Interpretable knowledge tracking method with time position codes

By introducing time position coding and rotation position coding in the knowledge tracking model, the attention calculation is optimized, and the problem of not taking into account time factors and low interpretability of the existing model is solved, and higher prediction accuracy and interpretability are achieved, and enhanced The robustness and adaptability of the model.

CN118967389AInactive Publication Date: 2024-11-15GUANGDONG UNIV OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411088495.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-09
Publication Date
2024-11-15
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The existing knowledge tracking model based on attention structure fails to effectively consider time factors when analyzing students' learning behavior, resulting in low interpretability, single model, and homogeneity problems, making it difficult to provide global and local interpretability.

Method used

Using an interpretability knowledge tracking method with time position encoding, the time attention knowledge tracking model (TAKT) is constructed, and the rotation position encoding and multi-head attention mechanism are introduced, combined with the information encoding module and the cognitive ability prediction module, the attention calculation is optimized. , enhance the global and local interpretability of the model.

Benefits of technology

Improve the prediction accuracy and interpretability of the model, better focus on recent behaviors, reduce the impact of homogeneity, enhance the robustness and adaptability of the model, and provide a clear analysis of the contribution of behavior to the predicted results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118967389A_ABST
    Figure CN118967389A_ABST
Patent Text Reader

Abstract

The invention belongs to the technical field of knowledge tracking, and particularly relates to an interpretability knowledge tracking method with time position coding, which comprises the following specific steps: constructing a training data set, constructing a time attention knowledge tracking model TAKT, training the model and testing, and performing interpretability practical application of the model. Time-based location coding is introduced. By analyzing the occurrence time of historical behaviors, the model can allocate small attention to long behaviors and allocate high attention to recent behaviors, so that recent behaviors are more effectively concerned, and the prediction accuracy is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of knowledge tracing, and in particular to an interpretable knowledge tracing method with time position coding. Background Art

[0002] With the development of educational technology, knowledge tracing (KT) has become an important research field, which aims to predict and understand students' mastery of knowledge components (KCs) by analyzing their historical learning interaction data. In recent years, a variety of deep learning-based models have been developed and applied to KT to improve the understanding of students' learning process and the accuracy of prediction.

[0003] The current mainstream structures can be divided into the following four categories: Deep Sequential Models, Memory Augmented Models, Adversarial Based Models, and Attention Based Models. These models are highly complex deep learning models that typically contain a large number of parameters and multiple layers of nonlinear transformations. This design enables the model to capture complex patterns and relationships, and often has extremely high accuracy in predicting student abilities, but it also means that their decision-making process is difficult to represent in a simple, human-understandable way. Therefore, although these models may perform well in terms of predictive performance, they have the following disadvantages:

[0004] (1) Most knowledge tracking task data contain the time when behaviors occur. However, existing attention-based models do not consider the time factor between behaviors when analyzing attention. Instead, they simply encode the behaviors according to the order in which they occur, ignoring the fact that the intervals between behaviors are not equal.

[0005] (2) Existing models have a single explanatory angle: Models often only explain prediction results or model preferences from a single angle, without taking into account both overall and local interpretability, resulting in low practical value of these models.

[0006] (3) Homogeneity problem: Currently, it can only be explained at the knowledge point level. When students answer questions on the same knowledge point, the knowledge increments they gain are the same, and the answers to questions with the same knowledge point will be the same.

[0007] Therefore, an interpretable knowledge tracing method with temporal position encoding is invented to integrate the interpretable method into the deep learning knowledge tracing system. Summary of the invention

[0008] In view of the above-mentioned problems and / or existing problems in the existing interpretable knowledge tracking methods with time position coding, the present invention is proposed.

[0009] Therefore, the purpose of the present invention is to provide an interpretable knowledge tracking method with time position coding, which can solve the above-mentioned existing problems.

[0010] To solve the above technical problems, according to one aspect of the present invention, the present invention provides the following technical solutions:

[0011] The explainable knowledge tracking method with time position encoding includes the following specific steps:

[0012] S1: build training dataset;

[0013] S2: Construct the temporal attention knowledge tracking model TAKT;

[0014] S3: train the model and test it;

[0015] S4: Practical application of model interpretability.

[0016] As a preferred solution of the interpretable knowledge tracking method with time position coding described in the present invention, the specific steps of S1 are as follows:

[0017] S11: For the data of the student's answer record exceeding the set length, it is segmented and only the part that meets the length requirement is retained. For the data of the student's answer record less than the target length, a zero padding operation is performed, which is equivalent to recording the state of the student when there is no behavior, so as to unify the length of the data sequence;

[0018] S12: After unifying the data length, classify and organize the information involved in the students' answer records.

[0019] As a preferred solution of the explainable knowledge tracking method with time position coding described in the present invention, the classification and organization includes: student information, test answer information, test question records before answering the target test question, and whether the student can answer the target test question correctly.

[0020] As a preferred solution of the interpretable knowledge tracking method with time position coding described in the present invention, the specific steps of S2 are as follows:

[0021] S21: The overall structure of the model consists of an information encoding module, a temporal attention analysis module, and a cognitive ability prediction module;

[0022] S22: Model interpretability analysis.

[0023] As a preferred solution of the interpretable knowledge tracking method with time position coding described in the present invention, the information encoding module encodes various information types into vectors, specifically including: converting student information into vectors using Embedding for analyzing students' cognitive abilities, converting test answer information into vectors using Embedding for distinguishing each test question, and converting test question records before answering the target test question into vectors through Embedding to analyze students' cognitive abilities through their historical test answer records.

[0024] As a preferred solution of the explainable knowledge tracking method with time position coding described in the present invention, the time attention analysis module is based on the multi-head attention mechanism, integrates the time factor, and through the optimized relative position coding, the time when the behavior occurs can affect the attention distribution, so that when analyzing the cognitive ability of students, priority can be given to the most recent behavior.

[0025] As a preferred solution of the interpretable knowledge tracking method with time position encoding described in the present invention, the cognitive ability prediction module uses an unbiased feedforward neural network to ensure the linearity of the calculation, so as to integrate the attention results and the student vectors into the feedforward neural network to output the predicted values, and then use the sigmoid function to convert them into probability values, so as to analyze the specific contribution of each input to the predicted results.

[0026] As a preferred solution of the interpretable knowledge tracking method with time position coding of the present invention, the specific steps of S22 are as follows:

[0027] S221: Assume that the model has input student behavior information vector Test question information vector Student information vector

[0028] S222: The time of each behavior is encoded by rotation position encoding, and time information is introduced into the attention calculation, so that the attention result calculated by the model is more accurate and has better global interpretability. The attention calculation formula is as follows:

[0029]

[0030]

[0031] Among them: Rope is the rotation position embedded code, They are The query and key when computing attention, represent Calculated attention coefficient

[0032] S223: To analyze the impact of individual student behavior on test questions, it is explained that there are x student behavior records for testing the question with serial number y, and the impact of the z-th behavior on the prediction result is selected for analysis (where (x < y)), and they are respectively input All vectors except are set to zero, and then input for attention calculation to obtain the attention result only related to the behavior Related attention result The specific formula is as follows:

[0033]

[0034] Where: Is the value when calculating attention, W q 、W k And W v Are weight parameters to be trained by the model;

[0035] S224: After multi-head calculation, the student information vector Is set to 0, and then concatenated with And input into a linear feed-forward neural network (LinearFeedForward) to obtain p z ,p z Represents the model prediction output only under the influence of the behavior The specific formula is as follows:

[0036]

[0037] S225: By dividing the predicted value p z Calculated for a single behavior by the predicted value p calculated for the overall behavior, weight z Is obtained, representing The contribution to the prediction result, and the specific formula is as follows:

[0038]

[0039] Based on this, a weight is provided for each behavior, indicating its impact on the prediction result, so as to clearly understand the contribution of each behavior to the model prediction, thereby enhancing the local interpretability of the model and helping teachers and algorithm engineers better understand the decision-making process and feature importance of the model.

[0040] As a preferred solution of the interpretable knowledge tracing method with time position encoding described in the present invention, wherein: The specific steps of S3 are as follows:

[0041] S31, split data: divide the data set into training set and test set with a ratio of 8:2;

[0042] S32, training model: using the temporal attention knowledge tracking model TAKT to analyze students’ historical interaction data and basic information, and predict the probability of students answering questions correctly;

[0043] S33, evaluate the model: use the trained model to predict the test set and obtain the predicted probability of each sample. Then, evaluate the classification ability of the model by drawing the ROC (Receiver Operating Characteristic Curve) curve and calculating the AUC (Area Under the Curve) value. The closer the AUC value is to 1, the better the classification performance of the model.

[0044] S34, parameter adjustment and optimization: construct a hyperparameter search space, define multiple hyperparameter combinations, then try each hyperparameter combination one by one through grid search, evaluate its performance through the evaluation model, and record the AUC value of each combination;

[0045] S35, model verification: Use the test set for verification and re-evaluate the AUC value of the model on the test set to ensure that the model's performance is stable and reliable. Then, select 5 mainstream knowledge tracking models in the current field, use the same test set to evaluate these models, and calculate their AUC values.

[0046] Compared with existing technologies:

[0047] 1. This invention optimizes the traditional attention mechanism and introduces time-based position encoding. By analyzing the time of occurrence of historical behaviors, the model can allocate less attention to older behaviors and more attention to recent behaviors, thereby more effectively focusing on recent behaviors and improving the accuracy of predictions.

[0048] 2. Aiming at the problem of single interpretable perspective, the present invention adopts a strategy combining global and local interpretation methods to enhance the diversity of local interpretability of the model, making the prediction results not only accurate but also easy to understand and analyze. Among them, the time factor is input into the position encoding, making the output attention more accurate, so the global is better.

[0049] 3. In order to solve the homogeneity problem, the present invention reduces the homogeneity effect caused by data uniformity by constructing more diverse input data and designing new position encoding, thereby improving the robustness and adaptability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is the framework diagram of the TAKT knowledge tracking model of the present invention;

[0051] Figure 2 is the attention distribution diagram of the present invention;

[0052] Figure 3 This is a diagram showing the influence of student behavior on knowledge mastery in the present invention. DETAILED DESCRIPTION

[0053] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0054] The present invention provides an interpretable knowledge tracking method with time position coding, please refer to Figure 1-Figure 3 , including the following specific steps:

[0055] S1: build training dataset;

[0056] Its knowledge tracking evaluates students' cognitive abilities by analyzing their historical behaviors. In practical applications, the students' success rate in answering questions on a given knowledge point is usually used to represent their mastery of that knowledge point. Most knowledge tracking datasets record students' answer records. In order to use these data for training models, the following processing is required:

[0057] S11: For the data of the student's answer record exceeding the set length, it is segmented and only the part that meets the length requirement is retained. For the data of the student's answer record less than the target length, a zero padding operation is performed, which is equivalent to recording the state of the student when there is no behavior, so as to unify the length of the data sequence;

[0058] S12: after unifying the data length, classify and sort the information involved in the student's answer record, including: student information, answer question information, record of questions done before answering the target question, and whether the student can answer the target question correctly;

[0059] Student information: including the school the student belongs to, teacher, current grade, etc.;

[0060] Answer the test questions: including the knowledge points, textbooks, subjects, etc. involved in the test questions;

[0061] Records of test questions you have taken before answering the target test questions: including the knowledge points, difficulty, subjects, etc. of the test questions;

[0062] Whether the student can answer the target test questions correctly: This data is usually recorded as yes or no, with correct answers converted to 1 and wrong answers converted to 0, which is equivalent to the label in model training;

[0063] S2: Construct the temporal attention knowledge tracking model TAKT;

[0064] The specific steps of S2 are as follows:

[0065] S21: The overall structure of the model is composed of an information encoding module, a temporal attention analysis module, and a cognitive ability prediction module. Its temporal attention knowledge tracking model TAKT aims to predict the probability of students answering questions correctly by analyzing students' historical interaction data and basic information. The model is designed to handle complex student behavior records, maintain global interpretability, and analyze the impact of each behavior on the prediction results;

[0066] in:

[0067] Information encoding module: The information encoding module is responsible for effectively encoding various information processed by the model. Unlike the traditional KT model, it not only processes the sequence of students answering questions, but also considers other relevant information, such as student-related information (school, teacher) and question-related information (textbook, difficulty level). By encoding this additional information and integrating it with the behavior encoding vector, the module can more accurately evaluate the specific impact of each factor on the prediction result in the subsequent local interpretability analysis;

[0068] Set H n contains multiple behavior records. Each answer record consists of the question ID, basic information, and whether the answer is correct based on historical behavior. We split these records and convert them into a series of student behavior information vectors through Embedding. The formula is as follows:

[0069]

[0070] Where n is the number of the behavior.

[0071] Then, the ID of the test problem and its basic information (E m ) is converted into a test information vector through Embedding

[0072]

[0073] Where m represents the number of the test question.

[0074] Students and their basic information (S) are also converted into student information vectors through Embeding

[0075]

[0076] Temporal Attention Analysis Module: The purpose of the temporal attention analysis module is to incorporate the time factor into the attention analysis, optimize the global interpretable structure of the attention mechanism, adjust the attention to behaviors occurring at different time intervals, and reduce the impact of older behaviors on the current prediction results;

[0077] In the field of KT, if time is regarded as a measurement factor, absolute position encoding cannot handle the infinite possibilities of time, so relative position encoding is needed; therefore, we optimize it based on Rotational Position Embedding (RoPE) encoding, replacing the original text position information with the time information in KT. This structural adjustment enables the model to process the information of the time series position and adjust the attention allocation according to the length of the time interval;

[0078] In this module, one head in the multi-head attention mechanism is analyzed. However, in practical applications, the number of heads is usually more than one to achieve better model performance.

[0079] The steps to update the attention coefficient are as follows: first, calculate the attention (Query), (key) and (value), the student's historical behavior is converted into and prediction problems Translates to:

[0080]

[0081]

[0082]

[0083] Where W q , W k and W v It is the weight parameter that needs to be trained by the model. In order to ensure that the attention calculation process remains linear, no bias term is added;

[0084] In order to reflect the relative time distance between different behaviors, RoPE is used to convert the occurrence time t n and t m Convert to Matrix and It represents the position coding information of the time, and will be respectively and k m This makes the output include relative time information during the attention calculation process

[0085]

[0086]

[0087]

[0088] in, The length of the vector is d, represent The i-th value of the vector, represents element-wise multiplication, represent The calculated attention coefficient, where θ i is a parameter used to calculate the rotation angle in position encoding, which determines the relative rotation between each position.

[0089] for This can be done by using replace and use replace To calculate, repeated description is omitted here.

[0090] Cognitive Ability Prediction Module: To simplify the computational complexity of the output model interpretability, the structure is simplified by removing the bias, normalization, and ReLU activation functions in the fully connected layer, thereby maintaining linearity and allowing precise analysis of how each input affects the prediction results. This linear model structure helps to work backward from the prediction results to determine the specific contribution of each input, thereby simplifying the evaluation process;

[0091] Results of the attention module and student vector The concatenated input is fed into a linear feedforward neural network LinearFeedForward; the network outputs a prediction value p, which is then converted into a predicted probability value y using the sigmoid function, where y represents the probability that the student correctly answers the target question;

[0092]

[0093] y = sigmoid(p)

[0094] S22: Model interpretability analysis;

[0095] In early knowledge tracking models, or when using techniques such as SHAP to determine the impact of individual behaviors on prediction results, detailed layer-by-layer analysis is required to understand how each input affects the calculation of each layer of the model. However, in this model, the contribution of each variable to the prediction result can be simply evaluated by operating the input, without the need to analyze the impact of each input on the output of each layer layer by layer; thus, from a holistic perspective, why the model has a posteriori local interpretability is analyzed. The specific steps are as follows:

[0096] S221: Assume that the model has input student behavior information vector Test question information vector Student information vector

[0097] S222: We introduce temporal information into the attention calculation by encoding the time of each behavior using Rotation Position Encoding, making the attention results calculated by the model

[0098] more accurate and globally interpretable. The attention calculation formula is as follows:

[0099]

[0100]

[0101] S223: To analyze the impact of an individual student's behavior on a test question, given that there are x student behavior records for testing question with serial number y, and we choose to analyze the impact of the z-th behavior on the prediction result (where x < y), we separately input by setting all vectors except to zero, and then input them for attention calculation to obtain the attention result related only to behavior The specific formula is as follows: Specifically, the formula is as follows:

[0102]

[0103] S224: After multi-head calculation, the student information vector is set to 0, and then concatenated with and input into a Linear FeedForward neural network to obtain p z where p z represents the model's predicted output under the influence of only behavior The specific formula is as follows:

[0104]

[0105] S225: By dividing the predicted value p z calculated for a single behavior by the predicted value p calculated for all behaviors, we get weight z which represents the contribution of

[0106]

[0107] Based on this, a weight is provided for each behavior to indicate its impact on the prediction result, enabling a clear understanding of the contribution of each behavior to the model's prediction, thereby enhancing the local interpretability of the model and helping teachers and algorithm engineers better understand the model's decision-making process and feature importance.

[0108] S3: train the model and test it;

[0109] The specific steps of S3 are as follows:

[0110] S31, data segmentation: data segmentation is a key step in training and evaluating models. In order to ensure the generalization ability and performance of the model, the data set is divided into a training set and a test set with a ratio of 8:2, which means that 80% of the data is used to train the model and the remaining 20% ​​is used to test the model. Specifically, we first need to randomly shuffle the data set to ensure the randomness of the data distribution, thereby avoiding the deviation caused by the data order. Then, the data set is divided into a training set and a test set according to the set ratio to ensure that the data distribution of the training set and the test set is similar to avoid the problem of inconsistent data distribution in the training and testing stages. After data segmentation, the training set and the test set are respectively subjected to necessary preprocessing and feature engineering, which includes processing missing values ​​and unifying the length of the data set sequence. The uniformly processed data will be more conducive to model learning and evaluation. In this process, it is also necessary to pay attention to whether the distribution of the training set and the test set is consistent to avoid large differences between the training data and the test data caused by data segmentation, so that the performance of the model can be evaluated more fairly.

[0111] S32, training model: After the data segmentation is completed, the model training phase begins. Here, the temporal attention knowledge tracking model TAKT is used to predict the probability of students answering questions correctly by analyzing students' historical interaction data and basic information. First, the structure of the temporal attention knowledge tracking model TAKT needs to be loaded and its parameters initialized. Next, the loss function and optimizer are defined. For predicting the correct probability of students answering questions, binary cross-entropy is selected as the loss function, and the Adam optimizer is used to optimize the model parameters. In addition, the training process is carried out through multiple rounds of iterations. In each iteration (epoch), the training data is input into the model, the loss function is calculated, and the model parameters are updated through back propagation. During this process, the training process needs to be monitored and the loss value and AUC (Area Under Curve) value of each epoch are recorded. In order to prevent overfitting, if the loss value and AUC value no longer decrease significantly in multiple epochs, the training can be stopped in advance (early stopping strategy). Through these steps, the model parameters can be gradually optimized to continuously improve the performance of the model on the training set.

[0112] S33, Evaluate the model: The purpose of evaluating the model is to measure its performance on the test set and ensure that the model has good generalization ability. Here, AUC is used as the evaluation indicator. First, the trained model is used to predict the test set to obtain the prediction probability of each sample. Then, the classification ability of the model is evaluated by drawing the ROC curve and calculating the AUC value. The closer the AUC value is to 1, the better the classification performance of the model. In the evaluation process, the AUC value provides a comprehensive evaluation standard, which not only considers the accuracy of the model, but also considers its performance under different thresholds. This makes the AUC value one of the important indicators for evaluating the performance of the classification model. By calculating the AUC value, we can fully understand the performance of the model on the test set and ensure that the model has a good effect in practical applications.

[0113] S34, parameter adjustment and optimization: After the initial evaluation of the model, it is necessary to further optimize the model parameters and improve the model performance. The grid search method is used here. First, a hyperparameter search space is constructed and multiple hyperparameter combinations are defined, such as learning rate, hidden layer size, number of attention heads, etc. Then, each hyperparameter combination is tried one by one through grid search, the model is trained and its performance is evaluated, and the AUC value of each combination is recorded. Among them, although the grid search has a large amount of calculation, it can fully cover all possible hyperparameter combinations and find the hyperparameter combination with the best performance. After finding the best hyperparameter combination, the model is retrained to ensure its best performance under the new parameters. Through grid search, the best model hyperparameter combination can be found to improve the accuracy and generalization ability of the model, so that the model performs better on the test set.

[0114] S35, model verification: Finally, the model needs to be verified through an independent test set and compared with other mainstream models to ensure its performance in practical applications. First, the test set is used for verification, and the AUC value of the model on the test set is re-evaluated to ensure that the model's performance is stable and reliable. Then, five mainstream knowledge tracking models in the current field are selected, and these models are evaluated using the same test set to calculate their AUC values. After that, the AUC value of the temporal attention knowledge tracking model TAKT is compared with the AUC values ​​of the five mainstream models to analyze their relative performance. This comparison can not only verify the superiority of the temporal attention knowledge tracking model TAKT, but also help discover the shortcomings of the model and provide a reference for further optimization. Through comprehensive model verification and comparison, it can be ensured that the temporal attention knowledge tracking model TAKT has good generalization ability and stability in practical applications, laying the foundation for subsequent model deployment and monitoring;

[0115] S4: Practical application of model interpretability;

[0116] The specific steps of S4 are as follows:

[0117] S41, Interpretability of Attention Mechanism: In order to verify the influence of time information on the interpretability of attention mechanism, a student data record was extracted and the attention distribution outputs with and without time factors were compared. Figure 2 The attention distribution effect under the position encoding "with time factor" and the attention distribution "without time factor" are compared; "without time factor" refers to excluding time information from TAKT. Obviously, in the figure "with time factor", the color on the right is darker than the left, which shows that the attention distribution becomes more biased towards the most recent behavior over time. On the contrary, in the figure without time position encoding, although the color at the top is darker than the bottom, this change is due to the increase in the number of behaviors leading to the dispersion of attention; these findings can help educators and researchers understand the dynamic changes of students' attention in the long-term learning process, which may affect the way they design courses and evaluate students' learning progress. By introducing time position encoding, the interpretability of the attention mechanism can be optimized, so that it can take the time of behavior as a consideration target, and output a more accurate attention distribution map to help teachers understand the influence relationship between different behaviors;

[0118] S42, Application of IRT to evaluate the temporal attention knowledge tracking model TAKT: Item Response Theory (IRT) uses a model to analyze the interaction between student behavior and test questions, and evaluates individual abilities through responses to various questions. Its application in educational settings is highly regarded for its ability to provide in-depth insights into student assessment and item characteristics. Therefore, IRT was used to test the explanatory power and accuracy of the proposed temporal attention knowledge tracking model TAKT in assessing student knowledge status. In addition, the changes in students' knowledge mastery during the answering activities were analyzed. Therefore, a student's answer question sequence was input into the temporal attention knowledge tracking model TAKT, and the changes in the knowledge status of the four knowledge components of the student when answering 50 questions were observed. Specifically, Figure 3 As shown, according to Figure 3 The following conclusions can be drawn:

[0119] For some knowledge points, such as K3 and K4, students continuously answered related questions correctly, and the model showed consistently high prediction accuracy, which indicated that students had a good grasp of these knowledge points;

[0120] For knowledge points K1 and K2, students initially frequently gave incorrect answers, resulting in the model's continuous prediction accuracy being less than 0.5. However, as students gradually increased the proportion of correct answers to these knowledge points, the prediction accuracy also improved, especially on questions 9 and 10. When students answered these questions correctly, the prediction accuracy of knowledge points K1 and K2 rose briefly, but because students did not perform more relevant exercises, the prediction probability began to decline.

[0121] In addition, by observing the mastery of knowledge points K1 and K2, the student gave a lot of wrong answers in the early stage, and the model's prediction also showed that the student did not master these two knowledge points. The predicted probability of the output correct answer was low. However, as the student answered more and more questions correctly, the probability of correctly answering questions about these two knowledge points also increased. It can be seen that the temporal attention knowledge tracking model TAKT prioritizes recent behavior;

[0122] S43, Conclusion: Through these interpretable case studies, the effectiveness of the temporal attention knowledge tracking model TAKT in practical applications is verified. The temporal attention knowledge tracking model TAKT not only performs well on various datasets, but also demonstrates its accuracy and strong generalization ability in predicting students' knowledge mastery level; the advantage of its temporal attention knowledge tracking model TAKT lies in its interpretability and individual-centric attention mechanism. As an interpretable model, the temporal attention knowledge tracking model TAKT not only provides accurate predictions, but also allows users to understand the reasons behind these predictions, which is particularly valuable in educational environments because both educators and students can use these insights to improve learning and teaching methods; in addition, by taking individual differences into account, the temporal attention knowledge tracking model TAKT provides each student with personalized knowledge state tracking, which helps to create a more customized learning experience and potentially improve students' learning outcomes.

[0123] The feasibility of the present invention has been verified through a series of experiments and simulations, and the specific results are as follows:

[0124] 1. Baseline model comparison:

[0125] In order to evaluate the performance of the temporal attention knowledge tracking model TAKT compared with the existing mainstream knowledge tracking models, a comparative experiment was conducted on the ASSIST2009 dataset. The ASSIST2009 dataset contains 34,860 interactions of 4,217 students on 26,688 different questions, covering 123 knowledge points. Among them, five mainstream DLKT models were selected as comparison benchmarks, including DKVMN, DKT+, simpleKT, ATKT and AKT, to verify the superiority of the temporal attention knowledge tracking model TAKT in predicting students' knowledge mastery. The following table shows the AUC (Area Under Curve) values ​​of each model.

[0126]

[0127] The experimental results show that the AUC value of the temporal attention knowledge tracking model TAKT is 0.8322, which is significantly higher than other models. This proves that the temporal attention knowledge tracking model TAKT has higher accuracy and strong generalization ability in predicting students' knowledge mastery, demonstrating its significant advantages in practical applications.

[0128] 2. Ablation study:

[0129] In order to analyze the role of different information in improving the prediction accuracy of the model, an ablation experiment was designed to evaluate the model performance after removing different information sources. It compared four scenarios: no additional information, only basic information of students, only basic information of questions, and both question and student information. Through these experiments, the effect of integrating multiple information sources on improving the prediction performance of the model was verified. The following table shows the AUC values ​​of each method on the ASSIST2009 dataset.

[0130] Model Methods AUC No questions and student information 0.8087 No problem information 0.8116 No student information 0.8236 TAKT 0.8322

[0131] The results show that when both question and student information are included, the temporal attention knowledge tracking model TAKT has the highest AUC value of 0.8322. This shows that integrating multiple information sources can significantly improve the prediction performance of the model, proving the importance of considering multiple information sources in the knowledge tracking model, which helps to improve prediction accuracy.

[0132] 3. Effectiveness of TIPE

[0133] In order to explore the impact of different position encodings on model performance, it compares absolute position encoding, relative position encoding, and temporal information position encoding (TIPE) used in the temporal attention knowledge tracking model TAKT. Through these experiments, the effect of temporal information position encoding in improving model prediction accuracy and interpretability is verified. The following table shows the AUC values ​​of each method on the ASSIST2009 dataset.

[0134] Model Methods AUC Absolute position encoding 0.8231 Relative position encoding 0.8154 TAKT 0.8322

[0135] The results show that the use of TIPE significantly improves the accuracy of the model, and the AUC value of the temporal attention knowledge tracking model TAKT is 0.8322, which is higher than the absolute position encoding and relative position encoding. This proves that incorporating temporal information into relative position encoding can better capture the characteristics of behavioral intervals changing over time and improve the prediction accuracy and interpretability of the model.

[0136] In summary, the feasibility of the present invention has been verified through a series of experiments and simulations. The results show that the temporal attention knowledge tracking model TAKT has significant advantages and high accuracy in predicting students' knowledge mastery, and can effectively improve the performance and interpretability of the knowledge tracking model.

[0137] Although the present invention has been described above with reference to the embodiments, various modifications may be made thereto and parts thereof may be replaced by equivalents without departing from the scope of the present invention. In particular, as long as there is no structural conflict, the various features in the embodiments disclosed in the present invention may be used in combination with each other in any manner, and the fact that these combinations are not exhaustively described in this specification is only for the sake of omitting space and saving resources. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. An interpretable knowledge tracking method with temporal position encoding, characterized in that: The specific steps are as follows: S1: build training dataset; S2: Construct the temporal attention knowledge tracking model TAKT; S3: train the model and test it; S4: Practical application of model interpretability.

2. The interpretable knowledge tracking method with time position coding according to claim 1 is characterized in that: The specific steps of S1 are as follows: S11: For the data of the student's answer record exceeding the set length, it is segmented and only the part that meets the length requirement is retained. For the data of the student's answer record less than the target length, a zero padding operation is performed, which is equivalent to recording the state of the student when there is no behavior, so as to unify the length of the data sequence; S12: After unifying the data length, classify and organize the information involved in the students' answer records.

3. The interpretable knowledge tracking method with time position coding according to claim 2 is characterized in that: The classification and arrangement includes: student information, answer question information, test question records before answering the target test question, and whether the student can answer the target test question correctly.

4. The interpretable knowledge tracking method with time position coding according to claim 1 is characterized in that: The specific steps of S2 are as follows: S21: The overall structure of the model consists of an information encoding module, a temporal attention analysis module, and a cognitive ability prediction module; S22: Model interpretability analysis.

5. The interpretable knowledge tracking method with time position coding according to claim 4 is characterized in that: The information encoding module encodes various information types into vectors, specifically including: converting student information into vectors using Embedding for analyzing students' cognitive abilities, converting test answer information into vectors using Embedding for distinguishing each test question, and converting test question records before answering the target test question into vectors through Embedding to analyze students' cognitive abilities through their historical test answer records.

6. The interpretable knowledge tracking method with time position coding according to claim 5 is characterized in that: The temporal attention analysis module is based on a multi-head attention mechanism and integrates the time factor. Through optimized relative position encoding, the time when the behavior occurs can affect the attention distribution, so that when analyzing students' cognitive abilities, priority can be given to the most recent behavior.

7. The interpretable knowledge tracking method with time position coding according to claim 5 is characterized in that: The cognitive ability prediction module uses an unbiased feedforward neural network to ensure the linearity of the calculation, integrates the attention results and the student vector into the feedforward neural network to output the predicted value, and then uses the sigmoid function to convert it into a probability value to facilitate the analysis of the specific contribution of each input to the predicted result.

8. The interpretable knowledge tracking method with time position coding according to claim 5 is characterized in that: The specific steps of S22 are as follows: S221: Assume that the model has input student behavior information vector Test question information vector Student information vector S222: Encode the time of each behavior by rotating the position, and introduce time information into the attention calculation, so that the attention result calculated by the model is more accurate and has better global interpretability. The attention calculation formula is as follows: Among them: Rope is the rotation position embedded code, They are The query and key when computing attention, represent Calculated attention coefficient; S223: In order to analyze the impact of a single student behavior on the test question, explain that there are x student behavior records for the test question with the sequence number y, and select the zth behavior to analyze the impact of the predicted result. Effect (where (x < y)), input separately Set all vectors except to zero, and then input for attention calculation to obtain the attention result only related to the behavior The specific formula is as follows: Specific formula is as follows: in: is the value when calculating attention, W q , W k and W v are the weight parameters that need to be trained by the model; S224: After multi-head calculation, the student information vector Set to 0, and then The concatenation input is fed into a linear feed-forward neural network to obtain p z , p z Indicates that only in behavior The model prediction output under the influence of is as follows: S225: The predicted value p calculated by a single behavior z Divide by the predicted value p calculated by the overall behavior to get weight z ,express The contribution to the prediction results is as follows: Based on this, a weight is provided for each behavior to indicate its impact on the prediction result, so that the contribution of each behavior to the model prediction can be clearly understood, thereby enhancing the local interpretability of the model and helping teachers and algorithm engineers to better understand the decision-making process and feature importance of the model.

9. The interpretable knowledge tracking method with time position coding according to claim 1 is characterized in that: The specific steps of S3 are as follows: S31, split data: divide the data set into training set and test set with a ratio of 8:2; S32, training model: using the temporal attention knowledge tracking model TAKT to analyze students’ historical interaction data and basic information, and predict the probability of students answering questions correctly; S33, evaluate the model: use the trained model to predict the test set and obtain the predicted probability of each sample. Then, evaluate the classification ability of the model by drawing the ROC curve and calculating the AUC value. The closer the AUC value is to 1, the better the classification performance of the model. S34, parameter adjustment and optimization: construct a hyperparameter search space, define multiple hyperparameter combinations, then try each hyperparameter combination one by one through grid search, evaluate its performance through the evaluation model, and record the AUC value of each combination; S35, model verification: Use the test set for verification and re-evaluate the AUC value of the model on the test set to ensure that the model's performance is stable and reliable. Then, select 5 mainstream knowledge tracking models in the current field, use the same test set to evaluate these models, and calculate their AUC values.