Digital student model construction method and system based on multi-role large model

Abstract

The invention relates to the field of calculation models, in particular to a digital student model construction method and system based on a multi-role large model, and the method comprises the steps: dividing a student set according to features in a behavior vector, selecting a score sequence and a mastery degree sequence of students of any length, dividing the score sequence and the mastery degree sequence, predicting the mastery degree, and obtaining a mastery degree absolute error sequence; the learning cycle of the features is obtained through the grasp absolute error sequence, the fairness of the features is obtained by using the grasp absolute error sequence of the students among different student sets in the learning cycle, and the comprehensive grasp is obtained based on the learning cycle and the periodic fairness of the features, so that the model prediction accuracy is improved.

Description

Technical Field

[0001] This application relates to the field of computational models, and in particular to a method and system for constructing digital student models based on multi-role large models. Background Technology

[0002] With the rapid development of artificial intelligence technology and the increasing demand for multimodal data processing, large-scale pre-trained models have demonstrated outstanding capabilities in integrating various types of data. For example, related models are gradually evolving from relying on empirical rules to becoming more adaptive and differentiated. As one of the core mechanisms, individual behavior simulation models, by modeling cognitive paths, interaction habits, and emotional feedback, can characterize potential tendencies and predict evolutionary trends, thereby supporting key aspects such as path generation, strategy optimization, and effect measurement.

[0003] However, early modeling approaches mainly revolved around "knowledge tracking," which involved constructing static or semi-dynamic user profiles based on serialized inputs (such as answer records and operation logs) and then using these profiles for state estimation and trend inference. These methods typically assume that individual characteristics remain stable within a fixed period and use a uniform time window for state updates, but they neglect the differences in response and weight shifts of different features under different periodic patterns.

[0004] Therefore, traditional models suffer from a "bias" problem in periodic modeling: by applying equal weights or fixed decay mechanisms to all features, they fail to capture the asymmetry where some features fluctuate sharply in short periods while others evolve slowly in long periods. This mechanistic bias leads to insufficient sensitivity of the model to state transitions in cross-period predictions, thus amplifying the cumulative error effect and reducing the reliability of evolutionary trend estimation. In particular, traditional digital student models typically predict students' subject mastery based on fixed-period inputs, ignoring the bias of different features across different periodic patterns, resulting in larger prediction errors. Summary of the Invention

[0005] To improve the accuracy of model prediction, this application provides a method and system for constructing a digital student model based on a multi-role large model.

[0006] Firstly, this application provides a method for constructing a digital student model based on a multi-role large model, employing the following technical solution: Collect data on each student's native language type, learning style, role labels, personality and emotions, as well as each student's score sequence and mastery sequence in each subject within the target class; The collected native language type, learning style, role tags, and personality and emotion data are converted and concatenated using one-hot encoding to obtain the behavior vector corresponding to each student. Select any feature from the behavior vector as the target feature, and classify all students according to the target feature to obtain multiple target student sets; For all the target student sets, select any length as the target period to divide and predict the mastery of each student's score sequence and mastery sequence in each subject, and obtain the corresponding mastery absolute error sequence. The learning cycle of the target feature is obtained by using the mastery absolute error sequence; The periodic fairness corresponding to the target feature is obtained by using the absolute error sequence of students’ mastery levels among different target student sets under the learning cycle. The overall mastery level is obtained based on the learning cycle of the target characteristics and the fairness of the cycle. The predicted overall mastery level of each student in the target class at the next moment is input into the operation controller for analysis and returned to the student with corresponding subject learning and suggestion instructions.

[0007] Optionally, for all the target student sets, an arbitrary length is selected as the target period to divide and predict the mastery of each student's score sequence and mastery sequence in each subject, obtaining the corresponding mastery absolute error sequence, including: Select any length as the target period in the target student set, divide the score sequence and mastery sequence of each student in the target student set into a sliding window, and input the score sequence, mastery sequence and student behavior vector in any window as training samples into the LSTM model, and output the mastery prediction value for the next time step. The mean squared error is chosen as the model loss function. The trained LSTM model is used to traverse the mastery sequence of any student to obtain the subject mastery prediction value at each time step. The absolute error sequence of mastery is obtained by comparing the absolute difference between the predicted mastery value and the collected mastery value of the subject.

[0008] Optionally, after obtaining the absolute error sequence of mastery based on the absolute difference between the predicted subject mastery value and the collected subject mastery values, the process includes: Calculate the Euclidean distance between the absolute error sequences of mastery of any two students in any target student set, and use the standard deviation of the Euclidean distance as the first eccentricity distance of the target student set. Different lengths are selected to construct training samples and train LSTM models, where each length corresponds to one LSTM model, and the absolute error sequence of mastery for each student under each length is obtained.

[0009] Optionally, the learning period of the target feature is obtained through the mastery absolute error sequence, including: Based on the mastery prediction error of the target student set with the same feature under different lengths, the periodic evaluation score of each length is analyzed, and the learning period for each feature is selected accordingly.

[0010] Optionally, the formula for calculating the periodic evaluation score is: in, This represents the periodic evaluation score of the target feature within the target period. This represents the number of student sets divided according to the target characteristics. Indicates the first The number of students contained in a student set Indicates the first digit under the target length. The first eccentric distance of the set of students Indicates the first digit under the target length. The first student in the set The average absolute error sequence of each student's mastery level Indicates the first among all lengths The first student in the set The sum of the average of the absolute error sequences of the mastery levels of each student.

[0011] Optionally, the formula for calculating periodic fairness is: in, This indicates the periodic fairness of the target feature within the target period. This represents the number of students obtained based on the target features. Indicates the first The student set and the first The absolute value of the difference in the number of students between student sets Indicates the first The sum of the absolute values ​​of the differences in the number of students between the current student set and all student sets. Indicates the first The student set and the first The second eccentric distance between sets of students It represents the maximum Euclidean distance between the absolute error sequences of mastery levels of any two students in all sets.

[0012] Optionally, the second eccentricity distance includes: For multiple target student sets divided by target features, select the Euclidean distance between the absolute error sequences of the mastery degree of students in any two target student sets under the target feature learning cycle; The average Euclidean distance between all students in any two target student sets is used as the second eccentric distance.

[0013] Optionally, the formula for calculating the overall mastery level is: in, This indicates the overall level of understanding of the predicted next moment. Indicates the number of learning cycles. Indicates the first The predicted mastery level at the next moment after one learning cycle. Indicates the first The sum of periodic fairness of all features over a learning cycle. This represents the average of the periodic fairness across all learning cycles. It measures the relative magnitude of the sum of the periodic fairness of all features under each learning cycle pattern.

[0014] Optionally, the target feature values ​​are the same in the behavior vectors of each student within the student set.

[0015] Secondly, this application provides a digital student model construction system based on a multi-role large model, which adopts the following technical solution: The data collection module is used to collect information on each student's native language type, learning style, role tags, personality and emotions, as well as each student's score sequence and mastery sequence in each subject within the target class. The splicing module is used to perform one-hot encoding conversion and splicing on the collected native language type, learning style, role tags, personality and emotion to obtain the behavior vector corresponding to each student; The classification module is used to select any one feature from the behavior vector as the target feature, and classify all students according to the target feature to obtain multiple target student sets; The period acquisition module is used to select any length as the target period for all the target student sets; The error sequence prediction module divides and predicts the mastery of each student's score sequence and mastery sequence for each subject, and obtains the corresponding absolute error sequence of mastery. The periodic fairness calculation module is used to obtain the learning period of the target feature through the mastery absolute error sequence, and to obtain the periodic fairness corresponding to the target feature by using the mastery absolute error sequence of students among different target student sets under the learning period. The comprehensive mastery analysis module is used to obtain the comprehensive mastery based on the learning cycle of the target characteristics and the fairness of the cycle, and input the predicted comprehensive mastery of each student in the target class at the next moment into the operation controller for analysis and return to the student the corresponding subject learning and suggestion instructions.

[0016] This application has the following technical advantages: 1. Compared with existing technologies that predict digital students’ subject mastery based on fixed periods, this application can obtain the potential learning period of each feature by analyzing the relative differences in the absolute error of mastery among student sets divided by a single feature under different periods. It can identify the periodic learning influence features of different features and perform multi-period integrated prediction to improve the accuracy of model prediction.

[0017] 2. Compared with existing technologies that predict subject mastery based on the initial behavioral vectors of digital students, this application obtains the fairness of features by analyzing the differences in mastery error predicted by different features under different learning cycles. This can balance the subject bias phenomenon of traditional models for some features and improve the fairness and accuracy of model prediction. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating steps S1-S7 of a digital student model construction method based on a multi-role large model in this application. Detailed Implementation

[0019] This application discloses a method for constructing a digital student model based on a multi-role large model, referring to... Figure 1 ,include: S1. Collect information on each student's native language type, learning style, role labels, personality and emotions, as well as each student's score sequence and mastery sequence in each subject within the target class.

[0020] It should be noted that for multiple students within the target class, data on each student's native language type, learning style, role label, and personality / emotional state are collected via an online platform or mobile app. The student's score and mastery sequence for each subject are also collected weekly. The collected native language type, learning style, role label, and personality / emotional state are then transformed and concatenated using one-hot encoding to obtain a 4-dimensional behavioral vector containing these four features. For example, a student might be labeled as "Native Language Type: Chinese, Learning Style: Proactive, Role Label: Excellent, Personality / Emotional State: Cheerful".

[0021] For a student's subject, there is a score sequence and a mastery sequence, and the time scale of the elements in the two sequences is the same.

[0022] As an example, native language type includes, but is not limited to, Chinese, English, French, German, etc.; learning style includes, but is not limited to, proactive, reflective, intuitive, holistic, etc.; role label includes, but is not limited to, excellent, average, poor, etc.; personality and emotion include, but is not limited to, cheerful, extroverted, introverted, confident, diligent, etc.

[0023] This application specifically relates to a method for predicting a student's mastery of a subject in a digital student model. This method is applicable to any subject, and subsequent data processing steps only involve data from a single subject. A corresponding mastery prediction model can be obtained for each subject when this method is applied.

[0024] S2. The collected native language type, learning style, role label, personality and emotion are converted and concatenated using one-hot encoding to obtain the behavior vector corresponding to each student.

[0025] It should be noted that one-hot encoding transforms non-numerical categorical features into binary vectors containing only 0s and 1s. Each feature value corresponds to a "position" within the vector; for a given student, the feature value is marked as 1 only at the corresponding position, with 0s marked at the other positions.

[0026] As an example, suppose there are 3 students in a class, and the collected feature information is as follows: Student A, Student B, and Student C; their corresponding native languages ​​are Chinese, English, and French, respectively; their corresponding learning styles are proactive, reflective, and holistic, respectively; their corresponding role labels are excellent, average, and poor, respectively; and their corresponding personality traits are outgoing, introverted, and confident, respectively.

[0027] Furthermore, based on the feature information collected from the above three students, one-hot encoding of individual features is performed. First, encoding positions are assigned to all possible values ​​of each feature (in practical applications, all potential values ​​must be included): Native language type (values: Chinese, English, French): encoding length 3, Chinese: [1,0,0], English: [0,1,0], French: [0,0,1]; Learning style (values: active, reflective, intuitive, global): encoding length 4, active: [1,0,0,1], [0], Reflection: [0,1,0,0], Global: [0,0,0,1]; Role Tag (Values: Excellent, Average, Poor): Encoding length 3, Excellent: [1,0,0], Average: [0,1,0], Poor: [0,0,1]; Personality and Emotion (Values: Cheerful, Extroverted, Introverted, Confident, Diligent): Encoding length 5, Cheerful: [1,0,0,0,0], Introverted: [0,0,1,0,0], Confident: [0,0,0,1,0].

[0028] Then, the above feature codes are concatenated into corresponding behavior vectors. The four types of feature codes for each student are concatenated in sequence to form a behavior vector of uniform dimension (total length = 3 + 4 + 3 + 5 = 15): Student A's behavior vector: [1,0,0] (Chinese) + [1,0,0,0] (proactive) + [1,0,0] (excellent) + [1,0,0,0,0] (cheerful) = [1,0,0,1,0,0,0,1,0,0,1,0,0,0,0]; Student B's behavior vector: [0,1, [0] (English) + [0,1,0,0] (Reflection) + [0,1,0] (Medium) + [0,0,1,0,0] (Introversion) = [0,1,0,0,1,0,0,0,1,0,0,0,1,0,0]; Student C's behavior vector: [0,0,1] (French) + [0,0,0,1] (Global) + [0,0,1] (Poor) + [0,0,0,1,0] (Confidence) = [0,0,1,0,0,0,1,0,0,1,0,0,0,1,0,0,0,1,0].

[0029] The encoding length is determined by all potential values ​​of the feature, rather than the number currently collected. The concatenation order is fixed (e.g., native language → learning style → role label → personality and emotion), ensuring that the behavioral vector dimension of all students is consistent. One-hot encoding can avoid the "priority bias" of non-numerical features, allowing the model to treat the different values ​​of each feature equally.

[0030] S3. Select any feature from the behavior vector as the target feature, and classify all students according to the target feature to obtain multiple target student sets.

[0031] As an example, using "native language type" as the target feature, the potential values ​​for "native language type" are: Chinese, English, and French. The classification logic is as follows: students with the same "native language type" are grouped into one set, resulting in three target student sets: Set 1 (native language = Chinese): {001, 003}, where all students in this set have "native language type" of Chinese, and other features may differ; Set 2 (native language = English): {002, 005}, where all students in this set have "native language type" of English, and other features may differ; Set 3 (native language = French): {004}, where only student 004 has French as their native language, forming a separate set.

[0032] For example, the potential values ​​for "learning style" are: proactive, reflective, global, and intuitive (only the first three appear in this example). The classification logic is as follows: students with the same "learning style value" are grouped into one set, resulting in three target student sets: Set 1 (learning style = proactive): {001, 005}, where students 001 and 005 both have the "proactive" learning style; Set 2 (learning style = reflective): {002, 004}, where students 002 and 004 both have the "reflective" learning style; and Set 3 (learning style = global): {003}, where only student 003 has the "global" learning style.

[0033] Therefore, based on the above classification, we can subsequently analyze the impact of this feature on the learning cycle and mastery error for each "target student set".

[0034] S4. For all target student sets, select any length as the target period to divide and predict the mastery of each student's score sequence and mastery sequence in each subject, and obtain the corresponding mastery absolute error sequence.

[0035] As an example, an arbitrary length is selected as the target period in the target student set. The score sequence and mastery sequence of each student in the target student set are divided into sliding windows. The score sequence, mastery sequence and student behavior vector within any window are used as training samples and input into the LSTM model. The output is the mastery prediction value for the next time step. The mean squared error is selected as the model loss function. The trained LSTM model is used to traverse the mastery sequence of any student to obtain the subject mastery prediction value at each time step. The absolute error sequence of mastery is obtained based on the absolute value of the difference between the subject mastery prediction value and the collected subject mastery.

[0036] In one embodiment, for all student sets, a sliding window is used to divide the score sequence and mastery sequence of each student in the student set into segments of any length as the target period. Target subject: Mathematics; data collection frequency: once a week for a total of 8 weeks; target student set: set 1 (learning style = proactive) selected above as the target feature, containing 2 students: student 001 and student 005. Among them, the math score sequence of student 001 (weekly score, full score 100): [65,72,70,78,82,80,85,88], and the math mastery sequence (0-1,1 indicates complete mastery): [0.65,0.72,0.70,0.78,0.82,0.80,0.85,0.88]; the math score sequence of student 005 (weekly score, full score 100): [58,65,62,70,75,73,78,82], and the math mastery sequence (0-1,1 indicates complete mastery): [0.58,0.65,0.62,0.70,0.75,0.73,0.78,0.82].

[0037] The target period length is set to "3 weeks". In practice, "any length" can be tested for 2 weeks, 4 weeks, etc. This application takes 3 weeks as an example. The sliding window is divided as follows: the "target period length" is used as the window. Each time, the window slides for 1 week to cut the sequence and obtain training samples. For example, if the 3-week window is cut into 8 weeks of data, 6 windows can be obtained.

[0038] Furthermore, the score sequence, mastery sequence, and student behavior vector within a window are input as training samples into an LSTM (Long Short-Term Memory) model. The output is the predicted mastery value for the next time step, and the mean squared error is chosen as the model loss function. The trained LSTM model is used to iterate through the mastery sequence of any student to obtain the predicted subject mastery value at each time step. The absolute mastery error sequence is obtained based on the absolute value of the difference between the predicted subject mastery value and the collected subject mastery value. When the model input data is insufficient, the window length is padded using nearest neighbor interpolation. Similarly, the absolute mastery error sequence for each student is obtained.

[0039] In one embodiment, the performance / mastery sequence is divided using a target period (sliding window method). The "performance sequence" and "mastery sequence" of each student are divided into 3-week windows. Each window corresponds to "input features" and "prediction target". Input features: 3-week performance sequence + 3-week mastery sequence + student behavior vector within the window. Prediction target: mastery of the week following the window. For example, the prediction target of the 1st to 3rd week window is the mastery of the 4th week.

[0040] Specifically, taking student 001 as an example, the segmentation results are as follows: Window 1 corresponds to the following in-window performance sequence (input): [65, 72, 70] (weeks 1-3), and the corresponding in-window mastery sequence (input): [0.65, 0.72, 0.70]; Window 2 corresponds to the following in-window performance sequence (input): [72, 70, 78] (weeks 2-4), and the corresponding in-window mastery sequence (input): [0.72, 0.70, 0.78]; Window 3 corresponds to the following in-window performance sequence (input): [70, 78, 82] (weeks 3-5), and the corresponding in-window mastery sequence (input): [0.70, ... [0.78, 0.82]; the input sequence of grades in window 4 is [78, 82, 80] (weeks 4-6), and the input sequence of mastery in window 5 is [0.78, 0.82, 0.80]; the input sequence of grades in window 5 is [82, 80, 85] (weeks 5-7), and the input sequence of mastery in window 6 is [0.82, 0.80, 0.85]; the input sequence of grades in window 6 is [80, 85, 88] (weeks 6-8), and the input sequence of mastery in window 6 is [0.80, 0.85, 0.88]. Similarly, the segmentation logic for student 005 is exactly the same, only replaced with its own grade / mastery sequence, such as the input of [58, 65, 62] and [0.58, 0.65, 0.62] for window 1, and the prediction target is 0.70 for week 4.

[0041] Then, the LSTM model is trained and the mastery level is predicted. Model training: The first 5 windows of data (windows with real prediction targets) of all target student sets (students 001, 005) are used as training samples and input into the LSTM model: Input layer: 3-week performance sequence within the window + 3-week mastery level sequence + student behavior vector (15-dimensional vector after one-hot encoding), Output layer: predicted "mastery level 1 week after the window", Loss function: mean squared error. Model Prediction: Using the trained LSTM model, output "predicted mastery" for all windows (including the 6th window with no true value). Assume the model prediction results are as follows (only core data is shown): Student 001: Window 1 predicted mastery 0.76, Window 2 predicted mastery 0.80, Window 3 predicted mastery 0.79, Window 4 predicted mastery 0.84, Window 5 predicted mastery 0.87, Window 6 predicted mastery 0.90; Student 005: Window 1 predicted mastery 0.68, Window 2 predicted mastery 0.64, Window 3 predicted mastery 0.72, Window 4 predicted mastery 0.74, Window 5 predicted mastery 0.79, Window 6 predicted mastery 0.83.

[0042] Furthermore, the absolute error sequence of mastery is calculated, where absolute error of mastery = |predicted mastery - actual mastery|. For each student, the errors of all windows are arranged in chronological order to obtain the "absolute error sequence of mastery". Window errors without true values ​​are not calculated. Student 001: Window 1 error: |0.76-0.78|=0.02, Window 2 error: |0.80-0.82|=0.02, Window 3 error: |0.79-0.80|=0.01, Window 4 error: |0.84-0.85|=0.01, Window 5 error: |0.87-0.88|=0.01, Student 001's absolute error sequence of mastery: [0.02,0.02,0.01,0.01,0.01]; Student 005: Window 1 error: |0.68-0.70|=0.02, Window 2 error: |0.64-0.62|=0.02, Window 3 error: |0.72-0.75|=0.03, Window 4 error: |0.74-0.73|=0.01, Window 5 error: |0.79-0.82|=0.03, Student 005's absolute error sequence of mastery: [0.02,0.02,0.03,0.01,0.03].

[0043] S5. Obtain the learning cycle of the target feature through the mastery absolute error sequence.

[0044] S6. Using the absolute error sequence of mastery among students in different target student sets under the learning cycle, obtain the cycle fairness corresponding to the target feature.

[0045] It should be noted that in the construction of the digital student model, the model predicts the student's mastery level in real time by inputting the student's behavior vector, subject scores, and mastery level, and assigns relevant exercises based on the changing trend of mastery level. However, because different features in the behavior vector have varying periods of influence on the digital student's learning status, and the model exhibits a certain degree of bias towards different features, existing technologies ignore the potential subject learning periods of different behavioral features and the fairness of the features. This results in the model favoring a particular feature to dominate the direction of mastery level prediction, leading to significant deviations in the predicted subject mastery levels.

[0046] Therefore, this application uses the mastery prediction error analysis of a student set with the same feature at different lengths to analyze the periodic evaluation scores for each length, and thereby selects the learning period for each target feature.

[0047] In one embodiment, the formula for calculating the periodic evaluation score is: in, This represents the periodic evaluation score of the target feature within the target period. This represents the number of student sets divided according to the target characteristics. Indicates the first The number of students contained in a student set Indicates the first digit under the target length. The first eccentric distance of the set of students Indicates the first digit under the target length. The first student in the set The average absolute error sequence of each student's mastery level Indicates the first among all lengths The first student in the set The sum of the average absolute error sequences of each student's mastery level. Similarly, the periodic evaluation scores of the target feature at any length are obtained, and the length with the largest periodic evaluation score is selected as the learning period of the target feature.

[0048] As an example, for any target student set, calculate the Euclidean distance between the mastery absolute error sequences of any two students within the target student set, and calculate the standard deviation of the obtained Euclidean distance as the first eccentricity distance of the target student set. Select different lengths to construct training samples and train an LSTM model, with each length corresponding to one LSTM model, to obtain the mastery absolute error sequence for each student under each length.

[0049] It should be noted that the first eccentricity distance It measures the similarity of mastery error among different students within a student set. The smaller the first eccentricity distance, the more similar the trend of mastery prediction error among students within the student set obtained from the same type of features, the better the prediction effect, and the higher the evaluation score. It measures the relative magnitude of the absolute error of students' mastery at the target length to the absolute error of mastery at other lengths. The larger the ratio, the smaller the prediction error of mastery at the target length, the better the prediction effect, and the more likely it is to be the learning cycle of the target feature.

[0050] Furthermore, for multiple target student sets segmented by target features, the Euclidean distance between any two target student sets is selected as the absolute error sequence of mastery levels of students under the target feature learning cycle. The average Euclidean distance between all students in any two target student sets is used as the second eccentricity distance. In the prediction of student subject mastery in the digital student model, since the mastery of different subjects is affected by the degree of eccentricity corresponding to different behavioral features, the prediction relies on single or partial features. Differences in eccentricity features lead to large errors in the predicted mastery values ​​of different students. For example, the model assumes that students whose native language is English have a higher mastery of English than students who are Chinese or French, and that high-achieving students have a higher mastery of Chinese than lower-achieving students, ignoring the influence of different students' own learning abilities and actual mastery of history.

[0051] Therefore, this application analyzes the periodic fairness of each behavioral feature for mastery prediction by examining the differences in the absolute error sequence of mastery among student sets divided based on target features and the differences in the number of students between sets during the learning cycle.

[0052] In one embodiment, the formula for calculating periodic fairness is: in, This indicates the periodic fairness of the target feature within the target period. This represents the number of students obtained based on the target features. Indicates the first The student set and the first The absolute value of the difference in the number of students between student sets Indicates the first The sum of the absolute values ​​of the differences in the number of students between the current student set and all student sets. Indicates the first The student set and the first The second eccentric distance between sets of students It represents the maximum Euclidean distance between the absolute error sequences of mastery levels of any two students in all sets.

[0053] In one embodiment, the formula for calculating periodic fairness can also be: in, This indicates the periodic fairness of the target feature within the target period. This represents the number of students obtained in step a based on the target features. Indicates the first The number of students in a student set This represents the average number of students in all student sets. Indicates the first The first eccentricity distance within the set of students and the second The absolute difference between the first eccentric distances of students within a set of students. Indicates the first The student set and the first The second eccentric distance between sets of students It represents the maximum Euclidean distance between the absolute error sequences of mastery levels of any two students in the set of all students.

[0054] It should be noted that the difference in the number of students can reflect the difference in the distribution of different feature labels of the same target feature in the digital student model. The greater the difference between student sets, the more biased the two student sets are in terms of student size in terms of target feature. The second bias distance measures the difference in the absolute error sequence of mastery between two student sets. The larger the second bias distance, the greater the difference in absolute error of mastery, the more serious the bias phenomenon, and the lower the periodic fairness. The first bias distance can measure the relative difference in the absolute error of mastery of different students in the two sets.

[0055] Therefore, compared with existing technologies that directly use behavior vectors with fixed periods as input to the mastery prediction model, this application can analyze the potential bias patterns of different features under different learning cycle modes, obtain the cycle fairness of each feature based on the difference in absolute mastery error and the difference in student set size, and improve the prediction accuracy of the prediction model. Thus, the cycle fairness of each feature under any learning cycle is obtained.

[0056] S7. Based on the learning cycle and cycle fairness of the target features, obtain the comprehensive mastery level, and input the predicted comprehensive mastery level of each student in the target class at the next moment into the operation controller for analysis and return to the student the corresponding subject learning and suggestion instructions.

[0057] It should be noted that, for a learning cycle, a window is divided according to the learning cycle length to obtain the behavior vector. The elements in the behavior vector are multiplied by the corresponding cycle fairness of the feature within that learning cycle to obtain the fair behavior vector. The fair behavior vector, performance sequence, and mastery sequence within a window are considered as one sample. Similarly, multiple training samples are obtained, and the LSTM model is trained a second time. The model output is the mastery prediction value for the next time step, and the mean squared error is used as the loss function. Furthermore, for a learning cycle, the fair behavior vector, performance sequence, and mastery sequence of the real-time window are obtained according to the learning cycle length and input into the trained LSTM model, outputting the mastery prediction value for the next time step.

[0058] As an example, the trends in student mastery and cycle fairness differ under different learning cycle models, leading to relative biases in prediction. Therefore, this application integrates the relative magnitudes of the predicted mastery values ​​and cycle fairness under different learning cycle models to obtain a comprehensive mastery score. The formula for calculating the comprehensive mastery score is as follows: in, This indicates the overall level of understanding of the predicted next moment. Indicates the number of learning cycles. Indicates the first The predicted mastery level at the next moment after one learning cycle. Indicates the first The sum of periodic fairness of all features over a learning cycle. This represents the average of the periodic fairness across all learning cycles. It measures the relative magnitude of the sum of the periodic fairness of all features under each learning cycle pattern.

[0059] At this point, we obtain the overall mastery level of each student in the target class as predicted for the next moment.

[0060] It should be further explained that the core logic of the operation controller in this application can be: first, map the "overall mastery level" to the "mastery level", and then combine the weaknesses of the knowledge modules of the subject (mathematics) (through the analysis of the score sequence) to generate "targeted learning content + intensity suggestions".

[0061] As an example, the preset mastery level classification rules are as follows: 0.90 and above is excellent, indicating mastery of more than 85% of the knowledge points at this stage, and further improvement is possible; 0.80-0.89 is good, indicating mastery of 70%-85% of the knowledge points, and weak modules need to be strengthened; 0.70-0.79 is average, indicating mastery of only 50%-70% of the knowledge points, and a systematic review of the basics is needed; below 0.70 is weak, indicating mastery of less than 50% of the knowledge points, and the core content needs to be relearned.

[0062] In one embodiment, learning suggestions are generated for student 001. Basic analysis: Overall mastery: 0.90 (excellent level); Weaknesses in performance sequence: Slight fluctuations in scores during weeks 3 and 6 (70 points and 80 points respectively), corresponding to lower accuracy rates in the math modules "Solving Quadratic Equations" and "Auxiliary Lines in Geometric Proofs". The operation controller outputs the following instructions: Math learning suggestions (student 001): Learning content: Prioritize mastering "Advanced Applications of Quadratic Equations" (e.g., profit problems) and "Constructing Complex Auxiliary Lines in Geometric Proofs" (e.g., the method of doubling the median); Learning intensity: 15 minutes of focused practice daily (2 medium-difficulty questions + 1 challenging question), and completing one comprehensive extension exam on weekends (emphasizing the final, most difficult question); Resource recommendation: Use the online platform's "Advanced Special Courses" (quadratic equation application chapter), and record problem-solving strategies in a corresponding error notebook; Feedback requirements: Submit focused practice results every Friday, and the system will adjust the learning difficulty for the following week based on the accuracy rate.

[0063] Based on the learning suggestions generated for Student 005, the basic analysis is as follows: Overall mastery: 0.86 (good level); Weaknesses in performance sequence: Lower scores in weeks 1 and 3 (58 points and 62 points respectively), corresponding to a weak foundation in the modules of "Solving Linear Equations in One Variable" and "Triangle Congruence Criteria". The operation controller outputs the following instructions: Mathematics Learning Suggestions (Student 005): Learning Content: Prioritize consolidating "Standardized Steps for Solving Linear Equations in One Variable" (common mistakes in removing denominators and parentheses) and "Distinguishing Triangle Congruence Criteria Theorems" (e.g., the difference between SSA and SAS); Learning Intensity: 20 minutes of basic practice daily (3 basic questions + 2 medium-difficulty questions), and complete one set of basic consolidation exercises on weekends (focusing on calculation and proof standardization); Resource Recommendation: Use the online platform's "Basic Micro-Lessons" (summary of common mistakes in linear equations in one variable), along with the accompanying "Revisit Incorrect Questions" function to strengthen weak points; Feedback Requirements: Submit basic practice results daily, and the system will push similar questions for reinforcement training in real time until the accuracy rate reaches over 90%.

[0064] This application also discloses a digital student model construction system based on a multi-role large model, including: The data collection module is used to collect information on each student's native language type, learning style, role tags, personality and emotions, as well as each student's score sequence and mastery sequence in each subject within the target class. The splicing module is used to perform one-hot encoding conversion and splicing on the collected native language type, learning style, role tags, personality and emotion to obtain the behavior vector corresponding to each student; The classification module is used to select any feature from the behavior vector as the target feature, and classify all students according to the target feature to obtain multiple target student sets; The period acquisition module is used to select any length as the target period for all target student sets; The error sequence prediction module divides and predicts the mastery of each student's score sequence and mastery sequence for each subject, and obtains the corresponding absolute error sequence of mastery. The periodic fairness calculation module is used to obtain the learning period of the target feature through the mastery absolute error sequence, and to obtain the periodic fairness corresponding to the target feature by using the mastery absolute error sequence of students among different target student sets under the learning period. The comprehensive mastery analysis module is used to obtain the comprehensive mastery based on the learning cycle and cycle fairness of the target characteristics, and inputs the predicted comprehensive mastery of each student in the target class at the next moment into the operation controller for analysis and returns the corresponding subject learning and suggestion instructions to the students.

[0065] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A method for constructing a digital student model based on a multi-role large model, characterized in that, include: Collect data on each student's native language type, learning style, role labels, personality and emotions, as well as each student's score sequence and mastery sequence in each subject within the target class; The collected native language type, learning style, role tags, and personality and emotion data are converted and concatenated using one-hot encoding to obtain the behavior vector corresponding to each student. Select any feature from the behavior vector as the target feature, and classify all students according to the target feature to obtain multiple target student sets; For all the target student sets, select any length as the target period to divide and predict the mastery of each student's score sequence and mastery sequence in each subject, and obtain the corresponding mastery absolute error sequence. The learning cycle of the target feature is obtained by using the mastery absolute error sequence; The periodic fairness corresponding to the target feature is obtained by using the absolute error sequence of students’ mastery levels among different target student sets under the learning cycle. The overall mastery level is obtained based on the learning cycle of the target characteristics and the fairness of the cycle. The predicted overall mastery level of each student in the target class at the next moment is input into the operation controller for analysis and returned to the student with corresponding subject learning and suggestion instructions.

2. The method for constructing a digital student model based on a multi-role large model according to claim 1, characterized in that, For all the target student sets, select any length as the target period to divide and predict the mastery of each student's score sequence and mastery sequence for each subject, and obtain the corresponding mastery absolute error sequence, including: Select any length as the target period in the target student set, divide the score sequence and mastery sequence of each student in the target student set into a sliding window, and input the score sequence, mastery sequence and student behavior vector in any window as training samples into the LSTM model, and output the mastery prediction value for the next time step. The mean squared error is chosen as the model loss function. The trained LSTM model is used to traverse the mastery sequence of any student to obtain the subject mastery prediction value at each time step. The absolute error sequence of mastery is obtained by comparing the absolute difference between the predicted mastery value and the collected mastery value of the subject.

3. The method for constructing a digital student model based on a multi-role large model according to claim 2, characterized in that, After obtaining the absolute error sequence of subject mastery based on the absolute difference between the predicted subject mastery value and the collected subject mastery values, the process includes: Calculate the Euclidean distance between the absolute error sequences of mastery of any two students in any target student set, and use the standard deviation of the Euclidean distance as the first eccentricity distance of the target student set. Different lengths are selected to construct training samples and train LSTM models, where each length corresponds to one LSTM model, and the absolute error sequence of mastery for each student under each length is obtained.

4. The method for constructing a digital student model based on a multi-role large model according to claim 1, characterized in that, The learning period of the target feature is obtained through the mastery absolute error sequence, including: Based on the mastery prediction error of the target student set with the same feature under different lengths, the periodic evaluation score of each length is analyzed, and the learning period for each target feature is selected accordingly.

5. The method for constructing a digital student model based on a multi-role large model according to claim 4, characterized in that, The formula for calculating the periodic evaluation score is as follows: in, This represents the periodic evaluation score of the target feature within the target period. This represents the number of student sets divided according to the target characteristics. Indicates the first The number of students contained in a student set Indicates the first digit under the target length. The first eccentric distance of the set of students Indicates the first digit under the target length. The first student in the set The average absolute error sequence of each student's mastery level Indicates the first among all lengths The first student in the set The sum of the average of the absolute error sequences of the mastery levels of each student.

6. The method for constructing a digital student model based on a multi-role large model according to claim 1, characterized in that, The formula for calculating periodic fairness is: in, This indicates the periodic fairness of the target feature within the target period. This represents the number of students obtained based on the target features. Indicates the first The student set and the first The absolute value of the difference in the number of students between student sets Indicates the first The sum of the absolute values ​​of the differences in the number of students between the current student set and all student sets. Indicates the first The student set and the first The second eccentric distance between sets of students It represents the maximum Euclidean distance between the absolute error sequences of mastery levels of any two students in all sets.

7. The method for constructing a digital student model based on a multi-role large model according to claim 6, characterized in that, The second eccentricity includes: For multiple target student sets divided by target features, select the Euclidean distance between the absolute error sequences of the mastery degree of students in any two target student sets under the target feature learning cycle; The average Euclidean distance between all students in any two target student sets is used as the second eccentric distance.

8. The method for constructing a digital student model based on a multi-role large model according to claim 1, characterized in that, The formula for calculating the overall mastery level is: in, This indicates the overall level of understanding of the predicted next moment. Indicates the number of learning cycles. Indicates the first The predicted mastery level at the next moment after one learning cycle. Indicates the first The sum of periodic fairness of all features over a learning cycle. This represents the average of the periodic fairness across all learning cycles. It measures the relative magnitude of the sum of the periodic fairness of all features under each learning cycle pattern.

9. The method for constructing a digital student model based on a multi-role large model according to claim 1, characterized in that, The target feature values ​​are the same in the behavior vectors of each student within the student set.

10. A digital student model construction system based on a multi-role large model, characterized in that, include: The data collection module is used to collect information on each student's native language type, learning style, role tags, personality and emotions, as well as each student's score sequence and mastery sequence in each subject within the target class. The splicing module is used to perform one-hot encoding conversion and splicing on the collected native language type, learning style, role tags, personality and emotion to obtain the behavior vector corresponding to each student; The classification module is used to select any one feature from the behavior vector as the target feature, and classify all students according to the target feature to obtain multiple target student sets; The period acquisition module is used to select any length as the target period for all the target student sets; The error sequence prediction module divides and predicts the mastery of each student's score sequence and mastery sequence for each subject, and obtains the corresponding absolute error sequence of mastery. The periodic fairness calculation module is used to obtain the learning period of the target feature through the mastery absolute error sequence, and to obtain the periodic fairness corresponding to the target feature by using the mastery absolute error sequence of students among different target student sets under the learning period. The comprehensive mastery analysis module is used to obtain the comprehensive mastery based on the learning cycle of the target characteristics and the fairness of the cycle, and input the predicted comprehensive mastery of each student in the target class at the next moment into the operation controller for analysis and return to the student the corresponding subject learning and suggestion instructions.

Images