sensors. A fusion architecture of convolutional neural network (CNN) and recurrent neural network (RNN) is adopted. CNN is used to extract spatial features of head posture and eye gaze direction, and RNN is used to extract temporal features of emotion polarity and attention index changes over time. The features output by both are concatenated into vectors and then processed by BatchNormalization (to avoid the influence of feature distribution differences).

[0076] Environmental state vector extraction: Taking temperature, humidity, and ambient noise decibel values ​​collected by environmental sensors as input, a simple fully connected network is used to standardize the original data (mapped to the 0-1 interval), and then generate an environmental state vector through linear transformation and nonlinear activation. The vector dimension is consistent with the number of environmental parameters.

[0077] Learning performance vector extraction: Taking partial features of knowledge point mastery vector, physiological state vector, and environmental state vector (such as focus index and environmental noise decibel value) and the number of interactions in the learning behavior log as input, a simplified architecture of graph convolutional network (GCN) is adopted to capture the correlation between each input feature and output the learning performance vector. Each element in the vector represents the student's performance score in different learning scenarios (such as classroom practice performance and group discussion performance).

[0078] After the above four types of concept vectors are generated, they are transmitted to the causal graph synthesis unit as the feature basis of the causal graph nodes.

[0079] Implementation of the Symbolic Rule Injection Unit: The symbolic rule injection unit stores prior knowledge symbolic rules in the field of education. These rules are derived from the consensus of experts in the field of education and the summary of teaching practices, ensuring the scientific and practical nature of the rules. The rule content is mainly divided into two categories:

[0080] Rules governing the prerequisite and successor relationships between knowledge points: Define the dependency relationships between different knowledge points, such as "solving quadratic equations in one variable" being a prerequisite for "properties of quadratic function graphs", and "the triangle angle sum theorem" being a prerequisite for "the polygon angle sum theorem". The expression of this type of rule is "if a student has not mastered knowledge point A, then he / she cannot effectively master knowledge point B", which corresponds to the directed edge constraint in the causal graph from node A (prerequisite knowledge point) to node B (successor knowledge point).

[0081] Rules governing the influence of physiological and environmental factors on learning performance: Define the relationship between physiological and environmental states on learning performance, such as "when the environmental noise level is greater than 60 dB, students' learning performance will decrease significantly" or "when the concentration index is less than 0.4, students' learning performance will decrease". The form of such rules is "if physiological / environmental parameter X satisfies condition Y, then learning performance Z will show trend W", which corresponds to the directed edge constraint from the physiological / environment node to the learning performance node in the causal graph.

[0082] The symbolic rule injection unit transforms the aforementioned symbolic rules into graph structure constraints: for the "pre-requisite, subsequent-successor" rule, an existence constraint of directed edges is added between the corresponding knowledge point nodes (i.e., an edge from the pre-requisite node to the subsequent node must exist); for physiological / environmental constraints, a weight constraint of the edges is added between the corresponding physiological / environmental nodes and the learning performance nodes (i.e., the weight of the edges must conform to the influence trend defined by the rule). These graph structure constraints are transmitted to the causal graph synthesis unit as the rule basis for constructing the causal graph.

[0083] Implementation of the Causal Graph Synthesis Unit: The causal graph synthesis unit receives the concept vector from the neural network feature extraction unit and the graph structure constraints from the symbolic rule injection unit. Combined with a causal discovery algorithm (using the PC algorithm, which determines the direction of causal relationships between variables by testing the conditional independence between variables, possessing high causal discovery accuracy), it constructs a neural symbolic causal graph model. The core is to achieve the fusion of neural reasoning and symbolic logic, specifically accomplished through the message passing process of the graph neural network. Its key formulas and principles are as follows:

[0084] In the message passing process of a graph neural network, nodes to neighboring nodes The message being delivered It must simultaneously satisfy the logical constraints of parameter-driven neural networks and symbol rules. The message calculation formula is:

[0085]

[0086] in: This is the weight matrix of the neural network, which is trained and optimized through a subsequent federated learning process to capture the numerical correlation between node features;

[0087] , They are nodes , The feature vector (i.e., the feature of the corresponding node in the concept vector);

[0088] This represents the vector concatenation operation, which... and The vectors are concatenated into a high-dimensional vector to simultaneously consider the feature information of both nodes.

[0089] A function based on predefined symbol rules, used to determine nodes. and Does a causal or logical relationship exist between them as defined by the rules? If so (e.g.) yes Prerequisite knowledge points, or (Environmental noise nodes) (If the learning performance node has the influence of the rule definition), then output a positive value (ranging from 0.5 to 1.0, with a larger value indicating a stronger rule constraint); if it does not exist, output a negative value (ranging from -1.0 to -0.5).

[0090] The sigmoid function is used to... The output is mapped to weight coefficients in the 0-1 range, realizing the weight adjustment of message passing by the symbol rule—when When it is a positive value, An output close to 1 strengthens message passing; a negative value... Output close to 0 weakens message passing strength;

[0091] The ReLU function is used as a non-linear activation function to perform a non-linear transformation on the result of message passing, thereby enhancing the model's ability to express complex causal relationships and avoiding the gradient vanishing problem.

[0092] Through the aforementioned message passing process, the causal graph synthesis unit maps concept vectors to nodes in the causal graph (each node corresponds to a type of concept vector, such as a knowledge point node corresponding to a knowledge point mastery vector, and a physiological state node corresponding to a physiological state vector). The edges between nodes are determined by the message passing strength (if the message passing strength exceeds a set threshold, a directed edge is formed, and the weight of the edge is the normalized value of the message passing strength), ultimately constructing a neural symbolic causal graph model. This model ensures the interpretability of causal relationships through symbolic rules (such as the inevitable influence of prior knowledge points on subsequent knowledge points) and captures personalized correlations in the data through neural network parameters (such as the differences in sensitivity of different students to the same environmental interference), providing a reliable model foundation for subsequent counterfactual reasoning.

[0093] Detailed Implementation of the Counterfactual Reasoning and Robustness Assessment Module: The core function of the Counterfactual Reasoning and Robustness Assessment module is based on the neural symbolic causal graph model. Through intervention operations and counterfactual calculations, it removes the interference of confounding factors (such as environmental noise and momentary emotional fluctuations) on students' learning performance, accurately assesses the robustness of students' true knowledge mastery, and locates the causal root causes of knowledge gaps. Its specific implementation includes the collaborative work of the following five units:

[0094] Implementation of the Intervention Execution Unit: The role of the intervention execution unit is to perform intervention operations on specified confounding factor nodes in the neural symbolic causal graph model, simulating the model state under ideal or interference-free environments, and providing a basis for counterfactual calculations. First, the system determines the types of confounding factor nodes through preliminary data statistics and judgment by educational experts, mainly including environmental confounding factor nodes (such as environmental noise nodes, temperature nodes) and physiological confounding factor nodes (such as transient emotion nodes, short-term attention nodes). Then, specific intervention values ​​are set for different types of confounding factor nodes: for environmental noise nodes, the intervention value is set to the ideal classroom noise level (ensuring that the noise decibel value does not affect learning); for transient emotion nodes, the intervention value is set to the feature value corresponding to neutral emotion. The intervention operation is implemented by forcibly setting the feature value of the confounding factor node to the intervention value, while fixing the edge weights between this node and other nodes (to avoid the intervention operation affecting the causal relationship related to non-confounding factors), generating a modified neural symbolic causal graph model.

[0095] Implementation of the Counterfactual Calculation Unit: Based on the modified causal graph model generated by the intervention execution unit, the counterfactual calculation unit performs forward inference calculations to obtain the predicted learning performance under counterfactual conditions. Forward inference employs a node state update algorithm of a graph neural network: taking the feature values ​​of all nodes in the modified model (excluding confusion factor nodes, knowledge point nodes, and stable physiological state nodes) as input, the feature states of each node are updated through a message passing process (consistent with the message passing formula in the causal graph construction phase), ultimately focusing on the learning performance node—using the feature vector of the learning performance node, combined with a linear regression model (this model is trained on historical learning performance data, with the input being the feature vector of the learning performance node and the output being the predicted learning performance score), the predicted learning performance under counterfactual conditions is calculated. This predicted value reflects the learning performance that students should exhibit based on their true knowledge mastery level when there is no confusion factor interference, providing a benchmark for subsequent effect quantification.

[0096] Implementation of the Effect Quantification Unit: The effect quantification unit calculates the average causal effect of the confounding factor by comparing the actual observed learning performance values ​​(i.e., the learning performance values ​​calculated based on the original causal graph model without intervention) with the predicted learning performance values ​​under counterfactual conditions. The core of this unit employs a dual machine learning model to achieve unbiased estimation. The specific steps are as follows:

[0097] The first step is to fit the dependency of the confounding factor on the covariates: Let... The confusion factor to be evaluated (such as the characteristic value of the environmental noise node). To remove All other covariates (including the knowledge mastery vector, stable physiological state vector, and feature values ​​of other non-target confusion factor nodes) are analyzed using a random forest model. Fitting right The dependency relationship, to obtain Predicted value And calculate the residuals The random forest model was chosen because of its ability to handle high-dimensional covariates, capture nonlinear relationships, and accurately fit the data. and Complex dependencies between them.

[0098] The second step is to fit the dependence of performance on covariates: Let... The gradient boosting tree model is used to represent the actual observed learning performance values. Fitting right The dependency relationship, to obtain Predicted value And calculate the residuals The gradient boosting tree model was chosen because of its high prediction accuracy in regression tasks, effectively capturing... and The relationship between them.

[0099] The third step is to calculate the average causal effect: through the residuals. For residuals Perform linear regression and obtain the regression coefficients This is the confusion factor. Regarding learning performance The average causal effect estimate is calculated using the following formula:

[0100]

[0101] in, Residual and The covariance reflects the degree of linear correlation between the two. Residual The variance reflects The degree of dispersion. The core logic of this formula is to eliminate covariates through residuals. right and The combined influence of these factors ensures Only reflects right The net causal effect is obtained, achieving unbiased estimation.

[0102] Implementation of the Robust Mastery Index Calculation Unit: The Robust Mastery Index Calculation Unit is used to assess the stability of students' knowledge mastery under various simulated intervention environments. Its core is the calculation of knowledge points. Robustness mastery index The specific implementation is as follows:

[0103] First, determine the total number of simulated environmental intervention scenarios. Based on common environmental disturbances in the classroom (such as different noise levels, temperature fluctuations, and changes in lighting), set up... Ten typical intervention scenarios were identified to ensure coverage of the main environmental disturbances students might face; for each intervention scenario... , implement intervention operations ( A set of environment variables. For the first The combinations of values ​​for environmental variables in various intervention scenarios, such as The noise level is 40dB and the temperature is 25℃. (e.g., "noise level 50dB + temperature 26℃").

[0104] Secondly, calculate the knowledge points for each intervention scenario. Predicted performance values: for each Knowledge points are obtained through counterfactual calculation units. Corresponding learning performance prediction value This value reflects the first Under various environmental interventions, students' understanding of knowledge points The performance of mastery.

[0105] Finally, the robustness mastery index is calculated. The calculation formula is:

[0106]

[0107] in, The threshold for satisfactory learning performance is set based on the assessment standards of the course syllabus (e.g., 80 points on a percentage scale; ensuring that students reach this threshold indicates that their mastery of the knowledge points meets the teaching requirements). For indicator functions, when the condition inside the parentheses... When true, Returns 1 otherwise returns 0. The value ranges from 0 to 1. The closer the value is to 1, the better the student's understanding of the knowledge points under different environmental interferences. The more stable the student's performance in mastering the knowledge points, the higher their true level of mastery; the closer the value is to 0, the better the student's understanding of the knowledge points. The mastery of knowledge is significantly affected by environmental factors, or the actual level of mastery is insufficient.

[0108] Implementation of the Root Cause Tracing Unit: The Root Cause Tracing Unit is used to locate the root cause knowledge points leading to knowledge gaps. Its core is achieved through iterative generation of counterfactual questions and reasoning. The specific implementation process is as follows:

[0109] Identify target knowledge points: First, identify the target knowledge points that students have not yet mastered. (Right now If the value is below a set threshold and the effect quantification unit determines that the influence of the confusion factor is not significant, it is identified as a true knowledge gap, such as "the application of quadratic functions".

[0110] Generate counterfactual questions: targeting specific knowledge points The system automatically generates a counterfactual question: "What if students have fully mastered the prerequisite knowledge points?" ( For determination according to the symbol rules The direct prerequisite knowledge points, such as "the graph properties of quadratic functions"), whether they can be mastered. ?

[0111] Implementation of intervention and reasoning: Addressing the prerequisite knowledge nodes in the neural symbolic causal graph model Implement intervention operations, The mastery feature value is set to the maximum value (i.e., 1, representing complete mastery), while other nodes (except for) are kept at their maximum value. The eigenvalues ​​of the external (external) remain unchanged; based on the modified model, the target knowledge points are re-inferred through the counterfactual calculation unit. The degree of mastery (in terms of) (Using changes as a basis for judgment).

[0112] Determining the causal root cause: Several prognoses of Significant improvement (the improvement exceeds the set threshold, such as from 0.3 to 0.8), and exceeds the qualified threshold. Then determine To cause The causal roots of the knowledge gap; several prognoses of If there is no significant improvement (the improvement is below the set threshold), it indicates that... This is not the root cause; further investigation is needed. prerequisite knowledge points (e.g., "solving quadratic equations in one variable").

[0113] Iterative tracing: Repeat steps 2-4 above, to... As a new target knowledge point Its prerequisite knowledge points As a new Generate new counterfactual questions and intervene in the reasoning until a certain prior knowledge point is found. —When After intervention (set to full mastery), all subsequent knowledge points (including the initial ones) will be covered. )of All significantly improved and exceeded At this point, a judgment is made. To lead to the initial The most fundamental and weakest point of the knowledge gap. This process ensures that the root knowledge point traced is the true "source" of the knowledge gap, providing a precise direction for repair in the subsequent dynamic path generation.

[0114] Implementation of the dynamic path generation module: The core function of the dynamic path generation module is to use the robustness mastery index output by the counterfactual reasoning and robustness assessment module. ) and the causal roots of knowledge gaps ( The decision-making logic and implementation process for dynamically generating and adjusting personalized learning path sequences are as follows:

[0115] Step 1: Determine the impact of the confounding factor: The dynamic path generation module first receives the average causal effect of the confounding factor output by the effect quantification unit. ,judge Is it greater than a first set threshold (this threshold is set based on educational practice data statistics, such as 0.3, to ensure...) (A value greater than this threshold indicates that the confusion factor has a significant impact on learning performance).

[0116] like The first threshold is set: if the student's current learning status assessment is significantly affected by confusion factors, the generated personalized learning path will prioritize the recommendation of auxiliary learning content to reduce the impact of these confusion factors, such as "noise isolation method guidance" and "concentration training videos" to address environmental noise interference, and "emotion regulation skills micro-lessons" to address emotional fluctuations. The learning duration and order of the auxiliary content are dynamically adjusted based on the student's physiological state data (such as concentration index and emotional polarity) to ensure that the interference of confusion factors is reduced in the short term, creating favorable conditions for subsequent knowledge learning.

[0117] Step 2: Identify the actual knowledge gap: If The initial threshold indicates that the confusion factor has no significant impact; at this point, the module further assesses the knowledge points. of Is it below the second set threshold (this threshold is set based on teaching assessment requirements, such as 0.6, to ensure...) Below this threshold indicates that the student's understanding of the knowledge point is insufficient. There is a real knowledge gap).

[0118] like The second threshold setting: The module receives the root cause knowledge points located by the causal root cause tracing unit. , generate from Starting point, gradually leading to the target knowledge points The backtracking learning path sequence. The generation of the path sequence follows the principle of "first fixing the root weaknesses, then gradually advancing," and the specific order is determined according to the precedence relationship of the symbol rules, such as... (Solving quadratic equations in one variable) → Direct follow-up knowledge points ("Graphical properties of quadratic functions") → Target knowledge point ("Applications of quadratic functions"); The learning content (such as micro-lecture videos, exercises, and interactive experiments) for each knowledge point is selected personalized based on students' learning behavior data (such as time spent answering questions and number of interactions). For example, for knowledge points that take a long time to answer questions, more detailed explanation videos are recommended; for students with fewer interactions, group collaborative learning tasks are recommended.

[0119] Dynamic path adjustment: After the personalized learning path is generated, the module receives updated data from the multimodal perception module and the counterfactual reasoning module in real time (such as students' answer data after learning new content, etc.). (Changes), if it is found that students have difficulty with a certain knowledge point If the level rises above the second set threshold, the subsequent reinforcement content for that knowledge point will be automatically skipped, and the student will move on to the next knowledge point. If it is found that a student has difficulty with a certain knowledge point... If the learning progress does not improve or even declines, supplementary learning content for that knowledge point (such as supplementary practice questions and one-on-one Q&A resources) will be added to ensure that the learning path always adapts to the student's real-time learning status.

[0120] Implementation of the Federated Learning Architecture: The federated learning architecture is used to train and update the system model. Its core principle is to collaboratively optimize model performance using client data from multiple schools while ensuring student data privacy. Its implementation involves the interaction between the central server and edge clients.

[0121] Edge Client Implementation: Edge clients are deployed on local servers in each school, storing anonymized multimodal data of the school's students (anonymization includes removing student names, student IDs, and other identifying information; anonymizing and encrypting sensitive data such as EEG signals and facial videos to ensure the data cannot be associated with any specific individual). The core task of the edge client is to utilize local data to process neural network components in the neural symbolic causal graph model (such as CNNs and RNNs for feature extraction, and weight matrices for message passing). Local training is performed: The training process uses the gradient descent algorithm, with the mean squared error (MSE) between the predicted and actual values ​​as the loss function. The local model parameters are optimized through multiple iterations to generate gradient information of the model parameters. After the gradient information is generated, it is encrypted using an encryption algorithm (such as isomorphic encryption) to ensure the security of the gradient transmission process.

[0122] Central Server Implementation: The central server is deployed in the cloud and has data aggregation and model update capabilities. The central server periodically (e.g., weekly, based on the school's teaching cycle) receives encrypted gradient information from each edge client and aggregates the gradient information using a secure averaging algorithm (FedAvg algorithm). This algorithm first assigns weights to the gradient information from each client (the weights are proportional to the amount of local data on the client, ensuring that clients with larger data volumes contribute more to the global model). Then, it calculates the weighted average gradient, which serves as the direction for updating the global model parameters. Based on the global average gradient, the central server updates the global parameters of the neural symbolic causal graph model (mainly the weight matrices of the neural network components). (Parameters of the feature extraction model, etc.), keep the rules of the symbol rule injection unit unchanged (to ensure the stability of prior educational knowledge).

[0123] Model Distribution and Iteration: After the central server completes the global model parameter update, it encrypts and distributes the updated global model parameters to each edge client. Upon receiving the global model parameters, the edge client merges them with its local model parameters (using model parameter replacement to ensure consistency between the local and global models) and continues the next round of local training based on the new local data. This process iterates continuously, achieving a closed loop of "local training - gradient upload - global aggregation - model distribution." This not only leverages the massive amounts of data from multiple schools to optimize the model's adaptability to different student learning characteristics but also ensures that the original student data remains locally on the edge client, preventing cross-institutional transmission, effectively protecting data privacy, and complying with relevant regulations on educational data security.

[0124] In summary, this embodiment details the complete implementation process of an AI-based personalized classroom learning path dynamic generation system. From multi-dimensional data collection by the multimodal perception module, to model construction by the neural symbol causal graph construction module, and then to interference removal, robustness assessment, and root cause tracing by the counterfactual reasoning and robustness assessment module, the system ultimately generates a learning path tailored to each individual through the dynamic path generation module. Data privacy and model performance optimization are ensured by relying on a federated learning architecture. Throughout the implementation process, each module, through clear unit collaboration, scientific algorithm selection (such as dual machine learning and federated averaging algorithms), and rigorous formula application (such as message passing formulas and robustness mastery index formulas), addresses the core shortcomings of existing technologies, such as interference from confounding factors in knowledge assessment and inaccurate identification of knowledge gap root causes. The system implementation process is operable and repeatable, and the generated personalized learning paths balance accuracy, robustness, and interpretability. The application of the federated learning architecture further ensures the system's compliance and privacy security in the education field.

[0125] Example 2

[0126] like Figure 1 As shown in Example 1, this example elaborates on the specific steps of a dynamic generation system for personalized classroom learning paths based on artificial intelligence during operation. The specific steps are as follows:

[0127] 1. Multimodal data acquisition:

[0128] The system collects multimodal data in real time, including students' facial expressions, eye gaze direction, voice emotion, environmental temperature, humidity and noise, EEG signals, and answering behavior, through various sensors deployed in the classroom (including visual, audio, environmental and physiological signal sensors) and the learning platform interface.

[0129] 2. Data preprocessing and feature extraction:

[0130] The collected raw data were preprocessed by cleaning, denoising, and standardization, and four types of feature vectors were extracted: knowledge mastery, physiological state, environmental state, and learning performance.

[0131] 3. Construction of neural symbolic causal graph:

[0132] By combining prior educational knowledge (such as the pre- and post-learning relationships between knowledge points and the rules governing the influence of the environment on learning) with data-driven causal discovery algorithms, a neural symbolic causal graph model is constructed and updated. This model uses knowledge points, physiological states, environmental states, and learning performance as nodes, and causal relationships as edges.

[0133] 4. Counterfactual reasoning and robustness assessment:

[0134] Intervention simulations were conducted on the confusing factor nodes (such as environmental noise and emotional fluctuations) in the causal graph to calculate the counterfactual learning performance under ideal conditions. In turn, the robust mastery index (RMI) of each knowledge point was evaluated, and the root cause of the knowledge gap was traced to the knowledge points.

[0135] 5. Personalized learning path generation:

[0136] Based on the robustness mastery index and causal root cause analysis results, a learning path is dynamically generated: if the confusion factor has a significant impact, supplementary content to alleviate the interference is recommended first; if there is a real knowledge gap, a backtracking learning sequence is generated that gradually advances from the root knowledge points.

[0137] 6. Path execution and real-time adjustment:

[0138] The generated learning path is pushed to students, and feedback data (such as answer results and behavioral changes) is continuously collected during the learning process. The content and order of the path are dynamically adjusted to ensure that the path always matches the students' real-time learning status.

[0139] 7. Federated learning model update:

[0140] Each edge client trains the model using local data, uploads the encrypted gradients to the central server, the server aggregates the global gradients and updates the model parameters, and then distributes the new model to each client, thus achieving iterative optimization of the model and protection of data privacy.

[0141] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A dynamic generation system for personalized classroom learning paths based on artificial intelligence, characterized in that, The system consists of: a multimodal perception module, a neural symbol causal graph construction module, a counterfactual reasoning and robustness assessment module, and a dynamic path generation module; The multimodal perception module is used to collect students' physiological state data, environmental data, and learning behavior data. The neural symbol causal graph construction module is used to construct and maintain a causal graph model containing knowledge point nodes, physiological state nodes, environmental nodes, and learning performance nodes, wherein the causal relationships between nodes are constrained by predefined symbol rules and causal discovery algorithms. The counterfactual reasoning and robustness assessment module is used to perform intervention operations based on the causal graph model, calculate counterfactual results, eliminate the influence of confounding factors, and output the student's robustness mastery index. The dynamic path generation module is used to dynamically generate and adjust a personalized learning path sequence based on the robustness mastery index and the causal root causes of knowledge gaps located by counterfactual reasoning.

2. The AI-based dynamic generation system for personalized classroom learning paths according to claim 1, characterized in that, The neural symbol causal graph construction module specifically includes: The neural network feature extraction unit is used to extract features from the raw data collected by the multimodal perception module to obtain corresponding concept vectors. The concept vectors include knowledge point mastery vectors, physiological state vectors, environmental state vectors, and learning performance vectors. The symbol rule injection unit stores symbol rules for prior educational knowledge. These symbol rules are used to define the logical relationships between knowledge points and the constraints of physiological and environmental factors on learning performance. The symbol rule injection unit converts the symbol rules into graph structure constraints. The causal graph synthesis unit receives the concept vector and the graph structure constraints, and constructs the neural symbolic causal graph model by combining it with the causal discovery algorithm. Nodes in the causal graph model correspond to the concept vectors, and edges represent the causal relationships between nodes. The causal graph synthesis unit achieves the fusion of neural reasoning and symbolic logic through the following computational process: In the message passing process of a graph neural network, nodes to neighboring nodes The message being delivered Not only by the neural network parameter weights and node features , The decision is made, and is also constrained by the symbolic rule function. The output of , whose message calculation formula is: in, It is a non-linear activation function. This represents a vector concatenation operation. The sigmoid function is used to map the output of a rule function to weight coefficients. It is a function based on predefined symbol rules, if the node and If at least one of the causal relationship and logical relationship defined by the rules exists, output a positive value; otherwise, output a negative value.

3. The AI-based dynamic generation system for personalized classroom learning paths according to claim 1, characterized in that, The counterfactual reasoning and robustness assessment module specifically includes: An intervention execution unit is used to perform intervention operations on a specified confusion factor node in the neural symbol causal graph model, and set the value of the confusion factor node to a specific intervention value, thereby simulating the state in an ideal, interference-free target environment. The counterfactual calculation unit is used to perform forward reasoning based on the modified graphical model after the intervention is implemented, and to calculate the predicted value of learning performance under counterfactual conditions. The effect quantification unit is used to compare the actual observed learning performance value with the predicted learning performance value under the counterfactual condition, and to calculate the average causal effect of the confusion factor. The robustness mastery index calculation unit is used to assess the stability of students' knowledge mastery under various simulated intervention environments. The robustness mastery index calculation unit calculates knowledge points using the following formula. Robustness mastery index : in, This represents the total number of simulated environmental intervention scenarios. Indicates the environment variable Implementation of the Intervention, setting its value to , This indicates the knowledge points under the intervention. The predicted performance value, To represent the acceptable threshold, This is an indicator function that returns 1 if the condition within the parentheses is true, and 0 otherwise.

4. The AI-based dynamic generation system for personalized classroom learning paths according to claim 1, characterized in that, The counterfactual reasoning and robustness assessment module also includes a causal root cause tracing unit. This unit is used to locate the root cause knowledge points leading to the knowledge gap. The causal root cause tracing unit achieves this by iteratively generating counterfactual questions and performing reasoning. The process is as follows: For the target knowledge points that are not yet mastered The system automatically generates a hypothetical question, the core of which is: assuming the student has fully mastered the prerequisite knowledge points. To determine whether they can master the target knowledge points. ; Subsequently, the knowledge point nodes in the neural symbolic causal graph model were analyzed. Implement intervention, setting its level of mastery to the maximum value; Based on the intervention, the target knowledge points are re-reasoned. The degree of mastery; If after intervention If the mastery reasoning result significantly improves and exceeds a predetermined threshold, then it is determined that... for The causal roots of the knowledge gap; This process proceeds iteratively in reverse along the cause-effect graph until the most fundamental weakness in prior knowledge is found.

5. The AI-based dynamic generation system for personalized classroom learning paths according to claim 1, characterized in that, The decision logic of the dynamic path generation module is based on the robustness mastery index and the output of the causal root cause tracing unit. The dynamic path generation module first determines whether the low robustness mastery index is mainly caused by the confusion factor effect; If the average causal effect of the confusion factor is greater than the first set threshold, it is determined that the current knowledge state assessment is significantly disturbed. In this case, the generated path will prioritize the recommendation of auxiliary learning content to reduce the impact of the confusion factor. If the average causal effect of the confusion factor is less than the first set threshold and the robustness mastery index is lower than the second set threshold, then it is determined that there is a real lack of knowledge mastery. At this time, the dynamic path generation module receives the root cause knowledge point located by the causal root cause tracing unit and generates a backtracking learning path sequence that starts from the root cause knowledge point and gradually leads to the target knowledge point.

6. The AI-based dynamic generation system for personalized classroom learning paths according to claim 3, characterized in that, The effect quantification unit employs a dual machine learning model to estimate the average causal effect of the confounding factor in order to achieve unbiased estimation. The estimation process of the dual machine learning model consists of the following steps: The first step is to use a machine learning model. Fitting Confusion Factor For covariates Based on the dependency relationship, the predicted value is obtained. And calculate the residuals ; The second step is to use another machine learning model. Fitting outcome variables For covariates Based on the dependency relationship, the predicted value is obtained. And calculate the residuals ; The third step is to use the residuals. For residuals Perform linear regression and obtain the regression coefficients This is the confusion factor. Regarding the results The estimated average causal effect is calculated using the following formula: in, Represents the confounding factor to be evaluated. Represents learning performance. Representative except All other covariates, Describing covariance, Indicates variance.

7. The AI-based dynamic generation system for personalized classroom learning paths according to claim 1, characterized in that, The multimodal sensing module includes: A non-invasive vision sensor is used to acquire video streams of students' faces and calculates the head posture offset angle and eye gaze direction in real time through a lightweight facial feature point detection model and an eye state classification model. An audio sensor is used to collect classroom environment sounds and student voices, and to distinguish between teacher voices and student voices and identify the emotional polarity in student voices through a voice activity detection and emotion recognition model. Environmental sensors are used to collect data on classroom temperature, humidity, and ambient noise levels in decibels. The physiological signal sensor uses a low-power EEG signal acquisition unit integrated into a head-mounted device to collect raw EEG signals from students and calculates attention index and cognitive load index in real time through embedded algorithms. The learning behavior log collection unit is used to obtain data on students' answers, answer time, and number of interactions on the learning platform in real time through the application programming interface.

8. A dynamic generation system for personalized classroom learning paths based on artificial intelligence, as described in any one of claims 1 to 7, characterized in that, The system employs a federated learning architecture for model training and updating; The federated learning architecture comprises a central server and multiple edge clients distributed across different schools; The edge client locally stores the desensitized multimodal data of its students and uses the local data to train the neural network components in the neural symbol causal graph model to generate gradient information of the model parameters. The central server periodically aggregates encrypted gradient information from each edge client, updates the global model parameters using a secure averaging algorithm, and then distributes the updated global model parameters to each edge client.

Images