User learning quality evaluation method and system based on education platform data analysis

Abstract

The invention discloses a user learning quality evaluation method and system based on education platform data analysis, and relates to the technical field of data analysis, and the method comprises the steps: constructing a dynamic knowledge graph sequence based on the overall learning process of a user on an education platform, analyzing the abnormal degree of a user learning path, and calculating a path behavior abnormal value; a path embedding vector is extracted through a neural network graph embedding method, the stability of a learning track is analyzed, and a convergence abnormal value is calculated; and judging whether the user learning quality evaluation result has fraud or not by combining the two. And if the fraud exists, performing exception identification on all paths, distinguishing an abnormal path from a normal path, and re-evaluating the learning quality based on the normal path. According to the method, fraudulent behaviors are effectively recognized and corrected through the dynamic knowledge graph, the deviation of learning quality evaluation is reduced, the judgment accuracy of an education platform on the real learning condition of a user is improved, and misguidance on the teaching intervention effect is reduced.

Description

Technical Field

[0001] The present invention relates to the field of data analysis technology, and in particular to a method and system for evaluating user learning quality based on education platform data analysis. Background Art

[0002] With the rapid development of educational informatization, there are more and more educational platforms on the market. Educational platforms can not only help users learn, but also analyze users' behavioral data to evaluate and quantify the quality of users' learning. Common user learning quality assessment methods are usually based on comprehensive analysis of indicators such as users' click behavior, video viewing time, homework completion, test scores, etc. to evaluate users' learning quality; they are used to respond to educational suggestions and educational guidance based on the assessed learning quality, thereby achieving better and more timely teaching interventions.

[0003] However, most existing assessment methods assume that user behavior data is authentic and reliable, ignoring interference from abnormal behaviors such as human intervention, malicious cheating, and scripted automated operations. When platform users engage in fraudulent behaviors such as false clicks, skipping learning, and repeatedly cheating, their learning behavior data is still recorded as normal by the platform. This makes it impossible for computers to accurately identify users' true learning status, resulting in an overestimation or distortion of overall learning quality assessment results. Summary of the Invention

[0004] The purpose of the present invention is to solve the above-mentioned problems and provide a user learning quality assessment method and system based on education platform data analysis.

[0005] In a first aspect of the present invention, a method for evaluating user learning quality based on education platform data analysis is first proposed, the method comprising:

[0006] Build a dynamic knowledge graph sequence based on the user's overall learning process and learning content, analyze the abnormality of the user's learning path based on the dynamic knowledge graph sequence, and calculate the path behavior abnormality value;

[0007] The neural network-based graph embedding method is used to extract the path embedding vectors corresponding to the paths in each knowledge graph. The stability of the user's learning trajectory is analyzed based on the embedding vectors corresponding to the paths in each knowledge graph, and the convergence outlier is calculated.

[0008] Analyze the evaluation results of user learning quality based on path behavior anomalies and convergence anomalies to determine whether there is fraud in the current evaluation results of user learning quality;

[0009] If there is fraud in the evaluation results of the current user's learning quality, the paths of all knowledge graph sequences are analyzed to identify abnormal paths and normal paths, and the overall learning quality of the user is re-evaluated based on the abnormal paths and normal paths.

[0010] Optionally, constructing a dynamic knowledge graph sequence based on the user's overall learning process and learning content includes:

[0011] Record the total time users spend studying on the education platform and divide the total time into several sub-time periods;

[0012] For each sub-time period, extract the learning content of the users in the corresponding sub-time period as a node of the knowledge graph of the corresponding sub-time period; the learning content includes knowledge point teaching and viewing content, knowledge point practice content, and knowledge point test content;

[0013] For several nodes in the knowledge graph, based on the learning time sequence, directed edges between adjacent nodes are added, and the weights of directed edges between adjacent nodes are determined based on the preset standard learning path to obtain a complete learning path;

[0014] The knowledge graph containing the complete learning path is used as the knowledge graph of the corresponding sub-time period; a dynamic knowledge graph sequence is constructed based on the knowledge graph of each sub-time period.

[0015] Optionally, the weights of directed edges between adjacent nodes may be determined based on a preset standard learning path:

[0016] Extract the user's initial learning content on the education platform as the starting node, match it to the preset standard learning path, and determine the number of jump levels between all adjacent nodes in the knowledge graph compared to the preset standard learning path;

[0017] Calculate the weight of the directed edge between adjacent nodes. The calculation formula is: w i,i+1 =max(0,1-β×k), where w i,j is the weight of the directed edge between adjacent nodes i and i+1; k is the number of jump layers between adjacent nodes i and i+1 compared to the preset standard learning path; β is the preset penalty coefficient, which is 0.1-0.2.

[0018] Optionally, the abnormality of the user's learning path is analyzed according to the dynamic knowledge graph sequence, and the steps of calculating the abnormal value of the path behavior are as follows:

[0019] Extract the learning path from the knowledge graph sequence corresponding to each sub-time period to obtain the set of all paths P t ={P1,P2,P3....P k}, where k represents the total number of sub-time periods;

[0020] The path corresponding to each knowledge graph is P j = <v1,v2....v m >, where v n Indicates the order of the nodes, indicating the nth node in the path;

[0021] Calculate the absolute difference between the edge weights of consecutive nodes in each path and the standard edge weight, add up all the absolute differences, and use them as the jump value of each path; the standard edge weight value is 1;

[0022] The fluctuation degree of the edge weights between consecutive nodes in each path is defined as the transition tension. The transition tension of each path is calculated using the following formula: Where, T(P j ) is the transition tension of the j-th path; is the mean edge weight of the j-th path;

[0023] The abnormality of each path is calculated based on the jump value and transition tension of each path. The calculation formula is: D(P j )=S(P j )×T(P j ), where D(P j ) is the abnormality of the jth path, S(P j ) is the jump value of the j-th path;

[0024] Calculate the abnormality difference between adjacent paths, measure the abnormality of the overall path over time, and calculate the overall path behavior abnormality. The calculation formula is: Where YU is the path behavior anomaly value.

[0025] Optionally, the stability of the user's learning trajectory is analyzed based on the embedding vectors corresponding to the paths in each knowledge graph, and the steps for calculating the convergence outlier are as follows:

[0026] The embedding vector set z corresponding to the path in the knowledge graph t ={z1,z2,z3....z k}, where k represents the total number of sub-time periods;

[0027] Calculate the difference vector V between the path embeddings of two adjacent time periods t , which represents the local direction of path evolution:

[0028] Calculate the embedding vector difference V between the start period and the end period trend , indicating the general direction of the overall path evolution;

[0029] Calculate the local path direction vector V t and the overall trend vector V trend The angle θt ,θ t The range is [0,π];

[0030] Calculate the median f of all angles θ , reflects the overall degree of deviation and calculates the interquartile range IQR of all angles θ , reflecting the discrete degree of deviation;

[0031] Normalized angle median and interquartile range, according to the normalized angle median E a and interquartile range E b Calculate the convergence anomaly index. The calculation steps are: ED is the convergence abnormality index;

[0032] Calculate the angle of change in the direction of the adjacent path as the jump angle Φ t ; Calculate the average jump angle The average value of the normalized jump angle Get the jump abnormality index E c ;

[0033] Combine the convergence anomaly index and the jump anomaly index to calculate the convergence anomaly value. The calculation formula is: HY = ED × E c , where HY is the convergence outlier.

[0034] Optionally, the steps of analyzing the evaluation results of the user's learning quality based on the path behavior anomaly value and the convergence anomaly value to determine whether the evaluation results of the current user's learning quality contain fraudulent phenomena are as follows:

[0035] Add the path behavior anomaly value and the convergence anomaly value to obtain the fraud index, and compare the fraud index with the preset fraud index threshold. If the fraud index is not less than the preset fraud index threshold, it means that the current user learning quality evaluation result contains fraudulent phenomena and the user learning quality needs to be re-evaluated.

[0036] If the fraud index is less than the preset fraud index threshold, it means that there is no fraud in the evaluation result of the current user's learning quality, and the evaluation result of the current user's learning quality is taken as the final evaluation result.

[0037] Optionally, the steps for analyzing the paths of all knowledge graph sequences and identifying abnormal paths and normal paths are as follows:

[0038] Compare the abnormality degree corresponding to all knowledge graph sequence paths with the preset abnormality threshold. If the abnormality degree is not less than the preset abnormality threshold, the path is directly recorded as an abnormal path;

[0039] If the abnormality is less than the preset abnormality threshold, the path is directly recorded as a temporary normal path;

[0040] Filter out all temporarily normal paths, calculate the mean and standard deviation of abnormality corresponding to all temporarily normal paths, and calculate the absolute difference between the mean abnormality and the preset abnormality threshold as the critical difference;

[0041] Compare the critical absolute difference with the preset critical difference threshold, and compare the standard deviation with the preset standard deviation. If the critical absolute difference is less than the preset critical difference threshold or the standard deviation is not less than the preset standard deviation, all temporarily normal paths are recorded as abnormal paths.

[0042] If the critical absolute difference is not less than the preset critical difference threshold and the standard deviation is less than the preset standard deviation, all temporary normal paths are recorded as normal paths;

[0043] For the learning content corresponding to the abnormal path, the learning quality assessment results are directly recorded as invalid learning;

[0044] For the learning content corresponding to the normal path, a secondary evaluation is conducted, and the results of the secondary evaluation of the normal path are used as the overall learning quality evaluation of the user on the education platform.

[0045] In a second aspect of the present invention, a user learning quality assessment system based on education platform data analysis is proposed, the system comprising:

[0046] Behavior anomaly module: Builds a dynamic knowledge graph sequence based on the user's overall learning process and learning content, analyzes the degree of abnormality of the user's learning path based on the dynamic knowledge graph sequence, and calculates the path behavior anomaly value;

[0047] Convergence anomaly module: This module extracts the path embedding vectors corresponding to the paths in each knowledge graph through a neural network-based graph embedding method. It then analyzes the stability of the user's learning trajectory based on the embedding vectors corresponding to the paths in each knowledge graph and calculates the convergence anomaly value.

[0048] Judgment module: Analyzes the evaluation results of user learning quality based on path behavior anomalies and convergence anomalies to determine whether there is fraud in the current user learning quality evaluation results;

[0049] Evaluation module: If there is fraud in the evaluation results of the current user's learning quality, analyze the paths of all knowledge graph sequences, identify abnormal paths and normal paths, and re-evaluate the user's overall learning quality based on the abnormal paths and normal paths.

[0050] Beneficial effects of the present invention:

[0051] The present invention proposes a user learning quality assessment method and system based on education platform data analysis, constructs a dynamic knowledge graph sequence according to the user's overall learning process and learning content, analyzes the abnormality of the user's learning path according to the dynamic knowledge graph sequence, and calculates the path behavior anomaly value; extracts the path embedding vector corresponding to the path in each knowledge graph through a neural network-based graph embedding method, analyzes the stability of the user's learning trajectory according to the embedding vector corresponding to the path in each knowledge graph, and calculates the convergence anomaly value; analyzes the user learning quality assessment results according to the path behavior anomaly value and the convergence anomaly value, and determines whether there is fraud in the current user learning quality assessment results; if there is fraud in the current user learning quality assessment results, analyzes the paths of all knowledge graph sequences, identifies abnormal paths and normal paths, and re-evaluates the overall user learning quality based on the abnormal paths and normal paths. In this way, by constructing a dynamic knowledge spectrum evaluation, it is possible to determine whether there is fraud in the evaluation results of the current user's learning quality. If there is fraud, the user's learning quality will be re-evaluated to ensure that the computer can accurately identify the user's true learning status and reduce the overall bias or distortion of the learning quality evaluation results, thereby weakening the influence of the education platform on the accuracy of the user's learning status and reducing the impact on the effectiveness of teaching intervention. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] The present invention will be further described below with reference to the accompanying drawings.

[0053] Figure 1 A flowchart of the user learning quality assessment method based on education platform data analysis;

[0054] Figure 2 This is a framework diagram of the user learning quality assessment system based on education platform data analysis. DETAILED DESCRIPTION

[0055] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0056] The embodiment of the present invention provides a method for evaluating user learning quality based on education platform data analysis. Figure 1 , Figure 1 Flowchart of a method for evaluating user learning quality based on education platform data analysis provided in an embodiment of the present invention. The method comprises the following steps:

[0057] Build a dynamic knowledge graph sequence based on the user's overall learning process and learning content, analyze the abnormality of the user's learning path based on the dynamic knowledge graph sequence, and calculate the path behavior abnormality value;

[0058] The neural network-based graph embedding method is used to extract the path embedding vectors corresponding to the paths in each knowledge graph. The stability of the user's learning trajectory is analyzed based on the embedding vectors corresponding to the paths in each knowledge graph, and the convergence outlier is calculated.

[0059] Analyze the evaluation results of user learning quality based on path behavior anomalies and convergence anomalies to determine whether there is fraud in the current evaluation results of user learning quality;

[0060] If there is fraud in the evaluation results of the current user's learning quality, the paths of all knowledge graph sequences are analyzed to identify abnormal paths and normal paths, and the overall learning quality of the user is re-evaluated based on the abnormal paths and normal paths.

[0061] Based on the user learning quality assessment method based on education platform data analysis provided by an embodiment of the present invention, through the above-mentioned method, by constructing a dynamic knowledge spectrum, it is determined whether there is fraud in the assessment results of the current user's learning quality. If fraud exists, the user's learning quality is re-evaluated to ensure that the computer can accurately identify the user's true learning status, reduce the overall high or distortion of the learning quality assessment results, thereby weakening the influence of the education platform on the judgment of the accuracy of the user's learning status and reducing the impact on the teaching intervention effect.

[0062] In one embodiment, a dynamic knowledge graph sequence is constructed based on the user's overall learning process and learning content, and the abnormality of the user's learning path is analyzed based on the dynamic knowledge graph sequence to calculate the path behavior abnormality value;

[0063] In one implementation, building a dynamic knowledge graph sequence based on the user's overall learning process and learning content includes:

[0064] Record the total time users spend studying on the education platform and divide the total time into several sub-time periods;

[0065] For each sub-time period, extract the learning content of the users in the corresponding sub-time period as a node of the knowledge graph of the corresponding sub-time period; the learning content includes knowledge point teaching and viewing content, knowledge point practice content, and knowledge point test content;

[0066] For several nodes in the knowledge graph, based on the learning time sequence, directed edges between adjacent nodes are added, and the weights of directed edges between adjacent nodes are determined based on the preset standard learning path to obtain a complete learning path;

[0067] The knowledge graph containing the complete learning path is used as the knowledge graph of the corresponding sub-time period; a dynamic knowledge graph sequence is constructed based on the knowledge graph of each sub-time period.

[0068] In one implementation, the weights of directed edges between adjacent nodes are determined based on a preset standard learning path as follows:

[0069] Extract the user's initial learning content on the education platform as the starting node, match it to the preset standard learning path, and determine the number of jump levels between all adjacent nodes in the knowledge graph compared to the preset standard learning path;

[0070] Calculate the weight of the directed edge between adjacent nodes. The calculation formula is: w i,i+1 =max(0,1-β×k), where w i,j is the weight of the directed edge between adjacent nodes i and i+1; k is the number of jump layers between adjacent nodes i and i+1 compared to the preset standard learning path; β is the preset penalty coefficient, which is generally set to 0.1-0.2.

[0071] It should be noted that the preset standard learning path is the standard learning path set by the education platform. For example, suppose a user learns about "force" and the standard learning path is: Concept of force → Gravity teaching → Gravity practice → Gravity test → Elasticity teaching → Elasticity practice → Elasticity test → Friction teaching → Friction practice → Friction test → Newton's first law teaching → Newton's first law practice → Newton's first law test → Newton's second law teaching → Newton's second law practice → Newton's second law test. Assuming the user starts learning about the concept of force and skips "Elasticity teaching → Elasticity practice → Elasticity test" during the actual learning process, jumping directly from "Gravity test" to "Friction teaching," the number of skip levels in the user's learning path is 3. Substituting this into the calculated edge weight, the weight of the directed edge between "Gravity test" and "Friction teaching" is 1-0.2*3=0.4. By matching this preset standard learning path, the weights of the directed edges are calculated for all adjacent nodes in the knowledge graph.

[0072] It should be noted that the benefits of constructing knowledge graphs for users in different time periods are: it can directly and dynamically depict the user's actual learning path and behavior patterns, thereby enabling fine-grained analysis of the learning process. On the one hand, dividing user learning behavior into several sub-time periods and constructing independent knowledge graphs for each sub-time period can accurately reflect the user's learning content, sequence, and depth at different times, helping to identify changing trends in the learning stage; on the other hand, introducing a preset standard learning path and the number of jump levels as the basis for setting edge weights can quantify the degree of deviation between the user's learning path and the ideal path, thereby providing a better foundation for subsequent analysis of the degree of abnormality in the user's learning path and calculating the path behavior anomaly value.

[0073] In one implementation, the abnormality of the user's learning path is analyzed based on the dynamic knowledge graph sequence, and the steps for calculating the abnormal value of the path behavior are as follows:

[0074] Extract the learning path from the knowledge graph sequence corresponding to each sub-time period to obtain the set of all paths P t ={P1,P2,P3....P k}, where k represents the total number of sub-time periods;

[0075] The path corresponding to each knowledge graph is P j = <v1,v2....v m >, where v n Indicates the order of the nodes, indicating the nth node in the path;

[0076] Calculate the absolute difference between the edge weights of consecutive nodes in each path compared to the standard edge weight, add up all the absolute differences, and use them as the jump value of each path; measure the jump degree of the edge weight of the entire path compared to the standard edge weight; the standard edge weight value is 1;

[0077] The “degree of fluctuation” of the edge weights between consecutive nodes in each path is defined as the transition tension. The transition tension of each path is calculated using the following formula: Where, T(P j ) is the transition tension of the j-th path; is the mean edge weight of the j-th path;

[0078] The abnormality of each path is calculated based on the jump value and transition tension of each path. The calculation formula is: D(P j )=S(P j )×T(P j ), where D(P j ) is the abnormality of the jth path, S(P j ) is the jump value of the j-th path;

[0079] Calculate the abnormality difference between adjacent paths, measure the abnormality of the overall path over time, and calculate the overall path behavior abnormality. The calculation formula is: Where YU is the path behavior anomaly value.

[0080] It should be noted that the path behavior anomaly value is used to measure whether the overall degree of anomaly increases over time during a user's learning process on the education platform. This is because in the early stages of learning, a user's learning path may be unstable, jumpy, or deviate from the normal knowledge structure. This may be due to unfamiliarity with the education platform and a lack of clear understanding of the knowledge points, and the user's learning behavior may show large fluctuations or tentative exploration. However, as learning progresses, the user should gradually become familiar with the platform's operating procedures and master the logical relationships between knowledge points. The learning path should also become stable and orderly, and the learning path should move closer to the standard path. Therefore, if the path behavior anomaly value continues to rise in subsequent time periods, exceeding the previous level, it indicates that the user's learning behavior is abnormal, with a high probability of fraudulent behavior such as false clicks, skipping learning, and repeated course review. In this case, the current evaluation of the user's learning quality is overstated or distorted. Further teaching intervention based on the current user learning quality evaluation will reduce the user's learning effect. Therefore, a re-evaluation of the user's learning quality is necessary.

[0081] It's important to note that the benefits of combining jump value and transition tension to calculate each path's anomaly are: the jump value reflects the overall deviation of the path's edge weights from the standard edge weights, quantifying the "jumps" or "deviations" in the knowledge structure of the learning path and identifying whether the user is learning along a reasonable knowledge progression sequence. Furthermore, transition tension measures the intensity of fluctuations in the path's edge weights, reflecting the stability and continuity of learning behavior. Multiplying the two together to form each path's anomaly takes into account both the degree of deviation and the fluctuation, amplifying the impact of both anomalies and making the anomaly measurement more accurate.

[0082] In one embodiment, a neural network-based graph embedding method is used to extract the path embedding vectors corresponding to the paths in each knowledge graph, and the stability of the user's learning trajectory is analyzed based on the embedding vectors corresponding to the paths in each knowledge graph, and the convergence outlier is calculated;

[0083] In one implementation, the steps of extracting the path embedding vector corresponding to the path in each knowledge graph by using a graph embedding method based on a neural network are as follows:

[0084] For each time period t, the user’s learning path at that time is extracted from the complete knowledge graph to form a subgraph G t =(v t ,r t ), v t Represents a path node in the subgraph; r t Represents the association relationship between knowledge points, that is, directed edges and connection relationships;

[0085] For each node v∈v tGenerate the initial embedding vector h (0) v ∈R d , R d represents a d-dimensional real vector space;

[0086] Initialization method: If the node has a pre-trained representation, use it directly; otherwise, use one-hot encoding and then multiply the embedding matrix for conversion; h (0) v =Embedding(v);

[0087] The GCN layer propagates and aggregates features of the graph structure and updates the representation of each node. For the lth layer, the representation update formula of node v is as follows: Where h (l) v is the representation of node v at layer l, represents the neighbor set of node v, including itself; d v The degree of node v is the sum of the number of in-degree and out-degree of node v; W (l) represents the weight matrix of the lth layer, σ is the activation function, such as ReLU; u represents the node to be updated currently, Represents all nodes adjacent to v; GCN collects the previous layer features h of all neighboring nodes u u (l-1) , and aggregate them to update the representation h of the current layer of node v (l) v ; Stack 2 to 3 layers of GCN to complete information aggregation, and finally obtain the high-order structure perception representation h of each node (L) v ;

[0088] The entire subgraph G t All node embedding vectors in are aggregated into a path embedding vector z t ∈R d , the aggregation method can be mean pooling aggregation, maximum pooling aggregation or attention aggregation, which is aggregated into the embedding vector z of each knowledge graph t ;

[0089] Finally, the path embedding vector corresponding to each sub-time period is obtained: t ={z1,z2,z3....z k}.

[0090] It should be noted that, by improving the above method, we can obtain the path embedding vector sequence z t ,This sequence represents the change curve of the user’s learning trajectory, ,which is used for convergence anomaly detection such as stability, ,jumpiness, and offset in subsequent analysis.

[0091] In one implementation, the stability of the user's learning trajectory is analyzed based on the embedding vectors corresponding to the paths in each knowledge graph, and the steps for calculating the convergence outlier are as follows:

[0092] The embedding vector set z corresponding to the path in the knowledge graph t ={z1,z2,z3....z k}, where k represents the total number of sub-time periods;

[0093] Calculate the difference vector of the path embedding between two adjacent time periods to represent the local direction of path evolution: V t =z t -z t-1 , where t = 2, 3...k;

[0094] Calculate the embedding vector difference V between the start period and the end period trend , indicating the general direction of the overall path evolution, V trend =z k -z1;

[0095] Calculate the local path direction vector V t and the overall trend vector V trend The angle θ t , the calculation formula is: θ t =arccos(cosθ t ), the range is [0,π];

[0096] Calculate the median f of all angles θ , reflects the overall degree of deviation and calculates the interquartile range IQR of all angles θ , reflecting the discrete degree of deviation;

[0097] Normalized angle median and interquartile range, normalized angle median E a for: Normalized interquartile range of angle E b for:

[0098] According to the normalized median angle E a and interquartile range E b Calculate the convergence anomaly index. The calculation steps are: ED is the convergence abnormality index;

[0099] Calculate the angle of change in the direction of the adjacent path as the jump angle t=3,4...k;

[0100] Calculate the average jump angle The calculation formula is:

[0101] The average value of the normalized jump angle is used to obtain the jump anomaly index E c for:

[0102] Combine the convergence anomaly index and the jump anomaly index to calculate the convergence anomaly value. The calculation formula is: HY = ED × E c , where HY is the convergence outlier.

[0103] It should be noted that the convergence outlier is a comprehensive anomaly metric used to measure whether a user's learning trajectory exhibits stable evolutionary characteristics during the learning process. This metric determines whether a user's learning trajectory exhibits natural evolution and continuity by analyzing the directional change trends of the embedding vectors of the paths traversed by the user in the knowledge graph over time. Specifically, a normal learning trajectory should exhibit a relatively smooth and coherent evolutionary trend. That is, the direction of change of the path embedding vector should be generally consistent with the overall trend direction, with small local angles and mild jumps. However, if a user's learning behavior exhibits abnormalities, such as frequently switching between different topics in a short period of time or jumping to non-related knowledge points, the angles between adjacent path directions will increase significantly, and the deviations between the overall trend and the local evolutionary direction will become more discrete, manifested as an increase in the median and interquartile range of the angles, as well as an increase in the mean of the directional jump angles. Therefore, when multiple indicators, such as the normalized median angle, interquartile range, and directional jump, collectively reflect significant instability in path evolution, the final aggregated convergence anomaly value will increase, indicating that the current user's learning trajectory has significantly deviated from the normal evolutionary pattern. A larger convergence anomaly value indicates that the user's learning behavior is structurally unnatural, exhibiting characteristics of manipulation or abnormal generation. This is likely associated with fraudulent behavior, such as using automated programs to brush learning records or simulating behavior to obtain learning achievement certification. Therefore, an increase in the convergence anomaly value becomes an important indicator for identifying the credibility of a user's learning behavior and whether there is suspicion of fraud.

[0104] In one implementation, the aforementioned convergence anomaly indicator, combined with the jump anomaly indicator, calculates the convergence anomaly value throughout the learning process. The advantage lies in the following: the convergence anomaly indicator focuses on whether the evolutionary trend of the overall learning path deviates from the norm, reflecting the coherence and consistency of the user's learning goals; while the jump anomaly indicator focuses on the continuity of local learning behaviors, effectively capturing issues such as frequent jumps and sudden behavioral changes. This combination of the two can detect whether users are gradually accumulating knowledge and deepening their learning along reasonable paths. It can also identify seemingly normal trajectories that actually exhibit unusual jumps and lack of learning depth, effectively screening out potential fabricated and inauthentic learning behaviors.

[0105] In one embodiment, the steps of analyzing the evaluation results of the user's learning quality based on the path behavior anomaly value and the convergence anomaly value to determine whether the current evaluation results of the user's learning quality contain fraudulent phenomena are as follows:

[0106] Add the path behavior anomaly value and the convergence anomaly value to obtain the fraud index, and compare the fraud index with the preset fraud index threshold. If the fraud index is not less than the preset fraud index threshold, it means that the current user learning quality evaluation result contains fraudulent phenomena and the user learning quality needs to be re-evaluated.

[0107] If the fraud index is less than the preset fraud index threshold, it means that there is no fraud in the evaluation result of the current user's learning quality, and the evaluation result of the current user's learning quality is taken as the final evaluation result.

[0108] It's important to note that the fraud index, derived by adding the path behavior anomaly value and the convergence anomaly value, simultaneously accounts for two different types of anomalies: the former reflects jumps and disconnections in the user's learning path's knowledge structure, while the latter reflects the continuity and convergence of the learning trajectory over time. When the sum of the two exceeds a preset threshold, it indicates not only that the user's behavior is jumpy or disordered, but also that their overall trajectory fails to converge gradually like a normal learner. This combined effect suggests a high likelihood of fraudulent behavior, such as cheating, idle time, and simulated clickbait. The threshold can be set based on offline statistics of a large number of "normal learning" samples and "known anomaly" samples, ensuring that the fraud index distribution of most normal samples is below the threshold, while the fraud index distribution of anomalous samples is above the threshold. If the user's current fraud index is at least this threshold, it indicates that their learning quality assessment results are severely distorted, necessitating a reassessment process. Conversely, if the fraud index remains consistently below the threshold, the user's learning behavior is considered normal, and the current learning quality assessment result is used as the final conclusion. This ensures that the assessment is neither overly lenient nor overly strict.

[0109] In one embodiment, if there is fraud in the evaluation results of the current user's learning quality, the embedding vectors of all knowledge graph sequences are analyzed to identify abnormal paths and normal paths, and the overall learning quality of the user is re-evaluated based on the abnormal paths and normal paths.

[0110] In one implementation, the steps for analyzing the paths of all knowledge graph sequences and identifying abnormal and normal paths are as follows:

[0111] Compare the abnormality degree corresponding to all knowledge graph sequence paths with the preset abnormality threshold. If the abnormality degree is not less than the preset abnormality threshold, the path is directly recorded as an abnormal path;

[0112] If the abnormality is less than the preset abnormality threshold, the path is directly recorded as a temporary normal path;

[0113] Filter out all temporarily normal paths, calculate the mean and standard deviation of abnormality corresponding to all temporarily normal paths, and calculate the absolute difference between the mean abnormality and the preset abnormality threshold as the critical difference;

[0114] Compare the critical absolute difference with the preset critical difference threshold, and compare the standard deviation with the preset standard deviation. If the critical absolute difference is less than the preset critical difference threshold or the standard deviation is not less than the preset standard deviation, all temporarily normal paths are recorded as abnormal paths.

[0115] If the critical absolute difference is not less than the preset critical difference threshold and the standard deviation is less than the preset standard deviation, all temporary normal paths are recorded as normal paths;

[0116] For the learning content corresponding to the abnormal path, the learning quality assessment results are directly recorded as invalid learning;

[0117] For the learning content corresponding to the normal path, a secondary evaluation is conducted, and the results of the secondary evaluation of the normal path are used as the overall learning quality evaluation of the user on the education platform.

[0118] It should be noted that the calculation method of the abnormality degree corresponding to each path has been described above and will not be repeated here.

[0119] It's important to note that when identifying abnormal and normal paths across all knowledge graph sequences, the abnormality score of each path is compared against a preset abnormality threshold. If a path's abnormality score is higher than or equal to the threshold, the education platform will directly mark it as an abnormal path, deeming it to present a significant learning problem or fraud risk. Otherwise, the path will be temporarily classified as normal. The mean and standard deviation of the abnormality scores for these temporarily normal paths are then calculated, and the difference between the mean abnormality score and the preset threshold is used as the critical difference. If this critical difference is small or the standard deviation is large, indicating significant fluctuation in the abnormality scores among temporarily normal paths and overall instability, all temporarily normal paths will be classified as abnormal, thereby improving the robustness of detection. For example, if the abnormality threshold is 0.5 and the mean abnormality score of temporarily normal paths is 0.48, the difference from the threshold is only 0.02, and the standard deviation is high, this indicates that while the average appears normal, there is significant fluctuation, indicating that some paths may harbor abnormalities, and therefore are classified as abnormal. Conversely, if the mean is significantly below the threshold and the standard deviation is small, indicating that these paths have stable and low abnormality, they can be confirmed as normal paths. For the learning content corresponding to abnormal paths, the learning quality assessment results are directly invalidated to avoid misleading subsequent decision-making. For normal paths, a more accurate secondary assessment is performed based on these paths, ultimately resulting in an overall assessment of the user's learning quality on the education platform. This ensures strict screening of abnormal behavior while maintaining the accuracy of the assessment of reasonable and normal learning trajectories.

[0120] It should be noted that for the learning content corresponding to the normal path, the secondary evaluation usually adopts a more detailed and multi-dimensional analysis method. For example, it combines the time evolution characteristics of the path embedding vector, the continuity and consistency of the learning behavior, and the depth and breadth of knowledge points, and uses machine learning models or deep learning models for comprehensive scoring. At the same time, learning outcome data such as test scores and homework completion status can also be introduced to verify the consistency between path behavior and actual learning effects. Through such multi-level and multi-angle evaluation, the user's true learning quality can be more accurately reflected, avoiding misjudgment caused by a single indicator, and ultimately providing the education platform with more reliable and scientific learning quality evaluation results.

[0121] Based on the same inventive concept, the present invention also provides a user learning quality evaluation system based on education platform data analysis. Figure 2 , Figure 2 This is a framework diagram of a user learning quality assessment system based on education platform data analysis provided by an embodiment of the present invention. The system includes:

[0122] Behavior anomaly module: Builds a dynamic knowledge graph sequence based on the user's overall learning process and learning content, analyzes the degree of abnormality of the user's learning path based on the dynamic knowledge graph sequence, and calculates the path behavior anomaly value;

[0123] Convergence anomaly module: This module extracts the path embedding vectors corresponding to the paths in each knowledge graph through a neural network-based graph embedding method. It then analyzes the stability of the user's learning trajectory based on the embedding vectors corresponding to the paths in each knowledge graph and calculates the convergence anomaly value.

[0124] Judgment module: Analyzes the evaluation results of user learning quality based on path behavior anomalies and convergence anomalies to determine whether there is fraud in the current user learning quality evaluation results;

[0125] Evaluation module: If there is fraud in the evaluation results of the current user's learning quality, analyze the paths of all knowledge graph sequences, identify abnormal paths and normal paths, and re-evaluate the user's overall learning quality based on the abnormal paths and normal paths.

[0126] Based on the user learning quality assessment system based on education platform data analysis provided by the embodiment of the present invention, through the above-mentioned method, by constructing a dynamic knowledge spectrum, it is determined whether there is fraud in the assessment results of the current user's learning quality. If fraud exists, the user's learning quality is re-evaluated to ensure that the computer can accurately identify the user's true learning status, reduce the overall high or distortion of the learning quality assessment results, thereby weakening the influence of the education platform on the accuracy of the user's learning status and reducing the impact on the teaching intervention effect.

[0127] The above is a detailed description of an embodiment of the present invention. However, the content described is only a preferred embodiment of the present invention and should not be considered to limit the scope of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the patent coverage of the present invention.

Claims

1. A user learning quality assessment method based on education platform data analysis, characterized in that: The following steps are involved: Build a dynamic knowledge graph sequence based on the user's overall learning process and learning content, analyze the abnormality of the user's learning path based on the dynamic knowledge graph sequence, and calculate the path behavior abnormality value; The neural network-based graph embedding method is used to extract the path embedding vectors corresponding to the paths in each knowledge graph. The stability of the user's learning trajectory is analyzed based on the embedding vectors corresponding to the paths in each knowledge graph, and the convergence outlier is calculated. Analyze the evaluation results of user learning quality based on path behavior anomalies and convergence anomalies to determine whether there is fraud in the current evaluation results of user learning quality; If there is fraud in the evaluation results of the current user's learning quality, the paths of all knowledge graph sequences are analyzed to identify abnormal paths and normal paths, and the overall learning quality of the user is re-evaluated based on the abnormal paths and normal paths.

2. The user learning quality assessment method based on education platform data analysis according to claim 1 is characterized in that: Building a dynamic knowledge graph sequence based on the user's overall learning process and learning content includes: Record the total time users spend studying on the education platform and divide the total time into several sub-time periods; For each sub-time period, extract the learning content of the users in the corresponding sub-time period as a node of the knowledge graph of the corresponding sub-time period; the learning content includes knowledge point teaching and viewing content, knowledge point practice content, and knowledge point test content; For several nodes in the knowledge graph, based on the learning time sequence, directed edges between adjacent nodes are added, and the weights of directed edges between adjacent nodes are determined based on the preset standard learning path to obtain a complete learning path; The knowledge graph containing the complete learning path is used as the knowledge graph of the corresponding sub-time period; a dynamic knowledge graph sequence is constructed based on the knowledge graph of each sub-time period.

3. The user learning quality assessment method based on education platform data analysis according to claim 2 is characterized in that: The weights of directed edges between adjacent nodes are determined based on the preset standard learning path: Extract the user's initial learning content on the education platform as the starting node, match it to the preset standard learning path, and determine the number of jump levels between all adjacent nodes in the knowledge graph compared to the preset standard learning path; Calculate the weight of the directed edge between adjacent nodes. The calculation formula is: w i,i+1 =max(0,1-β×k), where w i,j is the weight of the directed edge between adjacent nodes i and i+1; k is the number of jump layers between adjacent nodes i and i+1 compared to the preset standard learning path; β is the preset penalty coefficient, which ranges from 0.1 to 0.

2.

4. The user learning quality assessment method based on education platform data analysis according to claim 1 is characterized in that: The abnormal degree of the user's learning path is analyzed according to the dynamic knowledge graph sequence. The steps for calculating the abnormal value of the path behavior are as follows: Extract the learning path from the knowledge graph sequence corresponding to each sub-time period to obtain the set of all paths P t ={P1,P2,P3....P k }, where k represents the total number of sub-time periods; The path corresponding to each knowledge graph is P j = <v1,v2....v m >, where v n Indicates the order of the nodes, indicating the nth node in the path; Calculate the absolute difference between the edge weights of consecutive nodes in each path and the standard edge weight, add up all the absolute differences, and use them as the jump value of each path; the standard edge weight value is 1; The fluctuation degree of the edge weights between consecutive nodes in each path is defined as the transition tension. The transition tension of each path is calculated using the following formula: Where, T(P j ) is the transition tension of the j-th path; is the mean edge weight of the j-th path; The abnormality of each path is calculated based on the jump value and transition tension of each path. The calculation formula is: D(P j )=S(P j )×T(P j ), where D(P j ) is the abnormality of the jth path, S(P j ) is the jump value of the j-th path; Calculate the abnormality difference between adjacent paths, measure the abnormality of the overall path over time, and calculate the overall path behavior abnormality. The calculation formula is: Where YU is the path behavior anomaly value.

5. The user learning quality assessment method based on education platform data analysis according to claim 1 is characterized in that: The stability of the user's learning trajectory is analyzed based on the embedding vectors corresponding to the paths in each knowledge graph. The steps for calculating the convergence outlier value are as follows: The embedding vector set z corresponding to the path in the knowledge graph t ={z1,z2,z3....z k }, where k represents the total number of sub-time periods; Calculate the difference vector V between the path embeddings of two adjacent time periods t , which represents the local direction of path evolution: Calculate the embedding vector difference V between the start period and the end period trend , indicating the general direction of the overall path evolution; Calculate the local path direction vector V t and the overall trend vector V trend The angle θ t ,θ t The range is [0,π]; Calculate the median f of all angles θ , reflects the overall degree of deviation and calculates the interquartile range IQR of all angles θ , reflecting the discrete degree of deviation; Normalized angle median and interquartile range, according to the normalized angle median E a and interquartile range E b Calculate the convergence anomaly index. The calculation steps are: ED is the convergence abnormality index; Calculate the angle of change in the direction of the adjacent path as the jump angle Φ t ; Calculate the average jump angle The average value of the normalized jump angle Get the jump abnormality index E c ; Combine the convergence anomaly index and the jump anomaly index to calculate the convergence anomaly value. The calculation formula is: HY = ED × E c , where HY is the convergence outlier.

6. The user learning quality assessment method based on education platform data analysis according to claim 1 is characterized in that: The steps for analyzing the evaluation results of user learning quality based on path behavior anomalies and convergence anomalies to determine whether there is fraud in the current evaluation results of user learning quality are as follows: Add the path behavior anomaly value and the convergence anomaly value to obtain the fraud index, and compare the fraud index with the preset fraud index threshold. If the fraud index is not less than the preset fraud index threshold, it means that the current user learning quality evaluation result contains fraudulent phenomena and the user learning quality needs to be re-evaluated. If the fraud index is less than the preset fraud index threshold, it means that there is no fraud in the evaluation result of the current user's learning quality, and the evaluation result of the current user's learning quality is taken as the final evaluation result.

7. The user learning quality assessment method based on education platform data analysis according to claim 1 is characterized in that: The steps to analyze the paths of all knowledge graph sequences and identify abnormal paths and normal paths are as follows: Compare the abnormality degree corresponding to all knowledge graph sequence paths with the preset abnormality threshold. If the abnormality degree is not less than the preset abnormality threshold, the path is directly recorded as an abnormal path; If the abnormality is less than the preset abnormality threshold, the path is directly recorded as a temporary normal path; Filter out all temporarily normal paths, calculate the mean and standard deviation of abnormality corresponding to all temporarily normal paths, and calculate the absolute difference between the mean abnormality and the preset abnormality threshold as the critical difference; Compare the critical absolute difference with the preset critical difference threshold, and compare the standard deviation with the preset standard deviation. If the critical absolute difference is less than the preset critical difference threshold or the standard deviation is not less than the preset standard deviation, all temporarily normal paths are recorded as abnormal paths. If the critical absolute difference is not less than the preset critical difference threshold and the standard deviation is less than the preset standard deviation, all temporary normal paths are recorded as normal paths; For the learning content corresponding to the abnormal path, the learning quality assessment results are directly recorded as invalid learning; For the learning content corresponding to the normal path, a secondary evaluation is conducted, and the results of the secondary evaluation of the normal path are used as the overall learning quality evaluation of the user on the education platform.

8. A user learning quality assessment system based on education platform data analysis, used to implement the user learning quality assessment method based on education platform data analysis according to any one of claims 1 to 7, characterized in that: The system comprises: Behavior anomaly module: Builds a dynamic knowledge graph sequence based on the user's overall learning process and learning content, analyzes the degree of abnormality of the user's learning path based on the dynamic knowledge graph sequence, and calculates the path behavior anomaly value; Convergence anomaly module: This module extracts the path embedding vectors corresponding to the paths in each knowledge graph through a neural network-based graph embedding method. It then analyzes the stability of the user's learning trajectory based on the embedding vectors corresponding to the paths in each knowledge graph and calculates the convergence anomaly value. Judgment module: Analyzes the evaluation results of user learning quality based on path behavior anomalies and convergence anomalies to determine whether there is fraud in the current user learning quality evaluation results; Evaluation module: If there is fraud in the evaluation results of the current user's learning quality, analyze the paths of all knowledge graph sequences, identify abnormal paths and normal paths, and re-evaluate the user's overall learning quality based on the abnormal paths and normal paths.

Images