A method and system for providing feedback on adolescent learning data

Abstract

The application discloses a kind of evaluation feedback method and system of adolescent learning data, it is related to data analysis technical field.A kind of evaluation feedback system of adolescent learning data, including have: learning data classification module and learning data analysis module.The application is by dividing learning behavior into learning signal node and learning noise node, realizes the effective distinction of focused learning and distraction behavior, ensures the accuracy of learning process identification;Introduce time attenuation weight and feature vector calculation mechanism, model uniformly and construct learning signal-to-noise ratio to learning internal effectiveness and learning external deviation, so that learning state evaluation has objectivity and dynamics;Learning cognitive path graph based on graph theory constructs learning process from path structure level, can identify the jump and back-off behavior in learning, make up the deficiency that traditional method only pays attention to result.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, and in particular to an evaluation and feedback method and system for adolescent learning data. Background Technology

[0002] In the current digital learning environment, adolescents' learning behavior exhibits highly non-linear and fragmented characteristics. Existing technologies for assessing learning status often focus on outcome indicators, such as correct answer rates, test scores, or simple statistics on online learning time. However, these objective data are difficult to reflect the cognitive quality and behavioral characteristics during the learning process, ignoring the distracting operations of users when using electronic devices for learning, and failing to effectively identify whether users are engaged in deep thinking or ineffective mechanical operations.

[0003] Furthermore, traditional assessment methods lack a detailed analysis of the efficiency of learning paths, making it difficult to quantify and identify inefficient behaviors such as distraction or logical regression that frequently occur in adolescents during the learning process. Due to the lack of a unified assessment model that can simultaneously characterize the level of focus and the flow of thought, existing teaching feedback is often lagging behind and cannot provide targeted environmental adjustments or intervention suggestions based on the user's immediate cognitive state. Summary of the Invention

[0004] This invention aims to provide a method and system for evaluating and providing feedback on learning data of teenagers. By integrating the learning signal-to-noise ratio and the learning cognitive return rate, it solves the problem that existing technologies cannot effectively distinguish between focused learning behavior and distracted behavior, and achieves objective identification of the quality of the learning process.

[0005] An assessment and feedback method for adolescent learning data includes the following steps: Acquire learning behavior sequence data of adolescent users within a preset time period; analyze the learning behavior sequence data to obtain learning signal node data and learning noise node data; Data analysis is performed on the learning signal node data and learning noise node data to obtain the user's learning signal-to-noise ratio; a learning cognitive path graph is constructed based on graph theory algorithms and learning behavior sequence data; and the learning cognitive return rate is obtained by analyzing the learning cognitive path graph. A two-dimensional user learning profile space is constructed with the user learning signal-to-noise ratio as the x-axis and the learning cognitive return rate as the y-axis. Based on the two-dimensional user learning profile space, the learning status of adolescent users is evaluated, and user learning evaluation feedback results are obtained.

[0006] As a preferred embodiment of the present invention, the specific steps for analyzing learning behavior sequence data to obtain learning signal node data and learning noise node data include: Obtain user learning task requirements within a preset time period; parse user learning task requirements to obtain several learning task nodes with a sequential order; obtain learning derivative nodes of all learning task nodes based on existing learning task nodes and a preset knowledge association database. A main learning task path is constructed based on several learning task nodes that have a sequential order; learning derivative nodes are then reconstructed based on the learning task nodes corresponding to the main learning task path to obtain a learning task path diagram; historical behavior patterns of adolescent users are obtained; cluster analysis is performed on the historical behavior model to obtain the user's learning deviation tolerance. Semantic parsing is performed on the learning behavior sequence data to obtain several learning behavior nodes; the learning behavior nodes are judged according to the learning task path graph and the user's learning deviation tolerance to obtain learning signal nodes and non-learning signal nodes; the corresponding learning signal node data and learning noise node data are obtained based on the learning signal nodes and non-learning signal nodes.

[0007] As a preferred embodiment of the present invention, the specific steps for data analysis of learning signal node data and learning noise node data include: Extract the learning behavior feature vector of each learning signal node from the learning signal node data; extract the noise behavior feature vector of each non-learning signal node from the learning noise node data; Obtain the timestamps of all learning signal nodes and non-learning signal nodes; weight the learning signal nodes and non-learning signal nodes according to their contribution based on the timestamps to obtain the corresponding node time decay weights; For each learning signal node, a quantization calculation is performed based on the corresponding learning behavior feature vector and the node time decay weight to obtain the unit learning efficiency value; all unit learning efficiency values ​​are accumulated to obtain the total learning efficiency value. For each non-learning signal node, quantization is performed based on the corresponding noise behavior feature vector and node time decay weight to obtain the unit deviation efficiency value; all unit deviation efficiency values ​​are accumulated to obtain the total out-of-learning efficiency value. The user's learning signal-to-noise ratio is obtained by performing data smoothing calculations based on the total efficiency value within and outside the learning process.

[0008] As a preferred embodiment of the present invention, the specific steps for analysis based on the learning and cognitive path diagram include: Each learning behavior node in the learning behavior sequence data is taken as a vertex, and directed edges are generated between adjacent vertices according to the timestamp order of the corresponding vertices to obtain the learning cognitive path graph; the learning cognitive transfer cost between any two vertices connected by a directed edge is calculated as the edge weight of the corresponding directed edge; where the learning cognitive transfer cost is directly proportional to the semantic distance between adjacent vertices and inversely proportional to the difference in timestamps between adjacent vertices. Extract semantic content vectors from learning behavior nodes; perform cluster analysis on all learning behavior nodes based on semantic content vectors to obtain several learning concept clusters; in the learning cognitive path graph, identify invalid learning paths based on learning concept clusters, and calculate the learning cognitive return rate based on the identified invalid learning paths.

[0009] As a preferred embodiment of the present invention, invalid learning paths include cross-cluster backtracking paths and intra-cluster oscillation paths. The specific steps for identifying cross-cluster foldback paths include: the cross-cluster foldback path starts from a vertex in the first learning concept cluster, passes through one or more vertices belonging to the second learning concept cluster, and finally returns to any vertex in the first learning concept cluster; wherein, the first learning concept cluster and the second learning concept cluster are any two different learning concept clusters among a number of learning concept clusters. The specific steps for identifying oscillating paths within a cluster include: within a cluster of learning concepts, the oscillating path is a loop path in which the sum of the edge weights of the directed edges between two or more vertices exceeds a preset foldback threshold. The total learning return cost is calculated based on cross-cluster return paths and intra-cluster oscillation paths. The learning return rate is obtained by normalizing the total learning return cost over a preset time period.

[0010] As a preferred embodiment of the present invention, the specific steps for assessing the learning status of adolescent users based on a two-dimensional user learning profile space include: Based on the two-dimensional user learning profile space, four learning behavior pattern quadrants are obtained; the user learning evaluation feedback results are obtained by using the pre-trained learning profile model, learning behavior sequence data, and learning behavior pattern quadrants.

[0011] An assessment and feedback system for adolescent learning data, comprising: The learning data classification module includes a data classification unit; the data classification unit is used to acquire learning behavior sequence data of adolescent users within a preset time period; based on the learning behavior sequence data, the learning signal node data and learning noise node data are obtained; The learning data analysis module includes a learning cognition analysis unit and a learning evaluation unit. The learning cognition analysis unit analyzes learning signal node data and learning noise node data to obtain the user's learning signal-to-noise ratio (SNR). It constructs a learning cognition path graph based on graph theory algorithms and learning behavior sequence data. It then analyzes the learning cognition path graph to obtain the learning cognition return rate. The learning evaluation unit constructs a two-dimensional user learning profile space with the user's learning SNR as the x-axis and the learning cognition return rate as the y-axis. Based on the two-dimensional user learning profile space, it evaluates the learning status of adolescent users and obtains user learning evaluation feedback results.

[0012] The present invention has the following advantages: 1. This invention effectively distinguishes between focused learning and distracted behavior by dividing learning behavior into learning signal nodes and learning noise nodes, ensuring the accuracy of learning process identification; it introduces time decay weights and feature vector calculation mechanisms to uniformly model internal learning efficacy and external learning deviation and construct the learning signal-to-noise ratio, making the evaluation of learning state objective and dynamic; the learning cognitive path graph constructed based on graph theory describes the learning process from the path structure level, which can identify jumping and regression behaviors in learning, making up for the shortcomings of traditional methods that only focus on learning results.

[0013] 2. By defining cross-cluster return paths and intra-cluster oscillation paths and quantifying the learning cognitive return rate, the system reveals ineffective learning paths and enhances the ability to identify inefficient learning behaviors. A two-dimensional learning profile space constructed using the learning signal-to-noise ratio and the learning cognitive return rate maps complex learning behaviors into visual features, ensuring the interpretability and application value of the evaluation results. Combining the learning behavior pattern quadrant with a pre-trained model, the system achieves differentiated evaluation for different types of learners, providing a scientific basis for personalized learning intervention and teaching optimization. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the structure of an evaluation and feedback system for adolescent learning data used in an embodiment of the present invention. Detailed Implementation

[0015] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of this invention.

[0016] Example 1: A method for evaluating and providing feedback on adolescent learning data, comprising the following steps: Acquire learning behavior sequence data of adolescent users within a preset time period; analyze the learning behavior sequence data to obtain learning signal node data and learning noise node data; The data is obtained from the front end of the digital learning platform, capturing all interactive behaviors of adolescent users during specific learning tasks. The information obtained only includes the user's actual operation records. All acquired learning behavior sequence data has been strictly and irreversibly anonymized and de-identified on the client side to ensure that it cannot be reverse-linked to the user's real personal identity information through any technical means. All data is only used for internal algorithm analysis as described in this embodiment, which aims to optimize the learning experience and provide personalized feedback, and will never be used for any third-party sharing or commercial purposes.

[0017] Learning behavior sequence data is a digital footprint that is arranged in chronological order and contains detailed records of every user action, including the user's final submitted answer. It more completely captures the entire process of information flow, including mouse clicks, page jumps, video drags, text input and deletion, and the duration of stay on various learning resource pages.

[0018] The specific steps for analyzing learning behavior sequence data to obtain learning signal node data and learning noise node data include: Obtain user learning task requirements within a preset time period; parse user learning task requirements to obtain several learning task nodes with a sequential order; obtain learning derivative nodes of all learning task nodes based on existing learning task nodes and a preset knowledge association database. A main learning task path is constructed based on several learning task nodes that have a sequential order; learning derivative nodes are then reconstructed based on the learning task nodes corresponding to the main learning task path to obtain a learning task path diagram; historical behavior patterns of adolescent users are obtained; cluster analysis is performed on the historical behavior model to obtain the user's learning deviation tolerance. Semantic parsing is performed on the learning behavior sequence data to obtain several learning behavior nodes; the learning behavior nodes are judged according to the learning task path graph and the user's learning deviation tolerance to obtain learning signal nodes and non-learning signal nodes; the corresponding learning signal node data and learning noise node data are obtained based on the learning signal nodes and non-learning signal nodes.

[0019] At the front end of the digital learning platform, the specific learning tasks that teenage users need to complete within a specified learning period are obtained. These learning task requirements are structured or unstructured texts that describe the learning objectives. For example, mastering the four methods of solving a quadratic equation and completing related exercises after class. The user's learning task requirements clarify the core objectives and scope of this learning activity.

[0020] After acquiring the learning task requirements, the system utilizes its built-in natural language processing and educational knowledge graph parsing engine to perform in-depth analysis of the task requirement text. A learning task node refers to the smallest, indivisible unit of knowledge or skill that must be mastered to complete the entire learning task. The parsing engine identifies the core knowledge points implicit in the task requirements through keyword extraction, dependency parsing, and alignment with the educational knowledge graph, and understands the inherent logical sequence between them. For example, for mastering the four methods of solving quadratic equations, the system will parse out four learning task nodes: completing the square, using the formula, factoring, and graphical methods, and may assign a recommended learning order based on the subject's teaching principles.

[0021] It should be noted that the built-in natural language processing and education knowledge graph parsing engine is a domain-specific language model that is finely tuned based on a general language model (such as GPT) and using a corpus specifically for the education field (covering textbooks, supplementary teaching materials, academic papers, and online course texts for various subjects). The engine pre-acquires an educational text dataset; it then uses this dataset to continuously perform domain-adaptive training on the general language model, enabling it to more accurately understand professional terminology and contextual logic within the education field; finally, it combines the trained general language model with a large educational knowledge graph containing millions of knowledge points and their interrelationships, using entity linking and relation extraction techniques to obtain the final natural language processing and education knowledge graph parsing engine.

[0022] Centered on the parsed learning task nodes, the system performs extended queries in a pre-defined knowledge association database to obtain learning derivative nodes related to these core nodes. This knowledge association database is a complex network that pre-stores a massive amount of knowledge points and uses a graph structure to describe the various relationships between them, such as prerequisite knowledge, similar concepts, application scenarios, and common errors. Learning derivative nodes are peripheral knowledge points in this network that are directly or indirectly connected to the core learning task nodes and are helpful for deepening understanding or expanding applications. For example, for the learning task node of formula method, the system may query discriminant, Vieta's formulas, etc., as important learning derivative nodes.

[0023] After determining all the core learning task nodes and their order, these nodes are connected sequentially as vertices to conceptually construct a clear main learning task path. Correspondingly, all the previously obtained learning derivative nodes are connected to the main learning task path as branches or bypasses based on their relationship with each learning task node on the main path, thus obtaining a learning task path diagram.

[0024] Based on the entire learning behavior sequence data of the adolescent user over a period of time, a personal historical behavior pattern is formed. The historical behavior pattern is a set of data that has been preliminarily processed and feature extracted, which can reflect the user's learning habits and style. It includes, but is not limited to, the user's average answering time under different difficulty tasks, preference for different types of learning resources (videos, texts, interactive experiments), and a series of quantitative indicators such as the average thinking time and operation behavior when encountering difficult problems.

[0025] User learning deviation tolerance is a quantifiable value used to indicate the extent to which a user's behavior deviating from the preset learning path can be accepted as still being in a learning state rather than being useless learning due to distraction. The calculation steps are as follows: First, the system filters all historical deviation behaviors from the current learning task path map from the user's historical behavior patterns and analyzes whether these deviation behaviors have brought positive subsequent effects; that is, whether the user's performance on the relevant main task (such as accuracy or problem-solving speed) has significantly improved after the deviation occurred, thereby calculating the user's deviation behavior positive outcome conversion rate. The system uses the two dimensions of deviation behavior frequency and deviation behavior positive outcome conversion rate as features to perform unsupervised clustering analysis on the historical behavior models of massive users. For example, using the K-Means algorithm, this algorithm automatically divides all users into different clusters through iterative calculation, forming different user profile groups. The system presets a deviation tolerance value for each cluster group, thereby assigning a specific tolerance threshold that reflects the current user's learning style.

[0026] Based on the current learning behavior sequence data of adolescent users, semantic analysis is performed on each behavior in the sequence to obtain learning behavior nodes. Here, a learning behavior node is a point in the user's actual behavior trajectory, which clearly identifies which learning resource the user accessed at a certain moment. For example, if a user clicks to watch a video explaining Vieta's formulas, this behavior will be parsed as a learning behavior node pointing to the knowledge point of Vieta's formulas.

[0027] For each learning behavior node parsed in real time, the system determines whether it can match any node in the pre-constructed learning task path graph (whether it's a learning task node on the main path or a learning derivative node on a branch). If it matches, the learning behavior node is immediately classified as a learning signal node. If it doesn't match, the system initiates a second verification step: deviation tolerance judgment. It calculates the semantic distance between the learning behavior node and the nearest node on the learning task path graph and compares this distance value with the user's unique user learning deviation tolerance value. If the distance is less than or equal to the tolerance, it means that the deviation is within the user's acceptable exploration range, so the behavior node is also classified as a learning signal node. Only when a learning behavior node is neither on the learning task path graph nor exceeds the user's user learning deviation tolerance will it be finally determined as a non-learning signal node (i.e., a noise node).

[0028] Data analysis is performed on the learning signal node data and the learning noise node data to obtain the user's learning signal-to-noise ratio. The specific steps for this data analysis include: Extract the learning behavior feature vector of each learning signal node from the learning signal node data; extract the noise behavior feature vector of each non-learning signal node from the learning noise node data; Multi-dimensional feature extraction is performed on the classified learning signal node data and learning noise node data. For each learning signal node determined to be within the learning path, the system constructs a learning behavior feature vector. This vector is a set of numerical values ​​containing quantitative indicators such as the node access duration, total page scroll distance, and mouse hover frequency in key knowledge areas, to accurately characterize the user's investment depth in that knowledge point. At the same time, for each learning noise node determined to be off the path, the system also constructs a noise behavior feature vector. This vector focuses on recording the user's operational characteristics on non-learning pages, such as the frequency of page jumps and the number of times irrelevant keywords are entered. Through this vectorization, the system transforms abstract learning behavior into structured data with physical meaning that can be processed by a computer, providing standardized input for subsequent performance calculations.

[0029] Obtain the timestamps of all learning signal nodes and non-learning signal nodes; weight the learning signal nodes and non-learning signal nodes according to their contribution based on the timestamps to obtain the corresponding node time decay weights; After obtaining the feature vector, the system synchronously retrieves the precise time point at which each node was accessed in the original sequence, i.e., the timestamp information. To reflect the volatility and immediacy of adolescents' learning status, the system introduces a time decay weighting mechanism. The core logic of this mechanism is that the closer the exploration behavior is to the current time, the greater its reference value for assessing the user's immediate focus, while the influence of earlier behaviors should weaken over time. In specific operation, the system uses a preset exponential decay algorithm to calculate the difference between the timestamp of each node and the benchmark point, using the current assessment time as the benchmark point. This difference is then substituted into the decay function to generate a value between zero and one for each node, i.e., the node time decay weight. In this way, the system ensures that the final signal-to-noise ratio can dynamically reflect the user's recent learning status, rather than simply averaging historical data indiscriminately.

[0030] It should be noted that the specific operational steps of the preset exponential decay algorithm include: the system determines a current time reference point, usually set as the moment when the last learning behavior occurred within the preset time period or the instant when the system executes the evaluation task; for each learning behavior node or noisy behavior node to be processed, the system calculates the time difference between its original timestamp and the current time reference point, which is expressed in seconds or minutes, objectively quantifying how outdated the behavior is; the system introduces a preset decay constant, which is a positive number between zero and one, used to control the rate at which the weights decrease over time; and in the case of... In the numerical processing stage, the system uses the natural constant as the base and the negative value of the product of the time difference and the decay constant as the exponent to perform a power operation, thereby calculating an original decay coefficient for each behavior node that approaches zero as the time difference increases. The system then normalizes this original decay coefficient to make it fall within the range of zero to one, obtaining the final node time decay weight. Through this algorithm step, the system can ensure that the more recent the behavior, which represents the user's current instantaneous state, the closer its weight is to one; while the weight of behaviors that occurred a long time ago decreases exponentially, thus achieving a scientific quantification of the timeliness of learning behaviors.

[0031] For each learning signal node, a quantization calculation is performed based on the corresponding learning behavior feature vector and the node time decay weight to obtain the unit learning efficiency value; all unit learning efficiency values ​​are accumulated to obtain the total learning efficiency value. The unit learning efficacy value refers to the degree to which a learning behavior contributes substantially to achieving a learning goal within a specific time period. The specific calculation steps are as follows: The system normalizes the various indicators (such as duration and interaction depth) in the learning behavior feature vector of the node, and calculates a basic efficacy score using a preset weighted summation model. Then, the basic efficacy score is multiplied by the time decay weight corresponding to the node to obtain the final unit learning efficacy value. After completing the calculation for all individual signal nodes, the system linearly accumulates all unit learning efficacy values ​​generated within a preset time period, ultimately obtaining a value representing the intensity of the adolescent user's total investment in the learning main line within the task time, i.e., the total learning efficacy value. The higher this value, the deeper the user's investment and the newer their state on the effective learning path.

[0032] It should be noted that a set of weighting coefficients for each sub-item in the learning behavior feature vector needs to be pre-defined. These coefficients are empirical values ​​derived from big data analysis that reflect the contribution of different behaviors to learning. For example, a higher weight is assigned to the interaction depth to reflect the level of cognitive engagement, and a moderate weight is assigned to the page scroll distance to reflect the amount of information viewed. The system performs dimensionless normalization processing on the extracted original feature indicators (such as second-level dwell time, pixel-level scroll distance, and number of clicks). A linear mapping algorithm is used to uniformly transform the original values ​​of different units to a standardized numerical range between zero and one to eliminate the interference of differences in physical units on the calculation results. Subsequently... The system then enters the core weighted summation calculation stage, multiplying each standardized feature index value by its corresponding weight allocation coefficient to obtain the independent contribution score of each index. The system linearly accumulates these independent contribution scores and inputs them into a preset S-shaped growth function for non-linear compression to prevent individual extreme abnormal behaviors (such as excessively long dwell time caused by long-term idle time) from causing the score to be too large, thereby outputting a basic performance score within a stable value range that can objectively represent the actual input intensity of the behavior node. Through this series of structured mathematical processing steps, the system achieves a precise transformation from messy raw clickstream data to a unified and comparable learning performance metric.

[0033] For each non-learning signal node, quantization is performed based on the corresponding noise behavior feature vector and node time decay weight to obtain the unit deviation efficiency value; all unit deviation efficiency values ​​are accumulated to obtain the total out-of-learning efficiency value. Meanwhile, the system uses the same logical architecture to quantify non-learning signal nodes to obtain unit deviation efficiency values. These unit deviation efficiency values ​​quantify the cognitive energy consumed by the user in events other than learning. The system multiplies the noise behavior feature vector of each noise node with its corresponding time decay weight to obtain the negative impact of the deviation behavior. Subsequently, the system accumulates all unit deviation efficiency values ​​within the period to form an overall total efficiency value outside of learning. This value objectively records the total energy consumption of the user due to distraction, idleness, or irrelevant operations.

[0034] The user's learning signal-to-noise ratio is obtained by performing data smoothing calculations based on the total efficiency value within and outside the learning process.

[0035] When calculating the ratio, a very small constant is added to the denominator (i.e., the total effectiveness value outside of learning) as a smoothing factor, and the results are scaled and adjusted in conjunction with the preset difficulty coefficient of the current task. The final ratio of signal energy to noise energy is the user's learning signal-to-noise ratio. This indicator is no longer just the proportion of times, but the ratio of depth of investment. It is used to measure the energy distribution relationship between adolescent users' internal learning behaviors and their external activities during the learning process. It is the core vertical axis data for constructing the subsequent learning profile space.

[0036] The smoothing factor is a non-zero positive number between 0.00001 and 0.1, typically 0.001. The main technical purpose of this smoothing factor is to prevent the denominator (i.e., the total learning efficiency value) from becoming zero in specific learning scenarios where adolescent users exhibit extremely high concentration without any deviation from the learning process. This would prevent engineering problems such as abnormal division calculations or numerical divergence in the system. In the specific numerical processing steps, the system uses the accumulated total learning efficiency value as a base and performs arithmetic addition with the preset smoothing factor to obtain a corrected denominator value that is always greater than zero. The system then uses the total learning efficiency value as the numerator and the corrected denominator value as the denominator to perform a ratio calculation. Based on this, a task difficulty decay constant reflecting the challenge intensity of the current learning task is used to non-linearly scale and adjust the obtained ratio.

[0037] A learning cognitive path graph is constructed based on graph theory algorithms and learning behavior sequence data; the learning cognitive return rate is obtained by analyzing the learning cognitive path graph. The specific steps for analysis based on the learning and cognitive path map include: Each learning behavior node in the learning behavior sequence data is taken as a vertex, and directed edges are generated between adjacent vertices according to the timestamp order of the corresponding vertices to obtain the learning cognitive path graph; the learning cognitive transfer cost between any two vertices connected by a directed edge is calculated as the edge weight of the corresponding directed edge; where the learning cognitive transfer cost is directly proportional to the semantic distance between adjacent vertices and inversely proportional to the difference in timestamps between adjacent vertices. Each parsed learning behavior node is transformed into a vertex in graph theory, and these vertices together form the basis for subsequent path analysis. Then, based on the timestamps of these behavior nodes in the original learning behavior sequence data, the system creates a directed edge between each adjacent vertex in strict chronological order, with the edge pointing from the previous behavior to the next behavior. In this way, the system successfully transforms one-dimensional, linear time series data into a directed graph structure that can intuitively reflect the user's thought process flow path. This graph structure is the learning cognitive path graph.

[0038] Edge weights are defined as the learning and cognitive transfer cost, designed to quantify the time and effort a user expends when jumping from one knowledge point to another. The calculation steps are as follows: First, the system uses a natural language processing model to extract the semantic vectors of the knowledge content represented by the starting and ending vertices of the edge, and calculates the cosine distance between these two vectors to obtain a semantic distance value reflecting the degree of correlation between the two contents. Second, the system calculates the difference between the timestamps corresponding to these two vertices, i.e., the time interval between the user's two actions. The semantic distance value and the reciprocal of the time interval are then weighted and fused to obtain the final learning and cognitive transfer cost. The design logic of this cost value is that if the content span between two actions is large (long semantic distance) and the thinking time is short (small time interval), it means a drastic leap in thinking, and its transfer cost is high. By assigning a cost to each path, the entire learning and cognitive path graph is upgraded from a simple trajectory graph to a weighted graph containing energy consumption information.

[0039] Extract semantic content vectors from learning behavior nodes; perform cluster analysis on all learning behavior nodes based on semantic content vectors to obtain several learning concept clusters; in the learning cognitive path graph, identify invalid learning paths based on learning concept clusters, and calculate the learning cognitive return rate based on the identified invalid learning paths.

[0040] This is achieved by calling a pre-trained, education-oriented deep language model (e.g., a BERT model fine-tuned with educational literature). For each learning behavior node, the system inputs all the text content of the learning resource it represents (such as web page text, video subtitles, and question descriptions) into the model. The model, through its complex internal neural network structure, transforms this unstructured text information into a high-dimensional mathematical vector containing rich semantic information, namely, a semantic content vector.

[0041] After obtaining the semantic content vectors of all learning behavior nodes, the system will launch an unsupervised clustering analysis algorithm, such as the density-based DBSCAN algorithm, to automatically classify these vectors. Here, the learning concept cluster refers to a set of learning behavior nodes that are highly similar in semantic content and have been focused on by the user during the learning process. It objectively reflects a specific topic or conceptual category that the user focuses on thinking about over a period of time. The DBSCAN algorithm can effectively group semantically closely connected nodes into the same cluster by automatically finding high-density regions in the vector space, without having to pre-set the number of clusters.

[0042] Ineffective learning paths include cross-cluster backtracking paths and intra-cluster oscillation paths; The specific steps for identifying cross-cluster foldback paths include: the cross-cluster foldback path starts from a vertex in the first learning concept cluster, passes through one or more vertices belonging to the second learning concept cluster, and finally returns to any vertex in the first learning concept cluster; wherein, the first learning concept cluster and the second learning concept cluster are any two different learning concept clusters among a number of learning concept clusters. The specific steps for identifying oscillating paths within a cluster include: within a cluster of learning concepts, the oscillating path is a loop path in which the sum of the edge weights of the directed edges between two or more vertices exceeds a preset foldback threshold. The total learning return cost is calculated based on cross-cluster return paths and intra-cluster oscillation paths. The learning return rate is obtained by normalizing the total learning return cost over a preset time period.

[0043] In the specific invalid path identification process, the system first searches for a pattern called a cross-cluster backtracking path. The identification algorithm is as follows: the system traverses each path in the learning cognitive path graph and checks the concept cluster labels to which the vertex sequence on the path belongs. Once it finds that the starting point and ending point of a path have the same cluster label (for example, both belonging to the cluster of discriminant of the roots of a quadratic equation), while at least one vertex in the middle of the path has a different cluster label (for example, belonging to the cluster of symmetry of function graphs), the system marks this path, which starts from one learning concept cluster, goes to another learning concept cluster, and then returns to the original learning concept cluster, as a cross-cluster backtracking event. This behavior pattern objectively shows that after exploring a new concept, users have to perform a retrospective review or confirmation because their understanding of the original concept is not solid enough. This is a typical manifestation of cognitive incoherence.

[0044] In addition to cross-cluster backtracking, the system also identifies another type of invalid path occurring within a single learning concept cluster: intra-cluster oscillation paths. Its identification algorithm focuses on finding local loop structures in the graph. Specifically, within each learning concept cluster, the system checks if there exists a path whose starting and ending points are the same vertex. If such a loop path is found, the system further calculates the sum of the edge weights (i.e., learning cognitive transfer costs) of all directed edges on that loop path. Only when this total weight exceeds a preset backtracking threshold is the loop confirmed as an intra-cluster oscillation event. This threshold is set to filter out normal, brief, and low-cost back-and-forth viewing behaviors, and only capture learning behaviors where users hesitate and compare repeatedly for extended periods between several highly related knowledge points at high cost—a strong signal of local cognitive obstacles.

[0045] After identifying all cross-cluster backtracking paths and intra-cluster oscillation paths, the system performs a unified quantitative evaluation of these marked invalid paths. For each invalid path, the system sums the edge weights (i.e., learning cognitive transfer costs) of all edges on the path to obtain the single cost of that invalid backtracking event. Subsequently, the system sums the single costs of all invalid backtracking events that occur within a preset time period to obtain an overall, quantitative total learning backtracking cost. This total cost value is no longer a simple number of backtracking attempts, but a comprehensive cognitive waste indicator that integrates time consumption, cognitive span, and path complexity. Its magnitude is proportional to the ineffective effort that users expend in their mental activities during the learning process.

[0046] The specific operation involves calculating the ratio of the total cost of learning return to the user's total learning time or total number of steps within the preset time period. For example, even if a student who studies for 10 hours and a student who studies for 1 hour have the same total cost of return, the former's return rate will be much lower than the latter's. The final normalized ratio is the learning cognitive return rate, which reflects the clarity, logic, and efficiency of the thinking path of adolescent users during the learning process. It is the key horizontal axis data for constructing the subsequent two-dimensional learning profile space.

[0047] It should be noted that when performing weighted fusion to calculate the cognitive transfer cost of learning, the system sets dynamically adjusted weight coefficients for the semantic distance value and the reciprocal of the time interval. These weight coefficients are not fixed but are related to the user's current learning stage and task type. In the initial exploration stage of a learning task, the system assigns a higher weight to semantic distance to encourage users to explore knowledge extensively and across a wide range. In the later convergence or review stage of the task, the system increases the weight of the reciprocal of the time interval to encourage users to make quick and efficient associations and switches between related concepts. In addition, for open-ended tasks that require creative thinking, the system also increases the weight of semantic distance. For computational tasks that require proficiency and speed, the system focuses on the numerical content of the time interval. The initial values ​​of these dynamic weight coefficients are based on a preset educational psychology model and are continuously optimized and adjusted by machine learning algorithms based on massive amounts of user behavior data during system operation to ensure that the calculated cognitive transfer cost of learning most accurately reflects the true measurement of learning cost in different contexts.

[0048] A two-dimensional user learning profile space is constructed with the user learning signal-to-noise ratio as the x-axis and the learning cognitive return rate as the y-axis. Based on this two-dimensional user learning profile space, the learning status of adolescent users is assessed, and user learning assessment feedback results are obtained. The specific steps for assessing the learning status of adolescent users based on the two-dimensional user learning profile space include: Based on the two-dimensional user learning profile space, four learning behavior pattern quadrants are obtained; the user learning evaluation feedback results are obtained by using the pre-trained learning profile model, learning behavior sequence data, and learning behavior pattern quadrants.

[0049] Before constructing a two-dimensional user learning profile space, the system first needs to define and normalize the basic quantitative indicators of the horizontal and vertical axes. Among these, the user learning signal-to-noise ratio (SNR) refers to the ratio of the time a learner spends in an effective learning interaction state to the time spent engaging in irrelevant distracting behaviors within a specific learning cycle, reflecting the user's level of focus. The learning cognitive return rate refers to the frequency with which learners revisit or repeat previously learned content while advancing a learning task, reflecting the fluency of knowledge acquisition and cognitive resistance. To ensure the scientific validity of the coordinate system, the system uses massive amounts of historically collected adolescent learning behavior data as a baseline to extract the statistical distribution characteristics of these two indicators. Subsequently, the system employs a max-min linear scaling algorithm... The system performs data standardization, specifically by extracting the maximum and minimum values ​​of each indicator from the reference baseline. The minimum value is then subtracted from the original calculated value, and the result is divided by the difference between the maximum and minimum values. This maps the data to a dimensionless standardized percentage score between zero and one hundred. This process effectively eliminates dimensional interference caused by differences in task difficulty across disciplines or individual time investment. Furthermore, the system dynamically anchors the origin of the coordinate system to the statistical median level of the reference baseline and introduces a feedback verification mechanism during operation. It periodically calculates the root mean square error between the system output and the actual situation using manually evaluated samples. If the error exceeds a preset threshold, the system automatically calls back and updates the baseline parameters in the normalization formula, thereby ensuring continuous evaluation accuracy.

[0050] Meanwhile, to avoid assessment distortion caused by widespread fluctuations in student performance over a certain period, the system introduces an absolute efficacy benchmark preset by the expert system as a second reference frame in the two-dimensional user learning profile space, which works in conjunction with the dynamically fluctuating origin of the group median. This absolute efficacy benchmark is a fixed parameter calculated based on educational psychology norms and a massive amount of historical best learning samples, representing the minimum ideal signal-to-noise ratio and the maximum tolerable return rate that adolescents should achieve under specific subject tasks. When performing quadrant division and score mapping, the system adopts a dual-weight fusion algorithm to calculate the user's actual coordinates relative to the origin of the dynamic median and the absolute efficacy benchmark, respectively. When there is a general decline in learning performance across the entire domain, although the user's dynamic relative position may shift towards the higher-scoring quadrant, the system will perform logical hard constraint correction on the final output learning efficacy score by detecting the change in the Euclidean distance between the user and the absolute efficacy benchmark. Specifically, if a user's absolute physical distance does not reach the preset passing threshold, even if the user is in a leading position in the current group, the system will retain an absolute performance warning label in the generated structured data vector and limit the automatic increase of the user's performance score.

[0051] After establishing the dynamic median as the origin of the intersection of the horizontal and vertical axes, the entire two-dimensional user learning profile space is naturally divided into four quadrants of learning behavior patterns, and each quadrant corresponds to an objective data space of a teenager's learning state. Specifically, the first quadrant represents high signal-to-noise ratio and high return rate, the second quadrant represents low signal-to-noise ratio and high return rate, the third quadrant represents low signal-to-noise ratio and low return rate, and the fourth quadrant represents high signal-to-noise ratio and low return rate.

[0052] Based on the learning behavior pattern quadrant, the system extracts the objective accuracy rate of answering questions from the learning behavior sequence data, and uses it as the core dimension to measure the knowledge mastery result. At the same time, it obtains the user's standardized coordinates (x, y) in the profile space, which are the quantitative scores of the learning signal-to-noise ratio and the learning cognitive return rate. Then, the pre-trained learning profile model uses a spatial geometric mapping algorithm to calculate the Euclidean distance between the user's current coordinate point and the ideal state anchor point in the fourth quadrant, and substitutes it into a preset monotonically decreasing linear mapping function to transform it into a behavioral performance score that reflects the quality of the learning process.

[0053] During this process, the model will automatically adjust the weight allocation logic according to the user's current learning behavior pattern quadrant: if the user is in the high signal-to-noise ratio quadrant, the model will increase the weight of the accuracy of answering questions in the comprehensive evaluation to verify whether high concentration is effectively converted into knowledge output; if the user is in the high return rate quadrant, the model will introduce a difficulty coefficient reflecting the intensity of the task challenge as a correction operator, and by comparing the user's return frequency with the norm level under the same difficulty task, accurately determine whether it belongs to deep thinking or cognitive impairment.

[0054] Finally, the model nonlinearly fuses the behavioral performance score, corrected by quadrant features, with the accuracy rate of answering questions to calculate a quantitative comprehensive score of the student's learning status. This quantitative comprehensive score not only includes a total score out of 100, but also includes targeted environmental adjustment parameters matched from a pre-set strategy library based on the score range and coordinate movement trajectory. For example, for users with low comprehensive scores and in the low signal-to-noise ratio quadrant, the system will output a structured data vector containing enhanced visual guidance or reduced granularity of knowledge points. Ultimately, the system outputs this structured vector, which integrates the quantitative comprehensive score, process sub-indicators, and intervention strategy suggestions, as the user learning assessment feedback result, realizing the automated transformation from objective behavioral data to in-depth quantitative assessment.

[0055] It should be noted that the pre-trained learning profile model is a composite artificial intelligence model trained on massive amounts of adolescent learning trajectory data. It integrates multiple functions such as data feature extraction, spatial location analysis, and behavioral tendency prediction. The core function of this model is to transform the underlying discrete learning interaction behaviors into cognitive feature representations in a high-dimensional space, thereby identifying the deep correlation between different behavioral patterns and learning outcomes. In actual operation, this model acts as an intelligent evaluation center, which can receive the user's spatial coordinate data in real time and combine it with educational psychology knowledge graphs to automatically calibrate and evaluate the user's level of focus, coherence of thought, and knowledge absorption efficiency in multiple dimensions, ultimately outputting quantitative data that reflects the user's true cognitive state.

[0056] The spatial geometric mapping algorithm's operation process refers to the mathematical processing steps that convert the user's positional characteristics in a two-dimensional coordinate system into a computable distance. Its specific execution logic is as follows: First, the algorithm obtains the user's real-time coordinates in a standardized profile space, denoted as the horizontal axis learning signal-to-noise ratio and the vertical axis learning cognitive return rate. The algorithm then performs spatial positioning to determine the geometric orientation of this coordinate point relative to the origin and the boundaries of each quadrant. Finally, the algorithm calls the Euclidean distance formula, calculating the sum of the squares of the horizontal and vertical displacements between the user's current coordinates and the target reference point, and then performing a square root operation to obtain a spatial physical length that characterizes the degree to which the user's current state deviates from the ideal state. This algorithm transforms the abstract difference in learning states into a numerical value with rigorous geometric meaning, providing a precise mathematical basis for subsequent quantitative scoring.

[0057] The ideal state anchor point in the fourth quadrant refers to the extreme reference coordinate point representing the theoretically efficient learning performance in the two-dimensional user learning profile space. In the standardized percentage system space, this anchor point is usually set at the intersection of the maximum value of the user learning signal-to-noise ratio on the horizontal axis and the minimum value of the learning cognitive return rate on the vertical axis. By setting this fixed or semi-fixed spatial benchmark, the system provides a unified benchmark for all users at different learning stages, so that any coordinate movement can be quantified as the degree of approach or deviation from the ideal state.

[0058] The pre-defined monotonically decreasing linear mapping function is a mathematical operator used to inversely scale spatial distance values ​​to a percentage score. Specifically, the performance score is expressed as the difference between 100 multiplied by 1 and the ratio of the current distance to the maximum theoretical spatial distance. In this function, the independent variable is the spatial distance between the user's coordinates and the ideal anchor point, while the dependent variable is the final behavioral performance score. Because the function is monotonically decreasing, this means that as the distance between the user's coordinates and the ideal anchor point decreases, the calculated score increases linearly; conversely, if the distance increases, the score decreases proportionally. This function design ensures consistency between spatial geometry and evaluation logic, resulting in users closer to the ideal state receiving higher scores in the quantitative results, thus achieving a smooth conversion from spatial location to quality evaluation. The maximum theoretical spatial distance refers to the geometric straight-line distance from the origin (or an extreme vertex) to the extreme point on the opposite diagonal of a standardized square (or high-dimensional hypercube) image space, representing the maximum possible range of state differences in the system.

[0059] The conversion process of behavioral efficacy scores involves the specific operational steps of using the aforementioned mapping function to ultimately confirm spatial physical indicators as cognitive quality scores. The system extracts the Euclidean distance value output by the spatial geometric mapping algorithm and calculates a standardized ratio by combining it with the maximum diagonal length of the coordinate space. By performing a linear transformation of a preset function, this ratio is mapped to the range of zero to one hundred, yielding a preliminary behavioral efficacy score. The system then performs logical verification on this score based on the user's quadrant attribute. For example, for users in the first quadrant, the system identifies whether their learning backtracking behavior constitutes benign deep confirmation and fine-tunes the mapping result accordingly. This ensures that the final converted behavioral efficacy score accurately eliminates interference factors caused by task difficulty or individual cautiousness, truly reflecting the user's input quality and logical efficacy during the learning process.

[0060] In this embodiment, all computation and simulation processes are implemented based on a cloud-deployed hybrid artificial intelligence architecture. The underlying core consists of a deep attention mechanism model for semantic understanding and a heuristic graph search algorithm for path analysis. The former is responsible for encoding unstructured text information into computer-computable high-dimensional numerical vectors, while the latter is responsible for establishing logical connections between discrete behavioral nodes and assigning specific edge weights based on semantic relationships. To address the computational conflicts and technical incomparability issues that may arise from different data units, the system implements a rigorous multi-dimensional data standardization process as a solution. The specific execution steps are as follows: First, the system performs metadata annotation and classification on all input feature data, clearly distinguishing which are continuous physical quantities (e.g., duration in seconds), which are discrete counts (e.g., click frequency), and which are ratio data (e.g., accuracy percentage). Then, in the normalization adaptation stage, the system intelligently selects the appropriate numerical transformation operator based on the statistical distribution characteristics of each type of data. For time data that follows an approximately uniform distribution, a linear range standardization algorithm is used. The original value is subtracted from the minimum value and then divided by the range, thus mapping it to a closed interval between zero and one. For interaction frequency data that follows a skewed distribution and has abnormally large values, a robust quantile mapping method is used to weaken the interference of extreme values ​​and flatten the distribution curve. On this basis, the system further performs a feature dimension alignment operation, converting all standardized feature values ​​of different dimensions into standard scores relative to the global historical mean, so that data of different properties have the same measurement and physical meaning on the same virtual number axis. Finally, these standardized vectors processed with a unified scale are input into the backend fully connected fusion computing module, and the final evaluation index is generated through nonlinear combination operations using an adaptive weight allocation mechanism. This completely eliminates the numerical bias caused by differences in measurement units, ensuring the mathematical validity of key parameters such as user learning signal-to-noise ratio and learning cognitive return rate, as well as the objectivity and fairness of cross-group evaluation results.

[0061] In this embodiment, the preset data is configured using a hierarchical approach that combines expert experience initialization with dynamic calibration based on big data. The specific steps are as follows: First, based on educational subject standards and cognitive psychology theories, domain experts initially calibrate basic parameters such as decay constants, clustering thresholds, and intervention strategy weights and store them in the rule engine. Then, a training set is constructed using the large-scale desensitized historical behavioral data accumulated by the platform. The initial parameters are fitted and optimized using a machine learning regression model to obtain the optimal benchmark value suitable for a specific learning stage. During the operation phase after system deployment, a feedback closed-loop mechanism is introduced to periodically collect deviation data between evaluation results and actual learning outcomes. Adaptive adjustment algorithms are used to continuously update and iterate the preset parameters, thereby ensuring that the preset data can continuously adapt to the ever-changing group learning characteristics and individual development needs.

[0062] In practical applications, the embodiments of this invention were tested against a pre-defined comparative learning platform. For example, in a 60-minute online learning test on quadratic equations, the pre-defined comparative learning platform concluded that the user was in a good and focused state based solely on the correct answer rate (85%) and total learning time (58 minutes). However, the embodiments of this invention, by constructing a two-dimensional user learning profile space, found that the user's standardized learning signal-to-noise ratio score was only 42 points. This was because the system identified a large number of non-learning signal nodes, such as frequent irrelevant page jumps and ineffective hovering for up to 12 minutes. These behaviors were incorrectly counted as effective learning time in traditional time statistics. Simultaneously, in the efficiency evaluation of the cognitive path, the pre-defined platform... The existing comparative learning platform cannot detect user thought fluctuations. However, the learning cognitive path graph constructed using graph theory algorithms in this invention shows that the user's cognitive backtracking rate is as high as 68 points, placing it in the second quadrant (low signal-to-noise ratio, high backtracking rate). Specific comparative data shows that the system identified three cross-cluster backtracking paths, indicating that the user repeatedly backtracked due to a weak grasp of basic discriminant concepts while attempting to solve the problem. Furthermore, four instances of intra-cluster oscillations occurred within the factorization logic, where the sum of edge weights exceeded the threshold. The total cost of these ineffective backtrackings accounted for nearly 70% of the total cognitive cost. In contrast, if the evaluation results of the pre-set comparative learning platform are used, the system will maintain the original difficulty level. However, the composite learning profile model of this invention, based on structured... The quantified performance score in the data vector automatically triggered a visual guidance intervention strategy, setting the intervention weight to 0.8 to guide users back to basic knowledge points. This proportion clearly demonstrates that this invention, by quantifying the signal-to-noise ratio and return rate, reveals the inefficient learning essence masked by traditional high-scoring data, ensuring the logical consistency and scientific nature of the evaluation feedback. In this embodiment, the system of this invention and a pre-set traditional comparative learning platform are tested in parallel under completely identical hardware environments and task backgrounds. Both platforms are connected to the same sample database of adolescent users and are uniformly configured with learning resource sets containing standard subject tasks such as solving quadratic equations and the properties of function graphs, ensuring that network bandwidth, terminal response speed, and initial learning tasks are within acceptable limits during the experiment. The goal is complete consistency. The only difference between the system of this invention and the comparative learning platform lies in the technical solution of the evaluation feedback: the comparative learning platform only uses a traditional statistical model based on the correct answer rate and total learning time for outcome evaluation; while the system of this invention applies a core architecture consisting of a learning data classification module and a learning data analysis module, and performs in-depth evaluation within a dynamically anchored two-dimensional user learning profile space by calculating the quantified user learning signal-to-noise ratio and learning cognitive return rate. Throughout the entire testing period, except for the internal parameter configuration of the evaluation algorithm, the interface interaction logic, the content coverage of the knowledge association database, and the initial benchmark of user deviation tolerance of the two platforms are kept synchronized to eliminate the interference of uncontrolled variables on the accuracy of the evaluation results.

[0063] Example 2, an evaluation and feedback system for adolescent learning data, see [link to example]. Figure 1 As shown, it includes: The learning data classification module includes a data classification unit; the data classification unit is used to acquire learning behavior sequence data of adolescent users within a preset time period; based on the learning behavior sequence data, the learning signal node data and learning noise node data are obtained; The learning data analysis module includes a learning cognition analysis unit and a learning evaluation unit. The learning cognition analysis unit analyzes learning signal node data and learning noise node data to obtain the user's learning signal-to-noise ratio (SNR). It constructs a learning cognition path graph based on graph theory algorithms and learning behavior sequence data. It then analyzes the learning cognition path graph to obtain the learning cognition return rate. The learning evaluation unit constructs a two-dimensional user learning profile space with the user's learning SNR as the x-axis and the learning cognition return rate as the y-axis. Based on the two-dimensional user learning profile space, it evaluates the learning status of adolescent users and obtains user learning evaluation feedback results.

[0064] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims. Parts not described in detail in this specification are prior art known to those skilled in the art.

Claims

1. A method for evaluating and providing feedback on adolescent learning data, characterized in that, Includes the following steps: Acquire learning behavior sequence data of adolescent users within a preset time period; Based on the analysis of the learning behavior sequence data, learning signal node data and learning noise node data are obtained; Data analysis is performed on the learning signal node data and the learning noise node data to obtain the user's learning signal-to-noise ratio; A learning cognitive path graph is constructed based on graph theory algorithms and learning behavior sequence data; the learning cognitive return rate is obtained by analyzing the learning cognitive path graph. A two-dimensional user learning profile space is constructed with the user learning signal-to-noise ratio as the x-axis and the learning cognitive return rate as the y-axis. Based on the two-dimensional user learning profile space, the learning status of adolescent users is evaluated, and user learning evaluation feedback results are obtained.

2. The method for evaluating and providing feedback on adolescent learning data according to claim 1, characterized in that, The specific steps for analyzing learning behavior sequence data to obtain learning signal node data and learning noise node data include: Obtain user learning task requirements within a preset time period; parse user learning task requirements to obtain several learning task nodes with a sequential order; obtain learning derivative nodes of all learning task nodes based on existing learning task nodes and a preset knowledge association database. A main learning task path is constructed based on several learning task nodes that have a sequential order; learning derivative nodes are then reconstructed based on the learning task nodes corresponding to the main learning task path to obtain a learning task path diagram; historical behavior patterns of adolescent users are obtained; cluster analysis is performed on the historical behavior model to obtain the user's learning deviation tolerance. Semantic parsing is performed on the learning behavior sequence data to obtain several learning behavior nodes; the learning behavior nodes are judged according to the learning task path graph and the user's learning deviation tolerance to obtain learning signal nodes and non-learning signal nodes; the corresponding learning signal node data and learning noise node data are obtained based on the learning signal nodes and non-learning signal nodes.

3. The method for evaluating and providing feedback on adolescent learning data according to claim 2, characterized in that, The specific steps for data analysis of learned signal node data and learned noise node data include: Extract the learning behavior feature vector of each learning signal node from the learning signal node data; extract the noise behavior feature vector of each non-learning signal node from the learning noise node data; Obtain the timestamps of all learning signal nodes and non-learning signal nodes; weight the learning signal nodes and non-learning signal nodes according to their contribution based on the timestamps to obtain the corresponding node time decay weights; For each learning signal node, a quantization calculation is performed based on the corresponding learning behavior feature vector and the node time decay weight to obtain the unit learning efficiency value; all unit learning efficiency values ​​are accumulated to obtain the total learning efficiency value. For each non-learning signal node, quantization is performed based on the corresponding noise behavior feature vector and node time decay weight to obtain the unit deviation efficiency value; all unit deviation efficiency values ​​are accumulated to obtain the total out-of-learning efficiency value. The user's learning signal-to-noise ratio is obtained by performing data smoothing calculations based on the total efficiency value within and outside the learning process.

4. The method for evaluating and providing feedback on adolescent learning data according to claim 3, characterized in that, The specific steps for analysis based on the learning and cognitive path map include: Each learning behavior node in the learning behavior sequence data is taken as a vertex, and directed edges are generated between adjacent vertices according to the timestamp order of the corresponding vertices to obtain the learning cognitive path graph; the learning cognitive transfer cost between any two vertices connected by a directed edge is calculated as the edge weight of the corresponding directed edge; where the learning cognitive transfer cost is directly proportional to the semantic distance between adjacent vertices and inversely proportional to the difference in timestamps between adjacent vertices. Extract semantic content vectors from learning behavior nodes; perform cluster analysis on all learning behavior nodes based on semantic content vectors to obtain several learning concept clusters; in the learning cognitive path graph, identify invalid learning paths based on learning concept clusters, and calculate the learning cognitive return rate based on the identified invalid learning paths.

5. The method for evaluating and providing feedback on adolescent learning data according to claim 4, characterized in that, Ineffective learning paths include cross-cluster backtracking paths and intra-cluster oscillation paths; The specific steps for identifying cross-cluster foldback paths include: the cross-cluster foldback path starts from a vertex in the first learning concept cluster, passes through one or more vertices belonging to the second learning concept cluster, and finally returns to any vertex in the first learning concept cluster; wherein, the first learning concept cluster and the second learning concept cluster are any two different learning concept clusters among a number of learning concept clusters. The specific steps for identifying oscillating paths within a cluster include: within a cluster of learning concepts, the oscillating path is a loop path in which the sum of the edge weights of the directed edges between two or more vertices exceeds a preset foldback threshold. The total learning return cost is calculated based on cross-cluster return paths and intra-cluster oscillation paths. The learning return rate is obtained by normalizing the total learning return cost over a preset time period.

6. The method for evaluating and providing feedback on adolescent learning data according to claim 5, characterized in that, The specific steps for assessing the learning status of adolescent users based on a two-dimensional user learning profile space include: Based on the two-dimensional user learning profile space, four learning behavior pattern quadrants are obtained; the user learning evaluation feedback results are obtained by using the pre-trained learning profile model, learning behavior sequence data, and learning behavior pattern quadrants.

7. An evaluation and feedback system for adolescent learning data, characterized in that, The system applies an assessment and feedback method for adolescent learning data as described in any one of claims 1-6, comprising: The learning data classification module includes a data classification unit; the data classification unit is used to acquire learning behavior sequence data of adolescent users within a preset time period; based on the learning behavior sequence data, learning signal node data and learning noise node data are obtained; The learning data analysis module includes a learning cognition analysis unit and a learning evaluation unit. The learning cognition analysis unit analyzes learning signal node data and learning noise node data to obtain the user's learning signal-to-noise ratio (SNR). It constructs a learning cognition path graph based on graph theory algorithms and learning behavior sequence data. It then analyzes the learning cognition path graph to obtain the learning cognition return rate. The learning evaluation unit constructs a two-dimensional user learning profile space with the user's learning SNR as the x-axis and the learning cognition return rate as the y-axis. Based on the two-dimensional user learning profile space, it evaluates the learning status of adolescent users and obtains user learning evaluation feedback results.

Images