Knowledge tracking method for multi-dimensional capability construction based on learning behavior decoupling

Abstract

The application discloses a kind of multi-dimensional ability construction knowledge tracking method based on learning behavior decoupling, belong to knowledge tracking technical field.The method includes: obtaining test question embedding and test question-response embedding, learning bias behavior representation and multi-scale behavior representation are obtained based on test question-response embedding;Learning bias behavior representation and multi-scale behavior representation are input into binary state model, and knowledge construction representation is obtained;Obtain test question sequence structure representation based on test question embedding, and obtain multi-scale test question association representation by multi-scale association for test question sequence structure representation;Multi-scale test question association representation and knowledge construction representation are fused by the student ability modeling of frequency domain perception, and obtain differentiated learning ability;Student test question performance is predicted based on differentiated learning ability, and student test question performance prediction value is obtained.The method improves the accuracy of student test question performance prediction.

Description

Technical Field

[0001] This application belongs to the field of knowledge tracing technology, and in particular relates to a knowledge tracing method based on the decoupling of learning behavior and the construction of multi-dimensional capabilities. Background Technology

[0002] In the process of intelligent transformation of education driven by artificial intelligence technology, accurate tracking of students' knowledge status is a core requirement for achieving large-scale personalized teaching. In recent years, deep knowledge tracking has focused on using deep learning technology to model the dynamic evolution of students' knowledge status during problem-solving. Therefore, researching how to mine the frequency domain characteristics of learners' knowledge status fluctuations from learning interaction data, and modeling the knowledge status evolution process from both time and frequency domain dimensions, has significant research significance and application value for improving the accuracy of knowledge tracking and prediction.

[0003] Existing deep knowledge tracing models can be broadly categorized into three types: recurrent neural network (RNN)-based models, graph neural network (Graph NNN)-based models, and attention-based models. RNN-based models utilize gating mechanisms to capture temporal changes in knowledge states and enhance the interpretability of some parameters through specific designs, but they fail to adequately model complex behavioral characteristics in interactions. Graph NNN-based models excel at capturing complex relationships between questions and knowledge points, exhibiting advantages in knowledge state construction; however, these models generally neglect individual differences among students in learning speed, forgetting characteristics, and other aspects. Attention-based models attempt to simulate the forgetting process by incorporating decay mechanisms, but they overemphasize global information while neglecting local details and fail to effectively model student behavioral characteristics.

[0004] The relevant technologies mainly rely on capturing the temporal patterns of learning interactions, while ignoring the frequency domain fluctuation characteristics (such as periodic oscillations, high-frequency jumps and low-frequency slow changes) and individual differences in the process of knowledge state changes. As a result, they have a low ability to perceive fluctuations in students' knowledge state, make it difficult to fully depict the dynamic evolution of students' knowledge mastery, and have low accuracy in predicting students' future test performance. Summary of the Invention

[0005] This application aims to address at least one of the technical problems existing in the prior art. To this end, this application proposes a multi-dimensional capability construction knowledge tracking method based on learning behavior decoupling, which improves the accuracy of predicting students' test performance.

[0006] Firstly, this application provides a multi-dimensional capability construction knowledge tracing method based on learning behavior decoupling, the method comprising:

[0007] Obtain the item embedding and item-response embedding, and obtain learning bias behavior representation and multi-scale behavior representation based on the item-response embedding;

[0008] By inputting the learning bias behavior representation and the multi-scale behavior representation into the binary state model, a knowledge construction representation is obtained.

[0009] The test sequence structure representation is obtained based on the test item embedding, and the test sequence structure representation is used to obtain the multi-scale test item association representation through multi-scale association.

[0010] By fusing multi-scale test item association representations and knowledge construction representations with frequency domain perception, student ability is modeled to obtain differentiated learning ability;

[0011] Based on differentiated learning ability, students' test performance is predicted to obtain predicted values ​​for student test performance.

[0012] According to one embodiment of this application, the step of obtaining learning bias behavior representations and multi-scale behavior representations based on question-response embedding includes:

[0013] The question-response embedding is input into the monotonic attention module for feature extraction, resulting in a question-response embedding that incorporates the context.

[0014] The test question-response is embedded into the learning deviation behavior extraction module to obtain the learning deviation behavior representation in the student interaction process. The learning deviation behavior representation includes guessing representation and error representation.

[0015] The question-response embedding with fused context is input into a dual-channel multi-scale causal convolution module to filter guessed and erroneous representations, thus obtaining multi-scale behavioral representations.

[0016] According to one embodiment of this application, the step of inputting the learning bias behavior representation and the multi-scale behavior representation into a binary state model to obtain a knowledge construction representation includes:

[0017] Multi-scale behavioral representations are input into a multilayer perceptron for correlation to obtain multi-scale behavioral correlation representations.

[0018] Multi-scale behavioral association representations and learning bias behavioral representations are input into a binary state model for fusion to obtain a knowledge construction representation.

[0019] According to one embodiment of this application, the step of obtaining a test item sequence structure representation based on test item embedding, and obtaining a multi-scale test item association representation through multi-scale association, includes:

[0020] The test questions are embedded into the monotonic attention module for embedding association and fusion, resulting in a global test question association representation of the fusion context;

[0021] The global item association representation with fused context is input into a third-order Chebyshev KAN network to obtain the item sequence structure representation.

[0022] The test item sequence structure representation is input into a learnable gating network to obtain the test item gating representation;

[0023] After normalizing the gating representation of the test items, residual fusion is performed with the global test item association representation of the fusion context to obtain a multi-scale association representation.

[0024] According to one embodiment of this application, the step of modeling student abilities by fusing multi-scale test item association representations and knowledge construction representations through frequency domain perception to obtain differentiated learning abilities includes:

[0025] The multi-scale test item association representation is used as the query in the attention module, the knowledge construction representation is used as the value in the attention module, and the multi-scale test item association representation and the knowledge construction representation are used as the keys in the attention module. The input is then fed into the dual attention module for knowledge retrieval to obtain dynamic knowledge representation.

[0026] The dynamic knowledge representation is transformed from a time series to a frequency domain knowledge state through causal fast Fourier transform and then decoupled. After inverse Fourier transform, the solidified knowledge state, the periodically forgotten knowledge state, and the remembered knowledge state are obtained.

[0027] The frequency domain knowledge state representation is obtained by splicing together the solidified knowledge state, the periodically forgotten knowledge state, and the remembered knowledge state.

[0028] The dynamic knowledge representation and the frequency domain knowledge state representation are concatenated, and feature fusion and dimensionality reduction are performed through a fully connected layer to obtain a multimodal knowledge state representation.

[0029] The multimodal knowledge state representation is input into a two-parameter logic model to obtain differentiated learning capabilities.

[0030] According to one embodiment of this application, the step of inputting the multimodal knowledge state representation into a two-parameter logic model to obtain differentiated learning capability includes:

[0031] The multimodal knowledge state representation is input into the monotonic attention module, and the weights of different knowledge states are dynamically allocated through self-attention to obtain the true knowledge state.

[0032] Item discrimination is obtained based on item-response embedding;

[0033] By inputting the test item discrimination index and the actual knowledge state into a two-parameter logic model, differentiated learning ability is obtained.

[0034] According to one embodiment of this application, the step of predicting student test performance based on differentiated learning ability to obtain predicted student test performance values ​​includes:

[0035] Construct a pre-defined prediction model;

[0036] Acquire different learning abilities as datasets;

[0037] Based on the loss function, the preset prediction model is trained on the dataset to obtain the trained preset prediction model. The loss function is constructed based on the binary cross-entropy loss function and the joint loss function of item difficulty L2 regularization.

[0038] By inputting differentiated abilities into a pre-trained, pre-defined prediction model, predicted values ​​of students' test performance are obtained.

[0039] Secondly, this application provides a knowledge tracking device for constructing multi-dimensional capabilities based on learning behavior decoupling, the device comprising:

[0040] The acquisition module is used to acquire item embeddings and item-response embeddings, and to obtain learning bias behavior representations and multi-scale behavior representations based on the item-response embeddings.

[0041] The first processing module is used to input the learning bias behavior representation and multi-scale behavior representation into the binary state model to obtain the knowledge construction representation;

[0042] The second processing module is used to obtain the structural representation of the question sequence based on the question embedding, and to obtain the multi-scale question association representation by multi-scale association of the question sequence structural representation.

[0043] The third processing module is used to model student abilities by fusing multi-scale test item association representations and knowledge construction representations through frequency domain perception, thereby obtaining differentiated learning abilities.

[0044] The prediction module is used to predict students' test performance based on their differentiated learning abilities, and obtain predicted values ​​for students' test performance.

[0045] Thirdly, this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the multidimensional capability construction knowledge tracking method based on learning behavior decoupling as described in the first aspect above.

[0046] Fourthly, this application provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the multi-dimensional capability construction knowledge tracking method based on learning behavior decoupling as described in the first aspect above.

[0047] Fifthly, this application provides a chip including a processor and a communication interface, the communication interface being coupled to the processor, the processor being used to run programs or instructions to implement the multi-dimensional capability construction knowledge tracking method based on learning behavior decoupling as described in the first aspect.

[0048] Sixthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the multi-dimensional capability construction knowledge tracking method based on learning behavior decoupling as described in the first aspect above.

[0049] Additional aspects and advantages of this application 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 this application.

[0050] The present invention provides a multi-dimensional capability construction knowledge tracking method based on learning behavior decoupling, which has the following advantages over the prior art:

[0051] (1) This invention obtains question embeddings and question-response embeddings, and obtains learning deviation behavior representations and multi-scale behavior representations based on the question-response embeddings, and inputs them into a binary state model to obtain the student's knowledge construction representation. Question sequence structure representations are obtained based on question embeddings, and multi-scale question association representations are obtained through multi-scale association. By modeling student abilities through frequency domain perception fusion of multi-scale question association representations and knowledge construction representations, the differentiated learning abilities of students are obtained. Based on these differentiated learning abilities, the future performance of students can be dynamically evaluated and better predicted. This effectively simulates the evolution of students' knowledge states during problem-solving, improves the accuracy of predicting students' test performance, provides personalized learning guidance for students, optimizes teaching strategies, and enhances the learning experience.

[0052] (2) This invention embeds test questions into a monotonic attention module for embedding association and fusion. The global test question association representation with the fusion context is then input into a third-order Chebyshev KAN network to obtain a test question sequence structure representation. A learnable gating network is used to generate a test question gating representation, which, after normalization, is residually fused with the global test question association representation to obtain a multi-scale association representation. This effectively captures the structural features and multi-scale association information in the test question sequence, enhancing the accuracy and comprehensiveness of test question sequence modeling. By using Chebyshev polynomials as the basis functions of the KAN network, the extraction of complex relationships is strengthened, making the dynamic capture of associations between test questions more effective. This allows for more accurate prediction of students' learning trajectories and performance, providing more personalized learning paths and feedback.

[0053] (3) This invention performs knowledge retrieval using a dual attention module on multi-scale test item association representation and knowledge construction representation. It then combines causal fast Fourier transform to convert dynamic knowledge representation into frequency domain knowledge states and decouples them. From these states, it extracts fixed knowledge states, periodically forgotten knowledge states, and remembered knowledge states, further concatenating them to obtain frequency domain knowledge state representations. Finally, it uses a fully connected layer for feature fusion and dimensionality reduction to obtain multimodal knowledge state representations. By inputting these multimodal knowledge state representations into a two-parameter logic model, it obtains differentiated learning abilities, effectively capturing students' differentiated learning abilities under different knowledge states. By integrating knowledge state information from multiple dimensions, it improves the accuracy of predicting students' learning abilities, providing strong support for personalized teaching strategies and optimizing the design of learning paths, learning experience, and teaching effectiveness. Attached Figure Description

[0054] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0055] Figure 1 This is a flowchart illustrating the multi-dimensional capability construction and knowledge tracking method based on learning behavior decoupling provided in an embodiment of this application;

[0056] Figure 2 This is a schematic diagram of the learning behavior bias deconstruction and fusion representation module provided in the embodiments of this application;

[0057] Figure 3 This is a schematic diagram of the structure of the multi-scale temporal feature decoupling modeling module provided in the embodiments of this application;

[0058] Figure 4 This is a schematic diagram of the student ability modeling module with frequency domain awareness fusion provided in an embodiment of this application;

[0059] Figure 5 This is a schematic diagram of the structure of the knowledge tracking device for building multidimensional capabilities based on learning behavior decoupling provided in the embodiments of this application;

[0060] Figure 6 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0061] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0062] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class and the number of objects is not limited; for example, a first object can be one or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0063] The following description, in conjunction with the accompanying drawings, details the multidimensional capability construction knowledge tracking method, the multidimensional capability construction knowledge tracking device, the electronic device, and the readable storage medium provided in this application, through specific embodiments and application scenarios.

[0064] Among them, the knowledge tracking method based on the decoupling of learning behavior and the construction of multi-dimensional capabilities can be applied to the terminal, specifically by the hardware or software in the terminal.

[0065] The terminal includes, but is not limited to, portable communication devices such as mobile phones or tablets with touch-sensitive surfaces (e.g., touchscreen displays and / or touchpads). It should also be understood that, in some embodiments, the terminal may not be a portable communication device, but rather a desktop computer with touch-sensitive surfaces (e.g., touchscreen displays and / or touchpads).

[0066] The following embodiments describe a terminal including a display and a touch-sensitive surface. However, it should be understood that the terminal may include one or more other physical user interface devices such as a physical keyboard, mouse, and joystick.

[0067] The multi-dimensional capability construction knowledge tracing method based on learning behavior decoupling provided in this application embodiment can be executed by an electronic device or a functional module or entity in an electronic device that can implement the multi-dimensional capability construction knowledge tracing method based on learning behavior decoupling. The electronic devices mentioned in this application embodiment include, but are not limited to, mobile phones, tablets, computers, cameras, and wearable devices. The following uses an electronic device as the execution subject to illustrate the multi-dimensional capability construction knowledge tracing method based on learning behavior decoupling provided in this application embodiment.

[0068] The rapid development of artificial intelligence technology has accelerated the intelligent transformation of education. Knowledge tracking, as an important research direction in the field of intelligent education, utilizes artificial intelligence technology to dynamically track students' knowledge status during the learning process, thereby assisting intelligent teaching systems in implementing large-scale personalized instruction. This has become a significant trend.

[0069] In recent years, deep knowledge tracing based on deep learning technology has gradually focused on how to model changes in students' knowledge state during problem-solving. Existing mainstream deep knowledge tracing models can be broadly classified into three categories: deep knowledge tracing models based on recurrent neural networks, deep knowledge tracing models based on graph neural networks, and deep knowledge tracing models based on attention.

[0070] Deep knowledge tracing models based on recurrent neural networks primarily rely on gating mechanisms to capture the temporal changes in students' knowledge states. For example, the Progressive Knowledge Tracing (PKT) model, based on Long Short-Term Memory (LSTM) networks, decomposes the student's learning process into multiple stages and, based on Item Response Theory (IRT) and educational psychology, enhances the interpretability of parameters such as guessing and error behaviors by designing multi-level loss functions. However, this model fails to model the complex behavioral characteristics of students' learning interactions, making it difficult to accurately model the changes in students' knowledge states during the learning process.

[0071] Graph neural networks and their variants (such as GKT (Graph-based Knowledge Tracing) and SKT (Structure-based Knowledge Tracing)) have been widely used in the field of knowledge tracing due to their strong ability to capture complex relationships between questions and knowledge points, giving them a unique advantage in constructing knowledge states. GKT, by inputting student interactions and questions into a graph neural network, can automatically extract the dependencies between questions and knowledge points and has achieved excellent prediction performance on various datasets. However, graph neural network models, including GKT, ignore the individual differences between students, making it difficult to reflect the differences in knowledge mastery, learning speed, and forgetting characteristics among different students, resulting in less accurate modeling of the knowledge states of different students.

[0072] Knowledge states are dynamically evolving, meaning that students continuously learn new knowledge and forget old knowledge over time during problem-solving. Therefore, some studies have attempted to simulate forgetting characteristics and model knowledge mastery by incorporating exponential decay into attention mechanisms. Among these, context-aware attention-based knowledge tracking is widely used due to its excellent performance; however, this method focuses too much on capturing global information while neglecting the extraction of local information, and it fails to effectively model students' behavioral characteristics, leading to biases in predicting student performance.

[0073] While existing mainstream deep knowledge tracing models have made significant progress in the accuracy and interpretability of predicting students' future learning performance by modeling the temporal evolution of students' knowledge states, they still have a key limitation: they typically rely on capturing temporal patterns of learning interaction records, neglecting the frequency domain fluctuations and differential characteristics of individual students' knowledge states during the learning process. This results in a deficiency in perceiving knowledge state fluctuations, making it difficult to model the periodic oscillations, high-frequency jumps, and low-frequency gradual changes in students' knowledge states during learning, thus leading to accuracy biases in future learning performance predictions. In fact, from a frequency domain perspective, the process of students' knowledge state changes during learning is periodic and hierarchical. Therefore, researching how to extract frequency domain features from these interactive behavior data and modeling the evolution of learners' knowledge states from both the temporal characteristics of learning behavior and the frequency domain characteristics of knowledge state fluctuations is of significant research importance and application value for improving the accuracy of knowledge tracing predictions.

[0074] Figure 1 This is a flowchart illustrating the multi-dimensional capability construction and knowledge tracking method based on learning behavior decoupling provided in this application embodiment, as shown below. Figure 1 As shown, the multidimensional capability construction knowledge tracking method based on learning behavior decoupling includes steps 110, 120, 130, 140 and 150.

[0075] Step 110: Obtain the item embedding and item-response embedding, and obtain the learning bias behavior representation and multi-scale behavior representation based on the item-response embedding;

[0076] The process is straightforward: obtaining the question embedding and question-response embedding, and then using the question-response embedding to extract learning bias behavior and fuse multi-scale behavior to obtain learning bias behavior representations and multi-scale behavior representations. Learning bias behavior extraction is a method for extracting student guessing and error factors from the student's learning interaction process. By inputting the question-response embedding into a monotonic attention module based on exponential decay, a question-response embedding with fused context is obtained. This fused context question-response embedding is then passed through a fully connected layer and an activation function layer respectively to obtain student learning bias behavior representations such as guessing and errors.

[0077] Multi-scale behavior fusion is a method that uses multi-scale causal convolution to extract local information and filter out biased information. The question-response relationship of the fusion context is embedded into a dual-channel multi-scale causal convolution module. While filtering out cognitive errors, single and multiple behavioral representations are extracted separately, and then fused to obtain the student-question multi-scale behavioral representation.

[0078] Step 120: Input the learning bias behavior representation and multi-scale behavior representation into the binary state model to obtain the knowledge construction representation;

[0079] It is easy to understand that by inputting multi-scale behavioral representations and learning bias behavioral representations into a binary state model, knowledge construction representations are obtained. Knowledge construction behavior modeling is a method that uses a binary state model and two important parameters, guessing and error, to quantify the dynamic performance of students in the learning process in a fine-grained manner.

[0080] Step 130: Obtain the structural representation of the question sequence based on the question embedding, and obtain the multi-scale question association representation through multi-scale association;

[0081] It is easy to understand that test item temporal structure decoupling modeling is a method that uses attention mechanisms and Chebyshev KAN (Kolmogorov–Arnold Network) to extract contextual relationships and structural trends of test items. Test items are embedded into an input based on an exponentially decaying monotonic attention module to obtain a contextual relationship representation. This representation is then input into a 3rd-order Chebyshev KAN network. By decoupling the input features, the KAN network learns a different function for each feature and performs a weighted summation of these functions. This better fits the temporal dependencies and periodicity between test item sequences, thus better extracting the structural representation of the test item sequences.

[0082] By inputting the test item structure representation into a gated residual network, a multi-scale test item association representation of the fusion context is obtained. Test item association gated information fusion is a method that uses a gated residual network to capture the relationship between test items within a range.

[0083] Step 140: Model student abilities by fusing multi-scale test item association representations and knowledge construction representations through frequency domain perception to obtain differentiated learning abilities;

[0084] It should be noted that the student ability modeling of frequency domain perceptual fusion includes three processes: dual-attention knowledge matching, frequency domain knowledge state decoupling, and differentiated learning ability modeling. Dual-attention knowledge matching mainly utilizes a dual-attention model to retrieve knowledge states from multiple perspectives. Multi-scale test item association representations and adaptive information are input into a dual monotonic attention network, and a dynamic knowledge representation is output.

[0085] Frequency domain knowledge state decoupling is a method that extracts and reassembles different frequency features from the knowledge state using fast causal Fourier transform. The dynamic knowledge representation is input into a causal fast Fourier transform module to decouple fixed knowledge, periodically forgotten knowledge, and remembered knowledge. These three modules are then reassembled and subjected to MLP dimensionality reduction. Finally, they are concatenated with the dynamic knowledge representation for feature enhancement to obtain a multimodal knowledge state representation.

[0086] Differentiated learning ability modeling is a method that uses 2PLM (Two-Parameter Logistic Model) to transform knowledge states into ability representations. The knowledge state encoder is input into a monotonic attention module to obtain the knowledge state, and then the knowledge state is input into the 2PLM model to obtain the differentiated learning ability.

[0087] Step 150: Based on differentiated learning ability, predict students' test performance to obtain predicted values ​​for student test performance.

[0088] Finally, the differentiated learning ability is integrated with the question embedding, and the student's performance prediction on the question is obtained through feature dimensionality reduction. The final loss function is obtained through the binary cross-entropy loss function with Logits and the L2 regularization loss function of question difficulty, which together optimize the model's performance and guessing.

[0089] The multi-dimensional ability construction and knowledge tracking method based on learning behavior decoupling provided in this application obtains question embeddings and question-response embeddings. Learning deviation behavior representations and multi-scale behavior representations are obtained from the question-response embeddings and input into a binary state model to obtain the student's knowledge construction representation. Question sequence structure representations are obtained from the question embeddings, and multi-scale question association representations are obtained through multi-scale association. By fusing the multi-scale question association representations and knowledge construction representations with frequency domain perception to model student abilities, differentiated learning abilities are obtained. Based on these differentiated learning abilities, the method can dynamically evaluate and better predict students' future performance, effectively simulate the evolution of students' knowledge states during problem-solving, improve the accuracy of predicting students' test performance, provide personalized learning guidance, optimize teaching strategies, and enhance the learning experience.

[0090] In some embodiments, obtaining the learning bias behavior representation and multi-scale behavior representation based on the question-response embedding includes:

[0091] The question-response embedding is input into the monotonic attention module for feature extraction, resulting in a question-response embedding that incorporates the context.

[0092] The test question-response is embedded into the learning deviation behavior extraction module to obtain the learning deviation behavior representation in the student interaction process. The learning deviation behavior representation includes guessing representation and error representation.

[0093] The question-response embedding with fused context is input into a dual-channel multi-scale causal convolution module to filter guessed and erroneous representations, thus obtaining multi-scale behavioral representations.

[0094] It is easy to understand that by embedding the test item-response model into the learning behavior bias deconstruction and fusion representation module, learning bias behavior representations and multi-scale behavior representations are obtained. Figure 2 This is a schematic diagram of the learning behavior bias deconstruction and fusion representation module provided in the embodiments of this application, as shown below. Figure 2 As shown, the main function of the learning behavior deviation deconstruction and fusion representation module is to reconstruct students' behavior trajectories from a fine-grained perspective.

[0095] First, a monotonic attention module is used to process the question-response embedding to obtain a question-response embedding with fused context. Then, a learning bias behavior extraction module (including two MLP channels and two sigmoid functions) is used to process the question-response embedding with fused context to obtain the learning bias behavior representations of guesses and errors during student interaction. On the other hand, a multi-scale behavior fusion module (including a 1×1 convolution and a 3×3 causal convolution concatenated, then passed through an MLP and a Swish activation function) filters out these incidental information to obtain multi-scale behavior representations. Finally, a knowledge building behavior modeling module (including a normalization layer and an MLP) reorganizes these two parts of information to simulate the student's learning path, ultimately obtaining the knowledge building representation.

[0096] The computational process of deconstructing and fusing learning behavior bias representations includes: learning bias behavior extraction, multi-scale behavior fusion, and knowledge construction behavior modeling, as detailed below:

[0097] (1) Extraction of learning deviation behaviors

[0098] For time step t, the difficulty coefficient μ and the question answer difficulty embedding are used. The test questions and answer Embedding to obtain test questions - response embedding Embedding using μ and question difficulty Will Embedding to obtain test question embedding L is the question length, and D is the embedding dimension.

[0099] Then, the question-response embedding is input to the multi-head monotonic attention to obtain the question-response embedding with fused context. The calculation formula is as follows:

[0100]

[0101]

[0102]

[0103]

[0104]

[0105]

[0106] in, Let y represent the question-response embedding with fused context, and H represent the attention weights. Represents the mapping projection matrix. Let q represent the embedding dimension, k represent the query, and v represent the value. Indicates learnable parameters, The normalized weights representing the attention scores, Indicates the degree of correlation between distances. The monotonic attention score represents the incorporation of forgetting factors. The normalized weights of the monotonic attention score that incorporate the forgetting factor, where t represents the current time and τ represents any time before t.

[0107] Finally, the question-response embedding with fused context is processed through two different MLP channels to obtain the learned biased behavioral representations, guesses, and errors. The calculation formula is shown below:

[0108]

[0109]

[0110] in, This represents the Sigmoid activation function, guess means to guess, and slip means to miss. and This refers to guessing multilayer perceptrons and erroneous multilayer perceptrons.

[0111] (2) Multi-scale behavior fusion

[0112] Multi-scale behavior fusion utilizes causal convolution to filter out learned bias behaviors in the question-response embeddings of the fusion context. Therefore, a dual-channel multi-scale causal convolution module is designed for multi-scale behavior fusion. One channel undergoes a 1×1 convolution to filter out erroneous behaviors in the question-response embeddings of the fusion context, while the other channel undergoes a 3×3 causal convolution to filter out guessing behaviors in the question-response embeddings of the fusion context. The outputs of the two channels are then stacked in the third dimension, and finally, the channel features are fused through an MLP layer to obtain the multi-scale behavior representation. The calculation formula is shown below:

[0113]

[0114]

[0115]

[0116] Where W1 and W3 represent the kernel sizes, which are 1 and 3 respectively, stride represents the stride, padding represents zero padding, Y1 represents a single behavioral representation, Y3 represents multiple behavioral representations, and YC represents a joint representation. This represents the Swish activation function. This indicates the fusion of multilayer perceptrons. This represents multi-scale behavioral representations.

[0117] In some embodiments, the step of inputting the learning bias behavior representation and the multi-scale behavior representation into the binary state model to obtain the knowledge construction representation includes:

[0118] Multi-scale behavioral representations are input into a multilayer perceptron for correlation to obtain multi-scale behavioral correlation representations.

[0119] Multi-scale behavioral association representations and learning bias behavioral representations are input into a binary state model for fusion to obtain a knowledge construction representation.

[0120] It is easy to understand that although students only have intuitive information such as right or wrong in the actual process of doing questions, we cannot ignore the information of deviation behavior in the interaction process. By extracting and filtering the information of behavioral deviation, the information is input into the binary state model. This model is based on the dual processing theory in cognitive science and truly reflects the dynamic evolution process of students' learning trajectory.

[0121] First, the multi-scale behavioral representation is input into a multi-layer perceptron to obtain a multi-scale behavioral association representation. Then, the multi-scale behavioral association representation, the learning bias behavioral representation, the guesses and errors are input into a binary state model to obtain a knowledge construction representation.

[0122]

[0123]

[0124] in, This represents an associative multilayer perceptron, where O represents the knowledge construction representation. This represents a multi-scale behavioral association representation.

[0125] In this embodiment, by inputting multi-scale behavioral representations into a multilayer perceptron for correlation, and then fusing the multi-scale behavioral correlation representations with learning deviation behavioral representations into a binary state model to obtain a knowledge construction representation, the multi-dimensional behavioral information of students can be effectively integrated, improving the precision and accuracy of knowledge construction. This allows for a deeper analysis of students' learning deviations and multi-scale behavioral characteristics, enhancing the behavioral analysis and prediction capabilities during the knowledge construction process, providing strong support for personalized education and teaching, and improving learning outcomes and experience.

[0126] In this embodiment, the question-response embedding is input into a monotonic attention module for feature extraction, resulting in a context-integrated question-response embedding. This embedding is then input into a learning bias behavior extraction module to obtain guessing and error representations during student interaction. Finally, the context-integrated question-response embedding is input into a dual-channel multi-scale causal convolution module to filter these guessing and error representations, yielding multi-scale behavioral representations. This approach effectively extracts valuable information from students' learning behavior, better models students' knowledge construction and learning biases, and improves the accuracy of knowledge tracking and student learning performance prediction.

[0127] In some embodiments, obtaining a test item sequence structure representation based on test item embedding, and obtaining a multi-scale test item association representation through multi-scale association, includes:

[0128] The test questions are embedded into the monotonic attention module for embedding association and fusion, resulting in a global test question association representation of the fusion context;

[0129] The global item association representation with fused context is input into a third-order Chebyshev KAN network to obtain the item sequence structure representation.

[0130] The test item sequence structure representation is input into a learnable gating network to obtain the test item gating representation;

[0131] After normalizing the gating representation of the test items, residual fusion is performed with the global test item association representation of the fusion context to obtain a multi-scale association representation.

[0132] It is easy to understand that embedding test questions into the multi-scale temporal domain feature decoupling modeling module yields a multi-scale test question association representation. Figure 3 This is a schematic diagram of the structure of the multi-scale temporal feature decoupling modeling module provided in the embodiments of this application, as shown below. Figure 3 As shown, the multi-scale temporal feature decoupling modeling module is used to construct multi-scale relationships between test items. By extracting the associations between the global contexts of the test items, and extracting the temporal dependencies and periodicity information between the question sequences, multi-granularity association information between the question sequences is constructed. The multi-scale temporal feature decoupling modeling module consists of two sub-modules: a test item sequence structure representation modeling sub-module (e.g., a third-order Chebyshev KAN network) and a test item association gating information fusion sub-module (including a residual network, a learnable gating parameter, and a normalization layer).

[0133] Since test questions are interconnected rather than independent, extracting these connections and accurately reflecting their dependencies is crucial for accurately and efficiently predicting student learning performance. Leveraging the powerful ability of attention models to capture global contextual relationships and the robust fitting capabilities of KAN networks, and employing Chebyshev multinomials as the basis functions of the KAN network, the model's ability to extract complex relationships is enhanced, making the dynamic capture of connections between test questions more effective.

[0134] First, the test questions are embedded into the monotonic attention module of the input test questions to obtain a global test question association representation that incorporates the context. The representation is then input into a third-order Chebyshev KAN network, decoupling each input feature into individual features. Feature function fitting is then performed for each individual input feature. Finally, all feature functions are summed and passed through a dropout layer to obtain a test sequence structure representation that integrates test periodicity and test dependency. The calculation formula is shown below:

[0135]

[0136]

[0137]

[0138]

[0139] in, Denotes Chebyshev basis functions. This represents the learnable parameters, where n represents the Chebyshev order (e.g., 3), and D represents the dimension of the input vector. Let k represent the o-th output component of the Chebyshev KAN network, and k represent the structural representation of the test sequence. Represents the mapping projection matrix. The global question association representation represents the fusion context, and x represents the question embedding.

[0140] By introducing a residual network, the contextual relationships of the test items processed by monotonic attention are preserved, avoiding feature loss caused by forward processing of the model, while effectively enhancing the dependency relationships and periodic information in the test item sequence; at the same time, by utilizing a learnable gating network, the adaptability of the model to different datasets is improved, thereby improving the overall robustness of the model.

[0141] Furthermore, the structural features of the test item sequence are input into a learnable gating network to obtain the test item gating representation. The calculated value is then passed through a normalization layer, and finally, residual fusion is performed with the global item association representation obtained through monotonic attention processing to obtain a multi-scale item association representation. The calculation formula is as follows:

[0142]

[0143]

[0144] Where gate_param represents the parameters to be learned. Let represent the Sigmoid activation function, `gate` represent the gating parameter, `m` represent the multi-scale item association representation, and `k` represent the item sequence structure representation. A global question association representation that integrates context.

[0145] In this embodiment, test questions are embedded and fused into a monotonic attention module. The global test question association representation with the fused context is then input into a third-order Chebyshev KAN network to obtain a test question sequence structure representation. A learnable gating network is used to generate a test question gating representation, which, after normalization, is residually fused with the global test question association representation to obtain a multi-scale association representation. This effectively captures the structural features and multi-scale association information in the test question sequence, enhancing the accuracy and comprehensiveness of test question sequence modeling. By using Chebyshev polynomials as the basis functions of the KAN network, the extraction of complex relationships is strengthened, making the dynamic capture of associations between test questions more effective. This allows for more accurate prediction of students' learning trajectories and performance, providing more personalized learning paths and feedback.

[0146] In some embodiments, the step of modeling student abilities by fusing multi-scale test item association representations and knowledge construction representations through frequency domain perception to obtain differentiated learning abilities includes:

[0147] The multi-scale test item association representation is used as the query in the attention module, the knowledge construction representation is used as the value in the attention module, and the multi-scale test item association representation and the knowledge construction representation are used as the keys in the attention module. The input is then fed into the dual attention module for knowledge retrieval to obtain dynamic knowledge representation.

[0148] The dynamic knowledge representation is transformed from a time series to a frequency domain knowledge state through causal fast Fourier transform and then decoupled. After inverse Fourier transform, the solidified knowledge state, the periodically forgotten knowledge state, and the remembered knowledge state are obtained.

[0149] The frequency domain knowledge state representation is obtained by splicing together the solidified knowledge state, the periodically forgotten knowledge state, and the remembered knowledge state.

[0150] The dynamic knowledge representation and the frequency domain knowledge state representation are concatenated, and feature fusion and dimensionality reduction are performed through a fully connected layer to obtain a multimodal knowledge state representation.

[0151] The multimodal knowledge state representation is input into a two-parameter logic model to obtain differentiated learning capabilities.

[0152] What's easy to understand is that by fusing multi-scale test item association representations and knowledge construction representations with frequency domain perception to model student abilities, differentiated learning abilities can be obtained. Figure 4 This is a schematic diagram of the student ability modeling module with frequency domain awareness fusion provided in an embodiment of this application, as shown below. Figure 4 As shown, the main function of the frequency domain-aware fusion student ability modeling module is to construct students' ability performance from multimodal data composed of frequency and time domains. By constructing multi-angle dynamic knowledge representations, decoupling knowledge states from the frequency domain dimension, and reorganizing knowledge states in the time domain dimension, a multi-dimensional ability representation is finally obtained using the 2PLM model, and this representation is used to predict students' performance on test questions. The frequency domain-aware fusion student ability modeling module includes a dual-attention knowledge retrieval module, a frequency domain knowledge state decoupling module, and a differentiated learning ability modeling module.

[0153] (1) Dual Attention Knowledge Retrieval

[0154] The dual-attention knowledge retrieval module uses multi-scale question association representation as the query in attention and knowledge construction representation as the value of attention. Within this module, both multi-scale question association representation and knowledge construction representation are used as keys. This allows the model to consider not only the contextual representation of the questions but also the learner's question mastery when acquiring dynamic knowledge representations. This ensures that the obtained dynamic knowledge representations comprehensively and accurately model the student's knowledge state. The calculation formula is shown below:

[0155]

[0156]

[0157] in, , , , This represents the attention weight for each head. , Represents the projection mapping matrix, tm represents knowledge mapping, and tm represents dynamic knowledge representation.

[0158] (2) Frequency domain knowledge state decoupling

[0159] It should be noted that the evolution of knowledge states contains certain periodic characteristics, such as forgetting and review cycles. By using causal fast Fourier transform, the dynamic knowledge representation is transformed from a time series to a frequency domain feature, and the information that was originally a single module is transformed into a multimodal feature, which enhances the expressive power of the feature.

[0160] First, window extraction is performed on the dynamic knowledge representation. The time series is divided into sub-windows of equal length, with overlap between them to ensure the continuity of the time series. Then, frequency domain transformation and piecewise filtering are performed. Specifically, a Fast Fourier Transform is applied to each window, and a frequency band mask is used to extract low-frequency, mid-frequency, and high-frequency signals. Finally, feature aggregation is performed. The signals of each frequency band are recovered using inverse FFT, and the mean is calculated according to the dimensions of the window to obtain the frequency domain features corresponding to each time step.

[0161] The frequency domain knowledge state decoupling module decouples dynamic knowledge representation into three distinct knowledge states based on different frequency bands: solidified knowledge, periodically forgotten knowledge, and remembered knowledge. By analyzing the different components of these knowledge states, a more granular evolution process of knowledge states is obtained, improving the interpretability of knowledge tracing and the accuracy of predicting student performance. The calculation formula is shown below:

[0162]

[0163]

[0164]

[0165]

[0166] in, These are solidified knowledge, periodically forgotten knowledge, and memorized knowledge. It is a window feature. Frequency domain representation of window features Indicates Fourier transform, Indicates the inverse Fourier transform. Indicates the mask. Indicates the filtered signal. This indicates taking the real part, and N represents the window size.

[0167] After decoupling the different knowledge states of students, it is necessary to reassemble the solidified knowledge, periodically forgotten knowledge, and memorized knowledge so that the weights of solidified knowledge, periodically forgotten knowledge, and memorized knowledge can be redistributed to more accurately reflect the students' knowledge states. The calculation formula is as follows:

[0168]

[0169] in, Let G denote the GeLU activation function, and Q denote the frequency domain knowledge state representation. This indicates a multilayer perceptron.

[0170] When analyzing knowledge states, we do not rely solely on frequency domain signals for processing. Instead, we process knowledge states together from both the frequency and time domains. The features introduced by Fourier transform and the dynamic knowledge representation originally in the time domain together form multimodal knowledge state features, thereby comprehensively and accurately depicting the evolution process of knowledge states.

[0171] By concatenating dynamic knowledge representation and frequency domain knowledge representation, and then using a fully connected layer for feature fusion and dimensionality reduction, the calculation formula is as follows:

[0172]

[0173] Wherein, MU represents the multimodal knowledge state representation. This represents a multimodal knowledge state multilayer perceptron. This represents the GeLU activation function.

[0174] In some embodiments, inputting the multimodal knowledge state representation into a two-parameter logic model to obtain differentiated learning capabilities includes:

[0175] The multimodal knowledge state representation is input into the monotonic attention module, and the weights of different knowledge states are dynamically allocated through self-attention to obtain the true knowledge state.

[0176] Item discrimination is obtained based on item-response embedding;

[0177] By inputting the test item discrimination index and the actual knowledge state into a two-parameter logic model, differentiated learning ability is obtained.

[0178] The process is straightforward: multimodal knowledge state representations are input into a monotonic attention module, which processes these representations and dynamically assigns weights to each knowledge state component using self-attention, thereby capturing the student's true knowledge state. Then, the discrimination index is extracted from the question-response embedding. Finally, a two-parameter logic model is used to calculate the student's differentiated learning ability, as shown in the following formula:

[0179]

[0180]

[0181]

[0182] Where Z represents the true state of knowledge. GeLU activation function is represented, and DSC represents the item discrimination factor. This indicates that the test questions differentiate between multilayer perceptrons. For the mapping projection matrix, This is a hyperparameter, for example, 1.7. This indicates differentiated learning ability.

[0183] In this embodiment, by inputting multimodal knowledge state representations into a monotonic attention module and dynamically allocating weights for different knowledge states using self-attention, the true knowledge state is obtained. Item discrimination is calculated based on item-response embedding, and the item discrimination and the true knowledge state are input into a two-parameter logic model to model differentiated learning abilities. This approach can more accurately identify differences in students' learning abilities under different knowledge states. By dynamically adjusting the weights of knowledge states, the prediction accuracy of learning abilities is effectively improved. By considering the combination of item discrimination and student knowledge states, the characteristics of items can be more meticulously depicted, and the probability of answering questions can be dynamically assessed for different student ability levels, achieving a more accurate assessment of student abilities.

[0184] In this embodiment, knowledge retrieval is performed using a dual attention module on multi-scale question association representation and knowledge construction representation. Combined with causal fast Fourier transform, the dynamic knowledge representation is transformed into frequency domain knowledge states and decoupled. From these, fixed knowledge states, periodically forgotten knowledge states, and remembered knowledge states are extracted and further concatenated to obtain a frequency domain knowledge state representation. A fully connected layer is then used for feature fusion and dimensionality reduction to obtain a multimodal knowledge state representation. By inputting this multimodal knowledge state representation into a two-parameter logic model, differentiated learning abilities are obtained. This effectively captures students' differentiated learning abilities under different knowledge states. By integrating knowledge state information from multiple dimensions, the accuracy of predicting students' learning abilities is improved, providing strong support for personalized teaching strategies and optimizing the design of learning paths, learning experience, and teaching effectiveness.

[0185] In some embodiments, predicting student test performance based on differentiated learning ability to obtain predicted student test performance values ​​includes:

[0186] Construct a pre-defined prediction model;

[0187] Acquire different learning abilities as datasets;

[0188] Based on the loss function, the preset prediction model is trained on the dataset to obtain the trained preset prediction model. The loss function is constructed based on the binary cross-entropy loss function and the joint loss function of item difficulty L2 regularization.

[0189] By inputting differentiated abilities into a pre-trained, pre-defined prediction model, predicted values ​​of students' test performance are obtained.

[0190] The concept is easy to understand: by inputting differentiated abilities into a pre-trained predictive model, the predicted values ​​of student test performance are obtained, and the calculation formula is shown below:

[0191]

[0192]

[0193] Where, pred represents the predicted value of student test performance. It's a hyperparameter. This indicates a predictive multilayer perceptron. denoted by the true label, loss represents the loss function, and μ is the difficulty coefficient.

[0194] In this embodiment, the prediction accuracy of students' future performance is improved by using differentiated learning ability to predict students' performance on test questions. The robustness of the prediction model is improved by jointly constraining the training process of the model through the binary cross-entropy loss function and the joint loss function of L2 regularization for test difficulty.

[0195] The multi-dimensional capability construction knowledge tracking method based on learning behavior decoupling provided in this application can be executed by a multi-dimensional capability construction knowledge tracking device based on learning behavior decoupling. This application embodiment uses the execution of the multi-dimensional capability construction knowledge tracking method based on learning behavior decoupling by the multi-dimensional capability construction knowledge tracking device based on learning behavior decoupling as an example to illustrate the multi-dimensional capability construction knowledge tracking device based on learning behavior decoupling provided in this application embodiment.

[0196] This application also provides a knowledge tracking device for constructing multi-dimensional capabilities based on learning behavior decoupling, such as... Figure 5 As shown, the multi-dimensional capability construction knowledge tracking device based on learning behavior decoupling includes: an acquisition module 510, a first processing module 520, a second processing module 530, a third processing module 540, and a prediction module 550.

[0197] The acquisition module 510 is used to acquire the question embedding and the question-response embedding, and to obtain the learning bias behavior representation and multi-scale behavior representation based on the question-response embedding.

[0198] The first processing module 520 is used to input the learning bias behavior representation and the multi-scale behavior representation into the binary state model to obtain the knowledge construction representation;

[0199] The second processing module 530 is used to obtain a test item sequence structure representation based on the test item embedding, and to obtain a multi-scale test item association representation through multi-scale association of the test item sequence structure representation.

[0200] The third processing module 540 is used to model student abilities by fusing multi-scale test item association representations and knowledge construction representations through frequency domain perception, thereby obtaining differentiated learning abilities.

[0201] The prediction module 550 is used to predict students' test performance based on their differentiated learning abilities, and obtain predicted values ​​of students' test performance.

[0202] The multi-dimensional ability construction and knowledge tracking method based on learning behavior decoupling provided in this application obtains question embeddings and question-response embeddings. Learning deviation behavior representations and multi-scale behavior representations are obtained from the question-response embeddings and input into a binary state model to obtain the student's knowledge construction representation. Question sequence structure representations are obtained from the question embeddings, and multi-scale question association representations are obtained through multi-scale association. By fusing the multi-scale question association representations and knowledge construction representations with frequency domain perception to model student abilities, differentiated learning abilities are obtained. Based on these differentiated learning abilities, the method can dynamically evaluate and better predict students' future performance, effectively simulate the evolution of students' knowledge states during problem-solving, improve the accuracy of predicting students' test performance, provide personalized learning guidance, optimize teaching strategies, and enhance the learning experience.

[0203] The multi-dimensional capability construction and knowledge tracking device based on learning behavior decoupling provided in this application embodiment can achieve… Figures 1 to 4 The various processes implemented in the knowledge tracking method based on decoupling of learning behavior for building multidimensional capabilities will not be described in detail here to avoid repetition.

[0204] In some embodiments, such as Figure 6 As shown, this application embodiment also provides an electronic device 600, including a processor 601, a memory 602, and a computer program stored on the memory 602 and executable on the processor 601. When the program is executed by the processor 601, it implements the various processes of the above-described embodiment of the multidimensional capability construction knowledge tracking method based on learning behavior decoupling, and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0205] It should be noted that the electronic devices in the embodiments of this application include the mobile electronic devices and non-mobile electronic devices described above.

[0206] This application also provides a non-transitory computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the various processes of the above-described embodiments of the multi-dimensional capability construction knowledge tracking method based on learning behavior decoupling, and achieves the same technical effect. To avoid repetition, it will not be described again here.

[0207] The processor is the processor in the electronic device described in the above embodiments. The readable storage medium includes computer-readable storage media, such as computer read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0208] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described multidimensional capability construction knowledge tracking method based on learning behavior decoupling.

[0209] The processor is the processor in the electronic device described in the above embodiments. The readable storage medium includes computer-readable storage media, such as computer read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0210] This application embodiment also provides a chip, which includes a processor and a communication interface. The communication interface and the processor are coupled. The processor is used to run programs or instructions to implement the various processes of the above-described multi-dimensional capability construction knowledge tracking method embodiment based on learning behavior decoupling, and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0211] It should be understood that the chip mentioned in the embodiments of this application may also be referred to as a device-level chip, device chip, chip device, or on-chip device chip, etc.

[0212] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0213] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a computer software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, or network device, etc.) to execute the multi-dimensional capability construction knowledge tracking method based on learning behavior decoupling of the various embodiments of this application.

[0214] In the description of this application, "first feature" and "second feature" may include one or more of the features.

[0215] In the description of this application, "multiple" means two or more.

[0216] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

[0217] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0218] Although embodiments of this application have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of this application, the scope of which is defined by the claims and their equivalents.

Claims

1. A multi-dimensional capability-based knowledge tracing method based on learning behavior decoupling, characterized in that, The method includes: Obtain the item embedding and item-response embedding, and obtain learning bias behavior representation and multi-scale behavior representation based on the item-response embedding; By inputting the learning bias behavior representation and the multi-scale behavior representation into the binary state model, a knowledge construction representation is obtained. The test sequence structure representation is obtained based on the test item embedding, and the test sequence structure representation is used to obtain the multi-scale test item association representation through multi-scale association. By fusing multi-scale test item association representations and knowledge construction representations with frequency domain perception, student ability is modeled to obtain differentiated learning ability; Based on differentiated learning abilities, students' test performance is predicted to obtain predicted values ​​for student test performance. The learning bias behavior representation and multi-scale behavior representation obtained based on question-response embedding include: The question-response embedding is input into the monotonic attention module for feature extraction, resulting in a question-response embedding that incorporates the context. The test question-response is embedded into the learning deviation behavior extraction module to obtain the learning deviation behavior representation in the student interaction process. The learning deviation behavior representation includes guessing representation and error representation. The question-response embedding with fused context is input into a dual-channel multi-scale causal convolution module to filter guessed and erroneous representations, thereby obtaining multi-scale behavioral representations. The method of modeling student abilities by fusing multi-scale test item association representations and knowledge construction representations through frequency domain perception yields differentiated learning abilities, including: The multi-scale test item association representation is used as the query in the attention module, the knowledge construction representation is used as the value in the attention module, and the multi-scale test item association representation and the knowledge construction representation are used as the keys in the attention module. The input is then fed into the dual attention module for knowledge retrieval to obtain dynamic knowledge representation. The dynamic knowledge representation is transformed from a time series to a frequency domain knowledge state through causal fast Fourier transform and then decoupled. After inverse Fourier transform, the solidified knowledge state, the periodically forgotten knowledge state, and the remembered knowledge state are obtained. The frequency domain knowledge state representation is obtained by splicing together the solidified knowledge state, the periodically forgotten knowledge state, and the remembered knowledge state. The dynamic knowledge representation and the frequency domain knowledge state representation are concatenated, and feature fusion and dimensionality reduction are performed through a fully connected layer to obtain a multimodal knowledge state representation. The multimodal knowledge state representation is input into a two-parameter logic model to obtain differentiated learning capabilities.

2. The multi-dimensional capability construction knowledge tracking method based on learning behavior decoupling according to claim 1, characterized in that, The process of inputting learning bias behavior representations and multi-scale behavior representations into a binary state model to obtain knowledge construction representations includes: Multi-scale behavioral representations are input into a multilayer perceptron for correlation to obtain multi-scale behavioral correlation representations. Multi-scale behavioral association representations and learning bias behavioral representations are input into a binary state model for fusion to obtain a knowledge construction representation.

3. The multi-dimensional capability construction knowledge tracking method based on learning behavior decoupling according to claim 1, characterized in that, The process of obtaining a test sequence structure representation based on test item embedding, and then obtaining a multi-scale test item association representation through multi-scale association, includes: The test questions are embedded into the monotonic attention module for embedding association and fusion, resulting in a global test question association representation of the fusion context; The global item association representation with fused context is input into a third-order Chebyshev KAN network to obtain the item sequence structure representation. The test item sequence structure representation is input into a learnable gating network to obtain the test item gating representation; After normalizing the gating representation of the test items, residual fusion is performed with the global test item association representation of the fusion context to obtain a multi-scale association representation.

4. The multi-dimensional capability construction knowledge tracing method based on learning behavior decoupling according to claim 1, characterized in that, The step of inputting the multimodal knowledge state representation into a two-parameter logic model to obtain differentiated learning capabilities includes: The multimodal knowledge state representation is input into the monotonic attention module, and the weights of different knowledge states are dynamically allocated through self-attention to obtain the true knowledge state. Item discrimination is obtained based on item-response embedding; By inputting the test item discrimination index and the actual knowledge state into a two-parameter logic model, differentiated learning ability is obtained.

5. The multi-dimensional capability construction knowledge tracking method based on learning behavior decoupling according to claim 1, characterized in that, The method of predicting student test performance based on differentiated learning ability to obtain predicted student test performance values ​​includes: Construct a pre-defined prediction model; Acquire different learning abilities as datasets; Based on the loss function, the preset prediction model is trained on the dataset to obtain the trained preset prediction model. The loss function is constructed based on the binary cross-entropy loss function and the joint loss function of item difficulty L2 regularization. By inputting differentiated abilities into a pre-trained, pre-defined prediction model, predicted values ​​of students' test performance are obtained.

6. A multi-dimensional capability construction knowledge tracking device based on learning behavior decoupling, implemented using the multi-dimensional capability construction knowledge tracking method based on learning behavior decoupling as described in any one of claims 1 to 5, characterized in that, The device includes: The acquisition module is used to acquire item embeddings and item-response embeddings, and to obtain learning bias behavior representations and multi-scale behavior representations based on the item-response embeddings. The first processing module is used to input the learning bias behavior representation and multi-scale behavior representation into the binary state model to obtain the knowledge construction representation; The second processing module is used to obtain the structural representation of the question sequence based on the question embedding, and to obtain the multi-scale question association representation by multi-scale association of the question sequence structural representation. The third processing module is used to model student abilities by fusing multi-scale test item association representations and knowledge construction representations through frequency domain perception, thereby obtaining differentiated learning abilities. The prediction module is used to predict students' test performance based on their differentiated learning abilities, and obtain predicted values ​​for students' test performance.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the multi-dimensional capability construction knowledge tracking method based on learning behavior decoupling as described in any one of claims 1 to 5.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the multidimensional capability construction knowledge tracking method based on learning behavior decoupling as described in any one of claims 1 to 5.

Images