Knowledge tracking method and system based on cognitive decoupling
By decomposing students' learning behavior into stable cognitive patterns and random factors and combining it with the attenuated attention mechanism, the noise sensitivity and robustness problems of existing knowledge tracking models are solved, and more accurate and stable knowledge state tracking is achieved.
Patent Information
- Application Number
- CN202511270860.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-08
- Publication Date
- 2025-10-03
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing knowledge tracking models are highly dependent on data quality, prone to overfitting non-essential patterns, ignoring interference from random factors, and lacking abnormal data detection and correction mechanisms, resulting in poor robustness and unstable predictions.
A method based on cognitive decoupling is adopted to decompose students' historical learning interaction data into stable cognitive patterns and random factor components. The attenuated attention mechanism is used for time series modeling to simulate the short-term disturbance of random factors on the knowledge state, generate dynamic knowledge state representation, and predict the probability of students' correct answer to the next question.
It can effectively distinguish students' actual knowledge level from temporary interference, improve the accuracy and stability of the model in noisy environments, and enhance prediction accuracy and robustness, making it suitable for real-world educational scenarios.
Smart Images
Figure CN120746793A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of knowledge tracing, and in particular relates to a knowledge tracing method and system based on cognitive decoupling. Background Art
[0002] Knowledge tracing (KT) is a sequential prediction task that leverages students' historical learning interaction data to predict future performance. Early Bayesian knowledge tracing (BKT) approaches, based on probabilistic graphical models, modeled students' mastery of knowledge points as a Markov process, establishing independent mastery estimates for each knowledge point. With the development of deep learning, KT research has entered a new phase. For example, Deep Knowledge Tracing (DKT) first applied recurrent neural networks (RNNs) to KT tasks, dynamically representing students' knowledge states through latent vectors. Subsequently, researchers discovered the limitations of traditional RNNs in capturing long-term dependencies. Consequently, attention mechanisms, such as Self-Attentive Knowledge Tracing (SAKT) and Context-Aware Attentive Knowledge Tracing (AKT), have been introduced to further improve KT performance. These models utilize different attention modules to model the underlying relationships between questions, interactions, and student behavior, effectively improving the accuracy and personalization of knowledge tracing.
[0003] However, existing technologies still have many problems in practical applications. First, deep learning models are highly dependent on data quality and are prone to learning non-essential patterns in training data. Once there are anomalies, errors, or deviations in the training data, the model may overfit these "noises", resulting in a decrease in generalization ability. Secondly, students' learning behaviors are often affected by non-cognitive factors such as carelessness, fatigue, and stress. These random errors can cause students' historical learning records to deviate from their true knowledge status. Existing KT models generally regard student responses as a direct reflection of their cognitive level, ignoring the interference of this random factor. Finally, most current KT models lack a detection and correction mechanism for abnormal data and cannot effectively distinguish between cognitive patterns and random factors, making the model less robust in the face of noisy data, unstable prediction results, and lack of interpretability. Summary of the Invention
[0004] In order to solve the problem of how to extract cognitive models that accurately reflect students' knowledge mastery from error-prone data, the present invention proposes a knowledge tracking method and system based on cognitive decoupling, which decouples cognitive models and random factors during the modeling process, thereby improving the model's ability to accurately track students' knowledge status in a noisy environment, thereby solving the problems existing in the above-mentioned existing technologies.
[0005] To achieve the above objectives, in a first aspect, the present invention provides a knowledge tracking method based on cognitive decoupling, comprising:
[0006] Encode students' historical learning interaction data to generate question feature vectors and answer feature vectors;
[0007] Decomposing the question feature vector and the answer feature vector into a stable cognitive pattern component and a random factor component respectively;
[0008] Based on the decaying attention mechanism, the stable cognitive pattern component is temporally modeled, and based on the decaying attention mechanism, the short-term disturbance of the random factor component on the knowledge state is simulated to obtain a dynamic knowledge state representation;
[0009] Based on the dynamic knowledge state representation and the current question feature vector, the probability of the student's correct answer to the next question is predicted.
[0010] Preferably, the process of generating the question feature vector and the answer feature vector includes:
[0011] Generate a question feature vector through a question encoder based on the question number, knowledge point number and difficulty code;
[0012] According to the question number, knowledge point number, answer result and difficulty code, the answer feature vector is generated by the answer encoder.
[0013] Preferably, the formula for generating the question feature vector and the answer feature vector is:
[0014] ;
[0015] ;
[0016] in, Represents the time step question and the concepts involved The latest code; Code the difficulty of the questions; To encode,capture the changes in KC; for n-dimensional one-hot encoding; and Represent element-wise multiplication and element-wise addition operations respectively; Response status for participating students The next pair Enhanced coding; Express and Joint coding of Indicates coverage and The problem changes the encoding.
[0017] Preferably, the process of decomposing the question feature vector and the answer feature vector into a stable cognitive pattern component and a random factor component respectively includes:
[0018] Performing smoothing operations on the question feature vector and the answer feature vector respectively to obtain stable cognitive pattern components;
[0019] The stable cognitive pattern component is subtracted from the question feature vector and the answer feature vector to obtain a random factor component.
[0020] Preferably, the smoothing operation includes at least one of causal convolution, average pooling, Fourier transform and inverse Fourier reconstruction.
[0021] Preferably, the formulas for the Fourier transform and inverse Fourier reconstruction are:
[0022] ;
[0023] ;
[0024] in, For the input sequence frequency domain representation of the image; high-frequency components are selectively attenuated, while low-frequency components are retained.
[0025] Preferably, the formula of the attenuated attention mechanism is:
[0026] ;
[0027] in, is the attenuation function; 𝛿 is an adjustable hyperparameter; = Represents the problem at time t and the concepts involved Encoding Latest encoding after cognitive decoupling; for The transpose of = Indicates that the student's answer is considered at time t Afterwards, for The enhanced encoding of is the latest encoding after cognitive decoupling.
[0028] Preferably, the process of predicting the probability of the student's correct answer to the next question includes:
[0029] The dynamic knowledge state representation is concatenated with the current question feature vector and input into a two-layer fully connected neural network;
[0030] The Sigmoid function is used to output the probability that the student will answer the next question correctly.
[0031] Preferably, the output formula of the two-layer fully connected neural network is:
[0032]
[0033] in, , Represent Sigmoid and ReLU functions respectively; The vector representation of the student behavior output by the attenuated attention mechanism at time t+1; represents the problem at time t+1 and the concepts involved The encoding; 、 、 、 is a trainable parameter.
[0034] In a second aspect, the present invention further discloses a knowledge tracking system based on cognitive decoupling, for implementing the method described in the first aspect, the system comprising:
[0035] The interactive representation learning module is used to encode students' historical learning interaction data and generate question feature vectors and answer feature vectors;
[0036] a cognitive decoupling module, configured to decompose the question feature vector and the answer feature vector into a stable cognitive pattern component and a random factor component, respectively;
[0037] A knowledge state diagnosis module is used to perform time series modeling on the stable cognitive pattern component based on a decaying attention mechanism, and simulate the short-term disturbance of the random factor component on the knowledge state based on the decaying attention mechanism to obtain a dynamic knowledge state representation;
[0038] The knowledge tracking prediction module is used to predict the probability of the student's correct answer to the next question based on the dynamic knowledge state representation and the current question feature vector.
[0039] Compared with the prior art, the present invention has the following advantages and technical effects:
[0040] The present invention provides a knowledge tracking method based on cognitive decoupling, comprising: first, encoding students' historical learning interaction data to generate question feature vectors and answer feature vectors; second, decomposing the question feature vectors and answer feature vectors into stable cognitive pattern components and random factor components, respectively; third, performing time series modeling on the stable cognitive pattern components based on a decaying attention mechanism, and simulating the short-term disturbance of the random factor components on the knowledge state based on the decaying attention mechanism to obtain a dynamic knowledge state representation; finally, predicting the probability of the student's correct answer to the next question based on the dynamic knowledge state representation and the current question feature vector.
[0041] To address the technical problem that "traditional knowledge tracking models (such as DKT and SAKT) are prone to overfitting or prediction bias due to random errors in the data (such as carelessness and guesswork), the present invention decomposes learning behavior into stable cognitive patterns (knowledge mastery trends) and random factors (such as carelessness and fatigue). The present invention can effectively distinguish between students' actual knowledge level and temporary interference, reduce the impact of abnormal data on knowledge status estimation, and improve the stability of the model in real educational scenarios.
[0042] To address the technical problem that "traditional attention mechanisms (such as SAKT) are difficult to capture the disturbance of short-term random behaviors (such as fluctuations in answering caused by fatigue) on the knowledge state", the present invention adopts a decaying attention mechanism to give higher weights to adjacent time-step behaviors, simulate the short-term impact of random factors, and more accurately track the temporal changes of students' knowledge states, especially short-term fluctuations caused by temporary interference, thereby improving the granularity of prediction results.
[0043] This invention addresses existing issues such as noise sensitivity, insufficient dynamic modeling, and limited feature representation, ultimately achieving higher prediction accuracy, greater robustness, and more fine-grained knowledge state tracking. It has significant application value in real-world educational scenarios. BRIEF DESCRIPTION OF THE DRAWINGS
[0044] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of this application. The exemplary embodiments and descriptions of this application are intended to explain this application and do not constitute an improper limitation on this application. In the accompanying drawings:
[0045] Figure 1 This is a flow chart of a knowledge tracking method based on cognitive decoupling according to an embodiment of the present invention;
[0046] Figure 2 Schematic diagram of a knowledge tracking system based on cognitive decoupling according to an embodiment of the present invention. DETAILED DESCRIPTION
[0047] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.
[0048] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.
[0049] Example 1
[0050] like Figure 1 As shown, this embodiment provides a knowledge tracking method based on cognitive decoupling, including:
[0051] S1. Encode students’ historical learning interaction data to generate question feature vectors and answer feature vectors;
[0052] Furthermore, the process of generating the question feature vector and the answer feature vector includes:
[0053] Generate a question feature vector through a question encoder based on the question number, knowledge point number and difficulty code;
[0054] According to the question number, knowledge point number, answer result and difficulty code, the answer feature vector is generated by the answer encoder.
[0055] Specifically, in the context of an education system, a sequence of student interactions consists of corresponding knowledge concepts (KCs), questions associated with the KCs, and the student's answers. Due to the complex interdependencies and varying difficulty levels of these questions, the model must accurately simulate student interactions to optimize the effectiveness of the KT model. To achieve this goal, this embodiment uses a two-branch structure consisting of a question encoder and an answer encoder to meticulously encode the student's answer behavior for each question. The question encoder considers the knowledge points and difficulty level of the question, while the answer encoder further incorporates the student's answer results to effectively represent the student's individualized learning behavior characteristics.
[0056] The student interaction sequence is represented as .in represents the time step, Indicates the question number, Indicates the concept number involved in the problem, Indicates whether the student answered correctly, 0 indicates an incorrect answer, and 1 indicates a correct answer.
[0057] The student interaction representation can be expressed as:
[0058] (1)
[0059] (2)
[0060] in, Represents the time step question and the concepts involved The latest code; Code the difficulty of the questions; To encode,capture the changes in KC; for n-dimensional one-hot encoding; and Represent element-wise multiplication and element-wise addition operations respectively; Response status for participating students The next pair Enhanced coding; Express and Joint coding of Indicates coverage and The problem changes the encoding.
[0061] S2. Decomposing the question feature vector and the answer feature vector into a stable cognitive pattern component and a random factor component respectively;
[0062] Furthermore, the process of decomposing the question feature vector and the answer feature vector into a stable cognitive pattern component and a random factor component respectively includes:
[0063] Performing smoothing operations on the question feature vector and the answer feature vector respectively to obtain stable cognitive pattern components;
[0064] The stable cognitive pattern component is subtracted from the question feature vector and the answer feature vector to obtain a random factor component.
[0065] Furthermore, the smoothing operation includes at least one of causal convolution, average pooling, Fourier transform and inverse Fourier reconstruction.
[0066] Specifically, perfect historical learning interaction data is unattainable. Students' problem-solving abilities are not always fully utilized, as random factors such as carelessness, fatigue, or stress may lead to errors even on problems they have already mastered. These inevitable errors may result in data that does not accurately reflect students' true cognitive state, leading to a decline in the overall quality of the dataset. Inspired by cognitive science and educational measurement theory, this embodiment designs a cognitive decoupling module that uses a smoothing operation to divide the original learning behavior sequence into two parts: stable cognitive patterns (such as knowledge mastery and forgetting trends) and random factors (such as temporary behavioral interference such as carelessness and fatigue). Subsequently, a weighted fusion strategy is used to adaptively recombine the two parts of information, allowing the model to retain the main cognitive features while reducing the interference of abnormal factors.
[0067] First, cognitive pattern extraction. In order to separate meaningful cognitive patterns from noisy learning data, this embodiment assumes that the learning trajectory λ observed at time step λ can be expressed as:
[0068] (3)
[0069] in, To represent the cognitive characteristics of students’ knowledge mastery and smoothing behavior, is random noise caused by random factors. This decomposition allows the model to focus on capturing , which represents patterns that are structured and interpretable while minimizing the impact of noise.
[0070] Specifically, in order to extract cognitive patterns, this embodiment respectively and To apply a smoothing operation:
[0071] (4)
[0072] (5)
[0073] Smoothing can be achieved using causal convolution, average pooling layers, or Fourier transforms. Specifically, the Fourier transform decomposes the signal into the frequency domain:
[0074] (6)
[0075] in, For the input sequence Frequency domain representation of the signal. High-frequency components (typically associated with noise) can be selectively attenuated, while low-frequency components (representing meaningful cognitive patterns) are preserved. After processing in the frequency domain, the signal can be reconstructed in the time domain using an inverse Fourier transform:
[0076] (7)
[0077] This approach ensures that the reconstruction Only important cognitive patterns are retained while minimizing noise interference.
[0078] After extracting cognitive patterns, this embodiment focuses on identifying random factors in the cognitive process, such as carelessness and guesswork. To this end, this embodiment adopts a simple yet effective method to decouple random factors by subtracting the output of the cognitive pattern extractor from the original input:
[0079] (8)
[0080] (9)
[0081] In order to further capture the subtle structure of cognitive patterns, the extracted input signal can be modeled as a combination of multi-scale features:
[0082] (10)
[0083] in, Represents features at different time scales, capturing short-term sliding behavior and long-term grasping trends, is the weighting coefficient, which indicates the importance of each time scale feature. is noise, corresponding to random error. Using this multi-scale decomposition, this embodiment isolates key cognitive features from noise, enabling the model to learn global trends and local changes. Specifically, this embodiment achieves this through the following methods:
[0084] (11)
[0085] (12)
[0086] (13)
[0087] S3. Performing temporal modeling on the stable cognitive pattern component based on a decaying attention mechanism, and simulating the short-term disturbance of the random factor component on the knowledge state based on the decaying attention mechanism to obtain a dynamic knowledge state representation;
[0088] Specifically, this embodiment uses a decaying attention mechanism to model students' historical behaviors. Compared to traditional attention mechanisms, this mechanism emphasizes the influence of temporally adjacent behaviors on the current prediction, better simulating the short-term perturbations of random factors on the knowledge state, thereby achieving more accurate cognitive state estimation.
[0089] In order to effectively model the relationship between questions, knowledge points, and student answers, this embodiment adopts an attention-based structure as the building block of the knowledge state diagnosis of this embodiment. The existing attention mechanism uses the dot product attention pattern to model the relationship between input tags and is particularly good at capturing long-range dependencies in the data. However, this mechanism is not suitable for modeling students' cognitive patterns. The main reason is that the randomness retained in cognitive behavior undermines the assumption of long-term dependencies. Students' future performance is likely to be affected by random behaviors on adjacent time steps, such as carelessness and guessing. To solve this problem, this embodiment designs an attenuation-based attention mechanism that incorporates a attenuation factor based on the positional distance between inputs into the attention score. This approach allows this embodiment to more effectively simulate the randomness in cognitive behavior. Specifically, attenuation-based attention is defined as:
[0090] (14)
[0091] (15)
[0092] in, is the attenuation function; 𝛿 is an adjustable hyperparameter; = Represents the problem at time t and the concepts involved Encoding Latest encoding after cognitive decoupling; for The transpose of = Indicates that the student's answer is considered at time t Afterwards, for The enhanced encoding of is the latest encoding after cognitive decoupling.
[0093] S4. Predict the probability of the student's correct answer to the next question based on the dynamic knowledge state representation and the current question feature vector.
[0094] Furthermore, the process of predicting the probability of the student's correct answer to the next question includes:
[0095] The dynamic knowledge state representation is concatenated with the current question feature vector and input into a two-layer fully connected neural network;
[0096] The Sigmoid function is used to output the probability that the student will answer the next question correctly.
[0097] Specifically, the model uses a two-layer fully connected neural network to predict the probability of a student's next correct answer, which is used as the output of the KT task. The training process is optimized using a binary cross-entropy loss function.
[0098] In order to construct the training loss of the entire model, this embodiment designs a prediction module to predict the student's answer. This enables this embodiment to formulate a loss function to effectively train the entire model. This embodiment uses a two-layer fully connected neural network for prediction and minimizes the true answer. and predicted probability The binary cross entropy loss between them is used to optimize the prediction function:
[0099] (16)
[0100] (17)
[0101] in, , Represent Sigmoid and ReLU functions respectively; The vector representation of the student behavior output by the attenuated attention mechanism at time t+1; represents the problem at time t+1 and the concepts involved The encoding; 、 、 、 is a trainable parameter.
[0102] Beneficial effects of this embodiment:
[0103] This embodiment is based on smoothing operations, which effectively separates stable cognitive patterns from random factors in students' learning behaviors, and adaptively reconstructs features through gating mechanisms or weighted strategies, solving the problem that existing knowledge tracking models have difficulty distinguishing between "real mastery" and "temporary errors", and significantly enhancing the model's robustness and explanatory power for abnormal data.
[0104] This embodiment introduces a decaying attention mechanism that integrates location information to assign higher weights to students' adjacent behaviors, thereby strengthening the model's ability to identify short-term random behaviors. Combined with the cognitive feature decomposition structure, it can still stably predict knowledge states even when random anomalies (such as carelessness and fatigue) exist in students' answer data, effectively improving the model's predictive reliability and application value in real-world educational environments.
[0105] Example 2
[0106] Based on the same inventive concept, such as Figure 2 As shown, this embodiment further provides a knowledge tracking system based on cognitive decoupling, which is used to implement the method described in the first embodiment. The system includes:
[0107] The interactive representation learning module is used to encode students' historical learning interaction data and generate question feature vectors and answer feature vectors;
[0108] a cognitive decoupling module, configured to decompose the question feature vector and the answer feature vector into a stable cognitive pattern component and a random factor component, respectively;
[0109] A knowledge state diagnosis module is used to perform time series modeling on the stable cognitive pattern component based on a decaying attention mechanism, and simulate the short-term disturbance of the random factor component on the knowledge state based on the decaying attention mechanism to obtain a dynamic knowledge state representation;
[0110] The knowledge tracking prediction module is used to predict the probability of the student's correct answer to the next question based on the dynamic knowledge state representation and the current question feature vector.
[0111] The knowledge tracking system based on cognitive decoupling provided by this embodiment has all the advantages of the knowledge tracking method based on cognitive decoupling provided by the first embodiment.
[0112] The above are merely preferred embodiments of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.
Claims
1. A knowledge tracking method based on cognitive decoupling, characterized in that: The following steps are involved: Encode students' historical learning interaction data to generate question feature vectors and answer feature vectors; Decomposing the question feature vector and the answer feature vector into a stable cognitive pattern component and a random factor component respectively; Based on the decaying attention mechanism, the stable cognitive pattern component is temporally modeled, and based on the decaying attention mechanism, the short-term disturbance of the random factor component on the knowledge state is simulated to obtain a dynamic knowledge state representation; Based on the dynamic knowledge state representation and the current question feature vector, the probability of the student's correct answer to the next question is predicted.
2. The method according to claim 1, characterized in that The process of generating question feature vectors and answer feature vectors includes: Generate a question feature vector through a question encoder based on the question number, knowledge point number and difficulty code; According to the question number, knowledge point number, answer result and difficulty code, the answer feature vector is generated by the answer encoder.
3. The method according to claim 1, characterized in that The formula for generating the question feature vector and the answer feature vector is: ； ； in, Represents the time step question and the concepts involved The latest code; Code the difficulty of the questions; To encode,capture the changes in KC; for n-dimensional one-hot encoding; and Represent element-wise multiplication and element-wise addition operations respectively; Response status for participating students The next pair Enhanced coding; Express and Joint coding of Indicates coverage and The problem changes the encoding.
4. The method according to claim 1, wherein The process of decomposing the question feature vector and the answer feature vector into a stable cognitive pattern component and a random factor component respectively includes: Performing smoothing operations on the question feature vector and the answer feature vector respectively to obtain stable cognitive pattern components; The stable cognitive pattern component is subtracted from the question feature vector and the answer feature vector to obtain a random factor component.
5. The method according to claim 4, characterized in that The smoothing operation includes at least one of causal convolution, average pooling, Fourier transform and inverse Fourier reconstruction.
6. The method according to claim 5, characterized in that The formulas for the Fourier transform and inverse Fourier reconstruction are: ； ； in, For the input sequence frequency domain representation of the image; high-frequency components are selectively attenuated, while low-frequency components are retained.
7. The method according to claim 1, characterized in that The formula of the attenuated attention mechanism is: ； in, is the attenuation function; 𝛿 is an adjustable hyperparameter; = Represents the problem at time t and the concepts involved Encoding Latest encoding after cognitive decoupling; for The transpose of = Indicates that the student's answer is considered at time t Afterwards, for The enhanced encoding of is the latest encoding after cognitive decoupling.
8. The method according to claim 1, characterized in that The process of predicting the probability that a student will answer the next question correctly involves: The dynamic knowledge state representation is concatenated with the current question feature vector and input into a two-layer fully connected neural network; The Sigmoid function is used to output the probability that the student will answer the next question correctly.
9. The method according to claim 8, characterized in that The output formula of the two-layer fully connected neural network is: in, , Represent Sigmoid and ReLU functions respectively; The vector representation of the student behavior output by the attenuated attention mechanism at time t+1; represents the problem at time t+1 and the concepts involved The encoding; 、 、 、 is a trainable parameter.
10. A knowledge tracking system based on cognitive decoupling, characterized in that: For implementing the method according to any one of claims 1 to 9, the system comprises: The interactive representation learning module is used to encode students' historical learning interaction data and generate question feature vectors and answer feature vectors; a cognitive decoupling module, configured to decompose the question feature vector and the answer feature vector into a stable cognitive pattern component and a random factor component, respectively; A knowledge state diagnosis module is used to perform time series modeling on the stable cognitive pattern component based on a decaying attention mechanism, and simulate the short-term disturbance of the random factor component on the knowledge state based on the decaying attention mechanism to obtain a dynamic knowledge state representation; The knowledge tracking prediction module is used to predict the probability of the student's correct answer to the next question based on the dynamic knowledge state representation and the current question feature vector.