Education interaction behavior recognition system oriented to online learning environment

Abstract

The invention relates to the technical field of education, and discloses an online learning environment-oriented education interaction behavior recognition system, which is characterized in that a standardized cognitive model is constructed offline based on interaction data of a high-performance learning group; in the online stage, interaction events of individual learners are obtained in real time, high-dimensional cognitive trajectories representing cognitive states of the learners are generated, a trajectory dynamics analysis module executes double reference analysis, and internal dynamics indexes representing stability of the trajectories are obtained by calculating the maximum Lyapunov index, recognizing phase change points and the like; the external deviation measure of the difference between the trajectory and the normalized cognitive model is obtained by calculating the state deviation degree, the path efficiency and the like, and finally, the cognitive mode diagnosis module integrates the indexes and generates a structured diagnosis description according to a rule base. According to the method, through combination of internal dynamics and an external reference system, deep quantification and fine identification of the cognitive process of the learner can be realized, and an objective basis is provided for adaptive intervention.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of education, in particular to an educational interactive behavior recognition system for online learning environment. BACKGROUND

[0002] With the advancement of information technology, online learning has become an important part of the education field. How to effectively evaluate and optimize the online learning process and improve the learning effect is one of the core challenges of current education technology research. Traditional online learning behavior analysis methods often focus on statistical learning duration, answer accuracy, click frequency and other surface interaction data. Such methods can provide a macro overview of learner behavior, but have limitations in understanding the deep cognitive mechanisms of learners. For example, when a learner stays on a certain resource for a long time, the system cannot distinguish whether he is in deep thinking, effective exploration, or is encountering obstacles and getting lost. The limitations of such surface data analysis make it difficult for existing adaptive learning systems to truly provide precise and timely personalized interventions.

[0003] In existing technologies, there is a lack of fine-grained capturing ability for the dynamic evolution of learner cognitive states. Learning is not a linear accumulation of knowledge process, but a complex dynamic process full of exploration, struggle, insight and even loss. Existing technologies generally regard the cognitive state of learners as discrete or static entities, ignoring their continuity and internal dynamics in the multi-dimensional feature space. This makes it difficult for the system to identify key turning points in the cognitive process, such as phase transitions from ineffective attempts to effective understanding, or critical states from stable mastery to new challenges.

[0004] In addition, existing technologies often lack a scientific and interpretable reference standard when evaluating individual learning behavior. Even if a certain behavior pattern can be identified, it is difficult to determine whether this pattern is efficient, acceptable, or needs to be corrected. The lack of a standardized cognitive path and state model based on a large amount of high-performance learning data makes the diagnosis and evaluation of individual learning behavior subjective and uncertain. This limitation ultimately affects the ability of online learning systems to provide truly personalized and adaptive teaching support, making it difficult for them to effectively guide learners to achieve more efficient cognitive development. Therefore, a new method is needed that can dynamically capture the cognitive trajectory of learners, analyze their internal dynamics, and make accurate diagnoses by referring to standardized models. SUMMARY

[0005] To overcome the shortcomings of existing technologies, the present application provides an educational interactive behavior recognition system for online learning environment, which solves the problem that existing online learning behavior analysis technologies mostly rely on static and isolated interaction result indicators, lack in-depth characterization of the dynamic nature and internal structure of the learning process, and are difficult to effectively distinguish different types of learning states.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The first aspect of the present invention provides an educational interactive behavior recognition system for online learning environments, comprising: a cognitive manifold construction module, an event processing module, a cognitive trajectory generation module, a trajectory dynamics analysis module, and a cognitive pattern diagnosis module.

[0007] The cognitive manifold construction module is used to acquire multi-source interaction events of a predetermined high-performance learning group offline, process the events into cognitive trajectory data, and construct a normalized cognitive manifold based on the cognitive trajectory data through a manifold learning algorithm, and generate a normalized cognitive model.

[0008] In one embodiment, the module performs feature engineering on the sequence of interaction events of high-performing groups and then uses a nonlinear dimensionality reduction algorithm to embed them into a low-dimensional manifold space, which is the normalized cognitive manifold. High-density regions on the manifold are identified as attractors representing stable cognitive states, and high-probability paths connecting different attractors constitute the optimal path. The resulting normalized cognitive model includes the manifold structure, attractor positions, and optimal path information.

[0009] The event processing module is used to acquire multi-source interaction events of the current individual learner online in real time and process them into a feature vector sequence containing contextual information.

[0010] The cognitive trajectory generation module is used to generate a real-time high-dimensional cognitive trajectory representing the learner's cognitive state based on the feature vector sequence and through a time-series coding model.

[0011] In one embodiment, the temporal coding model may employ a variational autoencoder or a similar deep generative model to encode the input feature vector sequence into a high-dimensional vector in a latent space, i.e., an instantaneous cognitive state vector. A series of instantaneous cognitive state vectors arranged in chronological order The combination constitutes the real-time high-dimensional cognitive trajectory.

[0012] The core function of the trajectory dynamics analysis module is to perform a dual-reference analysis to obtain intrinsic dynamic indicators and extrinsic deviation measures.

[0013] When acquiring intrinsic dynamic indicators, this module treats the real-time high-dimensional cognitive trajectory itself as a dynamic system for analysis, without relying on external references.

[0014] To quantize the local stability of the trajectory, this module estimates the maximum Lyapunov exponent of the trajectory through phase space reconstruction. A commonly used estimation method is: ; in, It is the distance between a pair of nearest neighbors in phase space at the initial moment. It is this pair of nearest neighbors that are evolving. Distance after step, It is the time step. This represents the average over all valid point pairs. (Positive) The value indicates that the trajectory has chaotic properties; a negative value indicates that the trajectory has chaotic properties. The value indicates that the trajectory is stable.

[0015] In addition, to capture sudden changes in state, this module can use recursive quantitative analysis to identify the moment when a phase transition occurs by calculating the time series change rate of indicators such as the recursion rate.

[0016] When obtaining external deviation metrics, this module calls the standardized cognitive model as a reference benchmark. Specifically, this is used to calculate the state deviation degree. This module will use the current cognitive state vector Project onto the normalized cognitive manifold and calculate the Euclidean distance between them: ; in, yes The nearest neighbor projection point on the normalized cognitive manifold.

[0017] To evaluate path efficiency This module calculates the actual length of an observed trajectory segment. The optimal path length connecting the same origin and destination projections in the normalized cognitive model. The ratio: ; The cognitive pattern diagnosis module is used to integrate the intrinsic dynamic indicators and the extrinsic deviation measures, and generate a structured diagnostic description of the current learning behavior based on a preset pattern rule base.

[0018] In one embodiment, the module first will Phase transition point identification , , The indicators are concatenated into a diagnostic feature vector. Subsequently, by... Match the rules with the "if" and "then" rules in the pattern rule base, and output the corresponding cognitive pattern label and explanation.

[0019] A second aspect of this invention provides a method for recognizing educational interactive behaviors in online learning environments, the method comprising the following steps: Offline acquisition of multi-source interaction events of a predetermined high-performing learning group; processing of the events into cognitive trajectory data; construction of a normalized cognitive manifold based on the cognitive trajectory data using a manifold learning algorithm; and generation of a normalized cognitive model. The system acquires multi-source interaction events of the current individual learner in real time online and processes them into a sequence of feature vectors containing contextual information. Based on the feature vector sequence, a real-time high-dimensional cognitive trajectory representing the learner's cognitive state is generated through a time-series coding model. The normalized cognitive model is invoked to analyze the real-time high-dimensional cognitive trajectory in order to obtain at least one intrinsic dynamic index characterizing the stability of the trajectory and at least one extrinsic deviation measure characterizing the difference between the trajectory and the normalized cognitive manifold. By integrating the intrinsic dynamic indicators and the extrinsic deviation measures, and based on a preset pattern rule base, a structured diagnostic description of the current learning behavior is generated.

[0020] This invention provides an educational interactive behavior recognition system for online learning environments. It has the following beneficial effects: 1. This invention generates high-dimensional cognitive trajectories online in real time and analyzes them from two dimensions: the intrinsic dynamics of trajectory stability and the external deviation of the trajectory from the normalized manifold. This effectively distinguishes whether learners are in a highly focused state, an active exploratory state, or a lost or inefficient state. This dual-reference analytical framework provides a deeper insight into understanding learners' cognitive processes and avoids the limitation of generalizing all behaviors that deviate from the normative pattern.

[0021] 2. This invention integrates complex quantitative indicators into diagnostic feature vectors and matches them according to a preset pattern rule base. It can automatically generate explanatory cognitive pattern labels and detailed descriptions, such as efficient focus, lost exploration, or cognitive breakthrough. It intuitively reveals the learner's current learning status and potential problems, rather than just providing abstract numerical values. This explanatory diagnostic output can be directly used as the basis for personalized teaching recommendations, learning strategy adjustments, or early warning interventions. It avoids the burden of teachers or systems having to perform additional complex data interpretation, thereby providing learners with more timely and appropriate feedback and support.

[0022] 3. This invention constructs a standardized cognitive manifold and attractor network structure by learning from the multi-source interaction events of high-performing groups offline. It establishes a reference system composed of successful learning paths and stable cognitive states. The evaluation of individual learners' behavior is no longer blind or based on a single presupposition, but is carried out against the experience patterns of how to learn efficiently. It can discover the commonalities and differences between individual learners and university groups in their cognitive patterns, thereby effectively identifying potential learning risks or optimization opportunities, and guiding learners to evolve towards better learning strategies and cognitive states. Attached Figure Description

[0023] Figure 1 This is a system architecture diagram of the present invention; Figure 2 This is a flowchart of the method of the present invention.

[0024] Among them, 10 is the interactive event vectorization module; 20 is the normalized cognitive manifold construction module; 30 is the cognitive trajectory generation module; 40 is the trajectory dynamics analysis module; and 50 is the cognitive pattern generation module. 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] Example: Please see the appendix Figure 1 This invention provides an educational interactive behavior recognition system for online learning environments, including: an interactive event vectorization module 10, a normalized cognitive manifold construction module 20, a cognitive trajectory generation module 30, a trajectory dynamics analysis module 40, and a cognitive pattern generation module 50.

[0027] The interactive event vectorization module 10 is used to receive the original multi-source events generated by the interaction between the learner and the online learning environment, and process the original multi-source events into a feature vector sequence containing contextual semantic information and temporal information according to the preset feature engineering rules.

[0028] The standardized cognitive manifold construction module 20 runs offline before system deployment. This module is used to acquire and process the interaction data of a high-performing learning group selected according to predetermined criteria, construct a standardized cognitive manifold that can represent the internal geometric structure of the efficient learning process through a manifold learning algorithm, and solidify the manifold and its structured information into a standardized cognitive model.

[0029] The cognitive trajectory generation module 30 operates online during system runtime. This module receives the feature vector sequence of an individual learner generated by the interaction event vectorization module 10 in real time, and generates high-dimensional trajectory data representing the learner's cognitive state via a time-series coding model.

[0030] The trajectory dynamics analysis module 40, which runs online, is the core processing unit of this system. This module receives real-time trajectory data generated by the cognitive trajectory generation module 30 and calls the normalized cognitive model generated by the normalized cognitive manifold construction module 20.

[0031] This module performs a dual analysis task: first, it analyzes the inherent dynamic characteristics of the trajectory data itself to obtain quantitative indicators characterizing its stability; Secondly, the trajectory data is projected onto a normalized cognitive manifold to calculate its deviation from the ideal learning state.

[0032] The cognitive pattern generation module 50 is used to integrate the intrinsic dynamic indicators and extrinsic deviation measures output by the trajectory dynamics analysis module 40 to construct a comprehensive diagnostic feature vector. Subsequently, the module matches the diagnostic feature vector with a preset pattern rule base to generate a structured diagnostic description of the current learning behavior.

[0033] See attached document Figure 2 , Figure 2 This invention provides a method for recognizing educational interactive behaviors in online learning environments, the method comprising the following steps: S101: Offline acquisition of multi-source interaction events of a predetermined high-performing learning group, and processing of the events into cognitive trajectory data; based on the cognitive trajectory data, constructing a normalized cognitive manifold through a manifold learning algorithm, and generating a normalized cognitive model.

[0034] S102 acquires multi-source interaction events of the current individual learner in real time online and processes them into a sequence of feature vectors containing contextual information.

[0035] S103, based on the feature vector sequence, generates a real-time high-dimensional cognitive trajectory representing the learner's cognitive state through a time-series coding model.

[0036] S104, call the normalized cognitive model generated in step S101 to analyze the real-time high-dimensional cognitive trajectory generated in step S103, so as to obtain at least one intrinsic dynamic index representing the stability of the trajectory and at least one extrinsic deviation measure representing the difference between the trajectory and the normalized cognitive manifold.

[0037] S105 integrates intrinsic dynamic indicators and extrinsic deviation measures, and generates a structured diagnostic description of the current learning behavior based on a preset pattern rule base.

[0038] In one embodiment of the present invention, the primary task of the interaction event vectorization module 10 is to comprehensively collect and structurally define the multi-source interaction events generated during the interaction between the learner and the online learning environment. This step is the basis for all subsequent analyses, and its purpose is to record the learner's behavior in a machine-readable and unambiguous format.

[0039] Specifically, the collection and definition of multi-source interaction events may include the following sub-steps: Define and categorize the types of interactive events to be collected. In order to comprehensively represent learning behavior, the event types cover multiple dimensions of learners' activities within the platform.

[0040] A standardized data structure is defined for each collected interaction event. This data structure exists in the form of log entries, ensuring data consistency and parsability.

[0041] The data acquisition mechanism is deployed and executed. The implementation of the data acquisition mechanism can be completed by the collaboration of front-end event listening and back-end log recording. The front-end listens to the user's interface operation through the script deployed in the webpage, and sends the captured events to the server after encapsulating them according to the defined structure. While receiving the front-end data, the server back-end also records the events triggered by the server itself. Finally, all event logs are persistently stored in the database or log system. For the specific technical implementation of data acquisition, those skilled in the art can complete it through standard web development technology, which is a well-known technology in the field and will not be described in detail here.

[0042] After the interaction event vectorization module 10 completes the collection and definition of the original interaction events, it will perform context-aware feature vectorization processing. The purpose of this processing is to convert each discrete, structured event log into a real number vector containing rich context information that can be processed by the downstream time series model.

[0043] For timestamps Events that occurred This process generates a 3D feature vector In one specific embodiment, the feature vector From the event's own feature vector Contextual semantic feature vector and micro-temporal feature vectors It is pieced together. Its formal representation is: ; in, This represents a vector concatenation operation, which connects multiple vectors end-to-end to form a higher-dimensional vector.

[0044] The specific generation methods for each component of the feature vector are as follows: Generate the event's own feature vector This vector is used to represent the basic categories of interactive events. Specifically, the system pre-counts all possible event types, forming a vector containing... A dictionary of event types. For events... The type is generated using one-hot encoding. 2D binary vector In this vector, only the dimension corresponding to this event type has a value of 1; the rest... The values ​​in all dimensions are 0.

[0045] Generate contextual semantic feature vectors This vector aims to quantify the deep semantic information of the objects affected by the event, rather than merely their identifiers. To achieve this, the system pre-constructs or introduces a knowledge graph about the course content, with entities as nodes and relationships as edges. The system uses knowledge graph embedding techniques, such as TransE or RotatE models, to train this graph, learning a low-dimensional, dense embedding vector for each entity in the graph.

[0046] The training methods for knowledge graph embedding are well-known techniques in this field and will not be elaborated upon here. During the vectorization process, the system starts from events... The identifier of the target object is extracted from the event target field, and the corresponding embedding vector is found in the pre-trained embedding model. This vector is then used as the context semantic feature vector. In this way, interactive objects are no longer isolated symbols, but are given information about their position and relationships within the entire knowledge system.

[0047] Generate micro-temporal feature vectors This vector is used to capture the local temporal dynamics of an event. In one embodiment, the vector consists of multiple numerical features, including at least: Interaction time interval : Calculate the difference between the current event timestamp and the previous event timestamp, i.e. .

[0048] Page dwell time : Calculate the event from when a self-learner enters the current page or opens the current learning resource. The duration of the event.

[0049] Session duration : Calculate from the start of the current session to the event The total duration elapsed since the occurrence of the event; to eliminate the dimensional differences between different time-series features, the system combines these numerical features into a vector. Previously, they would be standardized, for example, by using min-max normalization or Z-score normalization.

[0050] Through the above steps, the system will generate the original event sequence. Time series converted into a feature vector Each vector in this sequence contains a comprehensive representation of the corresponding interactive event in terms of behavior, semantics, and time, providing high-quality input for subsequent cognitive trajectory generation.

[0051] When the standardized cognitive manifold construction module 20 runs offline, it first performs the screening of standardized samples and the preparation of trajectory data.

[0052] Specifically, the process may include the following sub-steps: To define and implement the selection of high-performing learners, and to ensure the standardization of the constructed manifold, it is necessary to establish a set of objective, multi-dimensional evaluation criteria to define high-performing learners. In one embodiment, the system calculates a comprehensive performance score for each learner who has completed the learning task. This score is a weighted combination of multiple quantitative indicators: ; in: It is the final overall performance score.

[0053] It is the total number of quantitative indicators used.

[0054] It is the first The preset weight coefficients of each indicator, and satisfy the following: .

[0055] It is the first The value of each original quantitative indicator.

[0056] norm(·) is a normalization function, such as min-max normalization, used to map raw index values ​​of different dimensions to an interval to eliminate the influence of dimensions.

[0057] The system calculates the overall performance score of all candidate learners. Then, the high-performing learners were selected based on the preset percentile threshold, which is the standard sample.

[0058] Extract and process the interaction data from the standardized samples. For each high-performing learner selected above, the system extracts a complete, time-ordered sequence of interaction events throughout the entire learning cycle from the raw log database. To ensure data quality, the system can perform data cleaning operations, such as removing session data with abnormal durations or repairing log entries with incorrect data formats.

[0059] Generate a dataset of cognitive trajectory data for standardized samples. This step transforms the interaction event sequence of each standardized sample obtained in the previous step into a trajectory in a high-dimensional cognitive phase space. Specifically, the system sequentially calls the functions of the interaction event vectorization module 10 and the cognitive trajectory generation module 30. First, the interaction event sequence is input into the interaction event vectorization module 10 to generate the corresponding feature vector sequence. Subsequently, the feature vector sequence Input the temporal coding model in the cognitive trajectory generation module 30, and output the learner's complete cognitive trajectory. .

[0060] By repeating this step on all standardized samples, a set consisting of multiple high-dimensional cognitive trajectories is eventually obtained, denoted as . normativesampleset This trajectory dataset will be used as input data for the next stage of manifold learning.

[0061] After the normalized cognitive manifold construction module 20 completes the trajectory data preparation, the manifold learning process based on geodesic distance begins. The core technical assumption of this process is that the high-dimensional cognitive trajectory data set generated above... Although embedded in a high-dimensional Euclidean space In the middle, but its data points are densely distributed in one dimension far lower than Intrinsic low-dimensional manifold Above. The purpose of this process is to promptly discover and construct this low-dimensional manifold that can characterize the intrinsic geometry of efficient learning behavior. .

[0062] In one specific embodiment, the process is implemented using an isometric mapping algorithm, which may include the following sub-steps: Constructing a nearest neighbor graph is a step aimed at determining the local connectivity between trajectory data points in a high-dimensional space, with the input being a set of trajectory data. ,in It is the total number of data points contained in all the canonical sample trajectories, and each data point The system assigns each data point Under its Euclidean distance metric, find the nearest distance among the remaining data points. Each point, that is, its - Neighbors.

[0063] Based on this, construct an undirected graph. In the figure In, each data point As a node; if node It is a node of -One of the neighbors, or yes of -One of the neighbors, then in and An edge is established between the two points. The weight of the edge is set to the Euclidean distance between the two points. .

[0064] Calculate the geodesic distance matrix; this step aims to approximate the manifold. The system uses the true distance between any two points, i.e., the geodesic distance, rather than the straight-line distance in high-dimensional space, to construct the nearest neighbor graph. Considered as a manifold A discrete approximation, for any two nodes and The geodesic distance between them is approximated by their distance in the figure. The shortest path length on the [theory / path].

[0065] The system employs a shortest path algorithm to compute the graph. Find the shortest path length between all pairs of nodes in the algorithm. Finally, this step outputs a... Symmetric distance matrix , of which elements Represents a node and Approximate geodesic distance between them.

[0066] The final output of this step is a new data set. Each of them It is the original high-dimensional data points Mapped coordinates in a low-dimensional space, this set of low-dimensional data points. This refers to the standardized cognitive manifold constructed in this invention. A specific numerical representation of.

[0067] In obtaining a standardized cognitive manifold Low-dimensional numerical representation Subsequently, the normalized cognitive manifold construction module 20 performs the analysis of the manifold structure. This analysis process aims to extract structured information with clear pedagogical significance from the low-dimensional geometric structure, specifically including normalized attractors representing stable cognitive states and optimal paths representing efficient cognitive state transitions.

[0068] In one embodiment, the DBSCAN algorithm is employed and applied to low-dimensional datasets. The DBSCAN algorithm operates using two core parameters: Neighborhood radius And the minimum number of points MinPts, the system will satisfy its condition. Points within a neighborhood containing at least MinPts data points are identified as core points. Starting from these core points, the algorithm expands, merging data points with achievable density into clusters. This algorithm can discover clusters of arbitrary shapes without pre-specifying the number of clusters. The specific implementation of the DBSCAN algorithm is a well-known technique in the field of data mining and will not be elaborated upon here. The output of this step is a set of data point clusters, each of which is defined as a normalized attractor. .

[0069] The optimal path is determined by connecting any two normalized attractors. and The geodesic path between these points represents the most efficient trajectory for high-performing learners to transition from one stable cognitive state to another. To determine this path, the system reuses the nearest neighbor graph constructed in the step of calculating the geodesic distance. And the shortest path algorithm.

[0070] Specifically, for any pair of normalized attractors and The system first calculates their respective centroids. and In low-dimensional space The coordinates in the original high-dimensional data point set. In the middle, find the center of mass respectively. and Corresponding nearest neighbor data point and The system is based on Starting from, As the endpoint, in the nearest neighbor graph Then, Dijkstra's shortest path algorithm is executed again to find a path consisting of a series of graph nodes connecting the two nodes. This path is defined as the connection attractor. and Optimal path .

[0071] Through the above steps, the normalized cognitive manifold construction module 20 ultimately generates and stores a structured normalized cognitive model, which at least includes: a normalized cognitive manifold. Low-dimensional representation All normalized attractors The definition of attractors and the optimal path connecting them. This model will serve as the core reference for the online analysis phase.

[0072] In the cognitive trajectory generation module 30, its core task is to receive the feature vector sequence generated by the interaction event vectorization module 10 and transform it into a high-dimensional vector that can represent the learner's comprehensive cognitive state at a specific moment. This representation process is completed through a pre-trained temporal coding model.

[0073] Specifically, at any point in time The cognitive trajectory generation module 30 receives the learner's complete feature vector sequence up to that moment. As input. The time-series coding model, denoted as... The input sequence is processed and output as a fixed-dimensional vector. This vector represents the learner's position at time point. The cognitive state is represented. This process can be formally represented as: ; in: It is the deadline. The input feature vector sequence.

[0074] It is a time-series coding model.

[0075] It is output. Dimensional cognitive state vector.

[0076] In one specific embodiment, the time-series coding model This can be an encoder based on the Transformer architecture. Through its internal self-attention mechanism, this encoder can effectively capture the long-distance dependencies between various interaction events in the input sequence, thereby generating a cognitive state representation that comprehensively considers the historical behavioral context. The computation of its self-attention mechanism can be summarized as follows: ; in , , These are the query, key, and content matrices generated from the input sequence, respectively. It is the dimension of the key vector. Through this mechanism, the model can compute a weighted representation for each element in the sequence. This weighted representation integrates information from all other elements in the sequence. The specific network structure and training method of the Transformer encoder are well-known techniques in the field of deep learning and will not be elaborated here.

[0077] Finally, this step outputs a cognitive state vector. Its dimensions Maintaining consistency with the high-dimensional phase space dimension used by the normalized cognitive manifold construction module 20 ensures that individual cognitive states can be mapped to the same space as the normalized samples for subsequent comparison and analysis.

[0078] The cognitive trajectory generation module 30 generates instantaneous cognitive state vectors through a temporal coding model. Next, the dynamic trajectory is constructed in real time. The purpose of this process is to connect a series of discrete, time-ordered cognitive state vectors into a continuous path that reflects the temporal evolution of the learner's cognitive state, i.e., the cognitive trajectory.

[0079] In one embodiment, the process of constructing the dynamic trajectory is as follows: The trajectory data container is initialized when a learner's session begins and their first interaction event is processed, generating the first cognitive state vector. At that time, the system initializes an empty, ordered trajectory data container for the learner, which is used to store all the state vectors generated subsequently.

[0080] The state vector is appended in real time. During the learning session, whenever a new interaction event occurs, a new cognitive state vector is generated by the aforementioned steps. At that time, the cognitive trajectory generation module 30 will generate the vector Appended to the end of the trajectory data container for the corresponding learner, this operation continues, causing the vector sequence stored in the container to grow dynamically over time.

[0081] Therefore, at any time The learner's complete cognitive trajectory can be represented as a series of... An ordered sequence of cognitive state vectors : ; in, It is in the A cognitive state vector is generated when an interactive event occurs.

[0082] In an optional embodiment, to balance computational efficiency with a focus on recent behavior, the trajectory data container can be implemented as a fixed-length container. In this implementation of the sliding window, when a new state vector is added... When the container is full, the oldest state vector in the container is... It will be removed. This method ensures that the trajectory used by the system for analysis is always the most recent. The evolution of cognitive states corresponding to each interactive event helps reduce the complexity of data processing and improve the system's response speed.

[0083] The final output of this step is the dynamic cognitive trajectory. Both the complete historical trajectory and the recent trajectory based on the sliding window will be used as core data input and passed to the trajectory dynamics analysis module 40 for subsequent in-depth analysis.

[0084] After receiving the real-time cognitive trajectory output by the cognitive trajectory generation module 30, the trajectory dynamics analysis module 40 executes its core analysis function, which is based on a dual-reference analysis system to quantitatively evaluate the learner's cognitive process from two complementary dimensions.

[0085] By integrating the results of internal and external reference analysis, this system can form a comprehensive and three-dimensional behavioral diagnosis. It not only reveals the immediate nature of the learner's cognitive state, but also measures the gap between the learning process and the ideal model, providing multi-dimensional quantitative basis for the subsequent generation of accurate cognitive pattern diagnosis.

[0086] In the trajectory dynamics analysis module 40, the intrinsic dynamics analysis aims to reveal the instantaneous stability and predictability of the learner's cognitive process by quantifying the intrinsic structure and evolutionary characteristics of the real-time cognitive trajectory. This analysis does not rely on any external reference, but rather dissects the trajectory itself as the output of a dynamic system. In one embodiment, the analysis mainly includes the quantification of local stability and the capture of state abrupt changes.

[0087] To quantify the local stability of the cognitive trajectory, that is, to measure whether the system tends to converge or diverge under small perturbations, this system calculates the maximum Lyapunov exponent of the trajectory.

[0088] A positive MLE value indicates that the trajectory is chaotic, and neighboring states will separate exponentially over time, corresponding to the learner being in a highly exploratory but unpredictable cognitive state; a negative MLE value indicates that the trajectory is stable, and neighboring states will tend to converge, corresponding to the learner being in a cognitive state with clear goals and fixed behavioral patterns.

[0089] The calculation process can specifically include the following sub-steps: Phase space reconstruction, due to the original cognitive trajectory It is a time series, and to reveal its inherent dynamic structure, the phase space needs to be reconstructed first according to Takens' embedding theorem. The system selects the embedding dimension... and time delay Two parameters, one-dimensional time series Converted into a series of high-dimensional vectors, the reconstructed first... Each phase space vector Defined as: ; in The range of values ​​for is .

[0090] Find the nearest neighbor and calculate the initial distance for each point in the reconstructed phase space. The system employs a fast nearest neighbor search algorithm to find its nearest neighbor in Euclidean distance. To avoid false nearest neighbors caused by autocorrelation in time series data, a time window constraint is imposed during the search, requiring... The initial distance between two points is denoted as greater than one average period. .

[0091] By tracking the evolution of nearest neighbor pairs and calculating the divergence rate, the system evolves forward in time. Step, calculate the nearest neighbor points in Distance after step .

[0092] The maximum Lyapunov exponent can be estimated by averaging the logarithms of the divergence rates of all initial point pairs. A commonly used estimation method uses the following formula: ; in: This is the time step between events, which can be set to 1 here.

[0093] It is the number of evolutionary steps.

[0094] It is the set of all valid initial point indices.

[0095] and They are the first The evolution of nearest neighbors The distance after the step and the initial distance.

[0096] For the selection of phase space reconstruction parameters and the specific estimation algorithm of MLE, those skilled in the art can refer to the standard methods in nonlinear time series analysis, which are well-known techniques in this field and will not be elaborated here.

[0097] To capture the key points where learners' cognitive states undergo drastic changes, namely phase transition points, this system employs recursive quantitative analysis.

[0098] The analysis process can specifically include the following sub-steps: Construct a recursive graph. The system's cognitive trajectory Set a width of A sliding window. For each trajectory segment within the window, the system constructs a recursive graph, which is a... matrix Its elements Defined as: ; in: and These are two state vectors within the window.

[0099] It is a preset distance threshold.

[0100] It is the Heaviside step function, which has a value of 1 when the independent variable is greater than or equal to zero, and a value of 0 otherwise. The point in the recursion graph... If it is 1, it indicates a state. and As they approached each other, the system recursively initiated at that moment.

[0101] The system calculates recursive quantification metrics. It extracts a series of quantification metrics from the recursion graph. A key metric for detecting phase transitions is the recursion rate, which is defined as the density of recursion points in the recursion graph.

[0102] ; The system identifies phase transition points and continuously calculates as the sliding window moves along the entire trajectory. value, thus obtaining a In a time series, when the value of the series rises or falls sharply and the rate of change exceeds a preset threshold, the time point is identified as a phase transition point, which marks a significant change in the dynamic characteristics of the learner's cognitive process.

[0103] To quantify the local stability in the intrinsic dynamic characteristics, the trajectory dynamics analysis module 40 performs the estimation of the maximum Leapnov exponent, MLE, a core indicator derived from dynamical system theory, used to measure the average exponential rate at which two adjacent trajectories in the system evolve towards or away from each other over time.

[0104] In the application scenario of this invention, a positive MLE indicates that the learner's cognitive trajectory has chaotic characteristics, indicating that its cognitive state is sensitive to initial conditions and its behavioral pattern is difficult to predict, possibly corresponding to a divergent exploration or disorientation state. Conversely, a negative or near-zero MLE indicates the stability of the trajectory, indicating that the cognitive state is not easily affected by small perturbations and the behavioral pattern is relatively fixed, possibly corresponding to a goal-oriented and convergent learning state.

[0105] In a specific embodiment, this estimation process can employ a time-series-based reconstructed phase space method, which specifically includes the following sub-steps: Perform phase space reconstruction and input real-time cognitive trajectory It is a high-dimensional time series. To reveal its embedded dynamics, it is first necessary to reconstruct it into a higher-dimensional phase space according to Takens' embedding theorem. This process requires determining two key parameters: Embedded Dimension and time delay In one embodiment, time delay It can be determined by calculating the average mutual information function of the time series and taking its first minimum point; Embedded Dimension The pseudo-nearest neighbor method can be used to determine the parameters. The specific implementation of these two parameter selection methods is a well-known technique in the field of nonlinear time series analysis, and will not be elaborated here.

[0106] After determining the parameters, the system converts the original trajectory sequence into a set of reconstructed phase space vectors. : ; in The range of values ​​is .

[0107] For each vector in the phase space, find the nearest neighbor. For each vector in the reconstructed phase space... The system traverses all other vectors in the phase space. And based on Euclidean distance Find its nearest neighbor .

[0108] During this search process, to eliminate false nearest neighbors found due to autocorrelation of time series data, the system applies a time window constraint, which requires the time index of the neighboring points to be specified. Time index of the current point The absolute value of the difference It must be greater than a preset average period The initial distance between the two points is denoted as . .

[0109] By tracking the distance evolution of nearest neighbor pairs, the system progresses forward on the timeline. Each time step, and calculates at... After the step, the original nearest neighbor pairs New position evolved The distance between them, which is denoted as .

[0110] Calculate and fit the mean log-divergence; for all valid initial point pairs, calculate their log-divergence. And by averaging these values, we obtain the mean logarithmic divergence. With evolution time According to the definition of the Lyapunov index, the slope of the changing curve in the initial linear growth region is the maximum Lyapunov index. A robust estimate.

[0111] Therefore, the system performs linear regression fitting on the linear region of the curve and uses the resulting slope as the final MLE estimate. This relationship can be expressed as: ; in: It is the largest Lyapunov index to be estimated.

[0112] It is the discrete evolution time step.

[0113] It is the actual time interval corresponding to each time step, which can be set to 1 in this embodiment.

[0114] It is the intercept of the fitted line.

[0115] Through the above steps, the system calculates a quantitative intrinsic dynamic index for the current learner's cognitive trajectory, namely the maximum Lyapunov index. This metric will be used for subsequent cognitive pattern generation.

[0116] To capture abrupt changes in the intrinsic dynamic characteristics, the trajectory dynamics analysis module 40 performs phase transition detection. This process aims to identify key time points in the learner's cognitive trajectory where structural or qualitative shifts occur. These time points, i.e., phase transition points, typically correspond to fundamental changes in learning strategies, insights, or rapid transitions from one stable state to another. In embodiments of the present invention, this detection process is implemented through recursive quantitative analysis.

[0117] Specifically, the detection process may include the following sub-steps: Construct a temporal recursive graph. The system receives real-time cognitive trajectories. Above, define a variable with a fixed width. A sliding window. For trajectory segments located within this window... The system constructs a recursion graph. The recursion graph is a graph of size ... binary matrix Its matrix elements The calculation method is as follows: ; in: These are the row and column indices of the matrix, with values ​​ranging from 1 to... .

[0118] and These are two cognitive state vectors within a sliding window.

[0119] This represents the Euclidean distance between two vectors.

[0120] It is a pre-set distance threshold used to determine whether two states are recursive (i.e., sufficiently similar).

[0121] It is the Heaviside step function. The function value is 1 when its independent variable is greater than or equal to zero; otherwise, it is 0.

[0122] Therefore, matrix One point If it is 1, it means at time point and Their cognitive states are similar.

[0123] The recursive quantization index is calculated because the recursive graph itself provides a visualization of the system's dynamics, while the recursive quantization index transforms its structural characteristics into numerical values. The system progresses from each recursive graph... A series of metrics are calculated, and in one embodiment, a key diagnostic metric is the recurrence rate, which quantifies how frequently a state is repeated within a window: ; This formula calculates the proportion of points with a value of 1 in the recursive graph.

[0124] Identifying phase transition points, the system follows the sliding window throughout the cognitive trajectory. Move upwards at each time point Perform the above steps to obtain a time series of recurrence rates. .

[0125] Phase transitions in cognitive states cause drastic changes in the dynamic characteristics of the system, which are directly reflected in the structure of the recursion graph, and consequently affect the recursion rate. The value changes drastically and non-linearly. Therefore, the system monitors... The first difference (i.e., the rate of change) of the sequence is used to identify phase transition points. A phase transition point is identified when the following conditions are met: Marked as a phase transition point: ; in This is a preset sensitivity threshold for phase transition detection. All marked phase transition points will be used as an important intrinsic dynamic indicator for subsequent cognitive pattern generation.

[0126] In the trajectory dynamics analysis module 40, external projection and deviation measurement are the second core components of the dual reference analysis system. This process compares the current learner's real-time cognitive trajectory with the normalized cognitive model pre-generated by the normalized cognitive manifold construction module 20. The aim is to quantitatively evaluate the effectiveness and normalization of the current learning behavior from the perspective of an external reference system. This process mainly includes projecting the real-time cognitive state onto the normalized manifold and calculating two types of measurement indicators: state deviation and path deviation.

[0127] In one specific embodiment, the process may include the following sub-steps: Perform a manifold projection of the real-time cognitive state. For the cognitive trajectory generation module 30 at time points... Output current cognitive state vector The system is based on the original high-dimensional data point set corresponding to the normalized cognitive manifold. In the middle, searching for and The closest point.

[0128] The search process is essentially a nearest neighbor search, and the found point is defined as... The projection onto the normalized cognitive manifold is denoted as Its mathematical definition is: ; in: It is the learner's real-time cognitive state vector.

[0129] It is the set of all high-dimensional data points that constitute a normalized cognitive manifold.

[0130] This represents the Euclidean distance between two vectors.

[0131] yes The projection point on the manifold.

[0132] The state deviation metric is calculated to quantify the geometric distance between the current cognitive state and the region of efficient learning states. This metric is directly defined as the current state point. Its projection point on the manifold The Euclidean distance between them. This is indicated by... : ; A larger A value of [value] indicates that the learner's current cognitive state deviates significantly from any known normative learning state, and may be in an inefficient or ineffective learning pattern. Conversely, a smaller value indicates that their cognitive state aligns with that of a high-performing learner. It will serve as a key external reference indicator.

[0133] The path deviation metric is calculated to assess whether the evolution of a learner's cognitive state follows the optimal path in the canonical model. The system first extracts segments of the learner's recent cognitive trajectory, such as a sequence containing... Time window for each state point .

[0134] The system determines the starting point of the trajectory segment. and the end point Projection on the manifold and Subsequently, the system queries and retrieves the optimal path connecting the attractors containing the two projection points from the pre-parsed normalized cognitive model. Finally, the system employs an algorithm capable of measuring the similarity between two time-series curves to calculate the trajectory segment. with the optimal path The differences between them.

[0135] In one embodiment, a Dynamic Time Warping (DTW) algorithm can be used. The DTW algorithm can find the optimal alignment between two sequences of potentially different lengths and calculate the minimum cumulative distance. The specific implementation of the DTW algorithm is a well-known technique in the fields of signal processing and data mining, and will not be elaborated upon here. The calculation result is the path deviation metric. : ; A larger The values ​​indicate that learners' transitions between two cognitive states differ significantly from the efficient paths employed by high-performing learners, suggesting that they may have adopted roundabout or ineffective exploratory methods.

[0136] As the core of the external projection and deviation measurement process, the calculation of cognitive deviation aims to accurately quantify and compare the learner's current cognitive process with the standardized model. Cognitive deviation is a comprehensive external reference indicator that characterizes the difference between the learner's current behavior and high-performance learning patterns from both state and path perspectives. Specifically, the calculation process can be broken down into separate calculations of state deviation and path deviation.

[0137] Calculate the state deviation, an indicator used to measure the degree to which a learner's instantaneous cognitive state deviates from the normalized cognitive manifold at a specific moment.

[0138] First, the system obtains the current time. Cognitive state vector .

[0139] Subsequently, the system needs to work on the high-dimensional data point set of the normalized cognitive manifold. In the middle, for Find a unique projection point .

[0140] The projection point is a set Zhongyu The point with the minimum Euclidean distance. To efficiently perform this nearest neighbor search among massive amounts of data points, in one specific embodiment, the system can pre-utilize... Construct a spatial index structure such as a kd-tree or a ball tree, or employ a library of approximate nearest neighbor search algorithms. The implementation of such search algorithms is well-known in computer science and will not be elaborated upon here.

[0141] Finally, state deviation Defined as Its projection point Euclidean distance between them: ; This value directly reflects the normativity of the current cognitive state itself. The larger the value, the rarer the state is in the normative learning pattern, and the higher the degree of deviation.

[0142] The path deviation is calculated, which measures the similarity between the learner's cognitive state evolution trajectory and the optimal path in the normalized model over a period of time.

[0143] First, the system captures the current moment. The previous segment was of length The cognitive trajectory segment is denoted as .

[0144] The system determines the starting point of the trajectory segment. and the end point And following the method described above, their projection points on the normalized cognitive manifold are calculated. and .

[0145] The system queries a pre-built, standardized cognitive model to find models that contain... and Two normalized attractors are identified, and the optimal path connecting these two attractors is extracted. .

[0146] The system employs a dynamic time warping algorithm to calculate the observed trajectory segments. Optimal path with reference To measure the similarity between two sequences of unequal length, the DTW algorithm constructs a cost matrix and finds a regular path that minimizes the cumulative distance. The minimum cumulative distance calculated by this algorithm is the path deviation. .

[0147] ; This value reflects the efficiency and rationality of the cognitive state transition process. The larger the value, the more circuitous or inefficient the path taken by the learner to reach the current cognitive state is compared with the normative path.

[0148] Through the above steps, the system at each time point Both yielded a two-dimensional vector consisting of state deviation and path deviation. These vectors together constitute a comprehensive quantification of the cognitive deviation in the learner's cognitive process.

[0149] As a complement to and enhancement of path deviation quantification, the trajectory dynamics analysis module 40 also performs a path efficiency assessment, a process designed to provide a more intuitive and standardized metric for measuring the economy of learners’ movement in the cognitive state space.

[0150] In one specific embodiment, the evaluation process may include the following sub-steps: The system first extracts and measures the length of the observed trajectory segment from the real-time cognitive trajectory for evaluation. This trajectory segment can be compared with the trajectory segment used in the path deviation calculation. To maintain consistency, the system calculates the path length of the observed trajectory segment. This length is defined as the sum of the Euclidean distances between all adjacent cognitive state vectors that constitute this trajectory segment: ; in: and These are two temporally adjacent cognitive state vectors within a trajectory segment.

[0151] It is the length of the selected trajectory segment.

[0152] This represents the Euclidean distance.

[0153] The value quantifies the total distance a learner travels in the cognitive phase space over a recent period.

[0154] Obtain and measure the length of the corresponding optimal path. The system reuses the path deviation calculation step, which determines the length of the observed trajectory segment. The optimal path corresponding to the start and end points The optimal path It is an ordered sequence of data points on a normalized cognitive manifold. The system calculates the length of the optimal path in the same way. : ; in and These are two adjacent points in the optimal path sequence. This represents the most economical path length required to complete the transition from the starting point to the end point of the same cognitive state in the high-performing learner model.

[0155] The path efficiency index is calculated. Path efficiency is defined as the ratio of the optimal path length to the observed path length. The system calculates the path efficiency index. : ; This indicator The value of is between 0 and 1. A value close to 1 indicates that the learner's path length is similar to that of the optimal path, reflecting extremely high path efficiency. A value close to 0 indicates that the learner's path length is much longer than the optimal path, suggesting that there are a lot of detours, repetitions or invalid explorations during the state transition process.

[0156] Through the above steps, the system generates a quantitative path efficiency index for the current learning process. This indicator, together with the aforementioned state deviation and path deviation, constitutes a complete characterization of external reference analysis, providing a comprehensive and multi-dimensional quantitative basis for subsequent cognitive pattern diagnosis.

[0157] In the cognitive pattern diagnosis module 50, the first step is to construct a diagnostic feature vector. This process aims to integrate and normalize multiple heterogeneous quantitative indicators output from the trajectory dynamics analysis module 40, which originate from the dual-reference analysis system, into a unified and structured data representation for use by subsequent pattern recognition algorithms.

[0158] This process occurs at any given time. The input is a series of intrinsic dynamic indicators and extrinsic reference indicators calculated by the trajectory dynamics analysis module 40 at that moment. In a specific embodiment, these indicators include at least: The maximum Liapnov exponent for quantizing the local stability of the trajectory .

[0159] Binary variables used to identify state changes When a phase transition point is detected, its value is 1; otherwise, it is 0.

[0160] State deviation degree, which quantifies the degree to which an instantaneous state deviates from the normalized cognitive manifold. .

[0161] Path deviation of quantifying the difference between cognitive trajectory segments and the optimal path .

[0162] Path efficiency for evaluating the economy of state transition processes .

[0163] The process of constructing diagnostic feature vectors involves taking these independent scalar indicators at a specific time... The vectors are then concatenated in an ordered manner to form a multi-dimensional row vector. This vector is the diagnostic feature vector at that moment, denoted as... Its composition can be formally represented as: ; Each component of the vector corresponds to a specific diagnostic indicator mentioned above.

[0164] In an optional embodiment, since the sources and physical meanings of the above indicators are different, their numerical ranges may vary greatly. In order to avoid the feature with a large numerical range dominating the distance calculation or model training in the subsequent pattern recognition algorithm, the system can perform normalization processing on some or all indicators before constructing the vector.

[0165] Through the above steps, the cognitive pattern diagnosis module 50 successfully transforms the multi-dimensional analysis results from the upstream module into a unified, numerical diagnostic feature vector. Over time, the system will generate a sequence of diagnostic feature vectors composed of these vectors. This sequence will serve as the direct input for the final cognitive pattern classification or clustering.

[0166] Diagnostic feature vectors were constructed in the cognitive pattern diagnosis module 50. The system then performs pattern interpretation and generation based on a rule base. The core task of this process is to transform the purely numerical diagnostic feature vectors generated in the previous stage into cognitive pattern diagnostic conclusions that have clear educational or psychological meanings and can be understood by humans. This transformation process is accomplished through a predefined rule base constructed from the knowledge of domain experts.

[0167] In one specific embodiment, the rule base consists of a series of logical rules in the form of "if" and "then". Each rule will diagnose the feature vector. A specific combination of numerical values ​​is mapped to a specific cognitive pattern label and its interpretation.

[0168] To perform rule matching, the system iterates through each rule in the rule base and uses the current diagnostic feature vector. The values ​​of each component are used to evaluate whether the condition defined in the if part of the rule is true. The condition part consists of one or more logical expressions about the characteristic components.

[0169] Through this rule-based matching and generation process, the system's final output is no longer an abstract numerical value, but a cognitive pattern diagnostic result that clearly reflects the learner's current learning status, the problems, and has clear guiding significance. This result will directly serve as the decision-making basis for the cognitive intervention generation module 60.

Claims

1. An educational interactive behavior recognition system for online learning environments, characterized in that: include: The cognitive manifold construction module is used to acquire multi-source interaction events of a predetermined high-performance learning group offline, process the events into cognitive trajectory data, construct a normalized cognitive manifold based on the cognitive trajectory data through a manifold learning algorithm, and generate a normalized cognitive model. The event processing module is used to acquire multi-source interaction events of the current individual learner in real time online and process them into a sequence of feature vectors containing contextual information. The cognitive trajectory generation module is used to generate a real-time high-dimensional cognitive trajectory representing the learner's cognitive state based on the feature vector sequence and through a time-series coding model. The trajectory dynamics analysis module is used to call the normalized cognitive model to analyze the real-time high-dimensional cognitive trajectory in order to obtain at least one intrinsic dynamic index characterizing trajectory stability and at least one extrinsic deviation measure characterizing the difference between the trajectory and the normalized cognitive manifold. The cognitive pattern diagnosis module is used to integrate the intrinsic dynamic indicators and the extrinsic deviation measures, and generate a structured diagnostic description of the current learning behavior based on a preset pattern rule base.

2. The educational interactive behavior recognition system for online learning environments according to claim 1, characterized in that, The cognitive manifold construction module is specifically used for: Acquire multi-source interaction events of the high-performing learning group, including but not limited to learning duration, answer results, click behavior, and browsing sequence; The event is encoded into a high-dimensional feature vector sequence; Based on the high-dimensional feature vector sequence, the normalized cognitive manifold is identified through a dimensionality reduction manifold learning algorithm, and the normalized cognitive manifold is expressed as an attractor network structure. The attractor network structure consists of multiple attractor state nodes and edges connecting the attractor state nodes, and the edges represent cognitive state transition paths. Generate the standardized cognitive model, which includes the attractor network structure and the representation of the attractor state nodes.

3. The educational interactive behavior recognition system for online learning environments according to claim 1, characterized in that, The event handling module is specifically used for: The real-time acquired multi-source interaction events are parsed and feature extracted to obtain the original feature vector; The original feature vector is context-encoded using a self-attention mechanism to generate the feature vector sequence containing context information.

4. The educational interactive behavior recognition system for online learning environments according to claim 1, characterized in that, The cognitive trajectory generation module is specifically used for: The feature vector sequence is encoded using the temporal coding model to generate an instantaneous cognitive state vector representing the learner's cognitive state. The instantaneous cognitive state vector is appended to the learner's corresponding trajectory data container in real time in chronological order to construct the real-time high-dimensional cognitive trajectory.

5. The educational interactive behavior recognition system for online learning environments according to claim 4, characterized in that, The trajectory data container is a fixed-length sliding window. When a new instantaneous cognitive state vector is added, the oldest instantaneous cognitive state vector in the sliding window is removed.

6. The educational interactive behavior recognition system for online learning environments according to claim 1, characterized in that, The trajectory dynamics analysis module is specifically used for: Based on the real-time high-dimensional cognitive trajectory, the maximum Liapunov exponent of the trajectory is estimated using the phase space reconstruction method as one of the intrinsic dynamic indicators; Based on the real-time high-dimensional cognitive trajectory, the phase transition points in the trajectory are identified as one of the intrinsic dynamic indicators through recursive quantification analysis.

7. The educational interactive behavior recognition system for online learning environments according to claim 1, characterized in that, The trajectory dynamics analysis module is specifically used for: The current cognitive state vector in the real-time high-dimensional cognitive trajectory is projected onto the normalized cognitive manifold to obtain the Euclidean distance between the current cognitive state vector and the normalized cognitive manifold as the state deviation degree in the external deviation metric. The recent trajectory segment in the real-time high-dimensional cognitive trajectory is compared with the corresponding optimal path in the normalized cognitive model to obtain the difference between the recent trajectory segment and the optimal path as the path deviation degree in the external deviation metric.

8. The educational interactive behavior recognition system for online learning environments according to claim 7, characterized in that, The trajectory dynamics analysis module is also used for: Calculate the actual length of the recent trajectory segment; Obtain the length of the optimal path; Based on the actual length and the length of the optimal path, the path efficiency of the real-time high-dimensional cognitive trajectory is calculated as one of the external deviation measures.

9. The educational interactive behavior recognition system for online learning environments according to claim 1, characterized in that, The cognitive pattern diagnosis module is specifically used for: The intrinsic dynamics index and the extrinsic deviation metric are integrated into a diagnostic feature vector; The diagnostic feature vector is matched with the pattern rule base to generate the structured diagnostic description. The pattern rule base contains if and then rules that map specific numerical combinations of the diagnostic feature vector to cognitive pattern labels and descriptions.

10. A method for recognizing educational interactive behaviors in online learning environments, and a system for recognizing educational interactive behaviors in online learning environments according to any one of claims 1-9, characterized in that, Includes the following steps: Offline acquisition of multi-source interaction events of a predetermined high-performing learning group; processing of the events into cognitive trajectory data; construction of a normalized cognitive manifold based on the cognitive trajectory data using a manifold learning algorithm; and generation of a normalized cognitive model. The system acquires multi-source interaction events of the current individual learner in real time online and processes them into a sequence of feature vectors containing contextual information. Based on the feature vector sequence, a real-time high-dimensional cognitive trajectory representing the learner's cognitive state is generated through a time-series coding model. The normalized cognitive model is invoked to analyze the real-time high-dimensional cognitive trajectory in order to obtain at least one intrinsic dynamic index characterizing the stability of the trajectory and at least one extrinsic deviation measure characterizing the difference between the trajectory and the normalized cognitive manifold. By integrating the intrinsic dynamic indicators and the extrinsic deviation measures, and based on a preset pattern rule base, a structured diagnostic description of the current learning behavior is generated.

Images