A knowledge tracking method based on temporal convolution

Abstract

The present invention provides a knowledge tracking method based on temporal convolution, which relates to the field of learning analysis. The specific implementation steps of the method are as follows: (1) Collect learning interaction records of multiple students and construct a learning interaction record dataset D. (2) Construct a temporal convolution attention knowledge tracking model TCAKT‑E, which is divided into a representation module and a learning modeling module. (3) Divide the dataset D into a training dataset and a test dataset in proportion, input the training dataset into TCAKT‑E to train and optimize the model, and then use the test dataset to verify the performance of the model. (4) Use the trained and verified TCAKT‑E model to predict student performance.

Description

Technical Field

[0001] The present invention relates to the field of learning analysis, and more specifically, to a knowledge tracking method based on time convolution. Background Art

[0002] With the rapid development of big data technology, data-driven intelligent education services have made personalized learning possible for learners. However, how to accurately and comprehensively model learners through their learning process data on online learning platforms is a difficult problem for personalized learning. Knowledge tracking has opened up new ideas for the study of this problem. Knowledge tracking is a technology that models learners' answering behavior sequences and predicts learners' knowledge mastery status. Through knowledge tracking, teachers can timely grasp students' mastery of knowledge, further assign exercises to students in a targeted manner, and achieve teaching in accordance with their aptitude and improve learning efficiency.

[0003] Knowledge tracking algorithms can be roughly divided into two categories: traditional machine learning-based methods and deep learning knowledge tracking. In the first category of algorithms, Bayesian knowledge tracking is a representative algorithm, which assumes that the student's knowledge state is a set of binary variables and uses hidden Markov chains to update the knowledge state based on the student's answer scores. Another typical method is factor analysis, and the classic model is item response theory, which measures students' abilities and the difficulty of questions, and then evaluates students' abilities and the probability of them answering questions correctly.

[0004] Knowledge tracking based on deep learning methods tracks students' real-time changes in knowledge mastery, thereby revealing the strength of students' knowledge mastery. Scholars have conducted in-depth data mining on students' historical learning data and developed many complex modeling techniques. The deep knowledge tracking model uses a recurrent neural network to track students' knowledge status and has powerful representation capabilities. Subsequently, a large number of improvement solutions based on the deep knowledge tracking model have emerged.

[0005] Chinese patent document number CN118115329A, published on May 31, 2024, records "A long-short term knowledge state adaptive knowledge tracking method based on self-attention mechanism". This invention proposes a Long-Short Term Knowledge State Adaptive Knowledge Tracing Model based on self-attention mechanism, aiming at the characteristics that the changes in knowledge state during students' learning process are jointly affected by long-term and short-term knowledge states. First, a long-short term knowledge state extractor is constructed based on the self-attention mechanism. At the same time, a dynamic weight strategy is designed for the characteristics of long-term knowledge state changes, and a dynamic window strategy is designed for the characteristics of short-term knowledge state changes. Then, a long-term and short-term knowledge state adaptive structure is designed by neural network structure search. Finally, the long-term and short-term knowledge states are fused to output the student's knowledge state through a pre-layer. The knowledge tracking method proposed in the present invention can more effectively extract the characteristics of long-term and short-term knowledge states in the student's learning process, and further improve the prediction accuracy.

[0006] Chinese patent document number CN117973527A, published on May 3, 2024, records "a knowledge tracking method based on GRU to capture question context features". The invention includes: based on a preset data structure and historical interaction records, determining the embedding matrix corresponding to the historical interaction records to embed the historical interaction records; determining the knowledge state features and exercise features corresponding to the historical interaction records based on the GRU unit; determining the answer prediction results based on the knowledge state features and the exercise features, which effectively solves the technical problem that the related technology destroys the underlying semantic information of the interaction encoding and reduces the benefits brought by the sequence information, resulting in poor accuracy in predicting students' answers, and achieves the technical effect of accurately predicting the learners' answer results.

[0007] The current focus of deep learning knowledge tracking is mainly on recurrent neural networks. Due to their structural characteristics, recurrent neural networks suffer from gradient vanishing and gradient exploding when training long sequence data. As a result, knowledge tracking based on this network and its improved models cannot learn dependencies with a large time span during the training process, causing long-term dependency problems.

[0008] Current deep learning knowledge tracking mostly relies on students' historical answer data, while ignoring differential learning features such as question similarity and exercise difficulty, which are crucial for understanding students' learning behavior and predicting their future learning performance. Summary of the invention

[0009] The purpose of the present invention is to propose a knowledge tracking method based on time convolution in view of the problems existing in the above background technology.

[0010] Terminology explanation:

[0011] One-Hot Encoding: maps each value in a finite set to a vector in which only one element is 1 and the others are 0. This encoding method is often used to process categorical variables and convert them into numerical data that machine learning models can understand.

[0012] Residual Block: It is a deep learning network. Its core idea is to add "residual" so that the network can better learn the mapping relationship between input and output. Its basic structure includes a sub-network and a skip connection. The sub-network consists of multiple convolutional layers.

[0013] Dilated Convolution: Also known as Atrous Convolution, it is a convolution operation commonly used in deep learning. Compared with traditional convolution operations, dilated convolution increases the effective receptive field of the network while maintaining the size of the receptive field, thereby improving the effect of feature extraction.

[0014] Weight normalization: is a technique used to train deep learning models. It normalizes the model's weight vector to unit length, and then puts the weight length information into a learnable parameter. The model's weight can adaptively adjust the length during training, thereby improving the model's performance.

[0015] Dropout: It is a regularization technique used to train deep learning models. It randomly discards some neurons of the model during the training process, thereby reducing the complexity of the model and avoiding overfitting of the model.

[0016] Channel Attention Mechanism: It is used to adaptively adjust the feature response of different channels to improve the sensitivity to different features. Its steps include: using average pooling to compress the feature map, using The function calculates the final weight of the feature channel and inputs the obtained weight into the original feature channel.

[0017] Global Average Pooling (GAP): A common pooling operation, the specific steps are to calculate the average value of each channel in the feature map to obtain a one-dimensional vector. The length of this vector is equal to the number of channels, which can be directly input into the Softmax layer for classification. Since GAP does not require additional parameters, it can effectively reduce the number of parameters of the model and improve the generalization ability of the model.

[0018] The knowledge tracking method based on time convolution has the following steps:

[0019] S100: Collect learning interaction records of multiple students and construct a learning interaction record dataset D;

[0020] S200: constructing a temporal convolutional attention knowledge tracking model TCAKT-E, wherein the temporal convolutional attention knowledge tracking model is divided into a representation module and a learning modeling module;

[0021] S300: Divide the data set D into a training data set and a test data set in proportion, input the training data set into TCAKT-E to train and optimize the model, and then use the test data set to verify the performance of TCAKT-E;

[0022] S400: Use the trained and validated TCAKT-E model to predict student performance.

[0023] Furthermore, the learning interaction records of the multiple students specifically include:

[0024] Each learning interaction record includes students, question information, answer status, and skill points.

[0025] Furthermore, for students, Information about the problem, skill points, The construction of the temporal convolutional attention knowledge tracking model TCAKT-E specifically includes:

[0026] S210: Use the one-hot encoding method to encode the question information 、Answer situation Encode them separately to get the problem vector and the answer vector , splicing and Get the student's performance at time t , the calculation formula is as follows:

[0027]

[0028] in, Indicates students Answered Question information, Indicates students For The answer to the question information, a value of 1 indicates a correct answer, and a value of 0 indicates an incorrect answer;

[0029] S220: Obtaining a representation of the student experience , and Encode and get the student experience vector ;

[0030] S230: Joint student response performance , Student Experience Vector Get student behavior vector , the calculation formula is as follows:

[0031]

[0032] in, represents a joint function;

[0033] S240: Using Temporal Convolutional Attention Network TCAN on Student Behavior Data Conduct learning modeling.

[0034] Furthermore, the acquisition of student experience representation The details are as follows:

[0035]

[0036] in, E xp s k i 1~ m ∈[0,c] From time 1 to time Students At skill points Empirical characterization on is an empirical constant representing a positive integer, From time 1 to time Students At skill points The difference between the number of correct answers and the number of incorrect answers, and The calculation formulas are as follows:

[0037]

[0038]

[0039] in, From time 1 to time Students At skill points The number of correct answers, From time 1 to time Students At skill points The number of incorrect answers.

[0040] Furthermore, the use of TCAN for student behavior data Conduct learning modeling, including:

[0041] S241: Multi-scale feature information extraction: student behavior data Input to In the stacked residual blocks, the coefficients used for the dilated convolution in the stacked residual blocks increase exponentially. After different residual block processing, learning behavior feature maps of different time scales are obtained; the residual blocks are composed of dilated convolution, ReLU activation, weight normalization, and Dropout operations. After each residual block processing, a learning behavior feature map is output. ), fusion The learning behavior characteristic diagram of different time scales is as follows:

[0042]

[0043] in, Represents the learning behavior feature map after fusion, represents the fusion function, represents the learning behavior feature map obtained by the first residual block processing, Represents the learning behavior feature map obtained by the second residual block processing, Indicates The learning behavior feature map obtained by processing the residual blocks;

[0044] S242: Output after processing by S241 ( Indicates height, Indicates width, Represents the total number of channels) after fusion learning behavior characteristic diagram ,right Perform global average pooling GAP operation, the calculation formula is as follows:

[0045]

[0046] in, represents the global average pooling operation, express Middle The characteristics of each element;

[0047] S243: Determine the scope of channel interaction information : According to the total number of channels The size of the adaptive channel interaction information is determined , the calculation formula is as follows:

[0048]

[0049] in, Indicates the number of neighboring channels that interact with the current channel. Indicates departure Recently and A large odd number, C represents the total number of channels, and yes Positive integer constants in calculation formulas;

[0050] S244: Channel attention adjustment: Adaptive confirmation After that, all channels share the learned attention weight parameters. Channels Weight The calculation formula is as follows:

[0051]

[0052] in, represents the weight parameter shared by all channels, Representation and Adjacent A collection of channels.

[0053] Furthermore, the dividing of the data set D and training and validating TCAKT-E specifically includes:

[0054] Set the hyperparameters required for TCAKT-E, such as the batch input size of the training dataset, learning rate, residual block dropout ratio, etc. The loss function used is calculated as follows:

[0055] L us =- 1 n ∑ i [ r t l and y t + 1- r t l and 1- y t ]

[0056] in, represents the loss function, express Real-time answer to questions, express The prediction at the moment answers the question.

[0057] Beneficial effects:

[0058] The present invention provides a knowledge tracking method based on time convolution, by adding student experience representation , which can enhance the model's differentiated learning representation of student performance prediction; by using TCAN to learn and model student behavior data, it can adapt to the learning feature map Assign different channel weights to speed up the convergence of model training and effectively solve the gradient vanishing problem. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 It is a flowchart of the knowledge tracking method based on temporal convolution.

[0060] Figure 2 is a flowchart for constructing the TCAKT-E model.

[0061] Figure 3 This is the structural diagram of the TCAKT-E model. DETAILED DESCRIPTION

[0062] The present invention will be further defined below in conjunction with the accompanying drawings and embodiments, but is not limited thereto.

[0063] Example

[0064] Those skilled in the art will appreciate that the following examples are only used to illustrate the present invention and should not be considered to limit the scope of the present invention. Where specific techniques or conditions are not indicated in the examples, the techniques or conditions described in the literature in the art or the product specifications are used. Where the materials or equipment used do not indicate the manufacturer, they are all conventional products that can be purchased.

[0065] A knowledge tracking method based on time convolution has the following process: Figure 1 As shown, the method specifically includes:

[0066] S100: Collect learning interaction records of multiple students and construct a learning interaction record dataset D.

[0067] Preferably, each of the learning interaction records contains students, question information, answer status, and skill points.

[0068] In this example, learning interaction records of multiple students may be collected from, including but not limited to, online education platforms, and each learning interaction record includes the student, question information, answer status, and skill points.

[0069] S200: Construct a temporal convolutional attention knowledge tracking model TCAKT-E. The temporal convolutional attention knowledge tracking model is divided into a representation module and a learning modeling module. The construction process is as follows: Figure 2 As shown, the model structure is as Figure 3 shown.

[0070] Preferably, for students, Information about the problem, skill points, The construction of the temporal convolutional attention knowledge tracking model TCAKT-E specifically includes:

[0071] S210: Use the one-hot encoding method to encode the question information 、Answer situation Encode them separately to get the problem vector and the answer vector , splicing and Get the student's performance at time t , the calculation formula is as follows:

[0072]

[0073] in, Indicates students Answered Question information, Indicates students For The answer to the question information, A value of 1 indicates a correct answer. A value of 0 indicates an incorrect answer;

[0074] S220: Obtaining a representation of the student experience , and Encode and get the student experience vector ; Student experience representation The calculation of is as follows:

[0075]

[0076] in, E xp s k i 1~ m ∈[0,c] From time 1 to time Students At skill points Empirical characterization on is an empirical constant representing a positive integer, From time 1 to time Students At skill points The difference between the number of correct answers and the number of incorrect answers, and The calculation formulas are as follows:

[0077]

[0078]

[0079] in, From time 1 to time Students At skill points The number of correct answers, From time 1 to time Students At skill points The number of incorrect answers.

[0080] In this example, the calculation formula of the student experience representation is applied to the data set D constructed by S100, and the experience representation constant is set As the student’s experience representation value increases, the student’s accuracy in answering questions at the corresponding skill point also increases.

[0081] S230: Joint student response performance , Student Experience Vector Get student behavior vector , the calculation formula is as follows:

[0082]

[0083] in, represents a joint function;

[0084] S240: Using Temporal Convolutional Attention Network TCAN on Student Behavior Data Conduct learning modeling.

[0085] Preferably, the temporal convolutional attention network TCAN is used to train student behavior data Learning modeling specifically includes:

[0086] S241: Multi-scale feature information extraction: student behavior data Input to In the stacked residual blocks, the coefficients used for the dilated convolution in the stacked residual blocks increase exponentially. After different residual block processing, learning behavior feature maps of different time scales are obtained; the residual blocks are composed of dilated convolution, ReLU activation, weight normalization, and Dropout operations. After each residual block processing, a learning behavior feature map is output. ), fusion The learning behavior characteristic diagram of different time scales is as follows:

[0087]

[0088] in, Represents the learning behavior feature map after fusion, represents the fusion function, represents the learning behavior feature map obtained by the first residual block processing, Represents the learning behavior feature map obtained by the second residual block processing, Indicates The learning behavior feature map obtained by processing the residual blocks;

[0089] In this example, student behavior data The input is sent to three stacked residual blocks with exponentially increasing expansion coefficients. After being processed by different residual blocks, learning behavior feature maps of different time scales are obtained. The residual blocks are composed of dilated convolution, ReLU activation, weight normalization, and Dropout operations. After each residual block is processed, a learning behavior feature map is output. ), fuse the learning behavior feature graphs of three different time scales, and output the fused learning behavior feature graph , the formula is as follows:

[0090]

[0091] is the learning behavior feature map output by the first residual block, is the learning behavior feature map output by the second residual block, is the learning behavior feature map output by the third residual block;

[0092] S242: Output after processing by S241 The learning behavior characteristics after fusion ,right Perform global average pooling GAP operation, the calculation formula is as follows:

[0093]

[0094] in, Indicates height, Indicates width, Indicates the total number of channels, represents the global average pooling operation, express Middle The characteristics of each element;

[0095] In this example, enter The learning behavior characteristics after fusion , after the global average pooling GAP operation, The output dimension becomes ;

[0096] S243: Determine the scope of channel interaction information : According to the total number of channels The size of the adaptive channel interaction information is determined , the calculation formula is as follows:

[0097]

[0098] in, Indicates the number of neighboring channels that interact with the current channel. Indicates departure Recently and A large odd number, C represents the total number of channels, and yes Positive integer constants in calculation formulas, The value is 2. The value is 1;

[0099] S244: Channel attention adjustment: Adaptive confirmation After that, all channels share the learned attention weight parameters. Channels Weight The calculation formula is as follows:

[0100]

[0101] in, represents the weight parameter shared by all channels, Representation and Adjacent A collection of channels.

[0102] Specifically, first Perform GAP operation to obtain all features without dimensionality reduction, and then adaptively select the size The fast one-dimensional convolution is used to capture the local cross-channel interaction information, and then the The function generates the weight ratio of each channel, and then converts the original input Combined with the channel weight, the features after channel attention adaptive adjustment are obtained.

[0103] S300: Divide the data set D into a training data set and a test data set in proportion, input the training data set into TCAKT-E to train and optimize the model, and then use the test data set to verify the performance of TCAKT-E.

[0104] Preferably, the dividing of the data set D and training and validating TCAKT-E specifically includes:

[0105] The dataset D is divided into a training dataset and a test dataset in a ratio of 70%:30%. The 70% training dataset is input into TCAKT-E to train and optimize the model, and the 30% test dataset verifies the performance of TCAKT-E. The batch input size of the training set is set to 32, the learning rate is set to 0.001, and the expansion coefficient of the dilated convolution operation in the residual block is Set to ,and Take a positive integer, and the Dropout ratio is initially set to 0.05;

[0106] Using the cross entropy loss function Optimize the various parameters of the model by adjusting the loss function Perform back propagation and continuously optimize the weight parameters of each module in the model until convergence. The calculation formula is as follows:

[0107] L us =- 1 n ∑ i [ r t l and y t + 1- r t l and 1- y t ]

[0108] in, represents the loss function, express Real-time answer to questions, express The prediction at the moment answers the question.

[0109] S400: Use the trained and validated TCAKT-E model to predict student performance.

[0110] In this example, given a student Moment-by-moment learning interaction records and The problem information at the moment is input into the TCAKT-E model, and after the student experience representation is obtained and encoded, the student behavior vector is obtained, the multi-scale feature information is extracted, and the adaptive channel weight is adjusted, the student's Always give the correct probability prediction result based on the current question information .

[0111] The above is a detailed description of the specific implementation of the present invention in conjunction with the accompanying drawings, but the present invention is not limited to the above implementation, and various changes can be made within the knowledge of ordinary technicians in the field without departing from the purpose of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A knowledge tracking method based on temporal convolution, characterized in that: include: S100: Collect learning interaction records of multiple students and construct a learning interaction record dataset D; S200: For students, Question information, skill points, , build the temporal convolutional attention knowledge tracking model TCAKT-E, which includes: S210: Use the one-hot encoding method to encode the question information 、Answer situation Encode them separately to get the problem vector and the answer vector , splicing and Get the student's performance at time t , the calculation formula is as follows: in, Indicates students Answered Question information, Indicates students For The answer to the question information, A value of 1 indicates a correct answer. A value of 0 indicates an incorrect answer; S220: Obtaining a representation of the student experience , and Encode and get the student experience vector ; The specific calculation formula is as follows: in, From time 1 to time Students At skill points Empirical characterization on is an empirical constant representing a positive integer, From time 1 to time Students At skill points The difference between the number of correct answers and the number of incorrect answers, and The calculation formulas are as follows: in, From time 1 to time Students At skill points The number of correct answers, From time 1 to time Students At skill points The number of wrong answers; S230: Joint student response performance , Student Experience Vector Get student behavior vector , the calculation formula is as follows: in, represents a joint function; S240: Using Temporal Convolutional Attention Network TCAN on Student Behavior Data Conduct learning modeling, including: S241: Multi-scale feature information extraction: student behavior data Input to In the stacked residual blocks, the coefficients used for the dilated convolution in the stacked residual blocks increase exponentially. After different residual block processing, learning behavior feature maps of different time scales are obtained; the residual blocks are composed of dilated convolution, ReLU activation, weight normalization, and Dropout operations. After each residual block processing, a learning behavior feature map is output. ) , fusion The learning behavior characteristic diagram at different time scales is calculated as follows: in, Represents the learning behavior feature map after fusion, represents the fusion function, represents the learning behavior feature map obtained by the first residual block processing, Represents the learning behavior feature map obtained by the second residual block processing, Indicates The learning behavior feature map obtained by processing the residual blocks; S242: Output after processing by S241 The learning behavior characteristics after fusion ,right Perform global average pooling GAP operation, the calculation formula is as follows: in, Indicates height, Indicates width, Indicates the total number of channels, represents the global average pooling operation, express Middle The characteristics of each element; S243: Determine the scope of channel interaction information : According to the total number of channels The size of the adaptive channel interaction information is determined , the calculation formula is as follows: in, Indicates the number of neighboring channels that interact with the current channel. Indicates departure Recently and Large odd numbers, C Indicates the total number of channels, and yes Positive integer constants in calculation formulas; S244: Channel attention adjustment: Adaptive confirmation After that, all channels share the learned attention weight parameters. Channels Weight The calculation formula is as follows: in, represents the weight parameter shared by all channels, Representation and Adjacent A collection of channels; S300: Divide the data set D into a training data set and a test data set in proportion, input the training data set into TCAKT-E to train and optimize the model, and then use the test data set to verify the performance of TCAKT-E; S400: Use the trained and validated TCAKT-E model to predict student performance.

2. The knowledge tracking method based on time convolution according to claim 1 is characterized in that: Each learning interaction record specifically includes students, question information, answer status, and skill points.

3. The knowledge tracking method based on time convolution according to claim 1 is characterized in that: The loss function calculation formula used in training TCAKT-E is as follows: in, represents the loss function, express Real-time answer to questions, express The prediction at the moment answers the question.

Images