Online education robust learning mastery estimation and adaptive verification method

Abstract

The application discloses an online education robust learning mastery estimation and self-adaptive verification method, and mainly solves the problem of inaccurate evaluation, which comprises the following steps: obtaining answer data information of students participating in online education; performing reputation calibration processing according to the answer data information to obtain a calibrated reputation value; performing mastery estimation by using the calibrated reputation value to obtain a student mastery weight, a question parameter and a target value; performing risk assessment by using a behavior log and a social relationship graph to obtain individual abnormal scores and social risks; obtaining instant benefits and comprehensive benefits according to the individual abnormal scores and the social risks, adopting attention weighted replication dynamics evaluation to obtain a risk heat map; and performing strategy linkage and output according to the student mastery weight, the question parameter and the risk heat map. Through the above scheme, the application has the advantages of simple logic, accuracy and reliability, and has high practical value and popularization value in the field of online education technology.

Description

Technical Field

[0001] This invention relates to the field of online education technology, and in particular to a robust learning mastery estimation and adaptive verification method for online education. Background Technology

[0002] Online education platforms assess students' mastery of content based on their submitted answer records, and then use algorithms to generate personalized practice recommendations and learning paths. However, in practice, several common issues arise: First, due to inconsistent test difficulty, simply considering accuracy rates cannot accurately reflect true mastery. Second, students engage in cheating or other abnormal behaviors: in real-world environments, students may use screen switching, proxy answering, or group answer sharing to cheat, causing assessments to deviate from their true level. Third, conventional platforms often lack adequate or no anti-cheating mechanisms. Even when they exist, they often rely on fixed thresholds or a fixed "suspicious percentage," failing to reflect changes in student behavior over time, or instances of cheating in offline classes or on social networks. Therefore, online education platforms need to introduce robust and reliable learning mastery estimation methods to ensure the accuracy of subsequent path planning.

[0003] Currently, existing online education platforms suffer from the following technical problems: First, the system's reputation and ability assessments lack a difficulty dimension: the difficulty of the questions is not taken into account, meaning that "doing easy questions" can quickly improve the corresponding reputation and ability. Second, there is a lack of constraints related to students' historical performance: students achieving significantly higher scores in a short period can easily inflate their assessment values, making them inconsistent with their long-term historical performance, thus artificially inflating their ability assessments. Third, there is a lack of ability to address cheating in offline classes and on social networks: students at risk of cheating in offline classes and virtual classes on social networks lack identification and constraints, and these students spread cheating through classes and social networks, leading to widespread cheating. Fourth, the diagnostic sampling mechanism lacks differentiation: the allocation of diagnostic sampling does not consider the risks of individual students and class communities, and cannot dynamically adjust the intensity and frequency of verification.

[0004] Therefore, there is an urgent need to propose a logically simple, accurate, and reliable method for providing robust learning mastery estimation and adaptive verification in online education. Summary of the Invention

[0005] To address the aforementioned problems, the present invention aims to provide a robust learning mastery estimation and adaptive verification method for online education. The technical solution adopted by the present invention is as follows:

[0006] A robust learning mastery estimation and adaptive verification method for online education, characterized by the following steps:

[0007] Obtain the answer data of students participating in online education; the answer data includes answer records, answer flags, question difficulty, behavior logs, and community relationship diagrams;

[0008] The credit rating is obtained by performing credit calibration based on the answer data.

[0009] Mastery is estimated using the calibrated reputation score, and the student mastery weights, question parameters, and target values ​​are obtained.

[0010] Risk assessment was conducted using behavioral logs and community relationship graphs to obtain individual anomaly scores and community risk.

[0011] Immediate and comprehensive benefits are derived from individual anomaly scores and community risks. An attention-weighted replication dynamics assessment is used to obtain a risk heatmap.

[0012] The strategy is linked and output based on the weight of students' mastery, question parameters, and risk heat map.

[0013] Compared with the prior art, the present invention has the following beneficial effects:

[0014] (1) This invention performs reputation calibration processing based on the answer data to obtain a calibrated reputation value. The calibrated reputation value is then used to estimate mastery, yielding student mastery weights, question parameters, and target values. When calculating student mastery, not only is the answer accuracy rate considered, but also question difficulty and Wilson's lower bound correction are introduced. This avoids the "artificially high" reputation caused by students repeatedly answering low-difficulty questions and provides a more stable reputation estimate even with insufficient sample size. Compared to existing evaluation methods that rely solely on accuracy, this significantly reduces the misjudgment rate in small sample situations.

[0015] (2) By introducing quantile equivalence regularization (QOC), this invention can maintain the consistency between the overall student ability assessment and their long-term historical performance when there is a short-term cramming or accidental high score. This design avoids the distortion of assessment values ​​due to short-term fluctuations and ensures the continuity of learning effect tracking, which is difficult to achieve with existing single statistical methods.

[0016] (3) This invention introduces graph smoothing constraints and attention-weighted replication dynamics model into the relationship network of classes or learning groups, which can identify and monitor the spread trend of abnormal behavior of "key students" and apply constraints in a timely manner. This not only improves the accuracy of cheating detection, but also prevents the large-scale spread of abnormal scores at the group level, while traditional methods mostly stay at the single user dimension and cannot take into account the group effect.

[0017] (4) This invention proposes an adaptive diagnostic sampling scheduling method that avoids a one-size-fits-all approach. Instead, it dynamically allocates verification tasks based on student reputation scores, abnormal characteristics, and group risk. Verification results are directly written back to the reputation score, forming a self-correcting closed loop of "evaluation-verification-re-evaluation." Compared with existing fixed-frequency or random sampling, this invention reduces the system burden while concentrating computing resources on high-risk objects, achieving more efficient and fairer risk control.

[0018] (5) This invention forms a complete mechanism covering individuals, groups, and feedback loops by organically combining four aspects: reputation calibration, historical constraints, group inhibition, and adaptive verification. Compared with traditional methods that only focus on one dimension, this synergistic effect of the invention enables online education platforms to simultaneously ensure the accuracy of assessment and the effectiveness of prevention and control in real-world large-scale applications, demonstrating significant practical value and promising prospects for promotion.

[0019] In summary, this invention has the advantages of simple logic and high accuracy and reliability, and has high practical and promotional value in the field of online education technology. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope of protection. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a logic flowchart of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this application clearer, the present invention will be further described below with reference to the accompanying drawings and embodiments. The embodiments of the present invention include, but are not limited to, the following embodiments. All other embodiments obtained by those skilled in the art based on the embodiments in this application without inventive effort are within the scope of protection of this application.

[0023] In this embodiment, the term "and / or" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can represent three situations: A exists alone, A and B exist simultaneously, and B exists alone.

[0024] The terms "first" and "second," etc., used in the specification and claims of this embodiment are used to distinguish different objects, not to describe a specific order of objects. For example, "first target object" and "second target object," etc., are used to distinguish different target objects, not to describe a specific order of target objects.

[0025] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design that is described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design. Specifically, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0026] In the description of the embodiments in this application, unless otherwise stated, "multiple" means two or more. For example, multiple processing units means two or more processing units; multiple systems means two or more systems.

[0027] like Figure 1 As shown, this embodiment provides a robust learning mastery estimation and adaptive verification method for online education, including reputation-constrained mastery estimation, adaptive diagnostic sampling scheduling, and attention-weighted replication dynamics evaluation. First, the robustness of mastery assessment is improved; then, dynamic diagnostic sampling is implemented; and finally, the spread of abnormal behavior is monitored and suppressed, forming a continuous linkage mechanism of assessment, correction, and prevention.

[0028] In this embodiment, reputation calibration considers the difficulty of questions when processing students' answer accuracy and mastery levels. It uses Wilson's lower bound to correct historical reputation, addressing the issue of inflated reputation caused by "doing easy questions." Additionally, structural constraints are implemented through quantile-based regularization of QOC, ensuring consistency between student mastery weights and historical reputation, reducing the impact of short-term cramming (abnormal fluctuations) on overall assessment. In the learning group relationship network within a class learning or online teaching system, graph smoothing constraints are introduced to prevent collective cheating and abnormal score shifts. This embodiment employs a near-end alternation minimization method, simultaneously updating student mastery and question parameters to ensure stable and reliable results. Furthermore, this embodiment adds a diagnostic sampling mechanism based on individual student risk and class community risk, dynamically allocating diagnostic sampling and writing the verification results back to the reputation value for continuous dynamic correction. Finally, this embodiment employs diffusion assessment, utilizing an attention-weighted replication dynamics model to monitor the diffusion trend of abnormal behavior within the class group and inversely influencing verification and recommendation strategies.

[0029] Specifically, the online education robust learning mastery estimation and adaptive verification method in this embodiment includes the following steps:

[0030] The first step is to obtain the answer data of students participating in online education. This answer data includes answer records, answer markers, question difficulty, behavior logs, and community relationship diagrams.

[0031] The second step involves performing reputation calibration based on the answer data to obtain the calibrated reputation value, which includes:

[0032] (21) Calculate the difficulty-weighted accuracy rate based on the answer data. Its expression is:

[0033] (1)

[0034] Here, formula (1) is used to calculate the student. The difficulty-weighted accuracy rate ranges from [0,1]. In actual teaching, students may improve their accuracy rate by completing a large number of easy test questions, but such results cannot reflect the students' true level. To address this issue, this embodiment assigns different weights to questions of varying difficulty when calculating students' accuracy rates. The more difficult the question, the greater its impact on the overall evaluation; the easier the question, the smaller its impact on the result. This allows for a better assessment of students' actual mastery and avoids inflated figures. In formula (1), Students Do you want to answer the question? The logo; This represents the problem's weight function; Students For the question The answer result (1 = correct, 0 = incorrect); Indicates the reference difficulty level; Indicates the topic Difficulty parameters; This represents a coefficient that controls the intensity of the impact of differences in difficulty levels, and is used to control the degree to which these differences are amplified.

[0035] (22) Reputation calibration (lower bound of Wilson-IRT for difficulty perception): weighted accuracy based on question difficulty. Based on the number of responses and the confidence coefficient, the calibrated reputation value is calculated. Its expression is:

[0036] (2)

[0037] When students answer a limited number of questions, their accuracy rate is likely to be unstable, meaning they might appear highly skilled simply because they guessed a few questions correctly. To avoid this bias, this implementation uses a weighted accuracy rate based on question difficulty. Based on this, a lower bound of the Wilson confidence interval is introduced for calibration. This corrects the problem of inflated values ​​due to insufficient sample size, and is used as the number of responses. When it is small, the reputation value The value will be suppressed to avoid the possibility of inflated values ​​due to a small sample size. This embodiment provides a robust lower bound estimate, which is a conservative estimate of the students' true level. In addition, this embodiment combines the difficulty of the questions to make the results fairer and avoid "relying on easy questions to improve credibility". In formula (2), Students The number of validly answered questions; This represents the confidence coefficient, typically taken as 1.96 (95% confidence interval) or 1.64 (90% confidence interval).

[0038] The third step involves using the calibrated reputation score to estimate mastery, obtaining the student mastery weights, item parameters, and target values, including:

[0039] (31) During learning assessment, if a student achieves unusually high scores in a short period or engages in occasional cramming, the weight of their mastery may be overestimated, becoming inconsistent with their long-term reputation. To avoid this, this embodiment adds quantile rank regularization when calculating the weights. Here, quantile rank regularization is constructed by sorting and ranking based on the calibrated reputation values. This quantile rank regularization term... The expression is:

[0040] (3)

[0041] Formula (3) is used to constrain the weight of student mastery. Compared with the calibrated reputation value Consistency. In formula (3), Students The mastery weight; B represents the number of quantile intervals; This represents the average weight of the students in the b-th quantile block; This represents the average weight of the students in the (b+1)th quantile block; This represents the relaxation parameter, used to allow for small fluctuations in the weights; Let f(b) represent the set of students in the b-th quantile block after sorting by reputation.

[0042] In this embodiment, students are first ranked according to their calibrated reputation scores. Sort from low to high and divide into B quantile blocks. Then calculate the average weight within each interval. Next, we analyze whether the average weight of the high reputation quantile is significantly lower than that of the next level quantile. If so, the system will impose a penalty. Finally, this constraint term is added to the objective function to maintain the overall consistency between "reputation ranking and weight distribution".

[0043] (32) Establish the main objective function using the quantile equivalence regularization term. Its expression is:

[0044] (5)

[0045] Here, formula (5) is used to estimate the weights of students' mastery. With question parameters Here, formula (5) is divided into four terms: the first term is answer consistency, which ensures that the students' predicted results are as consistent as possible with the actual students' answers, and is the basic fitting part of the model; the second term is reputation constraint, which uses quantile rank regularization to require that the weight distribution be consistent with the reputation ranking, and prevent short-term abnormal scores from raising the overall evaluation; the third term is community smoothing, which uses the smoothness of the community graph to constrain the weights of adjacent students to be close, suppress gang cheating deviation, and learn the community graph. If two students are found With students For students with strong connections (such as being in the same class or having frequent interactions), the algorithm requires that their weight differences not be too large. By penalizing weight differences, overall distortion caused by small-group cheating can be prevented, which aligns with the educational principle that "students with close relationships should have similar learning levels." Specifically:

[0046] (6)

[0047] The fourth item is the question prior, which introduces priors or regularization into the learning of question parameters to avoid overfitting of the question parameters.

[0048] In this embodiment, the entire objective function combines the accuracy of individual fitting, the robustness of reputation ranking, the rationality at the community level, and the reliability of question parameters, thereby ensuring the stability of mastery assessment even in complex scenarios. In formulas (5) and (6), This indicates a quantile ordinal regularization term; Represents the loss function; Represents the Sigmoid function; Let represent the parameter vector of question j; λ represent the weight coefficient of the reputation constraint; μ represent the weight coefficient of community smoothing; ρ represent the weight coefficient of the question prior. This indicates the graph smoothing regularization term; This indicates a priori or regular expression term for the problem parameters. Representing students on the community map With students Attention weights; Students Weight of mastery; This represents the set of edges in the learning community graph.

[0049] (33) The student mastery weights and question parameters are updated iteratively by minimizing the proximal alternation, and the expression is as follows: (7)

[0050] (8)

[0051] in, Indicates step size is Proximal operators; Represents a comprehensive regularization term; This represents the weight vector of student mastery at the (t+1)th iteration. This represents the weight vector of student mastery at the t-th iteration; This represents the weight vector representing the degree of student mastery. The gradient.

[0052] (34) In actual teaching, if all students are uniformly assigned diagnostic sampling, it will increase the burden and easily cause resentment. This embodiment uses an adaptive scheduling mechanism to dynamically determine whether each student needs to be verified at a certain time step t. This embodiment uses adaptive verification scheduling for diagnostic sampling, and its expression is:

[0053] (9)

[0054] in, Students The probability of being assigned a diagnostic sample at time step t; The scheduling coefficient represents the baseline probability; The scheduling coefficient representing the reputation penalty; The scheduling coefficient representing the impact of abnormal characteristics; The scheduling coefficient represents the impact of community risk. Represents the Sigmoid function; Students The reputation value at time step t; Students Scoring of abnormal behavior characteristics at time step t (such as screen switching, abnormal time consumption for answering questions, abnormal operation behavior, etc.); Students Community risk measurement.

[0055] In this embodiment, students are taken into account. Reputation value at time step t ,student Community risk measurement ,student Score of abnormal behavior characteristics at time step t Individualized verification probabilities are generated in real time. In addition, the lower the student's reputation, the more obvious the abnormal characteristics, and the higher the community risk, the greater the verification probability, thus concentrating computing resources on key risk groups. The results of the diagnostic sampling test are recorded in the calculation of reputation value (see formula (2)), thus forming a closed-loop mechanism of "assessment, verification, and reassessment".

[0056] The fourth step involves using behavioral logs and community relationship diagrams to conduct risk assessments, obtaining individual anomaly scores and community risk levels. Details are as follows:

[0057] (41) Generating student data from behavior logs Score of abnormal behavior characteristics at time step t (Individualized verification probability).

[0058] (42) At time step t, for the student The behavioral logs extract features that can be collected and analyzed, and after robust normalization and weighting, an individual anomaly score in the interval [0,1] is obtained, the expression of which is: (10)

[0059] in, Represents the feature weight coefficients. This represents the m-th log feature function, such as "time consumption deviation", "screen switching frequency", "short-term connection", "abnormal operation", etc.

[0060] For example:

[0061] (a) Characteristics of time consumption deviation It is calculated to [0,1] based on the deviation between the time spent answering questions and the reference distribution, using robust scores. The calculation process is as follows: Let the average time spent by students on each question within the window be... Seconds. For a peer sample of the same problem or difficulty level, calculating the logarithm of the reference distribution takes time. Robust position and scale: Position (Median); Scale , MAD is a more robust measure of volatility / dispersion than standard deviation; let the robust Z-score be: , in Minimal number (avoids division by zero, default) Normalized to: , Default is 3.

[0062] (ii) Screen switching frequency characteristics The number of screen switches within the evaluation window divided by the maximum limit. Partially truncated to 1. This is the upper limit of saturation for normalizing the screen-switching frequency feature. (Number of system focus loss / screen switches within the window) Normalization is: , The default value is 10.

[0063] (iii) Short-term pairing characteristics The longest consecutive number of correct answers within the window divided by the upper limit. Normalized to [0,1] Saturation upper bound of short-term consecutive pairing feature normalization. Longest consecutive number of correct answers within a window. Normalization: , Default is 10.

[0064] (iv) Abnormal operation characteristics The system can detect abnormal behaviors such as copying and pasting answers, and pasting in browser developer mode. When these behaviors are triggered, a corresponding score will be given, ranging from [0,1]. For example, copying and pasting an answer will give 0.4 points, pasting in browser developer mode will give +0.3 points, with a maximum score of 1.

[0065] (v) Change the operating equipment or IP address, or change the characteristics. The number of times you can switch between devices or networks within the same evaluation window, divided by the maximum limit. Up to [0,1], Change the saturation limit of the operating device or IP address feature normalization. Here, , The default value is 2.

[0066] (vi) The mismatch between difficulty and corresponding time consumption The percentage of cases exhibiting both "low time consumption and high accuracy" on challenging problems was statistically analyzed and normalized to [0,1]. Here, ,in, Students A set of highly difficult questions that can be answered within time step t; student The time taken to answer question j; This represents the time threshold for question j at the corresponding difficulty level (determined by historical data or empirical quantiles). This indicates that the answer is correct; Represents the normalization factor, equal to The number of questions; This indicates an indicator function that takes the value 1 if the condition is true, and 0 otherwise.

[0067] Suppose a student is taking a 10-minute learning session within a 20-minute assessment, and the session is conducted in one window:

[0068] Steady score Screen switching 8 times: ;

[0069] Answer L=7 questions correctly in a row: ;

[0070] Suspicious pasting detected: ;

[0071] Change the operating device or IP address: ;

[0072] Difficulty mismatch: ;

[0073] Set weights ,but =0.30×0.70+0.25×0.80+0.20×0.70+0.10×0.40+0.05×0.30=0.625. If the threshold is 0.6, then the individual's abnormal score during this period is considered high-risk.

[0074] In addition, this embodiment generates community risk based on the community graph and high-risk neighbors. :

[0075] Based on the above-mentioned abnormal individuals Then, the community relationship G is analyzed based on "who one is close to", and a community risk score is given to student i at time t. ∈[0,1], for example, individual abnormality scores: ∈[0,1]; Community relationship strength (attention weight): (The higher the frequency of interaction among classmates and in the same group); Individual high-risk threshold: Let's set it to 0.6.

[0076] Then look for "high-risk neighbors" and include all the neighbors in the list. Those people were selected and weighted according to the strength of their relationship. The larger the number of high-risk neighbors, the greater the impact. Finally, community risk is calculated. Essentially, it is the weighted proportion and intensity of high-risk neighbors.

[0077] (11)

[0078] For example, students There are 4 neighbors. threshold The individual anomaly scores of the neighbors were (0.72, 0.55, 0.83, 0.20), with the first and third neighbors being the highest-risk neighbors. =0.4×0.72+0.2×0.83=0.454. If the community warning threshold is set to 0.35, then 0.454 is greater than 0.35, which will trigger additional testing or stricter measures.

[0079] The fifth step involves calculating immediate and overall benefits based on individual anomaly scores and community risk, and then using attention-weighted replication dynamics to assess the risk heatmap. In actual classroom teaching or online learning groups, abnormal answering behaviors (such as plagiarism, proxy answers, and answer sharing) often spread rapidly with increased interaction.

[0080] Among them, immediate benefits If a student's accuracy rate or score on difficult problems increases significantly in a short period of time, then The probability of a student being assigned more verification tasks by the platform will increase if their individual abnormal score is high, or if there are multiple high-risk peers in their community. This will reduce the likelihood of the student being assigned more verification tasks by the platform. Furthermore, students who frequently perform unusual actions such as screen switching, copying and pasting, or changing devices will also be flagged by the system, thereby reducing their chances of being penalized. .

[0081] (12)

[0082] Here, This represents the benefit of score improvement, with a value range of [0,1]. It is used to represent the short-term increase in difficulty-weighted accuracy, and its expression is:

[0083] (13)

[0084] in, This represents the weighted accuracy rate of student i at time t, taking into account the difficulty of the question.

[0085] in addition, The risk and pressure of students being caught and subjected to stricter measures for abnormal behavior, [0,1], is expressed as:

[0086] (14)

[0087] The formula means that an individual's abnormally high risk level, the risk level of the social circle, or the probability of being tested by the system is high. Any abnormally high level indicates that the student is under abnormal pressure, and their immediate benefits are suppressed.

[0088] in addition, The implementation cost, [0,1], is expressed as: (15)

[0089] The more frequently you switch screens, make suspicious pastes, change operating devices, or experience IP address changes, the higher the cost will be. This is the weighting coefficient, and the default values ​​are (0.4, 0.4, 0.2).

[0090] The overall benefits of this embodiment Based on the logic of immediate gains, the system enhances the impact of community relationships on individuals. If several of a student's neighbors recently exhibit abnormally high scores or high-risk behaviors, the student may be affected, thus increasing their overall gains. Conversely, if most of the student's neighbors behave normally and the platform's monitoring is strict, even if the student's individual immediate gains are positive, their overall gains may decrease due to a tightening of the overall atmosphere. The system will determine whether a certain type of abnormal behavior will spread within the group based on the level of overall gains and will display this information in a risk heatmap.

[0091] (16)

[0092] in, Indicates the trust decay coefficient; Representing students on the community map With students Attention weighting, students' neighbors in the community have abnormal behavior.

[0093] Connecting with replication dynamics (risk heatmap)

[0094] (17)

[0095] Where I represents the number of participating individuals; This represents the overall benefit of student u at time step t. The anomalous strategy is more "profitable" on student i, and the strength of the anomalous strategy increases; while If this happens, the strength of the abnormal strategy decreases.

[0096] For example, in a class: Student A1 completed many difficult questions in a short time with a high accuracy rate. However, they also exhibited many abnormal behaviors and had a high individual abnormality score, resulting in a negative immediate reward from the system. Student B1 had a low individual abnormality score and normal behavior, but three of their close classmates were identified as high-risk neighbors. Therefore, the system gave Student B1 a lower overall reward and increased the probability of B1 being sampled. Student C1 improved their accuracy rate through effort in a short time, exhibited no obvious abnormal behavior, and was not in a high-risk community. The system considered their immediate and overall rewards to be positive, thus not triggering additional verification.

[0097] This embodiment uses an attention-weighted replication dynamics model to dynamically simulate this diffusion trend and guide risk control strategies. The expression for the attention-weighted replication dynamics evaluation is as follows:

[0098] (18)

[0099] in, Students The strength of the abnormal policy at time step t+1; Students The strength of the abnormal policy at time step t; Students The overall benefit at time step t; This represents the average return of the group at time step t; Students The immediate benefit at time step t; This represents the immediate benefit of student k at time step t.

[0100] This embodiment employs replication dynamics, attention weighting, trust decay, risk assessment, and closed-loop linkage. Among these, replication dynamics is the student's... The strength of the anomalous policy at time step t It will be based on its relative returns ( - This is a dynamic phenomenon; if a student's individual benefit is higher than the class average, it is easily imitated and spreads; if it is lower than the average, it will gradually shrink. Furthermore, attention weighting is a factor for students... The overall benefit at time step t It is a weighted average of the neighbor's income. It is used to represent the strength of the connection between students. Trust decay is the trust decay coefficient. Controlling the degree of influence of neighboring benefits: the greater the attenuation, the more dependent students are on their own benefits; the smaller the attenuation, the easier it is for them to be influenced by those around them. Risk assessment is conducted by evaluating students... The strength of the anomalous policy at time step t. Evolutionary observation can determine whether abnormal strategies are spreading in the community, and adjust the diagnostic sampling schedule (Formula (8)) and path recommendation intensity accordingly. Closed-loop linkage is a verification scheduling mechanism that combines Formula (8) to achieve closed-loop prevention and control of "community evolution monitoring, high-risk early warning, enhanced verification, and updated reputation".

[0101] The fifth step involves implementing and outputting strategies based on student mastery weights, question parameters, and risk heatmaps. This includes:

[0102] (51) If a student The strength of the abnormal policy at time step t+1 or students Community risk measurement If the value exceeds the preset threshold, the scheduling coefficient for reputation penalty will be increased accordingly. Scheduling coefficients affected by abnormal characteristics Dispatch coefficients for the impact of community risks Or raise the students' At time step The probability of being assigned to a diagnostic sampling test The upper limit.

[0103] (52) Personalized practice / explanation recommendation list and difficulty ratio are made according to the weight of students’ mastery and the parameters of the questions.

[0104] (53) Using the main objective function The weighting coefficients λ for reputation constraints, μ for community smoothing, ρ for question priors, and the edge set of the learning community graph are used in the statistical analysis of diagnostic sampling. The number of quantile intervals (B) and the scheduling coefficient of the baseline probability. Dispatch coefficient for reputation penalty Scheduling coefficients affected by abnormal characteristics Dispatch coefficient of community risk impact Parameter updates and optimizations are performed. For example, in an online test, the system identified 20 students with abnormal behavior through verification sampling. Of these, 18 were confirmed as abnormal by manual review, resulting in a hit rate of 90%. If 5 normal students were also found to be misjudged, the false positive rate is 25%; if 2 actual abnormal students were not identified by the system, the false negative rate is 10%. Then, depending on the situation, the reputation constraint weight coefficient λ is reduced, and the community smoothing weight coefficient μ is increased to reduce false positives and improve detection sensitivity.

[0105] The above embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any changes made based on the design principles of the present invention, or any non-creative modifications made thereon, shall fall within the scope of protection of the present invention.

Claims

1. An online education robust learning mastery estimation and adaptive verification method, characterized in that, Comprise the following steps: Obtain the answer data information of the students participating in online education; the answer data information includes answer record, answer mark, question difficulty, behavior log and social relationship graph; According to the answer data information, the credit calibration processing is carried out, and the calibrated credit value is obtained, including: The weighted correct rate of the question difficulty is calculated according to the answer data information , and the expression is: , wherein, represents the flag of whether the student answers the question; represents the question weight function; represents the answer result of the student to the question, that is, 1 for correct answer and 0 for wrong answer; represents the reference difficulty level; represents the difficulty parameter of the question; represents the coefficient for controlling the influence strength of the question difficulty difference.​​​​​ The weighted correct rate of the question difficulty And according to the quantity of the question and the confidence coefficient, the calibrated credit value is obtained , which is expressed as: Wherein, The effective question quantity of the student ; The confidence coefficient Using the calibrated credit value to estimate the mastery degree, the student mastery weight and the question parameter are obtained, and the target value of the student mastery weight and the target value of the question parameter are obtained by optimization convergence; Using the behavior log and the social relationship graph to carry out risk assessment, obtaining individual abnormal score and social risk; According to the individual abnormal score and the social risk, the instant benefit and the comprehensive benefit are obtained, the attention weighted copy kinetics evaluation is adopted, and the risk heat map is obtained; According to the student mastery weight, the question parameter and the risk heat map, the strategy linkage and output are carried out, including: If a student The strength of the abnormal policy at time step t+1 or students Community risk measurement If the value exceeds the preset threshold, the scheduling coefficient for reputation penalty will be increased accordingly. Scheduling coefficients affected by abnormal characteristics Dispatch coefficients for the impact of community risks Or raise the students' The probability of being assigned a diagnostic sample at time step t The upper limit; Based on the student mastery weight and the question parameter, the recommendation list of personalized practice and explanation is dynamically generated, and the difficulty ratio is optimized; Utilizing the main objective function and the statistical weight coefficient λ of the reputation constraint, the weight coefficient μ of the community smoothing, the weight coefficient ρ of the question prior, the edge set of the learning community graph , the number B of the quantile interval, the scheduling coefficient of the benchmark probability , the scheduling coefficient of the reputation penalty , the scheduling coefficient of the abnormal feature influence , the scheduling coefficient of the community risk influence Parameter updating and optimization are performed.

2. The online education robust learning mastery estimation and adaptive verification method of claim 1, wherein, Using the calibrated credit value to estimate the mastery degree, the student mastery weight and the question parameter are obtained, and the target value of the student mastery weight and the target value of the question parameter are obtained by optimization convergence, including the following steps: According to the calibrated credit value, the sorting and the quantile division are carried out, and the quantile same sequence degree regularization term is constructed; Using the quantile same sequence degree regularization term to establish the main objective function; Adopting proximal alternating minimization iteration to update the student mastery weight and the question parameter.

3. The online education robust learning mastery estimation and adaptive verification method of claim 2, wherein, The quantile homogeneity regularizer The expression is: where, represents the mastery weight of a student; B represents the number of quantile bins; represents the average of student weights in the bth quantile bin; represents the average of student weights in the b+1th quantile bin; represents the relaxation parameter; represents the set of students in the bth quantile bin after ordering by reputation.

4. The online education robust learning mastery estimation and adaptive verification method according to claim 2 or 3, characterized in that, The main objective function is established by using quantile homogeneity regular term The expression is: Among them, Indicates the quantile homogeneity regular term; Indicates the loss function; Indicates the Sigmoid function; Indicates the parameter vector of question j; λ indicates the weight coefficient of reputation constraint; μ indicates the weight coefficient of community smoothing; ρ indicates the weight coefficient of question prior; Indicates the graph smoothing regular term; Indicates the question parameter prior or question parameter regular term; Indicates the attention weight of student i on the community graph; Indicates the attention weight of student i on the community graph; Indicates the attention weight of student i on the community graph; Indicates the mastery weight of student i; Indicates the mastery weight of student i; Indicates the edge set of the learning community graph.

5. The online education robust learning mastery estimation and adaptive verification method of claim 4, wherein, The proximal alternating minimization is used to iteratively update the student mastery weight and the item parameter, and the expression is as follows: wherein, denotes the proximal operator with step size ; denotes the comprehensive regularization term; denotes the student mastery weight vector at the t+1th iteration; denotes the student mastery weight vector at the tth iteration; denotes the gradient of the student mastery weight vector .

6. The online education robust learning mastery estimation and adaptive verification method of claim 5, wherein, Also include: Adaptive kernel verification scheduling is used for diagnosis sampling, and its expression is: wherein, denotes the student probability of being assigned a diagnosis sampling at time step t; denotes the scheduling coefficient of the baseline probability; denotes the scheduling coefficient of the reputation penalty; denotes the scheduling coefficient of the abnormal feature influence; denotes the scheduling coefficient of the community risk influence; denotes the Sigmoid function; denotes the student reputation value at time step t; denotes the student abnormal behavior feature score at time step t; denotes the student community risk measure.

7. The online education robust learning mastery estimation and adaptive verification method of claim 6, wherein, The attention-weighted replication dynamics evaluation, expressed as: where, denotes the student the abnormal strategy strength at time step t+1; denotes the student the abnormal strategy strength at time step t; denotes the student the comprehensive return at time step t; denotes the average return of the group at time step t; denotes the instantaneous return of student k at time step t; denotes the trust decay coefficient.

Images