Explanatable knowledge tracking method and device based on hypergraph convolution attention mechanism

Abstract

The invention provides an interpretable knowledge tracking method and device based on a hypergraph convolution attention mechanism. According to the method, a two-parameter input mechanism of a 2PLM (two-parameter logic model) is introduced, and the challenge that the skill difficulty difference in the same learning item is obvious is effectively coped with by the HCAKT, so that the flexibility of the model is improved, and the interpretability of the model is enhanced. In addition, a hypergraph network is further added into the model and serves as a key tool for parameter balance, potential interference of redundant information on the performance of the model is effectively avoided, and the prediction accuracy is further improved by optimizing the mutual relation between parameters. And finally, a convolution attention mechanism is also integrated into the HCAKT to simulate and capture the forgetting rule of the student. The mechanism can deeply analyze the memory decline characteristics of the students in the learning process, so that the learning states of the students can be reflected more truly.

Description

Technical Field

[0001] The present application relates to an explainable knowledge tracking method and device based on a hypergraph convolutional attention mechanism, and belongs to the field of artificial intelligence technology. Background Art

[0002] Knowledge Tracking (KT) is an intelligent education method that analyzes students' answer history, builds a dynamic model of knowledge mastery, and predicts future learning performance. It can help plan personalized learning paths, dynamically adjust teaching content, and instantly evaluate learning outcomes, providing precise support for education.

[0003] The purpose of knowledge tracking is to predict students' real-time grasp of knowledge concepts by analyzing their performance on previous problems, thereby promoting personalized and adaptive teaching. In the field of artificial intelligence, knowledge tracking has developed rapidly and continues to play a huge role in education. Various intelligent education service platforms have been continuously derived. Students have significantly improved their learning efficiency through unique and targeted services, and are more focused on their weaknesses. At the same time, teachers can also implement more accurate and effective teaching guidance based on the results of students' knowledge and ability assessment, and jointly promote the all-round development of students.

[0004] At present, the mainstream methods in the field of knowledge tracing include the following three categories: probabilistic methods, methods based on attention mechanisms, and methods based on deep learning. The probabilistic method BKT (Bayesian Knowledge Tracing) uses a hidden Markov model to predict students' knowledge status, but it is limited by the lack of discussion on the flexibility of students' knowledge changes and the single setting of skill dimensions. It cannot capture multidimensional skills, and the prediction results also have large errors. In subsequent studies, Item Response Theory (IRT) was introduced into the field of knowledge tracing, which not only improved the interpretability of the parameters in the model and made the model more convincing, but also showed through experimental results that the introduction of this module significantly improved the prediction effect. However, both BKT and IRT face the challenge of difficulty in fully capturing multidimensional skills, which affects the accuracy of prediction.

[0005] In order to overcome the limitations of traditional methods in capturing multi-dimensional skills, Piech et al. proposed a knowledge tracing method based on deep learning, DKT (Deep Knowledge Tracing). DKT uses long short-term memory (LSTM) to alleviate the limitations of skill separation and binary assumptions. However, due to the limitations of LSTM, DKT does not have the ability to accurately process each skill of students. In order to further improve DKT, a dynamic key-value memory network (DKVMN) is proposed. This method uses the potential relationship between skills and tracks the corresponding knowledge state. Moreover, DKVMN also has an update component for storing the knowledge state of students during the learning process. In addition, Deep-IRT, which is improved based on DKVMN, has improved parameter interpretability and is used to estimate project difficulty and skill difficulty, but it is still limited by the assumption that students' abilities are overly dependent on project characteristics, and fails to effectively cope with the challenges of projects with the same skills but different difficulties, affecting the accuracy and interpretability of predictions.

[0006] Ghosh et al. pointed out the error of assuming that items with the same skills are equivalent in the early KT methods, and proposed a KT method based on attention mechanism (Context-Aware Attentive Knowledge Tracing, AKT). It effectively screens and weights past learning data by finely optimizing parameter weights to reveal the deep connection between skills and solved items. Although this method can provide state-of-the-art performance in prediction, AKT cannot directly reflect the trajectory of students' ability changes for each skill, thus limiting the interpretability of model parameters.

[0007] In summary, the main challenge of KT at present is to improve the interpretability of the model and parameters and reduce the impact of redundant information on the model without reducing the prediction accuracy. Summary of the invention

[0008] The present application provides an explainable knowledge tracking method and device based on a hypergraph convolutional attention mechanism to solve the problem of not being able to grasp the material level of a fabric section in real time.

[0009] In a first aspect, an embodiment of the present application provides an explainable knowledge tracking method based on a hypergraph convolutional attention mechanism, comprising:

[0010] Predict students’ skill mastery through an interpretable model based on hypergraph convolutional attention mechanism;

[0011] The explainable model based on the hypergraph convolutional attention mechanism includes: an explainable module based on 2PLM, a hypergraph convolutional balance parameter module and a convolutional attention memory component module;

[0012] The 2PLM-based interpretable module includes an extraction module and a student module; the extraction module extracts exercise difficulty parameters and skill difficulty parameters independently based on the exercises attempted by the students and the skills required to solve the exercises, and merges the exercise difficulty parameters and the skill difficulty parameters to obtain a comprehensive difficulty parameter; the student module calculates the student's potential change of knowledge state through the knowledge mastered by the student in a set time period; the 2PLM-based interpretable module also predicts the probability of responding to the exercise based on the comprehensive difficulty parameter and the student's potential change of knowledge state;

[0013] The hypergraph convolution balance parameter module optimizes the input parameters through the hypergraph neural network model to obtain the optimized input parameters, and calculates the retention parameters, the forgetting parameters and the influence parameters based on the optimized input parameters; wherein the input parameters of the hypergraph neural network model include the vector representation of the student's answer results to the exercise and the student's past skill potential value;

[0014] The convolutional attention memory component module fuses the retention parameter, the forgetting parameter and the influence parameter based on the attention mechanism to obtain a forgetting vector;

[0015] The explainable model based on the hypergraph convolutional attention mechanism predicts the student's skill mastery based on the response probability to the exercise, the student's past skill potential value and the forgetting vector to obtain the student's skill potential value.

[0016] Based on the above method, optionally, the extraction module is based on the embedding vector of the jth exercise The process of calculating the difficulty parameter of the exercise using a feedforward neural network is as follows:

[0017]

[0018]

[0019] In the formula, and represents the weight vector, k represents the number of hidden layers, The last hidden layer It represents the exercise difficulty parameter corresponding to the jth exercise, and represents the bias parameter vector;

[0020] The extraction module is also based on the embedding vector of the skill required to solve the jth exercise The process of obtaining the skill difficulty parameter and calculating it using a feedforward neural network is as follows:

[0021]

[0022]

[0023] In the formula, and Represents the weight vector, the last layer of the hidden layer represents the skill difficulty parameter corresponding to the j-th skill, and represents the bias parameter vector.

[0024] Based on the above method, optionally, in the student module, the knowledge acquired by the students in the time period t is To calculate the student's potential change in knowledge state:

[0025]

[0026] in, Indicates the i-th row Vector, ω ji represents the attention weight of the i-th skill;

[0027] The process of predicting student ability is as follows:

[0028]

[0029]

[0030] in, represents the weight vector, m represents the number of hidden layers, and , represents the bias parameter vector.

[0031] Based on the above method, optionally, the 2PLM-based interpretable module uses the student's ability to solve exercise j at time t With two difficulty parameters and The difference between the sum of and is used to predict the probability of a student's response to an exercise:

[0032]

[0033] Among them, P j (q,s) represents the predicted response probability.

[0034] Based on the above method, optionally, the 2PLM-based interpretable module calculates the model loss through cross entropy loss:

[0035]

[0036] Among them, l(U j (q,s),P j (q,s)) represents the cross entropy loss.

[0037] Based on the above method, optionally, the potential value of the student's past skills after optimization of the neural network model based on the hypergraph architecture The calculation process is as follows:

[0038]

[0039] Where W is the weight vector and η is the bias parameter vector;

[0040] Vector representation of the student's answer to the exercise and the potential value of students’ past skills The optimization process in the hypergraph network is as follows:

[0041]

[0042] Among them, 1 ∈R,ζ 2 ∈R and r are hyperparameters, and r represents the number of recursions in the recursive architecture.

[0043] Based on the above method, optionally, the retention parameter α t , forget parameter e t and the influencing parameter z t The calculation formula is:

[0044]

[0045] Among them, W e1 , W e2 , W z1 , W e1 , W α1 and W α2 represents the weight vector, η e , η z and η α represents the bias parameter vector.

[0046] Based on the above method, optionally, the formula for obtaining the forgetting vector by fusing the convolutional attention memory component module is expressed as:

[0047]

[0048]

[0049] Among them, si is the representation vector of the i-th factor, D aez The reserved parameter α in the above module t , forget parameter e t and the influencing parameter z t Composition, L is the number of attention layers; τ represents the proportion of input in the attention layer;

[0050] In addition, a unique attention weight is assigned to the forgetting factor:

[0051]

[0052] Among them, τ i represents the attention weight.

[0053] Based on the above method, optionally, the formula for calculating the potential value of the skills mastered by the student using the interpretable model based on the hypergraph convolutional attention mechanism is expressed as:

[0054]

[0055] Among them, AC = 1dCNN(S′), which means using a one-dimensional convolutional network to capture the retention parameter α t and the forgetting parameter e t The mutual relationships among.

[0056] In a second aspect, the embodiment of the present application further provides an interpretable knowledge tracking device based on a hypergraph convolutional attention mechanism, which includes:

[0057] The prediction unit is used to predict students’ skill mastery through an interpretable model based on the hypergraph convolutional attention mechanism;

[0058] The explainable model based on the hypergraph convolutional attention mechanism includes: an explainable module based on 2PLM, a hypergraph convolutional balance parameter module and a convolutional attention memory component module;

[0059] The 2PLM-based interpretable module includes an extraction module and a student module; the extraction module extracts exercise difficulty parameters and skill difficulty parameters independently based on the exercises attempted by the students and the skills required to solve the exercises, and merges the exercise difficulty parameters and the skill difficulty parameters to obtain a comprehensive difficulty parameter; the student module calculates the student's potential change of knowledge state through the knowledge mastered by the student in a set time period; the 2PLM-based interpretable module also predicts the probability of responding to the exercise based on the comprehensive difficulty parameter and the student's potential change of knowledge state;

[0060] The hypergraph convolution balance parameter module optimizes the input parameters through the hypergraph neural network model to obtain the optimized input parameters, and calculates the retention parameters, the forgetting parameters and the influence parameters based on the optimized input parameters; wherein the input parameters of the hypergraph neural network model include the vector representation of the student's answer results to the exercise and the student's past skill potential value;

[0061] The convolutional attention memory component module fuses the retention parameter, the forgetting parameter and the influence parameter based on the attention mechanism to obtain a forgetting vector;

[0062] The explainable model based on the hypergraph convolutional attention mechanism predicts the student's skill mastery based on the response probability to the exercise, the student's past skill potential value and the forgetting vector to obtain the student's skill potential value.

[0063] In the interpretable knowledge tracking method (HCAKT) and device based on the hypergraph convolutional attention mechanism provided in this application, by introducing the dual-parameter input mechanism of 2PLM (dual-parameter logic model), HCAKT effectively responds to the challenge of significant differences in skill difficulty within the same learning project, which not only improves the flexibility of the model, but also enhances the interpretability of the model. In addition, a hypergraph network is added to the model. As a key tool for parameter balance, it effectively avoids the potential interference of redundant information on model performance, and further improves the accuracy of prediction by optimizing the relationship between parameters. Finally, HCAKT also incorporates a convolutional attention mechanism to simulate and capture students' forgetting patterns. This mechanism can deeply analyze the memory decay characteristics of students in the learning process, thereby more realistically reflecting the students' learning status. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] The accompanying drawings herein are incorporated into and constitute a part of the specification, illustrate embodiments consistent with the present application, and together with the specification are used to explain the principles of the present application. In addition, these drawings and text descriptions are not intended to limit the scope of the concept of the present application in any way, but to illustrate the concept of the present application for those skilled in the art by referring to specific embodiments.

[0065] Figure 1 A schematic diagram of the architecture of an interpretable model based on a hypergraph convolutional attention mechanism provided for one embodiment of the present application;

[0066] Figure 2 A dual parameter input framework diagram based on 2PLM provided for one embodiment of the present application;

[0067] Figure 3 A schematic diagram of the structure of a hypergraph balancing factor and a convolutional attention memory component provided for one embodiment of the present application;

[0068] Figure 4 A schematic diagram of the convolution mapping of a convolutional attention memory component provided for one embodiment of the present application. DETAILED DESCRIPTION

[0069] In order to make the purpose, technical solutions and advantages of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in combination with the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present application. In the absence of conflict, the following embodiments and features in the embodiments can be combined with each other.

[0070] In order to solve the problems raised in the background technology part, this solution proposes a new interpretable knowledge tracking method based on the hypergraph convolutional attention mechanism. This method predicts students' skill mastery through an interpretable model based on the hypergraph convolutional attention mechanism.

[0071] The interpretable model based on the hypergraph convolutional attention mechanism mainly consists of two modules: 1) the module for acquiring student learning features; 2) the module for updating the learning process. The dual parameter input used by this method is better than the previous Deep-IRT model in acquiring the student's ability parameter characteristics and project difficulty characteristics during the learning process, and has better interpretability.

[0072] In addition, when updating students' knowledge, they only rely on model input as the basis for forgetting and updating. When facing new projects, if their skills do not match their past knowledge well, the prediction effect will be easily affected. To this end, this solution also introduces a hypergraph neural network into the update module to achieve a deep fusion of students' historical knowledge and new project skills, thereby obtaining more accurate parameter settings. At the same time, combined with the attention convolution module, the ability to extract parameter features is further enhanced, significantly improving the prediction accuracy.

[0073] The performance of this method is compared with the earlier KT method on a public dataset. The results show that the method proposed in this scheme has significant improvements in both accuracy and interpretability. The main contributions of this scheme are as follows:

[0074] 1): By establishing a dual-parameter IRT logic module to estimate skill parameters and project parameters, the parameter interpretability of the KT model is improved, which can effectively cope with the challenge of large differences in skill difficulty in the same project, thereby improving the adaptability and accuracy of the model and making the KT model of this scheme more convincing.

[0075] 2): The hypergraph neural network is used to balance the model input with past knowledge parameters, effectively alleviating the potential interference of redundant information on model performance. The memory update component composed of convolutional attention is used to extract features of the balanced parameters, which is more in line with the law of forgetting and significantly improves the performance and predictive ability of the model.

[0076] Two-parameter IRT logistic model:

[0077] As we all know, the IRT model has a high interpretability. However, in the standard IRT model, the learning ability remains constant throughout the learning process, which is obviously inconsistent with the actual learning rules. Although many recent studies have expanded the standard IRT, the IRT models of these studies all assume a single dimension of ability. Specifically, they only consider exercises with independent multiple skills, but not exercises with multiple different skills. Therefore, this scheme adopts a more popular new IRT model: the two-parameter IRT logistic model (Two-Parameter Logistic Model, 2PLM). This new IRT model with two parameters allows exercises and skills to correspond at each step. This can not only handle exercises with independent multiple skills, but also exercises with multiple different skills, and further improve the interpretability of the model.

[0078] Hypergraph Neural Networks and Convolutional Attention:

[0079] In the field of natural language processing, many studies use hypergraph networks to optimize the hidden layer weights and hidden variables of LSTM to express its contextual transformation. This solution adopts the above-mentioned methods, which not only optimizes the forgetting parameters used in the subsequent update of students' skill mastery, but also balances the past mastery, which significantly improves the experimental results.

[0080] Considering that the memory update module extracted by this scheme using a simple extraction method cannot well highlight the important role of forgetting law in the learning process of students. Therefore, this scheme introduces the convolutional attention mechanism, which aims to combine the obtained forgetting factors to simulate the forgetting law of students based on the convolutional attention memory component. The effect obtained is more significant than the simple extraction method, which well proves the important influence of forgetting factor on the learning process of students.

[0081] For the convenience of explanation, the following assumptions are made:

[0082] There is a set of exercises that students can perform in a certain order.

[0083] Γ={(ρ 1 ,α 1 ),(ρ2 ,α 2 ),(ρ k ,α k ),....,(ρ t ,α t )} to study the problems in the exercises, where ρ k ∈Ι represents the question answered by the student at step k, t represents the total number of steps used by the student to answer the question, and α k indicates that the student's answer at step k is correct. That is, when the student answers ρ k When the answer is correct, α k =1, otherwise, α k =0.

[0084] In previous studies, there is still a problem of poor interpretability of model parameters, that is, it cannot represent the ability of each student at each stage in the past. To address this problem, this solution proposes a hypergraph convolutional attention mechanism interpretable model that is different from other Deep-IRT models. The model diagram is as follows: Figure 1 As shown in the figure. The model mainly consists of two main modules: 1) 2PLM-based interpretable module. 2) Hypergraph convolution balance parameter module and convolution attention memory component module. In summary, the structure and function of each module are as follows:

[0085] The interpretable module based on 2PLM includes an extraction module and a student module; the extraction module independently extracts the exercise difficulty parameter and the skill difficulty parameter based on the exercises attempted by the students and the skills required to solve the exercises, and merges the exercise difficulty parameter and the skill difficulty parameter to obtain a comprehensive difficulty parameter; the student module calculates the students' potential knowledge state through the knowledge mastered by the students in a set time period; the interpretable module based on 2PLM also predicts the probability of responding to the exercises based on the comprehensive difficulty parameter and the students' potential knowledge state;

[0086] The hypergraph convolution balance parameter module optimizes the input parameters through the hypergraph neural network model to obtain the optimized input parameters, and calculates the retention parameters, forgetting parameters and influence parameters based on the optimized input parameters; wherein the input parameters of the hypergraph neural network model include the vector representation of the student's answer results to the exercise and the student's past skill potential value;

[0087] The convolutional attention memory component module fuses the retention parameters, forgetting parameters and influence parameters based on the attention mechanism to obtain the forgetting vector;

[0088] The interpretable model based on the hypergraph convolutional attention mechanism predicts the students' skill mastery based on the response probability to the exercise, the students' past skill potential value and the forgetting vector, and obtains the potential value of the skills mastered by the students.

[0089] Specifically, for the 2PLM-based interpretable module:

[0090] In previous studies using IRT, it is generally believed that each parameter is item-dependent and the implicit problem difficulty and skill ability are assumed to be constant. This assumption does not hold when the problem difficulty and skill mastery differ greatly. Therefore, it is difficult to predict students' learning behavior in this situation. To solve this problem, this scheme proposes an interpretable IRT module based on 2PLM, which consists of two small modules: 1): an independent problem difficulty / skill difficulty extraction module; 2): a student module based on DKVMN, such as Figure 2 As shown in the figure. The extraction module starts with two inputs: the exercises that students try and the skills required to solve the exercises, and extracts the difficulty factor for the exercises and the mastery factor for the skills. Then, the student module is combined with the actual situation of the students to calculate the probability that the students answer the exercises correctly. Because the proposed module extracts exercises and skills separately, the accuracy of the prediction will not decrease.

[0091] In the retrieval module, there are two difficulty parameters in IRT, one of which is the practice difficulty parameter The other is the skill difficulty parameter This module combines these two difficulty features to generate a new difficulty feature That is, the comprehensive difficulty parameter. For a more intuitive explanation, the input in the extraction module is the embedding vector of the jth exercise The calculation process using a feedforward neural network is as follows:

[0092]

[0093]

[0094] In the formula, and represents the weight vector, k represents the number of hidden layers, The last hidden layer It represents the exercise difficulty parameter corresponding to the jth exercise, and represents the bias parameter vector.

[0095] Similarly, to calculate the difficulty of a skill, we take the same approach based on the embedding vector of the skill required to solve the jth exercise The process of obtaining the skill difficulty parameter and calculating it using a feedforward neural network is as follows:

[0096]

[0097]

[0098] in, and Represents the weight vector, the last layer of the hidden layer represents the skill difficulty parameter corresponding to the j-th skill, and represents the bias parameter vector.

[0099] In the student module, this method is based on the knowledge acquired by students in time period t. To calculate the student's potential change in knowledge state of:

[0100]

[0101] in, Indicates the i-th row vector, and ω ji represents the attention weight of the i-th skill.

[0102] Similar to the weight calculation method of DKVMN, the prediction process for student ability is as follows:

[0103]

[0104]

[0105] in, represents the weight vector, m represents the number of hidden layers, and Denotes the bias parameter vector. Unlike the previous Deep-IRT, this method predicts student ability through an independent student module and has no direct connection with the item difficulty parameter. In addition, The value of also indicates the student's potential learning ability. In other words, can be considered as a measurement model.

[0106] The proposed method uses the student's ability to solve exercise j at time t With two difficulty parameters and The difference between the sum of j (q,s):

[0107]

[0108] After that, the loss function used in this scheme is used to map the cross entropy of the classification error, that is, by predicting the response probability P j (q,s) and the true response probability Uj (q,s) to calculate:

[0109]

[0110] Among them, l(U j (q,s),P j (q,s)) represents the cross entropy loss.

[0111] Student Hypergraph Network and Convolutional Attention Mechanism Module

[0112] In the long-term learning process, we should not only make good use of the current input, but also consider how to use past information to optimize the forgetting parameters and update the students' knowledge status. The simplest method is to use the LSTM forget gate to perform simple data processing, and the second is to balance the input data by adding weights. and past potential value However, as the learning process progresses, the number of weights increases dynamically, resulting in too many parameters to successfully estimate the results.

[0113] In a standard LSTM, hidden variables change over time, but their update weights are fixed and not well targeted at different time points. To solve this problem, various hypergraph networks have been proposed in various fields to optimize LSTM, which supports the main recurrent neural network by optimizing the non-shared weights at each time point in the hidden layer. These studies have shown that the results of this new LSTM are better than those of the traditional LSTM. Inspired by these studies, this solution also adds this new hypergraph network to the memory update component. It can not only generate corresponding non-shared weights for each time point, but also avoid parameter redundancy.

[0114] Figure 3 The hypergraph architecture and the proposed convolutional attention memory component module are given. The hypergraph optimizes the weights generated based on past data, which not only improves the interpretability of the parameters in the model, but also provides more accurate input for the subsequent memory component to update the student's knowledge state.

[0115] Specifically, in this method update Previously, the hypergraph architecture was constructed by taking the latent variables from time point t-λ to t To balance the current input and past potential value where λ represents the number of attempts the student made on the latent variable in the past.

[0116] The input of the hypergraph architecture is the embedding vector and past potential value When students answer question j, according to the input (sj ,u jt ) Calculate the embedding vector also, The calculation process is as follows:

[0117]

[0118] Where W is the weight vector and η is the bias parameter vector;

[0119] Vector representation of the student's answer to the exercise and the potential value of students’ past skills The optimization process in the hypergraph network is as follows:

[0120]

[0121] Among them, 1 ∈R,ζ 2 ∈R and r are hyperparameters, r represents the number of recursions in the recursive architecture, W v and W M represents the weight vector.

[0122] If r = 1, then and Because of the repeated multiplication of formulas (11) and (12), and This scheme also optimizes the number of recursions in subsequent experiments.

[0123] The optimized hypergraph architecture and As input to estimate the retention parameter α t and the forgetting parameter e t , these two parameters play a crucial role in the convolutional attention memory component module. Compared with the previous Deep-IRT, the optimized and It can better balance the parameters to obtain better accuracy and results. In addition, this solution also extracts an influencing parameter z t To α t For these three parameters (retain parameter α t , forget parameter e t and the influencing parameter z t ) is calculated as follows:

[0124]

[0125] Among them, W e1 , W e2 , W z1 , We1 , W α1 and W α2 represents the weight vector, η e , η z and η α represents the bias parameter vector.

[0126] After obtaining the required forgetting parameters, considering that the forgetting module can better conform to the students' forgetting rules as much as possible, the convolutional attention adopted in this scheme can fuse the obtained parameters to make them more correlated, and can alleviate the impact on the data by reducing noise, making it more accurate. The formula for the convolutional attention memory component module to fuse the forgetting vector is expressed as:

[0127]

[0128]

[0129] Among them, s i is the representation vector of the i-th factor, D aez The reserved parameter α in the above module t , forget parameter e t and the influencing parameter z t Composition, L is the number of attention layers; τ represents the proportion of input in the attention layer;

[0130] In addition, a unique attention weight S′ is assigned to the forgetting factor:

[0131]

[0132] This scheme uses the reweighting in formula (18) to simulate the student's forgetting law. Specifically, a one-dimensional convolutional network is used to capture the retention parameter α t and the forgetting parameter e t The mutual relationship, such as Figure 4 As shown, it is simply defined as:

[0133] AC=1dCNN(S′) (19)

[0134] Finally, the obtained parameters are recorded as the forgetting vector. Together with the skill parameters and past potential values, the student's performance is predicted and the potential value of the skills mastered by the student is updated. That is, the formula for calculating the potential value of the skills mastered by the student based on the interpretable model of the hypergraph convolutional attention mechanism is expressed as:

[0135]

[0136] During the updating process, the model can obtain more potential value information and can better and accurately capture the changes in students' knowledge status.

[0137] In summary, this scheme proposes a knowledge tracking method (HCAKT) that integrates hypergraph network and convolutional attention mechanism, which significantly enhances the adaptability and prediction accuracy of the traditional knowledge tracking model (KT). By introducing the dual-parameter input mechanism of 2PLM (dual-parameter logistic model), HCAKT effectively copes with the challenge of significant differences in skill difficulty within the same learning project, which not only improves the flexibility of the model but also enhances the interpretability of the model. In addition, the hypergraph network is added to the model as a key tool for parameter balance, which effectively avoids the potential interference of redundant information on model performance and further improves the accuracy of prediction by optimizing the relationship between parameters. Finally, HCAKT also incorporates the convolutional attention mechanism to simulate and capture the forgetting rules of students. This mechanism can deeply analyze the memory decay characteristics of students in the learning process, thereby more realistically reflecting the learning status of students.

[0138] In addition, the embodiment of the present application provides an interpretable knowledge tracking device based on a hypergraph convolutional attention mechanism, a prediction unit, for predicting the student's skill mastery through an interpretable model based on the hypergraph convolutional attention mechanism;

[0139] Among them, the interpretable model based on the hypergraph convolutional attention mechanism includes: an interpretable module based on 2PLM, a hypergraph convolutional balance parameter module, and a convolutional attention memory component module;

[0140] The interpretable module based on 2PLM includes an extraction module and a student module; the extraction module independently extracts the exercise difficulty parameter and the skill difficulty parameter based on the exercises attempted by the students and the skills required to solve the exercises, and merges the exercise difficulty parameter and the skill difficulty parameter to obtain a comprehensive difficulty parameter; the student module calculates the students' potential knowledge state through the knowledge mastered by the students in a set time period; the interpretable module based on 2PLM also predicts the probability of responding to the exercises based on the comprehensive difficulty parameter and the students' potential knowledge state;

[0141] The hypergraph convolution balance parameter module optimizes the input parameters through the hypergraph neural network model to obtain the optimized input parameters, and calculates the retention parameters, forgetting parameters and influence parameters based on the optimized input parameters; wherein the input parameters of the hypergraph neural network model include the vector representation of the student's answer results to the exercise and the student's past skill potential value;

[0142] The convolutional attention memory component module fuses the retention parameters, forgetting parameters and influence parameters based on the attention mechanism to obtain the forgetting vector;

[0143] The interpretable model based on the hypergraph convolutional attention mechanism predicts the students' skill mastery based on the response probability to the exercise, the students' past skill potential value and the forgetting vector, and obtains the potential value of the skills mastered by the students.

[0144] Among them, regarding the specific implementation methods of each module of the above-mentioned explainable knowledge tracking device based on the hypergraph convolutional attention mechanism, reference can be made to the corresponding contents in the aforementioned method embodiments, which will not be repeated here.

[0145] It can be understood that the same or similar parts of the above embodiments can be referenced to each other, and the contents not described in detail in some embodiments can refer to the same or similar contents in other embodiments.

[0146] It should be noted that, in the description of the present invention, the terms "first", "second", etc. are only used for descriptive purposes and cannot be understood as indicating or implying relative importance. In addition, in the description of the present invention, unless otherwise specified, the meaning of "plurality" refers to at least two.

[0147] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, segment or portion of code that includes one or more executable instructions for implementing the steps of a specific logical function or process, and the scope of the preferred embodiments of the present invention includes alternative implementations in which functions may not be performed in the order shown or discussed, including performing functions in a substantially simultaneous manner or in the reverse order depending on the functions involved, which should be understood by those skilled in the art to which the embodiments of the present invention belong.

[0148] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, a plurality of steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0149] A person skilled in the art may understand that all or part of the steps of the method for implementing the above-mentioned embodiment may be completed by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium. When the program is executed, it includes one or a combination of the steps of the method embodiment.

[0150] In addition, each functional unit in each embodiment of the present invention may be integrated into a processing module, or each unit may exist physically separately, or two or more units may be integrated into one module. The above-mentioned integrated module may be implemented in the form of hardware or in the form of a software functional module. If the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it may also be stored in a computer-readable storage medium. The above-mentioned storage medium may be a read-only memory, a disk or an optical disk, etc.

[0151] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0152] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. An interpretable knowledge tracking method based on hypergraph convolutional attention mechanism, characterized in that: include: Predict students’ skill mastery through an interpretable model based on hypergraph convolutional attention mechanism; The explainable model based on the hypergraph convolutional attention mechanism includes: an explainable module based on 2PLM, a hypergraph convolutional balance parameter module and a convolutional attention memory component module; The 2PLM-based interpretable module includes an extraction module and a student module; the extraction module extracts exercise difficulty parameters and skill difficulty parameters independently based on the exercises attempted by the students and the skills required to solve the exercises, and merges the exercise difficulty parameters and the skill difficulty parameters to obtain a comprehensive difficulty parameter; the student module calculates the student's potential change of knowledge state through the knowledge mastered by the student in a set time period; the 2PLM-based interpretable module also predicts the probability of responding to the exercise based on the comprehensive difficulty parameter and the student's potential change of knowledge state; The hypergraph convolution balance parameter module optimizes the input parameters through the hypergraph neural network model to obtain the optimized input parameters, and calculates the retention parameters, the forgetting parameters and the influence parameters based on the optimized input parameters; wherein the input parameters of the hypergraph neural network model include the vector representation of the student's answer results to the exercise and the student's past skill potential value; The convolutional attention memory component module fuses the retention parameter, the forgetting parameter and the influence parameter based on the attention mechanism to obtain a forgetting vector; The explainable model based on the hypergraph convolutional attention mechanism predicts the student's skill mastery based on the response probability to the exercise, the student's past skill potential value and the forgetting vector to obtain the student's skill potential value.

2. The method according to claim 1, characterized in that The extraction module is based on the embedding vector of the jth exercise The process of calculating the difficulty parameter of the exercise using a feedforward neural network is as follows: In the formula, and represents the weight vector, k represents the number of hidden layers, The last hidden layer It represents the exercise difficulty parameter corresponding to the jth exercise, and represents the bias parameter vector; The extraction module is also based on the embedding vector of the skill required to solve the jth exercise The process of obtaining the skill difficulty parameter and calculating it using a feedforward neural network is as follows: In the formula, and Represents the weight vector, the last layer of the hidden layer represents the skill difficulty parameter corresponding to the j-th skill, and represents the bias parameter vector.

3. The method according to claim 2, characterized in that In the student module, the knowledge acquired by the students in the t time period To calculate the student's potential change in knowledge state: in, Indicates the i-th row Vector, ω ji represents the attention weight of the i-th skill; The process of predicting student ability is as follows: in, represents the weight vector, m represents the number of hidden layers, and represents the bias parameter vector.

4. The method according to claim 3, characterized in that The 2PLM-based interpretable module uses the student's ability to solve exercise j at time t With two difficulty parameters and The difference between the sum of and is used to predict the probability of a student's response to an exercise: Among them, P j (q,s) represents the predicted response probability.

5. The method according to claim 4, characterized in that The 2PLM-based interpretable module calculates the model loss through cross entropy loss: Among them, l(U j (q,s),P j (q,s)) represents the cross entropy loss.

6. The method according to claim 5, characterized in that The potential value of students’ past skills after optimization of the neural network model based on the hypergraph architecture The calculation process is as follows: Where W is the weight vector and η is the bias parameter vector; Vector representation of the student's answer to the exercise and the potential value of students’ past skills The optimization process in the hypergraph network is as follows: Among them, ζ1∈R, ζ2∈R and r are hyperparameters, r represents the number of recursions in the recursive architecture, and W v and W M represents the weight vector.

7. The method according to claim 6, characterized in that The retention parameter α t , forget parameter e t and the influencing parameter z t The calculation formula is: Among them, W e1 , W e2 , W z1 , W e1 , W α1 and W α2 represents the weight vector, η e , η z and η α represents the bias parameter vector.

8. The method according to claim 7, characterized in that The formula for obtaining the forgetting vector by fusion of the convolutional attention memory component module is expressed as: Among them, s i is the representation vector of the i-th factor, D aez The reserved parameter α in the above module t , forget parameter e t and the influencing parameter z t Composition, L is the number of attention layers; τ represents the proportion of input in the attention layer; In addition, a unique attention weight is assigned to the forgetting factor: Among them, τ i represents the attention weight.

9. The method according to claim 8, characterized in that The formula for calculating the potential value of the skills mastered by students by the interpretable model based on the hypergraph convolutional attention mechanism is expressed as: Among them, AC = 1dCNN(S′), which means using a one-dimensional convolutional network to capture the retention parameter α t and the forgetting parameter e t The mutual relationships among.

10. An interpretable knowledge tracking device based on hypergraph convolutional attention mechanism, characterized in that: include: The prediction unit is used to predict students’ skill mastery through an interpretable model based on the hypergraph convolutional attention mechanism; The explainable model based on the hypergraph convolutional attention mechanism includes: an explainable module based on 2PLM, a hypergraph convolutional balance parameter module and a convolutional attention memory component module; The 2PLM-based interpretable module includes an extraction module and a student module; the extraction module extracts exercise difficulty parameters and skill difficulty parameters independently based on the exercises attempted by the students and the skills required to solve the exercises, and merges the exercise difficulty parameters and the skill difficulty parameters to obtain a comprehensive difficulty parameter; the student module calculates the student's potential change of knowledge state through the knowledge mastered by the student in a set time period; the 2PLM-based interpretable module also predicts the probability of responding to the exercise based on the comprehensive difficulty parameter and the student's potential change of knowledge state; The hypergraph convolution balance parameter module optimizes the input parameters through the hypergraph neural network model to obtain the optimized input parameters, and calculates the retention parameters, the forgetting parameters and the influence parameters based on the optimized input parameters; wherein the input parameters of the hypergraph neural network model include the vector representation of the student's answer results to the exercise and the student's past skill potential value; The convolutional attention memory component module fuses the retention parameter, the forgetting parameter and the influence parameter based on the attention mechanism to obtain a forgetting vector; The explainable model based on the hypergraph convolutional attention mechanism predicts the student's skill mastery based on the response probability to the exercise, the student's past skill potential value and the forgetting vector to obtain the student's skill potential value.

Images