Knowledge tracing method based on perplexity enhancement using gated neural networks
By using a perplexity enhancement method based on gated neural networks, the problem of online learning platforms being unable to track students' emotional changes was solved, students' knowledge status was dynamically updated, and the teaching quality and the effectiveness of personalized learning were improved.
Patent Information
- Application Number
- CN202410635560.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-05-22
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2044-05-22
AI Technical Summary
Existing online learning platforms cannot effectively track students' emotional changes and their impact on learning outcomes, resulting in lower teaching quality compared to offline teaching. Furthermore, existing sentiment analysis methods are not applicable to numerical datasets.
A confusion enhancement method based on gated neural networks is adopted to predict confusion emotions through students' interaction records and distinguish the differences in learning benefits caused by confusion. The knowledge status of students is tracked by combining learning and forgetting factors, and the knowledge status of students is dynamically updated using a gated neural network including confusion gate, learning gate and forgetting gate.
It effectively predicts students' confusion and distinguishes learning gains, improving the model's predictive performance and interpretability, and supporting personalized learning recommendations.
Smart Images

Figure CN118504671B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of knowledge tracing technology, specifically to a knowledge tracing method based on gated neural networks with enhanced perplexity. Background Technology
[0002] With the development of the internet, online education has become a trendy learning method. It differs from traditional education, as the learning content is no longer limited to what human instructors teach, encompassing a vast amount of learning materials. It is no longer constrained by the location and time of traditional classrooms; students can learn anytime, anywhere using their mobile phones or computers. Currently, a number of online learning platforms have emerged, such as MOOCs and Alexa. Although online learning platforms have the above advantages, the current teaching quality is still not as good as offline teaching. This is reflected in the fact that online learning systems provide essentially the same or similar knowledge content to different students. For excellent students, they may be able to learn this knowledge independently, but for average students, if they encounter some concepts they cannot understand, they may feel confused and develop a dislike for learning. Human instructors, on the other hand, can create different learning plans for different students, constantly monitor their knowledge acquisition, assign tasks, and when students encounter difficult concepts, human instructors can observe their emotions, answer their questions, and create specific exercises to encourage them to learn these concepts. To achieve these functions of human tutors, an effective solution is to analyze students' mastery of each knowledge point through historical interaction records, that is, to track students' knowledge status, analyze students' emotional state, and provide timely matching exercises.
[0003] Existing research primarily focuses on the differences in student gains arising from the relationship between exercises and knowledge concepts, neglecting the gains from changes in student emotions. However, the impact of student emotional changes on learning outcomes is significant; for example, the learning gains students gain when encountering and resolving confusion are greater than ordinary learning gains. Current sentiment analysis methods in online learning platforms either study the emotional information contained in text from the perspective of natural language processing or analyze students' facial expressions to classify emotions from the perspective of computer vision. However, most knowledge tracing methods are designed based on numerical datasets and are not suitable for either of these approaches. Summary of the Invention
[0004] To address the problems of existing technologies and consider the impact of students' emotional changes on learning outcomes, this invention proposes a knowledge tracking method based on gated neural networks to enhance confusion levels. This method predicts students' confusion emotions in their interaction records and distinguishes the differences in learning benefits brought by confusion in exercises, thereby tracking students' knowledge status by combining learning and forgetting factors.
[0005] The technical solution of this invention is as follows:
[0006] A perplexity-enhanced knowledge tracing method based on gated neural networks includes the following steps:
[0007] Step 1: Obtain the dataset for training the model and perform data preprocessing;
[0008] Retrieve student interaction records from the database, including the ID of the practice question the student performed, the judgment value of the student's final answer (whether it was correct or not), the time spent on the question, and the number of attempts; and the sequence I of interaction records for the student with ID n. sn for:
[0009] I sn ={(e1,r1,at1,ms1,),…,(e t ,r t ,at t ms t ,),…,(e T ,r T ,at T ms T ,)}
[0010] Where e t The id represents the practice problem that the student is working on at time t, r. t The value representing whether the answer is correct or not, at t The student representative attempted to solve e t Number of times, ms t Representing students in e t The time spent on it;
[0011] Delete the interaction record sequences of students whose actual number of interaction records does not meet the requirements, as well as the interaction records in the interaction record sequences whose time consumption does not meet the requirements;
[0012] Step 2: Pre-train the Q-matrix, including the following steps:
[0013] Step 2.1: Read the practice question IDs and corresponding labeled knowledge concepts from the dataset, and count the total number of knowledge concepts;
[0014] Step 2.2: Create a vector of all zeros with the same number of elements as the number of knowledge concepts, and then set the position of each knowledge concept to 1 according to the order in which the exercises appear, thus linking each knowledge concept with each exercise.
[0015] Step 2.3: Combine the all-zero vectors established in Step 2.2 to obtain the Q matrix;
[0016] Step 2.4: The Q matrix was pre-trained using the EKPT pre-training partial order method to obtain the embedding matrix A; and based on the embedding matrix A, according to the formula...
[0017] E t =A T ×e t
[0018]
[0019] Get Embedding Exercises
[0020] Step 3: Construct a gated neural network:
[0021] The gated neural network includes a confusion gate, a learning gate, and a forgetting gate;
[0022] The confusion gate is constructed by summing the number of attempts and time taken by the student before the next correct answer, applying an activation function to obtain a predicted confusion value, and then mapping the confusion value to a range of 0-1 using a sigmoid function. The result is then used to determine whether the answer is correct or not, based on the given judgment value r. t Obtain the perplexity level, and from the perplexity level... Decide whether to pass the signal to the next gating mechanism;
[0023] The learning gate is based on perplexity. As input, the learning gains students made from this exercise, along with the learning gains from confusion, are integrated to obtain the total learning gains.
[0024] The forgetting gate outputs the negative learning benefit students experience due to forgetting. It is only related to the knowledge level and practice questions;
[0025] Step 4: Based on the results obtained in Step 3 The outputs of three gated neural networks are used to update the student's knowledge state ks at time t. t ;
[0026] Step 5: Based on the student's knowledge state ks at time t t Predicting students' performance in new exercises t+1 Performance on r t+1 ;
[0027] Step 6: Divide the dataset into a training set and perform unsupervised learning on the gated neural network:
[0028] For a student's interaction record sequence
[0029] I sn ={(e1,r1,at1,ms1),…,(e t ,rt ,at t ms t ),…,(e T ,r T ,at T ms T )}
[0030] For each student, a knowledge state ks0 is initialized. Within one time step, the loss value between the predicted value and the actual value of the student's correct or incorrect answer is calculated. After optimization iteration, the final knowledge tracking model based on gated neural network with enhanced perplexity is obtained.
[0031] Step 7: Obtain new student interaction record sequences. Input the knowledge tracking model based on gated neural network with enhanced perplexity obtained in Step 6 to obtain the knowledge state of each student at this time and predict the student's future performance in practice.
[0032] Furthermore, in step 1, student interaction record sequences with fewer than 15 actual interaction records are removed; interaction records in the interaction record sequence that take less than 20ms are also removed.
[0033] Furthermore, the specific expression of the perplexity gate in step 3 is as follows:
[0034] Confusion Gate conf t The calculation method at time t is as follows:
[0035] onf t =tanh(W1·at) t +W2·ms t +b c )
[0036] Where tanh is the activation function, and W1, W2, b c These are the parameters that the model needs to optimize.
[0037] Confusion Gate conf t Mapping to the range of 0-1 using the sigmoid function, and using r t The decision to pass the data to the next gating mechanism yields the perplexity level. for:
[0038]
[0039] Furthermore, the specific way to express the door in step 3 is as follows:
[0040]
[0041] Where tanh is the activation function, and W3, W4, W5, b lThese are the parameters that the model needs to optimize, ks t-1 This represents the knowledge state at time t-1.
[0042] Furthermore, the specific expression of the forgetting gate in step 3 is as follows:
[0043]
[0044] Where tanh is the activation function, W6, W7, b f These are the parameters that the model needs to optimize.
[0045] Furthermore, in step 4, the student's knowledge state ks at time t is updated. t for:
[0046]
[0047] Where tanh is the activation function, W6, W7, b f These are the parameters that the model needs to optimize.
[0048] Furthermore, in step 5, predicting the student's performance in the new exercise e t+1 Performance on r t+1 The process is as follows:
[0049]
[0050]
[0051] Where ReLU and sigmoid are activation functions, dot is the vector product, and W9 and b are also activation functions. 1 These are the parameters used for model training.
[0052] Furthermore, in step 6, the loss value between the predicted value and the true value is calculated using BCELoss:
[0053]
[0054] Where N is the total number of samples in the training set, y i p is the real data of the i-th sample. i is the predicted data for the i-th sample; the Adam optimization algorithm is used to optimize the model to obtain the final knowledge tracking model based on gated neural networks with enhanced perplexity.
[0055] Beneficial effects
[0056] As can be seen from the above technical solution, the present invention has the following beneficial effects: Using a gated neural network, it can predict students' confusion levels through numerical data, comprehensively considering the gains from practice, the additional gains from confusion, and the negative gains from forgetting to simulate students' knowledge states, and dynamically update these states during the learning process. Based on different knowledge states, it can predict a student's future performance on exercises, thereby tracking the updates of different knowledge states caused by confusion and further distinguishing the knowledge states among students. Furthermore, compared with other knowledge tracking models, the method of the present invention fully utilizes various records during student interactions, models students' confusion levels, predicts their confusion values, and achieves integrated modeling of confusion, learning, and forgetting behaviors, improving the model's predictive performance and interpretability, thus enabling more effective application to personalized recommendation learning.
[0057] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description
[0058] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:
[0059] Figure 1 This is a block diagram illustrating the principle of the method of the present invention;
[0060] Figure 2 This is a graph showing the accuracy of the predicted performance of this invention.
[0061] Figure 3 This is an area under the performance curve of the present invention;
[0062] Figure 4 A trend chart for predicting the direction of confusion. Detailed Implementation
[0063] To predict students' confusion and differentiate the learning benefits derived from confusion in exercises, this invention provides a knowledge tracking method based on gated neural networks that enhances confusion levels. This method can predict students' confusion levels from their interaction records and track their knowledge status by incorporating learning and forgetting factors.
[0064] Specifically:
[0065] First, student data is extracted from the source dataset. Records of students with fewer than 15 interaction records and responses less than 20ms are removed. Student records include the practice questions the student answered, their final answer (correct or incorrect), the time spent on the question, and the number of attempts. The interaction record for each student at time t is then represented by I = {(e...}. t ,rt ,at t ms t ,)}express.
[0066] Secondly, we obtain the embedding matrix A by pre-training a Q-matrix that associates labeled practice questions with knowledge concepts. This embedding matrix A allows us to easily obtain the current practice question e. t The knowledge concepts involved. The Q matrix here is specifically a j-row, n-column Q matrix containing j practice questions and n knowledge concepts. In particular, a value of 1 in the i-th column of the u-th row indicates that the practice question with id u covers the knowledge concept involving i. Of course, the other entries in the u-th row are 0.
[0067] We then use the first gated neural network to predict the student's confusion level (confusion gate), and then use the confusion level as input to the next gated neural network to simulate the learning gains the student will gain from this exercise and the additional gains from solving the confusion (learning gate). Finally, we use the last gated network to simulate the negative learning gains caused by forgetting (forgetting gate), and finally update the student's knowledge status.
[0068] Finally, the student's performance in the next practice is predicted based on this knowledge state. The model loss is calculated by comparing the predicted and actual results. In this way, the data in the dataset can be used to train the final knowledge tracking model based on gated neural networks with enhanced perplexity. Using this trained model, the student's mastery of each knowledge concept can be analyzed based on the student's historical answer records in the online learning platform, and the student's performance in future practice can be predicted.
[0069] This invention draws upon the theory of the impact of confusion on students in educational psychology. It acquires interactive data from students' online learning processes and utilizes a knowledge tracking model based on a gated neural network to enhance confusion levels. This model estimates students' knowledge states, providing strong support for differentiating students' knowledge states and providing personalized learning within the online learning platform. Specifically, the knowledge tracking model uses a triple-gated neural network to explicitly model students' confusion, learning, and forgetting behaviors during the learning process, dynamically updating students' knowledge states and improving the model's predictive performance and interpretability.
[0070] The embodiments of the present invention are described in detail below. These embodiments are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.
[0071] Step 1: Data Preprocessing
[0072] First, remove students with fewer than 15 actual interaction records. Then, retrieve student interaction records from the dataset based on their student IDs. These records include the ID of the practice question the student answered, the student's final answer (correct or incorrect), the time spent on the question, and the number of attempts. This yields all interaction records for each student, represented by the following sequence:
[0073] I sn ={(e1,r1,at1,ms1,),…,(e t ,r t ,at t ms t ,),…,(e T ,r T ,at T ms T ,)}#(1)
[0074] I sn e represents all interaction records of the student with ID n. t The id represents the practice problem that the student is working on at time t, r. t Represents the answer (1 for correct, 0 for incorrect), at t The student representative attempted to solve e t Number of times, ms t Representing students in e t The time spent on it.
[0075] Finally, we removed records in the sequence where the student's time (ms) was less than 20ms, and obtained the final data.
[0076] Step 2: Pre-train the Q-matrix;
[0077] Read the exercise question IDs and corresponding labeled knowledge concepts from the dataset, count the total number of knowledge concepts, create an all-zero vector with the same number of elements as the number of knowledge concepts, and set the position of each knowledge concept to 1 according to the order in which the exercise questions appear. In this way, each knowledge concept is associated with each exercise question. Finally, combine these one-hot vectors to obtain a Q matrix that corresponds one-to-one between exercise questions and knowledge concepts.
[0078] In the knowledge tracing task, the relationship between practice questions and corresponding knowledge concepts is found through the Q matrix labeled by experts. It is a one-hot vector, such as a j-row n-column Q matrix, which contains j practice questions and n knowledge concepts. Specifically, a value of 1 in the i-th column of the u-th row indicates that the practice question with id u covers the knowledge concept involving i. Of course, the other entries in the u-th row are 0.
[0079] However, the Q-matrix labeled by experts contains limited information and cannot reflect the difficulty of the exercises. Moreover, as a one-hot vector, it is not conducive to model training. Therefore, in this embodiment, the partial order method of EKPT pre-training is used to obtain the embedding matrix A.
[0080] Specifically, for the i-th exercise, if the u-th knowledge concept is marked (at this time Q...) i,u =1), but without labeling the v-th knowledge concept, we define this partial order relation as u> i v indicates that the association strength between exercise i and the u-th knowledge concept is greater than that between exercise i and the v-th knowledge concept. The goal of this method is to train an embedding matrix A∈R. M×K Where M is the number of practice questions and K is the number of knowledge concepts, the model is optimized using a generative model and Bayesian processing, and for each practice question's embedding vector A... i The posterior probability p(A|PQ) is calculated as follows:
[0081] p(A|PQ)∝p(PQ|A)×p(A)#(2)
[0082] Where p(A|PQ) is the likelihood distribution, p(A) is the prior distribution, and it is assumed that the prior distribution follows a zero-mean spherical Gaussian distribution:
[0083]
[0084] The likelihood distribution assumes that the partial order relations induced by the Q matrix are independent and identically distributed among the exercises. Specifically, for each exercise i, in order to make the exercise embedding matrix A... i Satisfying the induced partial order constraint, the likelihood distribution is defined as:
[0085] p(u> i v|A i )=σ(A i,U -A i,v )#(4)
[0086] Where σ is the sigmoid function, in this way, the embedding vector of each exercise is optimized to satisfy the partial order relation in the Q matrix, while maintaining its relevance to the knowledge concept. This helps improve the interpretability and accuracy of the model. The resulting embedding matrix A∈R M×K It can be used as a lookup table in the knowledge tracing process, as follows:
[0087] E t =A T ×e t #(5)
[0088] Building upon this, we further employed a response method that distinguishes between correct and incorrect answers, given the interaction record at time t (e t ,r t ), to obtain embedding exercises (A 2K-dimensional vector), specifically as follows:
[0089]
[0090] Here, 0 is a K-dimensional vector filled with zeros. Clearly, different answers will result in different embedding exercises.
[0091] Step 3: Construct a gated neural network
[0092] Suppose that at time t, the student answers exercise e t Try at t Cost per transaction (ms) t Time will answer. t Practice e t The relationship between knowledge points and knowledge points can be obtained through the embedding matrix A. Thus, based on an interaction record I = {(e^t ...} at time t... t ,r t ,at t ms t We calculated the level of confusion for each incorrect answer, and included the positive benefit of confusion when the student answered correctly. Specifically, for a student's answer sequence on a certain knowledge point, we summed the number of attempts and the time used before the next correct answer, applied an activation function to obtain the predicted confusion value, and finally calculated the value based on the answer r. t This determines the output of the gating network. Formally, the confusion gate (conf) t The calculation method at time t is as follows:
[0093] conf t =tanh(W1·at) t +W2·ms t +b c )#(7)
[0094] Here, tanh is the activation function, W1, W2, b. c These are the parameters that the model needs to optimize.
[0095] Finally, the result is mapped to the range of 0-1 using the sigmoid function, and then r... t This determines whether to pass the information to the next gating mechanism. Formally, the degree of confusion... The calculation method is as follows:
[0096]
[0097] Based on the obtained level of confusion This is then input into the next gating mechanism (the learning gate), which integrates the student's learning gains from this exercise as well as any learning gains from confusion (if any). The calculation method is as follows:
[0098]
[0099] Here, tanh is the activation function, W3, W4, W5, b. l These are the parameters that the model needs to optimize, ks t-1 This represents the knowledge state at time t-1.
[0100] Finally, we calculate the negative learning benefit students receive due to forgetting, which is only related to their knowledge state and the exercises. The calculation method is as follows:
[0101]
[0102] Here, tanh is the activation function, W6, W7, b. f These are the parameters that the model needs to optimize.
[0103] Step 4: Update knowledge status
[0104] The result obtained in step 3 The results from three gated neural networks simulated student confusion, learning gains, and forgetting losses, respectively. After calculating these three factors, we updated the student's knowledge state ks at time t. t The calculation method is as follows:
[0105]
[0106] Here, ReLU is the activation function, W8,b ks These are the parameters that the model needs to optimize.
[0107] Step 5: Predict future performance
[0108] We obtained the student's knowledge state ks at time t using three gating mechanisms. t Therefore, it is possible to predict how students will perform in new exercises. t+1 The performance on the surface. First, the practice E is obtained through a pre-trained practice embedding model. T+1 ks through a fully connected network t Mapping to E T+1 The vector space is used to represent the knowledge state ks using the ReLU function. t Convert all values to positive values, represented as therefore Each value in the table corresponds to the level of mastery of a knowledge point. Finally, the conversion... With E T+1 The dot product of these two variables is used to predict the result, which is then derived using the sigmoid function. Formally, the process is as follows:
[0109]
[0110]
[0111] Here, ReLU and sigmoid are activation functions, dot is the vector product, and W9 and b are also activation functions. 1 These are the parameters used for model training.
[0112] Step 6: Model training;
[0113] We randomly divided the dataset into 70% as the training set for unsupervised learning in a gated neural network and 30% as the test set to test the model's performance.
[0114] For each student's interaction record sequence
[0115] I sn ={(e1,r1,at1,ms1),…,(e t ,r t ,at t ms t ),…,(e T ,r T ,at T ms T )}
[0116] First, we initialize a knowledge state ks0 (a dk-dimensional random tensor) for each student to represent their initial knowledge state. Within one time step, we use BCELoss to calculate the loss between the predicted and actual values, as follows:
[0117]
[0118] Where N is the total number of samples in the training set, y i p is the real data of the i-th sample. i This is the predicted data for the i-th sample. Optimizing the model using the Adam optimization algorithm yields the final perplexity-enhanced knowledge tracking model based on a gated neural network.
[0119] Step 7: Use the model to predict students' performance on practice and their feelings of confusion.
[0120] We use 30% of the dataset as the test set to test the model's performance, with 20% used for testing and 10% for validation. We retrieve student interaction records from the dataset by student ID value, including the ID of the practice question the student performed, the student's last answer (correct or incorrect), the time spent on that question, and the number of attempts. This yields all interaction records for each student, represented by the following sequence:
[0121] I sn ={(e1,r1,at1,ms1,),…,(e t ,r t ,at t ms t ,),…,(e T ,r T ,at T ms T ,)}#14
[0122] I sn e represents all interaction records of the student with ID n. t The id represents the practice problem that the student is working on at time t, r. t Represents the answer (1 for correct, 0 for incorrect), at t The student representative attempted to solve e t Number of times, ms t Representing students in e t The time spent on it.
[0123] The actual student interaction records are used as input to the model into three gated neural networks to obtain the knowledge state of each student at that time. Finally, this knowledge state is input into the prediction module to obtain the student's future performance in practice. The model performance is evaluated by calculating the accuracy (ACC) and area under the curve (AUC) between the model's predicted student responses and the actual responses. The results are as follows: Figure 2 As shown. The focus of this invention is the design of the confusion gate. To verify the effectiveness of the confusion gate, we also predicted the changes in students' confusion emotions over a period of time, and the results are as follows. Figure 3 As shown.
[0124] The above results indicate that the knowledge tracing model based on gated neural networks designed in this invention has excellent performance in knowledge tracing tasks, and can predict the changes in students' confusion during practice while keeping the error within a small range.
[0125] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention without departing from the principles and spirit of the present invention.
Claims
1. A confusion augmented knowledge tracing method based on a gated neural network, characterized in that: Includes the following steps: Step 1: Obtain the dataset for training the model and perform data preprocessing; Obtain student interaction records from the database, including the student's performed exercise question id, the student's last answer correct or not correct judgment value, the student's time spent on the question, the number of attempts; the interaction record sequence of the student with ID n For: wherein represents the exercise question id that the student is currently working on at time t, represents the judgment value of whether the answer is correct or not, represents the number of times the student has tried to solve , represents the time spent by the student on ; Delete the interaction record sequences of students whose actual number of interaction records does not meet the requirements, as well as the interaction records in the interaction record sequences whose time consumption does not meet the requirements; Step 2: Pre-training The matrix includes the following steps: Step 2.1: Read the practice question IDs and corresponding labeled knowledge concepts from the dataset, and count the total number of knowledge concepts; Step 2.2: Create a vector of all zeros with the same number of elements as the number of knowledge concepts, and then set the position of each knowledge concept to 1 according to the order in which the exercises appear, thus linking each knowledge concept with each exercise. Step 2.3: Combine the all-zero vectors established in Step 2.2 to obtain... matrix; Step 2.4: The partial order method of EKPT pre-training was used to... The matrix is pre-trained to obtain the embedding matrix A; and based on the embedding matrix A, according to the formula... Get Embedding Exercises ; Step 3: Construct a gated neural network: The gated neural network includes a confusion gate, a learning gate, and a forgetting gate; The confusion gate is constructed by summing the number of attempts and the time taken by the student before the next correct answer, applying the sum to an activation function to obtain a predicted confusion value, and then passing the confusion value through... The function maps the result to a value between 0 and 1, indicating whether the answer is correct or not. Obtain the perplexity level, and from the perplexity level... Decide whether to pass the signal to the next gating mechanism; Confusion Gate The calculation method at time t is as follows: Where tanh is the activation function, and W1, W2, b c These are the parameters that the model needs to optimize. The door to confusion pass Functions are mapped to the range 0-1, through The decision to pass the data to the next gating mechanism yields the perplexity level. for: The learning gate is based on perplexity. As input, the learning gains students made from this exercise, along with the learning gains from confusion, are integrated to obtain the total learning gains. ; The forgetting gate outputs the negative learning benefit students experience due to forgetting. It is only related to the knowledge level and practice questions; Step 4: Based on the results obtained in Step 3 , , The outputs of the three gated neural networks are updated. Students' knowledge status at all times ; Step 5: Based on the student's knowledge status at time t Predicting students' performance in new exercises Performance on ; Step 6: Divide the dataset into a training set and perform unsupervised learning on the gated neural network: For a student's interaction record sequence Initialize a knowledge state for each student Within a time step, the loss value between the predicted value and the actual value of whether the student's answer is correct or not is calculated. After optimization and iteration, the final knowledge tracking model based on gated neural network with enhanced perplexity is obtained. Step 7: Obtain new student interaction record sequences. Input the knowledge tracking model based on gated neural network with enhanced perplexity obtained in Step 6 to obtain the knowledge state of each student at this time and predict the student's future performance in practice.
2. The knowledge tracing method based on gated neural networks with enhanced perplexity as described in claim 1, characterized in that: In step 1, student interaction record sequences with fewer than 15 actual interaction records are removed; interaction records in the interaction record sequence that take less than 20ms are also removed.
3. The knowledge tracing method based on gated neural networks with enhanced perplexity as described in claim 1, characterized in that: In step 3, the specific way to express a door is as follows: Where tanh is the activation function, and W3, W4, W5, b l These are the parameters that the model needs to optimize. This represents the knowledge state at time t-1.
4. The knowledge tracing method based on gated neural networks with enhanced perplexity as described in claim 1, characterized in that: The specific way to express the forgetting gate in step 3 is as follows: where tanh is an activation function, W6, W7, b f are parameters to be optimized by the model.
5. The knowledge tracing method based on gated neural networks with enhanced perplexity according to claim 1, characterized in that: In step 4, update Students' knowledge status at all times for: where tanh is an activation function, W6, W7, b f are parameters to be optimized by the model.
6. The knowledge tracing method based on gated neural networks with enhanced perplexity according to claim 1, characterized in that: In step 5, predict students' performance in new exercises. Performance on The process is as follows: where ReLU, sigmoid are activation functions, dot is the vector product, W9, b 1 are parameters of the model training.
7. The knowledge tracing method based on gated neural networks with enhanced perplexity according to claim 1, characterized in that: In step 6, BCELoss is used to calculate the loss value between the predicted value and the actual value: in It is the total number of samples in the training set. It is the first Real data for each sample It is the predicted data for the i-th sample; The Adam optimization algorithm was used to optimize the model, resulting in the final knowledge tracing model with enhanced perplexity based on a gated neural network.
